content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{Introduction}
The origin of the chiral fermions and the family replication has been a big mystery.
On a higher-dimensional space-time including an orbifold as an extra space,
the family unification based on a large gauge group or the structure of extra dimensions
can provide a possible
solution~\cite{BB&K,W&T,GMN1,GMN2,GLMN,GLMS,KK&O,M&N,K&M,FKMN&S,GK&M,AFKMN&S}.
Four-dimensional (4D) chiral fermions come from a higher-dimensional fermion,
after the elimination of mirror particles by orbifolding upon compactification.
The symmetry breaking mechanism on the orbifold has been originally used
in superstring theory~\cite{DHV&W1,DHV&W2}.
The family replication emerges from a few multiplets of a large gauge group
including the family group as a subgroup.
Hence, it is interesting to explore a nature of the family number
based on the orbifold family unification,
in the expectation that it offers a hint on the origin of three families
in the standard model (SM).
In this paper, we study the relationship between the family number of chiral fermions
and the Wilson line phases, based on the orbifold family unification.
We find that flavor numbers are independent of the Wilson line phases
relating extra-dimensional components of gauge bosons,
as far as the SM gauge symmetry is respected.
This feature originates from a hidden quantum-mechanical supersymmetry (SUSY).
The contents of this paper are as follows.
In Sec. II, we review a feature of the family number in the orbifold family unification,
and present a conjecture on flavor numbers.
In Sec. III, we give an example to support the conjecture
and show that it is understood from the viewpoint of the hidden SUSY.
Section IV is devoted to conclusions.
In Appendices A and B, we present several formulas concerning the combination ${}_n C_{l}$,
derived from the feature that flavor numbers are independent of the Wilson line phases.
\section{Family number in orbifold family unification}
\label{Family number in orbifold family unification}
\subsection{A feature}
\label{A feature}
In our previous work~\cite{GK&M}, we have studied the orbifold family unification
in $SU(N)$ gauge theory on 6D space-time, $M^4 \times T^2/Z_M$ $(M=2,3,4,6)$.
Here, $M^4$ is the 4D Minkowski space-time and $T^2/Z_M$ is the 2D orbifold.
We have derived enormous numbers of models with three families
of $SU(5)$ multiplets and the SM multiplets
from a pair of 6D Weyl fermions with different chiralities,
using the orbifold breaking mechanism,
after the breakdown of gauge symmetry such that
$SU(N) \rightarrow SU(5) \times SU(p_2) \times \cdots \times SU(p_n) \times U(1)^{n-1-m}$
and $SU(N) \rightarrow SU(3) \times SU(2) \times
SU(p_3) \times \cdots \times SU(p_n) \times U(1)^{n-1-m}$, respectively.
Here and hereafter, $m$ is the number of zero in $\{p_i\}$
and ``$SU(1)$'' unconventionally stands for $U(1)$, $SU(0)$ means nothing.
Through the analysis, we have found the feature that
{\it each flavor number obtained from a 6D Weyl fermion with $[N,k]$
is invariant under the change $\{p_i\}$ into $\{p'_i\}$
among the equivalent boundary conditions (BCs).}
Here, $[N, k]$ is the rank $k$ totally antisymmetric tensor representation
whose dimension is ${}_N C_k$.
Let us present several illustrations.
On $T^2/Z_2$, the numbers of 4D left-handed Weyl fermions
with the representations $\overline{\bm{5}}$ and $\bm{10}$
obtained from the
breaking pattern $SU(N) \to SU(5) \times SU(p_2) \times \cdots \times SU(p_8) \times U(1)^{7-m}$
are same as those from
$SU(N) \to SU(5) \times SU(p'_2) \times \cdots \times SU(p'_8) \times U(1)^{7-m}$,
if the following relations are satisfied,
\begin{align}
&p'_2 - p_2 = p'_7 - p_7 = p_3 - p'_3 = p_6 - p'_6~, \notag \\
&p'_4 = p_4~,~~ p'_5 = p_5~,~~ p'_8 = p_8~,
\label{equ-r1}
\end{align}
or
\begin{align}
&p'_2 - p_2 = p'_7 - p_7 = p_4 - p'_4 = p_5 - p'_5~,\notag \\
&p'_3 = p_3~,~~ p'_6 = p_6~,~~ p'_8 = p_8~,
\label{equ-r2}
\end{align}
or
\begin{align}
&p'_3 - p_3 = p'_6 - p_6 = p_4 - p'_4 = p_5 - p'_5~,\notag \\
&p'_2 = p_2~,~~ p'_7 = p_7~,~~p'_8 = p_8~.
\label{equ-r3}
\end{align}
In the same way, the flavor numbers of the SM fermions
obtained from
$SU(N) \to SU(3) \times SU(2) \times SU(p_3) \times \cdots \times SU(p_8) \times U(1)^{7-m}$
are same as those from
$SU(N) \to SU(3) \times SU(2) \times SU(p'_3) \times \cdots \times SU(p'_8) \times U(1)^{7-m}$,
if the following relations are satisfied,
\begin{align}
p'_3 - p_3 = p'_6 - p_6 = p_4 - p'_4 = p_5 - p'_5~,~~ p'_7 = p_7~,~~
p'_8 = p_8~.
\label{equ-r1-SM}
\end{align}
On $T^2/Z_3$, the numbers of 4D left-handed Weyl fermions
with $\overline{\bm{5}}$ and $\bm{10}$ obtained from
$SU(N) \to SU(5) \times SU(p_2) \times \cdots \times SU(p_9) \times U(1)^{8-m}$
are same as those from
$SU(N) \to SU(5) \times SU(p'_2) \times \cdots \times SU(p'_9) \times U(1)^{8-m}$,
if the following relations are satisfied,
\begin{eqnarray}
p'_2 - p_2 = p'_6 - p_6 = p'_7 - p_7 = p_3 - p'_3 = p_4 - p'_4 = p_8 - p'_8~,~~
p'_5 = p_5~,~~p'_9 = p_9~.
\label{equ-r1-Z3}
\end{eqnarray}
\subsection{A conjecture}
\label{A conjecture}
In the above cases with the relations (\ref{equ-r1}) -- (\ref{equ-r1-Z3}),
the BCs relating $\{p_i\}$
are connected with those relating $\{p'_i\}$
by singular gauge transformations,
and they are regarded as equivalent in the presence of the Wilson line phases
relating extra-dimensional components of gauge bosons.
This equivalence originates from the dynamical rearrangement
in the Hosotani mechanism~\cite{Hosotani1,Hosotani2,HHH&K,HH&K}.
For cases on $T^2/Z_2$, the equivalence of BCs is shown by the following relations
among the diagonal representatives
for $2 \times 2$ submatrices of $(P_0, P_1, P_2)$~\cite{K&M2},
\begin{align}
(\tau_3, \tau_3, \tau_3) \sim (\tau_3, \tau_3, -\tau_3) \sim (\tau_3, -\tau_3, \tau_3)
\sim (\tau_3, -\tau_3, -\tau_3)~,
\label{equ-Z2}
\end{align}
where $P_0$, $P_1$ and $P_2$ are representation matrices for the $Z_2$ reflections, and
$\tau_3$ is the third component of the Pauli matrices.
For case on $T^2/Z_3$, it is shown by the following relations among the diagonal representatives
for $3 \times 3$ submatrices of $(\Theta_0, \Theta_1)$~\cite{K&M2},
\begin{eqnarray}
(X, X) \sim (X, \overline{\omega} X) \sim (X, \omega X)~,
\label{equ-Z3}
\end{eqnarray}
where $\Theta_0$ and $\Theta_1$ are representation matrices for the $Z_3$ rotations,
and $X = \mathrm{diag}(1, \omega, \overline{\omega})$ with
$\omega = e^{2\pi i/3}$ and $\overline{\omega} = e^{4\pi i/3}$.
In \cite{GK&M}, we assume that the BCs are chosen as physical ones, i.e., the system with
the physical vacuum is realized with the vanishing Wilson line phases
after a suitable gauge transformation is performed.
Then, the feature is expressed by
\begin{eqnarray}
\left. N_{\bm{r}} \right|_{(\{ p_i \},a_k=0)}
= \left. N_{\bm{r}} \right|_{(\{ p'_i \},a_k=0)}~,
\label{Nr1}
\end{eqnarray}
where $N_{\bm{r}}$ is a net chiral fermion number (flavor number) for 4D fermions
with the representation $\bm{r}$ of the gauge group,
unbroken even in the presence of the Wilson line phases $(2\pi a_k)$,
and it is defined by
\begin{eqnarray}
N_{\bm{r}} \equiv n^0_{{\rm L}\bm{r}} - n^0_{{\rm R}\bm{r}}
- n^0_{{\rm L}\overline{\bm{r}}} + n^0_{{\rm R}\overline{\bm{r}}}~.
\label{Nr-def}
\end{eqnarray}
Here, $n^0_{{\rm L}\bm{r}}$, $n^0_{{\rm R}\bm{r}}$,
$n^0_{{\rm L}\overline{\bm{r}}}$ and $n^0_{{\rm R}\overline{\bm{r}}}$
are the numbers of 4D left-handed massless fermions with $\bm{r}$,
4D right-handed one with $\bm{r}$, 4D left-handed one with the complex conjugate representation
$\overline{\bm{r}}$ and 4D right-handed one with $\overline{\bm{r}}$, respectively.
Note that 4D right-handed fermion with $\overline{\bm{r}}$
and 4D left-handed one with $\bm{r}$
are transformed into each other under the charge conjugation.
On the other hand, the equivalence due to the dynamical rearrangement is expressed by
\begin{eqnarray}
\left. N_{\bm{r}} \right|_{(\{ p_i \},a_k \ne 0)}
= \left. N_{\bm{r}} \right|_{(\{ p'_i \},a_k=0)}~.
\label{Nr2}
\end{eqnarray}
From (\ref{Nr1}) and (\ref{Nr2}),
we obtain the relation,
\begin{eqnarray}
\left. N_{\bm{r}} \right|_{(\{ p_i \},a_k = 0)}
= \left. N_{\bm{r}} \right|_{(\{ p_i \},a_k \ne 0)}~,
\label{Nr3}
\end{eqnarray}
and find that each flavor number obtained from $[N,k]$ does not change
even though the vacuum changes different ones in the presence of the Wilson line phases.
In this way, we arrive at the conjecture that
{\it each flavor number in the SM
is independent of the Wilson line phases that respect the SM gauge group.}
If there were a Wilson line phase with a non-vanishing SM gauge quantum number,
(a part of) the SM gauge symmetry can be broken down.
Hence, we assume that such a Wilson line phase is vanishing or switched off.
\section{Fermion numbers and hidden supersymmetry}
\label{Fermion numbers and hidden supersymmetry}
On a higher-dimensional space-time $M^4 \times K^{D-4}$,
a massless fermion $\Psi =\Psi(x,y)$ satisfies the equation,
\begin{eqnarray}
i\Gamma^M D_M \Psi = 0~,
\label{D-eq}
\end{eqnarray}
where $K^{D-4}$ is an $(D-4)$-dimensional extra space,
$\Gamma^M$ $(M=0, 1, 2, 3, 5, \cdots, D)$ are matrices that satisfy
the Clifford algebra $\Gamma^M \Gamma^N + \Gamma^N \Gamma^M = 2 \eta^{MN}$,
$D_M \equiv \partial_M+ig A_M$ and
$\Psi$ is a fermion with $2^{[D/2]}$-components.
Here, $g$ is a gauge coupling constant, $A_M (=A_M^{\alpha} T^{\alpha})$ are gauge bosons,
and $[*]$ is the Gauss symbol.
The coordinates $x^{\mu}$ $(\mu = 0,1,2,3)$ on $M^4$
and $x^m$ $(m=5, \cdots, D)$ on $K^{D-4}$
are denoted by $x$ and $y$, respectively.
After the breakdown of gauge symmetry,
$\Psi$ is decomposed as
\begin{eqnarray}
\Psi(x, y)
= \sum_{\bm{r}_H} \sum_{\{n_i\}} \left[\psi_{{\rm L}\bm{r}_H}^{\{n_i\}}(x)
\phi_{{\rm L}\bm{r}_H}^{\{n_i\}}(y)
+ \psi_{{\rm R}\bm{r}_H}^{\{n_i\}}(x) \phi_{{\rm R}\bm{r}_H}^{\{n_i\}}(y)\right]~,
\label{Psi-decomp}
\end{eqnarray}
where $\psi_{{\rm L}\bm{r}_H}^{\{n_i\}}(x)$ and $\psi_{{\rm R}\bm{r}_H}^{\{n_i\}}(x)$
are 4D left-handed spinors and right-handed ones, respectively.
The subscript $\bm{r}_H$ stands for some representation of the unbroken gauge group $H$,
and the superscript $\{n_i\}$ represents a set of numbers relating massive modes
and those concerning components of multiplet $\bm{r}_H$.
The functions $\phi_{{\rm L}\bm{r}_H}^{\{n_i\}}(y)$ and $\phi_{{\rm R}\bm{r}_H}^{\{n_i\}}(y)$
form complete sets on $K^{D-4}$.
We define the chiral fermion number relating $\bm{r}$ as
\begin{eqnarray}
n_{\bm{r}} \equiv n^0_{{\rm L}\bm{r}} - n^0_{{\rm R}\bm{r}}~,
\label{nr-def}
\end{eqnarray}
where $\bm{r}$ is a representation of the subgroup
unbroken in the presence of the Wilson line phases.
The net chiral fermion number $N_{\bm{r}}$
is given by $N_{\bm{r}} = n_{\bm{r}} - n_{\overline{\bm{r}}}$.
In case that $n_{\bm{r}}$ is independent of the Wilson line phases $(2\pi a_k)$,
$n^0_{{\rm L}\bm{r}}$ and $n^0_{{\rm R}\bm{r}}$ must be expressed as
\begin{eqnarray}
n^0_{{\rm L}\bm{r}} = n'^0_{{\rm L}\bm{r}} + f_{\bm{r}}(a_k)~~~{\rm and}~~~
n^0_{{\rm R}\bm{r}} = n'^0_{{\rm R}\bm{r}} + f_{\bm{r}}(a_k)~,
\label{nLR}
\end{eqnarray}
respectively.
Here, $n'^0_{{\rm L}\bm{r}}$ and $n'^0_{{\rm R}\bm{r}}$ are some constants
irrelevant to $a_k$ and $f_{\bm{r}}(a_k)$ is a function of $a_k$.
\subsection{An example}
\label{An example}
Let us calculate $n^0_{{\rm L}\bm{r}}$ and $n^0_{{\rm R}\bm{r}}$,
and verify the relations (\ref{nLR}),
using an $SU(3)$ gauge theory on $M^4 \times S^1/Z_2$.
On 5D space-time, $\Psi$ is expressed as
\begin{eqnarray}
\Psi = \left( \begin{array}{c} \psi_{\rm L} \\ \psi_{\rm R} \end{array} \right)~,
\label{psi}
\end{eqnarray}
where $\psi_{\rm L}$ and $\psi_{\rm R}$ are components
containing 4D left-handed fermions and 4D right-handed ones, respectively.
The equation (\ref{D-eq}) is divided into two parts,
\begin{eqnarray}
i\overline{\sigma}^{\mu}D_{\mu} \psi_{\rm L} - D_y\psi_{\rm R} = 0~,~~
i\sigma^{\mu}D_{\mu} \psi_{\rm R} + D_y\psi_{\rm L} = 0~,
\label{5D-eqs}
\end{eqnarray}
where $D_y \equiv \partial_y + i g A_y$.
For $\psi_{\rm L}$ and $\psi_{\rm R}$, the BCs are given by
\begin{eqnarray}
&~& \psi_{\rm L}(x,-y) = \eta^0 P_0 \psi_{\rm L}(x,y)~,~~
\psi_{\rm L}(x,2\pi R-y) = \eta^1 P_1 \psi_{\rm L}(x,y)~,
\label{BC-L}\\
&~& \psi_{\rm R}(x,-y) = -\eta^0 P_0 \psi_{\rm R}(x,y)~,~~
\psi_{\rm R}(x,2\pi R-y) = -\eta^1 P_1 \psi_{\rm R}(x,y)~,
\label{BC-R}
\end{eqnarray}
where $P_0$ and $P_1$ are the representation matrices for the $Z_2$
transformation $y \to -y$
and the $Z_2$ transformation $y \to 2\pi R -y$, respectively.
$\eta^0$ and $\eta^1$ are the intrinsic $Z_2$ parities for the left-handed component.
Note that $Z_2$ parities for the right-handed one are opposite to those of the left-handed one.
For the gauge bosons, the BCs are given by
\begin{eqnarray}
&~& A_{\mu}(x,-y) = P_0 A_{\mu}(x,y)P_0^{\dagger}~,~~
A_{\mu}(x,2\pi R-y) = P_1 A_{\mu}(x,y)P_1^{\dagger}~,
\label{BC-Amu}\\
&~& A_y(x,-y) = -P_0 A_y(x,y)P_0^{\dagger}~,~~
A_y(x,2\pi R-y) = -P_1 A_y(x,y)P_1^{\dagger}~.
\label{BC-Ay}
\end{eqnarray}
We take the representation matrices,
\begin{eqnarray}
P_0 = \text{diag}(1,1,-1)~,~~
P_1 = \text{diag}(1,1,-1)~.
\label{P++-}
\end{eqnarray}
Then $SU(3)$ is broken down to $SU(2) \times U(1)$.
We consider the fermion with the representation $\bf{3}$ of $SU(3)$
and $(\eta^0,\eta^1)=(1,1)$.
Then, $\psi_{\rm L}$ and $\psi_{\rm R}$ are expanded as
\begin{eqnarray}
\psi_{\rm L}
= \left( \begin{array}{c} \textstyle\sum\limits^{\infty}_{n=0} \psi^1_{{\rm L}n}(x)\cos{\frac{n}{R}y} \\
\textstyle\sum\limits^{\infty}_{n=0} \psi^2_{{\rm L}n}(x)\cos{\frac{n}{R}y} \\
\textstyle\sum\limits^{\infty}_{n=1} \psi^3_{{\rm L}n}(x)\sin{\frac{n}{R}y}
\end{array} \right)~,~~
\psi_{\rm R}
= \left( \begin{array}{c} \textstyle\sum\limits^{\infty}_{n=1} \psi^1_{{\rm R}n}(x)\sin{\frac{n}{R}y} \\
\textstyle\sum\limits^{\infty}_{n=1} \psi^2_{{\rm R}n}(x)\sin{\frac{n}{R}y} \\
\textstyle\sum\limits^{\infty}_{n=0} \psi^3_{{\rm R}n}(x)\cos{\frac{n}{R}y}
\end{array} \right).
\label{Exp++-}
\end{eqnarray}
After a suitable $SU(2)$ gauge transformation,
the vacuum expectation value (VEV) of $A_y$ is parameterized as
\begin{equation}
\langle A_y \rangle
= \frac{-i}{gR} \left( \begin{array}{ccc} 0&0&a \\ 0&0&0 \\ -a&0&0 \end{array} \right)~,
\label{Ay++-}
\end{equation}
where $2 \pi a$ is the Wilson line phase.
From the periodicity, we limit the domain of definition for $a$ as $0 \le a < 1$.
In case with $a \ne 0$, $SU(2)$ is broken down to $U(1)$,
and then every 4D fermion becomes a singlet.
Inserting (\ref{Exp++-}) and (\ref{Ay++-}) into (\ref{5D-eqs}),
we obtain a set of 4D equations,
\begin{eqnarray}
&~& i\overline{\sigma}^{\mu}D_{\mu} \psi_{{\rm L}0}^1
- \frac{a}{R} \psi_{{\rm R}0}^3 = 0~,~~
i\sigma^{\mu}D_{\mu} \psi_{{\rm R}0}^3
- \frac{a}{R} \psi_{{\rm L}0}^1 = 0~,
\label{DLR0}\\
&~& i\overline{\sigma}^{\mu}D_{\mu} \psi_{{\rm L}0}^2 = 0~,
\label{DLR0-2}\\
&~& i\overline{\sigma}^{\mu}D_{\mu}\psi_{{\rm L}n}^1
- \frac{n}{R}\psi_{{\rm R}n}^1 - \frac{a}{R}\psi_{{\rm R}n}^3 = 0 ~~~~(n=1,2,\cdots)~,
\label{DL1} \\
&~& i\overline{\sigma}^{\mu}D_{\mu}\psi_{{\rm L}n}^2 - \frac{n}{R}\psi_{{\rm R}n}^2 = 0
~~~~~~~~~~~~~~~~~~~~~(n=1,2,\cdots)~,
\label{DL2} \\
&~& i\overline{\sigma}^{\mu}D_{\mu}\psi_{{\rm L}n}^3 + \frac{n}{R}\psi_{{\rm R}n}^3
+ \frac{a}{R}\psi_{{\rm R}n}^1 = 0
~~~~(n=1,2,\cdots)~,
\label{DL3} \\
&~& i\sigma^{\mu}D_{\mu} \psi_{{\rm R}n}^1 - \frac{n}{R}\psi_{{\rm L}n}^1
+ \frac{a}{R}\psi_{{\rm L}n}^3 = 0
~~~~(n=1,2,\cdots)~,
\label{DR1} \\
&~& i\sigma^{\mu}D_{\mu} \psi_{{\rm R}n}^2 - \frac{n}{R}\psi_{{\rm L}n}^2 =0
~~~~~~~~~~~~~~~~~~~~~(n=1,2,\cdots)~,
\label{DR2} \\
&~& i\sigma^{\mu}D_{\mu} \psi_{{\rm R}n}^3 + \frac{n}{R}\psi_{{\rm L}n}^3
- \frac{a}{R}\psi_{{\rm L}n}^1 = 0 ~~~~(n=1,2,\cdots)~.
\label{DR3}
\end{eqnarray}
Using the equations (\ref{DL1}), (\ref{DL3}), (\ref{DR1}) and (\ref{DR3}),
we derive a set of 4D equations,
\begin{eqnarray}
&~& i\overline{\sigma}^{\mu}D_{\mu}(\psi_{{\rm L}n}^1+\psi_{{\rm L}n}^3)
- \frac{n-a}{R}(\psi_{{\rm R}n}^1-\psi_{{\rm R}n}^3) = 0 ~~~~(n=1,2,\cdots)~,
\label{DL1+3} \\
&~& i\overline{\sigma}^{\mu}D_{\mu}(\psi_{{\rm L}n}^1-\psi_{{\rm L}n}^3)
- \frac{n+a}{R}(\psi_{{\rm R}n}^1+\psi_{{\rm R}n}^3) = 0 ~~~~(n=1,2,\cdots)~,
\label{DL1-3} \\
&~& i\sigma^{\mu}D_{\mu} (\psi_{{\rm R}n}^1+\psi_{{\rm R}n}^3)
- \frac{n+a}{R}(\psi_{{\rm L}n}^1-\psi_{{\rm L}n}^3) = 0 ~~~~(n=1,2,\cdots)~,
\label{DR1+3} \\
&~& i\sigma^{\mu}D_{\mu} (\psi_{{\rm R}n}^1-\psi_{{\rm R}n}^3)
- \frac{n-a}{R}(\psi_{{\rm L}n}^1+\psi_{{\rm L}n}^3) = 0 ~~~~(n=1,2,\cdots)~.
\label{DR1-3}
\end{eqnarray}
From (\ref{DLR0}), $\psi_{{\rm L}0}^1$ and $\psi_{{\rm R}0}^3$ form
a 4D Dirac fermion.
In the same way, we find that
$(\psi_{{\rm L}n}^2, \psi_{{\rm R}n}^2)$,
$(\psi_{{\rm L}n}^1 + \psi_{{\rm L}n}^3, \psi_{{\rm R}n}^1- \psi_{{\rm R}n}^3)$
and $(\psi_{{\rm L}n}^1 - \psi_{{\rm L}n}^3, \psi_{{\rm R}n}^1 + \psi_{{\rm R}n}^3)$
form 4D Dirac fermions for $n=1, 2, \cdots$
from (\ref{DL2}) and (\ref{DR2}), (\ref{DL1+3}) and (\ref{DR1-3}),
and (\ref{DL1-3}) and (\ref{DR1+3}), respectively.
.
The numbers of 4D massless fermions are evaluated as
\begin{eqnarray}
n^0_{\rm L} = 1 + \delta_{0 a}~,~~
n^0_{\rm R}= \delta_{0 a}~,
\label{nLR-ex}
\end{eqnarray}
where $\delta_{0 a}$ represents the Kronecker delta.
From (\ref{nLR-ex}), we confirm that
the fermion number $n(\equiv n^0_{\rm L}-n^0_{\rm R}= 1)$
does not depend on the Wilson line phase.
The mass spectrum for 4D fermions in this model is depicted as Figure \ref{Fig1}.
Strictly speaking, the figure describes the case with $0 < a < 1/2$.
\begin{figure}[h]
\begin{center}
\begin{picture}(350,140)
\linethickness{1pt}
\put(-10,-4){$0$}
\put(-23,26){$1/R$}
\put(-23,56){$2/R$}
\put(-23,86){$3/R$}
\put(27,110){\hbox{$\vdots$}}
\put(25,-20){$\psi_{\rm L}$}
\multiput(0,0)(0,30){4}{\line(1,0){60}}
\multiput(15,0)(0,30){4}{\circle*{5}}
\multiput(30,0)(0,30){4}{\circle*{5}}
\multiput(45,30)(0,30){3}{\circle*{5}}
\put(117,110){\hbox{$\vdots$}}
\put(115,-20){$\psi_{\rm R}$}
\multiput(90,0)(0,30){4}{\line(1,0){60}}
\multiput(105,0)(0,30){4}{\circle{5}}
\multiput(120,30)(0,30){3}{\circle{5}}
\multiput(135,30)(0,30){3}{\circle{5}}
\put(50,-45){$(a)~~a = 0$}
\put(190,-4){$0$}
\put(177,26){$1/R$}
\put(177,56){$2/R$}
\put(177,86){$3/R$}
\put(227,110){\hbox{$\vdots$}}
\put(225,-20){$\psi_{\rm L}$}
\multiput(200,0)(0,30){4}{\line(1,0){60}}
\multiput(200,0)(0,30){3}{\multiput(0,23)(6,0){10}{\line(1,0){4}}}
\multiput(200,0)(0,30){4}{\multiput(0,7)(6,0){10}{\line(1,0){4}}}
\multiput(215,0)(0,30){4}{\circle*{5}}
\multiput(245,23)(0,30){3}{\circle*{5}}
\multiput(230,7)(0,30){4}{\circle*{5}}
\put(350,4){\footnotesize $a/R$}
\put(350,20){\footnotesize $(1-a)/R$}
\put(350,34){\footnotesize $(1+a)/R$}
\put(350,50){\footnotesize $(2-a)/R$}
\put(350,64){\footnotesize $(2+a)/R$}
\put(350,80){\footnotesize $(3-a)/R$}
\put(350,94){\footnotesize $(3+a)/R$}
\put(317,110){\hbox{$\vdots$}}
\put(315,-20){$\psi_{\rm R}$}
\multiput(290,0)(0,30){4}{\line(1,0){60}}
\multiput(290,0)(0,30){3}{\multiput(0,23)(6,0){10}{\line(1,0){4}}}
\multiput(290,0)(0,30){4}{\multiput(0,7)(6,0){10}{\line(1,0){4}}}
\multiput(305,30)(0,30){3}{\circle{5}}
\multiput(335,23)(0,30){3}{\circle{5}}
\multiput(320,7)(0,30){4}{\circle{5}}
\put(250,-45){$(b)~~a \ne 0$}
\end{picture}
\end{center}
\abovecaptionskip=40pt
\caption{Mass spectrum of 4D fermions.
The filled circles and the open ones represent left-handed fermions
and right-handed ones, respectively.}
\label{Fig1}
\end{figure}
\subsection{Hidden quantum-mechanical supersymmetry}
\label{Hidden quantum-mechanical supersymmetry}
We explore a physics behind the feature
that the fermion numbers are independent of the Wilson line phases.
From Figure \ref{Fig1},
we anticipate that the feature originates from a hidden quantum-mechanical SUSY.
Here, the quantum-mechanical SUSY means the symmetry generated by the supercharge
$Q$ that satisfies the algebraic relations~\cite{Witten1,CK&S},
\begin{eqnarray}
H = Q^2~,~~ \left\{ Q , (-1)^F \right\} = 0~,~~ \left((-1)^F\right)^2 = I~,
\label{SUSY}
\end{eqnarray}
where $H$, $F$ and $I$ are the Hamiltonian, the $\lq\lq$fermion'' number operator
and the identity operator, respectively.
The eigenvalue of $(-1)^F$ is given by $+1$ for $\lq\lq$bosonic'' states
and $-1$ for $\lq\lq$fermionic'' states,
and $\text{Tr}~(-1)^F$ is a topological invariant, called the Witten index~\cite{Witten2}.
It is known that the system with 4D fermions has the hidden SUSY
where the 4D Dirac operator plays the role of $Q$~\cite{Al,F&W}.
The correspondences are given by
\begin{eqnarray}
Q \leftrightarrow i\gamma^{\mu}D_{\mu} = \left( \begin{array}{cc} 0 & i{\sigma}^{\mu}D_{\mu} \\
i\overline{\sigma}^{\mu}D_{\mu} & 0 \end{array} \right)~,~~
(-1)^F \leftrightarrow \gamma_5~,
\label{SUSY-Dirac}
\end{eqnarray}
where $\gamma_5$ is the chirality operator defined by
$\gamma_5 \equiv i \gamma^0 \gamma^1 \gamma^2 \gamma^3$.
The trace of $\gamma_5$ is the index of the 4D Dirac operator, and
the following relations hold,
\begin{eqnarray}
&~& \left. \text{Tr}~\gamma_5\right|_{\bm{r}}
= n^0_{{\rm R}\bm{r}}[A_{\mu}] - n^0_{{\rm L}\bm{r}}[A_{\mu}]
= \dim \ker {\sigma}^{\mu}D_{\mu}|_{\bm{r}}
- \dim \ker \overline{\sigma}^{\mu}D_{\mu}|_{\bm{r}}
\nonumber \\
&~& ~~~~~~~~~~~~~~ = \frac{1}{32\pi^2}\int \text{tr}_{\bm{r}} \epsilon_{\mu\nu\alpha\beta}
F^{\mu\nu}F^{\alpha\beta}d^4x~,
\label{ASindex}
\end{eqnarray}
from the Atiyah-Singer index theorem.
Here, $n^0_{{\rm R}\bm{r}}[A_{\mu}]$ and $n^0_{{\rm L}\bm{r}}[A_{\mu}]$
are the numbers of
normalizable solutions (massless fermions)
satisfying $ i{\sigma}^{\mu}D_{\mu} \psi_{{\rm R}\bm{r}} = 0$
and $ i\overline{\sigma}^{\mu}D_{\mu} \psi_{{\rm L}\bm{r}} = 0$, respectively.
Note that massive fermions exist in pairs ($\psi_{{\rm R}\bm{r}}$ and $\psi_{{\rm L}\bm{r}}$)
and do not contribute to the index.
The integral quantity in (\ref{ASindex}) is called the Pontryagin number,
and it is deeply connected to the configuration of gauge bosons $A_{\mu}$ on 4D space-time.
It is pointed out that higher-dimensional theories with extra dimensions also
possess the hidden SUSY~\cite{LNS&S,S}.
In the system with a 5D fermion, the Dirac operator relating the fifth-coordinate
plays the role of $Q$ and there are the correspondences,
\begin{eqnarray}
Q \leftrightarrow \tilde{D}_y = \left( \begin{array}{cc} 0 & D_y \\ -D_y & 0 \end{array} \right)~,~~
(-1)^F \leftrightarrow \tilde{\Gamma} \equiv \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right)~.
\label{SUSY-Dy}
\end{eqnarray}
Note that $\tilde{\Gamma} = -\gamma_5$.
The counterpart of the Witten index is given by
\begin{eqnarray}
\left. \text{Tr}~\tilde{\Gamma}\right|_{\bm{r}}
= \tilde{n}^0_{{\rm R}\bm{r}}(a) - \tilde{n}^0_{{\rm L}\bm{r}}(a)~,
\label{TrGamma}
\end{eqnarray}
where $\tilde{n}^0_{{\rm R}\bm{r}}(a)$ and $\tilde{n}^0_{{\rm L}\bm{r}}(a)$ are
the numbers of eigenfunctions,
that satisfy the equations,
\begin{eqnarray}
\tilde{D}_y \left( \begin{array}{c} 0 \\ \psi_{\rm R} \end{array} \right)
= \left( \begin{array}{c} D_y\psi_{\rm R} \\ 0 \end{array} \right)
= \left( \begin{array}{c} 0 \\ 0 \end{array} \right)
\label{DyR}
\end{eqnarray}
and
\begin{eqnarray}
\tilde{D}_y \left( \begin{array}{c} \psi_{\rm L} \\ 0 \end{array} \right)
= \left( \begin{array}{c} 0 \\ -D_y \psi_{\rm L} \end{array} \right)
= \left( \begin{array}{c} 0 \\ 0 \end{array} \right)~,
\label{DyL}
\end{eqnarray}
respectively.
Note that the eigenvalue equations are given by $D_y \psi_{\rm R} = \lambda \psi_{\rm R}$
and $D_y \psi_{\rm L} = \lambda' \psi_{\rm L}$,
eigenfunctions with non-zero eigenvalues exist in pairs,
which correspond to 4D massive fermions as seen from (\ref{5D-eqs}),
and they do not contribute to the index.
From the equations (\ref{5D-eqs}),
there is a one-to-one correspondence such that
\begin{eqnarray}
D_y \psi_{\rm R} = 0 \leftrightarrow i \overline{\sigma}^{\mu} D_{\mu} \psi_{\rm L} = 0~,~~
D_y \psi_{\rm L} = 0 \leftrightarrow i \sigma^{\mu} D_{\mu} \psi_{\rm R} = 0~.
\label{Dy-SlashD}
\end{eqnarray}
Let us generalize to a system with a fermion on a higher-dimensional space-time.
For the case that $D=2n$ $(n=3, 4, \cdots)$,
the correspondences are given by
\begin{eqnarray}
Q \leftrightarrow \tilde{D} \equiv \sum_{m=5}^{D} i \Gamma^m D_m~,~~
(-1)^F \leftrightarrow \tilde{\Gamma} \equiv -\Gamma_{D+1}~,
\label{SUSY-Dm-even}
\end{eqnarray}
where $\Gamma_{D+1}$ is the chirality operator defined by
$\displaystyle{\Gamma_{D+1} = (-i)^{n+1} \Gamma^0 \Gamma^1 \cdots \Gamma^D}$.
For the case that $D=2n+1$ $(n=2, 3, \cdots)$,
the correspondences are given by
\begin{eqnarray}
Q \leftrightarrow \tilde{D} \equiv U^{\dagger} \sum_{m=5}^{D} i \Gamma^m D_m U~,~~
(-1)^F \leftrightarrow \tilde{\Gamma} \equiv i \Gamma^{D}~,
\label{SUSY-Dm-odd}
\end{eqnarray}
where $U$ is the unitary matrix that satisfies the relation
$i\Gamma^{D} = U^{\dagger} \Gamma^1 U$,
and $i \Gamma^{D}$ is a diagonal matrix with the same form as the chirality operator
on $D(=2n)$-dimensions up to a sign factor.
The equation (\ref{D-eq}) is written by
\begin{eqnarray}
i\Gamma^{\mu} D_{\mu} \Psi + \sum_{m=5}^{D} i\Gamma^m D_m \Psi = 0~.
\label{D-eqs}
\end{eqnarray}
For the case that $D=2n+1$, after the unitary transformation
$\Gamma'^M = U^{\dagger} \Gamma^M U$ and $\Psi' = U^{\dagger} \Psi$
is performed, $\Gamma'^M$ and $\Psi'$ are again denoted as $\Gamma^M$ and $\Psi$
in (\ref{D-eqs}).
The counterpart of the Witten index is given by
\begin{eqnarray}
\left. \text{Tr}~\tilde{\Gamma}\right|_{\bm{r}}
= \tilde{n}^0_{{\rm R}\bm{r}}(a_k) - \tilde{n}^0_{{\rm L}\bm{r}}(a_k)~,
\label{TrGamma-ak}
\end{eqnarray}
where $\tilde{n}^0_{{\rm R}\bm{r}}(a_k)$ and $\tilde{n}^0_{{\rm L}\bm{r}}(a_k)$
are the numbers of eigenfunctions,
that satisfy
$\tilde{D} \psi_{\rm R} = 0$ and $\tilde{D} \psi_{\rm L} = 0$, respectively.
From (\ref{D-eqs}), there is a one-to-one correspondence such that
\begin{eqnarray}
\tilde{D} \psi_{\rm R} = 0 \leftrightarrow i \Gamma^{\mu} D_{\mu} \psi_{\rm L} = 0~,~~
\tilde{D} \psi_{\rm L} = 0 \leftrightarrow i \Gamma^{\mu} D_{\mu} \psi_{\rm R} = 0~.
\label{tildeD-SlashD}
\end{eqnarray}
Here $\psi_{\rm R}$ and $\psi_{\rm L}$ are a 4D right-handed spinor component
and a 4D left-handed one in $\Psi$,
that are eigenspinors of the 4D chirality operator
$\Gamma_5 \equiv i \Gamma^0 \Gamma^1 \Gamma^2 \Gamma^3$
whose eigenvalues are $1$ and $-1$, respectively.
Note that components with a different 4D chirality involve each other
through the equation (\ref{D-eqs}),
because $\Gamma_5$
is anti-commutable to $i\Gamma^{\mu} D_{\mu}$
but it is commutable to $\tilde{D}$.
From (\ref{tildeD-SlashD}),
the following relations hold,
\begin{eqnarray}
\tilde{n}^0_{{\rm R}\bm{r}}(a_k) = n^0_{{\rm L}\bm{r}}~,~~
\tilde{n}^0_{{\rm L}\bm{r}}(a_k) = n^0_{{\rm R}\bm{r}}~,
\label{tilde-n=n}
\end{eqnarray}
and, using (\ref{tilde-n=n}), we derive the relation,
\begin{eqnarray}
\left. \text{Tr}~\tilde{\Gamma}\right|_{\bm{r}}
= \tilde{n}^0_{{\rm R}\bm{r}}(a_k) - \tilde{n}^0_{{\rm L}\bm{r}}(a_k)
= n^0_{{\rm L}\bm{r}} - n^0_{{\rm R}\bm{r}}~.
\label{TrGamma-2}
\end{eqnarray}
Because $\displaystyle{\left. \text{Tr}~\tilde{\Gamma}\right|_{\bm{r}}}$ is a topological invariant
and the Wilson line phases determine the vacuum with $\langle F_{mn} \rangle = 0$
globally in our orbifold family unification models,
$n_{\bm{r}} (= n^0_{{\rm L}\bm{r}} - n^0_{{\rm R}\bm{r}})$
is independent of the Wilson line phases.
Hence, $N_{\bm{r}} (= n_{\bm{r}} - n_{\overline{\bm{r}}})$
is also independent of the Wilson line phases.
Finally, we give a comment on $\displaystyle{\left. \text{Tr}~\tilde{\Gamma}\right|_{\bm{r}}}$.
As seen from the Atiyah-Singer index theorem relating the Dirac operator for extra-dimensions,
fermion numbers are deeply connected to the topological structure on $K^{D-4}$
including the configurations of $A_m$ on $K^{D-4}$.
From this point of view, the family number has been studied
in the Kaluza-Klein theory~\cite{Witten-KK}
and superstring theory~\cite{CHS&W}.
\section{Conclusions}
\label{Conclusions}
We have studied the relationship between the family number of chiral fermions
and the Wilson line phases, based on the orbifold family unification.
We have found that flavor numbers are independent
of the Wilson line phases relating extra-dimensional components of gauge bosons,
as far as the SM gauge symmetry is respected.
This feature originates from a hidden quantum-mechanical SUSY.
From our observation, the previous analyses~\cite{KK&O,K&M,GK&M},
based on the assumption that the BCs are physical ones,
are justified in the orbifold family unification.
Concretely,
even if the BCs are not physical, we can obtain the same result
as that of the physical ones,
because the family number is invariant under the change
from the original BCs to the physical ones by singular gauge transformations.
\section*{Acknowledgments}
This work was supported in part by scientific grants from the Ministry of Education, Culture,
Sports, Science and Technology under Grant Nos.~22540272 and 21244036 (Y.K.).
|
\section{Introduction}
Comets are the most pristine remnants of the formation of the
solar system 4.6 billion years ago. Investigating the
composition of cometary nuclei ices provides clues to the physical
conditions and chemical processes at play in the primitive solar
nebula. Comets may also have played a role in the delivery of
water and organic material to the early Earth \citep{Har11}.
The recent years have seen significant improvement in the sensitivity
and spectral coverage of millimetre receivers. The EMIR
receivers \citep{Car12} at the Institut de radioastronomie millim\'etrique
(IRAM) are equipped with a fast Fourier-transform
spectrometer that offers a wide frequency coverage at a high
spectral resolution (0.2~MHz). The combination enables sensitive
spectral surveys of cometary atmospheres and searches for
complex molecules through their multiple rotational lines in the 1~mm band.
We report here the detection of ethylene glycol
(CH$_2$OH)$_2$ and formamide (NH$_2$CHO) in comets C/2012~F6
(Lemmon) (hereafter referred to as Lemmon) and C/2013~R1
(Lovejoy) (hereafter Lovejoy), for the first time since their discovery in
comet C/1995 O1 (Hale-Bopp) \citep{Cro04a,Boc00}. We also present
the detection of CH$_3$CHO and HCOOH in comet Lovejoy. Comet
Lemmon is a long-period comet \citep[initial orbital period of
9800 yr;
[]{MPEC2014E18}, which reached
perihelion at 0.731 AU on 24 March 2013. Comet Lovejoy passed
perihelion on 22 December 2013 at 0.812 AU from the Sun and
also is a high-inclination long-period comet originating from the
Oort Cloud \citep[initial orbital period of 7000 yr;]
[]{MPEC2014E18}. Comets Lemmon and Lovejoy became
unexpectedly bright naked-eye comets, reaching visual magnitudes of
4.5 and 4.8, and displayed high water production
rates near perihelion of 1$\times10^{30}$ and $1.5\times10^{29}$
\mbox{molec.~s$^{-1}$} \citep[][Crovisier et al., personal communication]{Com14}.
\section{Observations}
Data on comet Lemmon were acquired with the IRAM 30-m telescope on March
14--18 and April 6--8, 2013, when favourable weather conditions enabled
observations. Although one of the most productive
comets of the past years, comet Lemmon was relatively far from
Earth near perihelion ($\Delta$ = 1.5 AU).
Comet Lovejoy was discovered on 7 September 2013 \citep{Lov13},
which was two and a half months before its optimal observing conditions
around perigee at $\Delta$ = 0.397 AU on 19 November. It was
observed during three periods: 8--12 November, 27 November--1 December,
and 9--16 December, 2013, under average to good weather
conditions. Additional observations obtained on November 13 and 16,
which focused on the search for phosphine, were reported by
\citet{Agu14}. During this period, the water production rate of
comet Lovejoy increased from $\sim5$ to $\sim11\times 10^{28}$
\mbox{molec.~s$^{-1}$}~according to Nan\c{c}ay OH observations (Crovisier et al.,
personal communication). The best sensitivity was reached at the
end of November when the comet was still close to Earth and
more active than two weeks earlier.
For the two comets, we surveyed most of the 1~mm band from 210 to
272~GHz with five different double-sideband tunings. Each
tuning covers $2\times8$~GHz in two linear
polarizations, with the upper 8~GHz sideband separated by 8~GHz
from the lower sideband. The 166--170~GHz range was also
observed in both comets. The comets were tracked with the latest
available orbital elements, pointing was checked once every 1-2
hours, and residual pointing offsets were estimated from coarse maps
of the strongest cometary lines. The HCN $J$(3--2) line
at 265.886~GHz and methanol lines present in all tuning setups
were observed regularly to monitor the activity and track the
location of the peak
of gas emissions in the coma. Table~\ref{tabobsefe} provides the
geometric circumstances of the observations together with
reference coma and outgassing parameters derived from
observations.
\begin{table*}
\caption[]{Observing circumstances and reference
parameters.}\label{tabobsefe}\vspace{-0.2cm} \centering
\begin{tabular}{lccccccc}
\hline\hline\noalign{\smallskip}
UT date & $<r_{h}>$ & $<\Delta>$ & $v_{exp}$ & $T_{kin}$ & $Q_{\rm H_2O}$ & $Q_{\rm HCN}$ & $Q_{\rm CH_3OH}$ \\
$($aaaa/mm/dd.d--dd.d) & (AU) & (AU) & (\mbox{km~s$^{-1}$}) & (K) & (\mbox{molec.~s$^{-1}$}) & (\mbox{molec.~s$^{-1}$}) & (\mbox{molec.~s$^{-1}$}) \\
\hline\noalign{\smallskip}
\multicolumn{8}{l}{Comet C/2012~F6 (Lemmon):} \\
2013/04/06.4--08.6 & 0.78 & 1.60 & 1.00 & 100 & $\sim9\times10^{29}$$^a$ & $12.7\pm0.9\times10^{26}$ & $14.6\pm1.1\times10^{27}$ \\
\multicolumn{8}{l}{Comet C/2013~R1 (Lovejoy):} \\
2013/11/08.2--12.4 & 1.13 & 0.47 & 0.76 & 55 & $0.5\times10^{29}$$^b$ & $7.4\pm0.1\times10^{25}$ & $1.4\pm0.1\times10^{27}$ \\
2013/11/27.5--31.6 & 0.92 & 0.47 & 0.90 & 65 & $0.8\times10^{29}$$^b$ & $12.7\pm0.1\times10^{25}$ & $2.2\pm0.2\times10^{27}$ \\
2013/12/09.5--16.6 & 0.83 & 0.71 & 0.93 & 80 & $1.1\times10^{29}$$^b$ & $18.0\pm0.1\times10^{25}$ & $2.9\pm0.2\times10^{27}$ \\
\hline
\end{tabular}\vspace{-0.2cm}
\tablefoot{$^{(a)}$ From \citet{Com14} and in agreement with $Q_{\rm CH_3OH}$,
and the values of $Q_{\rm H_2O}$ and $Q_{\rm CH_3OH}$/$Q_{\rm
H_2O}$ measured by \citet{Pag14}. $^{(b)}$ From contemporaneous
Nan\c{c}ay OH observations (Crovisier et al., personal com.).}
\end{table*}
\begin{figure}[h]
\sidecaption
\includegraphics*[width=7.0cm,angle=0]{figglycol2013.ps}
\caption{Weighted average of the 9--13 strongest {\it aGg'} (CH$_2$OH)$_2$
lines observed in the 213--272 GHz range (online Table~\ref{tabglycol}):
(a) C/2012 F6 (Lemmon), 6--8 April 2013, (b) C/2013~R1 (Lovejoy),
27 Nov.--1 Dec. 2013; (c) Lovejoy, 9--16 Dec. 2013. The vertical
scale is the main beam brightness temperature and the
horizontal scale is the Doppler velocity in the comet frame.
\vspace{-0.5cm}} \label{figglycol}
\end{figure}
\begin{figure}[h]
\includegraphics*[width=6.3cm,angle=270]{fignh2cho2013.ps}
\caption{Weighted average of the 5--10 strongest NH$_2$CHO $J$(11--10) and
$J$(12--11) lines (online Table~\ref{tabformamide}). (a) C/2012~F6 (Lemmon),
6--8 April 2013; (b) C/2013~R1 (Lovejoy), 8--12 Nov. 2013; (c)
C/2013~R1 (Lovejoy), 27 Nov.--1 Dec. 2013; (d) C/2013~R1
(Lovejoy), 9--16 Dec. 2013. Scales as in Fig.\ref{figglycol}.}
\label{fignh2cho}
\end{figure}
\section{Data analysis}
\label{sec:reduction}
Spectra were corrected for the main beam efficiency (0.55
to 0.44 in the 210--272 GHz range).
Calibration was regularly checked on reference line sources (W3OH,
W51D and IRC+10216) and the main beam efficiency was checked
several times per day on planets (e.g., Mars, Uranus). Losses due
to bad focus (up to 20\% during some daytime observations) and
elevation dependence of the antenna gain were corrected. The beam
size varies between 11.7\arcsec~and 9.1\arcsec~from 210 to
272~GHz, which corresponds to $\approx$12\,000 and 3\,000--6\,000~km
at the distances of comets Lemmon and Lovejoy.
Several individual lines of ``classical'' molecules such as HCN, HNC,
CH$_3$CN, CH$_3$OH, H$_2$CO, CS, and H$_2$S (and CO in Lovejoy) were
clearly detected \citep{Biv13,Agu14}. To search for
molecules with weak signatures, we averaged the multiple lines
present in the covered frequency range, focusing on the
strongest lines with similar expected intensities. As far as
possible, only lines for which the noise level
was similar were considered. Intensities were computed following
the models of \citet{Biv99,Biv00,Biv06,Biv11}, using the coma
temperatures and velocities given in Table~\ref{tabobsefe}. For
comet Lemmon, the 240--272~GHz range was much less noisy, while
most of the 210--272~GHz range could be used for Lovejoy. Several
molecules previously identified in a few comets were detected using
this method, e.g., HNCO, OCS, HC$_3$N, SO, and SO$_2$ in comet
Lemmon, and HNCO and HC$_3$N in comet Lovejoy
\citep[e.g.,][and {\it in preparation}]{Biv13}.
Here, we focus on the detection of ethylene glycol
(CH$_2$OH)$_2$ in its lowest energy conformer {\it aGg'}, and of
formamide (NH$_2$CHO) in the two comets. Spectra showing the
average of 5 to 13 lines are shown in
Figs.~\ref{figglycol}--\ref{fignh2cho}. Table~\ref{tabqprod}
lists average line intensities and inferred production rates.
The detailed information on the lines considered in the analysis
is given in online Tables~\ref{tabglycol}--\ref{tabformamide}.
Excitation processes considered in
calculating the production rates of these two molecules are
collisional excitation (with neutrals and electrons)
and spontaneous decay. Infrared pumping by
Sun radiation is not considered because band strengths are not
available in the literature. Generally, infrared pumping tends
to decrease the effect of spontaneous radiative decay, but for the present
observations the difference with full LTE is very small.
The gas kinetic temperature
(Table~\ref{tabobsefe}) was estimated from the series of methanol
rotational lines around 251~GHz: 18 lines sampling energy levels
with $E_{\rm up}$ between 60 and 240~K, and 28 lines with $E_{\rm
up}$ in the range 60--360~K were detected in comets Lovejoy and
Lemmon. The retrieved kinetic temperatures
do not differ significantly (by more than 10\%), whether derived from other
series of methanol lines observed between 165 and 305 GHz, from
methanol lines at offset positions ($\approx$7\arcsec),
or from other molecules (H$_2$S, CS). Values derived from
251~GHz and 165~GHz methanol lines in comet Lovejoy agree within 3~K.
The gas expansion velocity
(Table~\ref{tabobsefe}) is deduced from the line width of the
strongest emission lines (e.g., HCN $J$(3--2), CS $J$(5--4) and
some strong CH$_3$OH lines).
Section~\ref{other} presents the detection of HCOOH and CH$_3$CHO and
searches for other CHO-bearing molecules.
\begin{table*}
\caption[]{Average line intensities and production rates of ethylene glycol,
formamide, acetaldehyde, and formic acid.}\label{tabqprod}\vspace{-0.2cm} \centering
\setlength{\tabcolsep}{0.1mm}
\begin{tabular}{lllccccccccc}
\hline\hline\noalign{\smallskip}
Molecule & Freq. & Lines$^a$ & \multicolumn{2}{c}{C/2012 F6 (Lemmon)} & \phantom{0} & & & \multicolumn{2}{c}{C/2013 R1 (Lovejoy)} && \\
\cline{4-5} \cline{7-12} \\[-0.2cm]
& range & & \multicolumn{2}{c}{6--8 Apr.} && \multicolumn{2}{c}{8--12 Nov.} & \multicolumn{2}{c}{27--31 Nov.} & \multicolumn{2}{c}{9--16 Dec.} \\
& & & $I$$^b$ & $Q$ & & $I$$^b$ & $Q$ & $I$$^b$ & $Q$ & $I$$^b$ & $Q$ \\
&(GHz) & & (mK~\mbox{km~s$^{-1}$}) &
($10^{26}$s$^{-1}$)\phantom{0}& &(mK~\mbox{km~s$^{-1}$}) & ($10^{26}$s$^{-1}$) &
(mK~\mbox{km~s$^{-1}$}) &
($10^{26}$s$^{-1}$) & (mK~\mbox{km~s$^{-1}$}) & ($10^{26}$s$^{-1}$) \\
\hline\noalign{\smallskip}
(CH$_2$OH)$_2$\phantom{0}& 213--241\phantom{0} & 7(6$^{c}$) &-- & -- & & $<42$ & $<5.3$ & $18\pm6$ & $2.5\pm0.8$ & $18\pm4$ &$3.8\pm0.8$ \\
(CH$_2$OH)$_2$ & 241--270 & 9 & $39\pm5$ & $22\pm3$ & &--& --& -- & --& -- & --\\
(CH$_2$OH)$_2$ & 245--268 & 6 & -- & -- & &$<21$ & $<3.9$ & $23\pm4$ & $4.0\pm0.7$ & $13\pm4$ & $3.0\pm0.9$ \\
NH$_2$CHO & 223--240 & 5 & -- & -- & &$<32$ & $<6.4$ & $38\pm12$& $3.2\pm1.0$ & $17\pm5$ & $3.1\pm0.9$ \\
NH$_2$CHO & 243--261 & 5 & $24\pm7$ & $15\pm4$ && $34\pm8$ & $2.2\pm0.5$ & $21\pm4$ & $1.7\pm0.3$ & $16\pm5$ & $2.4\pm0.7$ \\
NH$_2$CHO & 263--267 & 2 & $<54$ & $<28$ & & $<42$ & $<2.6$ & $22\pm8$ & $1.7\pm0.6$ & $<24$ & $<3.4$ \\
CH$_3$CHO & 242--271 & 26 & 9$\pm$4 & $<7$ & & -- & -- & -- & -- & -- & -- \\
CH$_3$CHO & 211--268 & 45(39$^{c}$) & -- & -- & & $<9$ & $<0.7$ & 9$\pm$2 & 0.8$\pm$0.2 & 7$\pm$2& 1.1$\pm$0.3 \\
HCOOH & 241--268 & 6 & $<24$ & $<6.0$ & & -- & -- & -- & -- & -- & -- \\
HCOOH & 215--271 & 20(16$^{c}$) & -- & -- & & $12\pm5$ & $<0.6$ & 25$\pm$3 & 1.3$\pm$0.2 & 10$\pm$2& 1.0$\pm$0.2 \\
\hline
\end{tabular}\vspace{-0.2cm}
\tablefoot{$^{(a)}$ Number of lines -- details are provided in online Tables~\ref{tabglycol}--\ref{tabformamide}.
$^{(b)}$ Weighted average of line integrated intensities $\int{T_{mb}dv}$.
$^{(c)}$ For the 27--31 Nov. period.}
\end{table*}
\subsection{Production rates of ethylene glycol and formamide}
The estimated production rates of ethylene glycol are given in
Table~\ref{tabobsefe}. The predicted intensities of the ethylene
glycol lines are very close to those expected for thermal
equilibrium (the difference is lower than 1\%), so that the
approximations used for the excitation of this molecule probably do not
affect our estimates of the production rates significantly. The
photodissociation rate at 1 AU was assumed to be $\beta_0 =
2.0\times10^{-5}$ s$^{-1}$ \citep{Cro04a}. Multiplying it by 2
increases the derived production rates by $\approx20$\% for Lemmon and
$\approx10$\% for Lovejoy.
The width and asymmetry of the ethylene glycol line detected in comet
Lemmon (Fig.~\ref{figglycol}a) is comparable to that of the other
species. The derived production rate is indicative of an abundance
relative to water of 0.24\%.
Ethylene glycol was detected in comet Lovejoy during the second
(27--31 Nov.) and third (9--16 Dec.) observing periods with
average productions rates of $3.4\pm1.1\times10^{26}$ and
$3.4\pm0.6\times10^{26}$~\mbox{molec.~s$^{-1}$}, yielding abundances
relative to water of $0.40\pm0.13$\% and $0.30\pm0.05$\%.
The marginal difference between the two periods might
be due to, e.g., an underestimate of the gas temperature
$T_{\rm kin}$ when the comet was closer to Earth at the end of
November. Using $T_{\rm kin}$ = 80 K instead of 65 K, the derived
abundance for late November is 0.36$\pm$0.11\%, and the
differences between the production rates derived from the low- and
high-frequency groups of lines (Table~\ref{tabqprod}) are also
reduced. On the other hand, we did not find any measurable
variation ($>$ 10\%) of $T_{\rm kin}$ in the coma, yielding
an uncertainty lower than 5\% on the production rate of
ethylene glycol.
For the first observing period, we derive an upper limit
of 0.62\%, which explains the non-detection.
The derived production rate of formamide strongly depends on the
assumed photodissociation rate, which is presumably high
\citep[$\beta_0=7\times10^{-4}$ s$^{-1}$ according to][]{Jac76a,Jac76b}.
In this case, most molecules lie in the
collision-dominated coma region, so that radiative processes do
not play an important role in the rotational excitation of the
molecule. Following \citet{Boc00}, we
investigated how the assumed photodissociation rate affects the
derived production rate for comet Lovejoy and its evolution with time.
Indeed, comet
Lovejoy was observed over a range of heliocentric and geocentric
distances (Table~\ref{tabobsefe}), which makes the evolution of the
signal sensitive to the lifetime of formamide. Abundances relative
to water for the three periods are ($7.2\pm5.0$; $5.4\pm1.9$;
$7.3\pm2.5$)$\times10^{-4}$, ($3.4\pm2.0$; $2.2\pm0.6$;
$2.0\pm0.6$)$\times10^{-4}$, and ($2.8\pm1.8$; $1.7\pm0.5$;
$1.3\pm0.4$)$\times10^{-4}$ for $\beta_0 = 7.0, 1.0$, and
$0.1\times10^{-4}$~s$^{-1}$, respectively. For
$\beta_0=1.0\times10^{-5}$~s$^{-1}$, the abundance of formamide
decreases with time, while for $\beta_0=7.0\times10^{-4}$ s$^{-1}$
the abundance increases from period 2 to 3. This means that $\beta_0$ is
more likely around $1.0\times10^{-4}$ s$^{-1}$, which implies an
abundance of formamide relative to water in Lovejoy of 0.021\%.
The detection of formamide in comet Lemmon is marginal, but this
value of $\beta_0$ yields a NH$_2$CHO/H$_2$O ratio of
$0.016\pm0.005$\%, which is close to that found in comet Lovejoy
and similar to the value measured in comet Hale-Bopp
\citep{Boc00}. Production rates given in Table~\ref{tabqprod} are
for $\beta_0$ = $1.0\times10^{-4}$ s$^{-1}$.
A 10\% variation of $T_{\rm kin}$ would only modify
the production rate of formamide in either comet by 3\%.
\subsection{Abundances of other complex molecules}\label{other}
Using the method outlined in
Sect.~\ref{sec:reduction}, we searched for a number of CHO-bearing
molecules in the spectra of comets Lemmon and Lovejoy. Production
rates, or upper limits, were determined assuming the local
thermodynamical equilibrium, and the corresponding abundances relative to
water are displayed in Table~\ref{tababund}. Formic acid (HCOOH)
is detected in comet Lovejoy (more than 6 lines with a S/N $>$3),
but not in comet Lemmon. Acetaldehyde (CH$_3$CHO) is detected at
the 4$\sigma$ level in comet Lovejoy, both at the end of
November (average of 39 lines) and in December
(45 lines) (Table~\ref{tabglycol}). CH$_3$CHO was previously
detected only in
comet Hale-Bopp before \citep{Cro04b}. Methyl formate (CH$_3$OCHO)
and glycolaldehyde (CH$_2$OHCHO) are not detected, but significant
upper limits are obtained. Methyl formate and formic acid were
detected in comet Hale-Bopp \citep{Boc00,Cro04b}. We also include
in Table~\ref{tababund} the abundances of HNCO and H$_2$CO
(Biver et al. {\it in preparation}).
\section{Discussion}
Chemical diversity is observed in the population of Oort Cloud
comets, with abundances varying by up to a factor of ten for
several simple species such as CO, CH$_3$OH, H$_2$S, CS, and
hydrocarbons \citep[e.g.,][]{Boc11}. The origin of this diversity
is unclear and might reflect comet formation at different places
and times in the early solar system. The abundances of ethylene
glycol and formamide in comets Lemmon and
Lovejoy are remarkably similar to the values measured in comet
Hale-Bopp (Table~\ref{tababund}). Conversely, HNCO and CO
are depleted by a factor of 4--5 in both comets with respect to
Hale-Bopp, whereas HCOOH is depleted in comet Lemmon and not in
Lovejoy, and acetaldehyde is enhanced in Lovejoy.
The comparison of cometary abundances with those measured around
protostars can shed light on the processes responsible for the
formation of complex molecules in protoplanetary disks. It is important
to decipher the role of grain-surface chemistry and radiation
processing versus gas-phase chemistry \citep{Walsh2014}. As
discussed by \citet{Cro04a} and \citet{Walsh2014}, an
important issue is the now confirmed high abundance
of ethylene glycol in comets, which is at least 5--6 times
higher than that of the chemically related species CH$_2$OHCHO
(glycolaldehyde) (Table~\ref{tababund}). By contrast,
both molecules are observed in similar abundances in the hot core
Sgr B2(N) \citep{Hollis2002}, and tentatively in the Class 0 solar-type
binary protostar IRAS 16293--2422, where ethylene glycol relative
to glycolaldehyde is 0.3--0.5 \citep{Jor2012}. In IRAS
16293--2422, the glycolaldehyde lines have their origin in the
warm (200--300 K) gas close to binary components \citep{Jor2012}.
Similarly, ethylene glycol, together with other complex molecules
is found to originate from a warm region of radius $\sim$ 60 AU in
the Class 0 source NGC 1333-IRAS2a. Hence, this supports a
scenario where these two molecules are released into the gas phase
by the sublimation of grain mantles. From laboratory experiments,
\citet{Oberg2009} showed that both molecules are produced by
UV-irradiation of methanol ices mixed with CO, with the amount of ethylene
glycol relative to glycolaldehyde increasing with
decreasing CO content and being sensitive to the temperature.
This suggests that complex molecules found in
cometary ices might have been formed from the irradiation of
CO-poor ices. We note, however, that there is no correlation between
the CO and ethylene glycol abundances in comets
(Table~\ref{tababund}).
Formamide, the simplest amide, may have been the starting
point for prebiotic synthesis \citep{Sal12}.
Formamide has recently been
detected in the Class 0 source IRAS 16293--2422, with an abundance
10 times lower than the Hale-Bopp value \citep{Kahane2013}.
Several formation routes are proposed for NH$_2$CHO in
protoplanetary disks \citep{Kahane2013,Walsh2014}. Grain-surface
chemistry satisfactorily explains the NH$_2$CHO abundance relative
to water measured in cometary ices, as well as that of most
CHO-bearing species detected in comets \citep[excluding ethylene
glycol;][]{Walsh2014}.
In summary, the new measurements presented here show that the high
abundance of ethylene glycol is most likely a specific property of comets.
Molecules present in cometary ices could have formed by a wealth of
chemical processes, which will hopefully be better constrained in
the coming years by the combination of new observations, models,
and laboratory works.
\begin{table}
\caption[]{Abundances relative to water}\label{tababund}\vspace{-0.5cm}
\begin{center}
\begin{tabular}{lcccccccccc}
\hline\hline\noalign{\smallskip}
Molecule & \multicolumn{3}{c}{Abundance (\%)} \\
& C/1995~O1$^a$ & C/2012~F6 & C/2013~R1 \\
& (Hale-Bopp) & (Lemmon) & (Lovejoy) \\
\hline\noalign{\smallskip}
HCN & 0.25 & 0.14 & 0.16 \\
CO & 23 & 4.0$^b$ & 7.2$^c$\\
H$_2$CO & 1.1 & 0.7$^c$ & 0.7$^c$ \\
CH$_3$OH & 2.4 & 1.6 & 2.6 \\
HCOOH & 0.09 & $<0.07$ & 0.12 \\
(CH$_2$OH)$_2$ & 0.25 & 0.24 & 0.35 \\
HNCO & 0.10 & 0.025$^c$ & 0.021$^c$ \\
NH$_2$CHO & 0.02 & 0.016 & 0.021 \\
HCOOCH$_3$ & 0.08 & $<0.16$ & $<0.20$ \\
CH$_3$CHO & 0.025 & $<$ 0.07 & 0.10 \\
CH$_2$OHCHO & $<0.04$ & $<$ 0.08 & $<$ 0.07\\
\hline
\end{tabular}
\end{center}\vspace{-0.5cm}
\tablefoot{$^{(a)}$ \citet{Boc00,Cro04a,Cro04b}. $^{(b)}$\citet{Pag14}.$^{(c)}$ Biver et al. {\it in preparation}.} \\
\end{table}
\begin{acknowledgements}
This research has been supported by the Programme national de
planétologie de l'Institut des sciences de l'univers (INSU).
\end{acknowledgements}
|
\section{i. introduction}
The field of molecular electronics has been rapidly expanding during
the last two decades owing to continuous improvement of techniques
intended to electrically contact and control quantum dots and single
molecules in transport junctions. Recent advances in heat
measurements in nanoscale systems allow to study thermoelectric
properties of molecular junctions and similar systems. These studies
bring a deeper understanding of transport mechanisms \cite{1,2,3,4}
and additional information concerning electronic and vibrational
excitation spectra of molecules \cite{5,6}. Also, in the recent years
a new field of molecular thermoelectronics have emerged \cite{7}.
Thermal analogs of molecular transistors and heat-into-electricity
converters were proposed \cite{5,8,9,10,11,12}.
Heat-to-electric power converters operate due to Seebeck effect which
appears provided that thermal and electric driving forces
simultaneously affect electron transport through the considered
system. When a temperature gradient $ \Delta T $ is applied across the
system, a thermovoltage $ V_{th} $ emerges under the condition of
zero net current thus indicating the energy conversion. At small
temperature gradients $(\Delta T \ll T_0,\ T_0$ being the temperature
of the cool region) the system operates within the linear regime and
$ V_{th} = - S\Delta T. $ Within this regime, thermopower $ S $ is the
decisive quantity determining the extent of energy conversion.
Accordingly, the thermopower was intensively studied
\cite{5,10,13,14,15,16} along with the thermoelectric figure of merit
\cite{15,17,18}. Another interesting quantity characterizing
thermoelectric transport through molecular junctions and other systems
of similar kind is thermocurrent $ I_{th} . $ The latter may be
defined as a difference between the electron tunnel current flowing
through a biased thermoelectric junction where the electrodes are
kept at different temperatures $ (T_L = T_0 + \Delta T,\ T_R = T_0)$
and the current flowing in absence of temperature gradient \cite{19}:
\begin{equation}
I_{th} = I(V,T_0, \Delta T) - I(V,T_0, \Delta T = 0). \label{1}
\end{equation}
As well as the thermovoltage, the thermocurrent is simultaneously
controlled by electric and thermal driving forces. The combined
effect of these forces depends of the type of charge carriers involved
in transport and of the bias voltage polarity. When the bias voltage
is sufficiently strong, its effect predominates. Consequently, the
difference between two terms in the Eq. (\ref{1}) diminishes, and $
I_{th} $ approaches zero. At the same time, the thermocurrent flowing
through an unbiased or slightly biased molecule (or quantum dot) is
mostly controlled by the applied temperature gradient $ \Delta T. $
Assuming that $ V = 0 $ and $ \Delta T > 0 $ (the left electrode is
warmer than the right one), $ I_{th} $ takes on positive/negative
values when the charge carriers involved in transport process are,
respectively, holes/electrons. Finally, when $ V $ and $ \Delta T $
influence the transport to a similar extent, the resulting magnitude
and direction of $ I_{th} $ are determined by both $ \Delta T $ sign
and the bias voltage polarity. These two factors may cooperate or
counteract by pushing charge carriers in the same or opposite
directions. In the latter case, $ I_{th} $may change its sign at
certain values of $V$ and $ \Delta T $ when electric and thermal
driving forces counterbalance each other.
It was shown that (disregarding electron-phonon
interactions) maximum efficiency of molecular heat-to-electricity
converter could be reached under the condition of vanishing current $
I(V,T_0,\Delta T).$ Under this condition, $ I_{th} = -
I(V,T_0,\Delta T = 0).$ Therefore, relatively large in magnitude
thermocurrent flowing through a system may indicate that the system
operates in the regime favorable for energy conversion \cite{20}.
Properties of the thermovoltage in molecular junctions and quantum
dots were theoretically analyzed in numerous works
\cite{20,21,22,23,24,25,26,27,28,29}.
As yet, less attention was paid to studies of thermocurrent in spite
of the fact that $I_{th} $ is more straightforward to measure and
model than $ V_{th} ,$ as stated in the recent work \cite{19}.
However, theoretical analysis of $ I_{th} $ behavior within a weakly
nonlinear regime was recently suggested \cite{25}. The purpose of the
present work is to analyze the effect of the gate voltage,
electron-electron interactions and thermal phonons on the electrodes
on the characteristics of thermocurrent flowing through a
single-molecule junction or a quantum dot.
To properly analyze thermoelectric transport through
molecules/quantum dots one needs to use an approach including unified
treatment of electron and phonon dynamics in the considered system.
For this purpose, one may use diagrammatic technique or nonequilibrium
Green's function formalism (NEGF), as described in the review
\cite{30}. However, application of these advanced formalisms to
realistic models simulating molecular junctions is extremely
difficult. Several simplified approaches based on scattering theory
\cite{6} and on quantum rate equations \cite{5,14,20,31,32} were
developed and used to study thermoelectric properties of molecular
junctions taking into account contributions of vibrational phonons and
electron-vibron interactions. Very recently, a scattering theory based
approach was suggested to analyze weakly nonlinear thermoelectric
transport in mesoscopic systems \cite{25,26,27}. Nevertheless, these
studies are not completed so far.
In the present work we use a scattering theory first suggested by
Buttiker \cite{33} combined with certain NEGF based results. The
adopted approach allows to derive the expression for the electron
transmission which remains applicable for an arbitrary value of the
difference between the temperatures of the electrodes. Therefore, this
expression may be employed to analyze nonlinear effects in
thermoelectric properties of considered systems.
\section{ii. model and results}
The schematics of the suggested model is presented in the Fig. 1. To
simplify the computations we mimic the bridge linking the electrodes
by a single state with the energy $ E_0 .$
This energy is independent of the electrodes chemical potentials.
However, if the third (gate) electrode is attaches to the system, one
may shift the position of the bridge level by applying the gate
voltage. We assume that electrons tunnel from the electrodes to the
bridge and vice versa via the channels 1 and 2 thus maintaining an
elastic component of the charge flow. Incoming wave amplitudes $
a_1,a_2 $ and outgoing wave amplitudes $a_1', a_2' $ characterize this
process. Also, electrons in the electrodes may interact with thermal
phonons. An electron may enter the bridge by emitting or absorbing a
phonon. Such electrons contribute to inelastic component of the charge
flow. We assume that the total current is the sum of the elastic and
inelastic components. To describe the inelastic contribution to the
current we introduce a pair of dephasing/dissipative electron
reservoirs.
The channels 3,4 and/or 5,6 connect these reservoirs with the bridge.
While in a reservoir, the electron undergoes scattering by thermal
phonons and then it tunnels to the bridge, so there is no direct coupling of
the bridge to the thermal phonons. In the present analysis we assume
that the thermal phonons are associated with the left and right
electrodes which are kept at the temperatures $ T_L $ and $T_R. $ The labels indicated the reservoirs are accordingly chosen.
\begin{figure}[t]
\begin{center}
\includegraphics[width=4.6cm,height=6.9cm,angle=-90]{n251.eps}
\caption{(Color online) Schematics of the considered system.
Semicircles represent the left and right electrodes, square stands
for the molecule/quantum dot sandwiched in between.
Dephasing/dissipative reservoirs are associated with the electrodes
and characterized by temperatures $ T_L $ and $ T_R, $ respectively.
}
\label{rateI}
\end{center}\end{figure}
So, within the accepted model we have six transport channels. The
relations between incoming particle fluxes $ J_k $ and outgoing fluxes
$ J_k' $ take on the form $(1 \leq i \leq k \leq 6):$
\begin{equation}
J_k' = \sum T_{ik} J_k. \label{2}
\end{equation}
Here, the coefficients $ T_{ik} $ are related to matrix elements of
the scattering matrix $ M:\ T_{ik} = |M_{ik}|^2. $ To maintain the
charge conservation in the system, zero net current should flow in
the channels linking the bridge with the reservoirs:
\begin{align}
J_3 + J_4 - J_3' - J_4' = 0,
\nonumber\\
J_5 + J_6 - J_5' - J_6' = 0. \label{3}
\end{align}
The scattering matrix express outgoing wave amplitudes $ b_L', b_R',
a_3',a_4', a_5' $ and $a_6' $ in terms of incident ones $ b_L, b_R,
a_3,a_4, a_5, a_6 .$ To derive expressions for the matrix elements $
M_{ik} $ we first consider a part of the whole system shown in the
Fig. 1. This part consists of the reservoir of the Temperature $ T_L $
associated with the left electrode and the bridge site. Following the
Buttuker's approach \cite{33}, we find the expression far the matrix
$s_{(1)} $ relating wave amplitudes $ a_1', a_2', a_3', a_4'$ to wave
amplitudes $ a_1, a_2, a_3, a_4 $ namely
\begin{equation}
s^{(1)} = \left( \ba{cccc}
0 & \sqrt{1 - \epsilon_L} & \sqrt {\epsilon_L} & 0
\\
\sqrt{1 - \epsilon_L} & 0 & 0 & \sqrt {\epsilon_L}
\\
\sqrt{ \epsilon_L} & 0 & 0 & - \sqrt {1-\epsilon_L}
\\
0 & \sqrt{\epsilon_L} & -\sqrt {1-\epsilon_L} & 0
\\
\ea \right) \label{4}
\end{equation}
Here, $ \epsilon_L $ is the scattering probability associated with
the effect of thermal phonons concentrated on the left electrode (and
represented by the corresponding dephasing/dissipative reservoir).
When $ \epsilon_L = 0, $ the reservoir is detached from the bridge, so
thermal phonons do not affect electron's tunneling from the left
electrode to the bridge. Within the opposite limit $(\epsilon_L =1),$
electrons traveling from the left electrode to the bridge are
necessarily scattered into the reservoir which results in the overall
phase randomization and inelastic transport. Electron tunneling
through a barrier separating the left electrode from the bridge is
characterized by the transmission and reflection amplitudes $(t_L $
and $r_L ,$ respectively). These are matrix elements of a $ 2 \times 2
$ matrix:
\begin{equation}
s_L = \left( \ba{cc}
t_L & - r_L \\ r_L & t_L \ea \right) . \label{5}
\end{equation}
which relates $ b_L',a_1' $ to $ b_L, a_1. $ Combining Egs. (\ref{4})
and (\ref{5}) one obtained the following expression for the matrix $
M^{(1)} $ which relates $ b_L', a_2', a_3',a_4' $ to $ b_L, a_2, a_3,
a_4 :$
\begin{equation}
M^{(1)} = \left(\ba{cccc}
r_L & \alpha_Lt_L & \beta_Lt_L & 0
\\
\alpha_Lt_L & \alpha_L^2r_L & \alpha_L\beta_Lr_L & \beta_L
\\\beta_Lt_L & \alpha_L\beta_Lr_L & \beta_L^2r_L & -\alpha_L
\\
0 & \beta_L & -\alpha_L & 0 \ea \right). \label{6}
\end{equation}
Here, $ \alpha_L = \sqrt{1 - \epsilon_L},\ \beta_L =
\sqrt{\epsilon_L}. $ now, we take into consideration the remaining
part of the considered system. As shown in an earlier work \cite{34},
the matrix $M^{(2)} $ relating $a_2', b_R', a_5', a_6' $ to $a_2,
b_R, a_5, a_6 $ has the form:
\begin{equation}
M^{(2)} = \left( \ba{cccc}
\alpha_R^2 r_R & \alpha_R t_R & \beta_R & \alpha_R \beta_R r_R
\\
\alpha_R t_R & r_R & 0 & \beta_R t_R
\\
\beta_R & 0 & 0 & -\alpha_R
\\
\alpha_R \beta_R r_R & \beta_R & - \alpha_R & \beta_R^2 r_R
\ea \right) \label{7}
\end{equation}
where $ \alpha_R = \sqrt{1 - \epsilon_R},\ \beta_R =
\sqrt{\epsilon_R}, $ the scattering probability $ \epsilon_R $ is
associated with the effect of thermal phonons concentrated on the
right electrode, and the transmission $(t_R) $ and reflection $(r_R) $
amplitudes characterize electron tunneling between the bridge site and
tunneling between the bridge site the right electrode. Using Eqs.
(\ref{6}), (\ref{7}), one arrives at the following expression for the
scattering matrix \cite{35}:
\begin{widetext}
\begin{equation}
M = \frac{1}{Z} \left \{\ba{cccccc}
r_L + \alpha_L^2\alpha_R^2r_R &\alpha_L \alpha_R t_L t_R &
\beta_L t_L & \alpha_L\alpha_R^2 \beta_L t_L r_R &
\alpha_L \beta_R t_L & \alpha_L \alpha_R \beta_R t_L r_R
\\
\alpha_L \alpha_R t_L t_R & r_R + \alpha_L^2 \alpha_R^2 r_L &
\alpha_L \alpha_R \beta_L t_R r_L & \alpha_R \beta_L t_R &
\alpha_L^2 \alpha_R \beta_R t_R r_L & \beta_R t_R
\\
\beta_L t_L & \alpha_L \alpha_R \beta_L t_R r_L &
\beta_L^2 r_R & \alpha_L(\alpha_R^2 r_L r_R -1) &
\alpha_L \beta_L \beta_R r_L & \alpha_L \alpha_R \beta_L \beta_R r_L r_R
\\
\alpha_L\alpha_R^2 \beta_L t_L r_R & \alpha_R \beta_L t_R &
\alpha_L(\alpha_R^2 r_L r_{R}-1) & \alpha_R^2 \beta_L^2 r_R &
\beta_L \beta_R & \alpha_R \beta_L \beta_R r_R
\\
\alpha_L \beta_R t_L & \alpha_L^2 \alpha_R \beta_R t_R r_L &
\alpha_L \beta_L \beta_R r_L & \beta_L \beta_R &
\alpha_L^2 \beta_R^2 r_L & \alpha_R (\alpha_L^2 r_L r_R -1)
\\
\alpha_L \alpha_R \beta_R r_R t_L & \beta_R t_R &
\alpha_L\alpha_R \beta_L \beta_R r_L r_R & \alpha_R \beta_L \beta_R
r_R & \alpha_R(\alpha_L^2 r_L r_R-1) & \beta_R^2 r_R \\
\ea \right \} . \label{8}
\end{equation}
\end{widetext}
Here, $ Z = 1 - \alpha_L^2 \alpha_R^2 r_L r_R. $
Solving the equations (\ref{2}), (\ref{3}) one arrives at the
following expression for the electron transmission \cite{36}:
\begin{equation}
\tau (E) = \frac{J_2'}{J_1} = T_{21} + \sum_{m,n} K_m^{(2)}
(W^{-1})_{mn} K_n^{(1)}. \label{9}
\end{equation}
Within the considered model, $ 1 \leq m, \ n \leq 2, $
\begin{align}
K_m^{(1)} = T_{2m+1,1} + T_{2m +2,1},
\nonumber\\
K_m^{(2)} = T_{2,2m+1} + T_{2,2m +2}, \label{10}
\end{align}
and matrix elements of $ 2 \times 2 $ matrix $ W $ are given by
\begin{equation}
W_{mn} = (2 - R_{mm})\delta_{mn} - \tilde R_{mn}(1 - \delta_{mn}) \label{11}
\end{equation}
where
\begin{align}
R_{mm} =& T_{2m+1,2m+1} + T_{2m +2,2m+2}
\nonumber\\ &
+ T_{2m+2,2m+1} + T_{2m +1,2m+2},
\nonumber\\
\tilde R_{mn} = &T_{2m+1,2n+1} + T_{2m +1,2n+2}
\nonumber\\ &
+ T_{2m+2,2n+1} + T_{2m +2,2n+2}. \label{12}
\end{align}
It is known that thermoelectric efficiency of a thermoelectric
materials becomes reduced if the phonons thermal conductance takes on
significant values. In the expression of thermoelectric figure of
merit which characterizes the efficiency of energy conversion at small
temperature gradients,the denominator is the sum of electron and
phonon thermal conductances \cite{37,38}.
We remark that within the adopted model the phonon thermal conductance
through the junction is zero. This seems a reasonable assumption for
experiments give low values of phonon thermal conductance in several
molecular junctions \cite{39,40}. This may be attributed to the fact
that in many molecules the majority of vibrational transitions lie
above the range determined by thermal energy provided that relevant
temperatures take on values below room temperature \cite{15,41}.
In the following analysis we focus on a significantly coupled system.
Provided that the dephasing reservoirs are detached from the bridge
$(\epsilon_L = \epsilon_R = 0)$ and the barriers separating the
electrodes from the bridge are identical $(t_L = t_R = t,\ r_L = r_R =
r), $ the electron transmission determined by Eqs.
(\ref{8})-(\ref{12}) accepts a simple form:
\begin{equation}
\tau(E) = \frac{t^4}{(1 + r^2)^2} . \label{13}
\end{equation}
As known, in the case of coherent transport, the current flowing
through the system could be presented in the form:
\begin{align}
I = &\frac{ie}{h} \sum_\sigma \int dE \big \{ \big(\Gamma_L^\sigma f_L
- \Gamma_R^\sigma f_R \big)\big(G_\sigma^r - G_\sigma^a \big)
\nonumber \\ & +
\big(\Gamma_L^\sigma - \Gamma_R^\sigma \big) G_\sigma^<
\big \} \label{14}
\end{align}
Here, $ \Gamma_{L,R}^\sigma (E)$ are self-energy terms describing
coupling of an electron on the bridge (with a certain spin orientation
$\sigma $) to the electrodes, $ f_{L,R} $ are Fermi distribution
functions for the electrons on the electrodes, and $ G_\sigma^{r,a,<}
(E) $ are the retarded, advanced and lesser Green's functions
associated with the QD/molecule.
Further, we accept a wide-band approximation for the self-energy
terms, and we concentrate on a symmetrically coupled system assuming $
\Gamma_L^\sigma = \Gamma_R^\sigma = \Gamma. $ Then the term
proportional to the lesser Green's function disappears from Eq.
(\ref{14}), and we obtain:
\begin{equation}
I = \frac{e}{\pi\hbar} \int \tau(E) (f_L - f_R)dE \label{15}
\end{equation}
where the electron transmission function is given by
\begin{equation}
\tau(E) = \frac{1}{2}\Gamma \sum_\sigma \big[G_\sigma^r (E) -
G_\sigma^a(e) \big] \equiv g^2(E). \label{16}
\end{equation}
So, we have derived two expressions for the electron transmission
function appropriate to describe coherent electron transport through
the considered system. One of them (Eq. (\ref{13})) is obtained using
the scattering theory whereas another one (Eq. (\ref{16})) is ensued
employing NEGF formalism. Within the considered coherent limit these
expressions should be identical. Comparing them, we get:
\begin{equation}
t^2 = \frac{2g}{1 +g}. \label{17}
\end{equation}
Provided that the electron transport through the junction is
undisturbed by electron-phonon interactions and disregarding spin-flip
processes, $ G_\sigma^r(E) $ may be approximated as \cite{42}:
\begin{equation}
G_\sigma^r(E) = \frac{E - E_0 - \Sigma_{02}^\sigma - U\big(1-
\big<n_{-\sigma}\big>\big)}{(E - E_0 - \Sigma_{0\sigma})(E - E_0 - U -
\Sigma_{02}^\sigma) + U\Sigma_{1\sigma}} . \label{18}
\end{equation}
In this expression, $ U $ is the charging energy describing Coulomb
repulsion between electrons on the bridge, and $ \big<n_\sigma\big>$
are one-particle occupation numbers which could be computed by
integration of the lesser Green's function $ G_\sigma^< (E) $ over the
whole range of tunnel energy $ E $ values:
\begin{equation}
\big< n_\sigma \big > = -\frac{i}{2\pi} \int G_\sigma^< (E) dE.
\label{19}
\end{equation}
Self-energy terms $ \Sigma_{o\sigma},\ \Sigma_{1\sigma} $ and $
\Sigma_{2\sigma} $ appear in the Eq. (\ref{10}) due to the coupling of
the bridge to the leads:
\begin{align}
\Sigma_{0\sigma} = & \sum_{r\beta}\frac{|t_{r\beta;\sigma}|^2}{E -
\epsilon_{r\beta\sigma} + i\eta}, \label{20}
\\
\Sigma_{1\sigma} = & \sum_{r\beta} {|t_{r\beta;\sigma}|^2}
f_{r,-\sigma}^\beta \Big\{\frac{1}{E - \epsilon_{r\beta;-\sigma} +
i\eta}
\nonumber\\ & +
\frac{1}{E - 2E_0 - U +\epsilon_{r\beta;-\sigma} + i\eta} \Big\}, \label{21}
\\
\Sigma_{2\sigma} = & \sum_{r\beta} |t_{r\beta;\sigma}|^2
\Big\{\frac{1}{E - \epsilon_{r\beta;-\sigma} + i\eta}
\nonumber\\ & +
\frac{1}{E - 2E_0 - U +\epsilon_{r\beta;-\sigma} + i\eta} \Big\},
\label{22}
\\
\Sigma_{02} = & \Sigma_{0\sigma} + \Sigma_{2\sigma}. \label{23}
\end{align}
Here, $ t_{r\beta;\sigma}$ are parameters describing the coupling of $
r,\beta $ electron states on the electrode $ \beta(\beta = L,R) $ to
the bridge state, $ \epsilon_{r\beta;\sigma} $ are single-electron
energies in the electrode $ \beta, \ f_{r\sigma}^\beta $ is the Fermi
distribution function for the energy $ \epsilon_{r\beta;\sigma} ,$
chemical potential $ \mu_\beta $ and temperature $ T_\beta $ and
$\eta $ is a positive infinitisemal parameter.
Previously introduced coupling parameters $ \Gamma_{L,R}^\sigma $
are closely related to $ \Sigma_{0\sigma},$ namely: $ \Sigma_{0\sigma}
= \Sigma_{0\sigma}^L+ \Sigma_{0\sigma}^R,\ \Gamma_{L,R}^{\sigma} = - 2 \mbox{Im} \Sigma_{o\sigma}^{L,R}.$ We remark that $ \Sigma_{1\sigma} $
and $ \Sigma_{2\sigma} $ depend on the temperatures of electrodes
which is taken into account in further computations.
Scattering probabilities $ \epsilon_{L,R} $ may be given an explicit
physical meaning by expressing them in terms of relevant energies. In
the considered system, dephasing and energy dissipation originate from
interactions of charge carriers in the electrodes with thermal
phonons. So, one can approximate these parameters as follows:
\begin{equation}
\epsilon_\beta = \frac{\Gamma_{ph}^\beta}{\Gamma_\beta +
\Gamma_{ph}^\beta} . \label{24}
\end{equation}
Here, $ \Gamma_{ph}^\beta $ represents the self-energy term occurring
due to electron-phonon interactions in the electrode $\beta.$ Using
NEGF to compute the relevant electron and phonon Green's functions
within the self-consistent Born approximation, one may derive a
relatively simple approximation for $ \Gamma_{ph}^\beta$ \cite{43}:
\begin{align}
\Gamma_{ph}^\beta = & \frac{2\pi}{\Gamma_\beta^2} \int_0^\infty
d(\hbar\omega) \sum_\alpha |\lambda_{\alpha\beta}|^2
\delta(\hbar\omega - \hbar\omega_{\alpha\beta}) \sum_{r\sigma}
|t_{r\beta;\sigma}|^2
\nonumber\\ & \times
\Big\{ \big[1 - f_\sigma^\beta(E - \hbar\omega)\big]\big[1 +
N_\beta(\omega)\big] \delta(E - \epsilon_{r\beta;\sigma} - \hbar\omega)
\nonumber\\ & +
f_\sigma^\beta(E - \hbar\omega) N_\beta(\omega) \delta(E -
\epsilon_{r\beta;\sigma} - \hbar\omega)
\nonumber\\ & +
f_\sigma^\beta(E + \hbar\omega) \big[1 + N_\beta(\omega)\big]
\delta(E - \epsilon_{r\beta;\sigma} + \hbar\omega)
\nonumber\\ & +
\big[1 - f_\sigma^\beta(E + \hbar\omega)\big] N_\beta(\omega)
\delta(E - \epsilon_{r\beta;\sigma} + \hbar\omega) \Big\} \label{25} .
\end{align}
In writing down this expression, we had assumed that the coupling
strength $ \lambda_{\alpha\beta} $ characterizing interaction between
an electron in the electrode $ \beta $ and a thermal phonon with the
energy $ \hbar\omega_\alpha $ does not depend of the electron state $
r,\sigma. $
The electrodes are kept at different temperatures $T_\beta ,$ so we
introduce two phonon distribution functions $ N_\beta(\omega) = \big
\{\exp \big[\hbar\omega/kT_\beta\big] - 1 \big\}^{-1},\ ( k $ being
the Bolzmann's constant).
The phonon spectral function may be determined using molecular dynamic
simulations. However, to qualitatively analyze the effect of thermal
phonons on the transport characteristics, one may use the
approximation \cite{44}:
\begin{align}
\rho_{ph}^\beta(\omega) = & \sum_\alpha |\lambda_{\alpha\beta}|^2 \delta(\hbar\omega - \hbar\omega_{\alpha\beta})
\nonumber\\ = &
\lambda_\beta \left(\frac{\omega}{\omega_{c\beta}}\right)
e^{-\omega/\omega_{c\beta}} \theta(\omega)
\label{26}
\end{align}
where $ \theta(\omega) $ is the Heaviside step function, and $ \omega_{c\beta} $ characterize relaxation times for the thermal phonons.
Substituting the approximation (\ref{26}) into Eq. (\ref{25}) one may
see that the major contribution to the integral comes from the region
where $ \omega \sim \omega_c. $ On these grounds, one may replace $
\omega $ by $ \omega_{c\beta} $ in the arguments of slowly varying functions
in the integrand of Eq (\ref{25}). Also, one may assume that $
\hbar\omega_{c\beta} $ is much smaller than $ \mu_\beta. $ So,
corrections including $ \hbar\omega_{c\beta} $ in the arguments of the
Fermi distribution functions and delta functions may be omitted. Then
one may use the expression for $ \Gamma_\beta^\sigma $ which follows
from its definition:
\begin{equation}
\Gamma_\beta^\sigma = 2\pi \sum_r |t_{r\beta;\sigma}|^2 \delta (E -
\epsilon_{r\beta;\sigma}) \label{27}
\end{equation}
to reduce Eq. (\ref{25}) to the form:
\begin{equation}
\Gamma_{ph}^\beta = \frac{2\lambda_\beta}{\Gamma_\beta}
\hbar\omega_{c\beta}
\left(\frac{kT_\beta}{\hbar\omega_{c\beta}}\right)^2 \zeta
\left(2;\frac{kT_\beta}{\hbar\omega_{c\beta}} + 1\right)
\label{28}
\end{equation}
where $\zeta (x;q) $ is the Riemann's $\zeta $ function. If the
thermal energy $ kT_\beta $ significantly exceeds $ \hbar\omega_c, \
\Gamma_{ph}^\beta $ accepts an especially simple form, namely: $
\Gamma_{ph}^\beta \approx 2\lambda_\beta kT_\beta\big/\Gamma_\beta. $
In further calculations we assume for simplicity that $ \lambda_L =
\lambda_R = \lambda. $
For a considered symmetrically coupled junction, the thermocurrent is
described by the following expression:
\begin{align}
I_{th} = & \frac{e}{\pi\hbar} \int dE \big\{\tau(E,T_0,\Delta T)
f^L(E,T_0 + \Delta T)
\nonumber\\ & -
\tau(E,T_0,\Delta T =0) f^L(E,T_0) - \Delta\tau f^R(E,T_0) \big\} \label{29}
\end{align}
where $ \Delta \tau $ is given by:
\begin{equation}
\Delta\tau = \tau(E,T_0,\Delta T) - \tau(E,T_0, \Delta T = 0). \label{30}
\end{equation}
The suggested approach allows one to analyze the effect of thermal
phonons on characteristics of thermoelectric transport beyond the
linear regime. Using Eqs. (\ref{8})-(\ref{28}), one may compute the
electron transmission function for an arbitrary value of the ratio $
\Delta T/T_0 .$ This result may be used to calculate thermocurrent
and analyze how it is affected by various characteristics of the
considered junction (such as the quality of contact between the
electrodes and the linker, electron-electron and electron-phonon
interactions) and by external factors including the bias and gate
voltage and temperature gradient.
\section{iii. discussion}
The most interesting thermoelectric properties are better pronounced
in weakly coupled junctions where electron transmission exhibits sharp
maxima \cite{20}. Correspondingly, in further analysis we assume that
$ \Gamma < kT_0. $
Then, as shown in the Fig. 2, $ \Delta \tau $ displays sharp dips at
$ E = E_0 $ and $ E = E_0 + U, $ and the magnitudes of these features
increase as the difference in the temperatures of electrodes enhances.
Electron-phonon interactions significantly affect the transmission. As
the coupling between electrons and thermal phonons strengthens, $
\Delta \tau $ generally takes on greater values. However, too strong
electron-phonon interactions bring partial spreading of the resonance
features which is not favorable for nonlinear behavior of
thermocurrent to be revealed.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.8cm,height=4.3cm]{n252.eps}
\caption{(Color online) Energy dependencies of the electron
transmission through the junction. The function $ \Delta\tau =
\tau(E,\Delta T) - \tau(E,\Delta T = 0) $ is shown assuming $ E_0 =
-2meV,\ U = 6meV,\ \Gamma = 0.3meV,\ kT_0 = 0.65 meV $ at $\lambda =
0.5meV $(left panel) and $ \Delta T = 0.5 T_0 $ (right panel).
}
\label{rateI}
\end{center}\end{figure}
Thermocurrent dependence of the applied bias voltage is illustrated in
the left panel of the Fig. 3 assuming for certainty that the left
(hot) electrode is kept at higher voltage and the bridge energy level
$ E_0 $ is situated below the Fermi level for the electrodes. When the
bias is small so that $E_0 $ remains outside the conduction window
whose width is determined by the difference between chemical
potentials of electrodes $ \mu_L $ and $\mu_R ,$ the charge flow is
driven by the temperature gradient. Higher temperature of the left
electrode enhances probability for an electron to tunnel there from
the bridge provided that $ E_0 $ is rather close to the Fermi level,
so that their difference is of the same order as the thermal energy $
kT_0. $ Under these conditions, $ I_{th} $ takes on positive values.
When the bias becomes greater, $ E_0 $ enters the conduction window,
and the electric driving forces come into play. As a result, the
thermocurrent changes its sign. It remains negative at moderate bias.
In strongly biased junctions, electric driving forces predominate, and
$ I_{th} $ approaches zero. The depth of the dip appearing on the $
I_{th} - V $ curves, as well as its position, are controlled by
several factors. Assuming that the charging energy $ U $ and
electron-phonon coupling strength $ \lambda $ are fixed, it is
determined by the value of $ \Delta T. $ The greater becomes the
difference between the electrodes temperatures, the greater is the
maximum magnitude of $ I_{th} $ indicating more favorable conditions
for energy conversion in the considered system.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.8cm,height=4.3cm]{n253.eps}
\caption{(Color online) Left panel: Thermocurrent as the function of
bias voltage. The curves are plotted at $ U = 6meV,\ \lambda =
0.5meV,\ kT_0 = 0.65meV. $ Right panel: Temperature dependencies of
the thermocurrent. The curves are plotted assuming $ kT_0 = 0.65meV,\
\lambda = 0.5meV. $ Remaining parameters are the same as in the Fig.
2.
}
\label{rateI}
\end{center}\end{figure}
It was reported \cite{19} that thermocurrent flowing through
a quantum dot may exhibit a nonlinear dependence of $ \Delta T $ even
at small values of the latter. As recently suggested \cite{25}, the
nonlinearity appears due to renormalization of the energy $ E_0 $ in
the presence of temperature gradient. Taking into account the
suggested energy renormalization, we showed that weak nonlinearity of
$ I_{th} $ versus $ \Delta T $ curves may be actually traced (see
right panel of the Fig. 3). Presented curves demonstrate a
qualitative agreement with the experimental results of Ref. \cite{19}.
However, we remark that the suggested renormalization $ E_0 \to E_0 +
zk\Delta T $ where $|z| < 1 $ may significantly affect the
thermocurrent behavior only at certain values of the bias voltage,
when $ E_0 $ is very close to the boundary of the conduction window.
Coulomb repulsion between electrons on the bridge and the electron
interactions with thermal phonons may significantly influence
thermoelectric transport through the junction, as demonstrated in the
Fig. 4. The enhancement of charging energy narrows down the interval
where $ I_{th} $ accepts negative values and makes the dips in the $
I_{th} - V $ curves more shallow. This may be explained by considering
the role taken by Coulomb interactions. Electron-electron interactions
are hindering electron flow through the system in both directions,
which leads to the Coulomb blockade. These interactions could be
treated as a source of an effective force opposing any predominating
driving force (originating either from the bias voltage or from the
temperature gradient applied across the junction). As a result, the
characteristic features manifested in the shapes of $ I_{th} - V $
curves become less distinctly pronounced.
Electrons interactions with thermal phonons do not change the width of
the region where $ I_{th} $ remains negative. However, these
interactions may affect the magnitude of thermocurrent. In the right
panel of the Fig. 4, we present the $ I_{th} $ value at the bottom
of the dip as a function of the electron-phonon coupling strength. As
shown in the figure, the dip depth reduces as $ \lambda $ increases.
This effect is better pronounced at weak or moderate electron-phonon
coupling $(\lambda \lesssim \Gamma, kT_0),$ and it fades away when $
\lambda $ significantly exceeds $ \Gamma. $
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.8cm,height=4.3cm]{n254.eps}
\caption{(Color online) The effect of electron-electron (left panel)
and electron-phonon (right panel) interactions on the thermocurrent.
Curves are plotted at $ \Delta T = 0.25 T_0,\ \lambda = 0 $ (left
panel) and at $ V = 6meV $ (right panel). Other relevant parameters
are the same as in the previous figures.
}
\label{rateI}
\end{center}\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.8cm,height=4.3cm]{n255.eps}
\caption{(Color online) Thermocurrent as a function of the bridge
level position $ E_0. $ Curves are plotted assuming $ V = 4mV,\ \Delta
T = 0.25 T_0,\ \lambda = 2.5 meV $ (left panel) and $ U = 5meV $
(right panel). Other relevant parameters are the same as in the
previous figures.
}
\label{rateI}
\end{center}\end{figure}
As known, transport properties of molecular junctions and other
similar systems may be controlled by varying the positions of the
bridge energy levels. Practically, the levels may be shifted by a gate
voltage applied to the system. Within the accepted model, the bridge
in the considered junction is represented by a single energy level $
E_0. $ In the Fig. 5 we trace the thermocurrent dependencies of $
E_0. $ The presented results confirm those shown in the previous
figures. Again, the thermocurrent acquires a negative sign when the
bridge level moving upwards appears in the conduction window. For a
symmetrically coupled system this happens at $E_0 = - \frac{1}{2}V $
assuming that the Fermi level for unbiased junction $ \mu = 0. $As the
bridge level moves higher, $ I_{th} $ remains negative while the level
is still inside the conduction window. When it leaves the window but
remains near its upper boundary, the thermocurrent may change sign
once more, being influenced by electron-electron interactions and
electron-phonon interactions in the electrodes. However, when the
bridge level moves farther away from the conduction window, both terms
in the Eq. (\ref{1}) approach zero, so the thermocurrent disappears.
\section{iv. conclusion}
Finally, we repeat again that various aspects of energy conversion in
nanoscale systems attract significant interest of the research
community. The present work was inspired by this common interest.
Also, the present research was motivated by recent experimental
observations of nonlinear thermocurrent flowing through quantum dots.
The thermocurrent defined by Eq. (\ref{1}) is an important
characteristic of thermoelectric transport. The behavior of the thermocurrent brings some information concerning the efficiently of energy conversion in the considered systems. For instance, $ I_{th} $ reaches minimum at a certain value of the applied bias voltage, as shown in the Figs. 3,4. One may conjecture that this minimum occurs when the current $
I(V,T_0, \Delta T) $ becomes zero, so the corresponding magnitude of
bias voltage is close to the magnitude of thermovoltage $ V_{th} $ and
may be employed to estimate the latter. However, we remark that one
cannot extract sufficient information to properly estimate the
thermoelectric efficiency of the system basing on the thermocurrent
behavior alone. Even within the linear in temperature regime one
needs additional information concerning electron electrical and
thermal conductances and thermopower to make such estimates.
In a nanoscale junction consisting of a quantum dot or molecule
sandwiched in between two electrodes, the thermocurrent value is
controlled by several factors. These include temperatures of
electrodes, bias and gate voltage, charging energy characterizing
electron-electron interactions on the bridge and specific energies
characterizing the coupling of the bridge to the electrodes and
electron-phonon interactions. To theoretically analyze possible effect
of the above factors on thermocurrent we employ a single-particle
scattering approach pioneered by Landauer in the context of charge
transport in mesoscopic systems.
For simplicity, we simulate the bridge in the considered junction by
a single orbital, and we assume that it is symmetrically coupled to
the electrodes. Then the thermocurrent is expressed in terms of
electron transmission functions (which depend on all above mentioned
factors) and Fermi distribution functions for the electrodes. We
derive the expression for the electron transmission which remains
valid for an arbitrary value of the difference between the
temperatures of electrodes. This makes it suitable for analysis of
thermoelectric transport both within and beyond linear in temperature
regime. Using this expression, we show that varying the bias and gate
voltage one may create favorable conditions for heat-to-electricity
conversion in a junction assuming that the most important
characteristic energies $ kT_0,\ U $ and $ \lambda $ are fixed. Also,
we analyze how electron-electron interactions on the bridge and
electrons interactions with thermal phonons associated with electrodes
may affect the thermoelectric properties of the considered systems.
Coulomb repulsion between electrons opposes electron transport through
the junction at small bias voltage. We show that this brings a partial
suppression of the thermocurrent. Electron-phonon interactions in the
electrodes assist in the increase of scattering probabilities thus
destroying the coherence of electron transport and bringing additional
suppression of thermocurrent, as illustrated in Figs. 4,5.
The computational method employed in the present work may be further
generalized to include the effect of molecular vibrations. Also, one
may simulate the bridging molecule/quantum dot by several orbitals
thus opening the way to studies of quantum interference effects. So,
we believe that presented computational scheme and obtained results
may be helpful for further understanding of thermoelectric properties
of nanoscale systems.
\section{ Acknowledgments}
{\bf Acknowledgments:}
This work was supported by NSF-DMR-PREM 0934195. The author
thanks G. M. Zimbovsky for help with the manuscript.
|
\section{Introduction}
\label{Intro}
One of the most challenging issues of modern cosmology is to describe the positive late time acceleration~\cite{SNeIa1,SNeIa2,copeland}
through a single self-consistent theoretical scheme.
Indeed, the physical origin of the measured cosmic speed up is not well accounted on theoretical grounds, without invoking the existence of an additional fluid which drives
the universe dynamics, eventually dominating over the other species.
Any viable fluid differs from standard matter by manifesting negative equation of state parameters,
capable of counterbalancing the gravitational attraction at late times~\cite{antig}.
Thus, since no common matter is expected to behave anti-gravitationally, one refers to such a fluid as dark energy.
The simplest candidate for dark energy consists in introducing within Einstein equations a vacuum energy cosmological constant term, namely $\Lambda$~\cite{v1}.
The corresponding paradigm, dubbed the $\Lambda$CDM model, has been probed to be consistent with almost all experimental constraints~\cite{tests},
becoming the standard paradigm in cosmology.
One of the main advantages of $\Lambda$CDM is the remarkably small number of cosmological parameters that it introduces,
which suggests that any modifications of Einstein's gravity reduce to $\Lambda$CDM at small redshift~\cite{freddo}.
However, recent measurements of the Hubble expansion rate at redshift $z = 2.34$~\cite{Delubac:2014aqe}
and an analysis of linear redshift space distortions~\cite{Macaulay:2013swa} reside outside the $\Lambda$CDM expectations at $2.5\, \sigma$ and $0.99$ confidence level, respectively.
Due to these facts and to the need of accounting for the ultraviolet modifications of Einstein's gravity,
extensions of general relativity have been proposed so far.
Moreover, the standard cosmological model is plagued by two profound shortcomings.
First, according to observations, it is not clear why matter and $\Lambda$ magnitudes appear to be extremely close to each other, indicating an unexpected coincidence problem~\cite{chomp}.
Second, cosmological bounds on $\Lambda$ indicate a value which differs from quantum field calculations by a factor of 123 orders of magnitude, leading to a severe fine-tuning problem~\cite{problema2}.
Standard cosmology deems that the universe dynamics can be framed assuming that dark energy evolves as a perfect fluid,
with a varying equation of state, i.e. $\omega(z)\equiv P / \rho$, with total pressure $P$ and density $\rho$. So, in a Friedmann-Robertson-Walker (FRW)
picture, the universe dynamics is depicted through a pressureless matter term, a barotropic evolving dark energy density and a vanishing scalar
curvature, i.e. $\Omega_{k}=0$~\cite{freddo2}.
In lieu of developing a theory which predicts the dark energy fluid, cosmologists often try to reconstruct the universe expansion history,
by parameterizing the equation of state of dark energy~\cite{nexu}.
For example, polynomial fits, data depending reconstructions and cosmographic
representations~\cite{nexu2} are consolidated manners to reconstruct $\omega(z)$~\cite{capo,cpl,capo2,capo3,capo4,noi}.
All cosmological recontructions are based on inferring the properties of dark energy
without imposing \emph{a priori} a form for the equation of state.
In fact, any imposition would cause misleading results~\cite{mine1},
as a consequence of the strong degeneracy between cosmological models.
Therefore, it turns out that a reconstruction of $\omega(z)$ should be carried out as much as possible in a model independent manner~\cite{mine2}.
To this regard, a well established method is to develop a model independent parametrization by expanding $\omega(z)$
into a truncated Taylor series and fixing the corresponding free parameters through current data~\cite{cpl,altro}.
However, even though Taylor series are widely used to approximate known functions with polynomials around some point, they provide
bad convergence results over a large interval, since $\omega(z)$ is expanded around $z=0$,
while data usually span over intervals larger than the convergence radius~\cite{aggiungiamo,john}.
A more sophisticated technique of approximation, the Pad\'e approximation,
aims to approximate functions by means of a ratio between two polynomials~\cite{padex}.
Pad\'e approximation is usually best suited to approximate diverging functions and functions over a whole interval,
giving a better approximation than the corresponding truncated Taylor series~\cite{padexapprox}.
In~\cite{OrlChri}, Pad\'e approximations have been introduced in the context of cosmography, whereas applications have been discussed and extended in~\cite{Chinese}, but
the authors have focused principally on writing the dark energy equation of state as a Pad\'e function.
In this work we want to propose a new approach to cosmography, based on approximating the luminosity distance by means
of Pad\'e functions, instead of Taylor polynomials.
In this way we expect to have a better match of the model with cosmic data and to overcome possible divergences of the Taylor approach at $z\gg1$.
Indeed, using the Pad\'e approximation of the luminosity distances, we also show that one can improve the quality of the fits with
respect to the standard re-parametrizations of the luminosity distances by means of auxiliary variables.
We also propose how to deal numerically with such approximations and how to get the most viable Pad\'e expansions.
As a result, we will obtain a refined statistical analysis of the cosmographic parameters.
A large part of the work will be devoted to
outline the drawbacks and the advantages of this technique and compare it to more standard approaches as Taylor series and the use of auxiliary variables.
Moreover, we also include a discussion about the most adequate Pad\'e types among the wide range of possibilities.
Finally, we obtain a reconstruction of the dark energy equation of state which is only based on
the observational values of the luminosity distance, over the full range for the redshift in which data are given.
In this way, we demonstrate that Pad\'e approximations are actually preferred to fit high redshift cosmic data, thus representing a valid alternative technique to reconstruct
the universe expansion history at late times.
The paper is structured as follows: in Sec.~\ref{model} we highlight the role of cosmography in the description of the present time dynamics of the universe.
In particular, we discuss connections with the cosmographic series and the FRW metric.
In Sec.~\ref{sec:padeandcosmography} we introduce the Pad\'e formalism
and we focus on the differences between standard Taylor expansions and rational series
in the context of cosmography, giving a qualitative indication that a Pad\'e approximation could be preferred. We also enumerate some issues
related to cosmography in the context of the observable universe. For every problem, we point out possible solutions and we underline
how we treat such troubles in our paper, with particular attention to the Pad\'e formalism. All experimental results have been portrayed in Secs.~\ref{Sect:DataSet} and~\ref{ParameterEstimations}, in which we present
the numerical outcomes derived both from using the Pad\'e technique
and standard cosmographic approach. In Sec.~\ref{applicationsPade} we give an application of the Pad\'e recipe, that is,
we use the Pad\'e technique to estimate the free parameters of some known models. In Sec.~\ref{universeEoS}, we discuss the consequences on the equation of state for the universe which can be inferred from our numerical
outcomes derived by the use of Pad\'e approximants. Moreover, in Sec.~\ref{consequencespade} we discuss our numerical outcomes and we interpret the bounds obtained. Finally, the last section, Sec.~\ref{conclusions}, is devoted to
conclusions and perspectives of our approach.
\section{The role of cosmography in \emph{precision cosmology}}
\label{model}
In this section we briefly introduce the role of cosmography and its standard-usage techniques to fix cosmographic
constraints on the observable universe. The great advantage of the cosmographic method is that it permits
one to bound present time cosmology
without having to assume any particular model for the evolution of dark energy with time.
The \emph{cosmographic method} stands for a coarse grained model
independent technique to gather viable limits on the universe expansion history at late times, provided the cosmological
principle is valid~\cite{mine1,mine2,cosmus}. The corresponding requirements demanded by cosmography are
homogeneity and isotropy with spatial curvature somehow fixed.
Common assumptions on the cosmological puzzle provide a whole energy budget dominated by $\Lambda$,
(or by some sort of dark energy density), with cold dark matter in second place and baryons as a small fraction only.
Spatial curvature in case of time-independent dark energy density is actually constrained to be negligible.
However, for evolving dark energy contributions, observations are not so restrictive~\cite{rat}.
More details will be given later, as we treat the degeneracy between scalar curvature and variation of the acceleration.
From now on, having fixed the spatial curvature $\Omega_{k}$ to be zero, all cosmological observables can be expanded around present time.
Moreover, comparing such expansions to cosmological data allows one to fix bounds on the evolution of each
variable under exam. This strategy matches cosmological observations with theoretical expectations.
By doing so, one gets numerical outcomes which do not
depend on the particular choice of the cosmological model, since only Taylor expansions are compared with data.
Indeed, cosmography relates observations and theoretical predictions, and it is able to alleviate the degeneracy
among cosmological models. Cosmography is therefore able to distinguish between models that are compatible with
cosmographic predictions and models that have to be discarded, since they do not fit the cosmographic limits.
Hence, according to the cosmological principle, we assume the universe to be described by a Friedmann-Robertson-Walker (FRW) metric, i.e.
\begin{equation}\label{frw}
{\rm d} s^2={\rm d} t^2-a(t)^2\left({\rm d} r^2+r^2{\rm d}\Omega^2\right)\,,
\end{equation}
where we use the notation ${\rm d}\Omega^2\equiv {\rm d}\theta^2+\sin^2\theta {\rm d}\phi^2$.
As a first example of cosmographic expansions, we determine the scale factor $a(t)$ as a Taylor series~\cite{Weinberg2008} around present time $t_0$. We have
\begin{eqnarray}\label{serie1a}
a(t) & \sim & a(t_0)+ a'(t_0) \Delta t + \frac{a''(t_0)}{2} \Delta t^2+
\frac{a'''(t_0)}{6} \Delta t^3 +\nonumber\\
&+& \frac{a^{(iv)}(t_0)}{24} \Delta t^4 +\frac{a^{(v)}(t_0)}{120} \Delta t^5+\ldots\,,
\end{eqnarray}
which recovers signal causality if one assumes $\Delta t\equiv t-t_0>0$. From the above expansion of $a(t)$, one defines
\begin{subequations}\label{CSdef}
\begin{align}
H \equiv \frac{1}{a} \frac{{\rm d} a}{{\rm d} t}\,,\\
q \equiv -\frac{1}{a H^2} \frac{{\rm d}^2a}{{\rm d} t^2}\,,\\
j \equiv \frac{1}{a H^3} \frac{{\rm d}^3a}{{\rm d} t^3}\,,\\
s \equiv \frac{1}{a H^4} \frac{{\rm d}^4a}{{\rm d} t^4}\,,\\
l \equiv \frac{1}{a H^5} \frac{{\rm d}^5a}{{\rm d} t^5}\,.
\end{align}
\end{subequations}
Such functions are, by construction, model independent quantities, i.e. they do not depend on the form of the dark energy fluid,
since they can be directly bounded by observations.
They are known in the literature as the Hubble rate ($H$), the acceleration parameter ($q$), the jerk parameter ($j$),
the snap parameter ($s$) and the lerk parameter ($l$)~\cite{starobinskievisser}. Once such functions are fixed at present time,
they are referred to as the \emph{cosmographic series} (CS). This is the set of coefficients usually derived in cosmography from observations.
Rewriting $a(t)$ in terms of the CS gives
\begin{eqnarray}\label{serie1a2}
a(t) & \sim & 1+ H_0 \Delta t - \frac{q_0}{2} H_0^2\Delta t^2+\nonumber \\
&+&
\frac{j_0}{6} H_0^3 \Delta t^3 + \frac{s_0}{24} H_0^4\Delta t^4 +\frac{l_0}{120} H_0^5\Delta t^5\ldots\ ,
\end{eqnarray}
where we have normalized the scale factor to $a(t_0) = 1$. By rewriting Eq.~($\ref{serie1a}$) as in Eq.~($\ref{serie1a2}$), one can read out the meaning of each parameter.
In fact, each term of the CS displays a remarkable dynamical meaning.
In particular, the
snap and lerk parameters determine the shape of Hubble's flow at higher redshift regimes. The Hubble parameter must be
positive, in order to allow the universe to expand and finally $q$ and $j$ fix kinematic properties at lower redshift domains. Indeed,
the value of $q$ at a given time specifies whether the universe is accelerating or decelerating and also provides some hints on
the cosmological fluid responsible for the dynamics.
Let us focus on $q$ first. We can distinguish three cases, splitting the physical interval of viability for $q_0$:
\begin{enumerate}
\item
$\underline{q_0>0}$, shows an expanding universe which undergoes a deceleration phase.
This is the case of either a matter dominated universe or any pressureless barotropic fluid.
Observations do not favor $q_{0}>0$ at present times, which however appears relevant for early time cosmology,
where dark energy did not dominate over matter.
\item
$\underline{-1<q_0<0}$, represents an expanding universe which is currently speeding up. This actually represents
the case of our universe. The universe is thought to be dominated by some sort of anti-gravitational fluid, as stressed in Sec.~\ref{Intro}.
In turn, cosmography confirms such characteristics, without postulating any particular form of dark energy evolution.
\item
$\underline{q_0=-1}$, indicates that all the whole cosmological energy budget is dominated by a de Sitter fluid, i.e. a cosmic component with
constant energy density which does not evolve as the universe expands. This is the case of Inflation at the very early
universe. However at present-time this value is ruled-out by observations.
\end{enumerate}
Besides, the variation of the acceleration provides a way to understand whether the universe passes or not through a
deceleration phase. Precisely, the variation of acceleration, i.e. ${\rm d} q/{\rm d} z$, is related to $j$ as
\begin{equation}\label{jh35kdj}
\frac{{\rm d} q}{{\rm d} z}=\frac{j-2q^2-q}{1+z}\,.
\end{equation}
At present time, we therefore have $j_0=\frac{{\rm d} q}{dz}|_{0}+2q_0^2+q_0$ and since we expect $-1\leq q_0<-1/2$, we get that
$2q_0^2+q_0>0$. Thus if $q_0 < -1/2$, then $j_0$ is linked to the sign of the variation of $q$. We will confirm from observations that it actually lies in the interval $q_0<-1/2$.
Accordingly we can determine three cases:
\begin{enumerate}
\item $\underline{j_0<0}$ the universe does not show any departure from the present time accelerated phase.
This would indicate that dark energy influences early times dynamics, without any changes throughout the universe evolution.
Even though this may be a possible scenario, observations seem to indicate that this does not occur and it
is difficult to admit that the acceleration parameter does not change its sign as the universe expands.
\item $\underline{j_0=0}$ indicates that the acceleration parameter smoothly tends to a precise value, without changing
its behavior as $z\rightarrow\infty$. No theoretical considerations may discard or support this hypothesis, although
observations definitively show that a model compatible with zero jerk parameter badly fits current cosmological data.
\item
$\underline{j_0>0}$ implies that the universe acceleration started at a precise time during the evolution.
Usually, one refers to the corresponding redshift as the \emph{transition redshift}~\cite{transition}, at which dark energy effects
actually become significative. As a consequence, $j_0>0$ indicates the presence of a further cosmological \emph{resource}. By a direct measurement of the transition redshift $z_{tr}$, one would get relevant constraints on
the dark energy equation of state.
It turns out that the sign of $j_0$ corresponds to a change of slope of universe's dynamics. Rephrasing it differently, a positive $j_0$ definitively forecasts that the acceleration parameter should change sign at $z>z_{tr}$.
\end{enumerate}
A useful trick of cosmography is to re-scale the CS by means of the Hubble rate. In other words, it is possible to
demonstrate that if one takes into account $n$ cosmographic coefficients, only $n-1$ are really independent.
From definitions~(\ref{CSdef}), one can write
\begin{subequations}\label{Hpunto}
\begin{align}
\dot{H} =& -H^2 (1 + q)\,,\\
\ddot{H} =& H^3 (j + 3q + 2)\,,\\
\dddot{H} =& H^4 \left [ s - 4j - 3q (q + 4) - 6 \right]\,,\\
\ddddot{H} =& H^5 \left [ l - 5s + 10 (q + 2) j + 30 (q + 2) q + 24\right ]\,,
\end{align}
\end{subequations}
\noindent and we immediately see the correspondence between derivatives of the Hubble parameter and the CS
(note in particular the degeneracy, due to the fact that all these expressions are multiplied by $ H$).
As a consequence of the above discussion, one can choose a particular set of observable quantities and, expanding it,
as well as the scale factor, it is possible to infer viable limits on the parameters. To better illustrate this statement, by means
of Eqs.~(\ref{Hpunto}), one can infer the numerical values of the CS using the well-known luminosity distance.
In fact, keeping in mind the definition of the cosmological redshift $z$ in terms of the cosmic time $t$, that is
\begin{equation}\label{cosmoz}
\frac{{\rm d} { z} }{(1+z)}=-H(z){\rm d} t\,,
\end{equation}
then the luminosity distance in flat space can be expressed as
\begin{equation}\label{defDL}
d_L = (1+z)\chi(z),
\end{equation}
where
\begin{equation}\label{kjjgjhg}
\chi(z) =\int_{0}^{z}{\frac{dz'}{H(z')}}\,,
\end{equation}
is the comoving distance traveled by a photon from redshift $z$ to us, at $z=0$. The $d_L$ can be written as
\begin{eqnarray}\label{dLTaylor}
d_L = z\,d_H(H_0) \tilde{d}_L(z;q_0,j_0,\dots)\,,
\end{eqnarray}
where
\begin{equation}\label{hdhfkjd}
d_H(H_0)= \frac{1}{H_0}\,,
\end{equation}
and
\begin{eqnarray}\label{jhgkfh}
\tilde{d}_L(z;q_0,j_0,\dots) &=& 1 +\frac{1 - q_0}{2}z+\\
&-& \frac{1-q_0(1+3q_{0}) +j_0 }{6}z^2 + \mathcal{O}(z)\,.\nonumber
\end{eqnarray}
Further expansions up to order five in $z$ that will be used in this work are reported in the Appendix~\ref{appA}. Here, for brevity we reported in Eq. (\ref{jhgkfh}) the expansion up to the second order in $z$.
It is worth noticing that Eq.~(\ref{dLTaylor}) is general and applies to any cosmological model, provided it
is based on a flat FRW metric. Thus, by directly fitting the cosmological data for $d_L$, one gets physical bounds on $q_0$,
$j_0$, $s_0$ and $l_0$ \emph{for any cosmological model} (see~\cite{mine1,mine2}).
The above description, based on common Taylor expansions, represents only one of the possible approximations that one may use for the luminosity distance.
It may be argued that such an approximation does not provide adequate convergence for high redshift data.
Thus, our aim is to propose possible extensions of the standard Taylor treatment, i.e. Pad\'e approximants, that could accurately resolve the issues of standard cosmography.
In the next section we will present a different approximation for the luminosity distance, given by rational Pad\'e functions instead of Taylor polynomials,
we will analyze the relationship with the usual Taylor expansion
and argue that the rational approximation may be preferred from a theoretical point of view.
Later, in Secs.~\ref{Sect:DataSet} and~\ref{ParameterEstimations}
we will also perform the numerical comparison with observational data, in order to show that one can get improved results from this novel approach.
\section{Pad\'e approximations in the context of cosmography}\label{sec:padeandcosmography}
In this section we introduce the concept of Pad\'e approximants and we describe the applications of Pad\'e treatment to cosmography.
To do so, let us define the $(n,m)$ Pad\'e approximant of a generic function $f(z)$, which is given by the rational function
\begin{equation}\label{padedef}
P_{n m}(z) = \frac{a_0+a_1\,z+\dots+a_n\,z^n}{1+b_1\,z+\dots+b_m\,z^m}\,,
\end{equation}
with degree $n\geq0$ (numerator) and $m\geq0$ (denominator) that agrees with $f(z)$ and its derivatives at $z=0$ to the highest possible order, i.e. such that $P_{n m}(0)=f(0), P_{n m}'(0)=f'(0),\ldots, P_{n m}^{m+n}(0)=f^{m+n}(0)$~\cite{Padebook}. Pad\'e approximants for given $n$ and $m$ are unique up to an overall multiplicative constant. As a consequence, the first constant in the denominator is usually set to one, in order to face this scaling freedom. Hereafter, we follow this standard notation and indicate as $P_{nm}$ the Pad\'e approximant of degree $n$ at the numerator and $m$ in the denominator. As we see, in cosmology one may use direct data through the distance modulus $\mu$ of
different astronomical objects, such as e.g. supernovae. In the usual applications of cosmography, the luminosity distance
$d_L$, which enters $\mu$, is assumed to be a (truncated) Taylor series with respect to $z$ around present time $z=0$.
A problem with such a procedure occurs when one uses data out of the interval $z\geq1$. In fact, due to the divergence at high redshifts of the Taylor polynomials, this can possibly give non-accurate numerical results \cite{OrlChri}.
Consequently, data taken over $z>1$ are quite unlikely to accurately fit Taylor series. Pad\'e approximants can resolve this issue. In fact let us consider the general situation when one has to reconstruct a function, supposing to know the values of such a function, taken in the two limits where the independent variable is very small and very large respectively. Hence, let us consider two different approximate expansions of $d_L$. The first for small values of $z$ (around $z=0$), the second for large values of $z$ (around $1/z=0$).
The two approximations can be written as
\begin{subequations}\label{sus}
\begin{align}
d_L^0&\sim f_0+f'\,z+\frac{f''}{2}\,z^2\\
d_L^{\infty}&\sim g_0+g'\,\frac{1}{z}+\frac{g''}{2}\,\frac{1}{z^2}\,.
\end{align}
\end{subequations}
In this way, provided we construct a function that behaves as $d_L^0$ when $z \sim 0$ and as $d_L^\infty$ as $z \sim \infty$,
we are sure that in both limits such a function remains finite (when $z\sim 0$ and $z\sim \infty$ respectively). Given such a property, the most natural function able to interpolate our data between those two limits is naturally given by a \emph{rational function} of $z$. Pad\'e approximants are therefore adequate candidates to carry on this technique. In the next subsection we describe some problems associated to cosmography and to the Pad\'e formalism. Later, we also propose feasible solutions that we will adopt throughout this work.
\subsection{Pad\'e treatment to overcome cosmographic drawbacks}
We introduce this subsection to give a general discussion about several drawbacks plaguing the standard cosmographic approach.
For every single problem, we describe the techniques of solutions in the framework of Pad\'e approximations, showing how we treat Pad\'e approximants in order to improve the cosmographic analysis.
\begin{description}
\item[Degeneracy between coefficients]
Each cosmographic coefficient may be related to $H$, as previously shown.
This somehow provides that the whole list of independent parameters is really limited to $q_0,j_0,s_0,l_0,\ldots$. However,
one can think of measuring $ H_0$ through cosmography in any case, assuming $ H_0$ to be a cosmographic coefficient, without loss of generality.
The problem of degeneracy unfortunately leads to the impossibility of estimating $ H_0$ alone by using measurements of the distance modulus,
\begin{equation}
\mu(z) = 5\log_{10}\left(\frac{d_L(z)}{\text{Mpc}}\right) + 25\,,
\end{equation}
in the case of supernova observations, as we will see later. From Eq.~(\ref{dLTaylor}) it follows that $d_L$ can be factorized into two pieces: $d_H$ and $\tilde{d}_L$.
Since $d_H \equiv H_0^{-1}$, therefore it
depends only on $H_0$, thus becoming
an additive constant in $\mu(z)$ which cannot be estimated, its only effect is to act as a lever to
the logarithm of $\tilde{d}_L(z;q_0,j_0,\dots$).
In other words, $H_0$ degenerates
with the rest of the parameters. To alleviate such problem, we here make use of two different data sets, together with supernovae, i.e.
the Hubble measurements and the Hubble telescope measure. In this way we employ direct measures of $ H_0$, thus reducing
the errors associated to the degeneracy between cosmographic parameters.
\item[Degeneracy with scalar curvature] Spatial curvature of the FRW model
enters the luminosity distance, since the metric directly depends on it. Thus, geodesics of photons correspondingly change due to its value.
Therefore, any expansion of $d_L$ depends on $\Omega_k$ as well, degenerating the values of the CS with respect to $\Omega_k$.
The jerk parameter is deeply influenced by the value of the scalar curvature and degenerates with it. In our work, we
overcome this problem through geometrical bounds on $\Omega_k$, determined by early time observations.
According to recent measurements, the universe is considered to be spatially flat and any possible
small deviations will not influence the simple case $\Omega_k=0$.
This is the case we hereafter adopt, except for the last part in which we extensively investigate the role of $\Omega_k$ in the framework of the $\Lambda$CDM model.
\item[Dependence on the cosmological priors]
The choice of the cosmological priors may influence the numerical outcomes derived from our analyses.
This turns out to be dangerous in order to determine the signs of the cosmographic coefficients.
However, to alleviate this problem we may easily enlarge all the cosmological priors, showing that within convergence ranges
the CS are fairly well constrained.
The corresponding problem would indicate possible departures from convergence limits, if ranges are outside the theoretical expectations.
Hence, we found a compromise for each cosmological interval, and we report the whole list of numerical priors in Table~\ref{tab:priors}.
\item[Systematics due to truncated series]
Slower convergence in the best fit algorithm may be induced by choosing truncated series at a precise order,
while systematics in measurements occur, on the contrary, if series are expanded up to a certain order.
In other words, introducing additional terms would decrease convergence accuracy,
although lower orders may badly influence the analysis itself. To alleviate this problem, we will constrain the parameters
through different orders of broadening samples. In this way, different orders will be analyzed and we will show no significative departures from our truncated series order.
\item[Dependence on the Friedmann equations]
Dark energy is thought to be responsible of the present time acceleration.
However, cosmography is able to describe the current universe speeding up without the need of postulating a precise dark energy fluid \emph{a priori}.
This statement is clearly true only if a really barotropic fluid is responsible for the dark energy effects.
In case there is no significative deviations from a constant equation of state for pressureless matter and dark energy is provided by
some modification of gravitation, cosmography should be adjusted consequently.
This leads to the implicit choice of assuming general relativity as the specific paradigm to get constraints on the cosmographic observables.
One may therefore inquire to what extent cosmography is really independent of the Friedmann equations.
Rephrasing it differently, to reveal the correct cosmological model we do not fix further assumptions,
e.g. geometrical constraints, Lorentz invariance violations, and so forth, since we circumscribe our analysis to general relativity only.
Any possible deviations from the standard approach would need additional theoretical bounds and the corresponding
CS should be adjusted accordingly.
However, this problem does not occur in this work and we {can} impose limits without the need of particular assumptions at
the beginning of every analysis.
\item[Convergence] The convergence problem probably represents the most spinous issue of cosmography.
As we have previously stated, the problem of truncated series is intimately
intertwined to the order chosen for determining the particular Taylor expansion under exam.
Unfortunately, almost all cosmological data sets exceed the bound $z\simeq 0$,
which represents the value around which one expands $d_L$ into a series.
In principle, all Taylor series are expected to diverge when $z\gg 1$, as a consequence of the fact that they are polynomials.
Thus, finite truncations get problems to adapt to data taken at $z\gg1$, leading to possible misleading outcomes.
For example, this often provides additional systematic errors because it is probable that the increase of bad convergence may affect numerical results.
Here, we improve accuracy by adopting the union 2.1 supernovae data set and the two additional surveys based on measurements of $H(z)$,
i.e. direct Hubble measures and Hubble space telescope measurements.
Combining these data together naturally eases the issue of systematics,
whereas to overcome finite truncation problems we manage to develop the so called Pad\'e approximation for different orders.
By construction, since Pad\'e approximants represent a powerful technique to approximate functions by means of ratios of polynomials,
one easily alleviates convergence problems for $z\gg1$.
As such, we expect Pad\'e approximants can better approximate the luminosity distance with respect to standard Taylor treatments,
especially when high redshift data sets are employed in the analysis.
On the other hand, in order to overcome the problem of divergence, precision cosmology employs the use of several re-parametrizations of the redshift $z$,
in terms
of \emph{auxiliary variables} ($\mathcal Z_{new}$), which enlarge the convergence radius of the Taylor expansion to a sphere of radius $\mathcal Z_{new}<1$.
Rephrasing it differently, supposing that data lie within $z\in[0,\infty)$, any auxiliary variable restricts the interval in a more stringent (non-divergent) range.
A prototype of such an approach is for example given by $y_1 = \frac{z}{1+z}$ (see e.g.~\cite{Cattoen:2007sk}),
whose limits in the past universe (i.e. $z \in [0,\infty)$) read $y_1\in [0,1]$, while in the future (i.e. $z \in [-1,0]$) read $y_1\in (-\infty,0]$.
The construction of any auxiliary variable should satisfy some additional requirements.
It must be feasible to invert it, passing from the redshift $z$ to it, being one to one invertible.
Moreover, it should not diverge for any values of the redshift $z$
(in this sense, $y_1$ suffers from a divergence problem at future times).
Finally, any parametrization needs to behave smoothly as the universe expands, without showing any critical points.
In this work, we also compare Pad\'e expansions with the auxiliary variables proposed in the literature, namely $y_1$,
already cited, and $y_4=\arctan z$, which has been introduced in~\cite{mine2}.
The variable $y_4$ improves $y_1$, since it has been constructed by following the above mentioned recipe to build up \emph{ad hoc} auxiliary variables. One of our results is that auxiliary variables, albeit being well consolidated tricks for reducing the convergence problem,
behave worse than Pad\'e approximations.
This is probably due to the unknown form of the correct $\mathcal Z_{new}$, which is not known \emph{a priori}.
To do so, we describe in detail differences between our new technique of cosmographic investigations which uses Pad\'e approximations
versus standard approaches which make use of auxiliary variables,
showing that the convergence problem may be definitively healed through the use of rational approximations, instead of constructing auxiliary variables.
\end{description}
In the next subsection we demonstrate with the help of exactly soluble models that Pad\'e approximations
indeed improve the accuracy in approximating the luminosity distance. We stress the fact that this property is more significative as data span over larger intervals of $z$, i.e. $z\gg1$.
\subsection{Taylor versus Pad\'e for exact cosmological models}
\label{comparisonLCDM}
In this section, we give a qualitative representation of the improvements that one gets by performing a Pad\'e approximation of the luminosity distance.
To do so, we plot $d_L$ for known models and compare it with the numerical behavior of different Taylor and Pad\'e approximations over a large range of $z$.
As two significative examples, we work with the $\Lambda$CDM and $\omega$CDM models for our elucidative purposes~\cite{galaxyluongo}.
Afterwards, we infer a theoretical method to focus on viable Pad\'e approximations, by treating powers $n$ and $m$, without comparing with particular models.
\textbf{The $\Lambda$CDM model:}
The Hubble parameter for the $\Lambda$CDM model reads
\begin{equation}
H(z)=\sqrt{\Omega_m(1+z)^3+\Omega_\Lambda}\,,
\end{equation}
where $\Omega_\Lambda=1-\Omega_{m}$, representing the dark energy density, expressed in terms of a pure cosmological constant. Performing a numerical integration of $d_L$~(\ref{defDL}),
we can plot such function over the interval of interest, which is arbitrarily fixed inside $z\in[0,10]$. Besides, we can also compute different
Taylor and Pad\'e approximations for this function, and graph all the results, to show that the
approximation is generally improved with the use of rational functions. In Fig.~\ref{LCDMDL},
we present the plots of the exact $\Lambda$CDM luminosity distance,
compared with its approximations obtained using a Taylor polynomial and Pad\'e functions,
for different orders of approximation. In particular, the Taylor polynomial of degree three is plotted
together with the Pad\'e approximants of degree $(1,1)$, $(1,2)$ and $(2,1)$,
the polynomial of fourth degree together with the Pad\'e approximants $(1,3)$, $(3,1)$ and $(2,2)$
and finally the fifth order Taylor polynomial is compared with the Pad\'e functions $(1,4)$, $(3,2)$, $(2,3)$ and $(4,1)$.
We remind that e.g. the Taylor polynomials of third degree and the Pad\'e approximants $(1,2)$ and $(2,1)$
have the same number of free parameters and they agree by definition up to the third order of derivatives at present time.
The same holds to higher orders of both Taylor and Pad\'e approximations. Therefore, the situation described by the Taylor and Pad\'e approximants can also be seen as having
two different models which give approximately the same values for the CS parameters,
albeit providing different evolutions over the whole interval considered.
As one can immediately notice from all plots in Fig.~\ref{LCDMDL}, Taylor
approximations (in blue in the figures) are really accurate until $z$ stays small, whereas they rapidly diverge from the exact curve (in red) as $z>2$.
On the contrary, we can see from the first plot that the rational approximant $P_{21}$ keeps very close to the exact function over the complete interval analyzed.
Moreover, as we see in the second and third plots, the situation is the same as we increase the order of the approximants.
In fact, the Pad\'e functions $P_{22}$ and $P_{32}$ fairly approximate the exact $\Lambda$CDM luminosity distance over all the interval considered.
In particular, we remark also that the correctness of the approximation not necessarily increases by increasing the order of the approximants
(as expected, since all the possible Pad\'e functions have completely different behaviors, depending on the degrees of numerator and denominator)
and that $P_{21}$, $P_{22}$ and $P_{32}$ seem to be the best approximations, within the ones considered, giving excellent results.
\begin{figure*} [ht]
\begin{center}
\includegraphics[width=3.5in]{PadeLCDMT3.eps}
\includegraphics[width=3.5in]{PadeLCDMT4.eps}
\includegraphics[width=3.5in]{PadeLCDMT5.eps}
{\small \caption{Analytical curves for the luminosity distance of the $\Lambda$CDM model (in red) compared
to its Taylor and Pad\'e approximations. As we see, the Taylor polynomials $T3$, $T4$ and $T5$ tends to quickly diverge outside the region $z\leq2$.
At the same time, not all the Pad\'e approximants give good approximations of this model. For example, $P_{11}$, $P_{13}$ and $P_{23}$ give spurious singularities
when used to approximate the $\Lambda$CDM model. We will see how to avoid this problem in the numerical analysis. On the contrary, $P_{21}$, $P_{22}$ and $P_{32}$
give excellent approximations to the $d_{L}$ derived from assuming $\Lambda$CDM.
Additional comments have been reported in the text.}
\label{LCDMDL}}
\end{center}
\end{figure*}
As a good check for our conclusions, we repeat such considerations by using a different model, i.e. the $\omega$CDM model, probably representing the first step beyond the $\Lambda$CDM model.
\textbf{The $\omega$CDM model:}
The Hubble parameter resulting from the $\omega$CDM model reads
\begin{equation}
H(z)=\sqrt{\Omega_m(1+z)^3+\Omega_Q\,(1+z)^{3(1+\omega)}}\,,
\end{equation}
where $\Omega_Q=1-\Omega_{m}$ and $\omega$ is a
free parameter of the model that lies in the interval $\omega\in (-1,-0.8)$. The exact integral of $d_L$ involves here a hypergeometric function.
We plot it over the interval of interest, which is again $z\in[0,10]$. Besides, we can also compute different Taylor and Pad\'e approximations for this function, plotting all the results, to show that the
approximation is generally improved with the use of rational functions as well as in the $\Lambda$CDM case.
Moreover, all the comments presented for the $\Lambda$CDM model also apply for the $\omega$CDM case,
showing that the Pad\'e approximants give a better description of the exact luminosity distance over the full interval considered, as one can
see in Fig.~\ref{wCDMDL}.
\begin{figure*} [ht]
\begin{center}
\includegraphics[width=3.5in]{PadeT3.eps}
\includegraphics[width=3.5in]{PadeT4.eps}
\includegraphics[width=3.5in]{PadeT5.eps}
{\small \caption{Analytical curves for the luminosity distance of the $\omega$CDM model (in red) compared
to its Taylor and Pad\'e approximations. As we see, the Taylor polynomials $T3$, $T4$ and $T5$ tends to quickly diverge outside the region $z\leq2$.
At the same time, not all the Pad\'e approximants give good approximations of this model. For example, $P_{11}$, $P_{13}$, $P_{23}$ and $P_{41}$ give spurious singularities
when used to approximate the $\omega$CDM model. We will see how to avoid this problem before our numerical analysis. On the contrary, $P_{21}$, $P_{22}$ and $P_{32}$
give excellent approximations to the $d_{L}$ derived from assuming $\omega$CDM.
More comments in the text.}
\label{wCDMDL}}
\end{center}
\end{figure*}
Exactly as in the $\Lambda$CDM case, the best approximations are given by $P_{2 1}$, $P_{22}$ and $P_{3 2}$.
To conclude, Figs.~\ref{LCDMDL} and~\ref{wCDMDL} clearly show that, provided we are given data over a large interval of values for the cosmological redshift $z$,
it would be better to fit the observed luminosity distances with a rational function,
in order to get a more realistic function that fits such data over the whole interval.
By the same reasoning, the use of Pad\'e approximants seems to be also more convenient in order to infer the evolution of $d_L$
from knowledge of the CS. In particular, it seems that the Pad\'e approximants $P_{2 1}$, $P_{2 2}$ and $P_{3 2}$ give the best approximations,
which strongly suggests that the order of the numerator and that of the denominator for these models should be very close to each other,
with the former possibly being greater than the latter.
Given this fact, in the next section we will give a quantitative analysis of different Pad\'e approximations for the luminosity distance,
by comparing them with the astronomical data. In this way we get the best values for the CS parameters by a direct fit using different forms.
As we will see, this novel approach can give better bounds on the parameters, and takes better account of more distant objects,
as the Pad\'e approximation over a large interval is more reliable than Taylor's technique.
The above considerations suggest some theoretical conclusions to build up viable Pad\'e rational functions.
Here, we formalize a {possible} recipe to determine which Pad\'e rational functions are favored {with respect to} others.
First, the Pad\'e function should smoothly evolve in the redshift range chosen for the particular cosmographic analysis.
Naively, this suggests that any possible Pad\'e approximant should not have singularities in the observable redshift intervals.
Moreover, any Pad\'e {approximant for $d_{L}$} must be positive definite and cannot show negative regions,
otherwise the definition of magnitude would not hold at all.
{Finally, we expect that the degree of the numerator and that of the denominator should be close, with the former a little greater than the latter.}
Keeping this in mind, we are ready to perform our experimental analyses.
To do so, we consider some Pad\'e expansions as reported in the following sections.
\section{Experimental analysis with Pad\'e functions}\label{Sect:DataSet}
In this section we present the main aspects of our experimental analysis. We illustrate how we
directly fit general expressions of $d_L$, in terms of different types of approximations, i.e. Taylor
(standard cosmographic approach), auxiliary variables and Pad\'e expansions (our novel cosmographic technique).
In general all Pad\'e approximations, due to their rational forms, may show spurious singularities for certain
values of the redshift $z$ lying in the interval of data. In other words, the need of constructing precise Pad\'e
approximations which are not plagued by divergences due to poles, is actually one of the tasks of our analysis.
In particular, a simple manner to completely avoid such a problem consists in the choice of suitable
priors for the free parameters, built up \emph{ad hoc}, shifting any possible poles to future time cosmological
evolution. We show that data are confined inside intervals of the form $z\in[0,\infty)$, whereas possible divergences
of Pad\'e functions are limited to future
times, i.e. $z\leq-1$, and hence do not influence our experimental analysis.
Moreover, the cosmological
priors adopted here are perfectly compatible with the ones proposed in several previous papers and do not influence the numerical outcomes.
This shows that the Pad\'e method does not reduce the accuracy in fitting procedures and it is a good candidate to improve standard methods of cosmographic analyses. Thus, let us investigate
the improvements of the Pad\'e treatments with respect to standard techniques.
To do so, we denote the cosmographic parameters
by a suitable vector ${\theta}$, whose dimension changes depending on how many coefficients we are going
to analyze in a single experimental test. Estimations of the cosmographic parameters have been performed
through Bayesian techniques and best fits have been obtained by maximizing the likelihood function, defined as
\begin{equation}\label{jhfdkjdf}
\mathcal{L}
\propto \exp (-\chi^2/2 )\,,
\end{equation}
where $\chi^2$ is the common {\it (pseudo) $\chi$-squared function}, whose form is explicitly determined for
each data set employed. Maximizing the likelihood function leads to minimizing the pseudo-$\chi$-squared
function and it can be done by means of a direct comparison with each cosmological data set.
For our purposes, we describe three statistical data sets, characterized by different maximum order of parameters,
providing a hierarchy among parameters. This procedure leads to a broadening of the sampled distributions
if the whole set of parameters is wider, i.e. if the dimension of ${\theta}$ is higher. As a consequence, the numerical
outcomes may show deeper errors, which may be healed by means of the above cited priors. We make use of
the supernova union 2.1 compilation from the \emph{Supernovae Cosmology Project}~\cite{Suzuki:2011hu}, i.e.
free available data of the most recent and complete supernova survey. Further, we employ a Gaussian prior
on the present-time Hubble parameter, i.e. $H_0 = 73.8 \pm 2.4 \,\text{km/s/Mpc}$~\cite{Riess:2011yx} implied
by the \emph{Hubble Space Telescope} (HST) measurements, and we also consider the almost model independent Baryonic Acoustic Oscillation, {\it BAO ratio},
as proposed in~\cite{Percival:2009xn}. In addition, we use relevant measurements of the Hubble parameter $H(z)$
at 26 different redshifts spanning from $z=0.09$ to
$z=2.3$~\cite{Simon:2004tf,Stern:2009ep,Moresco:2012jh,Gaztanaga:2008xz, Scrimgeour:2012wt, Busca:2012bu},
commonly named \emph{Observational Hubble Data} (OHD) or differential Hubble measurements.
The cosmological priors that we have employed here are summarized in Tab.~\ref{tab:priors}, in which we report the largest numerical interval developed for any single variable.
\begin{widetext}
\begin{center}
\newlength{\mywidth}
\setlength{\mywidth}{0.54\textwidth}
\begin{table}
\begin{center}
\begin{tabular}{ccccc}
\begin{tabular*}{\mywidth}{cccccc}
\hline\hline
\begin{tabular}{rclclrcl}
\phantom{ccccc} & Flat priors & \phantom{cc} & \vline & \phantom{cc} & Additional & Constraints & \\ \hline
$0.5\quad <$ & $h$ & $< \quad 0.9$ & \vline & \phantom{cc} & \phantom{cc} & & \\
$0.001 \quad < $&$\Omega_{\rm b}h^2$ & $< \quad 0.09$ & \vline & \phantom{cc} & $\Omega_{k}\quad =$&$0$ & \phantom{cc}\\
$0.01 \quad <$&$\Omega_{\rm dm}h^2$ & $< \quad 0.25$ & \vline & \phantom{cc} & $\Omega_{m}\quad<$&$0.5$ & \phantom{cc} \\
$-1000\, \quad <$&$ {\tilde \Theta}$ & $< \quad 1000\,$ & \vline &\phantom{cc} & \phantom{cc}& &\\
\end{tabular}\\\hlin
\end{tabular*}
\end{tabular}
\end{center}
\caption{Priors imposed on the free parameters involved in the Bayesian analysis for all cosmographic tests here employed. The parameter $h$ is the normalized Hubble rate, while $\tilde\Theta$ indicates a generic cosmographic coefficient ($q_0,j_0,s_0,\ldots$). We also report geometrical consequences on scalar curvature and the whole matter density.}\label{tab:priors}
\end{table}
\end{center}
\end{widetext}
Now, we are ready to investigate whether and how much Pad\'e approximants are favored for estimating bounds on the late time universe.
To better illustrate the procedure, we report below the $\chi^2$ function for each of the data sets adopted in the numerical analysis.
\subsection{Supernova type Ia compilation}
Type Ia supernovae observations have been extensively analyzed during the last decades for parameter-fitting
of cosmological models. They are considered as standard candles, i.e. quantities whose luminosity curves are intimately related
to distances. In our work, we employ the most recent survey of supernovae compilations, namely union 2.1~\cite{Suzuki:2011hu},
which extends previous versions union and union 2 data sets~\cite{Amanullah:2010vv,Kowalski:2008ez}. Here, systematics is reduced
and does not influence numerical outcomes, as for previous surveys.
The standard fitting procedure relies on using a Gaussian {\it $\chi$-squared} function, evaluating differences between
theoretical and observational distance modulus $\mu(z_i; {\theta}) - \mu_{obs}(z_i)$.
Nevertheless, the presence of nuisance parameters as the Hubble factor $ H_0$ and absolute magnitude $M$ enforces
to marginalize over. Straightforward calculations provide~\cite{GoliathAmanullah}
\begin{equation}
\chi^2_{SN} = A - \frac{B^2}{C} + \log \left( \frac{C}{2\pi}\right)\,,
\end{equation}
where we defined
\begin{eqnarray}
A &=& {\bf x}^T \mathcal{C}^{-1} {\bf x}\,, \nonumber \\
B &=& \sum_i (\mathcal{C}^{-1}{\bf x})_i\,, \\
C &=& \text{Tr}[\,\mathcal{C}^{-1}\,]\,. \nonumber
\end{eqnarray}
Here $\mathcal{C}$ represents the covariance matrix of observational data, including statistical and systematic errors
as well, and the $i$-th component of the vector ${\bf x}$ given by
\begin{equation}
{\bf x}_i= 5\log_{10}\left(\frac{d_L(z_i; { \theta})}{\text{Mpc}}\right) + 25 - \mu_{obs}(z_i)\,.
\end{equation}
\subsection{Ratio of baryonic acoustic oscillation}\label{subsec:BAO}
The ratio of baryonic acoustic oscillation (BAO) is slightly model dependent~\cite{Durrer2011}, since acoustic scales actually depend
on the redshift (drag time redshift), inferred from a first order perturbation theory. However, baryonic acoustic oscillations
determined from~\cite{Percival:2009xn} have been found in terms of a model independent quantity, i.e.
\begin{equation}
\mathcal{B_R}\equiv D_V(0.35)/ D_V(0.20) = 1.736 \pm 0.065\,,
\end{equation}
where the volumetric distance is defined as
\begin{equation}\label{defDV}
D_V(z) = \left[c \,z\, d_L(z)^2/(1+z)^2H(z) \right]^{1/3}\,.
\end{equation}
The BAO ratio $\chi$-squared function is simply given by
\begin{equation}
\chi^2_{BAOr}(\theta) = \frac{(D_V(0.35; \theta)/ D_V(0.20; \theta) - 1.736)^2}{0.065^2}\,.
\end{equation}
\noindent We describe below the procedure to compute $D_{V}$ by means of the Pad\'e expansions for $d_{L}$. First, one needs to compute the approximation of $H(z)$ in terms of $d_L$ by inverting Eq. (\ref{defDL}). It follows
\begin{equation}\label{HfromDL}
H(z)=\left[\frac{\rm d}{{\rm d} z}\left(\frac{d_L}{1+z}\right)\right]^{-1}\,.
\end{equation}
Afterwards, inserting $H(z)$ from Eq. ($\ref{HfromDL}$) and the Pad\'e expressions for $d_{L}$ into Eq. (\ref{defDV}), one obtains the corresponding approximations for $D_V$, as reported in Appendix A.
\subsection{Direct Hubble measurements}\label{ohd}
We use 26 independent OHD data from~\cite{Simon:2004tf,Stern:2009ep,Moresco:2012jh,Gaztanaga:2008xz, Scrimgeour:2012wt, Busca:2012bu}, as reported in the appendix of this work.
We use those data, following~\cite{Simon:2004tf,Stern:2009ep,Moresco:2012jh}, in which a novel approach to track the universe expansion history has
been proposed, employing massive early type galaxies as cosmic chronometers~\cite{JimenezLoeb}. The technique
allows one to estimate the quantity $dt/dz$, sometimes referred to as \emph{differential time}, which is related to the Hubble rate by
\begin{equation}\label{njvnjfk}
H(z) = - (1+z)^{-1}dz/dt\,.
\end{equation}
Bearing in mind Eq.~(\ref{njvnjfk}), a preliminary list of 19 numerical outcomes has been found, whereas the
other 7 data have been determined from the study of galaxy surveys: 2 from~\cite{Gaztanaga:2008xz}, 4 from the wiggle Z collaboration~\cite{Scrimgeour:2012wt}
and one more from~\cite{Busca:2012bu}. All Hubble estimates are uncorrelated, therefore
the $\chi$-$squared$ function is simply given by
\begin{equation}
\chi^2_{OHD}({\theta}) = \sum_{i} \frac{(H(z_i; {\theta}) - H_{obs}(z_i))^2 }{\sigma_i^2}\,.
\end{equation}
In the Appendix~\ref{appA}, as already stressed, we provide Tab. \ref{table:OHD}, where we summarize the OHD data used in this paper.
\subsection{The fitting procedure}\label{fitting}
Due to the fact that the different data sets are uncorrelated, the total $\chi$-$squared$ function is given by
\begin{equation}
\chi^2({\theta}) = \chi^2_{\text{SN}} + \chi^2_{\text{OHD}} + \chi^2_{\text{BAOr}} + \chi^2_{\text{HST}}\,.
\end{equation}
The best fit to the data is given by those parameters that maximize the likelihood function $\mathcal{L} \propto \exp (-\chi^2/2)$. We obtain them and their
respective confidence intervals
by using a Metropolis-Hasting Markov Chain Monte Carlo (MCMC) algorithm~\cite{Hastings70,Metropolis53} with the publicly
available CosmoMC code~\cite{Lewis:2002ah,cosmomc_notes}.
We run several independent chains and to probe their
convergence we use the Gelman-Rubin criteria $R \sim \text{``mean of chains variances''} / \text{``variance of chains means''}$ \cite{Gelman:1992zz}
with $R-1 < 0.01$. We accurately modify the priors for each ${\theta}$, within the interval of values reported in Tab.~\ref{tab:priors}.
\section{Estimation of the cosmographic series} \label{ParameterEstimations}
\begin{figure}
\begin{center}
\includegraphics[width=3.4in]{fig1a.eps}
{\small \caption{The Pad\'e approximants used for the different ${\theta}$. The approximations enclosed in the triangle give
conclusive results.}
\label{fig:PA}}
\end{center}
\end{figure}
\begin{figure*} [ht]
\begin{center}
\includegraphics[width=2in]{fig2a.eps}
\includegraphics[width=2in]{fig2b.eps}
\includegraphics[width=2in]{fig2c.eps}
{\small \caption{(color online) 1-dimensional marginalized posteriors for the
parameter set $\mathcal{A}$. $SC$ stands for standard cosmography.}
\label{fig:1dimModelA}}
\end{center}
\end{figure*}
For the parameters estimation we will use the CS combined
in three sets with different maximum order of
parameters:
\begin{eqnarray}
\mathcal{A} &=& \{ H_0, q_0, j_0\}\,, \nonumber\\
\mathcal{B} &=& \{ H_0, q_0, j_0, s_0 \}\,,\\
\mathcal{C} &=& \{ H_0, q_0, j_0, s_0, l_0 \}\,. \nonumber
\end{eqnarray}
For the parameters set $\mathcal{A}$ the corresponding Pad\'e approximants are $P_{12}$ and $P_{21}$,
as shown in Fig.~\ref{fig:PA}. Of those approximants, only $P_{21}$ gives
conclusive results. For $\mathcal{B}$ we obtain conclusive results for $P_{31}$ and $P_{22}$
and for $\mathcal{C}$ for $P_{32}$, $P_{23}$ and $P_{41}$.
In Tables~\ref{table:1DResults1},~\ref{table:1DResults2}, and~\ref{table:1DResults3} we show
the best fits and their $1\sigma$-likelihoods for the parameters sets
$\mathcal{A}$, $\mathcal{B}$ and $\mathcal{C}$ respectively.
We also show the estimated CS obtained by the standard cosmography ($SC$) or Taylor approach.
We worked out the $\Lambda$CDM model, which is for our purposes and the redshifts involved sufficiently described by two parameters: $\Omega_m h^2$ and $\Theta$.
Here $\Theta$ is defined as 100 times the ratio of the sound horizon to the angular diameter distance at recombination,
while as usual $\Omega_m$ is the abundance of matter density (both baryonic $\Omega_b$ and dark matter $\Omega_{dm}$), and $h$ is
the dimensionless Hubble parameter, as reported in Tab.~\ref{tab:priors}.
The best fits, using the same data sets as above, are given by $\Omega_m h^2 = 0.148${\tiny ${}_{-0.010}^{+0.012}$} and
$\Theta = 1.041${\tiny ${}_{-0.010}^{+0.011}$}. From these values and the formulas
\begin{eqnarray} \label{LCDMCS}
q_0 &=& -1 + \frac{3}{2}\Omega_m\,, \nonumber\\
j_0 &=& 1\,, \nonumber\\
s_0 &=& 1 - \frac{9}{2} \Omega_m\,, \\
l_0 &=& 1 + 3 \Omega_m - \frac{27}{2}\Omega_m^2\,, \nonumber
\end{eqnarray}
which are valid only for flat $\Lambda$CDM model, we have also estimated the cosmographic parameters and
we report them in Tables~\ref{table:1DResults1},~\ref{table:1DResults2}, and~\ref{table:1DResults3}.
\begin{table*}
\caption{{\small Table of best fits and their likelihoods (1$\sigma$) for the parameter set $\mathcal{A}$. $SC$ stands for the standard cosmography approach and
the $\Lambda$CDM derived columns are the parameters inferred assuming as valid the $\Lambda$CDM model.}}
\begin{tabular}{c|c|c|c}
\hline\hline
{\small $\quad$ Parameter $\quad$} & {\small $\qquad$ $P_{21}$ $\qquad$ }
& {\small $\qquad$ $SC$ $\qquad$} & {\small $\quad$ $\Lambda$CDM derived $\quad$}\\
\hline
{\small$H_0$} & {\small $70.64$}{\tiny ${}_{-2.63}^{+2.77}$}
& {\small $71.98$}{\tiny ${}_{-2.55}^{+2.48}$} & {\small $71.68$}{\tiny${}_{-2.16}^{+2.25}$}\\[0.8ex]
{\small$q_0$} & {\small $-0.4712$}{\tiny ${}_{-0.1106}^{+0.1224}$}
& {\small $-0.5701$}{\tiny ${}_{-0.0928}^{+0.1057}$} & {\small $-0.6117$}{\tiny${}_{-0.0365}^{+0.0401}$}\\[0.8ex]
{\small$j_0$} & {\small $0.593$}{\tiny ${}_{-0.210}^{+0.216}$}
& {\small $0.766$}{\tiny ${}_{-0.207}^{+0.211}$} & {\small $1$}\\[0.8ex]
\hline \hline
\end{tabular}
{\tiny Notes.
a. $H_0$ is given in Km/s/Mpc units.
}
\label{table:1DResults1}
\end{table*}
\begin{table*}
\caption{{\small Table of best fits and their likelihoods (1$\sigma$) for the parameter set $\mathcal{B}$. $SC$ stands for the standard cosmography approach and
the $\Lambda$CDM derived columns are the parameters inferred assuming as valid the $\Lambda$CDM model.}}
\begin{tabular}{c|c|c|c|c}
\hline\hline
{\small $\quad$ Parameter $\quad$} & {\small $\qquad$ $P_{31}$ $\qquad$ } & {\small $\qquad$ $P_{22}$ $\qquad$ }
& {\small $\qquad$ $SC$ $\qquad$} & {\small $\quad$ $\Lambda$CDM derived $\quad$}\\
\hline
{\small$H_0$} & {\small $71.76$}{\tiny${}_{-3.46}^{+ 3.38}$} & {\small $71.71$}{\tiny${}_{-3.15 }^{+3.35}$}
& {\small $72.53$}{\tiny ${}_{-3.51}^{+3.53}$} & {\small $71.68$}{\tiny${}_{-2.16}^{+2.25}$}\\[0.8ex]
{\small$q_0$} & {\small $-0.6483$}{\tiny${}_{-0.1623}^{+0.2589}$} & {\small $-0.6767$}{\tiny${}_{-0.2580}^{+0.2395}$}
& {\small $-0.6642$}{\tiny${}_{-0.1963}^{+0.2050}$} & {\small $-0.6117$}{\tiny${}_{-0.0365}^{+0.0401}$}\\[0.8ex]
{\small$j_0$} & {\small $1.313$}{\tiny${}_{-0.917}^{+0.521}$} & {\small $1.500$}{\tiny${}_{-1.009}^{+0.973}$}
& {\small $1.223$}{\tiny ${}_{-0.664}^{+0.644}$} & {\small $1$}\\[0.8ex]
{\small$s_0$} & {\small $0.425$}{\tiny${}_{-0.841}^{+1.079}$} & {\small $0.681$}{\tiny${}_{-1.055}^{+2.367}$}
& {\small $0.394$}{\tiny ${}_{-0.731}^{+1.335}$} & {\small $-0.165$}{\tiny${}_{-0.120}^{+0.109}$}\\[0.8ex]
\hline \hline
\end{tabular}
{\tiny Notes.
a. $H_0$ is given in Km/s/Mpc units.
}
\label{table:1DResults2}
\end{table*}
\begin{table*}
\caption{{\small Table of best fits and their likelihoods (1$\sigma$) for the parameter set $\mathcal{C}$. $SC$ stands for the standard cosmography approach and
the $\Lambda$CDM derived columns are the parameters inferred assuming as valid the $\Lambda$CDM model.}}
\begin{tabular}{c|c|c|c|c|c}
\hline\hline
{\small $\quad$ Parameter $\quad$} & {\small $\qquad$ $P_{41}$ $\qquad$ } & {\small $\qquad$ $P_{32}$ $\qquad$ }
& {\small $\qquad$ $P_{23}$ $\qquad$ }
& {\small $\qquad$ $SC$ $\qquad$} & {\small $\quad$ $\Lambda$CDM derived $\quad$}\\
\hline
{\small$H_0$} & {\small $71.56$}{\tiny${}_{-3.95}^{+3.95}$} & {\small $71.83$}{\tiny ${}_{-3.64}^{+3.53}$}
& {\small $70.75$}{\tiny ${}_{-3.12}^{+3.41}$}
& {\small $71.38$}{\tiny ${}_{-3.68}^{+4.01}$} & {\small $71.68$}{\tiny${}_{-2.25}^{+2.16}$}\\[0.8ex]
{\small$q_0$} & {\small $-0.5516$}{\tiny${}_{-0.4556}^{+0.3190}$} & {\small $-0.7189$}{\tiny ${}_{-0.4631}^{+0.3397}$}
& {\small $-0.5539$}{\tiny ${}_{-0.2171}^{+0.2966}$}
& {\small $-0.6173$}{\tiny ${}_{-0.3139}^{+0.3658}$} & {\small $-0.6117$}{\tiny${}_{-0.0401}^{+0.0365}$}\\[0.8ex]
{\small$j_0$} & {\small $0.721$}{\tiny${}_{-1.982}^{+2.489}$} & {\small $1.959$}{\tiny ${}_{-2.516}^{+3.290}$}
& {\small $0.710$}{\tiny ${}_{-1.499}^{+1.389}$}
& {\small $0.949$}{\tiny ${}_{-1.686}^{+1.374}$} & {\small $1$}\\[0.8ex]
{\small$s_0$} & {\small $-1.060$}{\tiny${}_{-3.193}^ {+7.341}$} & {\small $1.950$}{\tiny ${}_{-5.524}^{+14.072}$}
& {\small $-1.203$}{\tiny ${}_{-2.864}^{+3.073}$}
& {\small $-0.797$}{\tiny ${}_{-3.585}^{+2.962}$} & {\small $-0.165$}{\tiny${}_{-0.109}^{+0.120}$}\\[0.8ex]
{\small$l_0$} & {\small $4.43$}{\tiny${}_{-2.98}^{+19.79}$} & {\small $8.14$}{\tiny ${}_{-7.05}^{+71.96}$}
& {\small $4.82$}{\tiny ${}_{-3.98}^{+13.07}$}
& {\small $4.47$}{\tiny ${}_{-3.76}^{+18.67}$} & {\small $2.681$}{\tiny${}_{-0.277}^{+0.235}$}\\[0.8ex]
\hline \hline
\end{tabular}
{\tiny Notes.
a. $H_0$ is given in Km/s/Mpc units.
}
\label{table:1DResults3}
\end{table*}
\begin{figure*}
\begin{center}
\includegraphics[width=2in]{fig3a.eps}
\includegraphics[width=2in]{fig3b.eps}
\includegraphics[width=2in]{fig3c.eps}
\includegraphics[width=2in]{fig3d.eps}
{\small\caption{(color online) 1-dimensional marginalized posteriors for
parameter set $\mathcal{B}$. $SC$ stands for standard cosmography.}
\label{fig:1dim2ModelB}}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=2in]{fig4a.eps}
\includegraphics[width=2in]{fig4b.eps}
\includegraphics[width=2in]{fig4c.eps}
\includegraphics[width=2in]{fig4d.eps}
\includegraphics[width=2in]{fig4e.eps}
{\small\caption{(color online) 1-dimensional marginalized posteriors for
parameter set $\mathcal{C}$. $SC$ stands for standard cosmography.}
\label{fig:1dim2ModelC}}
\end{center}
\end{figure*}
As we can observe from Figs.~\ref{fig:1dimModelA},~\ref{fig:1dim2ModelB} and~\ref{fig:1dim2ModelC},
the Pad\'e approximants give similar results to standard cosmography, with the advantage of the convergence properties discussed in the previous sections.
We note that in particular $P_{21}$, $P_{31}$ and $P_{23}$ draw better samples with narrower
dispersion. For this reason, we plot the contours for those approximants in Figs.~\ref{fig:Pade21},~\ref{fig:Pade31}
and~\ref{fig:Pade32}. It is remarkable that the same degeneracies
among the parameters are found in all cases, even in other cases which have not been investigated here, see, e.g.~\cite{mine2}.
\begin{figure*}
\begin{center}
\includegraphics[width=3.4in]{fig5.ps}
{\small\caption{(color online) Marginalized posterior constraints for
parameter set $\mathcal{A}$ using $P_{21}$.}
\label{fig:Pade21}}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=3.4in]{fig6.ps}
{\small\caption{(color online) Marginalized posterior constraints for
parameter set $\mathcal{B}$ using $P_{31}$.}
\label{fig:Pade31}}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=3.6in]{fig7.ps}
{\small\caption{(color online) Marginalized posterior constraints for
parameter set $\mathcal{C}$ using $P_{23}$.}
\label{fig:Pade32}}
\end{center}
\end{figure*}
\section{Applications of Pad\'e's approach}\label{applicationsPade}
In flat FRW metric~(\ref{frw}), the
Friedmann equation for the energy density $\rho\equiv\sum_i\rho_i$ reads $3H^2 = 8\pi G \sum_i \rho_i$, where
the sum is over all cosmic species contributing to the whole energy budget. Afterwards, recovering Bianchi identities, one gets the continuity equation in the form
$\dot{\rho_i} = -3H(1+w_i)\rho_i$, in the absence of energy transfer
among the different components. In order to determine a specific model, one must specify the
cosmic fluids and their equations of state~\cite{galaxyluongo}. Below we test some models by means of our cosmographic results, inferred from the Pad\'e formalism.
We deal with implicit propagation of errors, since it is convenient to work with expected values and variances of
the cosmographic parameters, instead of their probability distribution functions. Thus, for example
\begin{equation}
\langle q_0 \rangle = \int q_0 p(q_0) dq_0\,,
\end{equation}
where $p(q_0) = f(q_0)/ \int f(q_0)dq$ and $f(q_0)$ is the non-normalized posterior distribution
found in section~\ref{ParameterEstimations} by the MCMC analysis.
The variance is
\begin{equation}
\sigma^2_{q_0} = \langle q_0^2 \rangle - \langle q_0 \rangle^2\,,
\end{equation}
and similar equations hold for the other cosmological parameters.
For the Pad\'e approximants $P_{21}$, $P_{31}$ and $P_{23}$ we obtain
\begin{itemize}
\item Pad\'e approximant $P_{21}$:
\begin{eqnarray}
\langle q_0 \rangle&=& -0.4623 \pm 0.0677\,, \nonumber \\
\langle j_0 \rangle &=& 0.5834 \pm 0.1215\,, \label{P21expectedv}
\end{eqnarray}
\item Pad\'e approximant $P_{31}$:
\begin{eqnarray}
\langle q_0 \rangle&=& -0.6040 \pm 0.1051\,, \nonumber \\
\langle j_0 \rangle &=& 1.1597 \pm 0.3690\,, \label{P31expectedv} \\
\langle s_0 \rangle &=& 0.2858 \pm 0.4866\,. \nonumber
\end{eqnarray}
\item Pad\'e approximant $P_{23}$:
\begin{eqnarray}
\langle q_0 \rangle&=& -0.7511 \pm 0.1737\,, \nonumber \\
\langle j_0 \rangle &=& 2.1968 \pm 1.1828\,, \nonumber \\
\langle s_0 \rangle &=& 3.2038 \pm 4.0459\,, \label{P32expectedv} \\
\langle l_0 \rangle &=& 15.9014 \pm 16.1370\,, \nonumber
\end{eqnarray}
\end{itemize}
where the reported error values are the standard deviations of the probability distributions, $\sigma = \sqrt{\sigma^2}$.
Using these results we can approximate the probability distributions as Gaussians centered around their mean values
and with variance $\sigma^2$. Now we are ready to investigate the implications of the results obtained using Pad\'e on some relevant cosmological models.
\subsection{The case of the $\Lambda$CDM model}
Concerning the flat $\Lambda$CDM model, the parameters to estimate are only $H_0$ and $\Omega_m$.
It is easy to demonstrate that, while $H_0$ is actually one of the CS, the matter density can be related to $q_0$ as $\Omega_m(q_0) = 2 (q_0 +1) /3$.
We have found by the MCMC algorithm the distribution functions $f(q_0)$ for $q_0$, obtained using the results for the Pad\'e approximant $P_{21}$.
The expected value for $\Omega_m$ is given by
\begin{equation}
\langle \Omega_m \rangle = \int \Omega_m(q_0) p(q_0) dq_0\,,
\end{equation}
and its variance is $ \sigma^2 = \langle \Omega_m^2 \rangle - \langle \Omega_m \rangle^2$, obtaining
\begin{equation} \label{LCDMOm0}
\Omega_m = 0.36\pm 0.05\,.
\end{equation}
Using the results for the Pad\'e approximant $P_{31}$ we obtain, from $\Omega_m(q_0) = 2 (q_0 +1) /3$, see Eqs.~(\ref{LCDMCS}),
\begin{equation}
\Omega_m(q_0) = 0.26\pm 0.07\,. \label{LCDMOm1}
\end{equation}
This procedure leads to a projection from a 4-parameter model (the Pad\'e $P_{31}$) to a 2-parameter model (late time flat-$\Lambda$CDM model), providing a broadening on the estimated parameters.
In analogy to the case in which $\Omega_m=\Omega(q_0)$, it is easy to show that $\Omega_m(s_0) = 2(1-s_0)/9$. Keeping in mind this expression, we obtain:
\begin{equation}
\Omega_m(s_0) = 0.16\pm 0.11\,. \label{LCDMOm2}
\end{equation}
The combination of the two results~(\ref{LCDMOm1}) and~(\ref{LCDMOm2}) should give tighter constraints.
If the probability distribution functions of $s_0$ and $q_0$ are
independent, the distribution function of $\Omega_m$ is simply the
product of the two distributions $\Omega_m(q_0)$ and $\Omega_m(s_0)$.
If we further assume Gaussian distributions, all the statistical information
is given by Eqs.~(\ref{LCDMOm1}) and~(\ref{LCDMOm2}), obtaining a rough estimate:
\begin{equation}\label{mah}
\Omega_m \simeq 0.23 \pm 0.06\,.
\end{equation}
In this way, we did a 3-parameter to 2-parameter projection. We cannot do anything better than Eq.~(\ref{mah}) for the flat-$\Lambda$CDM model due to the fact that $j_0$ is fixed to $j_0=1$.
As an example, to go beyond the case $j_0=1$, one can consider a generic additional cosmic component $X$, relevant at late times.
To do so, its equation of state parameter should lie in the interval
$-1 < w_X < 0$; but to avoid large degeneracies with a cosmological constant or with dust fluids we cannot be very close to $-1$ or to $0$.
A possible example is offered by the scalar curvature $\Omega_k$, that we neglected in all our previous numerical outcomes.
In such case, one can choose the equation of state $P_X = -\rho_X/3$, thus the corresponding Hubble rate takes the simple form
\begin{equation}
H(z) = H_0 \sqrt{\Omega_{\Lambda} + \Omega_X(1+z)^2 + \Omega_m(1+z)^3},
\end{equation}
with $\Omega_\Lambda = 1 - \Omega_{X} - \Omega_m$. The process of measurement indeed differs,
since $d_L$ has a different expressions for flat and non-flat cases. In general, the luminosity distance
equation is
\begin{eqnarray}
d_L &=& \frac{(1+z)}{\sqrt{\Omega_{k}}} \sinh (\sqrt{\Omega_{k}} \chi(z)) \nonumber \\
&\approx& (1+z) \left( \chi(z) + \frac{1}{3!}\Omega_{k} \chi(z)^3 \right)\,, \label{lumdistNonFlat}
\end{eqnarray}
with $\chi(z)$ the comoving distance to redshift $z$ given by Eq.~(\ref{kjjgjhg}).
For sufficiently small $\Omega_{k}$, the second equality in equation~(\ref{lumdistNonFlat}) is a good approximation. For illustration purposes,
we can consider $d_L = \chi(z)$ and assume that
the estimated values for the parameters are good for small $\Omega_{k}$ and definitively identify $\Omega_X$ with curvature.
The cosmographic parameters up to $s_0$ are in this case
\begin{eqnarray}
q_0 &=& -1 + \Omega_X + \frac{3}{2}\Omega_m\,, \nonumber \\
j_0 &=& 1-\Omega_X \,, \\
s_0 &=& (1-\Omega_X)^2 - \frac{3}{2}(3 - \Omega_X)\Omega_m\,. \nonumber
\end{eqnarray}
From the second equation and the results for the Pad\'e approximant $P_{31}$ (Eq.~(\ref{P31expectedv})), we have $\Omega_X(j_0) = -0.16 \pm 0.37$.
In the case of $\Omega_m$, using the first and second equations, $\Omega_m(q_0) = 0.37 \pm 0.31$,
and using the second and third equations, $\Omega_m(s_0) = 0.29 \pm 0.32$. Joining these results we obtain
\begin{eqnarray}
\Omega_m &=& 0.32 \pm 0.22\,, \\
\Omega_X &=& -0.16 \pm 0.37\,.
\end{eqnarray}
\subsection{The case of the Chevallier-Polarski-Linder parametrization}
The Chevallier-Polarski-Linder (CPL)~\cite{cpl} dark energy parametrization assumes that the universe is composed by baryons,
cold dark matter and dark energy with an evolving equation of state of the form:
\begin{equation}
w_{de} = w_0 + w_a\frac{z}{1+z}\,.
\end{equation}
The background cosmology cannot distinguish between dark matter and baryons, thus we write
\begin{eqnarray}
H(z) & = & H_0 \Bigg(\Omega_m (1+z)^3 + \Omega_{de} (1+z)^{3(1+w_0+w_a)} \times\nonumber\\
& & \exp \left(-\frac{3w_a z}{1+z}\right) \Bigg)^{1/2}\,, \label{CPLH}
\end{eqnarray}
where $\Omega_m = \Omega_b + \Omega_c$ and $\Omega_{de} = 1-\Omega_m$. Using~(\ref{CPLH}) and Eqs.~(\ref{Hpunto}) we obtain
\begin{equation}
q_0 = \frac{1}{2} \left(1 + 3 w_0 \left(1 - \Omega_m \right)\right)\,,
\end{equation}
\begin{equation}
j_0 = \frac{3}{2} \big( (3 w_0 \left(w_0+1\right) + w_a )(1 -\Omega_m ) \big) +1\,,
\end{equation}
\begin{eqnarray}
s_0 &=& \frac{1}{4} \Big[ 9 w_0 (1-\Omega_m) (w_a
(\Omega_m-7)-9) \nonumber\\ &-& 33 w_a (1 - \Omega_m) - 27 w_0^3 (1-\Omega_m) (3-\Omega_m)
\nonumber\\ & +& 9 w_0^2 (1 - \Omega_m) (3 \Omega_m-16) - 14 \Big]\,.
\end{eqnarray}
Thus, if we use the $P_{31}$ results, we have to estimate 3 parameters out of other 3 parameters.
This is done numerically by propagating errors in Eqs.~(\ref{P31expectedv}) obtaining
\begin{eqnarray}
\Omega_m &=& 0.26 \pm 0.19\,, \nonumber\\
w_0 &=& -1.04 \pm 0.16\,, \\
w_a &=& 0.08 \pm 0.28\,. \nonumber
\end{eqnarray}
\subsection{The case of unified dark energy}
One relevant approach to dark energy suggests that the universe is composed by a single fluid,
which unifies dark matter and dark energy in a single description.
A barotropic perfect fluid with vanishing adiabatic sound speed reproduces the $\Lambda$CDM behavior at the background,
as proposed in~\cite{Luongo:2011yk} and it is compatible with small perturbations, as shown in~\cite{Aviles:2011ak}. The corresponding equation of state reads
\begin{equation}
w_{df} = -\frac{1}{1+A (1+z)^3}\,,
\end{equation}
while the total equation of state for the $\Lambda + \text{dm}$ total dark fluid in the $\Lambda$CDM model reads
\begin{equation}
w_{\Lambda + \text{dm}} =\frac{ \sum_{i} \rho_i w_i}{\sum_{i} \rho_i} = -\frac{1}{1+\frac{\Omega_{\text{dm}}}{\Omega_{\Lambda}}(1+z)^3}\,. \label{TEOSLCDM}
\end{equation}
Thus, both models, i.e. $\Lambda$CDM and the negligible sound speed model,
exactly behave in the same way and hence they are degenerate. There are several other options for a unified dark fluid which does not
degenerate with $\Lambda$CDM. One of these frameworks is represented by the Chaplygin
gas~\cite{Kamenshchik:2001cp,Bilic:2001cg} and its generalizations~\cite{Bento:2002ps,Aviles:2011ak,Aviles:2011sd}
and constant adiabatic speed of sound models~\cite{Xu:2011bp}, among others. Therefore, we parameterize the dark fluid equation of state by a Taylor series:
\begin{equation}
w_{df}(z) = w_0 + w_1 z + w_2 z^2 + w_3 z^3 +\mathcal{O}(4)\,,
\end{equation}
where
\begin{equation}
w_i = \frac{1}{i!}\frac{d^i w(z)}{dz^i}\Big|_{z=0}\,.
\end{equation}
Knowing the value of $\Omega_b$, in order to estimate up to $w_i$, we need to use the first $i$th cosmographic parameters.
If we use Pad\'e approximants up to $P_{23}$, then we truncate the expansion series
at third order. The Hubble rate easily reads
\begin{equation}
H(z) = H_0 \sqrt{\Omega_b(1+z)^3+ \Omega_{df}F(z)}\,\,,
\end{equation}
where $\Omega_{df} = 1-\Omega_{b}$ and we define
\begin{eqnarray}
F(z) &=& (1+z)^{3(1+w_0-w_1+w_2-w_3)} \times \nonumber\\
\!\!\!\!\!\ \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! & & \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! \!\!\!\!\! \!\!\!\!\!
\exp \left[ 3(w_1-w_2+w_3)z + \frac{3}{2}(w_2-w_3) z^2 + w_3 z^3\right]. \label{fofZ}
\end{eqnarray}
The parameters to estimate are given implicitly by
\begin{eqnarray}
q_0 &=& \frac{1}{2} + \frac{2}{3}w_0(1-\Omega_b)\,, \label{DFq0} \\
j_0 &=& 1 + \frac{9}{2} w_0(1+w_0) + \frac{3}{2}w_1 \nonumber \\
& & - \frac{3}{2}\Omega_b \big( 3w_0(1+w_0) + w_1 \big)\,, \label{DFj0} \\
s_0 &=& -\frac{7}{2} - \frac{9}{4}w_0(1-\Omega_b)(9+w_1(7-\Omega_b)) \nonumber\\
& & -\frac{9}{4} w_0^2(1-\Omega_b)(16-3\Omega_b) \nonumber\\
& & -\frac{27}{4} w_0^3(3-\Omega_b)(1-\Omega_b) \nonumber\\
& &- \frac{45}{2}w_1(1-\Omega_b) - 3w_2(1-\Omega_b)\,, \label{DFs0}
\end{eqnarray}
which reduce to the flat-$\Lambda$CDM values when $w_0 = -1$, $w_1=w_2=0$, and by considering $\Omega_b \rightarrow \Omega_m$.
From several independent observations we have measurements of the baryon species in the universe. In this section we will take the best-fit
from the Planck Collaboration $\Omega_b = 0.0488$~\cite{Ade:2013zuv}. We report the estimated values from Pad\'e approximants $P_{21}$,
$P_{31}$ and $P_{23}$:
\begin{enumerate}
\item Pad\'e approximant $P_{21}$:
\begin{eqnarray}
w_0 &=& -0.67 \pm 0.05\,, \nonumber \\
w_1 &=& 0.37 \pm 0.13\,.
\end{eqnarray}
\item Pad\'e approximant $P_{31}$:
\begin{eqnarray}
w_0 &=& -0.77 \pm 0.07\,, \nonumber \\
w_1 &=& 0.63 \pm 0.37\,, \\
w_2 &=& 0.06 \pm 0.50\,. \nonumber
\end{eqnarray}
\item Pad\'e approximant $P_{23}$:
\begin{eqnarray}
w_0 &=& -0.87 \pm 0.12\,, \nonumber \\
w_1 &=& 1.13 \pm 1.09\,, \nonumber\\
w_2 &=& 0.23 \pm 2.71\,, \\
w_3 &=& -0.95 \pm 2.21\,. \nonumber
\end{eqnarray}
\end{enumerate}
These results should be compared with the best fit values for the $\Lambda$CDM model, $w_0 = -0.76$, $w_1 = 0.55$, $w_2 = 0.15$ and
$w=-0.32$, obtained by substituting in equation~(\ref{TEOSLCDM}) the values $\Omega_m = 0.2880$, $\Omega_{\Lambda} = 0.7119$,
estimated in Sec.~\ref{ParameterEstimations} for the $\Lambda$CDM model, and the value $\Omega_b = 0.0488$ from Planck.
\section{The universe equation of state}\label{universeEoS}
Now, let us consider an arbitrary collection of fluids (baryons, cold dark matter, dark energy, ...) with total energy density
$\rho = \sum_i \rho_i$ which comprises all possible species present in the universe. The Friedmann equation is thus $3H^2 = 8\pi G \rho$, as already mentioned.
We want to estimate the total equation of state of the universe given by
\begin{equation}
w_{T}(z) = w_0 + w_1 z + w_2 z^2 + w_3 z^3 + \mathcal{O}(4)\,. \label{EOSUwT}
\end{equation}
The Friedmann equation can be recast as
\begin{equation}
H(z) = H_0 \sqrt{F(z)}\,,
\end{equation}
where $F(z)$ is given again by Eq.~(\ref{fofZ}). The cosmographic parameters are equal to Eqs~(\ref{DFq0})-(\ref{DFs0}), by imposing $\Omega_b=0$.
At late times, the total equation of state of the universe is given by
\begin{equation}
w_{T} = -\frac{1}{1+\frac{\Omega_m}{\Omega_{\Lambda}}(1+z)^3}\,. \label{TEOSLCDM2}
\end{equation}
We report the estimated values from Pad\'e approximants $P_{21}$,
$P_{31}$ and $P_{23}$:
\begin{enumerate}
\item Pad\'e approximant $P_{21}$:
\begin{eqnarray}
w_0 &=& -0.64 \pm 0.05\,, \nonumber \\
w_1 &=& 0.41 \pm 0.12\,.
\end{eqnarray}
\item Pad\'e approximant $P_{31}$:
\begin{eqnarray}
w_0 &=& -0.73 \pm 0.07\,, \nonumber \\
w_1 &=& 0.67 \pm 0.34\,, \\
w_2 &=& -0.02 \pm 0.50\,. \nonumber
\end{eqnarray}
\item Pad\'e approximant $P_{23}$:
\begin{eqnarray}
w_0 &=& -0.83 \pm 0.12\,, \nonumber \\
w_1 &=& 1.17 \pm 1.02\,, \nonumber\\
w_2 &=& 0.10 \pm 2.51\,, \\
w_3 &=& -0.47 \pm 1.93\,. \nonumber
\end{eqnarray}
\end{enumerate}
These results should be compared with the best fit values for the $\Lambda$CDM model, $w_0 = -0.71$, $w_1 = 0.62$, $w_2 = 0.08$ and
$w=-0.39$, obtained by substituting in equation~(\ref{TEOSLCDM2}) the values $\Omega_m = 0.2880$, $\Omega_{\Lambda} = 0.7119$,
estimated in Sec.~\ref{ParameterEstimations} for the $\Lambda$CDM model.
\begin{table*}
\caption{{\small Table of best fits and their likelihoods (1$\sigma$) for the redshifts functions $y_1$ and $y_4$, using the parameter sets
$\mathcal{B}$ and $\mathcal{C}$.}}
\begin{tabular}{c|c|c|c|c}
\hline\hline
{\small $\quad$ Parameter $\quad$} & {\small $\qquad$ $y_1 = z/(1+z)$ $\qquad$ } & {\small $\qquad$ $y_1 = z/(1+z)$ $\qquad$ }
& {\small $\qquad$ $y_4 = \tan z$ $\qquad$} & {\small $\quad$ $y_4 = \tan z$ $\quad$}\\
{\small $\quad$ $\quad$} & {\small set $\mathcal{B}$ } & {\small set $\mathcal{C}$ }
& {\small set $\mathcal{B}$ } & {\small set $\mathcal{C}$}\\
\hline
{\small$H_0$} & {\small $75.11$}{\tiny ${}_{-3.44}^{+3.29}$} & {\small $73.17$}{\tiny ${}_{-3.38}^{+3.92}$}
& {\small $72.34$}{\tiny ${}_{-3.97}^{+3.55}$} & {\small $72.58$}{\tiny ${}_{-4.31}^{+3.94}$}\\[0.8ex]
{\small$q_0$} & {\small $-1.0642$}{\tiny ${}_{-0.1958}^{+0.2216}$} & {\small $-0.8517$}{\tiny ${}_{-0.3695}^{+0.3795}$}
& {\small $-0.868$}{\tiny ${}_{-0.2763}^{+0.3165}$} & {\small $-0.7501$}{\tiny ${}_{-0.3839}^{+0.3891}$}\\[0.8ex]
{\small$j_0$} & {\small $2.991$}{\tiny ${}_{-1.109}^{+1.030}$} & {\small $1.983$}{\tiny ${}_{-2.772}^{+2.646}$}
& {\small $2.142$}{\tiny ${}_{-1.448}^{+1.411}$} & {\small $1.520$}{\tiny ${}_{-1.736}^{+2.123}$}\\[0.8ex]
{\small$s_0$} & {\small $4.919$}{\tiny ${}_{-3.198}^{+3.909}$} & {\small $1.591$}{\tiny ${}_{-6.469}^{+10.905}$}
& {\small $5.149$}{\tiny ${}_{-1.338}^{+2.210}$} & {\small $-0.206$}{\tiny ${}_{-4.256}^{+4.960}$}\\[0.8ex]
{\small$l_0$} & -- -- & {\small $7.96$}{\tiny ${}_{-4.79}^{+47.83}$}
& -- -- & {\small $-18.64$}{\tiny ${}_{-12.72}^{+21.60}$}\\[0.8ex]
\hline \hline
\end{tabular}
{\tiny Notes.
a. $H_0$ is given in Km/s/Mpc units.
}
\label{table:1DOtherRedshifts}
\end{table*}
\section{Consequences of the Pad\'e results for dark energy}\label{consequencespade}
We showed that the use of Pad\'e approximants in cosmography provides a new
model-independent technique for reconstructing the luminosity distance and the Hubble parameter $H(z)$.
This method is particulary valid since standard constructions in cosmography require to develop the luminosity distance $d_{L}$
as a Taylor series and then to match the data with this approximation.
In particular, when data are taken over $z>1$, Pad\'e functions work better than truncated Taylor series.
To make the argument consistent, we have performed in Sec.~\ref{Sect:DataSet} a detailed analysis of our models derived from Pad\'e approximants with respect to the
data taken from different observations.
The results have been elaborated in Secs.~\ref{Sect:DataSet} and~\ref{ParameterEstimations} and compared with
the standard cosmographic approach and to the values inferred from assuming the $\Lambda$CDM model.
As expected, not all the Pad\'e approximants work properly. For example, we have commented that one has to take special care of the possible spurious
divergences that may appear in $d_{L}$ when approximating with Pad\'e, due to the fact that such functions are rational functions.
Moreover, not all Pad\'e models can fit the data in the appropriate
way. Indeed, we have seen both theoretically and numerically that approximants
whose degrees of the numerator and of the denominator are similar seem to be preferred
(see Figs.~\ref{LCDMDL}-\ref{fig:PA} and Tabs.~\ref{table:1DResults1}-\ref{table:1DResults3}).
This fact suggests that the increase of the luminosity distance with $z$ has to be indeed slower than the one depicted by a Taylor approximation.
Interestingly, our numerical analysis has singled out the Pad\'e functions $P_{21}$, $P_{31}$ and $P_{23}$,
which are the ones that draw the best samples,
with narrowest dispersion (see Figs.~\ref{fig:1dim2ModelB}-\ref{fig:Pade32}).
As one can see from Tabs.~\ref{table:1DResults1}-\ref{table:1DResults3}, the best fit values and errors for the CS parameters
estimated using the approximants $P_{21}$, $P_{31}$ and $P_{23}$ are in good agreement with the SC results.
In particular, the approximant $P_{23}$ gives smaller relative errors than the corresponding SC analysis, thus suggesting that enlarging the approximation order, the
analysis by means of Pad\'e are increasingly more appropriate than the standard one.
The estimated values of the CS parameters, through the use of the Pad\'e approximants $P_{21}$, $P_{31}$ and $P_{23}$
seem to indicate that the value of $H_{0}$ is smaller than the one derived by means of the standard (Taylor) approximation. Our results therefore
agree with Planck results, which show smaller values of $H_0$ than previous estimations. On the contrary,
$q_{0}$ seems to be larger than the result obtained by standard cosmography, while for $j_{0}$ the situation is less clear ($P_{21}$ and $P_{23}$ indicate
a smaller value, while $P_{31}$ a larger one). In any cases, the sign of $j_0$ is positive at a $68\%$ of confidence level. This fact, according to Sec. II,
provides a universe which starts decelerating at a particular redshift $z_{tr}$, named the transition redshift.
\noindent From the above considerations, a comparison of our results with the ones obtained previously using Pad\'e expansions is essential.
In particular, in \cite{OrlChri} the authors employed a $P_{12}$ Pad\'e approximant, motivating their choice by noticing that for $z\ll1$ the requirement $m>n$ could be
appropriate to describe the behavior of $d_L$.
Their idea was to propose this Pad\'e prototype and to use it for higher redshift domains. Their heuristic guess has not been compared in that work with respect to other $P_{nm}$ approximants.
Hence, the need of extending their approach has been achieved in the present paper, where we analyzed thoroughly which extensions work better.
Moreover, the authors adopted the $P_{12}$ Pad\'e approximant as a first example to describe the convergence radius in terms of the Pad\'e formalism,
providing discrepancies with respect to standard Taylor treatments.
Their numerical analyses were essentially based on SNeIa data only, while in our paper we adopted different data sets, i.e. baryonic acoustic oscillation,
Hubble space telescope measurements and differential age data, with improved numerical accuracies developed by using the CosmoMC code \cite{Lewis:2002ah,cosmomc_notes}.
As a consequence, we found that the cosmographic results obtained using $P_{21}$ are significantly different from the ones obtained using $P_{12}$.
Indeed, in \cite{OrlChri} the authors employed the $P_{12}$ approximant only, whereas in our paper we reported in Fig. 4 the plots of $P_{21}$,
which definitively provide the differences between $P_{21}$ and $P_{12}$.
In general, our results seem to be more accurate and general than the numerical outcomes of \cite{OrlChri}.
However, we showed a positive jerk parameter, for sets $\mathcal A$ and $\mathcal B$, which is compatible with their results, albeit not strictly constrained to $j_0>1$, as they proposed.
Numerical outcomes for $H_0$ and $q_0$ lie in similar intervals with respect to \cite{OrlChri}.
Summing up, although the use of $P_{12}$ is possible \emph{a priori}, we demonstrated here that considering different models one can find
parameterizations that work better than $P_{12}$ and therefore are more natural candidates for further uses in upcoming works on cosmography.
Further, it is of special interest to look at the comparison of the numerical results obtained for the cases of $\Lambda$CDM, CPL and unified dark energy models by inserting
the values estimated by fitting the Pad\'e functions $P_{21}$, $P_{31}$ and $P_{23}$ (see Sec.~\ref{applicationsPade}).
From this analysis it turns out that all of them suggest small departures
from $\Lambda$CDM, as expected. Moreover, $P_{31}$ is the one that better reproduces the results of $\Lambda$CDM. However, we expect from Fig.~\ref{LCDMDL}
that $P_{22}$ and $P_{32}$ should match even better the $\Lambda$CDM predictions. Therefore, we consider that it is needed to repeat the analysis with a larger set of data in the region $z\gg1$
to get more reliable results in this sense. This indication will be object of extensive future works.
Finally, let us comment the fact that results of Tab.~\ref{table:1DOtherRedshifts}, compared with the ones in Tabs.~\ref{table:1DResults1}-\ref{table:1DResults3}, show that
the approximants $P_{21}$, $P_{31}$ and $P_{23}$ give values for the CS parameters that are much closer to the ones estimated by standard cosmography
and by the $\Lambda$CDM model, than the results provided by the introduction of auxiliary variables in the standard cosmographic approach.
This definitively candidates Pad\'e approximants to represent a significative alternative to overcome the issues of divergence in cosmography, without the need of any additional auxiliary parametrization.
\section{Final outlooks and perspectives}\label{conclusions}
In this work we
proposed the use of Pad\'e approximations in the context of observational cosmology.
In particular, we improved the standard cosmographic approach,
{which enables to} accurately determine refined cosmographic bounds {on the dynamical parameters of the models}.
We stressed the fact that the Pad\'e recipe
{can be used as a} relevant tool to extend standard Taylor treatments {of the luminosity distance.}
Our main goal was to introduce a class of Pad\'e approximants able to overcome all the problems plaguing modern cosmography.
To do so, we enumerated the basic properties and the most important demands of the Pad\'e treatment and we matched theoretical predictions with modern data.
In particular, the {main advantage} of the rational cosmographic method {is that} Pad\'e functions
{reduce the issue of convergence of the standard cosmographic approach based on} truncated Taylor series, {especially for data taken over
a larger redshift range.}
In other words, usual model independent expansions performed at $z=0$ suffer from divergences due to data spanning cosmic intervals with $z>1$.
Since Pad\'e {approximants are rational functions, thence they can} naturally overcome this issue. In particular, in our numerical treatment, we have considered all the possible Pad\'e approximants of the luminosity distance whose order
of the numerator and denominator sum up to three, four and five and compared them with the corresponding Taylor polynomials
of degree three, four and five in $z$. Among these models, it turned out that the Pad\'e technique can give results similar to those obtained by standard cosmography
and also improve the accuracy. In addition, the Pad\'e technique overcomes the need of introducing auxiliary variables,
as proposed in standard cosmography to reduce divergences at higher redshifts.
To do so, we compared Pad\'e results also with Taylor re-parameterized expansions. {In all the cases considered here,
our Pad\'e numerical outcomes appear to improve} the standard analyses.
Furthermore, we also considered to overcome the degeneracy problem by {employing} additional data sets. In particular, we assumed union 2.1 type Ia supernovae, baryonic acoustic oscillation, Hubble space telescope measurements and direct observations of Hubble rates,
based on the differential age method. Moreover, all cosmographic drawbacks have also been investigated and treated in terms of Pad\'e's recipe,
proposing for each problem a possible solution to improve the experimental analyses. Afterwards, we guaranteed our numerical outcomes to {lie} in viable intervals and we demonstrated that the refined cosmographic bounds almost confirm the standard cosmological paradigm,
{thus} forecasting the sign of the variation of acceleration, i.e. the jerk parameter. However, although the $\Lambda$CDM model {passes our experimental tests, we cannot conclude that} evolving dark energy terms are ruled out.
Indeed, we compared our Pad\'e results with a class of cosmological models, namely
the $\omega$CDM model, the Chevallier-Polarski-Linder parametrization and the unified dark energy models, finding a good agreement with those paradigms. Furthermore, we also investigated the consequences of Pad\'e's bounds on the universe equation of state. {To conclude, we have proposed and investigated here the use of Pad\'e approximants in} the field of precision cosmology,
with particular regards to cosmography. Future perspectives will be clearly devoted to describe the Pad\'e approach in other relevant fields.
For example, early time cosmology is expected to be {more easily} described in our framework, as well as additional epochs {related to} high redshift data. {Collecting all these results one could in principle definitively} reconstruct the universe expansion history, matching late with early time observations and also permitting to understand
whether the dark energy fluid evolves or remains a pure cosmological constant at all times.
\section*{Acknowledgements}
SC and OL are grateful to Manuel Scinta for important discussions on numerical and theoretical results. AB and OL want to thank Hernando Quevedo
and Christine Gruber for their support during the phases of this work.
AA is thankful to Jaime Klapp for discussions on the numerical outcomes of this work. SC is supported
by INFN through iniziative specifiche NA12, OG51. OL is supported by the European PONa3 00038F1 KM3NET
(INFN) Project. AA is supported by the project CONACyT-EDOMEX-2011-C01-165873. AB {wants to thank the} A. Della Riccia Foundation (Florence, Italy) {for support.}
\begin{widetext}
|
\section{Introduction}
The goal of an agnostic learning algorithm for a concept class $\C$ is to produce, for any distribution on examples, a hypothesis $h$ whose error on a random example from the distribution is close to the best possible by a concept from $C$. This model reflects a common empirical approach to learning, where few or no assumptions are made on the process that generates the examples and a limited space of candidate hypothesis functions is searched in an attempt to find the best approximation to the given data.
Agnostic learning of disjunctions (or, equivalently, conjunctions) is a fundamental question in learning theory and a key step in learning algorithms for other concept classes such as DNF formulas and decision trees. Algorithms for this problem, such as the Set Covering Machine \citep{MarchandS02}, are also used in practical applications. There is no known efficient algorithm for the problem, in fact the fastest algorithm that does not make any distributional assumptions runs in $2^{\tilde{O}(\sqrt{n})}$ time \citep{KKMS05}. Polynomial-time learnability is only known when the examples are very close to being consistent with some disjunction \citep{AwasthiBS10}.
While the problem appears to be hard, strong hardness results are known only if the hypothesis is restricted to be a disjunction or a linear threshold function \citep{BenDavidEL:03,BshoutyBurroughs:06,FeldmanGKP09,FeldmanGRW:12}, or for learning using $\ell_1$-regression \citep{KS10}. Weaker, quasi-polynomial lower bounds are known assuming hardness of learning sparse parities with noise (see Section~\ref{sec:agnostic-learn}) and, very recently, hardness of refuting random SAT formulas \citep{DanielyS14}. It is also well-known that distribution-independent agnostic learning of disjunctions implies PAC learning of DNF expressions \citep{KearnsSS:94} (similar results for distribution specific-learning are discussed below). Finally, agnostic learning of disjunctions is known to be closely related to the problem of differentially-private release of answers to conjunctive queries \citep{GHRU11}.
We consider this problem with an additional assumption that example points are distributed according to a symmetric or a product distribution. Symmetric and product distributions are two incomparable classes of distributions that generalize the well-studied uniform distribution. Theoretical study of learning over symmetric distributions was first done by \citet{Wim10} who gave $n^{O(1/\epsilon^4)}$ time agnostic learning algorithm for the class of halfspaces. Agnostic learning of disjunctions over symmetric distributions on $\zo^n$ also arises naturally in the well-studied problem of privately releasing answers to all short conjunction queries with low average error \citep{FeldmanKothari:14}.
\subsection{Our Results}
\label{sec:our-results}
We prove that disjunctions (and conjunctions) are learnable agnostically over any symmetric distribution in time $n^{O(\log(1/\eps))}$. This matches the well-known upper bound for the uniform distribution. Our proof is based on $\ell_1$-approximation of any disjunction by a linear combination of functions from a fixed set of functions. Such approximation directly gives an agnostic learning algorithm via $\ell_1$-regression based approach introduced by \citet{KKMS05}.
A natural and commonly used set of basis functions is the set of all monomials on $\zo^n$ of some bounded degree. It is easy to see that on product distributions with constant bias, disjunctions longer than some constant multiple of $\log(1/\eps)$ are $\eps$-close to the constant function $1$. Therefore, polynomials of degree $O(\log(1/\eps))$ suffice for $\ell_1$ (or $\ell_2$) approximation on such distributions. This simple argument does not work for general product distributions. However it was shown by \citet{BOW08} that the same degree (up to a constant factor) still suffices in this case. Their argument is based on the analysis of noise sensitivity under product distributions and implies additional interesting results.
Interestingly, it turns out that low-degree polynomials cannot be used to obtain the same result for all symmetric distributions: we show that there exists a symmetric distribution for which disjunctions are no longer $\ell_1$-approximated by low-degree polynomials.
\begin{theorem}
\label{lem:hardsym-intro}
There exists a symmetric distribution $\D$ such that for $c = x_1 \vee x_2 \vee \cdots \vee x_n$, any polynomial $p$ that satisfies $\E_{x \sim \D}[ |c(x) -p(x)|] \leq 1/3$ is of degree $\Omega(\sqrt{n})$.
\end{theorem}
To prove this, we consider the standard linear program \citep[see][]{KS10} to find the coefficients of a degree $r$ polynomial that minimizes pointwise error with the disjunction $c$. The key idea is to observe that an optimal point for the dual can be used to obtain a distribution on which the \textit{$\ell_1$ error} of the best fitting polynomial $p$ for $c$ is same as the value of minimum \textit{pointwise error} of any degree $r$ polynomial with respect to $c$. When $c$ is a symmetric function, one can further observe that the distribution so obtained is in fact symmetric. Combined with the degree lower bound for uniform approximation by polynomials by \citet{KS10}, we obtain the result. The details of the proof appear in Section~\ref{sec:poly-lower}.
Our approximation for general symmetric distributions is based on a proof that for the special case of the uniform distribution on $S_r$ (the points from $\on^n$ with Hamming weight $r$), low-degree polynomials still work, namely, for any disjunction $c$, there is a polynomial $p$ of degree at most $O(\log{(1/\epsilon)})$ such that the $\ell_1$ error $\E_{x \sim S_r}[ |c(x)-p(x)|] \leq \epsilon$.
\begin{theorem}
For $r \in \{0,\ldots,n\}$, let $S_r$ denote the set of points in $\zo^n$ that have exactly $r$ $1$'s and let $\D_r$ denote the uniform distribution on $S_r$. For every disjunction $c$ and $\eps > 0$, there exists a polynomial $p$ of degree at most $O(\log{(1/\epsilon)})$ such that $\E_{\D_r}[|c(x) - p(x)|] \leq \eps$.\label{monotone-sym}
\end{theorem}
This result can be easily converted to a basis for approximating disjunctions over arbitrary symmetric distributions. All we need is to partition the domain $\zo^n$ into layers as $\cup_{0 \leq r \leq n}S_r$ and
use a (different) polynomial for each layer. Formally, the basis now contains functions of the form $\mathrm{IND}(r) \cdot \chi$, where $\mathrm{IND}$ is the indicator function of being in layer of Hamming weight $r$ and $\chi$ is a monomial of degree $O(\log(1/\eps))$. We note that a related strategy, of constructing a collection of functions, one for each layer of the cube was used by \citet{Wim10} to give $n^{O(1/\epsilon^4)}$ time agnostic learning algorithm for the class of halfspaces on symmetric distributions. However, his proof technique is based on an involved use of representation theory of the symmetric group and is not related to ours.
Our proof technique also gives a simpler proof for the result of \citet{BOW08} that implies approximation of disjunction by low-degree polynomials on all product distributions.
\begin{theorem}\label{th:product-poly-intro}
For any disjunction $c$ and product distribution $\D$ on $\zo^n$, there is a polynomial $p$ of degree $O(\log{(1/\epsilon)})$ such that $\E_{x \sim \D}[|c(x)-p(x)|] \leq \epsilon.$ \label{th:product-poly}
\end{theorem}
\subsection{Applications}
Theorem \ref{monotone-sym} together with a standard application of $\ell_1$ regression \citep{KKMS05} yields an agnostic learning algorithm for the class of disjunctions running in time $n^{O(\log(1/\eps))}$.
\begin{corollary}
There is an algorithm that agnostically learns the class of disjunctions on arbitrary symmetric distributions on $\zo^n$ in time $n^{O( \log{(1/\epsilon)})}$. \label{agnostic-learning}
\end{corollary}
This learning algorithm was extended to the class of all coverage functions in \citep{FeldmanKothari:14}, and then applied to the well-studied problem of privately releasing answers to all short conjunction queries with low average error.
It was shown by \citet{KalaiKM:09} and \citet{Feldman:10ab} that agnostic learning of conjunctions over a distribution $D$ in time $T(n,1/\eps)$ implies learning of DNF formulas with $s$ terms over $D$ in time $\poly(n,1/\eps) \cdot T(n,(4s/\eps))$. Further, under the same conditions distribution-specific agnostic boosting \citep{KalaiKanade:09,Feldman:10ab} implies that there exists an agnostic learning algorithm for decision trees with $s$ leaves running in time $\poly(n,1/\eps) \cdot T(n,s/\eps)$. Therefore we obtain quasi-polynomial learning algorithms for DNF formulas and decision trees over symmetric distributions.
\begin{corollary}
\begin{enumerate}
\item DNF formulas with $s$ terms are PAC learnable with error $\eps$ in time $n^{O(\log(s/\eps))}$ over all symmetric distributions;
\item Decision trees with $s$ leaves are agnostically learnable with excess error $\eps$ in time $n^{O(\log(s/\eps))}$ over all symmetric distributions.
\end{enumerate}
\end{corollary}
We also observe that any algorithm that agnostically learns the class of disjunction on the uniform distribution in time $n^{o(\log{(\frac{1}{\epsilon})})}$ would yield a faster algorithm for the notoriously hard problem of Learning Sparse Parities with Noise. This is implicit in prior work \citep{KKMS05,Feldman:12jcss} and we provide additional details in Section \ref{sec:agnostic-learn}.
\citet{DachmanFTWW:15} recently showed that $\ell_1$ approximation by polynomials is necessary and sufficient condition for agnostic learning over a product distribution (at least in the statistical query framework of \citet{Kearns:98}). Our agnostic learning algorithm (Theorem \ref{agnostic-learning}) and lower bound for polynomial approximation (Theorem \ref{lem:hardsym-intro}) demonstrate that this equivalence does not hold for non-product distributions.
\section{Preliminaries}
\label{sec:prelims}
We use $\zo^n$ to denote the $n$-dimensional Boolean hypercube. Let $[n]$ denote the set $\{1,2,\ldots,n\}$. For $S\subseteq [n]$, we denote by $\ORR_S: \zo^n \rightarrow \zo$, the monotone Boolean disjunction on variables with indices in $S$, that is, for any $x \in \zo^n$, $\ORR_S(x) = 0 \Leftrightarrow \forall i \in S \ \ x_i = 0$.
One can define norms and errors with respect to any distribution $\D$ on $\zo^n$. Thus, for $f: \zo^n \rightarrow \R$, we write the $\ell_1$ and $\ell_2$ norms of $f$ as $\|f\|_1 = \E_{x \sim \D}[ |f(x)|] $ and $\|f\|_2 = \sqrt{\E[ f(x)^2]}$ respectively. The $\ell_1$ and $\ell_2$ error of $f$ with respect to $g$ are given by $\|f-g\|_1$ and $\|f-g\|_2$ respectively.
\subsection{Agnostic Learning}
The agnostic learning model is formally defined as follows \citep{Haussler:92,KearnsSS:94}.
\begin{definition}
\label{def:agnostic}
Let $\F$ be a class of Boolean functions and let $\D$ be any fixed distribution on $\zo^n$. For any distribution $\P$ over $\zo^n \times \zo$, let $\mbox{opt}(\P,\F)$ be defined as: $\mbox{opt}(\P,\F) = \inf_{f \in \F} \E_{(x,y) \sim \P} [|y - f(x)|] .$ An algorithm $\A$, is said to agnostically learn $\F$ on $\D$ if for every {\em excess error} $\epsilon> 0$ and any distribution $\P$ on $\zo^n \times \zo$ such that the marginal of $\P$ on $\zo^n$ is $\D$, given access to random independent examples drawn from $\P$, with probability at least $\frac{2}{3}$, $\A$ outputs a hypothesis $h:\zo^n \rightarrow [0,1]$, such that $\E_{(x,y) \sim \P} [ |h(x)- y| ] \leq \mbox{opt}(\P, \F) + \epsilon.$
\end{definition}
It is easy to see that given a set of $t$ examples $\{(x^i,y^i)\}_{i\leq t}$ and a set of $m$ functions $\phi_1, \phi_2, \ldots, \phi_m$ finding coefficients $\alpha_1,\ldots,\alpha_m$ which minimize $$\sum_{i\leq t} \left| \sum_{ j \leq m} \alpha_j \phi_j (x^i) - y^i \right|$$ can be formulated as a linear program. This LP is referred to as Least-Absolute-Error (LAE) LP or Least-Absolute-Deviation LP, or $\ell_1$ linear regression. As observed by \citet{KKMS05}, $\ell_1$ linear regression gives a general technique for agnostic learning of Boolean functions.
\begin{theorem}
\label{th:lae-lp}
Let $\C$ be a class of Boolean functions, $\D$ be distribution on $\zo^n$ and $\phi_1, \phi_2, \ldots, \phi_m: \zo^n \rightarrow \R$ be a set of functions that can be evaluated in time polynomial in $n$. Assume that there exists $\Delta$
such that for each $f \in \C$, there exist reals $\alpha_1, \alpha_2, \ldots, \alpha_m$ such that $${\E_{x \sim \D} \left[ \left|\sum_{ i \leq m} \alpha_i \phi_i (x) - f(x)\right|\right] \leq \Delta}.$$
Then there is an algorithm that for every $\eps >0$ and any distribution $\P$ on $\zo^n \times \zo$ such that the marginal of $\P$ on $\zo^n$ is $\D$, given access to random independent examples drawn from $\P$, with probability at least $2/3$, outputs a function $h$ such that $$\E_{(x,y) \sim \P} [ |h(x)- y| ] \leq \Delta + \epsilon.$$ The algorithm uses $O(m/\eps^2)$ examples, runs in time polynomial in $n$, $m$, $1/\eps$ and returns a linear combination of $\phi_i$'s.
\end{theorem}
The output of this LP is not necessarily a Boolean function but can be converted to a Boolean function with disagreement error of $\Delta + 2\eps$ using $``h(x) \geq \theta"$ function as a hypothesis for an appropriately chosen $\theta$ \citep{KKMS05}.
\newcommand{\q}{\mathbf{q}}
\section{$\ell_1$ Approximation on Symmetric Distributions} \label{sec:l1approx}
In this section, we show how to approximate the class of all disjunctions on any symmetric distribution by a linear combination of a small set of basis functions.
As discussed above, polynomials of degree $O(\log{(1/\epsilon)})$ can $\eps$-approximate any disjunction in $\ell_1$ distance on any product distribution. This is equivalent to using low-degree monomials as basis functions. We first show that this basis would not suffice for approximating disjunctions on symmetric distributions. Indeed, we construct a symmetric distribution on $\zo^n$, on which, any polynomial that approximates the monotone disjunction $c = x_1 \vee x_2 \vee \ldots \vee x_n$ within $\ell_1$ error of $1/3$ must be of degree $\Omega(\sqrt{n})$.
\subsection{Lower Bound on $\ell_1$ Approximation by Low-Degree Polynomials}
\label{sec:poly-lower}
In this section we give the proof of Theorem \ref{lem:hardsym-intro}.
\begin{proof}[Proof of Theorem \ref{lem:hardsym-intro}]
Let $d:[n] \rightarrow \zo$ be the predicate corresponding to the disjunction $x_1 \vee x_2 \vee \ldots \vee x_n$, that is, $d(0) = 0$ and $d(i) = 1$ for each $i > 0$.
Consider a natural linear program to find a univariate polynomial $f$ of degree at most $d$ such that $\|d-f\|_{\infty} = \max_{0 \leq i \leq n} |d(i) - f(i)| $ is minimized. This program (and its dual) often comes up in proving polynomial degree lower bounds for various function classes \citep[for example,][]{KS10}.
\begin{align*}
& \min \text{ } \epsilon & \text{ }\\
s.t. \text{ } &\epsilon \geq |d(m) - \sum_{i = 0} ^{r} \alpha_i \cdot m^i|& \forall\text{ } m \in \zeto{n}\\
&\alpha_i \in \R \text{ } &\forall \text{ } i\in \zeto{r}
\end{align*}
If $\{\alpha_0, \alpha_1, \ldots, \alpha_n\}$ is a solution for the program above that has value $\eps$ then $f(m) = \sum_{i = 0}^r \alpha_i m^i$ is a degree $r$ polynomial that approximates $d$ within an error of at most $\epsilon$ at every point in $\zeto{n}$. \citet{KS10} show that there exists an $r^* = \Theta(\sqrt{n})$, such that the optimal value of the program above for $r = r^*$ is $\epsilon^* \geq 1/3$. Standard manipulations \citep[see][]{KS10} can be used to produce the dual of the program.
\begin{align*}
& \max \text{ }\sum_{m=0}^n \beta_m \cdot d(m) &\\
s.t. \text{ } &\sum_{m=0}^n \beta_m \cdot m^i = 0 \text{ }& \forall \text{ } i\in \zeto{r}\\
&\sum_{m=0}^n |\beta_m| \leq 1&\\
& \beta_m \in \R \text{ }& \forall \text{ } m \in \zeto{n}\\
\end{align*}
Let $\beta^* = \{ \beta^*_m\}_{m \in \zeto{n}} $ denote an optimal solution for the dual program with $r = r^*$. Then, by strong duality, the value of the dual is also $\epsilon^*$. Observe that $\sum_{m=0}^n |\beta^*_m| = 1$, since otherwise we can scale up all the $\beta^*_m$ by the same factor and increase the value of the program while still satisfying the constraints.
Let $\rho:\zeto{n} \rightarrow [0,1]$ be defined by $\rho(m) = |\beta^*_m|$. Then $\rho$ can be viewed as a density function of a distribution on $\zeto{n}$ and we use it to define a symmetric distribution $\D$ on $\on^n$ as follows: $\D(x) = \rho(w(x))/{n \choose w(x)}$, where $w(x) = \sum_{i=1}^{n} x_i$ is the Hamming weight of point $x$. We now show that any polynomial $p$ of degree $r^*$ satisfies $\E_{x \sim \D}[ |c(x) - p(x)|] \geq 1/3$.
We now extract a univariate polynomial $f_p$ that approximates $d$ on the distribution with the density function $\rho$ using $p$. Let $p_{avg}:\on^n \rightarrow \R$ be obtained by averaging $p$ over every layer. That is, $p_{avg}(x) = \E_{z \sim \D_{w(x)}}[ p(z)]$, where $w(x)$ denotes the Hamming weight of $x$. It is easy to check that since $c$ is symmetric, $p_{avg}$ is at least as close to $c$ as $p$ in $\ell_1$ distance.
Further, $p_{avg}$ is a symmetric function computed by a multivariate polynomial of degree at most $r^*$ on $\zo^n$. Thus, the function $f_p(m)$ that gives the value of $p_{avg}$ on points of Hamming weight $m$ can be computed by a univariate polynomial of degree $r^*$. Further, $$\E_{x\sim \D}[|c(x)-p(x)|] \geq \E_{x\sim \D}[|c(x)-p_{avg}(x)|] = \E_{m \sim \rho}[ |d(m) - f_p(m)|].$$
Let us now estimate the error of $f_p$ w.r.t $d$ on the distribution $\rho$. Using the fact that $f_p$ is of degree at most $r^*$ and thus $\sum_{m = 0}^n f_p(m) \cdot \beta_m = 0$ (enforced by the dual constraints), we have:
\begin{align*}
\E_{m \sim \rho}[ |d(m) - f_p(m)|] &\geq \E_{m \sim \rho}[ (d(m) - f_p(m)) \cdot \sign ( \beta^*_m)]\\
&= \sum_{m = 0}^n d(m) \cdot \beta^*_m - \sum_{m = 0}^n f_p(m) \cdot \beta^*_m\\
& = \epsilon^* - 0 = \epsilon^* \geq 1/3.
\end{align*}
Thus, the degree of any polynomial that approximates $c$ on the distribution $\D$ with error of at most $1/3$ is $\Omega(\sqrt{n})$.
\end{proof}
\subsection{Upper Bound}
\label{sec:symmetric}
In this section, we describe how to approximate disjunctions on any symmetric distribution by using a linear combination of functions from a set of small size. Recall that $S_r$ denotes the set of all points from $\zo^n$ with weight $r$.
As we have seen above, symmetric distributions can behave very differently when compared to (constant bounded) product distributions. However, for the special case of the uniform distribution on $S_r$, denoted by $\D_r$, we show that for every disjunction $c$, there is a polynomial of degree $O(\log{(1/\epsilon)})$ that $\eps$-approximates it in $\ell_1$ distance on $\D_r$. As described in Section \ref{sec:our-results}, one can stitch together polynomial approximations on each $S_r$ to build a set of basis functions $\S$ such that every disjunction is well approximated by some linear combination of functions in $\S$. Thus, our goal is now reduced to constructing approximating polynomials on $\D_r$.
\eat{
\begin{lemma}
For any $0 \leq r \leq n$ and any disjunction $c$ there is a polynomial $p$ of degree at most $O( \log{(1/\epsilon)})$ such that
$\E_{\D_r}[|c(x) - p(x)|] \leq \eps$.
\label{monotone-sym}
\end{lemma}}
\begin{proof}[Proof of Theorem \ref{monotone-sym}]
We first assume that $c$ is monotone and without loss of generality $c=x_1 \vee \cdots \vee x_k$. We will also prove a slightly stronger claim that $\E_{\D_r}[|c(x) - p(x)|] \leq \E_{\D_r}[(c(x) - p(x))^2] \leq \eps$ in this case. Let $d:\zeto{k} \rightarrow \zo$ be the predicate associated with the disjunction, that is $d(i)=1$ whenever $i\geq 1$. Note that $c(x) = d\left(\sum_{i\in [k]} x_i\right)$. Therefore our goal is to find a univariate polynomial $f$ that approximates $d$ and then substitute $p_f(x) = f\left(\sum_{i\in [k]}x_i\right)$. This substitution preserves the total degree of the polynomial.
We break our construction into several cases based on the relative magnitudes of $r,k$ and $\epsilon$.
If $k \leq 2 \ln{(1/\epsilon)}$, then the univariate polynomial that exactly computes the predicate $d$ satisfies the requirements. Thus assume that $k > 2 \ln(1/\epsilon)$. If $r > n-k$, then, $c$ always takes the value $1$ on $S_r$ and thus the constant polynomial $1$ achieves zero error. If on the other hand, if $r \geq (n/k)\ln{(1/\epsilon)}$, then, $$\pr_{x \sim \D_r}[ c(x) = 0] = \frac{{{n-k} \choose r}}{{n \choose r}} = \prod_{i = 0}^{r-1} \left(1-\frac{k}{n-i}\right) \leq (1-k/n)^r \leq e^{-kr/n} \leq \epsilon.$$ In this case, the constant polynomial $1$ achieves an $\ell_2^2$ error of at most $\pr_{x \sim \D_r}[ c(x) = 0] \cdot 1 \leq \epsilon$.
Finally, observe that $r \leq (n/k) \ln{(1/\epsilon)}$ and $k > 2 \ln(1/\epsilon)$ implies $r \leq n/2$. Thus, for the remaining part of the proof, assume that $r < \min \{n-k, (n/k) \ln{(1/\epsilon), n/2}\}$.
Consider the univariate polynomial $f:\zeto{k} \rightarrow \R$ of degree $t$ (for some $t$ to be chosen later) that computes the predicate $d$ exactly on $\zeto{t}$. This polynomial is given by $$f(w) = 1 - \frac{1}{t!} \prod_{i=1}^t (i-w) = \left\{ \begin{subarray}[
1 - {w \choose t} \text{ for $w > t$}\\
1 \text{ for } 0 < w \leq t\\
0 \text{ for } w = 0\\
\end{subarray}\right. $$
Let $$\delta_j = \pr_{x \sim \D_r}[ |\{ i \cond x_i = 1\}| = j] = \frac{ {{n-k} \choose {r-j} } \cdot {k \choose j} }{{n \choose r}}.$$ The $\ell_2^2$ error of $p_f(x)$ on $c$ satisfies, $$||p_f-c||_2^2 = \E_{x \sim \D_r}[(c(x) - p_f(x))^2] = \sum_{j = t+1}^k \delta_j \cdot { {j \choose t}}^2 .$$ We denote the RHS of this equality by $\|d-f\|^2_2$.
We first upper bound $\delta_j$ as follows:
\begin{align*}\delta_j = \frac{{{n-k} \choose {r-j} } \cdot {k \choose j} }{{n \choose r} } &= \frac{(n-k)!}{(n-k-r+j)!(r-j)!} \cdot \frac{k!}{(k-j)!j!} \cdot \frac{(n-r)! r!}{n!}\\
&= \frac{1}{j!} \cdot \frac{r!}{(r-j)!} \cdot \frac{k!}{(k-j)!} \cdot \frac{(n-r)!}{n!} \cdot \frac{(n-k)!}{(n-k-r+j)!}\\
&\leq \frac{1}{j!} \cdot (rk)^j \cdot \frac{(n-k) \cdot (n-k-1) \cdots (n-k-r+j+1)}{n \cdot (n-1) \cdots (n-r+1)}\\
&\leq \frac{1}{j!} \cdot (n \ln{(1/\epsilon)})^j \cdot \frac{1}{(n-r+j) \cdot (n-r+j-1) \cdots (n-r+1)},
\end{align*}
where, in the second to last inequality, we used that $r <n/k \ln {(1/\epsilon)}$ to conclude that $rk \leq (n \ln{(1/\epsilon)})$.
Now, $r < n/2$ and thus $(n-r+1) > n/2$. Therefore,
\begin{align*}
\delta_j &\leq \frac{2^j \cdot (n\ln{(1/\epsilon)})^j}{n^j \cdot j!} = \frac{(2\ln{(1/\epsilon)})^j}{j!},
\end{align*}
and thus: $$\|d-f\|^2_2\leq \sum_{j = t+1}^k {j \choose t}^2 \frac{(2 \ln{(1/\epsilon)})^j}{j!}.$$
Set $t = 8e^2 \ln{(1/\epsilon)} $. Using $j! > (j/e)^j > (t/e)^j$ for every $j \geq t+1$, we obtain:
\begin{align}
\|d-f\|_2^2 &\leq \sum_{j = t+1}^k 2^{2j} \cdot \left( \frac{2 \ln{(1/\epsilon)}}{8e \ln{(1/\epsilon)}} \right)^j \leq \epsilon \cdot \sum_{j = t+1}^{\infty} 1/e^j \leq \epsilon. \label{eq:final-bound}
\end{align}
To see that $\E_{\D_r}[|c(x) - p(x)|] \leq \E_{\D_r}[(c(x) - p(x))^2]$ we note that in all cases and for all $x$, $|p(x)-c(x)|$ is either $0$ or $\geq 1$. This completes the proof of the monotone case.
We next consider the more general case when $c = x_1 \vee x_2 \vee \ldots \vee x_{k_1} \vee \bar{x}_{k_1+1} \vee \bar{x}_{k_1+2} \vee \ldots \vee \bar{x}_{k_1+k_2}$. Let $c_1 = x_1 \vee x_2 \vee \ldots \vee x_{k_1}$ and $c_2 = \bar{x}_{k_1+1} \vee \bar{x}_{k_1+2} \vee \ldots \vee \bar{x}_{k_1+k_2}$ and $k = k_1 + k_2$. Observe that $c = 1 - (1-c_1) \cdot (1-c_2) = c_1+c_2-c_1c_2$.
Let $p_1$ be a polynomial of degree $O(\log{(1/\epsilon)})$ such that $\|c_1 - p_1\|_1 \leq \|c_1 - p_1\|_2^2 \leq \epsilon/3$. Note that if we swap 0 and 1 in $\zo^n$ then $c_2$ will be equal to a monotone disjunction $\bar{c}_2 = x_{k_1+1} \vee x_{k_1+2} \vee \ldots \vee x_{k_1+k_2}$ and $\D_r$ will become $\D_{n-r}$. Therefore by the argument for the monotone case, there exists a polynomial $\bar{p}_2$ of degree $O(\log{(1/\epsilon)})$ such that $\|\bar{c}_2 - \bar{p}_2\|_1 \leq \epsilon/3$. By renaming the variables back we will obtain a polynomial $p_2$ of degree $O(\log{(1/\epsilon)})$ such that $\|c_2 - p_2\|_1 \leq \|c_2 - p_2\|_2^2 \leq \epsilon/3$.
Now let $p = p_1+p_2 - p_1p_2$. Clearly the degree of $p$ is $O( \log{(1/\epsilon)})$. We now show that $\|c-p\|_1\leq \epsilon$:
\alequn{
\E_{x\sim \D_r}[|c(x)-p(x)|] &= \E_{x\sim \D_r}[|(1-c(x))-(1-p(x))|] \\
& = \E_{x\sim \D_r}[|(1-c_1)(1-c_2)-(1-p_1)(1-p_2)|]\\
&= \E_{x\sim \D_r}[|(1-c_1)(p_2 -c_2) + (1-c_2)(p_1-c_1) - (c_1-p_1)(c_2-p_2)|] \\
&\leq \E_{x\sim \D_r}[|(1-c_1)(p_2 -c_2)|] + \E_{x\sim \D_r}[|(1-c_2)(p_1-c_1)|] +\E_{x\sim \D_r}[|(c_1-p_1)(c_2-p_2)|] \\
&\leq \E_{x\sim \D_r}[|p_2 -c_2|] + \E_{x\sim \D_r}[|p_1-c_1|] + \sqrt{\E_{x\sim \D_r}[(c_1-p_1)^2]\E_{x\sim \D_r}[(c_2-p_2)^2]} \\
& \leq \eps/3 +\eps/3+\eps/3 = \eps.
}
\end{proof}
\section{Polynomial Approximation on Product Distributions}
\label{sec:product}
In this section, we show that for every product distribution $\D = \prod_{i \in [n]} \D_i$, every $\epsilon > 0$ and every disjunction (or conjunction) $c$ of length $k$, there exists a polynomial $p: \zo^n \rightarrow \R$ of degree {$O(\log{(1/\epsilon)})$} such that $p$ $\eps$-approximates $c$ in $\ell_1$ distance on $\D$.
\eat{
\begin{lemma}
Let $\D = \prod_{i \in [n]} \D_i$ be any product distribution on $\zo^n$ with $\pr_{x_i \sim \D^i}[x_i = 1] = \mu_i $ for each $i \in [k]$ and let $c = x_1 \vee x_2 \vee \ldots \vee x_k$ be the monotone disjunction on all the variables. For any $ \epsilon > 0$, there exists a polynomial $f: \zo^k \rightarrow \R$ of degree $O( \log{(1/\epsilon)})$ such that $$\|c-f\|_1 = \E_{x \in \D} |c(x) - f(x)| \leq \epsilon.$$
\label{lem:product-poly}
\end{lemma}}
\begin{proof}[Proof of Theorem \ref{th:product-poly-intro}]
First, we note that without loss of generality we can assume that the disjunction $c$ is equal to $x_1 \vee x_2 \vee \cdots \vee x_k$ for some $k\in[n]$. We can assume monotonicity since we can convert negated variables to un-negated variables by swapping the roles of $0$ and $1$ for that variable. The obtained distribution will remain product after this operation. Further we can assume that $k=n$ since variables with indices $i > k$ do not affect probabilities of variables with indices $\leq k$ or the value of $c(x)$.
We first note that we can assume that $\pr_{x \sim \D} [x = 0^k] > \epsilon$ since, otherwise, the constant polynomial $1$ gives the desired approximation. Let $\mu_i = \pr_{x_i \sim \D^i}[x_i = 1]$. Since $c$ is a symmetric function, its value at any $x \in \zo^k$ depends only on the Hamming weight of $x$ that we denote by $w(x)$. Thus, we can equivalently work with the univariate predicate $d:\{0,1,\ldots,k\} \rightarrow \zo$, where $d(i) = 1$ for $i >0$ and $d(0) = 0$.
As in the proof of Theorem \ref{monotone-sym}, we will approximate $d$ by a univariate polynomial $f$ and then use the polynomial $p_f(x) = f(w(x))$ to approximate $c$.
Let $f:\{0,1, \ldots, k\} \rightarrow \R$ be the univariate polynomial of degree $t$ that matches $d$ on all points in $\{0,1, \ldots, t\}$. Thus, $$f(w) = 1 - \frac{1}{t!} \cdot \prod_{i=1}^t (w-i) = \left\{ \begin{subarray}[
1 - {w \choose t} \text{ for $w > t$}\\
1 \text{ for } 0 < w \leq t\\
0 \text{ for } w= 0\\
\end{subarray}\right. $$
We have,
$$\E_{x \sim \D_r}[(c(x) - p_f(x))^2] = \sum_{j = 0}^{k} \pr_{x \sim \D}[ w(x) = j] \cdot |d(j)-f(j)|$$
and we denote the RHS of this equation by $\|d-f\|_1$.
Then:
\begin{align}
\|d-f\|_1 &= \sum_{j = t+1}^k \pr_{ \D} [ w(x) = j] \cdot |1-f(j)| \nonumber\\
&= \sum_{j = t+1}^k \pr_{ \D} [ w(x) = j] \cdot {j \choose t} . \label{error-exp}
\end{align}
Let us now estimate $\pr_{\D}[w(x) = j]$.
\begin{align*}
\pr_{\D}[ w(x)= j] & = \sum_{S \subseteq [n]\text{, } |S| = j} \prod_{i \in S} \mu_i \cdot \prod_{i \notin S} (1-\mu_i ) \\
&\leq \sum_{S \subseteq [n]\text{, } |S| = j} \prod_{i \in S} \mu_i \\
\end{align*}
Observe that in the expansion of $(\sum_{i=1}^k \mu_i)^j$, the term $\prod_{i \in S} \mu_i$ occurs exactly $j!$ times. Thus, $$ \sum_{S \subseteq [n]\text{, } |S| = j} \prod_{i \in S} \mu_i \leq \frac{( \sum_{i=1}^k \mu_i)^j }{j!}.$$
Set $\mu_{avg} = \frac{1}{k} \sum_{i=1}^k \mu_i$. We have: $$\epsilon \leq \pr_{x \sim \D} [ x = 0^k] = \prod_{i=1}^k (1-\mu_i) \leq \left(1 - \frac{1}{k} \cdot \sum_{i=1}^k \mu_i\right)^k = (1-\mu_{avg})^k.$$
Thus, $\mu_{avg} = c/k$ for some $c \leq 2\ln{(1/\epsilon)}$ whenever $k \geq k_0$ where $k_0$ is some universal constant. In what follows, assume that $k \geq k_0$. (Otherwise, we can use the polynomial of degree equal to $k$ that exactly computes the predicate $d$ on all points).
We are now ready to upper bound the error $\|d-f\|_1$. From Equation \eqref{error-exp}, we have:
\begin{align*}
\|d-f\|_1 &= \sum_{j = t+1}^k \pr_{ \D} [ w(x) = j] \cdot {j \choose t} \leq \sum_{j = t+1}^k \frac{( \sum_{i=1}^k \mu_i)^j }{j!} \cdot {j \choose t}\\
& \leq \sum_{j = t+1}^k {j \choose t} \cdot \frac{(2\ln(1/\epsilon) )^j }{j!}
\end{align*}
Setting $t = 4e^2 \ln {(1/\epsilon)}$ and using the calculation from Equation \eqref{eq:final-bound} in the proof of Thm.~\ref{monotone-sym}, we obtain that the error $\|d-f\|_1 \leq \epsilon$.
\end{proof}
\section{Agnostic Learning of Disjunctions}
\label{sec:agnostic-learn}
Combining Thm.~\ref{th:lae-lp} with the results of the previous section (and the discussion in Section \ref{sec:our-results}), we obtain an agnostic learning algorithm for the class of all disjunctions on product and symmetric distributions running in time $n^{O(\log{(1/\epsilon)})}$.
\begin{corollary}[Cor.~\ref{agnostic-learning}, restated]
There is an algorithm that agnostically learns the class of disjunctions on any product or symmetric distribution on $\zo^n$ with excess error of at most $\epsilon$ in time $n^{O( \log{(1/\epsilon)})}$.
\end{corollary}
We now remark that any algorithm that agnostically learns the class of disjunctions (or conjunctions) on $n$ inputs on the uniform distribution on $\zo^n$ in time $n^{o(\log{(\frac{1}{\epsilon})})}$ would yield a faster algorithm for the notoriously hard problem of Learning Sparse Parities with Noise(SLPN). The reduction is based on the technique implicit in the work of \citet{KKMS05} and \citet{Feldman:12jcss}.
For $S \subseteq [n]$, we use $\chi_S$ to denote the parity of inputs with indices in $S$. Let $\U$ denote the uniform distribution on $\zo^n$. We say that random examples of a Boolean function $f$ have noise of rate $\eta$ if the label of a random example equals $f(x)$ with probability $1 - \eta$ and $1-f(x)$ with probability $\eta$.
\begin{problem}[Learning Sparse Parities with Noise]
For $\eta \in (0,1/2)$ and $k \leq n$ the problem of learning $k$-sparse parities with noise $\eta$ is the problem of finding (with probability at least $2/3$) the set $S \subseteq [n]$,$|S| \leq k$, given access to random examples with noise of rate $\eta$ of parity function $\chi_S$.
\end{problem}
The fastest known algorithm for learning $k$-sparse parities with noise $\eta$ is a recent breakthrough result of Valiant \citeyearpar{Val12} which runs in time $O(n^{0.8k} \poly(\frac{1}{1-2\eta}))$ .
\citet{KKMS05} and \citet{Feldman:12jcss} prove hardness of agnostic learning of majorities and conjunctions, respectively, based on correlation of concepts in these classes with parities. We state below this general relationship between correlation with parities and reduction to SLPN, a simple proof of which appears in \citep{FeldmanKV:13}.
\begin{lemma}
Let $\C$ be a class of Boolean functions on $\zo^n$. Suppose, there exist $\gamma >0$ and $k \in \N$ such that for every $S \subseteq [n]$, $|S| \leq k$, there exists a function, $f_S \in \C$, such that $|\E_{x \sim \U} [f_S(x) \chi_S(x)]|\geq \gamma(k)$. If there exists an algorithm $\A$ that learns the class $\C$ agnostically with excess error $\epsilon$ in time $T(n, \frac{1}{\epsilon})$ then, there exists an algorithm $\A'$ that learns $k$-sparse parities with noise $\eta < 1/2$ in time $\poly(n,\frac{1}{(1-2\eta)\gamma(k)}) + 2 T(n, \frac{2}{(1-2\eta)\gamma(k)})$. \label{cor2lpn}
\end{lemma}
The correlation between a disjunction and a parity is easy to estimate.
\begin{fact}
For any $S \subseteq [n]$, $|\E_{x \sim \U} [\ORR_S(x) \chi_S(x)]| = \frac{1}{2^{|S|-1}}$.
\end{fact}
We thus immediately obtain the following corollary.
\begin{theorem}
Suppose there exists an algorithm that learns the class of Boolean disjunctions over the uniform distribution agnostically with excess error of $\epsilon > 0$ in time $T(n,\frac{1}{\epsilon})$. Then there exists an algorithm that learns $k$-sparse parities with noise $\eta < \frac{1}{2}$ in time $\poly(n,\frac{2^{k-1}}{1-2\eta})+ 2T(n, \frac{2^{k-1}}{1-2\eta})$. In particular, if $T(n, \frac{1}{\epsilon}) = n^{o(\log{(1/\epsilon)})}$, then, there exists an algorithm to solve $k$-SLPN in time $n^{o(k)}$.
\end{theorem}
Thus, any algorithm that is asymptotically faster than the one from Cor.~\ref{agnostic-learning} yields a faster algorithm for $k$-SLPN.
\bibliographystyle{plainnat}
|
\section{Introduction}\label{Introduction}
The multiple tiling problem can be described as follows: cover every point in $ \mathbb R^d$ exactly $k$ times by translates of a convex polytope using a discrete multiset $\Lambda$ (also known as a \textbf{tiling set}) of translation vectors.
However, in the process of trying to cover every point in $ \mathbb R^d$, we may be forced to cover points in the boundary of its translates for more than $k$ times.
To avoid this technicality, we say that $P$ $k$-tiles $ \mathbb R^d$ with $\Lambda$ if every point that does not belong to the boundary of any translate of $P$ is covered exactly $k$ times.
We call a polytope $P$ that satisfies the condition above a $\mathbf{k}$\textbf{-tiler}.
In the special case when $k$ is equal to $1$, the multiple tiling problem becomes what is traditionally known as the translational tiling problem. For more details on the translational tiling problem, the reader is referred to \cite{Alexandrov,Gruber,KM10} for a nice overview of the topic.
The translational tiling problem is a classical topic in discrete geometry with several beautiful structural results.
In 1897, Minkowski \cite{Minkowski} gave a necessary condition for a polytope to be a $1$-tiler. He proved that if a convex polytope $P$ $1$-tiles $ \mathbb R^d$, then $P$ must be centrally symmetric and all facets of $P$ must be centrally symmetric.
It was not until 50 years later that Venkov \cite{Venkov} found a necessary and sufficient condition for a polytope $P$ to be a 1-tiler.
He showed that $P$ 1-tiles $ \mathbb R^d$ if and only if $P$ is centrally symmetric, all facets of $P$ are centrally symmetric, and each belt of $P$ contains four or six facets.
The same result was later rediscovered independently by McMullen \cite{McMullen} in 1980.
In contrast to the situation for $1$-tilers, there are still a lot of unsolved problems on the structure of $k$-tilers. This is partly because $k$-tilers have a much richer structure compared to $1$-tilers. For example, in two dimensions there are only two types of convex polytopes that $1$-tile $ \mathbb R^2$, namely centrally symmetric parallelograms and centrally symmetric hexagons. In contrast, all centrally symmetric integer polygons are $k$-tilers in $ \mathbb R^2$ \cite{Gravin-Robins-Shiryaev}. With that being said, there are several important results for multiple tiling that mirror the results for $1$-tiling. In 1994, Bolle \cite{Bolle} gave a necessary and sufficient condition for a polytope to $k$-tile $ \mathbb R^2$ with a lattice, and in 2012, Gravin, Robins, and Shiryaev \cite{Gravin-Robins-Shiryaev} proved that a $k$-tiler in $ \mathbb R^d$ must be centrally symmetric and all its facets must be centrally symmetric, providing a multiple tiling analogue for Minkowski's condition.
The main motivation for this paper comes from a classical result of McMullen \cite{McMullen} that if a convex polytope $P$ $1$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, then $P$ can, in fact, $1$-tile $ \mathbb R^d$ with a lattice $\mathcal{L}$.
A generalization of McMullen's theorem for $k$-tiling was conjectured by Gravin, Robins, and Shiryaev \cite{Gravin-Robins-Shiryaev}, and is stated below:
\begin{conjecture}\label{conjecture last boss}\textnormal{\cite[Conjecture 7.3]{Gravin-Robins-Shiryaev}}
If a convex polytope $P$ $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, then $P$ $m$-tiles $ \mathbb R^d$ with a lattice $\mathcal{L}$ for some $m$ (not necessarily equal to $k$).
\end{conjecture}
There has been some recent progress on Conjecture \ref{conjecture last boss} in lower dimensions.
We say that a set $\mathcal{Q}$ is a \textbf{quasi-periodic set} if $\mathcal{Q}$ is a finite union of translated
lattices, not necessarily of the same lattice.
Kolountzakis \cite{Kolountzakis} showed that if $P$ $k$-tiles $ \mathbb R^2$ with a discrete multiset $\Lambda$, then $P$ can $m$-tile $ \mathbb R^2$ with a quasi-periodic set for some $m$. Kolountzakis' result was later extended by Shiryaev \cite{Shiryaev}, who proved Conjecture \ref{conjecture last boss} when $d$ is equal to 2.
For the case when $d$ is equal to 3, Gravin, Kolountzakis, Robins, and Shiryaev \cite{GKRS} showed that every $k$-tiler $P$ in $ \mathbb R^3$ can $m$-tile $ \mathbb R^3$ with a quasi-periodic set for some $m$.
Motivated by the results of \cite{Kolountzakis} and \cite{GKRS},
we focus here on studying the
\textbf{quasi-periodic tiling} problem for general dimensions,
which is the multiple tiling problem with an additional assumption that every element of the tiling multiset $\Lambda$ is contained in a {quasi-periodic set} (note that the multiset $\Lambda$ is not necessarily a quasi-periodic set).
We approach the quasi-periodic tiling problem by studying an equivalent lattice-point enumeration problem, which will be described in Section \ref{section lattice point enumeration}.
This approach allows us to employ several tools from lattice-point enumeration that are otherwise
not available for general multiple tiling problems.
For more details regarding discrete-point enumeration of polytopes, the reader is referred to the work of Beck and Robins \cite{Beck-Sinai} and Barvinok \cite{Sasha-Barvinok}.
Throughout this paper, we will use two different notions for \textbf{general positions}, one for vectors and one for lattices. Let $P$ be a a fixed convex polytope and let $\partial P$ denote the boundary of $P$.
We say that a vector $v \in \mathbb R^d$ is in \textbf{general position} with respect to a discrete multiset $\Lambda$ if $v$ is not contained in $\partial P+\Lambda$, the union of translates of boundary of $P$ by $\Lambda$. The second notion is defined for lattices in $ \mathbb R^d$. Let $\mathcal{Q}$ be a union of $n$ translated lattices $\mathcal{L}_1$, $\mathcal{L}_2,\ldots,
\mathcal{L}_n$.
We say that $\mathcal{L}_i$ is in \textbf{ general position} with respect to $\mathcal{Q}$ if the set $ \mathbb R^d \setminus H_i$ is path-connected, where $H_i$ is defined as:
\begin{equation}\label{technical condition}
H_i:= \bigcup_{j=1, j\neq i}^{n} (\partial P+\mathcal{L}_i) \cap (\partial P+\mathcal{L}_j).
\end{equation}
We will refer to
the hypothesis that $\mathcal{L}_i$ is in general position with respect to $\mathcal{Q}$ as Hypothesis \ref{technical condition}.
The first result in this paper is that if Hypothesis \ref{technical condition} holds, then Conjecture \ref{conjecture last boss} is true.
\begin{theorem}\label{main theorem 1}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, and suppose that every element of $\Lambda$ is contained in a quasi-periodic set $\mathcal{Q}$. If a lattice $\mathcal{L}$ in $\mathcal{Q}$ is in general position with respect to $\mathcal{Q}$ and $\mathcal{L} \cap \Lambda$ is non-empty, then $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}$ for some $m$.
\end{theorem}
If the quasi-periodic set $\mathcal{Q}$ in Theorem \ref{main theorem 1} is also a lattice, then the lattice $\mathcal{Q}$ is in general position with respect to $\mathcal{Q}$ by definition, and Theorem \ref{main theorem 1} gives us the following corollary.
\begin{corollary}\label{corollary one lattice}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
If every element of $\Lambda$ is contained in a lattice $\mathcal{L}$, then $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}$ for some $m$. \qed
\end{corollary}
We note that Theorem \ref{main theorem 1} will fail to hold if the hypothesis that $\mathcal{L}$ is in general position with respect to $\mathcal{Q}$ is omitted.
In Example \ref{example rectangle}, we present a a polytope $P$ that $k$-tiles $ \mathbb R^d$ with a quasi-periodic set $\mathcal{Q}$ that contains a lattice $\mathcal{L}$, and yet $\mathcal{L}$ is not a tiling set of $P$.
The fact that Hypothesis \ref{technical condition} cannot be omitted from Theorem \ref{main theorem 1} naturally leads us to consider the case where every lattice in $\mathcal{Q}$ is not in general position with respect to $\mathcal{Q}$.
In Section \ref{section for main theorem 2} we address this scenario
for the case where $\mathcal{Q}$ is a union of two translated copies of a lattice, and we show that
Conjecture \ref{conjecture last boss} holds in this case.
\begin{theorem}\label{main theorem 2}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
If every element of $\Lambda$ is contained in a union of two translated copies of one single lattice,
then there is a lattice $\mathcal{L}$ in $ \mathbb R^d$ such that $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}$ for some $m$.
\end{theorem}
The paper is organized as follows. In Section \ref{definition} we introduce definitions and notations used in this paper.
We use Section \ref{section lattice point enumeration} to establish the connection between the $k$-tiling problem and a lattice-point enumeration problem.
Section \ref{section main theorem 1} is devoted to the proof of Theorem \ref{main theorem 1}, and Section \ref{section for main theorem 2} is devoted to the proof of Theorem \ref{main theorem 2}.
Finally, in Section \ref{conjectures} we discuss possible future research that can be done to prove Conjecture \ref{conjecture last boss}.
\section{Definitions and preliminaries}\label{definition}
Throughout this paper, we use $P$ to denote a convex polytope in $ \mathbb R^d$, $\text{Int}(P)$ to denote the interior of $P$, and $\partial P$ to denote the boundary of $P$ (the closure of $P$ minus the interior of P).
We note that there is no loss in generality in assuming that $P$ is a convex polytope, because
every convex body that $k$-tiles $ \mathbb R^d$ is necessarily a polytope \cite{McMullen}.
We use $\Lambda$ to denote a discrete multiset of vectors in $ \mathbb R^d$, $\mathcal{L}$ to denote a lattice in $ \mathbb R^d$, and $\mathcal{Q}$ to denote a \textbf{quasi-periodic set}, which is a finite union of translated lattices, not necessarily of the same lattice.
We use $\#(A)$ to denote the cardinality of a finite multiset $A$ (counted with multiplicities).
The intersection of a multiset $A$ and a set $S$, denoted by $A \cap S$, is the multiset that contains all elements $a$ in $A$ that are also contained in $S$.
The multiplicity of an element $a$ in $A \cap S$ is equal to the multiplicity of $a$ in $A$.
The complement of a set $S$ with respect to a multiset $A$, denoted by $A\setminus S$, is the set $A \cap S^c$.
A convex polytope $P$ is said to \textbf{ k-tile } $ \mathbb R^d$ ($k$ being a positive natural number) with a discrete multiset $\Lambda$ of vectors in $ \mathbb R^d$ if
\begin{equation}\label{equation k-tiling}\sum_{\lambda \in \Lambda} \boldsymbol{1}_{P+\lambda}(v)=k,\end{equation}
for all $v \notin \partial P+\Lambda$, where $\boldsymbol{1}_X$ is the indicator function of the set $X$.
Throughout this paper, we assume that $ \mathfrak{h}$ is a fixed vector in $ \mathbb R^d$ such that every line with direction vector $ \mathfrak{h}$ meets $\partial P$
at finitely many points.
The \textbf{half-open} counterpart $P^{ \mathfrak{h}}$ of a convex polytope $P$ is the subset of the closure of $P$ that contains all points $v \in \mathbb R^d$
which satisfies the property that
for a sufficiently small $\epsilon_v>0$, the ray $r_{\epsilon_v}:=\{v+ c \mathfrak{h} \ | \ 0 < c < \epsilon_v \} $ is contained in $\text{Int}(P)$.
Note that $P^{ \mathfrak{h}}$ consists of $\text{Int}(P)$ and a part of $\partial P$.
In the particular case when $P$ is a cube,
the polytope $P^{ \mathfrak{h}}$ is
the half-open cube defined in \cite{Stanley}.
For a a discrete multiset $\Lambda$ and a convex polytope $P$ in $ \mathbb R^d$, the \textbf{$\Lambda$-point enumerator} of $P$ is the integer $\#(\Lambda \cap P)$, which is the number of points of $\Lambda$ (counted with multiplicities) contained in $P$. When $\Lambda$ is a lattice, we refer to $\#(\Lambda \cap P)$ as a lattice-point enumerator.
We define two integer-valued functions, $L_{\Lambda}$ and $L_{\Lambda}^{ \mathfrak{h}}$, on every point $v$ in $ \mathbb R^d$ as follows:
\[L_{\Lambda}(v):= \# (\Lambda \cap \{-1 \cdot P+ v \}), \ L_{\Lambda}^{ \mathfrak{h}}(v):=\# (\Lambda \cap \{-1 \cdot P^{ \mathfrak{h}}+ v \}), \] i.e. $L_{\Lambda}(v)$ is the number of points of $\Lambda$ contained in the translate of $-1 \cdot P$ by $v$, and $L_{\Lambda}^{ \mathfrak{h}}(v)$ is the number of points of $\Lambda$ contained in the translate of $-1 \cdot P^{ \mathfrak{h}}$ by $v$. If the intended multiset $\Lambda$ is evident from the context,
we will use $L$ and $L^{ \mathfrak{h}}$ as a shorthand for $L_{\Lambda}$ and $L_{\Lambda}^{ \mathfrak{h}}$, respectively.
These two functions play an important role in relating the multiple tiling problem to the lattice-point enumeration problem, which is discussed in Section \ref{section lattice point enumeration}.
\section{Lattice-point enumeration of polytopes}\label{section lattice point enumeration}
In this section, we present a lattice-point enumeration problem that is equivalent to $k$-tiling problem.
This equivalence was first shown in \cite{Gravin-Robins-Shiryaev}, where it was employed to show that all rational $k$-tilers can $m$-tile with a lattice for some $m$.
However, we replace the polytope $P$ by its half-open counterpart $P^{ \mathfrak{h}}$ in the statement of the equivalence.
This is done so that we can drop the technical condition in the equivalence concerning vectors in general position.
\begin{lemma}\label{half open}
A $d$-dimensional convex polytope $P$ $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$ if and only if its half-open counterpart $P^{ \mathfrak{h}}$ $k-$tiles $ \mathbb R^d$ with $\Lambda$. Moreover, if $P^{ \mathfrak{h}}$ $k$-tiles $ \mathbb R^d$ with $\Lambda$, then
\begin{equation}\label{equation half open}
\sum_{\lambda \in \Lambda} \boldsymbol{1}_{P^{ \mathfrak{h}}+\lambda}(v)=k,
\end{equation}
for all $v $ in $ \mathbb R^d$.
\end{lemma}
\begin{proof}
Note that $P$ and $P^{ \mathfrak{h}}$ have the same interior, which implies that
\[\sum_{\lambda \in \Lambda}\boldsymbol{1}_{P^{ \mathfrak{h}}+\lambda}(v)=\sum_{\lambda \in \Lambda} \boldsymbol{1}_{P+\lambda}(v), \]
for all $v \notin \partial P+\Lambda$. Therefore, $P$ $k$-tiles $ \mathbb R^d$ with $\Lambda$ if and only if $P^{ \mathfrak{h}}$ $k$-tiles $ \mathbb R^d$ with $\Lambda$.
To prove the second part of the claim, let $v$ be an arbitrary point in $ \mathbb R^d$. By our assumption on $ \mathfrak{h}$, the ray $r_{\epsilon_v}= \{v+ c \mathfrak{h} \ | \ 0< c < \epsilon_v \}$ intersects $\partial P+\Lambda$ at finitely many points.
Hence for a sufficiently small $\epsilon_v>0$, the ray $r_{\epsilon_v}$ does not intersect $\partial P+\Lambda$.
Because $P^{ \mathfrak{h}}$(and hence $P$) $k$-tiles $ \mathbb R^d$ with $\Lambda$, this implies that there are exactly $k$ vectors $\lambda_1, \ldots, \lambda_k$ in $\Lambda$ such that $r_{\epsilon_v}$ is contained in the interior of $P+\lambda_i$ for all $i$. By the definition of half-open polytopes in Section \ref{definition}, this means that $v$ is contained in $P^{ \mathfrak{h}}+\lambda_i$ for all $i$, and hence we have:
\begin{equation*}\sum_{\lambda \in \Lambda} \boldsymbol{1}_{P^{ \mathfrak{h}}+\lambda}(v)=k,\end{equation*}
for all $v \in \mathbb R^d$.
\qed \end{proof}
The lemma below was shown in \cite{Gravin-Robins-Shiryaev} for the case when $v \in \mathbb R^d$ is in general position with respect to $\Lambda$;
our proof is virtually identical, with a minor adjustment for the case when $P$ is replaced by $P^{ \mathfrak{h}}$.
\begin{lemma}\label{lemma GRS}\textnormal{(c.f. \cite[Lemma 3.1]{Gravin-Robins-Shiryaev})} A convex polytope $P$ $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$ if and only $L_{\Lambda}( v)$ is equal to $k$
for every $v \in \mathbb R^d$ that is in general position with respect to $\Lambda$.
Moreover, $P^{ \mathfrak{h}}$ $k$-tiles $ \mathbb R^d$ with $\Lambda$ if and only $L_{\Lambda}^{ \mathfrak{h}}( v)$ is equal to $k$ for every $v$ in $ \mathbb R^d$.
\end{lemma}
\begin{proof}
For every $v$ in $ \mathbb R^d$, we can write
\[\sum_{\lambda \in \Lambda} \boldsymbol{1}_{P^ \mathfrak{h}+\lambda}(v)= \sum_{\lambda \in \Lambda} \boldsymbol{1}_{-1 \cdot P^ \mathfrak{h}+v}(\lambda)= \# (\Lambda \cap \{-1 \cdot P^ \mathfrak{h} + v\})=L_{\Lambda}( v).\]
By Equation \ref{equation half open} in Lemma \ref{half open}, this implies that $P^{ \mathfrak{h}}$ $k$-tiles $ \mathbb R^d$ if and only if $L_{\Lambda}^{ \mathfrak{h}}(v)$ is equal to $k$ for every $v$ in $ \mathbb R^d$.
By a similar argument,
Equation \ref{equation k-tiling} in the definition of $k$-tilers implies that $P$ $k$-tiles $ \mathbb R^d$ if and only if $L_{\Lambda}( v)
$ is equal to $k$ for every $v \in \mathbb R^d$ that is in general position with respect to $\Lambda$.
\qed \end{proof}
Here we list two properties of the function $L_\Lambda$ that will be used throughout this paper:
\begin{enumerate}[leftmargin=*]
\item \label{property periodic} If $\mathcal{L}$ is a lattice in $ \mathbb R^d$, then the function $L_{\mathcal{L}}$ is a periodic function of $\mathcal{L}$ (i.e. $L_{\mathcal{L}}(v+\lambda)=L_{\mathcal{L}}(v) $ for every $v$ in $ \mathbb R^d$ and every $\lambda$ in $\mathcal{L}$). This is because a lattice-point enumerator is invariant under translation by elements contained in the lattice.
\item \label{property general position} The function $L_{\Lambda}$ is a constant function in a sufficiently small neighborhood $B_v$ of $v$ in $ \mathbb R^d$ if $v$ is in general position with respect to $\Lambda$. This is because if $v \notin \partial P+\Lambda$, then $-1 \cdot P +v$ does not contain any points from $\Lambda$ in its boundary. Hence moving $v$ in any sufficiently small direction will not change the value of $L_{\Lambda}(v)$.
\end{enumerate}
It can be easily checked that the two properties above also hold for the function $L_{\Lambda}^{ \mathfrak{h}}$.
\section{Proof of Theorem \ref{main theorem 1} } \label{section main theorem 1}
We start this section by proving a functional equation involving the function $L_{\Lambda}$.
The proof borrows several ideas from asymptotic analysis of infinite sums in $ \mathbb R^d$.
\begin{lemma}\label{asymptotic method}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
Let $\mathcal{L}$ be a lattice and let $a_1,\ldots, a_n$ be vectors in $ \mathbb R^d$.
If every element of $\Lambda$ is contained in the
finite union $\bigcup_{i=1}^{n} a_i+\mathcal{L}$, then
there are non-negative real numbers $g_1, \ldots, g_n$ such that
\begin{equation}\label{equation in props asymptotic}\sum_{i=1}^{n} g_i \cdot L_{\mathcal{L}}^{ \mathfrak{h}}(v-a_i)=k,
\end{equation}
for all $v \in \mathbb R^d$.
\end{lemma}
\begin{proof}
Without loss of generality, we can assume that $\mathcal{L}= \mathbb Z^d$.
For a real vector $w=(w_1, \ldots, w_d)$ in $ \mathbb R^d$, we use $ \mathbb R^d_{\geq w}$ to denote the set $\{(w_1', \ldots, w_d'): w_i' \geq w_i \text{ for } i \in \{1,\ldots,d\} \}$.
We use $ \Lambda_{\geq 0}$ to denote the multiset $\Lambda \cap \mathbb R^d_{\geq0}$.
Let $v \in \mathbb R^d$ be an arbitrary vector.
Because $P^{ \mathfrak{h}}$ $k$-tiles $ \mathbb R^d$ with $\Lambda$, this implies that there is a vector ${\alpha(v)} \in \mathbb R^d$ such that every point in $ \mathbb R^{d}_{\geq {\alpha(v)}}$ is covered exactly $k$ times by $P^{ \mathfrak{h}}+v+\Lambda_{\geq 0}$. Also notice that because $\Lambda_{\geq 0}$ is in the positive orthant, there is a vector $\beta(v)$ in $ \mathbb R^d$ such that $P^{ \mathfrak{h}}+v+\Lambda_{\geq 0}$ is contained in $ \mathbb R^d_{\geq \beta(v)}$. Without loss of generality, we can assume that both ${\alpha(v)}$ and $\beta(v)$ are integer vectors.
Let $\Gamma(v)$ be the multiset $ ( P^{ \mathfrak{h}}+v+\Lambda_{\geq 0}) \setminus \mathbb R^d_{\geq {\alpha(v)}} $.
Because $ P^{ \mathfrak{h}}+v+\Lambda_{\geq 0}$ covers every point in $ \mathbb R^d_{\geq {\alpha(v)}}$ exactly $k$ times, we have the following equality:
\begin{align}\label{equation start}
\sum_{\substack{x \in P^{ \mathfrak{h}}+v+\Lambda_{\geq 0}, \\ x \in \mathbb Z^d}} z^x &= \sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}} } kz^x + \sum_{x \in \Gamma(v) \cap \mathbb Z^d}z^x,
\end{align}
where $z^x$ is the multivariable polynomial $z_1^{x_1}\ldots z_d^{x_d}$ and $x$ is an integer vector $(x_1, \ldots, x_d)$.
We assume $|z_j|<1$ so that all the sums in Equation \ref{equation start} converge to a well-defined value.
We define the multisets $\Lambda_i$ for $i\in\{1,\ldots, n\}$ recursively by $\Lambda_i:= ( \Lambda_{\geq 0} \cap \{a_i + \mathbb Z^d\} )\setminus \bigcup_{j=1}^{i-1} \Lambda_j$ for $ i \in \{1,\ldots, n\}$.
Note that $\boldsymbol{1}_{\Lambda_{\geq 0}}= \sum_{i=1}^{n} \boldsymbol{1}_{\Lambda_i}$, and every element of $\Lambda_i-a_i$ is contained in $ \mathbb Z^d$.
With this notation, Equation \ref{equation start} now becomes
\begin{align}
\sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}} } kz^x + \sum_{x \in \Gamma(v) \cap \mathbb Z^d}z^x
= \sum_{i=1}^{n} \sum_{\substack{x \in P^{ \mathfrak{h}}+v+\Lambda_i,\\ x \in \mathbb Z^d}} z^x
& =\sum_{i=1}^{n} \left(\sum_{\substack{x \in P^{ \mathfrak{h}}+v+a_i, \\ x \in \mathbb Z^d}} z^x \cdot \sum_{y \in \Lambda_i -a_i} z^y \right) \label{equation intermediate}\\
&= \sum_{i=1}^{n} \# ( \mathbb Z^d \cap \{ P^{ \mathfrak{h}}+a_i+v \} ) \cdot \sum_{y \in \Lambda_i -a_i} z^y. \label{equation important}
\end{align}
Let $s_i(z):=\sum_{y \in \Lambda_i -a_i} z^y \cdot \prod_{j=1}^d (1-z_j)$.
By
multiplying
Equation \ref{equation important} with $\prod_{j=1}^d (1-z_j) $, we obtain that:
\begin{align}\label{equation no limit}
\sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}} } kz^x \cdot \prod_{j=1}^d (1-z_j) + \sum_{x \in \Gamma(v) \cap \mathbb Z^d}z^x \cdot \prod_{j=1}^d (1-z_j)=\sum_{i=1}^{n} \# ( \mathbb Z^d \cap \{ P^{ \mathfrak{h}}+a_i+v \} ) \cdot s_i(z)
\end{align}
We now show that the left side of Equation \ref{equation no limit} converges to $k$ as $z$ converges to $(1,\ldots,1)$ from below.
First note that every element in $\Gamma(v)$ is contained in $ \mathbb R^d_{\geq \beta(v)} \setminus \mathbb R^d_{\geq {\alpha(v)}} $, and every element of $\Gamma(v)$ has multiplicity at most $k$.
Also note that $ \mathbb R^d_{\geq \beta(v)} \setminus \mathbb R^d_{\geq {\alpha(v)}} $
is contained in the set $\bigcup_{i=1}^d R_i(v)$, where $R_i(v)$ is the set
\[R_i(v):=\{(a_1, \ldots, a_d) \ : \ {\beta(v)}_i \leq a_i < {\alpha(v)}_i \text{ and } {\beta(v)}_j\leq a_j \text{ for all } j \neq i , \ 1\leq j \leq d \ \}.\]
Because $|z_i|<1$ for all $i$, we have the following closed-form expressions:
\begin{align}\label{equation stripes}
\sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}}} z^x= \prod_{j=1}^d \frac{z_j^{{\alpha(v)}_j}}{1-z_j} ; \
\sum_{x \in R_i(v) \cap \mathbb Z^d}z^x =(1-z_i^{{\alpha(v)}_i-\beta(v)_i}) \cdot \prod_{j=1}^{d}\frac{z_j^{{\beta(v)}_j}}{1-z_j}.
\end{align}
Because $\Gamma(v)$ is contained in $\bigcup_{i=1}^{d} R_i(v)$ and each element in $\Gamma(v)$ has multiplicity at most $k$, we conclude that:
\begin{align}\label{equation gamma}
\lim_{z \to (1,\ldots,1)^-} \sum_{x \in \Gamma(v) \cap \mathbb Z^d}z^x \cdot \prod_{j=1}^d (1-z_j)& \leq \lim_{z \to (1,\ldots,1)^-} \sum_{i=1}^{d} \sum_{x \in (R_i(v) \cap \mathbb Z^d)}kz^x \cdot \prod_{j=1}^d (1-z_j) \notag \\
&= \lim_{z \to (1,\ldots,1)^-} \sum_{i=1}^{d} k (1-z_i^{{\alpha(v)}_i-\beta(v)_i}) z^{{\beta(v)}}=0.
\end{align}
Combining
Equation \ref{equation stripes} and Equation \ref{equation gamma}, we get:
\begin{equation}\label{right hand side}
\lim_{z \to (1,\ldots,1)^-} \sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}} } kz^x \cdot \prod_{j=1}^d (1-z_j) + \sum_{x \in \Gamma(v) \cap \mathbb Z^d}z^x \cdot \prod_{j=1}^d (1-z_j) =k,
\end{equation}
which shows that the left side of Equation \ref{equation no limit} converges to $k$ as $z$ converges to $(1,\ldots,1)$ from below.
Note that
if $ \lim_{z \to (1,\ldots,1)^-} s_i(z)$ exists for all $i$, then
taking the limit of Equation \ref{equation no limit} as $z$
converges to $(1,\ldots,1)$ from below will give us
Equation \ref{equation in props asymptotic}, and the proof will be done.
However, the limit of $s_i(z)$ does not always exist, and hence we need a more subtle approach to derive Equation \ref{equation in props asymptotic}.
Since $|z_i|<1$ for all $i$, we have that $s_i(z)$ is a positive real number for all $i$.
Also note that
the multiplicity of every element in $\Lambda_i$ can not exceed $k$,
and every element of $\Lambda_i-a_i$ is contained in $ \mathbb Z^d_{\geq -a_i}$.
These facts allow us to derive the following inequality
\begin{align}\label{equation s is bounded above}
s_i(z)= \sum_{y\in \Lambda_i -a_i} z^y \cdot \prod_{j=1}^d (1-z_j)
&\leq \sum_{y \in \mathbb Z^{d}_{\geq -a_i}} k z^y \cdot \prod_{j=1}^d (1-z_j) \notag \\
&= k z^{[-a_i]} \cdot \prod_{j=1}^{d} \frac{1}{1-z_j} \cdot \prod_{j=1}^d (1-z_j) \notag\\
&= k z^{[-a_i]},
\end{align}
where $[x]$ is the integer part of the real vector $x$ in $ \mathbb R^d$.
Note that as $z$ converges to $(1,\ldots,1)$ from below, the right side of Equation \ref{equation s is bounded above} is bounded from above by $k+1$.
Hence as $z$ converges to $(1,\ldots,1)$ from below, the value of $s_i(z)$ is bounded between $0$ and $k+1$.
The Bolzano-Weierstrass theorem \cite{BS92} then implies that there exists a sequence $(\mathfrak z_u)_{u \in \mathbb N}$ that converges to $(1,\ldots,1)$ from below and with the property that $\lim_{u \to \infty} s_i(\mathfrak z_u)$ exists for all $i$.
Let $g_i:=\lim_{u \to \infty} s_i(\mathfrak z_u)$ for all $i$, note that $g_i$ is non-negative because $s_i(\mathfrak z_u)$ is a positive real number for all $u \in \mathbb N$.
Also note that the definition of $s_i(z)$ does not involve $v$,
and hence each $g_i$ is a constant that is independent from the choice of $v$.
By substituting $\mathfrak{z}_u$ into Equation \ref{equation no limit} and then taking the limit as $u$ goes to infinity, we get the following equality:
\begin{align}
& \sum_{i=1}^{n} \# ( \mathbb Z^d \cap \{ P^{ \mathfrak{h}}+a_i+v \} ) \cdot \lim_{u \to \infty}s_i(\mathfrak z_u) \notag \\
= & \lim_{u \to \infty} \sum_{x \in \mathbb Z^d_{\geq {\alpha(v)}} } k \mathfrak z_u^x \cdot \prod_{j=1}^d (1- {\mathfrak z_u}_j) + \sum_{x \in \Gamma(v) \cap \mathbb Z^d} \mathfrak z_u^x \cdot \prod_{j=1}^d (1-{\mathfrak z_u}_j) \notag.
\end{align}
Substituting Equation \ref{right hand side} into the right side of the equation above, we get:
\begin{align}\label{equation last in the lemma}
\sum_{i=1}^{n} \# ( \mathbb Z^d \cap \{ P^{ \mathfrak{h}}+a_i+v \} ) \cdot g_i = & k. \end{align}
To get Equation \ref{equation in props asymptotic} from Equation \ref{equation last in the lemma}, note that
\begin{equation}\label{equation technical}
\# ( \mathbb Z^d \cap \{ P^{ \mathfrak{h}}+a_i+v \} ) =\# (-1 \cdot \mathbb Z^d \cap \{ -1 \cdot P^{ \mathfrak{h}} -a_i-v )\} )=L_{ \mathbb Z^d }^{ \mathfrak{h}}(-a_i-v).
\end{equation}
Substituting Equation \ref{equation technical} into the left side of Equation \ref{equation last in the lemma}, we get:
\begin{align}\label{equation final in the lemma}
\sum_{i=1}^{n} L_{\Lambda}^{ \mathfrak{h}}(-a_i-v) \cdot g_i = & k.
\end{align}
As the choice of $v \in \mathbb R^d$ is arbitrary, we can replace $v$ in Equation \ref{equation final in the lemma} by $-v$ to get
Equation \ref{equation in props asymptotic} and the proof is now complete.
\qed \end{proof}
\begin{remark}
The proof of Lemma \ref{asymptotic method} can be made much shorter if we use the stronger assumption that $\Lambda$ is equal to the disjoint union of finitely many translates of one lattice.
Indeed, in this case, we have:
\begin{align*}
k=\# (\Lambda \cap \{-1 \cdot P^{ \mathfrak{h}} + v\}) &= \sum_{i=1}^n \# (
\{a_i+ \mathcal{L}
\} \cap \{-1 \cdot P^{ \mathfrak{h}} + v\}) \\&= \sum_{i=1}^n \# (\mathcal{L} \cap \{ -1 \cdot P^{ \mathfrak{h}} + v-a_i\})=\sum_{i=1}^{n} L_{\mathcal{L}}^{ \mathfrak{h}}(v-a_i), \notag
\end{align*}
for all $v$ in $ \mathbb R^d$.
\end{remark}
\begin{remark}
The proof of Lemma \ref{asymptotic method} no longer works if we use the weaker assumption that every element of $\Lambda$ is contained in a quasi-periodic set, because the original assumption is essential for deriving Equation \ref{equation intermediate} from Equation \ref{equation start}.
\end{remark}
Now we show that
the value $g_1,\ldots, g_n$ in Lemma \ref{asymptotic method} can, in fact, be chosen to be rational numbers with some value $m$ on the right side of Equation \ref{equation in props asymptotic}, which gives us the following theorem.
\begin{theorem}\label{main theorem 3}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
Let $\mathcal{L}$ be a lattice and let $a_1,\ldots, a_n$ be vectors in $ \mathbb R^d$.
If every element of $\Lambda$ is contained in the
finite union $\bigcup_{i=1}^{n} a_i+\mathcal{L}$,
then $P$ $m$-tiles in $ \mathbb R^d$ for some $m$ with a finite union of copies of the lattices $a_1+\mathcal{L}$, $, \ldots,$ $a_n+\mathcal{L}$.
\end{theorem}
\begin{proof}
By Lemma \ref{asymptotic method}, there are non-negative real numbers $g_1,\ldots, g_n$ such that
\begin{equation}\label{equation vector space}
\sum_{i=1}^{n} g_i \cdot L_{\mathcal{L}}^{ \mathfrak{h}}(v-a_i)=k,
\end{equation}
for all vectors $v$ in $ \mathbb R^d$.
For an arbitrary $v$ and $w$ in $ \mathbb R^d$, let $l_i(v,w)$ be the integer
\[l_i(v,w):= L_{\mathcal{L}}^{h}(v-a_i) -L_{\mathcal{L}}^{h}(w-a_i),\]
and let $V$ be the vector space in $ \mathbb R^n$ spanned by the following set of vectors:
\[ \{ (l_1(v,w),l_2(v,w), \ldots, l_n(v,w) ) \ : \ v,w \in \mathbb R^d \}.\]
Note that by Equation \ref{equation vector space}, the vector
$ (g_1, \ldots, g_n)$ is contained in the orthogonal complement $V^\perp$ of $V$, and hence $V^\perp$ contains a non-zero non-negative vector.
Also note that $V$ are generated by integer vectors, and hence $V$ and $V^{\perp}$ have a basis of integer vectors.
These two facts imply that
there is a non-negative non-zero integer vector $(g_1', \ldots g_n')$ that is contained in $V^{\perp}$.
By the construction of the vector space $V$, the statement that $(g_1', \ldots g_n')$ is orthogonal to $V$ is equivalent to the following equation:
\begin{equation}
m=\sum_{i=1}^{n} g_i' \cdot \#(\mathcal{L} \cap \{-1\cdot P^{ \mathfrak{h}} +v-a_1\})= \sum_{i=1}^{n} g_i' \cdot \#(a_i+\mathcal{L} \cap \{-1\cdot P^{ \mathfrak{h}} +v\}),
\end{equation}
for some positive integer $m$ and for all $v$ in $ \mathbb R^d$. By Lemma \ref{lemma GRS} this implies that $P$ $m$-tiles $ \mathbb R^d$ with the union of the translated lattices $a_1+\mathcal{L}$, $, \ldots,$ $a_n+\mathcal{L}$, where each element of $a_i+\mathcal{L}$ has multiplicity $g_i'$.
\qed \end{proof}
In the case when all elements of $\Lambda$ are contained in a lattice, Theorem \ref{main theorem 3} gives us the following corollary:
\begin{repcorollary}{corollary one lattice}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
If every element of $\Lambda$ is contained in a lattice $\mathcal{L}$, then $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}$ for some $m$. \qed
\end{repcorollary}
We now present a technical lemma that allows us to reduce Theorem \ref{main theorem 1} to the situation where all elements of $\Lambda$ are contained in a lattice, so that we can apply Corollary \ref{corollary one lattice} to prove Theorem \ref{main theorem 1}.
\begin{lemma}\label{general lambda}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, and suppose that $\Lambda$ is the disjoint union of two discrete multisets $\Lambda_1$ and $\Lambda_2$.
If the set $ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P+\Lambda_2)$ is path-connected and $\Lambda_1$ is non-empty, then $P$ $m$-tiles $ \mathbb R^d$ with $\Lambda_1$ for some $m$.
\end{lemma}
\begin{proof}
By Lemma \ref{lemma GRS}, the fact that $P$ $k$-tiles $ \mathbb R^d$ with $\Lambda=\Lambda_1 \sqcup \Lambda_2$ implies that
\begin{equation}\label{equation general lambda}
L_{\Lambda_1}^{ \mathfrak{h}}(v)+ L_{\Lambda_2}^{ \mathfrak{h}}(v)= L_{\Lambda}^{ \mathfrak{h}}(v)=k,
\end{equation}
for all $v$ in $ \mathbb R^d$.
Let $v_1$ and $v_2$ be two points in $ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P+\Lambda_2)$.
Because $ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P+\Lambda_2)$ is path-connected, there is a path $\mathfrak P: [0,1] \to \mathbb R^d$ starting at $v_1$ and ending at $v_2$ such that $\mathfrak P$ does not contain points from $(\partial P+ \Lambda_1) \cap (\partial P+\Lambda_2)$.
We claim that the function $L_{\Lambda_1}^{ \mathfrak{h}}(\mathfrak P(x))$ remains constant $x$ goes from $0$ to $1$.
Suppose to the contrary that $L_{\Lambda_1}^{ \mathfrak{h}}(\mathfrak P(x))$ is not a constant function.
This means that there is $\alpha \in [0,1]$ such that the function $L_{\Lambda_1}^{ \mathfrak{h}}(\mathfrak P(x))$ is not constant in every open neighborhood of $\alpha$.
Because $\mathfrak P (\alpha)$ is not contained in $(\partial P+ \Lambda_1) \cap (\partial P+ \Lambda_2)$, either one of the following scenarios will hold:
\begin{itemize}[leftmargin=*]
\item $\mathfrak P (\alpha) $ is not contained in $\partial P+\Lambda_1$. This means that $\mathfrak P (\alpha)$ is in general position with respect to $\Lambda_1$. By Property \ref{property general position} in Section \ref{section lattice point enumeration}, the function $L_{\Lambda_1}^{ \mathfrak{h}}$ is constant in a sufficiently small neighborhood of $\mathfrak P(\alpha)$, contradicting the assumption on $\alpha$.
\item $\mathfrak P (\alpha) $ is not contained in $\partial P+\Lambda_2$. This means that $\mathfrak P (\alpha)$ is in general position with respect to $\Lambda_2$.
By Property \ref{property general position} in Section \ref{section lattice point enumeration}, the function $L_{\Lambda_2}^{ \mathfrak{h}}$ is constant in a sufficiently small neighborhood of $\mathfrak P(\alpha)$.
By Equation \ref{equation general lambda} we have
$L_{\Lambda_1}^{ \mathfrak{h}}= k- L_{\Lambda_2}^{ \mathfrak{h}}$, and hence $L_{\Lambda_1}^{ \mathfrak{h}}$ is also a constant function in a sufficiently small neighborhood of $\mathfrak P(\alpha)$, contradicting to the assumption on $\alpha$.
\end{itemize}
Hence $L_{\Lambda_1}^{ \mathfrak{h}}(v)$ has a constant value $m$ for all $v$ in $ \mathbb R^d\setminus (\partial P+ \Lambda_1) \cap (\partial P+ \Lambda_2)$.
We now show that $m$ is a positive integer.
Because $\Lambda_1$ is non-empty, we have $L_{\Lambda}^{ \mathfrak{h}}(v_1)= \# (\Lambda \cap \{-1 \cdot P^{ \mathfrak{h}}+ v_1 \})$ is positive for some $v_1$ in an open set $B$ of $ \mathbb R^d$.
Because the set
$ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P +\Lambda_2)$ is dense in $ \mathbb R^d$, there exists $v_2 \in \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P +\Lambda_2)$ that is also contained in $B$.
This implies that $m=L_{\Lambda_1}^{ \mathfrak{h}}(v_2)=L_{\Lambda_1}^{ \mathfrak{h}}(v_1)>0$, and hence $m$ is a positive integer.
Now let $v \in \mathbb R^d$ be a vector in general position with respect to $\Lambda_1$.
By Property \ref{property general position} in Section \ref{section lattice point enumeration},
the function $L_{\Lambda_1}^{ \mathfrak{h}}(v)$ is a constant function in an open neighborhood $B_v$ of $v$ in $ \mathbb R^d$.
Because the set
$ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P +\Lambda_2)$ is dense in $ \mathbb R^d$, there exists $w$ in $ \mathbb R^d \setminus (\partial P+ \Lambda_1) \cap (\partial P +\Lambda_2)$ that is contained in $B_v$.
By the argument above, this implies that $L_{\Lambda_1}^{ \mathfrak{h}}(v)= L_{\Lambda_1}^{ \mathfrak{h}}(w)=m$.
Because the choice of $v$ is arbitrary, this implies that $L_{\Lambda_1}^{ \mathfrak{h}}(v)=m$ for every $v \in \mathbb R^d$ that is in general position with respect to $\Lambda_1$.
By Lemma \ref{lemma GRS}, we conclude that $P$ $m$-tiles $ \mathbb R^d$ with $\Lambda_1$.
\qed \end{proof}
We now proceed to prove Theorem \ref{main theorem 1}.
\begin{reptheorem}{main theorem 1}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, and suppose that every element of $\Lambda$ is contained in a quasi-periodic set $\mathcal{Q}$. If a lattice $\mathcal{L}$ in $\mathcal{Q}$ is in general position with respect to $\mathcal{Q}$ and $\mathcal{L} \cap \Lambda$ is non-empty, then $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}$ for some $m$.
\end{reptheorem}
\begin{proof}
Without loss of generality, we can assume that $\mathcal{L}_i$ is a lattice instead of a translate of a lattice.
Let $\Lambda_1=\Lambda \cap \mathcal{L}_i$ and $\Lambda_2= \Lambda \setminus \Lambda_1$.
We have that $\Lambda$ is a disjoint union of $\Lambda_1$ and $\Lambda_2$, and by the distributive law the set $H=(\partial P+\Lambda_1) \cap \partial (P+\Lambda_2)$ is contained in $H_i$ (where $H_i$ is as defined in Equation \ref{technical condition}).
Because $ \mathbb R^d \setminus H_i$ is path-connected by the assumption that $\mathcal{L}$ is in general position with respect to $\mathcal{Q}$, this implies that $ \mathbb R^d \setminus H$ is path-connected.
Also note that by assumption $\Lambda_1$ is a non-empty multiset.
Hence by Lemma \ref{general lambda} $P$ $m'$-tiles $ \mathbb R^d$ with $\Lambda_1$ for some $m'$.
Because every element of $\Lambda_1$ is contained in $\mathcal{L}_i$, we conclude that $P$ $m$-tiles $ \mathbb R^d$ with $\mathcal{L}_i$ for some $m$ by Corollary \ref{corollary one lattice} .
\qed \end{proof}
Note that the assumption that the lattice in Theorem \ref{main theorem 1} is in general position can not be omitted from the statement of the theorem, as seen in Example \ref{example rectangle} below.
\begin{example}\label{example rectangle}
Let $P$ be a rectangle in $ \mathbb R^2$ with $(0,0) , (0,\frac{1}{2}), (1,0), (1, \frac{1}{2})$ as vertices.
Let $\mathcal{L}_1$ be the lattice $ \mathbb Z^2$, let $\mathcal{L}_2$ be the translated lattice $( \frac{\sqrt 2}{2},\frac{1}{2})+ \mathbb Z^2$, and let $\mathcal{Q}=\mathcal{L}_1 \cup \mathcal{L}_2$.
It can be seen from Figure \ref{figure tiling} that $P$ $1$-tiles $ \mathbb R^2$ with $\mathcal{Q}$, but
$P$ does not $m$-tile $ \mathbb R^2$ with $\mathcal{L}_1$ or $\mathcal{L}_2$ for any $m$.
Also notice that $\mathcal{L}_1$ and $\mathcal{L}_2$ are not in general position to $\mathcal{Q}$, as the sets
$ \mathbb R^2 \setminus H_1= \mathbb R^2 \setminus H_2= \mathbb R^2 \setminus \{(x,y) \in \mathbb R^2 \ : \ x \in \mathbb Z \}$ are not path-connected.
\end{example}
\begin{figure}[ht]\label{figure tiling}
\caption{A rectangle $P$ that 1-tiles $ \mathbb R^2$ with $\mathcal{Q}=\mathcal{L}_1 \cup \mathcal{L}_2$, but does not $m$-tile with $\mathcal{L}_1$ or $\mathcal{L}_2$.}
\centering
\includegraphics{tiling}
\end{figure}
\section{Quasi-periodic tiling without Hypothesis \ref{technical condition}} \label{section for main theorem 2}
In this section, we discuss quasi-periodic tiling in the situation when Hypothesis \ref{technical condition} is omitted.
As we observed from Example \ref{example rectangle}, the condition that a lattice $\mathcal{L}$ is in general position with respect to the quasi-periodic set $\mathcal{Q}$ can not be dropped.
However, this does not
preclude the possibility that $P$ $m$-tiles $ \mathbb R^d$ for some $m$ with some lattice $\mathcal{L}$ that is not contained in $\mathcal{Q}$.
For example, the rectangle $P$ in Example \ref{example rectangle}
can $2$-tile $ \mathbb R^2$ with the lattice $\frac{1}{2} \mathbb Z\times \frac{1}{2} \mathbb Z$, even though $\frac{1}{2} \mathbb Z\times \frac{1}{2} \mathbb Z$ is not contained in $\mathcal{Q}$.
In the next theorem we present an approach to construct such a lattice $\mathcal{L}$ for the case when $\mathcal{Q}$ is a union of two translated copies of a lattice.
\begin{reptheorem}{main theorem 2}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$.
Let $\mathcal{L}_1$ and $\mathcal{L}_2$ be translates of one single lattice in $ \mathbb R^d$.
If every element in $\Lambda$ is contained in $\mathcal{L}_1 \cup \mathcal{L}_2$, then $P$ $m$-tiles $ \mathbb R^d$ with some lattice $\mathcal{L}$ for some $m$.
\end{reptheorem}
\begin{proof}
Without loss of generality, we can assume that $\mathcal{L}_1= \mathbb Z^d$ and $\mathcal{L}_2=a+ \mathbb Z^d$ for some $a$ in $ \mathbb R^d$.
Suppose that $a$ is a rational vector, let $N$
be the least common multiple of the denominator of entries of $a$.
Note that both $ \mathbb Z^d$ and $a+ \mathbb Z^d$ are now contained
$(\frac{1}{N} \mathbb Z)^d$, and by Corollary \ref{corollary one lattice} we have $P$ $m$-tiles $ \mathbb R^d$ with $(\frac{1}{N} \mathbb Z)^d$ for some $m$.
Hence we can assume that $a$ is not a rational vector.
By permuting the coordinates, we can assume that $a=(\alpha_1, \alpha_2, \ldots, \alpha_k, \beta_{k+1}, \ldots, \beta_d)$, where $\alpha_1, \ldots, \alpha_k$ are irrational numbers linearly independent over $ \mathbb Q$, and $\beta_{k+1}, \ldots, \beta_d$ are contained in $\langle \alpha_1, \ldots, \alpha_k,1 \rangle_{ \mathbb Q}$.
Because $a$ is not a rational vector, we have $k\geq 1$.
Because $\beta_i$ is contained in $\langle \alpha_1, \ldots, \alpha_k,1 \rangle_{ \mathbb Q}$
for all $i \in \{k+1,\ldots, d\}$,
there exists $c_{i,j} \in \mathbb Q $ for ${j \in \{1,\ldots,k\}}$ such that $\beta_i=c_{i,1}\alpha_1 +\ldots+ c_{i,k} \alpha_k+c_{i,k+1}$.
Let $N$ be the least common multiple of the denominators of $c_{i,j}$, where $i \in \{k+1,\ldots,d\}$ and $j \in \{1,\ldots,k\}$.
Note that $\beta_i$ is contained in $\langle \alpha_1, \ldots, \alpha_k,1 \rangle_{(\frac{1}{N} \mathbb Z)^d}$ for all $i \in \{k+1,\ldots,d\}$.
In the rest of this proof, we will use $L^{ \mathfrak{h}} $ as a shorthand for $L_{(\frac{1}{N} \mathbb Z)^d}^{ \mathfrak{h}} $.
By Lemma \ref{asymptotic method}, we have the following equation:
\begin{equation}\label{equation cute}
g_1 \cdot L^{ \mathfrak{h}}(v)+ g_2 \cdot L^{ \mathfrak{h}}(v-a)=k,\end{equation}
for some non-negative real numbers $g_1$ and $g_2$ and for all $v$ in $ \mathbb R^d$.
If $g_1=0$, then $ L^{ \mathfrak{h}}(v-a)=\frac{k}{g_2}$ for all $v \in \mathbb R^d$.
By Lemma \ref{lemma GRS}, this implies that $P$ $m$-tiles $ \mathbb R^d$ with $(\frac{1}{N} \mathbb Z)^d$ for $m=\frac{k}{g_2}$ and the claim is proved.
By symmetry we get the same conclusion for when $g_2=0$.
Hence we can assume that both $g_1$ and $g_2$ are non-zero.
Let $L_j=L^{ \mathfrak{h}}(v-a\cdot j)-\frac{k}{g_1+g_2}$ for all $j \in \mathbb Z$.
Substituting $v$ in Equation \ref{equation cute} with $v-a\cdot j$, we get the following relation for all $j \in \mathbb Z$:
\begin{equation}\label{equation L_j and L_j+1}
L_jg_1+L_{j+1}g_2=0,
\end{equation}
and we can without loss of generality assume that $g_2 \leq g_1$.
First, suppose that $g_2<g_1$.
Note that by Equation \ref{equation L_j and L_j+1} we have $L_{0}=(-\frac{g_2}{g_1})^jL_j$ for all $j \in \mathbb Z$.
On the other hand, the function $L^ \mathfrak{h}$ is a periodic function by Property \ref{property periodic} in Section \ref{section lattice point enumeration}, which implies that $L_j(v)$ (which is equal to $L^{ \mathfrak{h}}(v-a)-\frac{k}{g_1+g_2}$) is a bounded function.
Because $L_j$ is bounded and $g_2<g_1$, we have
\[L_0=\lim_{j \to \infty} \left(-\frac{g_2}{g_1} \right)^jL_j= 0.\]
This implies that
$ L^{ \mathfrak{h}}(v)=\frac{k}{g_1+g_2}$ for all $v$ in $ \mathbb R^d$, and by Lemma \ref{lemma GRS} we have that $P$ $m$-tiles $ \mathbb R^d$ with $(\frac{1}{N} \mathbb Z)^d$ for $m= \frac{k}{g_1+g_2} $, and the claim is proved.
Now suppose that $g_1=g_2$.
We claim that
$L_{2j+1}=L_0$ for some $j \in \mathbb Z$.
Note that if the claim holds, then we can conclude that $L_0=L_1$ (because $L_{2j+1}=L_1$ and $L_{2j}=L_0$ for all $j \in \mathbb Z$ by Equation \ref{equation L_j and L_j+1}).
Because we also have $L_0+L_1=0$ by Equation \ref{equation L_j and L_j+1}, this implies that $L^{ \mathfrak{h}}(v)-\frac{k}{g_1+g_2}=L_0=0$.
By Lemma \ref{lemma GRS}, we can then conclude that $P$ $m$-tiles $ \mathbb R^d$ with $(\frac{1}{N} \mathbb Z)^d$ for $m= \frac{k}{g_1+g_2} $, and the claim is proved.
For any two points $w$ and $w'$ in $ \mathbb R^d$ and a constant $\epsilon>0$, we say that $w$ is $\epsilon$-close to $w'$ modulo $ \mathbb Z^d$ if $|w-w'+\lambda|<\epsilon$ for some $\lambda$ in $ \mathbb Z^d$.
Let $a' \in \mathbb R^k $ be the vector $(\alpha_1, \ldots, \alpha_k)$,
where $\alpha_1,\ldots, \alpha_k$ are irrational numbers defined in the beginning of the proof.
We claim that for any $\epsilon>0$, there exists $j\in \mathbb Z$ such that $(2j+1)\cdot a'$ is $\epsilon$-close to $0$ modulo $ \mathbb Z^k$.
To prove this claim,
we use a powerful tool from number theory called the Weyl criterion for the multidimensional case.
We say that a sequence $(x_n )_{n\in \mathbb N} $ of vectors in $ \mathbb R^k$ is dense modulo $ \mathbb Z^k$ in $ \mathbb R^k$ if for any point $w$ in $ \mathbb R^k$ and a constant $\epsilon>0$, there exists a natural number $j$ such that $w$ is $\epsilon$-close to $x_j$ modulo $ \mathbb Z^k$.
\begin{theorem} \label{Weyl Criterion} \textnormal{(weak form of Weyl criterion, \cite[Theorem 6.2]{Kuipers})}
Let $(x_n )_{n\in \mathbb N} $ be a sequence of vectors in $ \mathbb R^k$.
The sequence $(x_n )_{n\in \mathbb N} $ is dense modulo $ \mathbb Z^k$ in $ \mathbb R^k$ if for every lattice point $h\in \mathbb Z^k$, $h \neq 0$,
\[\lim_{M \to \infty} \frac{1}{M} \sum_{n=1}^{M}e^{2\pi i \langle h,x_n \rangle}=0. \] \qed
\end{theorem}
Let $(x_n )_{n\in \mathbb N} $ be the sequence defined by $x_n=2n \cdot a'$ for all $n \in \mathbb N$.
Because $\alpha_1, \ldots, \alpha_k$ are irrational numbers that are linearly independent over $ \mathbb Q$, we have that ${ \langle h,a'\rangle}$ is not equal to $0$ for all $h \in \mathbb Z^k$.
Hence the limit of the sum in the Weyl criterion is equal to
\[ \lim_{M \to \infty} \frac{1}{M} \sum_{n=1}^{M}e^{2\pi i \langle h,x_n \rangle}= \lim_{M \to \infty} \frac{1}{M} \sum_{n=1}^{M}e^{4n \pi i \langle h,a' \rangle}= \lim_{M \to \infty} \frac{1}{M} \cdot \frac{e^{4\pi i \langle h,a' \rangle}(1- e^{4M\pi i \langle h,a' \rangle})}{1-e^{4\pi i \langle h,a' \rangle}} = 0.\]
Hence the Weyl criterion implies that the sequence $(2n \cdot a')_{n \in \mathbb N}$ is dense modulo $ \mathbb Z^k$ in $ \mathbb R^k$.
In the particular case when $w=-a'$, we have that for any $\epsilon>0$, there exists $j \in \mathbb N$ such that $-a'$ is $\epsilon$-close to $2j\cdot a'$ modulo $ \mathbb Z^k$.
Hence we conclude that $(2j+1) \cdot a'$ is $\epsilon$-close to $0$ modulo $ \mathbb Z^k$.
Because
$(2j+1) \cdot a'$ is $\epsilon$-close to $0$ modulo $ \mathbb Z^k$,
this implies that
$(2j+1)\alpha_1, \ldots, (2j+1)\alpha_k$ are all $\epsilon$-close to an integer. Because $\beta_i$ is contained in $\langle \alpha_1, \ldots, \alpha_k,1 \rangle_{(\frac{1}{N} \mathbb Z)^d}$, this also implies that $(2j+1)\beta_i$ is $O(\epsilon)$-close to $\frac{1}{N} \mathbb Z$ for $i \in \{k+1,\ldots, d\}$.
Hence we conclude that we can find an odd number $2j+1$ such that $(2j+1)\cdot a$ is $O(\epsilon)$-close to $(\frac{1}{N} \mathbb Z)^d$.
We will now show that $L_{2j+1}=L_0$.
Because we assume that $v \in \mathbb R^d$ is in general position with respect to $(\frac{1}{N} \mathbb Z)^d$, by Property \ref{property general position} Section \ref{section lattice point enumeration} there is a sufficiently small open neighborhood $B_v$ of $v$ such that $L^{ \mathfrak{h}}$ is a constant function in $B_v$.
By our previous argument, for any $\epsilon>0$, there is an odd number $2j+1$ such that $(2j+1) \cdot a$ is $O(\epsilon)$-close to $(\frac{1}{N} \mathbb Z)^d$.
By choosing a sufficiently small $\epsilon$, we conclude that $v-(2j+1)\cdot a$ is contained in $B_v$ modulo $(\frac{1}{N} \mathbb Z)^d$.
Because the function $L^{ \mathfrak{h}}$ has period $(\frac{1}{N} \mathbb Z)^d$ (Property \ref{property periodic} Section \ref{section lattice point enumeration}),
this implies that $L_{2j+1}=L^{ \mathfrak{h}}(v-(2j+1)\cdot a) = L^{ \mathfrak{h}}(v)=L_0$, and the proof is complete.
\qed \end{proof}
\begin{remark}
The proof of Theorem \ref{main theorem 2} can not be altered in any way to show that $P$ $m$-tiles $ \mathbb R^d$ with $ \mathbb Z^d$ instead of $(\frac{1}{N} \mathbb Z)^d$.
This can be seen from Example \ref{example rectangle}, where the rectangle in the example does not $m$-tile $ \mathbb R^2$ with $ \mathbb Z^2$ for any $m$.
\end{remark}
\section{Future research}\label{conjectures}
We conclude this paper by discussing possible future research problems that may lead to a proof of Conjecture \ref{conjecture last boss}.
\begin{problem} \label{conjecture theorem 1.2}
Let $P$ be a convex polytope that $k$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$, and suppose that every element of $\Lambda$ is contained in a quasi-periodic set $\mathcal{Q}$. Prove or disprove that $P$ $m$-tiles $ \mathbb R^d$ with a lattice $\mathcal{L}$ for some $m$.
\end{problem}
Problem \ref{conjecture theorem 1.2} is a generalization of Theorem \ref{main theorem 1} by removing Hypothesis \ref{technical condition}.
A more specific question to ask is whether Problem \ref{conjecture theorem 1.2} has a positive answer when $\mathcal{Q}$ is a finite union of translates of a single lattice.
A positive answer to this specific problem is given by Theorem \ref{main theorem 2} in the case where $\mathcal{Q}$ is a union of two translates of a single lattice.
\begin{problem}\label{conjecture bridge}
Prove or disprove that if a convex polytope $P$ $k$-tiles $ \mathbb R^d$, then $P$ $m$-tiles $ \mathbb R^d$ with a discrete multiset $\Lambda$ that is contained in a quasi-periodic set $\mathcal{Q}$ for some $m$.
\end{problem}
For dimension 2 and 3, Problem \ref{conjecture bridge} was positively answered by \cite{Kolountzakis} and \cite{GKRS} respectively.
This problem is open for dimensions higher than 3.
In particular,
a positive answer to both Problem \ref{conjecture theorem 1.2} and Problem \ref{conjecture bridge} will imply that Conjecture \ref{conjecture last boss} is true.
\section*{Acknowledgement}
The author would like to thank Sinai Robins for his advice and helpful discussions during the preparation of this paper and for introducing the author to this topic.
The author would like to thank the anonymous reviewers
for their valuable comments and suggestions to improve the
quality of the paper.
Lastly, the author would like to thank Henk Hollmann and Thomas Gavin for proofreading this paper.
\bibliographystyle{alpha}
|
\section{Supplemental Material: Mobile magnetic impurities in a Fermi superfluid: a route to designer molecules}
\begin{center}\emph{Sarang Gopalakrishnan, Colin V. Parker, and Eugene Demler}\end{center}
In what follows, we introduce two self-consistent approaches to computing the bound state energy and the impurity effective mass. First, we discuss an approach that is tailored for light impurities in three (or two) dimensions, and yields good results in both the weak and strong coupling limits; next, we turn to the one-dimensional case and discuss an approach that is reliable across the parity-changing transition whenever the impurity is sufficiently massive. Together, these approaches support the physical argument in the main text that the parity-changing transition should not be accompanied by any divergences, as it is strongly first-order.
\section*{Higher-dimensional T-matrix approach (light impurity)}
\subsection{Estimate of the parity transition}
In this section we begin with the Hamiltonian [Eq. (1) of main text] and compute the impurity-fermion T-matrix treating the BCS coherence factors [Eq.~(2) of main text] \emph{exactly} rather than approximately. This allows us to estimate the transition between the phase in which the molecule is an excited state and that in which the molecule is the ground state.
\begin{figure}[b]
\begin{center}
\includegraphics{tmatro-short.pdf}
\caption{Upper panel: equations for the self-consistent impurity propagator and for the T-matrix in the ladder approximation. Lower panel: illustration of a typical higher-order diagram of the kind resummed by our procedure.}
\label{tmatro}
\end{center}
\end{figure}
In general the T-matrix for impurity-quasiparticle scattering is given in the ladder approximation by the equation
\begin{equation}\label{ladder1}
\hat{T}(\omega) = \hat{J} + \hat{J} \left[ \int d^d k d\Omega \hat{G}(\mathbf{k}, \Omega) D(-\mathbf{k}, \omega - \Omega) \right] \hat{T}(\omega)
\end{equation}
where the hats denote Nambu matrix structure; matrix multiplication is implied; $\hat{J} \equiv J \mathbb{I}$ is the interaction; $\hat{G}$ is the bare fermion (i.e., $c^\dagger c$) propagator; and $D$ is the impurity propagator,
\begin{equation}\label{impprop}
\hat{G}(\omega, \mathbf{k}) = \frac{1}{\omega^2 - (\Delta^2 + v_F^2 k^2)} \left(\begin{array}{cc} \omega - v_F (q - k_F) & \Delta \\ \Delta & \omega + v_F (q - k_F) \end{array} \right); \quad D(\omega, \mathbf{k}) = \frac{1}{\omega - k^2/(2M)}.
\end{equation}
To find the parity transition we need to find the poles in $\hat{T}(0)$. Once $\Omega$ is integrated over we have
\begin{equation}\label{suppT}
\int dq \, d^{d-1} k \, \frac{1}{\sqrt{\Delta^2 + v_F^2 q^2} + k^2/(2M)} \left(\begin{array}{cc} \sqrt{\Delta^2 + v_F^2 q^2} - v_F (q - k_F) & \Delta \\ \Delta & \sqrt{\Delta^2 + v_F^2 q^2} + v_F (q - k_F) \end{array} \right).
\end{equation}
where $q$ is the direction normal to the Fermi surface and $k$ denotes all other directions. In this expression we have assumed that $v_F^2/\Delta \gg 1/M$, i.e., we neglect the change in curvature of the Bogoliubov dispersion minimum. This inequality is equivalent to requiring $E_F/\Delta \gg m/M$, which is always satisfied in the BCS limit.
We extract the $M$-dependence from this integral by substituting the variable $s = k/\sqrt{M}$. One sees then that the integral is given by $M^{(d-1)/2} \times F(\Delta, v_F)$, where the latter function depends exclusively on fermionic parameters and not on $M$. The remaining integral is formally divergent, but this divergence can be eliminated, and $F$ can be determined, if one requires the bound state energy to go smoothly to its infinite-mass limit. This yields the result in the main text.
\subsection{Effective mass at strong coupling}
We now turn to the effective mass $M^*$ of the impurity at strong coupling. We make two assumptions: (i)~that the superfluid is in the BCS limit, so that $E_F$ is much greater than any other energy scale in the problem; (ii)~that the impurity is sufficiently light that its renormalized recoil energy $\mathcal{E}^* \equiv 2 k_F^2/M^*$ remains greater than the coupling scale. Condition (ii) will be checked for self-consistency at the end of the calculation. When conditions (i) and (ii) are satisfied, the effective mass can be computed for arbitrary coupling if one replaces the bare interaction vertex with the T-matrix and resums the diagrams shown in Fig.~\ref{tmatro}.
These diagrams dominate because, in the light-impurity and BCS limits, the impurity can only emit and absorb pairs of quasiparticles with nearly opposite momenta. The diagrams in Fig.~\ref{tmatro}, in which successive each quasiparticle pair can lie anywhere on the dispersion minimum, are parametrically more important (by a factor $\sim \Delta/\mathcal{E}$) than diagrams in which the quasiparticle lines cross.
Specifically, the effective mass is given by the equation:
\begin{eqnarray}\label{ladder2}
\frac{1}{M^*(\hat{T})} & = & \frac{1}{M} - \frac{d^2}{dP^2} \int d\omega\, d\Omega \, d^3 k \, d^3 q\, D_{M^*}(-\Omega - \omega, \mathbf{P} - \mathbf{q} - \mathbf{k}) \nonumber \\ & & \quad \qquad \times \mathrm{Tr} \left[ \hat{G}(\omega, \mathbf{k})\hat{T}(\omega) \hat{G}(\Omega, \mathbf{q}) \hat{T}(\omega) \right].
\end{eqnarray}
where the impurity propagator on the right-hand side is to be computed using $M^*$. The equations~\eqref{ladder1}, \eqref{ladder2} together define a self-consistent theory incorporating the effects of the impurity recoil and the quasiparticle rearrangement on each other.
In this self-consistent approach, there are two ways for the impurity to emit a quasiparticle: either it can virtually occupy a Shiba state and a continuum state, or it can virtually occupy two continuum states. (There cannot be two Shiba states at once as at least one of the quasiparticles must have the wrong spin to form a Shiba state.) We first briefly consider processes in which an impurity creates a Shiba state and a continuum state. It is straightforward to see that, even in the extreme limit where the zero-quasiparticle state hybridizes resonantly with such a state \emph{and} the molecule is infinitely massive, this channel at most increases the effective mass by a factor of two. (This is the standard result for the hybridization between a dispersing band and a flat band.) As we shall see, the continuum channel is parametrically more important at strong coupling.
We now discuss the behavior of this continuum channel. From the discussion of the pair-creation diagram in the main text, we know that the dominant quasiparticle frequencies contributing to Eq.~\eqref{ladder2} are $\alt \Delta$. Therefore, we replace the T-matrix in this equation with its value at $\Delta$, i.e., the bottom of the two-particle continuum; this can be expressed in terms of an impurity-quasiparticle scattering length $a$~\cite{braaten}
\begin{equation}
T(k) \simeq \frac{1/M^*}{a^{-1} + i k}
\end{equation}
Crucially, in the regime of interest (i.e., that of strong coupling, where a midgap Shiba state is present), the scattering length is given by $a \sim 1/\sqrt{2 M^* (\Delta - E_b)}$. Thus, in terms of the bound state energy, the T-matrix element for pair creation is then given in the regime of interest by
\begin{equation}
T(\Delta) \simeq \frac{1}{2 \sqrt{\Delta - E_b} M^* \sqrt{\Delta/v_F^2}},
\end{equation}
where we have substituted in the anisotropic mass from the main text. We emphasize that this relation does \emph{not} rely on the ladder approximation for the T-matrix. Now, we evaluate Eq.~\eqref{ladder2}, to arrive at the expression
\begin{equation}
\frac{1}{M^*} = \frac{1}{M} - \frac{3}{\sqrt{2} \pi^4} T(\Delta)^2 m^{3/2} (M^*)^{1/2} \Delta \simeq \frac{1}{M} - \alpha \left(\frac{m}{M^*}\right)^{3/2} \frac{v_F^2}{(\Delta - E_b)},
\end{equation}
where $\alpha$ is a numerical constant of order unity. When $E_F$ is sufficiently large compared with the other scales, one finds that
\begin{equation}
M^* \simeq m (M v_F^2/(\Delta - E_b))^{2/3} \sim m \left(\frac{E_F^2}{\mathcal{E} \Delta}\right)^{2/3}
\end{equation}
For the theory to remain self-consistent, we require that $\Delta \ll k_F^2/(2M^*)$ (otherwise it would not be legitimate to use the light-impurity limit for the T-matrix calculation). One can easily see that this self-consistency condition is satisfied whenever $\mathcal{E}^2 \gg \sqrt{\Delta E_F}$, i.e., for relatively light impurities in the BCS limit.
This self-consistent approach relies on the physically reasonable assumption that successive pair-creation events are uncorrelated. This assumption is also made by various intermediate-coupling theories such as Ref.~\cite{leelowpines}; in contrast with such theories, however, we also include the ``vertex corrections'' (i.e., the T-matrix ladder diagrams) that are necessary to account correctly for the existence of the bound state. These corrections are much more important here than in the standard polaron problem because the impurity-fermion interaction is \emph{quadratic} rather than linear in the fermions.
\section*{One-dimensional T-matrix approach (heavy impurity)}
We now consider the case of a one-dimensional system. In this case, the dispersion minimum consists of two points rather than a surface. The resulting qualitative differences are: (i)~a midgap Shiba molecule exists for arbitrarily weak coupling, and has a binding energy $E_b \simeq - J^2 / (1/M + v_F^2/\Delta)$; (ii)~the perturbative effective-mass correction goes as $J^2/M$, and \emph{decreases} as the bare mass is increased [in contrast with the three-dimensional result Eqs.~\eqref{effmass1}, \eqref{effmass2}]. In one dimension, we can explicitly calculate (using a self-consistent method described below) the evolution of the bound-state energy and effective mass; as shown in Fig.~4 of the main text, the effective mass evolves smoothly across the parity-changing transition.
We now introduce the self-consistent T-matrix procedure. In effect, this method treats $J$ exactly and involves resummed perturbation theory in $1/M$, the inverse impurity mass. We begin by performing a standard, exact polaron transformation~\cite{leelowpines} on $\mathcal{H}$; it then takes the form
\begin{equation}
\mathcal{H}' = H_{\mathrm{BCS}} + \sum_{kk' \sigma} \frac{(V + J \sigma)}{\mathcal{V}} c^\dagger_{k} c_{k'} + \mathrm{h.c.} + \frac{1}{2M} \left(P_0 - \sum_{k \sigma} k c^\dagger_{k \sigma} c_{k \sigma} \right)^2
\end{equation}
In this equation, the first line is quadratic in the fermion operators, and corresponds to the solvable limit of a fixed ($M = \infty$) impurity. The second line includes a term that is \emph{quartic} in the fermions; this quartic term renormalizes both the dispersion and the scattering of the fermions, via the diagrams in Fig.~\ref{diagrams}. In general, the renormalized scattering will be $k$-dependent; however, we note that $k_F$ is much larger than the other momentum scales involved, which justifies approximating the total fermion momentum as the difference between the number of right- and left-movers, i.e.,
\begin{equation}\label{Lmovers}
\sum_{k \sigma} k c^\dagger_{k \sigma} c_{k \sigma} \rightarrow k_F \sum_{k, \sigma, \tau = \pm 1} \tau c^\dagger_{k \sigma \tau} c_{k \sigma \tau}.
\end{equation}
In this approximation, the T-matrix has a $2\times 2$ matrix structure in $(L,R)$ space but no other momentum-dependence.
We are now equipped to write out a self-consistent set of equations for the fermion Green's function and the T-matrix, shown diagrammatically in Fig.~\ref{diagrams}. As we would like to discuss the even-sector effective mass even when the even sector is not the ground state, we shall work within the Keldysh framework; thus, all the Green's functions in the procedure are the retarded component, except for the Green's function in the loop, which is the Keldysh component. (The Keldysh component of the Green's function is essentially a product of the spectral function and the distribution function; we choose the distribution function to be the ground-state Fermi function \emph{except} that we require the Shiba state to remain unfilled regardless of its energy.)
\begin{figure}[htbp]
\begin{center}
\includegraphics{tmatdiag.pdf}
\caption{Diagrammatic overview of the self-consistent T-matrix procedure. The circled X denotes the interaction $J$; the red square, the interaction $\sim 1/2M$; the thin lines are bare fermionic propagators; the bold lines are full fermionic propagators; and the other symbols are defined through the system of equations T1-T3. T1 shows how the impurity-fermion interaction vertex is renormalized, to lowest order in $1/M$, by impurity recoil. T2 accounts for repeated scattering events using this renormalized vertex by promoting it into a T-matrix, $T$. T3 constructs the full Green's function from the T-matrix via the standard relation~\cite{braaten} $G = G_0 + G_0 T G_0$.}
\label{diagrams}
\end{center}
\end{figure}
From these equations, it is straightforward to find the binding energy of the Shiba state (as one simply looks at the poles of the T-matrix in Eq. T2).
In addition, one can extract the effective mass from this diagrammatic system, using the following exact relations:
\begin{equation}
\frac{d E}{dP} = v = P/M^*; \quad \frac{d E}{dP} = \frac{1}{M} \left\langle P - \sum_k k c^\dagger_k c_k \right\rangle.
\end{equation}
Physically, this equation states that the group velocity of the polaron (viz. $P/M^*$) is on average the same as the velocity of the impurity, which is its momentum divided by its bare mass.
Once the impurity Green's function is computed from diagram T3 in Fig.~\ref{diagrams}, the expectation values on the right-hand side of this equation are known, and therefore it is straightforward to compute the effective mass.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width = 0.45\textwidth]{massdep.pdf}
\includegraphics[width = 0.45\textwidth]{jdep.pdf}
\caption{Dependence of the effective mass correction on the bare mass and on the coupling; there are small but noticeable deviations from the $J^2/M$-dependence predicted by perturbation theory.}
\label{massdep}
\end{center}
\end{figure}
Figs.~\ref{massdep} show the typical dependence of the effective-mass correction on the bare mass and on $J$. These results are remarkably close to those obtained via perturbation theory (as in the main text):
\begin{equation}\label{1dpert}
\frac{M^*}{M} - 1 \simeq \frac{J^2 \Delta}{M v_F^4} \nonumber
\end{equation}
\subsection*{Generalization of heavy-impurity approach to higher dimensions}
In principle the self-consistent approach discussed here can be generalized directly to higher dimensions. For simplicity we consider the case of two dimensions. The key step is to replace Eq.~\eqref{Lmovers} with the substitution
\begin{equation}
\sum_{\mathbf{k}} \mathbf{k} c^\dagger_{\mathbf{k}} c_{\mathbf{k}} \rightarrow k_F \sum_{k \theta} \cos(\theta) c^\dagger_{k\theta} c_{k\theta}.
\end{equation}
One can then proceed as above, except that the renormalized interaction becomes a continuous function of $\theta$ instead of just acquiring a $2 \times 2$ matrix structure. Note that this approximation also cures the ultraviolet divergences that would arise if one tried to work with the full theory.
We have not followed this route further in the present work because our analysis of the (computationally simpler) one-dimensional system suggests that the convergence of this approach is good only for $k_F^2/(M\Delta) \alt 1$, which is the opposite regime to that considered in the main text. However, we hope to adapt this approach to study the crossover between heavy and light impurities in future work.
|
\section{Introduction}
The Gauss-Bonnet (GB) dilaton gravity is known to be an active area of research in theoretical physics
through decades, which
was proposed to include the perturbative effects within effective theory based on the well known Einstein's
gravity at the two-loop level \cite{Berg:2005ja,Cicoli:2007xp,Bershadsky:1988mf,Kakushadze:1999bb,Green:1999pu,Roiban:2007jf}. For such theories the two-loop effective coupling signifies the strength of
the self-interaction between the spin 2 graviton degrees of
freedom below the Ultra-Violet (UV) cut-off of the quantum theory of gravity.
Usually such corrections originate naturally in string theory where power expansion in terms of inverse of Regge slope (or string tension) yields the
higher curvature corrections to pure Einstein’s gravity.
Supergravity, as the low energy limit
\cite{Nilles:1983ge,Lyth:1998xn,VanNieuwenhuizen:1981ae,Mazumdar:2010sa,Freedman:1976xh,Choudhury:2014sxa,Choudhury:2014uxa,Choudhury:2013jya,Choudhury:2013zna,Choudhury:2012ib,Choudhury:2011sq,Choudhury:2011rz}
of heterotic string
theory \cite{Green1,Green2,Polchinski1,Polchinski2,Evans:1986ada,Robb:1986de,Gross:1985rr,Candelas:1986tz,Cai:1986sa}, yields the Gauss-Bonnet (GB) term along with dilaton coupling at the leading order correction.
Consequently it became an active area of interest as a modified theory of gravity.
In the context of black hole it has been shown that GB correction suppresses graviton emission which makes
the black hole more stable.The correction to black hole entropy due to GB term has
also been explored.Moreover in search of extra dimensions, GB dilaton term in a warped braneworld
model has been studied in the context of first kaluza-klein graviton decay channel investigated
by ATLAS group in LHC experiments. Thus the Gauss-Bonnet dilaton gravity as a modified gravity theory has
been studied extensively in different contexts as a first step to include the higher curvature effects over
Einstein’ gravity.
Stability of the modulus in such models is an important issue from phenomenological point of view.
Goldberger and Wise (GW) \cite{Goldberger:1999uk,Goldberger:1999un,Goldberger:1999wh} first explicitly showed that the dynamics of a five dimensional
bulk scalar field in Randall Sundrum (RS) two brane setup can stabilize the size of the fifth (extra) dimension to a permissible value
to solve the gauge hierarchy problem. In this paper we examine such scenario in the context of higher curvature gravity,
where the usual Einstein's gravity is modified by the perturbative GB coupling and dilaton coupling.
In this theoretical prescription the stabilized effective potential for the bulk modulus is generated by
the presence of a bulk scalar field with quartic self interactions localized in two 3-branes.
This results in a modulus potential which after minimization
yields a compactification scale in terms of the VEV’s of the
scalar fields at the two branes. This concomitantly solves the gauge hierarchy problem without introducing any fine tuning
of the model parameters in the prescribed theoretical setup. Here we extend this study to include
higher curvature-dilaton term in the bulk space-time where we neglect the
effects of back reaction of the bulk scalar on the geometry as was done in case
of the original GW mechanism.Some critical studies have been made in this context \cite{DeWolfe:1999cp,Dey:2006px,
Das:2007mr,Das:2008uw,Maity:2006in}.
Broad aspects of the moduli stabilization mechanism in higher dimensions
\cite {Chacko:1999eb,Ponton:2001hq,Charmousis:2004zd, Burgess:2007vi},
specifically in the context of cosmological studies \cite {Cline:2000xn,Kanti:1999nz,Binetruy:2000wn,Ashcroft:2004rs}
from braneworlds i.e. inflation, dark energy and with
non minimal scalar fields coupled to the gravity sector have been reported in
\cite{Ghoroku:2001pi, Lewandowski:2001qp,Lesgourgues:2003mi,Kobayashi:2004aj,Eto:2004yk,Brevik:2004rt,Nunes:2005up,Ichinose:2004ys,Brummer:2005sh}.
The plan of this paper is as follows: In section \ref{l1} we study the framework of the modulus stabilization mechanism
in the context of GB dilaton gravity. First we propose the background model in higher curvature gravity from which we compute
the
the expression of the warp factor. Further using this warped solution we determine the analytical expression for the
stabilized potential for the bulk modulus field. To check the consistency of our
present analysis we
then study our setup in three distinct limiting situations namely in RS limit and limit when
either of GB coupling or dilaton coupling is present.
\section{Modulus Stabilization Mechanism in Gauss-Bonnet dilaton gravity}
\label{l1}
Here we generalize the analysis of modulus stabilization mechanism in warped geometry
in presence of Gauss-Bonnet coupling and gravidilaton coupling in a 5D bulk.
The background warped geometry model
is proposed by making use of the following sets of assumptions:
\begin{itemize}
\item The leading order Einstein's gravity sector is modified by the Gauss-Bonnet
\cite{Choudhury:2012yh,Choudhury:2012kw,Choudhury:2013qza,Choudhury:2013eoa,Choudhury:2013dia,Choudhury:2013yg,Kim:1999dq,Lee:2000vf,Kim:2000pz,Kim:2000ym}
and dilaton coupling \cite{Choudhury:2013eoa,Choudhury:2013dia,Choudhury:2013yg,Choudhury:2013aqa}
which originates from heterotic string theory.
\item The background warped metric has a RS like structure \cite{Randall:1999ee,Randall:1999vf} on a slice of ${\bf AdS_{5}}$ geometry.
For example,
from 10-dimensional string model compactified on ${\bf AdS_5\times S^5}$, one typically obtains
moduli from ${\bf S^5}$ as scalar degrees of freedom. Such moduli can be stabilized by fluxes.
In our model, which is similar to a 5-dimensional Randall-Sundrum (RS) model, it is assumed that these
degrees of freedom are frozen to their VEV and are non-dynamical at the energy scale under consideration \cite{Reece:2010xj}.
We therefore focus into the slice of ${\bf AdS_5}$ as is done for the 5-dimensional RS model.
\item The dilaton degrees of freedom is assumed to be confined within the bulk.
\item We allow the interaction between dilaton and the 5D bulk cosmological constant via dilaton coupling.
\item The Higgs field is localized at the visible (TeV) brane and the hierarchy problem is resolved via Planck to TeV scale warping.
\item Additionally while determining the values of the model parameters we require
that the bulk curvature to be less than the five dimensional Planck scale $M_{5}$ so that
the classical solution of the 5-dimensional gravitational equations can be trusted \cite{Davoudiasl:1999jd,Das:2013lqa}.
\end{itemize}
\subsection{The background setup}
\label{l1a}
We start our discussion with the following 5D action of the two brane warped geometry model \cite{Choudhury:2013yg}:
\begin{equation}\begin{array}{llll}\label{eq1}
\displaystyle S=\int d^{5}x \left[\sqrt{-g_{(5)}}\left\{\frac{M^{3}_{(5)}}{2}R_{(5)}+\frac{\alpha_{(5)}M_{(5)}}{2}
\left[R^{ABCD(5)}R^{(5)}_{ABCD}-4R^{AB(5)}R^{(5)}_{AB}+R^{2}_{(5)}\right]\right.\right.\\ \left.\left.
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~~+\frac{g^{AB}}{2}\partial_{A}\Phi\partial_{B}\Phi-\frac{m^{2}_{\Phi}}{2}\Phi^{2}
+\frac{g^{AB}}{2}\partial_{A}\chi(y)\partial_{B}\chi(y)-2\Lambda_{5}e^{\chi(y)}\right\}
\right.\\ \left. \displaystyle~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~-
\displaystyle \sum^{2}_{i=1}\sqrt{-g^{(i)}_{(5)}}\left[\lambda_{i}\left(\Phi^{2}-{\cal V}^{2}_{i}\right)^{2}+T_{i}\right]\delta(y-y_{i})\right]
\end{array}\end{equation}
with $A,B,C,D=0,1,2,3,4$.
Here $i$ signifies the brane index, $i=1(\text{hidden})$, $2(\text{visible})$.
${\cal V}_{i}$ and $\lambda_{i}$ signifies the VEV and self coupling of the bulk scalar fields on the {\it ith} brane where
$T_{i}$ is the brane tension and
$\Phi$ represent the bulk scalar degrees of freedom. Additionally $\alpha_{5}$ and $\chi(y)$ represent the
GB coupling and dilaton field.
The background metric describing slice of the ${\bf AdS_{5}}$ is given by,
\begin{equation}\begin{array}{llllll}\label{eq2}
\displaystyle ds^{2}_{5}=g_{AB}dx^{A}dx^{B}=e^{-2A(y)}\eta_{\alpha\beta}dx^{\alpha}dx^{\beta}+r^{2}_{c}dy^{2}
\end{array}\end{equation}
where $r_{c}$ represents the compactification radius of extra dimension.
Here the orbifold points are $y_{i}=[0,\pi]$
and periodic boundary condition is imposed in the closed interval $-\pi\leq y\leq\pi$. After orbifolding, the size of the extra dimensional
interval is $\pi r_{c}$. Moreover in the above metric ansatz
$e^{-2A(y)}$ represents the warp factor while $\eta_{\alpha\beta}=(-1,+1,+1,+1)$ is flat Minkowski metric.
A more general brane metric for a purely Einsteinian bulk has been discussed in \cite{Koley:2010za}.
\subsection{Warp factor}
\label{l1b}
After varying the model action stated in equation(\ref{eq1}) with respect to the background metric
the 5D bulk equation of motion turns out to be,
\begin{equation}\begin{array}{lllll}\label{eq3}
\displaystyle \sqrt{-g_{(5)}}\left[G^{(5)}_{AB}+\frac{\alpha_{(5)}}{M^{2}_{(5)}}H^{(5)}_{AB}\right]
=-\frac{e^{\chi(y)}}{M^{3}_{(5)}}\left[\Lambda_{(5)} \sqrt{-g_{(5)}}g^{(5)}_{AB}+
\sum^{2}_{i=1}T_{i}\sqrt{-g^{(i)}_{(5)}}g^{(i)}_{\alpha\beta}\delta^{\alpha}_{A}\delta^{\beta}_{B}\delta(y-y_{i})\right]
\end{array}\end{equation}
where the five dimensional Einstein's tensor and the Gauss-Bonnet tensor are given by
\begin{equation}\begin{array}{llll}\label{eq4}
G^{(5)}_{AB}=\left[R^{(5)}_{AB}-\frac{1}{2}g^{(5)}_{AB}R_{(5)}\right],
\end{array}\end{equation}
and
\begin{equation}\begin{array}{llll}\label{eq5}
H^{(5)}_{AB}=2R^{(5)}_{ACDE}R_{B}^{CDE(5)}-4R_{ACBD}^{(5)}R^{CD(5)}
-4R_{AC}^{(5)}R_{B}^{C(5)}+2R^{(5)}R_{AB}^{(5)}\\ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~-\frac{1}{2}g^{(5)}_{AB}
\left(R^{ABCD(5)}R^{(5)}_{ABCD}-4R^{AB(5)}R^{(5)}_{AB}+R^{2}_{(5)}\right).
\end{array}\end{equation}
Similarly varying equation(\ref{eq1}) with respect to the dilaton field the gravidilaton equation of motion turns out to be
\begin{equation}\begin{array}{llll}\label{eq6}
\displaystyle \frac{1}{M^{2}_{(5)}}\sum^{2}_{i=1}T_{i}\sqrt{-g^{(i)}_{(5)}}e^{\chi(y)}\delta(y-y_{i})
=\sqrt{-g_{(5)}}\left\{
\displaystyle
2\frac{\Lambda_{(5)}}{M^{2}_{(5)}}e^{\chi(y)}+\frac{\Box_{(5)} \chi}{M_{(5)}}\right\}
\end{array}\end{equation}
where the five dimensional D'Alembertian operator is defined as:
\begin{equation}\label{eq7} \Box_{(5)}\chi(y)=\frac{1}{\sqrt{-g_{(5)}}}\partial_{A}\left(\sqrt{-g_{(5)}}\partial^{A}\chi(y)\right).\end{equation}
Now using the ${\bf Z_{2}}$ orbifolding, we obtain
at the leading order of $\alpha_{(5)}$ \cite{Choudhury:2013yg}:
\begin{equation}\begin{array}{llll}\label{gradilatonic}
\displaystyle \chi(y)=\left(c_{1}|y|+c_{2} \right)
\end{array}\end{equation}
where $c_{1}$ and $c_{2}$ are arbitrary integration constants in which $c_{1}$ characterizes the strength of the dilaton self interaction within the bulk.
The corresponding warp factor turns out to be \cite{Choudhury:2013yg}:,
\begin{equation}\begin{array}{llll}\label{warp}
\displaystyle A(y):= A_{\pm}(y)=k_{\pm}(y)r_{c}|y|\end{array}\end{equation}
where
\begin{equation}\begin{array}{llll}\label{wsol}
\displaystyle k_{\pm}(y)=\sqrt{\frac{3M^{2}_{(5)}}{16\alpha_{(5)}}
\left[1\pm\sqrt{\left(1+\frac{4\alpha_{(5)}\Lambda_{5}e^{\chi(y)}}{9M^{5}_{(5)}}\right)}\right]}.\end{array}\end{equation}
In the small $\alpha_{(5)}$, $c_{1}$ and $c_{2}$ limit we retrieve the results as in the case of RS model with:
\begin{equation} \label{rs}
k_{-}(y)\rightarrow k_{RS}=\sqrt{-\frac{\Lambda_{5}}{24M^{3}_{(5)}}}.\end{equation}
Here we have discarded
the +ve branch of solution of $k_{+}$ which diverges in the small $\alpha_{(5)}$ limit, bringing
in ghost fields \cite{Rizzo:2004rq,Dotti:2007az,Torii:2005xu,Konoplya:2010vz,Kim:2000ym,Nojiri:2010wj}.
Now expanding Eq~(\ref{wsol}) in the perturbation series order by order around
$\alpha_{5}\rightarrow 0$, $c_{1}\rightarrow 0$ and $c_{2}\rightarrow 0$ we can write:
\begin{equation}\begin{array}{lll}\label{wsol1}
\displaystyle k_{\bf M}(y):=k_{-}(y)=k_{RS}~e^{\frac{\chi(y)}{2}}\left[1+\frac{4\alpha_{(5)} k^{2}_{RS}}{M^{2}_{(5)}}
+{\cal O}\left(\frac{\alpha^{2}_{(5)}k^{4}_{RS}}{M^{4}_{(5)}}\right)+\cdots\right].
\end{array}\end{equation}
\subsection{Stabilized potential for the modulus field}
\label{l1c}
Here we start with the background model action stated in Eq~(\ref{eq1}).
After varying the Eq~(\ref{eq1}) with respect to the scalar field $\Phi$ we get the
following equation of motion:
\begin{equation}\begin{array}{llll}\label{eom1}
\displaystyle -\frac{1}{r^{2}_{c}}\partial_{y}\left(e^{-4k_{\bf M}(y)r_{c}|y|}\partial_{y}\Phi\right)+m^{2}_{\Phi}e^{-4k_{\bf M}(y)r_{c}|y|}\Phi
+\frac{4}{r_{c}}\sum^{2}_{i=1}e^{-k_{\bf M}(y)r_{c}|y|}\lambda_{i}\Phi\left(\Phi^{2}-{\cal V}^{2}_{i}\right)\delta(y-y_{i})=0
\end{array}
\end{equation}
which clearly shows that the equation of motion changes from its RS counterpart
due an additional coordinate dependence of the function via the dilaton field $\chi(y)$ in $k_{\bf M}(y)$.
For convenience we introduce a set of parameters as:
\begin{equation}\begin{array}{llll}\label{parameter}
\displaystyle L=\frac{4 \alpha_{(5)} k^{2}_{RS}}{M^{2}_{(5)}},~~~
\displaystyle G=m_{\Phi}^2 r_{c}^2=M_{1}r_{c}\\
\displaystyle S= 4 k_{RS} c_{1} r_{c},~~~
\displaystyle Q=4 k_{RS} r_{c},\\
\displaystyle Z_{L}=(1+L+{\cal O}(L^2)).
\end{array}
\end{equation}
Further using Eq.~(\ref{parameter}) in Eq~(\ref{wsol1}) one can re-express the warp function $k_{\bf M}(y)$ as:
\begin{equation}\label{modified}
k_{\bf M}(y)=k_{RS}~e^{\frac{c_{1}|y|}{2}} Z_{L}.
\end{equation}
Now solving the Eq~(\ref{eom1}) we obtain the
solution for the bulk scalar field as,
\begin{equation}\begin{array}{llll}\label{soln}
\Phi(y)= A_{1} H_{-A} (B+ BSy) + B_{1}~ {}_1 F_1[\frac{A}{2}, \frac{1}{2},(BQ+BSy)^2] \\
\end{array}
\end{equation}
where,
\begin{equation}\begin{array}{llll}\label{cont}
\displaystyle A=\frac{G}{Z_{L}S},~~~~
B=\frac{\sqrt{Z_{L}}}{\sqrt{2 S}} .
\end{array}
\end{equation}
Here ${}_1 F_1$ represents the hypergeometric function of first kind and $H_{-A}$ represents
the Hermite function. Also $A_{1}$ and $B_{1}$ are the arbitrary integration constants which can be evaluated by using
appropriate boundary conditions at the locations of the branes in the prescribed two brane setup.
Since in the perturbative regime of the warping solution the GB coupling $\alpha_{(5)}$ and dilaton coupling $c_1$ is usually small, hence we
can expand the above solution in a series form and retain upto second order terms which enables us to
recast the solution for the bulk scalar field stated in Eq~(\ref{soln}) as,
\begin{equation}\begin{array}{llll} \label{soln1}
\displaystyle \Phi(y)= A_1 \frac{ \left[\left\{-2 B(Q+Sy) Z_{L}
S \Gamma\left[1+\frac{A}{4}\right]\right\}+ \left\{(B^2 G(Q+Sy)^2
+2Z_{L}S)\Gamma[\frac{1}{2}+\frac{A}{4}]\right\}\right]}{2 Z_{L} S \Gamma[1+\frac{A}{2}]}\\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\displaystyle+B_{1} \left(1+A B^2 (Q+Sy)^2 \right).
\end{array}
\end{equation}
The effective potential $V_{\Phi}(r_{c})$ can be obtained by substituting the above Eq (\ref{soln1})
into the scalar field action stated in Eq~(\ref{eq1}) and integrating out the extra dimensional
coordinate within $0\leq y\leq\pi$. This results in an effective potential for the modulus $r_c$ which
is given in the appendix.
\subsection{Some limiting cases of Einstein-GB-dilaton model}
We now discuss various limits that can emerge from our proposed model.
\label{l1d}
\subsubsection{Randall-Sundrum (RS) limit}
\label{as1}
Before discussing the effects of GB and dilaton term let us quickly recall that in absence of these terms
the action corresponds to the stabilization mechanism proposed by Goldberger and Wise.
In this case the modulus potential takes the form \cite{Goldberger:1999uk,Goldberger:1999un,Goldberger:1999wh}:
\begin{equation} \begin{array}{llll} \label{GWpot}
\displaystyle V_{\Phi}({r_c})= k_{RS}\epsilon {\cal V}^2_{h}+4k_{RS}e^{-4 k_{RS}r_c \pi}({\cal V}_{v}-{\cal V}_{h}e^{-\epsilon k_{RS}r_c \pi})^2\left(1+\frac{\epsilon}{4}\right)\\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~+ k_{RS} \epsilon {\cal V}_{h}e^{-(4+\epsilon) k_{RS}r_c \pi} (2{\cal V}_{v}-{\cal V}_{h} e^{-\epsilon k_{RS}r_c \pi})
\end{array}
\end{equation}
where $\epsilon=\frac{m^2_{\Phi}}{4 k^2_{RS}} <<1$ for which the terms of ${\cal O} (\epsilon^2)$
can be neglected.
One therefore obtains the minimum of the potential at:
\begin{equation}\label{minb}
k_{RS}r_c=\frac{4}{\pi}\frac{k^2_{RS}}{m^2_{\Phi}}\ln \left(\frac{{\cal V}_h}{{\cal V}_v}\right)
\end{equation}
Using Eq~(\ref{minb}) one can solve the hierarchy problem by choosing the ratio of VEVs at
$\frac{{\cal V}_h}{{\cal V}_v}=1.5$ and $\frac{m_{\Phi}}{k_{RS}}=0.2$. This choice
yields $k_{RS}r_{c} \sim 12$.
\subsubsection{Gauss-Bonnet (GB) gravity limit}
\label{as2}
In this case we choose the dilation coupling, $c_1= 0$, but the GB coupling $\alpha_{(5)}\neq 0$.
Substituting this in Eq~(\ref{parameter}) we get, $S=0$.
Here, the warp factor takes the form:
\begin{equation}
k_{M}(y)\rightarrow k_{L}=k_{RS}Z_{L}=k_{RS}(1+L+{\cal O}(L^2))
\end{equation}
This clearly implies that the warp factor in the RS case gets rescaled by a constant factor $Z_{L}=(1+L+{\cal O}(L^2))$ in pure GB limit.
One can obtain the same result as in the case of RS limit by replacing $k_{RS}$ to $k_{L}$ yielding the
stabilized potential:
\begin{equation} \begin{array}{llll} \label{GWpotGB}
\displaystyle V_{\Phi}({r_c})= k_{L}\epsilon_{L} {\cal V}^2_{h}+4k_{L}e^{-4 k_{L}r_c \pi}({\cal V}_{v}-{\cal V}_{h}e^{-\epsilon_{L} k_{L}r_c \pi})^2\left(1+\frac{\epsilon_{L}}{4}\right)\\
\displaystyle ~~~~~~~~~~~~~~~~~~~~~~~~~~+ k_{L} \epsilon_{L} {\cal V}_{h}e^{-(4+\epsilon_{L}) k_{L}r_c \pi} (2{\cal V}_{v}-{\cal V}_{h} e^{-\epsilon_{L} k_{L}r_c \pi})
\end{array}
\end{equation}
where $\epsilon_{L}=\frac{m^2_{\Phi}}{4 k^2_{L}} <<1$ for which the terms of ${\cal O} (\epsilon^2_{L})$
has been neglected.
Consequently the
minima appears at:
\begin{equation} \label {GWmin}
k_{L}r_{c}=\frac{4}{\pi}\frac{k^2_{L}}{m^2_{\Phi}}ln \left(\frac{{\cal V}_h}{{\cal V}_v}\right)
\Rightarrow k_{RS}r_{c}=\frac{4}{\pi}\frac{k^2_{RS}}{m^2_{\Phi}}Z_{L}\ln \left(\frac{{\cal V}_h}{{\cal V}_v}\right)
\end{equation}
where ${\cal O}(L^2)<<1$ terms can be neglected in the perturbative regime of the solution.
Since $k_{L}$ depends on both $\alpha_{(5)}$ and $k_{RS}$, we can get a family of solutions in terms of $k_{RS}$ and $\alpha_{5}$
to solve the gauge hierarchy problem in the Einstein Gauss-Bonnet gravity. This we shall discuss in a
more general set up later.
\subsubsection{Dilaton gravity limit}
\label{as3}
In this particular case, the GB coupling $\alpha_{(5)}=0$, but the dilaton coupling $c_1\neq 0$, which results in pure dilaton gravity limit.
Substituting this limit in Eq~(\ref{parameter}) we get, $L=0,Z_{L}=1$.
The warp factor in this case takes the form:
\begin{equation}\label{modifieddil}
k_{\bf M}(y)\rightarrow k_{D}(y)=k_{RS}~e^{\frac{c_{1}|y|}{2}}.
\end{equation}
The classical differential equation for scalar field
in the bulk turn out to be
\begin{equation} \begin{array}{llll} \label{GWeqndil}
\displaystyle -\frac{1}{r^2_{c}} \partial_{y}\left( e^{-4 k_{D}(y)r_c |y|} \partial_{y} \Phi(y)\right)+m^2_{\Phi} e^{-4 k_{D}(y)r_c |y|} \Phi(y)
+\frac{4}{r_{c}}\sum^{2}_{i=1}e^{-k_{D}(y)r_{c}|y|}\lambda_{i}\Phi\left(\Phi^{2}-{\cal V}^{2}_{i}\right)\delta(y-y_{i})=0
\end{array}
\end{equation}
Away from the boundaries at $y=0,\pi$, the general
solution of Eq~(\ref{GWeqndil}) can be written as:
\begin{equation}\begin{array}{llll} \label{soln1dil}
\displaystyle \Phi(y)= A_1 \frac{ \left[\left\{-2\sqrt{2S} \Gamma\left[\frac{2+\frac{G}{2S}}{4}\right]\right\}+ \left\{(\frac{G}{2S}(Q+Sy)^2
+2S)\Gamma[\frac{1}{2}+\frac{G}{4S}]\right\}\right]}{4S \Gamma[\frac{G}{2S}]}\\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\displaystyle+B_{1} \left(1+\frac{G(Q+Sy)^2}{2S^{2}} \right).
\end{array}
\end{equation}
This results in an effective potential which is explicitly given in the appendix.
\begin{figure*}[t]
\centering
\subfigure[$L=10^{-7},S=0.09,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=7.2cm, height=8.3cm] {GD1.pdf}
\label{fig:subfig1}
}
\subfigure[$L=10^{-1},S=0.09,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=7.2cm, height=8.3cm] {GD2.pdf}
\label{fig:subfig2}
}
\subfigure[$L=0.78,S=0.09,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=7.2cm, height=8.3cm] {GD4.pdf}
\label{fig:subfig3}
}
\subfigure[$L=0.92,S=0.09,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=7.2cm, height=8.3cm] {GD3.pdf}
\label{fig:subfig4}
}
\caption[Optional caption for list of figures]{Behaviour of the moduli stabilized potential with respect to the compactification radius $r_{c}$ in Gauss Bonnet dilaton (GBD)
limit.
}
\label{fig1}
\end{figure*}
\begin{figure*}[htb]
\centering
\subfigure[$L=0,S=50,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=11.1cm, height=8.3cm] {D1.pdf}
\label{fig:subfig6}
}
\subfigure[$L=0,S=0.4,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=11.1cm, height=8.3cm] {D2.pdf}
\label{fig:subfig7}
}
\caption[Optional caption for list of figures]{Behaviour of the modulus potential with respect to the compactification radius $r_{c}$ in pure dilaton
limit.
}
\label{fig2}
\end{figure*}
\begin{figure*}[htb]
\centering
\subfigure[$L=4\times 10^{-7},S=0,\frac{{\cal V}_h}{{\cal V}_v}=1.5$]{
\includegraphics[width=7.2cm, height=8.3cm] {GB1.pdf}
\label{fig:subfig8}
}
\subfigure[$L=10^{-1},S=0,\frac{{\cal V}_h}{{\cal V}_v}=1.25$]{
\includegraphics[width=7.2cm, height=8.3cm] {GB2.pdf}
\label{fig:subfig8a}
}
\subfigure[$L=0,S=0,\frac{{\cal V}_h}{{\cal V}_v}=1.5$]{
\includegraphics[width=10.1cm, height=8.3cm] {RS.pdf}
\label{fig:subfig9}
}
\caption[Optional caption for list of figures]{Behaviour of the modulus potential with respect to the compactification radius $r_{c}$
for Einstein-GB-dilaton gravity.}
\label{fig3}
\end{figure*}
\section{Features of the stabilized potential in higher curvature gravity}
\begin{table}[h]
\renewcommand{\tabcolsep}{0.06pc}
\small
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline
\hline\hline
Different & L & S & $\frac{{\cal V}_h}{{\cal V}_v}$& \multicolumn{2}{c|}{existence of } & \multicolumn{2}{c|}{value of $r_c$} & \multicolumn{2}{c|}{value of Potential$V(r_c) $} \\
\cline{5-10}
features & & & & $minima$ & $Maxima$ & $minima$ & $maxima$ & $minima$ & $maxima$ \\ \hline\hline\hline
Gauss-Bonnet & $10^{-7}$ & 0.09 & 1.25 & double &double& 0.3465,1.14 &0.7379,1.573& 0.004842,0.1855& 2.013,1.491 \\
Dilaton & $10^{-1}$ & 0.09 & 1.25 & double & double & 0.3461,1.07 &0.7031, 1.495 & 0.003442,0.1441 & 7.421, 3.621 \\
(GBD limit) & $0.78$ & 0.09 & 1.25 & Single &X& 0.4975 &X & 0.01214 & X \\
& $0.92$ & 0.09 & 1.25 & X &Single & X & 0.1281 & X & 0.4865 \\
\hline
Dilaton limit& 0& 50 & 1.25 &Single & single &2.873& 0.1019 & -8.719 & 17.79 \\
& 0&0.4 & 1.25 &Single &double &0.7119&0.2496, 1.312 & -0.01827 & 1.156, 1.192
\\ \hline
GB limit&$4\times 10^{-7}$ &0 & 1.5 & minima & X &12.77&X & -0.002217& X \\
&$10^{-1}$ &0 & 1.25 & minima & X &11.19&X & -0.001096& X \\\hline
RS limit & 0 & 0 & 1.5 &minima & X &12.74&X & -0.00067& X
\\ \hline\hline\hline
\end{tabular}
\caption{ Values of moduli radius and moduli potential in GBD, dilaton limit, GB and RS limit. } \label{table1}
\end{table}
\subsection{Case I: Einstein-Gauss Bonnet-dilaton bulk $(\alpha_{5} \neq 0, c_1 \neq 0)$:}
It is clear from the table~(\ref{table1}), fig~(\ref{fig:subfig1}) and fig~(\ref{fig:subfig2}) that in this case there
exists multiple (double) number of minima of the modulus potential obtained from the stabilization condition
of modulus within the interval, $10^{-1}\leq L\leq 10^{-7}$ for fixed
dilaton coupling at $S\sim 0.09$. In fig~(\ref{fig:subfig1}) and fig~(\ref{fig:subfig2}), the first minima
appears to be more stable than the second one. The presence of more than one minimum implies the possibility of tunneling
from one minimum to a more stable one i.e.the one with a lesser value of the moduli potential $V(r_{c})$. From
the table~(\ref{table1}) it may be seen that this causes decrease in the value of rc.
For a given ${\cal V}_h/{\cal V}_v$, this will result into an increase in the value of the warp factor
causing an enhancement of the value of the graviton Kaluza Klien (KK) mode masses and decrease
in the value of the KK graviton coupling to brane fields. As a result the cross section
for the KK graviton exchange will fall.Though the presence of two minima may imply
the possibility of tunneling, however as the two minima are separated by a width ${\cal O}(M_{p})$
one can rule out the possibility of tunneling from one stabilized minimum to the adjacent one.
We also observe from our analysis that if one increases the ratio of VEV,
then the position of the minimum of the potential slightly shifts toward the higher value of the $r_c$. We have seen that as the strength of the
GB coupling increases, one passes
from double minima to single minimum. Most significantly, the increase in GB coupling causes the
minima to disappear
while a maximum appears in the moduli potential.This signals disappearance of any stable value for the
modulus implying that large GB coupling leads to instability. See fig~(\ref{fig:subfig4}) for details. Moreover it can be seen that as the VEV decreases (ratio becomes $\sim
1.25$), the potential becomes deeper implying greater stability. Additionally, for $L=0.78,S=0.09$ and $L=0.92, S=0.09$ we get one minimum and one maximum respectively
as shown in fig~(\ref{fig:subfig3}-\ref{fig:subfig4}). Also we observe that when $L$ changes from $10^{-7}$
to $10^{-1}$, for $S \sim 0.2-0.9$ we get double minima of the potential. As $S$ increases from 0.9 the
double minima disappears and we have single minimum. On the other hand,
if $S$ decreases from 0.2, at about $S \sim 0.014$, we have an appearance of single minimum in the modulus
potential.We always keep $L$ from $10^{-7}$
to $10^{-1}$ since
$L\geq 1$ is not a feasible value as the perturbative setup will no longer be valid and
the theory goes to the non-perturbative regime of the solution which is beyond the scope of the present analysis.
\subsection{Case II: Dilaton limit $(\alpha_{5} = 0, c_1 \neq 0)$:}
If one considers the dilaton limit, then from the table~(\ref{table1}), one single minimum is observed.
In fig~(\ref{fig:subfig6}) and fig~(\ref{fig:subfig7}) we have depicted such features of stabilized potential with respect to modulus
for the weak and strong dilaton coupling fixed at $S=0.4$ and $S=50$ respectively.
We also observe from the present analysis that as in case of GBD scenario no such double minima appears in the scenario where only dilaton coupling is present.
Moreover as the strength of the dilaton coupling increases,
stability of the effective potential decreases.
\subsection{Case III: Gauss-Bonnet limit $(\alpha_{5} \neq 0, c_1 = 0)$:}
In GB limit, only single minimum is observed as mentioned in table~(\ref{table1}).
The behaviour of the modulus potential is depicted in fig~(\ref{fig:subfig8}) for the ratio of the VEV$\sim 1.5$.
Here we choose the value of the GB coupling $\sim {\cal O}(10^{-7})$ as constrained by various collider (i.e. higgs mass, $H\rightarrow
\gamma\gamma,\tau\bar{\tau}$ decay \cite{Choudhury:2013eoa} obtained from ATLAS \cite{ATLAS:2013mma,Aad:2012tfa,ATLAS:2011ab}
and CMS data \cite{Chatrchyan:2013lba}) and solar system observations \cite{Chakraborty:2012sd}.
There is no known dynamical origin of the small value of the Gauss-Bonnet coupling ${\cal O}(10^{−7})$. The consistency of the experimental
results points towards this value.
We have analyzed that as the VEV decreases (ratio becomes $\sim
1.25$) for a fixed GB coupling, the position of the minimum gets closer to the origin.
By adjusting the GB parameter $L$ , we can address the well known hierarchy problem. For example,
initially the ratio of VEV is fixed at $1.5$. In such a case $k_{L}r_{c}\sim {\cal O}(12.77)$
through which one can solve the hierarchy problem even in the weak
GB coupling $\sim {\cal O}(10^{-7})$. Now if the ratio of the VEV is decreased to $1.25$ then we observe that $k_{L}r_{c}\sim {\cal O}(6.98)$, which implies that fine tuning problem
cannot be addressed with a very weak GB coupling $\sim {\cal O}(10^{-7})$. But if we increase the GB coupling
to $\sim {\cal O}(10^{-1})$ within
the perturbative regime
then even with the decreased value of the ratio of VEV to $1.25$ the gauge hierarchy problem can be
addressed. See fig~(\ref{fig:subfig8a}) for the details.
Using the Eq~(\ref{GWmin}) we find that the ratio of the VEV can be expressed in the GB limit as:
\begin{equation} \frac{{\cal V}_h}{{\cal V}_v}=e^{\frac{\pi r_c m_{\phi}^2}{4 k_{RS}(1+L)}}.\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=12.3cm,height=10cm]{VEV.pdf}
\caption{\footnotesize Variation of the ratio of VEV with respect to the GB parameter $L$ for $\frac{\pi r_c m_{\phi}^2}{4 k_{RS}}\sim \frac{2}{5}$.
}
\label{figVEV}
\end{figure}
The variation of the ratio of VEV is given with respect to the GB parameter $L$ in Fig~(\ref{figVEV}).From
this figure, it can be clearly seen that in the limit $L \rightarrow 0$,we retrieve the RS limit.
Thus we can generate a parameter space consisting of the GB coupling and ratio of the VEV to
resolve the hierarchy problem.
Recently, in the context of radion phenomenology,\cite{Maitra:2013cta} it has been shown that in the presence
of GB coupling, radion VEV can be consistently adjusted to give first graviton excitation mass well above
$\sim$ $3$ TeV as required from the latest ATLAS data.
\subsection{Case IV: Randall-Sundrum limit $(\alpha_{5} = 0, c_1 = 0)$:}
In the RS limit, single minimum has been observed as mentioned in table~(\ref{table1}) The behaviour of the moduli potential is depicted in fig~(\ref{fig:subfig9})
for the ratio of the VEV $\sim 1.5$. To resolve the hierarchy problem, one should fix the
ratio to this prescribed value. If the ratio of the VEV decreases( ratio becomes $\sim
1.25$) , the position of the minimum gets closer to the origin and stability of the effective moduli potential
increases. However unlike the previous case now we have no parameter like GB parameter to the value of
$k_{RS}r_c$ so that a Planck to TeV scale warping can be achieved. Hence, we can conclude that in case of
zero GB coupling and zero dialton coupling, we have a specific choice for the ratio of the VEVs of the
bulk scalar to address the hierarchy issue. The presence of GB and dilaton in the bulk provide us with
flexibility in this choice.
\section{Conclusion}
In this work, we have studied the modulus stabilization mechanism in warped braneworld model when higher curvature gravity is present in the bulk via GB
and dilaton coupling (GBD). We have
also studied different limiting situations such as pure GB limit, pure dilaton limit and the RS limit. Analytical expressions for the stabilized
potentials are derived for different cases. We summarize our results as follows:
\begin{itemize}
\item We observe the existence of double minima when both GB and dilaton coupling are present. As the strength of the GB coupling increases the unstable minimum of these two double minima disappears,
resulting into a single minimum. If we go on increasing the strength of the GB coupling then it is
observed that the single minimum disappears and a single maximum in the modulus potential appears.
Thus increasing the GB coupling beyond a value leads to instability. Hence, in the perturbative regime of the solution we can always obtain a stabilized modulus potential although these stabilized values
of the modulus radius $r_c$ are
not effective in resolving the gauge hierarchy or fine-tuning problem as $k_{\bf M}r_c <<{\cal O}(12)$.
We observe that as $S$ goes beyond the value $\sim 0.9$ the minimum of the potential disappears
and we move to the region of instability. On the other hand, if value of the dilaton coupling
decreases from a value $\sim 0.2$ we have appearance of single minimum.
\item The existence of double minima of the moduli potential in higher curvature gravity may have interesting consequences in the context of stability of the model.
As the minima in GBD case are separated by a width ${\cal O}(M_{p})$ one can rule out the possibility
of tunneling from one stabilized minimum to the adjacent one.
\item In case of pure dilaton limit we observe that as the strength of the dilaton coupling increases the stability of the effective moduli potential increases.
Also we have only one minimum of the potential in this case.
\item In case of pure GB limit also only single minimum is observed. For a fixed weak GB coupling, as the ratio of the VEV decreases,
position of the minimum gets closer
to the origin. It is also observed that using weak GB coupling and large ratio of VEV one cannot solve the hierarchy
issue. However in the GB limit we observe that if the value of the GB coupling is increased then by decreasing the ratio of VEV it is
still possible to resolve the gauge hierarchy problem.
\item In the RS limit single minimum is observed as found in GW mechanism. One can solve the fine-tuning problem by taking a small value of the ratio of the VEV.
\item It is well known that in RS model the various KK graviton modes are important sources for phenomenological
signatures. The possible diphoton/dilepton decay channel of such gravitons are being studied by
ATLAS collaboration in LHC.
The most recent result has set stringent lower bounds on the 1st KK graviton $\sim 3$ TeV \cite{Das:2013lqa}.
With pure Einstein gravity in the bulk it is very difficult to satisfy this bound and it has been demonstrated
that the presence of higher curvature terms along with dilaton can explain the
ATLAS result. In this context the study of stability of our proposed model is of utmost importance.
Through this work we therefore undertake to present a detailed analysis of stabilizing the higher curvature modified
warped geometry model in presence of dilaton.
\end{itemize}
In summary, if we compare our findings with the original Goldberger-Wise stabilization mechanism
we observe that the presence of Gauss-Bonnet (GB) higher curvature term and dilaton term produces the following modification in the
modulus stabilization scenario. \\ \\
\begin{itemize}
\item If GB coupling $L$ increases beyond a desired value, for a given dilaton coupling $S$, then the minima of the potential disappears.\\
\item The value of the dilaton coupling $S$ should be below a critical value to avoid the appearance of
double minima which removes the possibility of tunneling.\\
\item The reduction in the stabilized value of the modulus $r_c$ (Please see the Table 1)
than Goldberger-Wise scenario implies an
improvement in reducing the hierarchy between $r_c$ and inverse of the 4D Planck scale $M^{-1}_{pl}$.
\end{itemize}
\section*{Acknowledgments}
SC thanks Council of Scientific and
Industrial Research, India for financial support through Senior
Research Fellowship (Grant No. 09/093(0132)/2010). SC and JM thanks Indian Association for the Cultivation of Sience (IACS)
where most of the work has been done.
\section{Appendix}
Let us explicitly write down the expression for the stabilized potential for the modulus in case of Gauss-Bonnet dilaton:
\begin{equation}\begin{array}{lllll}\label{potq}
V_{\Phi}(r_c)=V_{1}(r_c)+V_{2}(r_c),
\end{array}\end{equation}
where for Gauss-Bonnet dilaton gravity $V_{1}(r_c)$ and $V_{2}(r_c)$ are given by the following expressions:
\begin{equation}\small\begin{array}{llll} \label{pot1}
V_{1}(r_c)= -\frac{1}{\Gamma[(1+A/4)^2]}M_{1}(-1/(Z_{L}^5 Q^5)e^{-Z_{L}BQ^2}\\ ~~~~~~~~~~~~~~(-M^2(6A^2+6Z_{L}BQ^2 A^2+2Z^3_{L} B Q^4 A(1+B^2Q^2A/2)+Z^4_{L} Q^4(1+B^2Q^2A/2)^2 \\
~~~~+ 2Z^2_{L} Q^2 A (1+3 B^2Q^2 A/2))\Gamma(1+A/4)^
-Z_{L}^2 M^2 Q^2 (2+Z_{L}^2 B^2 Q^4+2 Z_{L} B Q^2)A^2 \Gamma[(1+A/2)/2]^2 \\
~~~~~~~~~~~~~~~~~~+ Z_{L} M Q A (6 A + 6 Z_{L} B Q^2 A+Z_{L}^3 B Q^4 (1+B^2 Q^2 A)\\ ~~~~~~~~~~~~~~~~~+Z_{L}^2 Q^2 (1+3 B^2 Q^2 A))W \Gamma[(1+A/2)/2]\Gamma[(1+A/2)] \\
~~~~~~~~~~- (24 A^2+24 Z_{L} B Q^2 A^2 +4 Z_{L}^3 B Q^4 A(1+B^2 Q^2 A)+ Z_{L}^4 Q^4 (1+B^2 Q^2 A)^2\\ ~~~~~~~~~~~~+ 4 Z_{L}^2 Q^2 A (1+3 B^2 Q^2 A))W^2 \Gamma[(1+A/2)^2] \\
~~~~~~~~+ 2 M \Gamma[(1+A/4)](Z_{L}MQA/2 (3A+3 Z_{L} B Q^2 A+Z_{L}^3 B Q^4(1+ B^2 Q^2 A/2)\\ ~~~~~~~~+Z_{L}^2 Q^2 (1+3 B^2 Q^2 A/2)) \Gamma[(1+A/2)/2] -
- (12 A^2+12 Z_{L} B Q^2 A^2 + 3 Z_{L}^2 Q^2 A (1+2 B^2 Q^2 A)\\~~~~~~~+ Z_{L}^3 B Q^4 A (3+ 2 B^2 Q^2 A) + Z_{L}^4 Q^4 (1+ 3 B^2 Q^2 A/2 + B^4 Q^4 A^2/2)) W \Gamma[(1+A/2)])) \\
~~~~~~~+ 1/(Z_{L}^5 Q^5)e^{-Z_{L}BQ(Q+\pi S)}(-M^2(6A^2+6Z_{L}BQ (Q+\pi S) A^2\\~~~~~~~~~+2z^3 B Q^3(Q+\pi S) A(1+B^2(Q+\pi S)^2 A/2)
+z^4 Q^4(1+B^2 (Q+\pi S)^2 A/2)^2 \\
~~~~~~~~~+ 2z^2 Q^2 A (1+3 B^2(Q+\pi S)^2 A/2))\Gamma(1+A/4)^2-(Z_{L}^2 M^2 Q^2/4) (2+2 Z_{L} B Q(Q+\pi S)\\~~~~~~~~+ Z_{L}^2 B^2 Q^2(Q+\pi S))A^2 \Gamma[(1+A/2)/2]^2
+ Z_{L} M Q A (6 A + 6 Z_{L} B Q(Q+\pi S) A \\ ~~~~~~~~~~+Z_{L}^3 B Q^3(Q+\pi S) (1+B^2 (Q+\pi S)^2 A/2)\\~~~~~~~~~+Z_{L}^2 Q^2 (1+3 B^2 (Q+\pi S)^2 A/2))W \Gamma[(1+A/2)/2]\Gamma[(1+A/2)] \\
~~~~~~~- (24 A^2+24 Z_{L} B Q(Q+\pi S) A^2 +4 Z_{L}^3 B Q^3(Q+\pi S) A(1+B^2 (Q+\pi S)^2 A)\\
~~~~~~~~~~~~+ Z_{L}^4 Q^4 (1+B^2 (Q+\pi S)^2 A)^2 + 4 Z_{L}^2 Q^2 A (1+3 B^2 (Q+\pi S)^2 A))W^2 \Gamma[(1+A/2)^2] \\
~~~~~~~~~~~~~~~~~~+ 2 M \Gamma[(1+A/4)](Z_{L} M Q A/2 (3A+3 Z_{L} B Q(Q+\pi S) A\\ ~~~~+Z_{L}^3 B Q^3(Q+\pi S)(1+ B^2 (Q+\pi S)^2 A/2)+Z_{L}^2 Q^2 (1+3 B^2 (Q+\pi S)^2 A/2)) \Gamma[(1+A/2)/2] - \\
~~~~~~~~- (12 A^2+12 Z_{L} B Q(Q+\pi S) A^2 + 3 Z_{L}^2 Q^2 A (1+2 B^2 (Q+\pi S)^2 A)\\~~~~~~~~~~ + Z_{L}^3 B Q^3(Q+\pi S) A (3+ 2 B^2 (Q+\pi S)^2 A) + Z_{L}^4 Q^4 (1+ 3 B^2
(Q+\pi S)^2 A/2 \\~~~~~~~~~~~~+ B^4 (Q+\pi S)^4 A^2/2)) W \Gamma[(1+A/2)])))
\end{array}
\end{equation}
\begin{equation}\small\begin{array}{llll} \label{pot2}
V_{2}(r_c)= -\frac{1}{\Gamma[(1+A/2)]^2} Z_{L} S A ( e^{-Z_{L} B Q^2} (-\frac{1}{4 S}Q A (-4 M \Gamma[(1+A/4)] + 3 Z_{L} M Q \Gamma[(1+A/2)/2]\\
~~~~~~~~~~~~-8 W \Gamma [(1+A/2)]
( M \Gamma[(1+A/4)] + 2 W \Gamma [(1+A/2)])+B^3 Q^3 A(M \Gamma [(1+A/4)]\\
~~~~~~~~~~~~+2 W \Gamma [(1+A/2)])^2-\frac{1}{Z_{L}^3 Q^3} (-4 M^2 A \Gamma [(1+A/4)]^2
- Z_{L}^2 M^2 Q^2 A/2 \Gamma[(1+A/2)/2]^2 \\
~~~~~~~~~~~~~+ Z_{L} M Q (Z_{L}^2 Q^2+4 A)W \Gamma[(1+A/2)/2] \Gamma[1+(A/2)]-16 A W^2 \Gamma[1+(A/2)]^2 \\
~~~~~~~~~~~~ + M \Gamma[(1+A/4)](Z_{L} M Q(Z_{L}^2 Q^2+2 A) \Gamma[(1+A/2)/2] -16 A W \Gamma[1+(A/2)]))\\
~~~~~~~~~~~~+ \frac{1}{2 Z_{L} B Q S}(2 M^2 (Z_{L}^2 Q^2+2 A)\Gamma[(1+A/4)]^2
+ Z_{L}^2 M^2 Q^2 A/2 \Gamma[1+(A/2)]^2 \\
~~~~~~~~~~~-4 Z_{L} M Q A W \Gamma[(1+(A/2)/2)] \Gamma[1+(A/2)]+ 4(Z_{L}^2 Q^2 +4 A)W^2 \Gamma[1+(A/2)]^2 \\
~~~~~~~~~~~~~ + 2 M \Gamma[(1+A/4)] (-Z_{L} M Q A \Gamma[(1+(A/2)/2)] + (3 Z_{L}^2 Q^2 +8 A) W \Gamma[(1+(A/2))])))\\
~~~~~~~~~~~~ + e^{-Z_{L} B Q(Q+\pi S)} (\frac{1}{4 Q S}(Q+\pi S)^2 A (-4 M \Gamma[(1+A/4)]\\
~~~~~~~~~~~ + 3 Z_{L} M Q \Gamma[(1+A/2)/2]-8 W \Gamma [(1+A/2)])
( M \Gamma[(1+A/4)] + 2 W \Gamma [(1+A/2)])\\
~~~~~~~~~~~~-B^3 (Q+\pi S)^3 A(M \Gamma [(1+A/4)]+2 W \Gamma [(1+A/2)])^2\\
~~~~~~~~~~~~+\frac{1}{Z_{L}^3 Q^3} (-4 M^2 A \Gamma [(1+A/4)]^2
- (Z_{L}^2 M^2 Q^2A/2) \Gamma[(1+A/2)/2]^2 \\~~~~~~~~~~~~+ Z_{L} M Q (Z_{L}^2 Q^2+4 A)W \Gamma[(1+A/2)/2] \Gamma[1+(A/2)]-16 A W^2 \Gamma[1+(A/2)]^2 \\
~~~~~~~~~~~~+ M \Gamma[(1+A/4)](Z_{L} M Q(Z_{L}^2 Q^2+2 A) \Gamma[(1+A/2)/2] -16 A W \Gamma[1+(A/2)]))\\
~~~~~~~~~~~~~- \frac{1}{2 Z_{L} B Q^2 S}(Q+\pi S)(2 M^2 (Z_{L}^2 Q^2+2 A)\Gamma[(1+A/4)]^2
+ Z_{L}^2 M^2 Q^2 A/2 \Gamma[(1+(A/2))/2]^2 \\
~~~~~~~~~~~-4 Z_{L} M Q A W \Gamma[(1+(A/2)/2)] \Gamma[1+(A/2)]+ 4(Z_{L}^2 Q^2 +4 A)W^2 \Gamma[1+(A/2)]^2 \\
~~~~~~~~~~~~~+ 2 M \Gamma[(1+A/4)] (-Z_{L} M Q A \Gamma[(1+(A/2)/2)] + (3 Z_{L}^2 Q^2 +8 A) W \Gamma[(1+(A/2))]))))
\end{array}
\end{equation}
where all the parameters $Q,S,L,Z_{L},A,B$ are deined in Eq~(\ref{parameter}) and Eq~(\ref{cont}).
Further if we substitute $Z_{L}=1$ in Eq~(\ref{potq}), Eq~(\ref{pot1}) and Eq~(\ref{pot2}) then it results in the stabilized potential for
modulus in case of pure dilaton gravity
limit as mentioned in \ref{as3}.
|
\section{Introduction}
Non-Minimal Flavor Violating (NMFV) processes in the scalar quark sector
of the Minimal Supersymmetric Standard Model
(MSSM)~\cite{mssm,Haber:1989xc,Gunion:1984yn,Gunion:1986nh}, provide
important probes to new physics involving non-vanishing flavor
mixing between the three generations. Within the Standard Model (SM),
the only source of flavor violation comes from the CKM matrix,
$V_{\rm CKM}$, and thus
in general leads to small contributions.
Within the MSSM there are clear candidates to produce
flavor mixings with important phenomenological implications.
The possible presence of soft Supersymmetry
(SUSY)-breaking parameters in the squark sector, which are
off-diagonal in flavor space (mass parameters as well as trilinear
couplings) are
the most general way to introduce squark flavor mixing within the
MSSM. The off-diagonality in the squark mass matrix reflects the
misalignment (in flavor space) between quark and squark mass
matrices, that cannot be diagonalized
simultaneously. This misalignment can be produced from various
origins, but we will not rely on any particular one in this work.
For instance, these off-diagonal squark mass matrix entries can be
generated by renormalization effects from the CKM
matrix, which can be obtained by means of the Renormalization
Group Equations
(RGE) running from a high energy scale, where gauge coupling unification
is achieved, down to the low energies where the NMFV effects
are explored.
In this work we will not investigate the possible dynamical origin
of this squark-quark misalignment, nor the particular predictions for
the off-diagonal squark soft SUSY-breaking mass terms in specific
SUSY models, but instead we parametrize the general
non-diagonal entries in the squark mass matrices
in terms of generic soft SUSY-breaking terms, and we explore
here their phenomenological implications on various precision
observables. In
particular, we explore the consequences of these general squark mass matrices
on the light MSSM Higgs boson mass, $M_h$, as well as on the three most
prominent $B$-physics observables, \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and \ensuremath{\Delta M_{B_s}}. Specifically, we
parametrize the non-diagonal squark mass matrix entries in terms of a
complete set of generic dimensionless parameters, $\ensuremath{\delta^{AB}_{ij}}$ ($A,B=L,R$;
$i,j=u,c,t$ or $d,s,b$) where $L,R$ refer to the
``left-'' and ``right-handed'' SUSY partners of the corresponding
quark degrees of freedom
and $i,j$ ($i \neq j$) are the involved generation indexes. For the
presentation of our theoretical framework and notation we follow
closely our previous work~\cite{mhNMFV} on this same subject, which
was done previous to the Higgs discovery.
The main aspect of this work is setting updated bounds on the
allowed values of the $\ensuremath{\delta^{AB}_{ij}}$'s in this
model-independent parametrization of general squark
flavor mixing.
In particular, this is done in view of the collected data at
LHC\cite{LHCHiggslast,LHCSusy}, which has
provided very important information and constraints for the MSSM,
including the absence of SUSY particle experimental signals and
the discovery of a Higgs boson with a mass close to $125 - 126 \,\, \mathrm{GeV}$.
We work consistently in MSSM scenarios that are compatible with
LHC data. It should be noted that the analyzed scenarios have relatively
heavy SUSY spectra, which are naturally in
agreement with the present MSSM particle mass bounds (although
substantially lower masses, especially in the electroweak sector, are
allowed by LHC data). Furthermore the analyzed scenarios are chosen such
that the light ${\cal CP}$-even MSSM Higgs mass is around $125 - 126 \,\, \mathrm{GeV}$ and
thus in agreement with the Higgs boson discovery~\cite{LHCHiggs}.
In addition we
require that our selected MSSM scenarios give a
prediction for the muon anomalous magnetic moment, $(g-2)_\mu$, in
agreement with current data~\cite{Bennett:2006fi}.
The paper is organized as follows: first we review the main features of
the MSSM with general squark flavor mixing and set the relevant
notation for the $\ensuremath{\delta^{AB}_{ij}}$'s in ~\refse{sec:nmfv}. The description
of the numerical scenarios that we choose here is also done in this
section. The selection of
relevant precision observables and flavor observables
we are working with are
presented in ~\refse{sec:obs}. A summary on the present
experimental bounds on NMFV, that will be used in our analysis are also
included in this section. \refse{sec:results} contains the main results
of our numerical analysis and present the updated constraints found on the
$\ensuremath{\delta^{AB}_{ij}}$'s. Our conclusions are summarized in
\refse{sec:conclusions}.
\section{Calculational basis for Non-Minimal Flavor Violation}
\label{sec:nmfv}
\subsection{Theoretical set-up}
We work in SUSY scenarios with the same particle content as the
MSSM, but with
general flavor mixing hypothesis in the squark sector. Within these
SUSY-NMFV scenarios, besides the usual flavor violation originated by
the CKM matrix of the quark sector, the general flavor mixing in the
squark mass matrices additionally generates flavor violation from the
squark sector.
These squark flavor mixings are usually described in terms of a set of
dimensionless parameters $\delta^{AB}_{ij}$ ($X,Y=L,R$; $i,j=u,c,t$ or $d,s,b$).
In this section we summarize the main features of the squark flavor
mixing within the SUSY-NMFV scenarios and set the notation.
The more theoretical background, including the derivation from the super
potential can be found in \citere{mhNMFV}.
The usual procedure to introduce general flavor mixing in the squark
sector is to include the non-diagonality in flavor space once the quarks
have been rotated to the physical basis,
namely, in the so-called Super-CKM basis. Thus, one usually writes
the $6\times 6$ non-diagonal mass matrices, ${\cal M}_{\tilde u}^2$
and ${\cal M}_{\tilde d}^2$, referred to the Super-CKM basis, being ordered
respectively as $(\tilde{u}_L, \tilde{c}_L, \tilde{t}_L, \tilde{u}_R, \tilde{c}_R, \tilde{t}_R)$ and
$(\tilde{d}_L, \tilde{s}_L, \tilde{b}_L, \tilde{d}_R, \tilde{s}_R, \tilde{b}_R)$, and write them in
terms of left- and right-handed blocks $M^2_{\tilde q \, AB}$
($\tilde q= \tilde u, \tilde d;$ $A,B=L,R$), which are
non-diagonal $3\times 3$ matrices,
\begin{equation}
{\cal M}_{\tilde q}^2 =\left( \begin{array}{cc}
M^2_{\tilde q \, LL} & M^2_{\tilde q \, LR} \\[.3em]
M_{\tilde q \, LR}^{2 \, \dagger} & M^2_{\tilde q \,RR}
\end{array} \right), \qquad \tilde q= \tilde u, \tilde d~,
\label{eq:blocks-matrix}
\end{equation}
where:
\begin{alignat}{5}
M_{\tilde u \, LL \, ij}^2
= & m_{\tilde U_L \, ij}^2 + \left( m_{u_i}^2
+ (T_3^u-Q_u\sin^2 \theta_W ) M_Z^2 \cos 2\beta \right) \delta_{ij}, \notag\\
M^2_{\tilde u \, RR \, ij}
= & m_{\tilde U_R \, ij}^2 + \left( m_{u_i}^2
+ Q_u\sin^2 \theta_W M_Z^2 \cos 2\beta \right) \delta_{ij} \notag, \\
M^2_{\tilde u \, LR \, ij}
= & \left< {\cal H}_2^0 \right> {\cal A}_{ij}^u- m_{u_{i}} \mu \cot \beta \, \delta_{ij},
\notag, \\
M_{\tilde d \, LL \, ij}^2
= & m_{\tilde D_L \, ij}^2 + \left( m_{d_i}^2
+ (T_3^d-Q_d \sin^2 \theta_W ) M_Z^2 \cos 2\beta \right) \delta_{ij}, \notag\\
M^2_{\tilde d \, RR \, ij}
= & m_{\tilde D_R \, ij}^2 + \left( m_{d_i}^2
+ Q_d\sin^2 \theta_W M_Z^2 \cos 2\beta \right) \delta_{ij} \notag, \\
M^2_{\tilde d \, LR \, ij}
= & \left< {\cal H}_1^0 \right> {\cal A}_{ij}^d- m_{d_{i}} \mu \tan \beta \, \delta_{ij}~,
\label{eq:SCKM-entries}
\end{alignat}
with, $i,j=1,2,3$, $Q_u=2/3$, $Q_d=-1/3$, $T_3^u=1/2$ and
$T_3^d=-1/2$.
$\sin^2\theta_W = 1 - M_W^2/M_Z^2$
with $M_{W,Z}$ denoting the masses of the $W$~and $Z$~boson mass,
respectively,
and $(m_{u_1},m_{u_2}, m_{u_3})=(m_u,m_c,m_t)$, $(m_{d_1},m_{d_2},
m_{d_3})=(m_d,m_s,m_b)$. $\mu$ is the usual Higgsino mass term and
$\tan \beta=v_2/v_1$
with $v_1=\left< {\cal H}_1^0 \right>$ and $v_2=\left< {\cal H}_2^0
\right>$ being the two vacuum expectation values of the corresponding
neutral Higgs boson in the Higgs $SU(2)_L$ doublets,
${\cal H}_1= ({\cal H}^0_1\,\,\, {\cal H}^-_1)$ and ${\cal H}_2= ({\cal H}^+_2 \,\,\,{\cal H}^0_2)$.
It should be noted that the non-diagonality in flavor comes from the
values of $m_{\tilde U_L \, ij}^2$, $m_{\tilde U_R \, ij}^2$,
$m_{\tilde D_L \, ij}^2$, $m_{\tilde D_R \, ij}^2$, ${\cal A}_{ij}^u$
and ${\cal A}_{ij}^d$ for $i \neq j$.
The general squark flavor mixing is introduced via the
non-diagonal terms in the soft breaking squark mass matrices and
trilinear coupling matrices, which are defined here as:
\begin{equation}
m^2_{\tilde U_L}= \left(\begin{array}{ccc}
m^2_{\tilde Q_{1}} & \delta_{12}^{LL} m_{\tilde Q_{1}}m_{\tilde Q_{2}} &
\delta_{13}^{LL} m_{\tilde Q_{1}}m_{\tilde Q_{3}} \\
\delta_{21}^{LL} m_{\tilde Q_{2}}m_{\tilde Q_{1}} & m^2_{\tilde Q_{2}} &
\delta_{23}^{LL} m_{\tilde Q_{2}}m_{\tilde Q_{3}}\\
\delta_{31}^{LL} m_{\tilde Q_{3}}m_{\tilde Q_{1}} &
\delta_{32}^{LL} m_{\tilde Q_{3}}m_{\tilde Q_{2}}& m^2_{\tilde Q_{3}}
\end{array}\right)~,
\label{mUL}
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde D_L}= V_{\rm CKM}^\dagger \, m^2_{\tilde U_L} \, V_{\rm CKM}~,
\label{mDL}
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde U_R}= \left(\begin{array}{ccc}
m^2_{\tilde U_{1}} & \delta_{uc}^{RR} m_{\tilde U_{1}}m_{\tilde U_{2}} &
\delta_{ut}^{RR} m_{\tilde U_{1}}m_{\tilde U_{3}}\\
\delta_{{cu}}^{RR} m_{\tilde U_{2}}m_{\tilde U_{1}} & m^2_{\tilde U_{2}} &
\delta_{ct}^{RR} m_{\tilde U_{2}}m_{\tilde U_{3}}\\
\delta_{{tu}}^{RR} m_{\tilde U_{3}} m_{\tilde U_{1}}&
\delta_{{tc}}^{RR} m_{\tilde U_{3}}m_{\tilde U_{2}}& m^2_{\tilde U_{3}}
\end{array}\right)~,
\end{equation}
\noindent
\begin{equation}
m^2_{\tilde D_R}= \left(\begin{array}{ccc}
m^2_{\tilde D_{1}} & \delta_{ds}^{RR} m_{\tilde D_{1}}m_{\tilde D_{2}} &
\delta_{db}^{RR} m_{\tilde D_{1}}m_{\tilde D_{3}}\\
\delta_{{sd}}^{RR} m_{\tilde D_{2}}m_{\tilde D_{1}} & m^2_{\tilde D_{2}} &
\delta_{sb}^{RR} m_{\tilde D_{2}}m_{\tilde D_{3}}\\
\delta_{{bd}}^{RR} m_{\tilde D_{3}} m_{\tilde D_{1}}&
\delta_{{bs}}^{RR} m_{\tilde D_{3}}m_{\tilde D_{2}}& m^2_{\tilde D_{3}}
\end{array}\right)~,
\end{equation}
\noindent
\begin{equation}
v_2 {\cal A}^u =\left(\begin{array}{ccc}
m_u A_u & \delta_{uc}^{LR} m_{\tilde Q_{1}}m_{\tilde U_{2}} &
\delta_{ut}^{LR} m_{\tilde Q_{1}}m_{\tilde U_{3}}\\
\delta_{{cu}}^{LR} m_{\tilde Q_{2}}m_{\tilde U_{1}} &
m_c A_c & \delta_{ct}^{LR} m_{\tilde Q_{2}}m_{\tilde U_{3}}\\
\delta_{{tu}}^{LR} m_{\tilde Q_{3}}m_{\tilde U_{1}} &
\delta_{{tc}}^{LR} m_{\tilde Q_{3}} m_{\tilde U_{2}}& m_t A_t
\end{array}\right)~,
\label{v2Au}
\end{equation}
\noindent
\begin{equation}
v_1 {\cal A}^d =\left(\begin{array}{ccc}
m_d A_d & \delta_{ds}^{LR} m_{\tilde Q_{1}}m_{\tilde D_{2}} &
\delta_{db}^{LR} m_{\tilde Q_{1}}m_{\tilde D_{3}}\\
\delta_{{sd}}^{LR} m_{\tilde Q_{2}}m_{\tilde D_{1}} & m_s A_s &
\delta_{sb}^{LR} m_{\tilde Q_{2}}m_{\tilde D_{3}}\\
\delta_{{bd}}^{LR} m_{\tilde Q_{3}}m_{\tilde D_{1}} &
\delta_{{bs}}^{LR} m_{\tilde Q_{3}} m_{\tilde D_{2}}& m_b A_b
\end{array}\right)~.
\label{v1Ad}
\end{equation}
\noindent
In all this work, for simplicity, we are assuming that all $\ensuremath{\delta^{AB}_{ij}}$
parameters are real, therefore, hermiticity of
${\cal M}_{\tilde q}^2$ implies $\delta_{ij}^{AB}= \delta_{ji}^{BA}$.
It should be noted that
we have used a common notation for the
$\del{LL}{ij}$'s with $i,j=1,2,3$ in the $\tilde U_L$ and $\tilde D_L$
sectors, due to the $SU(2)_L$ gauge invariance that relates
$m^2_{\tilde U_L}$ and $m^2_{\tilde D_L}$ via $V_{\rm CKM}$,
as given in \refeqs{mUL} and (\ref{mDL}).
\medskip
The next step is to rotate the squark states from the Super-CKM basis,
${\tilde q}_{L,R}$, to the physical basis.
If we set the order in the Super-CKM basis as above,
$(\tilde{u}_L, \tilde{c}_L, \tilde{t}_L, \tilde{u}_R, \tilde{c}_R, \tilde{t}_R)$ and
$(\tilde{d}_L, \tilde{s}_L, \tilde{b}_L, \tilde{d}_R, \tilde{s}_R, \tilde{b}_R)$,
and in the physical basis as
${\tilde u}_{1,..6}$ and ${\tilde d}_{1,..6}$, respectively, these last
rotations are given by two $6 \times 6$ matrices, $R^{\tilde u}$ and
$R^{\tilde d}$,
\begin{equation}
\left( \begin{array}{c} \tilde u_{1} \\ \tilde u_{2} \\ \tilde u_{3} \\
\tilde u_{4} \\ \tilde u_{5} \\\tilde u_{6} \end{array} \right)
\; = \; R^{\tilde u} \left( \begin{array}{c} \tilde{u}_L \\ \tilde{c}_L \\\tilde{t}_L \\
\tilde{u}_R \\ \tilde{c}_R \\ \tilde{t}_R \end{array} \right) ~,~~~~
\left( \begin{array}{c} \tilde d_{1} \\ \tilde d_{2} \\ \tilde d_{3} \\
\tilde d_{4} \\ \tilde d_{5} \\ \tilde d_{6} \end{array} \right) \; = \; R^{\tilde d} \left( \begin{array}{c} \tilde{d}_L \\ \tilde{s}_L \\ \tilde{b}_L \\
\tilde{d}_R \\ \tilde{s}_R \\ \tilde{b}_R \end{array} \right) ~,
\label{newsquarks}
\end{equation}
yielding the diagonal mass-squared matrices as follows,
\begin{eqnarray}
{\rm diag}\{m_{\tilde u_1}^2, m_{\tilde u_2}^2,
m_{\tilde u_3}^2, m_{\tilde u_4}^2, m_{\tilde u_5}^2, m_{\tilde u_6}^2
\} & = &
R^{\tilde u} \; {\cal M}_{\tilde u}^2 \;
R^{\tilde u \dagger} ~,\\
{\rm diag}\{m_{\tilde d_1}^2, m_{\tilde d_2}^2,
m_{\tilde d_3}^2, m_{\tilde d_4}^2, m_{\tilde d_5}^2, m_{\tilde d_6}^2
\} & = &
R^{\tilde d} \; {\cal M}_{\tilde d}^2 \;
R^{\tilde d \dagger} ~.
\end{eqnarray}
The corresponding Feynman rules in the physical basis for the vertices
including NMFV squarks had been
implemented into the program packages
{\tt FeynArts}/{\tt FormCalc}~\cite{feynarts,formcalc}
extending the previous MSSM model
file~\cite{famssm}. The Feynman rules of the NMFV MSSM that are
relevant for the present work can be found in \cite{mhNMFV}.
\subsection{Numerical scenarios}
\label{sec:scenarios}
Regarding our choice of MSSM parameters for our forthcoming numerical
analysis of the NMFV constraints, we have proceeded within two frameworks,
both compatible with present data, that we briefly describe in the following.
\subsubsection{Framework 1}
In the first framework, we have selected six specific points in the MSSM
parameter space, S1, \ldots, S6, as examples of points that are allowed by
present data, including recent LHC searches and the measurements of the
muon anomalous magnetic moment. In \refta{tab:spectra} the
values of the various MSSM parameters as well as the values of the
predicted MSSM mass spectra are summarized, with all $\ensuremath{\delta^{AB}_{ij}} = 0$.
They were evaluated with
the program {\tt FeynHiggs}~\cite{feynhiggs,mhiggsAEC}. For simplicity, and to
reduce the number of
independent MSSM input parameters we have assumed equal soft masses for
the squarks of the first and second generations (similarly for the
sleptons), equal soft masses for the left and right squark sectors
(similarly for the sleptons, where $\tilde L$ denotes the
``left-handed'' slepton sector, whereas $\tilde E$ denotes
the ``right-handed'' charged slepton sector)
and also equal trilinear couplings for
the stop, $A_t$, and sbottom squarks, $A_b$. In the slepton sector we
just consider the stau trilinear coupling, $A_\tau$. The other trilinear
sfermion couplings are set to zero value. Regarding the soft
SUSY-breaking parameters for the gaugino
masses, $M_i$ ($i=1,2,3$), we assume an approximate GUT relation. The
pseudoscalar Higgs mass $M_A$, and the $\mu$ parameter are also taken as
independent input parameters. In summary, the six points
S1, \ldots, S6 are
defined in terms of the following subset of ten input MSSM parameters
(plus the $\ensuremath{\delta^{AB}_{ij}}$, which will be analyzed below):
\pagebreak
\begin{eqnarray}
m_{\tilde L_1} &=& m_{\tilde L_2} \; ; \; m_{\tilde L_3} \;
(\mbox{with~} m_{\tilde L_{i}} = m_{\tilde E_{i}}\,\,,\,\,i=1,2,3)~, \nonumber \\
m_{\tilde Q_1} &=& m_{\tilde Q_2} \; ; \; m_{\tilde Q_3} \;
(\mbox{with~} m_{\tilde Q_i} = m_{\tilde U_i} = m_{\tilde D_i}\,\,,\,\,i=1,2,3)~,
\nonumber \\
A_t&=&A_b\,\,;\,\,A_\tau~, \nonumber \\
M_2&=&2 M_1\, =\,M_3/4 \,\,;\,\,\mu \nonumber~, \\
M_A&\,\,;\,\, &\tan \beta~.
\end{eqnarray}
\begin{table}[h!]
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& S1 & S2 & S3 & S4 & S5 & S6 \\\hline
$m_{\tilde L_{1,2}}$& 500 & 750 & 1000 & 800 & 500 & 1500 \\
$m_{\tilde L_{3}}$ & 500 & 750 & 1000 & 500 & 500 & 1500 \\
$M_2$ & 500 & 500 & 500 & 500 & 750 & 300 \\
$A_\tau$ & 500 & 750 & 1000 & 500 & 0 & 1500 \\
$\mu$ & 400 & 400 & 400 & 400 & 800 & 300 \\
$\tan \beta$ & 20 & 30 & 50 & 40 & 10 & 40 \\
$M_A$ & 500 & 1000 & 1000 & 1000 & 1000 & 1500 \\
$m_{\tilde Q_{1,2}}$ & 2000 & 2000 & 2000 & 2000 & 2500 & 1500 \\
$m_{\tilde Q_{3}}$ & 2000 & 2000 & 2000 & 500 & 2500 & 1500 \\
$A_t$ & 2300 & 2300 & 2300 & 1000 & 2500 & 1500 \\\hline
$m_{\tilde l_{1}}-m_{\tilde l_{6}}$ & 489-515 & 738-765 & 984-1018 & 474-802 & 488-516 & 1494-1507 \\
$m_{\tilde \nu_{1}}-m_{\tilde \nu_{3}}$& 496 & 747 & 998 & 496-797 & 496 & 1499 \\
$m_{{\tilde \chi}_1^\pm}-m_{{\tilde \chi}_2^\pm}$ & 375-531 & 376-530 & 377-530 & 377-530 & 710-844 & 247-363 \\
$m_{{\tilde \chi}_1^0}-m_{{\tilde \chi}_4^0}$& 244-531 & 245-531 & 245-530 & 245-530 & 373-844 & 145-363 \\
$M_{h}$ & 126.6 & 127.0 & 127.3 & 123.1 & 123.8 & 125.1 \\
$M_{H}$ & 500 & 1000 & 999 & 1001 & 1000 & 1499 \\
$M_{A}$ & 500 & 1000 & 1000 & 1000 & 1000 & 1500 \\
$M_{H^\pm}$ & 507 & 1003 & 1003 & 1005 & 1003 & 1502 \\
$m_{\tilde u_{1}}-m_{\tilde u_{6}}$& 1909-2100 & 1909-2100 & 1908-2100 & 336-2000 & 2423-2585 & 1423-1589 \\
$m_{\tilde d_{1}}-m_{\tilde d_{6}}$ & 1997-2004 & 1994-2007 & 1990-2011 & 474-2001 & 2498-2503 & 1492-1509 \\
$m_{\tilde g}$ & 2000 & 2000 & 2000 & 2000 & 3000 & 1200 \\
\hline
\end{tabular}
\caption{
Selected points in the MSSM parameter space (upper part)
and their corresponding spectra (lower part), with all $\ensuremath{\delta^{AB}_{ij}} = 0$.
All mass parameters and trilinear couplings are given in GeV.}
\label{tab:spectra}
\end{table}
The specific values of these ten MSSM parameters in \refta{tab:spectra},
to be used in the forthcoming NMFV analysis, are chosen to provide
different
patterns in the various sparticle masses, but all leading to rather
heavy spectra, thus they are naturally in agreement with the
absence of SUSY signals at LHC. In particular
all points lead to rather heavy squarks and gluinos above $1200\,\, \mathrm{GeV}$ and
heavy sleptons above $500\,\, \mathrm{GeV}$ (where the LHC limits would also
permit substantially lighter scalar leptons).
The values of $M_A$ within the interval
$(500,1500)\,\, \mathrm{GeV}$, $\tan \beta$ within the interval $(10,50)$ and a large
$A_t$ within $(1000,2500)\,\, \mathrm{GeV}$ are fixed such that a light Higgs boson~$h$
within the LHC-favored range $(123, 128)\,\, \mathrm{GeV}$ is obtaine
\footnote{
This range takes into account experimental uncertainties as well as
theoretical uncertainties, where the latter would permit an even larger
interval~\cite{mhiggsAEC,ehowp}. However, for the phenomenological
analyses later we will use a
correspondingly wider range.
~in the Minimal Flavor Violation (MFV) limit
\footnote{
Here, by MFV limit we mean setting all $\ensuremath{\delta^{AB}_{ij}}$'s to zero.
~It should also be noted that the large chosen values of $M_A \ge 500 \,\, \mathrm{GeV}$
place the Higgs sector of our scenarios in the so called decoupling
regime~\cite{Haber:1989xc},
where the couplings of~$h$ to gauge bosons and fermions are close to
the SM Higgs couplings, and the heavy~$H$ couples like the
pseudoscalar~$A$, and all heavy Higgs bosons are close in mass.
Increasing $M_A$ the heavy
Higgs bosons tend to decouple from low energy physics and the light~$h$
behaves like the SM Higgs boson.
This type of MSSM Higgs sector seems
to be in good agreement with recent LHC
data~\cite{LHCHiggslast}.
We have checked with the code {\tt HiggsBounds}~\cite{higgsbounds}
that the Higgs sector is in agreement with the LHC searches.
Particularly, the so far absence of gluinos at
LHC, forbids too low $M_3$ and, therefore, given the assumed GUT
relation, forbids also a too low $M_2$. Consequently, the
values of $M_2$ and $\mu$ are fixed as to get gaugino masses compatible
with present LHC bounds.
Finally, we have also required that all our points lead to a prediction
of the anomalous magnetic moment of the muon in the MSSM that can fill
the present discrepancy between the Standard Model prediction and the
experimental value. Specifically, we use \citeres{Bennett:2006fi} and
\cite{Davier:2010nc} to extract the size of this discrepancy,
see also \citere{gm2-Jegerlehner}:
\begin{equation}
(g-2)_\mu^{\rm exp}-(g-2)_\mu^{\rm SM}= (30.2 \pm 9.0) \times 10^{-10}.
\label{gminus2}
\end{equation}
We then require that the SUSY contributions from charginos and
neutralinos in the MSSM to one-loop level,
$(g-2)_\mu^{\rm SUSY}$, be within the interval defined by $3 \sigma$
around the central value in \refeq{gminus2}, namely:
\begin{align}
(g-2)_\mu^{\rm SUSY} &\in (3.2 \times 10^{-10},57.2 \times 10^{-10}) ~.
\label{gminus2interval}
\end{align}
\subsubsection{Framework 2}
\label{sec:f2}
In the second framework, several possibilities for the
MSSM parameters have been considered, leading to simple patterns
of SUSY masses with specific
relations among them and where the number of input parameters is strongly
reduced. As in framework 1, the scenarios selected in
framework~2 lead to predictions of $(g-2)_\mu$ and
$M_h$ (for all deltas equal to zero) that are compatible with present data
over a large part of the parameter space. To simplify the analysis of
the limits of the deltas, we will focus in scenarios where the
mass scales of the SUSY QCD sector that are relevant for the NMFV
processes are all set
relative to one mass scale, generically called here $m_{\rm SUSY-QCD}$.
These include the squark soft masses, the trilinear soft
squark couplings and the gluino soft mass, $M_3$.
Similarly, also the mass scales in the SUSY electroweak sector are set
in reference to one common value, $m_{\rm SUSY-EW}$.
These include the slepton soft masses, the gaugino soft
masses, $M_2$ and $M_1$, and the $\mu$ parameter.
It should also be noted that these latter mass
parameters are the relevant ones for $(g-2)_\mu$.
To further simplify the scenarios, we will relate $m_{\rm SUSY-QCD}$ and $m_{\rm SUSY-EW}$.
The remaining
relevant parameter in both NMFV and for the $M_h$ prediction
is $\tan \beta$. Since we wish to explore a wide range in $\tan \beta$,
from 5 to 40, $M_A$ is fixed to $1000 \,\, \mathrm{GeV}$ to ensure the agreement
with the present bounds in the $(\tan \beta, M_A)$ plane from LHC
searches~\cite{CMSpashig12050,Carena:2013qia}.
Finally, to reduce even further the number of
input parameters we will assume again an approximate GUT relation among
the gaugino soft masses, $M_2=2 M_1\, =\,M_3/4$ and the $\mu$ parameter
will be set equal to $M_2$. Regarding the (diagonal)
trilinear couplings, they will
all be set to zero except those of the stop and sbottom sectors, being
relevant for $M_h$, and that will be simplified to $A_t=A_b$.
All parameters are thus either fixed or set relative to $m_{\rm SUSY-QCD}$, where
the different relative settings exhibit certain mass patterns of the
MSSM. These kind of scenarios have
the advantage of reducing considerably the number of input parameters
respect to the MSSM and, consequently, making easier the analysis of
their phenomenological implications.
Similar scenarios have been analyzed in the context of Lepton Flavor
Violation observables in \citere{Arana-Catania:2013nha}.
For the forthcoming numerical analysis we consider the following
specific scenarios:
\begin{itemize}
\item[{\bf (a)}]
\begin{align}
m_{\tilde L}&= m_{\tilde E}=m_{\rm SUSY-EW},\nonumber \\
M_2&= m_{\rm SUSY-EW} \nonumber := 1/2 \, m_{\rm SUSY-QCD},\\
m_{\tilde Q} &= m_{\tilde U}=m_{\tilde D}=m_{\rm SUSY-QCD},\nonumber \\
A_t &= 1.3 \, m_{\rm SUSY-QCD}, \nonumber \\
M_3 &= 2 m_{\rm SUSY-QCD},
\label{Sa}
\end{align}
\item[{\bf (b)}]
\begin{align}
m_{\tilde L} &= m_{\tilde E}=m_{\rm SUSY-EW},\nonumber \\
M_2 &= 1/5 \, m_{\rm SUSY-EW} := 1/10 \, m_{\rm SUSY-QCD}, \nonumber \\
m_{\tilde Q} &= m_{\tilde U}=m_{\tilde D}=m_{\rm SUSY-QCD},\nonumber \\
A_t &= m_{\rm SUSY-QCD}, \nonumber \\
M_3 &= 2/5 \, m_{\rm SUSY-QCD},
\label{Sb}
\end{align}
\item[{\bf (c)}]
\begin{align}
m_{\tilde L} &= m_{\tilde E}=m_{\rm SUSY-EW},\nonumber \\
M_2 &= m_{\rm SUSY-EW} \nonumber := 1/4 \, m_{\rm SUSY-QCD}, \nonumber \\
m_{\tilde Q} &= m_{\tilde U}=m_{\tilde D}=m_{\rm SUSY-QCD},\nonumber \\
A_t &= m_{\rm SUSY-QCD}, \nonumber \\
M_3 &= m_{\rm SUSY-QCD},
\label{Sc}
\end{align}
\item[{\bf (d)}]
\begin{align}
m_{\tilde L} &= m_{\tilde E}=m_{\rm SUSY-EW},\nonumber \\
M_2 &= 1/3 \, m_{\rm SUSY-EW} \nonumber := 1/3 \, m_{\rm SUSY-QCD}, \nonumber \\
m_{\tilde Q} &= m_{\tilde U}=m_{\tilde D}=m_{\rm SUSY-QCD},\nonumber \\
A_t &= m_{\rm SUSY-QCD}, \nonumber \\
M_3 &= 4/3 \, m_{\rm SUSY-QCD}.
\label{Sd}
\end{align}
\end{itemize}
Here we have simplified the notation for the soft sfermion masses, by
using $m_{\tilde L}$ for
$m_{\tilde L}=m_{\tilde L_{1}}=m_{\tilde L_{2}}=m_{\tilde L_{3}}$, etc.
In the forthcoming numerical analysis of the limits of
the deltas within these scenarios, the most relevant parameters
$m_{\rm SUSY-QCD} \equiv m_{\rm SUSY}$ and $\tan \beta$ will be varied within the
intervals:
\begin{align}
1000 \,\, \mathrm{GeV} \leq m_{\rm SUSY} \leq 3000 \,\, \mathrm{GeV}, \quad
5 \leq \tan \beta \leq 40~.
\end{align}
The main results in this framework~2 will be presented in the
($m_{\rm SUSY}$, $\tan \beta$) plane. In the final analysis we will show the
compatibility with $(g-2)_\mu$, but
focus on the consequences of the changes in $M_h$ induced by non-zero
values for the deltas.
\subsubsection{Selected \boldmath{$\ensuremath{\delta^{AB}_{ij}}$} mixings}
Finally, for our purpose in this paper, we need to select the squark
mixings and to set the range of values for the explored
$\ensuremath{\delta^{AB}_{ij}}$'s. In principle, we work in a complete basis, that is we
take into account the full set of 21 $\ensuremath{\delta^{AB}_{ij}}$'s.
However, since the mixing between the first and second/third generation is
already very restricted, we focus here on the deltas that mix only
second and third generation (although our numerical code can handle any
kind of deltas).
For simplicity, we will assume real values for these flavor squark
mixing parameters. Concretely, the scanned interval
in our estimates of NMFV rates will be:
\begin{align}
-1 \le \delta^{AB}_{ij} \le +1
\end{align}
The above scan interval is simply meant to cover all possible ranges. Here
we do not take into account, for instance, constraints
on $\del{LR,RL}{ij}$'s from the requirement of vacuum
stability~\cite{Casas:1996de} or vacuum meta-stability~\cite{Park:2010wf},
which could invalidate large
values for these deltas, corresponding to large ${\cal A}_{ij}$-terms.
\section{The precision observables}
\label{sec:obs}
In this section we briefly review the current status of the precision
observables that we consider in our NMFV analysis.
Since we are mainly interested
in the phenomenological consequences of the flavor mixing between the
third and second generations we will focu
\footnote{
We have checked that electroweak precision observables, where NMFV
effects enter, for instance, via $\Delta\rho$~\cite{mhNMFVearly}, do not
lead to relevant additional constraints on the allowed parameter space. Our results on this constraint are in agreement with \citere{Cao1}.
~on the lightest Higgs boson mass in the (NMFV) MSSM and the
following three B
meson observables: 1) Branching ratio of the $B$ radiative decay \ensuremath{{\rm BR}(B \to X_s \gamma)},
2) Branching ratio of the $B_s$ muonic decay \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)},
and 3) $B_s-{\bar B_s}$ mass difference \ensuremath{\Delta M_{B_s}}. Another
$B$ observable of interest in the present context is \ensuremath{{\rm BR}(B \to X_s l^+ l^-)}.
However, we have not included this in our study, because the
predicted rates in NMFV-SUSY scenarios for this observable are closely
correlated with those from \ensuremath{{\rm BR}(B \to X_s \gamma)}\ due to the dipole
operators dominance in the photon-penguin diagrams mediating \ensuremath{{\rm BR}(B \to X_s l^+ l^-)}\
decays. It implies that the restrictions on the flavor
mixing $\delta^{AB}_{ij}$ parameters from \ensuremath{{\rm BR}(B \to X_s l^+ l^-)}\ are also expected to be correlated with those
from the radiative decays.
The summary of the relevant features for our analysis of these four
observables is given in the following.
\subsection{The lightest Higgs boson mass \boldmath{$M_h$}}
In the
Feynman diagrammatic approach that we are following here, the
higher-order corrected
${\cal CP}$-even Higgs boson masses are derived by finding the
poles of the $(h,H)$-propagator
matrix. The inverse of this matrix is given by
\begin{equation}
\left(\Delta_{\rm Higgs}\right)^{-1}
= - i \left( \begin{array}{cc} p^2 - m_{H,{\rm tree}}^2 + \hat{\Sigma}_{HH}(p^2) & \hat{\Sigma}_{hH}(p^2) \\
\hat{\Sigma}_{hH}(p^2) & p^2 - m_{h,{\rm tree}}^2 + \hat{\Sigma}_{hh}(p^2) \end{array} \right)~.
\label{higgsmassmatrixnondiag}
\end{equation}
Determining the poles of the matrix $\Delta_{\rm Higgs}$ in
\refeq{higgsmassmatrixnondiag} is equivalent to solving
the equation
\begin{equation}
\left[p^2 - m_{h,{\rm tree}}^2 + \hat{\Sigma}_{hh}(p^2) \right]
\left[p^2 - m_{H,{\rm tree}}^2 + \hat{\Sigma}_{HH}(p^2) \right] -
\left[\hat{\Sigma}_{hH}(p^2)\right]^2 = 0\,.
\label{eq:proppole}
\end{equation}
The NMFV parameters enter into the one-loop prediction of the various
(renormalized) Higgs-boson self-energies, where details can be found in
\citere{mhNMFV}. Numerically the results have been obtained using the
code {\tt FeynHiggs}~\cite{feynhiggs,mhiggsAEC}, which contains the complete set of one-loop
NMFV corrections
\footnote{
Not yet taken into account are the logarithmically resummed
corrections~\cite{Mh-logresum}, which could be relevant for the largest values
of $m_{\rm SUSY}$ as analyzed below.
}
The current experimental average for the (SM) Higgs boson mass
is~\cite{pdg-www},
\begin{align}
M_H^{\rm exp} &= 125.6 \pm 0.3 \,\, \mathrm{GeV}~.
\label{MHexp}
\end{align}
The intrinsic theoretical uncertainty is taken to be~\cite{mhiggsAEC,ehowp}
\begin{align}
\deM_h^{\rm th} = \pm 3 \,\, \mathrm{GeV}~,
\label{Mhintr}
\end{align}
and both uncertainties combined give an estimate of the total
uncertainty of $M_h$ in the MSSM.
\subsection{\boldmath{\ensuremath{{\rm BR}(B \to X_s \gamma)}}}
\label{sec:bsg}
For a more detailed description of the inclusion of NMFV effects into the
prediction of $B$-physics observables in general, and for \ensuremath{{\rm BR}(B \to X_s \gamma)}\ in
particular, we refer the reader to \citere{mhNMFV} and references therein.
The relevant effective
Hamiltonian for this decay is given in terms of the Wilson coefficients $C_i$ and operators $O_i$ by:
\noindent \begin{equation}
\mathcal{H}_{\rm eff}=-\frac{4G_{F}}{\sqrt{2}}V_{\rm CKM}^{ts*}V_{\rm CKM}^{tb}
\sum_{i=1}^{8}(C_{i}O_{i}+C'_{i}O'_{i}).
\end{equation}
Where the primed operators can be obtained from the unprimed ones by
replacing $L \leftrightarrow R$.
The complete list of operators can be found, for instance, in
\citere{Gambino:2001ew}.
In the context of SUSY scenarios with the MSSM particle content and
assuming NMFV, only four of these operators are relevant
(we have omitted the color indices here for brevity):
\noindent
\begin{align}
O_{7} &= \frac{e}{16\pi^{2}}m_{b}
\left(\bar{s}_{L}\sigma^{\mu\nu}b_{R}\right)F_{\mu\nu}~, \\
O_{8} & = \frac{g_{3}}{16\pi^{2}}m_{b}
\left(\bar{s}_{L}\sigma^{\mu\nu}T^{a}b_{R}\right)G_{\mu\nu}^{a}~, \\
O'_{7} &= \frac{e}{16\pi^{2}}m_{b}
\left(\bar{s}_{R}\sigma^{\mu\nu}b_{L}\right)F_{\mu\nu}~, \\
O'_{8} &= \frac{g_{3}}{16\pi^{2}}m_{b}
\left(\bar{s}_{R}\sigma^{\mu\nu}T^{a}b_{L}\right)G_{\mu\nu}^{a}.
\end{align}
We have included in our analysis the most relevant loop contributions to
the Wilson coefficient
\footnote{The RGE-running of the Wilson coefficients
is done in two steps: The first one is
from the SUSY scale down to the electroweak scale, and the second
one is from this electroweak scale down to the $B$-physics scale.
For the first step, we use the LO-RGEs for the
relevant Wilson coefficients as in \cite{Degrassi:2000qf} and fix six
active quark flavors in this running.
For the second running we use the NLO-RGEs
as in \cite{Hurth:2003dk} and fix, correspondingly, five active quark
flavors. For the charged Higgs sector, as in \citere{mhNMFV}, we use the
NLO formulas for the Wilson coefficients of
\citere{Borzumati:1998tg}.
, concretely:
1)~loops with Higgs bosons (including the resummation of large $\tan \beta$
effects~\cite{Isidori:2002qe}),
2)~loops with charginos and
3)~loops with gluinos.
It should be noted that, at one loop order, the gluino
loops do not contribute in MFV scenarios, but they are very relevant
(dominant in many cases) in the present NMFV scenarios.
The total branching ratio for this decay is finally estimated by adding
the new contributions from the SUSY and Higgs sectors to the
SM rate. More specifically, we use eq.(42) of \cite{Hurth:2003dk} for
the estimate of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ in terms of the ratios of
the Wilson coefficients $C_{7,8}$ and $C'_{7,8}$ (including all the
mentioned new contributions)
divided by the corresponding $C_{7,8}^{\rm SM}$ in the SM.
For the numerical estimates of \ensuremath{{\rm BR}(B \to X_s \gamma)}\ (and the other $B$-physics
observables) we use the {\tt FORTRAN}
subroutine {\tt BPHYSICS} (modified as to include the contributions from
$C'_{7,8}$ which were not included in its original version) included in
the {\tt SuFla} code, that incorporates all the above mentioned
ingredients~\cite{sufla}.
In order to obtain the updated limits on the NMFV parameters, the
following experimental measurement of
\ensuremath{{\rm BR}(B \to X_s \gamma)}~\cite{hfag:rad
\footnote{We have added the various contributions to the experimental
error in quadrature.
, and its prediction within the SM~\cite{Misiak:2009nr} have been used:
\begin{align}
\label{bsgamma-exp}
\ensuremath{{\rm BR}(B \to X_s \gamma)}_{\rm exp} &= (3.43 \pm 0.22)\times10^{-4}~, \\
\ensuremath{{\rm BR}(B \to X_s \gamma)}_{\rm SM} &= (3.15 \pm 0.23)\times10^{-4}~.
\label{bsgamma-SM}
\end{align}
\subsection{\boldmath{\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}}}
\label{sec:bmm}
The relevant effective Hamiltonian for this process is
\cite{Chankowski:2000ng,Bobeth:2002ch}:
\noindent
\begin{equation}
\mathcal{H}_{\rm eff}=-\frac{G_{F}\alpha}{\sqrt{2} \pi}
V_{\rm CKM}^{ts*}V_{\rm CKM}^{tb}\sum_{i} (C_{i}O_{i}+C'_{i} O'_{i}),
\end{equation}
where the operators $O_i$ are given by:
\begin{align}
{O}_{10} &= \left(\bar{s}\gamma^{\nu}P_Lb\right)
\left(\bar{\mu}\gamma_{\nu}\gamma_5\mu\right)~,
& {O}_{10}^{\prime} &= \left(\bar{s}\gamma^{\nu}P_Rb\right)
\left(\bar{\mu}\gamma_{\nu}\gamma_5\mu\right)~, \nonumber \\
{O}_{S} &= m_b\left(\bar{s}P_Rb\right)
\left(\bar{\mu}\mu\right)~,
& {O}_{S}^{\prime} &= m_s\left(\bar{s} P_Lb \right)
\left(\bar{\mu}\mu\right)~, \nonumber \\
{O}_{P} &= m_b\left(\bar{s} P_Rb \right)
\left(\bar{\mu}\gamma_5\mu\right)~,
& {O}_{P}^{\prime} &= m_s\left(\bar{s} P_Lb \right)
\left(\bar{\mu}\gamma_5\mu\right)~.
\label{bsm:Ops}
\end{align}
We have again omitted the color indices here for brevity.
The prediction for the decay rate is expressed by:
\begin{align}
\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)} &= \frac{G_F^2\alpha^2 m_{B_s}^2 f_{B_s}^2\tau_{B_s}}{64 \pi^3}
\lvert V_{\rm CKM}^{ts*}V_{\rm CKM}^{tb}\rvert^2\sqrt{1-4\hat{m}_{\mu}^2}
\nonumber \\
&\times\left[\left(1-4\hat{m}_{\mu}^2\right)\lvert F_S\rvert^2
+\lvert F_P+2\hat{m}_{\mu}^2 F_{10}\rvert^2\right],
\label{bsm:br}
\end{align}
where $\hat{m}_{\mu}=m_{\mu}/m_{B_s}$ and the
$F_i$ are given by
\begin{align}
F_{S,P}&=m_{B_s}\left[\frac{C_{S,P}m_b-C_{S,P}^{\prime}m_s}{m_b+m_s}\right],
&F_{10}=C_{10}-C_{10}^{\prime}.
\nonumber
\end{align}
In the context of NMFV MSSM, with no preference for large $\tan \beta$
values, there are in general three types of one-loop diagrams that contribute
to the previous $C_i$ Wilson coefficients for this $B_s \to \mu^+ \mu^-$
decay:
1)~Box diagrams,
2)~$Z$-penguin diagrams and
3)~neutral Higgs boson $\phi$-penguin diagrams, where $\phi$ denotes the
three neutral MSSM Higgs bosons, $\phi = h, H, A$ (again large resummed
$\tan \beta$ effects have been taken into account).
In our numerical estimates
we have included what are known to be the dominant contributions to
these three types of diagrams \cite{Chankowski:2000ng}: chargino
contributions to box and $Z$-penguin diagrams and chargino and gluino
contributions to $\phi$-penguin diagrams.
The present experimental value
for this observable \cite{Chatrchyan:2013bka,Aaij:2013aka},
and the prediction within the SM \cite{Buras:2012ru} are given by
\begin{align}
\label{bsmumu-exp}
\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}_{\rm exp} &= (3.0^{+1.0}_{-0.9})\times 10^{-9}~, \\
\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}_{\rm SM} &= (3.23\pm 0.27)\times 10^{-9}~.
\label{bsmumu-SM}
\end{align}
\subsection{\boldmath{\ensuremath{\Delta M_{B_s}}}}
\label{sec:dmbs}
The relevant effective Hamiltonian for $B_s-{\bar B_s}$ mixing and, hence, for
the $B_s/{\bar B_s}$ mass difference \ensuremath{\Delta M_{B_s}}\ is:
\begin{equation}
\mathcal{H}_{\rm eff}=
\frac{G_F^2}{16\pi^2}M_W^2
\left(V_{\rm CKM}^{tb*}{}V_{\rm CKM}^{ts}\right)^2
\sum_{i}C_i O_i.
\label{Ham}
\end{equation}
Within the NMFV MSSM the following operators are relevant (now including
the color indices explicitly):
\begin{align}
\label{SMOps}
O^{VLL} &= (\bar{b}^{\alpha}\gamma_{\mu}P_L s^{\alpha})
(\bar{b}^{\beta}\gamma^{\mu}P_L s^{\beta})~, \\
\label{Ops1}
O^{LR}_{1} &= (\bar{b}^{\alpha}\gamma_{\mu}P_L s^{\alpha})
(\bar{b}^{\beta}\gamma^{\mu}P_R s^{\beta})~,
& O^{LR}_{2} &= (\bar{b}^{\alpha}P_L s^{\alpha})
(\bar{b}^{\beta}P_R s^{\beta})~, \\
O^{SLL}_{1} &= (\bar{b}^{\alpha}P_L s^{\alpha})
(\bar{b}^{\beta}P_L s^{\beta}),
& O^{SLL}_{2} &= (\bar{b}^{\alpha}\sigma_{\mu\nu}P_L s^{\alpha})
(\bar{b}^{\beta}\sigma^{\mu\nu}P_L s^{\beta})~,
\label{Ops2}
\end{align}
and the corresponding operators $O^{VRR}$ and
$O^{SRR}_{i}$ that can be obtained by replacing $P_L \leftrightarrow P_R$
in~\refeq{SMOps} and~\refeq{Ops2}.
The mass difference $\Delta M_{B_s}$ is then evaluated by taking the matrix
element
\begin{align}
\ensuremath{\Delta M_{B_s}} &= 2\lvert\langle\bar{B}_s\lvert\mathcal{H}_{\rm eff}
\rvert B_s\rangle\rvert,
\label{delmb}
\end{align}
where $\langle\bar{B}_s\lvert\mathcal{H}_{\rm eff}\rvert B_s\rangle$ is given
by
\begin{align}
\langle\bar{B}_s\lvert\mathcal{H}_{\rm eff}\rvert B_s\rangle=&
\frac{G_F^2}{48\pi^2}M_W^2 m_{B_s} f^2_{B_s}
\left(V_{\rm CKM}^{tb*} V_{\rm CKM}^{ts}\right)^2
\sum_{i}P_i C_i\left(\mu_W\right).
\label{matel}
\end{align}
Here $m_{B_s}$ is the $B_s$ meson mass, and
$f_{B_s}$ is the $B_s$ decay constant.
The coefficients $P_i$ contain the effects due to RGE running between
the electroweak scale $\mu_W$ and $m_b$ as well as the relevant hadronic
matrix element. We use the coefficients $P_i$ from the lattice
calculation \cite{Becirevic:2001xt}:
\begin{align}
P^{VLL}_1=&0.73,
&P^{LR}_1=&-1.97,
&P^{LR}_2=&2.50,
&P^{SLL}_1=&-1.02,
&P^{SLL}_2=&-1.97.
\label{pcoef}
\end{align}
In the context of the NMFV MSSM, besides the SM contributions,
there are in general three types of one-loop diagrams that contribute:
1)~Box diagrams,
2)~$Z$-penguin diagrams and
3)~double Higgs-penguin diagrams (again including the resummation of
large $\tan \beta$ enhanced effects).
In our numerical estimates we have included what are known to be the
dominant contributions to these three types of diagrams in scenarios
with non-minimal flavor violation (for a review see, for instance,
\cite{Foster:2005wb}): gluino contributions to box
diagrams, chargino contributions to box and $Z$-penguin diagrams, and
chargino and gluino contributions to double $\phi$-penguin diagrams.
For the numerical estimates we have modified the {\tt BPHYSICS} subroutine
included in the {\tt SuFla} code~\cite{sufla} which incorporates all the
ingredients that we have pointed out above, except the contributions
from gluino boxes which we have added, see \citere{mhNMFV} for a
detailed discussion on these contributions.
The experimental result~\cite{hfag:pdg} and the SM
prediction (using the NLO expression of \cite{Buras:1990fn} and the
error estimate of \cite{Golowich:2011cx}) used to obtain our updated
bounds on the NMFV parameters are given by:
\begin{align}
\label{deltams-exp}
{\ensuremath{\Delta M_{B_s}}}_{\rm exp} &= (116.4 \pm 0.5) \times 10^{-10} \,\, \mathrm{MeV}~, \\
{\ensuremath{\Delta M_{B_s}}}_{\rm SM} &= (117.1^{+17.2}_{-16.4}) \times 10^{-10} \,\, \mathrm{MeV}~.
\label{deltams-SM}
\end{align}
\section{Numerical results}
\label{sec:results}
In this section we present our numerical results. First we analyze
the six scenarios of framework~1, exploring $\ensuremath{\delta^{AB}_{ij}} \neq 0$,
with respect to the flavor observables
and derive the corresponding bounds on the deltas.
In a second step we
will show which corrections to the Higgs boson masses can be found in
these scenarios, but bounds on the deltas are only derived from ``too
large'' corrections to the lightest Higgs boson mass, as will be defined
and discussed below. These ``too large'' corrections to $M_h$ indicate
that the light Higgs boson mass itself can serve as an additional
observable constraining further the deltas, which can therefore
complement the previous constraints from $B$-physics observables.
The heavy Higgs boson masses, on the other hand,
depend (to a good approximation) linearly on $M_A$ and can
thus easily avoid bounds by an appropriate choice of $M_A$.
Finally, having the new restrictions from $M_h$ in mind,
we then focus next on the simple scenarios of framework 2, where we have
performed a systematic study in the ($m_{\rm SUSY}$, $\tan \beta$) plane to conclude
on the maximum allowed deltas that are compatible with both the
$B$-physics data and the
present Higgs mass value. In this analysis we will consider also the
compatibility with the $(g-2)_\mu$ data.
\subsection{Framework 1: flavor observables}
\begin{figure}[htb!]
\centering
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-81-deltassimult-bsgs}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-82-deltassimult-bsgs}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-83-deltassimult-bsgs}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-84-deltassimult-bsgs}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-85-deltassimult-bsgs}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-86-deltassimult-bsgs}
\caption{Sensitivity to the NMFV deltas in \ensuremath{{\rm BR}(B \to X_s \gamma)}\ for the points S1\ldots S6.
The experimental allowed $3\sigma$ area is the horizontal
colored band. The SM prediction and the theory uncertainty $\ensuremath{\Delta^{\rm theo}}(\ensuremath{{\rm BR}(B \to X_s \gamma)})$
(red bar) is displayed on the right axis.}
\label{fig:bsg}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-81-deltassimult-sll_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-82-deltassimult-sll_parad}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-83-deltassimult-sll_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-84-deltassimult-sll_parad}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-85-deltassimult-sll_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-86-deltassimult-sll_parad}
\caption{Sensitivity to the NMFV deltas in \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ for the points S1\ldots S6.
The experimental allowed $3\sigma$ area is the horizontal
colored band. The SM prediction and the theory uncertainty $\ensuremath{\Delta^{\rm theo}}(\ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)})$
(red bar) is displayed on the right axis.}
\label{fig:bmm}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-81-deltassimult-deltabs_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-82-deltassimult-deltabs_parad}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-83-deltassimult-deltabs_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-84-deltassimult-deltabs_parad}\\
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-85-deltassimult-deltabs_parad}
\includegraphics[width=6.75cm,height=8cm,angle=270]{plot-86-deltassimult-deltabs_parad}
\caption{Sensitivity to the NMFV deltas in \ensuremath{\Delta M_{B_s}}\ for the points S1\ldots S6.
The experimental allowed $3\sigma$ area is the horizontal
colored band. The SM prediction and the theory uncertainty $\ensuremath{\Delta^{\rm theo}}(\ensuremath{\Delta M_{B_s}})$
(red bar) is displayed on the right axis.}
\label{fig:dmbs}
\end{figure}
In \reffis{fig:bsg} -- \ref{fig:dmbs} we show the results for the
three flavor observables discussed in \refses{sec:bsg} --
\ref{sec:dmbs}. The results are shown for the points S1\ldots S6, see
\refta{tab:spectra}, where the various \ensuremath{\delta^{AB}_{ij}}\ are varied individually.
We have also included in the right vertical axis of these figures, for
comparison, the respective SM prediction in \refeqs{bsgamma-SM},
(\ref{bsmumu-SM}), and (\ref{deltams-SM}). The red error bars displayed are
the corresponding $3\,\sigma$ SM uncertainties
(called $\Delta^{\rm theo}$).
The shadowed horizontal bands in all cases, \ensuremath{{\rm BR}(B \to X_s \gamma)}\ , \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and
\ensuremath{\Delta M_{B_s}}, are their corresponding experimental measurements
in \refeqs{bsgamma-exp}, (\ref{bsmumu-exp}) and (\ref{deltams-exp}),
expanded with $3\,\sigma_{\rm exp}$ errors.
In order to assess the total uncertainty the SM errors are also applied to
the MSSM predictions. If this error bar is outside the experimental band
the point can be regarded as excluded by the experimental
measurement. It should be noted that the theory uncertainties can be
larger in the MSSM than in the SM. However, estimates are much more
complicated than in the SM and strongly dependent on the chosen SUSY
parameters. Therefore we simply apply the SM uncertainty with $3\,\sigma$
errors.
Regarding the explored intervals for the deltas in the following
\reffis{fig:bsg} -- \ref{fig:dmbs}, these will be $-1 \leq \ensuremath{\delta^{AB}_{ij}}
\leq 1$, as discussed above. However, in some cases these
intervals are smaller: in computing the MSSM spectra with
non-vanishing $\ensuremath{\delta^{AB}_{ij}}$ the code does not accept points that lead either
to too low MSSM masses, excluded by experiment, or even non-physical
negative squared masses. This is, for instance, the case of
$\del{LR}{ij}$ and $\del{RL}{ij}$ with $ij=sb$ and $ij=ct$ that, as we
can see in \reffis{fig:bsg} -- \ref{fig:dmbs}, are explored in smaller
intervals since outside of them they lead to
negative squared scalar masses. In particular, the contributions from
the deltas leading to too low $M_h$ will be studied further in the
following sections.
The analysis for \ensuremath{{\rm BR}(B \to X_s \gamma)}\ is shown in \reffi{fig:bsg} in the scenarios
S1\ldots S6. In the MFV case (i.e. for all $\ensuremath{\delta^{AB}_{ij}} = 0$) we see
that all points, except S4, are in agreement
with experimental data. Only a very small variation with
\del{LR}{ct}, \del{RL}{ct}, \del{RR}{ct}, \del{RR}{sb} (except for S4)
is observed. A clear dependence on \del{LL}{23} can be seen, placing bounds of
\order{0.1} on this NMFV parameter in all five scenarios, S1, S2, S3, S5
and S6. A very strong
variation with \del{LR}{sb} and \del{RL}{sb} is found, which are
restricted to very small values $\leq$ \order{0.01} by the \ensuremath{{\rm BR}(B \to X_s \gamma)}\ measurement.
In scenario S4 the strong variation with \del{LL}{23} or
\del{LR}{sb} can bring the prediction into agreement with the
experimental data. Turning the argument around, the scenario S4, which
appears to be excluded by the \ensuremath{{\rm BR}(B \to X_s \gamma)}\ measurement is actually a valid
scenario for certain values of \del{LL}{23} and \del{LR}{sb}.
The results for \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ are shown in \reffi{fig:bmm} for the six
scenarios. All \ensuremath{\delta^{AB}_{ij}}\ show a relatively small impact, except for
\del{LL}{23}. From these plots we find the following allowed
intervals for \del{LL}{23}: $S1:(-0.3,0.7)$ , $S2:(-0.3,0.8)$,
$S3:(-0.1,0.2)$, $S4:(-0.3,0.3)$, $S6:(-0.3,0.8)$. Therefore, bounds
on this parameter ranging between $\sim - 0.1$ and
$\sim + 0.8$ can be set in all scenarios except in S5 where we do not
get any constraint. This scenario
is characterized by a very large value of $M_A = 1000 \,\, \mathrm{GeV}$ and a
relatively small value of $\tan \beta = 10$, leading to a strong suppression of
the contributions to \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}.
The predictions for \ensuremath{\Delta M_{B_s}}\ in the six scenarios are shown in
\reffi{fig:dmbs}. While the experimental precision is very high the
theoretical error is quite large, and the
bounds on the \ensuremath{\delta^{AB}_{ij}}\ are mainly given by the SM uncertainty in the
\ensuremath{\Delta M_{B_s}}\ prediction. All six scenarios for all $\ensuremath{\delta^{AB}_{ij}} = 0$ are in
agreement with the experimental data, once the SM uncertainty is taken
into account. Except for S4, which is sensitive to all deltas, the
other points are nearly insensitive to \del{RR}{ct}, \del{LR}{ct} and
\del{RL}{ct}, therefore we do not get any additional bound for them in
the allowed range from this observable.
An important variation can be observed for
\del{LR}{sb} and \del{RL}{sb}. However, due to the MSSM particle mass
restrictions commented above which shortened the allowed intervals,
hardly any new bounds are placed by \ensuremath{\Delta M_{B_s}}, except in S4
and S6. Some sensitivity is found for \del{RR}{sb}, especially in S4 and
S6 where $|\del{RR}{sb}|$ is bounded by $\leq \order{0.5}$. The
strongest variation is found for \del{LL}{23}, where due to the
particular 'W-shape' dependence, both
intermediate and large values can be excluded.
The overall allowed intervals for the seven \ensuremath{\delta^{AB}_{ij}}\ in the six
scenarios and considering the three observables together, \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}
and \ensuremath{\Delta M_{B_s}}, can be found in \refta{tab:boundsS1S6}. From this table we
then conclude on the strongest bounds that can be
obtained from the combination of all three $B$-physics observables.
As a general comment, the main restrictions to the deltas come from
\ensuremath{{\rm BR}(B \to X_s \gamma)}\ and in some cases from \ensuremath{\Delta M_{B_s}}\ and not yet from the young
measurement \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}.
The most restricted deltas are \del{LR}{sb} and \del{RL}{sb} that can
reach values at most of \order{0.01}, then \del{LL}{23}, \del{LR}{ct}
and \del{RL}{ct} that can be at most of \order{0.1}, with the first one
being slightly more restricted than the last two, and finally the less
restricted deltas are \del{RR}{ct} and \del{RR}{sb} that in general can
reach up to the largest explored values of \order{1}. Special
attention deserves scenario S4, where, as mentioned above, setting
$\ensuremath{\delta^{AB}_{ij}} = 0$ leads to experimentally excluded predictions. Only non-zero
values of \del{LR}{sb} can reconcile this scenario with experimental
data. Consequently, assuming only {\em one} $\ensuremath{\delta^{AB}_{ij}}$ different from zero
leads to an ``excluded'' scenario for all the other $\ensuremath{\delta^{AB}_{ij}}$ as shown
in \refta{tab:boundsS1S6}.
It can also be seen that larger constraints in the ``$sb$ sector'' than in
the ``$ct$ sector'' are obtained, since the $B$-physics
observables are in general
more sensitive to mixing among $b$-type squarks. We also see that the
\del{LR}{} become more restricted than the others, since they involve the
trilinear couplings that provide in general large corrections.
Regarding the comparison of our results with previous studies, we
conclude that the bounds on the squark mixing deltas that we find here
for the scenarios S1-S6 are more relaxed than in the set of benchmark
scenarios that were analyzed in \cite{mhNMFV} before the LHC started
operation. The scenarios investigated in the pre-LHC era contained
relatively light scalar quarks (now excluded), leading to relatively
large radiative corrections from NMFV effects. After the so far
unsuccessful search for beyond SM physics at the LHC, scalar quark
masses (in particular those of the first and second generation) have
substantially higher lower bounds. Benchmark scenarios that take this
into account (as our S1-S6) naturally permit larger values for the NMFV
deltas.
\renewcommand{\arraystretch}{1.1}
\begin{table}[H]
\begin{center}
\resizebox{9.0cm}{!} {
\begin{tabular}{|c|c|c|} \hline
& & Total allowed intervals \\ \hline
$\delta^{LL}_{23}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.27:0.28) \\ (-0.23:0.23) \\ (-0.12:0.06) (0.17:0.19) \\ excluded \\ (-0.83:-0.78) (-0.14:0.14) \\ (-0.076:0.14) \end{tabular} \\ \hline
$\delta^{LR}_{ct}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular}
& \begin{tabular}{c}
(-0.27:0.27) \\ (-0.27:0.27) \\ (-0.27:0.27) \\ excluded \\ (-0.22:0.22) \\ (-0.37:0.37) \end{tabular} \\ \hline
$\delta^{LR}_{sb}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.0069:0.014) (0.12:0.13) \\ (-0.0069:0.014) (0.11:0.13) \\ (-0.0069:0.014) (0.11:0.13) \\ (0.076:0.12) (0.26:0.30) \\ (-0.014:0.021) (0.17:0.19) \\ (0:0.0069) (0.069:0.076) \end{tabular} \\ \hline
$\delta^{RL}_{ct}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c}
(-0.27:0.27) \\ (-0.27:0.27) \\ (-0.27:0.27) \\ excluded \\ (-0.22:0.22) \\ (-0.37:0.37) \end{tabular}
\\ \hline
$\delta^{RL}_{sb}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c} (-0.034:0.034) \\ (-0.034:0.034) \\ (-0.034:0.034) \\ excluded \\ (-0.062:0.062) \\ (-0.021:0.021) \end{tabular}
\\ \hline
$\delta^{RR}_{ct}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} & \begin{tabular}{c}
(-0.99:0.99) \\ (-0.99:0.99) \\ (-0.98:0.97) \\ excluded \\ (-0.99:0.99) \\ (-0.96:0.94) \end{tabular} \\ \hline
$\delta^{RR}_{sb}$ & \begin{tabular}{c} S1 \\ S2 \\ S3 \\ S4 \\ S5 \\ S6 \end{tabular} &
\begin{tabular}{c} (-0.96:0.96) \\ (-0.96:0.96) \\ (-0.96:0.94) \\ excluded \\ (-0.97:0.97) \\ (-0.97:-0.94) (-0.63:0.64) (0.93:0.97)
\end{tabular} \\ \hline
\end{tabular}}
\end{center}
\caption{Present allowed intervals on the squark mixing parameters
$\delta^{AB}_{ij}$ for the selected S1-S6 MSSM points defined in
\refta{tab:spectra}.
}
\label{tab:boundsS1S6}
\end{table}
\renewcommand{\arraystretch}{1.55}
\subsection{Framework 1: effects on Higgs boson masses}
In this section we discuss the one-loop NMFV effects on the Higgs boson
masses. A more detailed description of the computation of these
one-loop NMFV effects in terms of one-loop diagrams and the
corresponding corrections to the involved self-energies can be found
in \citere{mhNMFV}. We are interested here mainly in the differences
between the predictions within NMFV and MFV.
We show, in \reffis{fig:h0}, \ref{fig:H0} and \ref{fig:Hp},
\begin{align}
\Delta m_\phi := M_\phi^{\rm NMFV} - M_\phi^{\rm MFV}, \quad \phi = h, H, H^\pm
\label{dmh}
\end{align}
as a function of \ensuremath{\delta^{AB}_{ij}}\ in the scenarios S1\ldots S6.
We start our investigation with $\Dem_h$ in \reffi{fig:h0}.
Bounds on the \ensuremath{\delta^{AB}_{ij}}\ can in principle only be placed by the $M_h$
prediction, since this is the only mass parameter that has been
measured experimentally so far.
It should be noted that the value of $M_h^{\rm NMFV}$ depends
strongly on the MFV SUSY parameters, in particular on $X_t$ (where
$\mtX_t$ is the off-diagonal entry in the scalar top mass matrix).
Consequently, delta values that produce an $M_h^{\rm NMFV}$ value
slightly outside the allowed range, see \refeqs{MHexp},
(\ref{Mhintr}), could be brought in agreement with
experimental data by a small change in the scenario (e.g.\ by slightly
changing the $X_t$ parameter).
As can be seen in \reffi{fig:h0},
a negligible variation is found for \del{RR}{sb} in all scenarios.
An enhancement of $M_h$ by up to $1 \,\, \mathrm{GeV}$ is found for \del{LL}{23} and
\del{RR}{ct} once the largest considered values of \order{1} are
reached. However, whereas these are possible for \del{RR}{ct}, such
large values are excluded in the \del{LL}{23} case, as we have seen in
\refta{tab:boundsS1S6}. The only
exception here is scenario S4, where \del{LL}{23} and \del{RR}{ct} lead
to a sizable reduction of $M_h$ once values larger than $\pm 0.5$ are
reached. The remaining \del{LR,RL}{ij} have a larger impact on the $M_h$
prediction. Again the corresponding trilinear couplings involved
play a relevant role here.
Small \del{LR,RL}{ct} values lead to
an enhancement of up to $1 \,\, \mathrm{GeV}$, and
larger values of \order{0.1} yield a large negative contribution to
$M_h$ (i.e.\ an effect similar to the dependence on $X_t$ can be
observed). Consequently, bounds of \order{0.2} can be placed on
\del{LR,RL}{ct}, predicting $M_h$ values that are
outside the allowed range, see \refeqs{MHexp}, (\ref{Mhintr}).
Similar bounds can be derived for
\del{LR,RL}{bs}, however, these are in general weaker than the
previous bounds found from the $B$-physics observables, as can be
seen in \refta{tab:boundsS1S6}.
The strong sensitivity to $LR$ and $RL$ parameters can be
understood due to the relevance of the ${\cal A}_{ij}$-terms in these
Higgs mass corrections.
It can be seen in the Feynman rules (i.e.\ see the coupling of two squarks
and one/two Higgs bosons in Appendix~A of \citere{mhNMFV}) that the
${\cal A}_{ij}$-terms enter directly into the couplings, and in some cases,
as in the couplings of down-type squarks to the ${\cal CP}$-odd Higgs boson,
enhanced by $\tan \beta$. Therefore, considering the relationship between the
${\cal A}_{ij}$-terms and these $LR$ and $RL$ parameters, as is shown in
\refeqs{eq:SCKM-entries}, (\ref{v2Au}) and (\ref{v1Ad}),
the strong sensitivity to these parameters can be
understood. A similar strong sensitivity to $\delta^{LR}_{ct}$ in $\Dem_h$
has been found in \cite{Cao1}.
The predictions for $\Dem_H$ and $\DeM_{H^\pm}$ are shown in \reffis{fig:H0}
and \ref{fig:Hp}. In general, only \del{LR}{sb} and \del{RL}{sb} lead to
sizable effects in $M_H$ and $M_{H^\pm}$, where large (negative for $M_H$,
and both negative and positive for $M_{H^\pm}$) contributions
are found for delta values exceeding $\sim 0.05$. However, since
these masses are mainly determined by the overall MSSM Higgs boson mass
scale, $M_A$, no strong conclusions (or bounds stronger than from the
$M_h$ prediction) can be drawn.
On the other hand, these corrections will become relevant {\em after}
a possible discovery of these heavy Higgs bosons.
\begin{figure}[htb!]
\centering
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-81-deltassimult-masas-h0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-82-deltassimult-masas-h0}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-83-deltassimult-masas-h0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-84-deltassimult-masas-h0}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-85-deltassimult-masas-h0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-86-deltassimult-masas-h0}
\caption{One-loop corrections to $M_h$ in the scenarios S1\ldots S6.
}
\label{fig:h0}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-81-deltassimult-masas-H0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-82-deltassimult-masas-H0}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-83-deltassimult-masas-H0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-84-deltassimult-masas-H0}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-85-deltassimult-masas-H0}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-86-deltassimult-masas-H0}
\caption{One-loop corrections to $M_H$ in the scenarios S1\ldots S6.
}
\label{fig:H0}
\end{figure}
\begin{figure}[htb!]
\centering
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-81-deltassimult-masas-Hp}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-82-deltassimult-masas-Hp}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-83-deltassimult-masas-Hp}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-84-deltassimult-masas-Hp}\\
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-85-deltassimult-masas-Hp}
\includegraphics[width=7.10cm,height=8cm,angle=270]{plot-86-deltassimult-masas-Hp}
\caption{One-loop corrections to $M_{H^\pm}$ in the scenarios S1\ldots S6.
}
\label{fig:Hp}
\end{figure}
\subsection{Framework 2:}
The main goal of this part is to investigate how the upper
bounds on the deltas can be placed from the corrections induced for the
light MSSM Higgs boson mass.
In order to explore the variation of these bounds for different choices
in the MSSM parameter space, we investigate the four qualitatively
different scenarios {\bf (a)}, {\bf (b)}, {\bf (c)} and {\bf (d)}
defined in \refeqs{Sa},
(\ref{Sb}), (\ref{Sc}) and (\ref{Sd}), respectively.
As explained above, the idea is to explore generic scenarios that are
compatible with present data, in particular with the measurement of
a Higgs boson mass, which we interpret as the mass of the light
${\cal CP}$-even Higgs boson in the MSSM (for all $\ensuremath{\delta^{AB}_{ij}} = 0$), and the present
experimental measurement of $(g-2)_\mu$. Taking these experimental
results into account, we have re-analyzed the full set of bounds for
the single deltas that are
extracted from the requirement that the corrections to $M_h$ do not
exceed $125.6 \pm 5 \,\, \mathrm{GeV}
\footnote{This is, allowing for a slightly larger
interval according to our discussion after \refeq{dmh}.}
~as a function of the
two most relevant parameters in our framework 2: the generic
SUSY mass scale $m_{\rm SUSY}$ $(\equiv m_{\rm SUSY-QCD})$ and $\tan \beta$.
In order to find $M_h$ around $125.6 \,\, \mathrm{GeV}$ for $\ensuremath{\delta^{AB}_{ij}} = 0$ the
scale $m_{\rm SUSY-QCD}$ as well as the trilinear couplings have been
chosen to sufficiently high values, see \refse{sec:f2}.
Alternatively one could choose
scenarios with a light Higgs boson mass {\em not} in agreement with the
experimental data and explore the regions of $\ensuremath{\delta^{AB}_{ij}}$ that reconcile
the $M_h$ prediction with the experimental data. However, we will not
pursue this alternative here.
We present the numerical results of our analysis in framework~2
in \reffi{msusytb-LRct}, where we restrict ourselves to the
analysis of \del{LR}{ct} and \del{RL}{ct}, which are the only parameters showing a strong
impact on $M_h$, apart from \del{LR}{sb} and \del{RL}{sb} that are
strongly restricted by $B$-physics observables, see the
previous subsection. Furthermore, almost identical
results are obtained for \del{LR}{ct} and \del{RL}{ct}, and
consequently, we restrict ourselves to one of those parameters.
In each plot we show the resulting contourlines in the
($m_{\rm SUSY}$, $\tan \beta$) plane of maximum allowed value of $|\del{LR}{ct}|$,
i.e.\ the ones that do not lead to contributions to $M_h$
outside $125.6 \pm 5 \,\, \mathrm{GeV}$.
The shaded areas in pink are the regions leading to a
$(g-2)_\mu^{\rm SUSY}$ prediction, from the SUSY one-loop contributions,
in the allowed interval of $(3.2,57.2) \times 10^{-10}$. The
interior pink dashed contourline corresponds to $(g-2)_\mu^{\rm SUSY}$
exactly at the central value of the discrepancy
$(g-2)_\mu^{\rm exp}-(g-2)_\mu^{\rm SM}=30.2 \times 10^{-10}$.
As in the previous framework~1, we use here again
{\tt FeynHiggs}~\cite{feynhiggs,mhiggsAEC} to evaluate
$M_h$ and {\tt SPHENO}~\cite{Porod:2003um} to evaluate
$(g-2)_\mu$ (where {\tt FeynHiggs}\ gives very similar results).
Due to the different relations between the SUSY-QCD and the SUSY-EW
scales in our
four scenarios the pink shaded areas differ substantially in the four
plots. In particular in scenario {\bf (d)}, where we have set
$m_{\rm SUSY-EW} := m_{\rm SUSY-QCD}$ only relatively small values of $m_{\rm SUSY}$ yield a good
prediction of $(g-2)_\mu^{\rm SUSY}$.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=7.10cm,height=8cm]{plot-deltalrctmax-msusytanb-91}
\includegraphics[width=7.10cm,height=8cm]{plot-deltalrctmax-msusytanb-90}\\[3em]
\includegraphics[width=7.10cm,height=8cm]{plot-deltalrctmax-msusytanb-92}
\includegraphics[width=7.10cm,height=8cm]{plot-deltalrctmax-msusytanb-93}
\end{center}
\caption{ Contourlines in the
($m_{\rm SUSY}$, $\tan \beta$) plane of maximum squark mixing
$|\del{LR}{ct}|_{\rm max}$ that are allowed by the requirement that
the correction to $M_h$ does not exceed $\pm 5 \,\, \mathrm{GeV}$
for the scenarios {\bf (a)}, {\bf (b)}, {\bf (c)} and
{\bf (d)} of our framework 2 The shaded (pink)
areas are the regions leading to a $(g-2)_\mu^{\rm SUSY}$
prediction in the $(3.2,57.2) \times 10^{-10}$ interval. The interior
pink dashed contourline corresponds to
$(g-2)_\mu^{\rm SUSY}$ exactly
at the central value of the discrepancy
$(g-2)_\mu^{\rm exp}-(g-2)_\mu^{\rm SM}=30.2 \times 10^{-10}$ .
}
\label{msusytb-LRct}
\vspace{1em}
\end{figure}
One can observe in \reffi{msusytb-LRct} that the bounds on
$|\del{LR}{ct}|$
depend only weakly on the chosen scenario, such that they can be
regarded as relatively general. For $m_{\rm SUSY} \sim 1 \,\, \mathrm{TeV}$ bounds around
$|\del{LR}{ct}| \lsim 0.28$ are found, whereas for $m_{\rm SUSY} \sim 3 \,\, \mathrm{TeV}$
only $|\del{LR}{ct}| \lsim 0.12$ is allowed. For most of the parameter
space the results are nearly independent of $\tan \beta$.
Only for $\tan \beta \lsim 7$
smaller bounds for smaller $m_{\rm SUSY}$ values are reached.
The results are consistent with previous findings, i.e.\ large SUSY mass
scales, leading to larger intergenerational mixing terms (and in
particular $A$-terms) lead to larger effects and thus to smaller allowed
\ensuremath{\delta^{AB}_{ij}}. Comparing the obtained contours, which depend on $m_{\rm SUSY-QCD}$, with
the $(g-2)_\mu$ preferred regions, which depend on $m_{\rm SUSY-EW}$,
slightly
smaller $|\del{LR}{ct}|_{\rm max}$ values as in {\bf (c)} or
slightly larger ones as in {\bf (d)} are
favored. However, this just reflects the choice of the hierarchy between
these two fundamental mass scales used in the respective scenario.
\section{Conclusions}
\label{sec:conclusions}
We presented an up-to-date comparison of the predictions for flavor and
Higgs observables based on NMFV parameters in the MSSM with the current
experimental data.
The flavor observables include \ensuremath{{\rm BR}(B \to X_s \gamma)}, \ensuremath{{\rm BR}(B_s \to \mu^+\mu^-)}\ and \ensuremath{\Delta M_{B_s}}. In the Higgs
sector we evaluated the corrections to the light and heavy
${\cal CP}$-even Higgs masses as well as to the charged Higgs boson mass.
Within the MSSM the calculations were performed at the full one-loop
level with the full (s)quark flavor structure, i.e.\ not relying on the
mass insertion or other approximations.
In the first part we analyzed six representative scenarios which are in
agreement with current bounds on the SUSY and Higgs searches at the
LHC. We derived the most up-to-date bounds on $\ensuremath{\delta^{AB}_{ij}}$ within
these six scenarios from flavor observables,
thus giving an idea of the overall size of these
parameters taking the latest experimental bounds into account.
The corresponding contributions indicate which level of higher-order
corrections are possible and allowed by the inclusion of NMFV. In
particular in the case of the light Higgs boson we find that
the prediction of $M_h$ can lead to additional new constraints on
the deltas, specifically on \del{LR}{ct} and \del{RL}{ct}.
This is due to the fact that
${\cal A}_{ij}$-terms enter directly into the couplings,
creating a strong sensitivity to these parameters.
In the second part we analyzed four different two-dimensional scenarios,
which are characterized by universal scales for the SUSY electroweak
scale, $m_{\rm SUSY-EW}$, that determines the masses of the scalar leptons and
electroweak particles, and for
the SUSY QCD scale, $m_{\rm SUSY-QCD}$, that determines the masses of the scalar
quarks. As additional free parameter we kept $\tan \beta$.
Within this simplified model it is possible to analyze the
behavior of the corrections to $M_h$, where at the same time agreement with
the anomalous magnetic moment of the muon, \ensuremath{(g-2)_\mu}\ is required.
We demanded that the correction to $M_h$ does not yield values
outside $125.6 \pm 5 \,\, \mathrm{GeV}$,
leading to new improved bounds on \del{LR}{ct} and \del{RL}{ct},
whereas no limits on the other \ensuremath{\delta^{AB}_{ij}} can be obtained.
The limits on $|\del{LR}{ct}|$ turn out to be relatively
independent on the choice of the scenario. For $m_{\rm SUSY-QCD} \sim 1 (3) \,\, \mathrm{TeV}$ bounds of
$|\del{LR}{ct}| \lsim 0.28 (0.12)$ were found. These bounds
on \del{LR}{ct} and \del{RL}{ct} are genuine from Higgs physics
and do not have competitive bounds from $B$-physics observables.
\subsection*{Acknowledgments}
The work of S.H.\ was supported
by the Spanish MICINN's Consolider-Ingenio 2010
Program under grant MultiDark CSD2009-00064.
The work of M.H.\ and M.A.-C.\ was partially supported by the
European Union FP7 ITN INVISIBLES (Marie Curie Actions, PITN- GA-2011-
289442), by the CICYT through the project FPA2012-31880,
by the Spanish Consolider-Ingenio 2010 Programme CPAN (CSD2007-00042)
and by the Spanish MINECO's ``Centro de Excelencia Severo Ocho''
Programme under grant SEV-2012-0249.
|
\section{Introduction}
Manifestation of localization of single particle states in low-dimensional
quantum systems has always been in limelight of research in condensed
matter physics since its prediction. Studies on this topic was basically
stimulated after the work of Anderson in 1958~\cite{anderson} which has
become a milestone in material science. It is well set that exponentially
localized states are the only allowed solutions for one-dimensional ($1$D)
systems with random on-site potentials~\cite{anderson,sharma}. Apart from
the above mentioned fact, another kind of localization phenomenon, the
so-called Wannier-Stark localization~\cite{wannier}, has also drawn
much attention of many physicists~\cite{borysowicz,zekri}. This
type of localizing behavior is observed in $1$D systems subjected to an
external electric field, even in absence of any kind of disorder. Although
the existence of localized surface states in crystals has been first
suggested by Tamm~\cite{tamm}, on the basis of a special $1$D model proposed
by James~\cite{james}, popularity of this phenomenon among scientists has
been started to grow following the work of Wannier~\cite{wannier} which
has won its way after the widespread application to the optical properties
of quantum wells~\cite{harwit,levine1}.
For both these two typical cases i.e., systems with random site energies
and samples with external electric field, one cannot find the existence of
{\em mobility edge} separating the conducting states from the non-conducting
region, since all the energy levels are localized. Therefore, for these
models no long-range transport will be obtained. However, a large number
of physically relevant models are available which exhibit mobility edges
at some typical energy values. For example, a $1$D tight-binding (TB) chain
composed of two uncorrelated random atomic sites exhibits extended energy
eigenstates for some specific electron energies over the entire material
under a suitable alignment of these random sites as originally explored
by Dunlap {\em et al.}~\cite{dunlap}. In this case, a short range positional
correlation among the atomic sites is established. On the other hand,
in quasi-periodic Aubry-Andre model~\cite{aubry}, where long range positional
correlation is found to exist between the atomic sites in $1$D geometry,
conducting energy levels are also noticed. All these special classes of
materials, the so-called correlated disordered materials, have provided
a new turn in the localization phenomena.
Although few attempts~\cite{eco,das,rolf,sch,san1,san2,san3} have been made
to explore the existence of mobility edges in some one- and two-dimensional
($2$D) systems, a detailed analysis of it is still missing which essentially
motivates us to investigate this phenomenon with a renewed interest. In this
work we investigate localization properties of a tight-binding ladder
network, constructed by coupling two superlattice chains laterally
(see Fig.1), in presence of an external electric field. Incidentally, many
interesting and novel features of electronic properties have already been
examined in different types of $1$D superlattice
chains~\cite{pai1,pai2,pai3,pai4,wang}, which are considered as the
equivalent representation of periodic structures of different metallic
layers~\cite{multi1,multi2,multi3}. The motivation of the present study is
two-fold. Firstly, we intend to reveal the interplay of the superlattice
configuration in the ladder network and the external electric field on
electronic conductance. If the mobility edge phenomenon in such a system
still persists at multiple electron energies like a $1$D superlattice chain
in presence of external electric field, then a ladder network of this
particular type, can be used in electronic circuits as a switching device
which might throw new light in the present era of nanotechnology. Secondly,
the ladder networks are extremely suitable for explaining charge transport
through double stranded DNA~\cite{yi,diaz,apalkov}. A clear understanding
of it is highly important due to its potential application in
nano-electronics, especially, to design DNA-based computers. Not only that,
it has also broad applications in biology since it is directly related to
cell division, protein binding processes, etc. Experimental results suggest
that DNA can have all possible conducting properties like,
metallic~\cite{rakitin}, semiconducting~\cite{cai,porath} or even
insulating~\cite{pablo,braun}. In spite of the numerous experimental and
theoretical~\cite{th1,th2,th3} progress of charge transport through double
helix DNA structures, the conducting nature of DNA molecules are not yet
well explained and a deeper analysis is still needed, in the context of
fundamental physics and technological interest. Nowadays scientists are
trying to develop artificial DNA molecules~\cite{doi}, based on DNA
synthesis equipment, those are highly stable and analogous to naturally
occurred DNA molecules. Therefore, the study of electronic transport in a
ladder network comprising of superlattices can give rise significant new
results.
To make the present model much more realistic one should consider the
effect of backbones and the environment. Though backbone sites are not
directly involved on electronic conduction, but the effect of environment
can be well analyzed since it essentially tunes the on-site potentials of
these sites which results some interesting patterns. In our present work
we consider different arrangements of backbone sites (see Fig.1) and
consider their effect on electronic conduction, to make it a self-contained
study.
The rest of the article is arranged as follows. In Section II we present
the model and the calculation method. The numerical results from our model
calculations are illustrated in Section III, and finally, in Section IV
we summarize the essential results.
\section{Model and theoretical framework}
Here we present the model quantum system considered for this work and the
theoretical formulation to describe electronic transport properties in
\begin{figure}[ht]
{\centering \resizebox*{8.7cm}{2.2cm}{\includegraphics{newladder.eps}}\par}
\caption{(Color online). A tight-biding ladder network coupled to finite
width source and drain electrodes. The filled circles with different colors
correspond to different atomic sites.}
\label{ladder2}
\end{figure}
presence of an external bias. We use a simple TB framework to describe the
system Hamiltonian, and, within non-interacting picture this approach is
highly suitable for analyzing electron transport through a bridge system.
\subsection{The TB model}
We consider a conducting junction in which a ladder network is clamped
between source and drain electrodes. A sketch of such a
source-conductor-drain junction is presented in Fig.~\ref{ladder2}, where
the filled circles with different colors correspond to different atomic
sites. The Hamiltonian of this full system can be written as sum of three
terms like,
\begin{equation}
\mbox{\boldmath$H$}=\mbox{\boldmath$H$}_{\mbox{\tiny lad}} +
\mbox{\boldmath$H$}_{\mbox{\tiny eltd}} +
\mbox{\boldmath$H$}_{\mbox{\tiny tun}}.
\label{hsum}
\end{equation}
The first term of Eq.~\ref{hsum} represents the Hamiltonian of the ladder
network, while the next two terms correspond to the Hamiltonians of the
side attached electrodes and electrode-to-ladder couplings, respectively.
Under nearest-neighbor hopping approximation the TB Hamiltonian of the
ladder network in site representation reads as,
\begin{eqnarray}
\mbox{\boldmath$H$}_{\mbox{\tiny lad}}&=&\sum_{i=I,II} \sum_j
\left[\epsilon_{i,j} \mbox{\boldmath$c$}_{i,j}^{\dagger}
\mbox{\boldmath$c$}_{i,j} + t \left( \mbox{\boldmath$c$}_{i,j}^{\dagger}
\mbox{\boldmath$c$}_{i,j+1} + \mbox{h.c.}\right) \right] \nonumber \\
& + & \sum_j v \left(\mbox{\boldmath$c$}_{I,j}^{\dagger}
\mbox{\boldmath$c$}_{II,j} + \mbox{\boldmath$c$}_{II,j}^{\dagger}
\mbox{\boldmath$c$}_{I,j} \right) \nonumber \\
& + & \sum_{i=I,II} \sum_j \left[\epsilon_{i,j}^b
\mbox{\boldmath$c$}_{i,j}^{b\dagger} \mbox{\boldmath$c$}_{i,j}^b +
t^b \left( \mbox{\boldmath$c$}_{i,j}^{b\dagger} \mbox{\boldmath$c$}_{i,j}^b
+ \mbox{h.c.}\right) \right]
\label{hladder}
\end{eqnarray}
where the index $i$($=I,II$) corresponds to two strands of the ladder
and the index $j$($=1,2,3,\dots$) refers to the atomic sites in these
strands. Thus, using the notation ($i,j$) we can locate atomic sites in
any strand. Each of these sites is again directly coupled to a backbone
site (see Fig.~\ref{ladder2}) which essentially captures the effect of the
environment. $\epsilon_{i,j}$ is the on-site energy of an electron at the
site $j$ of $i$-th strand, and $t$ and $v$ are the intra-strand and
inter-strand hopping integrals, respectively.
$\mbox{\boldmath$c$}^{\dagger}_{i,j}$ ($\mbox{\boldmath$c$}_{i,j}$) is the
creation (annihilation) operator of an electron at the site $(i,j)$. For
the backbones we use an additional index $b$ in our TB model to describe
the parameters like, on-site energies and hopping matrix elements, such
that they are distinguished from the parameters used in the parent ladder.
The backbone sites are modeled by the on-site energies $\epsilon_{i,j}^b$,
while their couplings to the atomic sites of the strands are characterized
by the nearest-neighbor hopping integrals $t^b$.
Depending on the arrangements of atomic sites in two different strands as
well as in backbones, several cases are analyzed to describe the localization
phenomena in presence of an external voltage bias. We essentially focus on
two different setups. In one case the upper and lower strands are arranged
with superlattice configurations, setting the uniform site energies in the
backbones, while in the other case the scenario gets reversed. Here, the
strands are arranged with identical lattice sites and the backbone sites
are configured with superlattice sites. In the present communication, we
describe the superlattice models considering two or three types of atoms,
in each unit cell, those are labeled as $\alpha$ and $\beta$ or $\alpha$,
$\beta$ and $\gamma$, and they are arranged following a particular sequence.
For example, it can be either $\alpha^p \beta^q$ or $\alpha^p \beta^q
\gamma^r$, where $p$, $q$ and $r$ are three positive integers. Repeating
such unit cells we will get the entire lattice chain and construct the
desired full system for our study.
In presence of a finite voltage bias $V$ between the source and drain
electrodes, an electric field is established which results the site
energies of the bridging conductor voltage dependent. Mathematically it
reads as,
\begin{eqnarray}
\epsilon_{i,j}=\epsilon_{i,j}(0)+\epsilon_{i,j}(V)
\label{ss}
\end{eqnarray}
where, $\epsilon_{i,j}(0)$ represents the site energy in absence of external
bias $V$. Depending on the nature of atomic sites $\epsilon_{i,j}(0)$'s are
called as $\epsilon_{\alpha}(0)$, $\epsilon_{\beta}(0)$,
$\epsilon_{\gamma}(0)$ corresponding to $\alpha$, $\beta$, $\gamma$ type
atomic sites, respectively. The voltage dependence in site energies
essentially comes from two regions. One from the bare electric field
\begin{figure}[ht]
{\centering \resizebox*{6cm}{3.5cm}{\includegraphics{profile.eps}}\par}
\caption{(Color online). Voltage dependent site energies for three different
potential profiles in a $200$-rung ladder network when the external bias $V$
is set at $0.2$.}
\label{profile}
\end{figure}
in the source-conductor-drain bias junctions and the other from the
screening in presence of long-range electron-electron (e-e) interactions,
which is not directly taken into account in the Hamiltonian Eq.~\ref{hladder}.
If initially we don't consider the effect of such screening due to e-e
interactions, then the electric field becomes uniform along the ladder
and it gets the form,
\begin{equation}
\epsilon_{i,j}(V)=V/2 -\frac{j V}{N_r+1}
\label{eqfld}
\end{equation}
where, $N_r$ is the total number of rungs in the ladder. Each of these
rungs is attached to two backbone sites in opposite sides of the strands
($i=I,II$) whose site energies are also identically modified, following
Eq.~\ref{eqfld}, in presence of voltage bias like the parent lattice sites
in individual rungs. Thus, site energies of four atomic sites (two parent
and two backbone sites) in each vertical line get equally modified in
presence of external bias and their voltage dependence depend only on the
distance from the finite width source electrode which results Eq.~\ref{eqfld}
$i$ independent. Now, if we consider the screening effect, then the electric
field will no longer be linear. Two such cases are shown in
Fig.~\ref{profile} for two different screening strengths, as illustrative
examples. Below, we will analyze the localization phenomena in the ladder
network considering both linear and non-linear bias drops.
The TB Hamiltonians of the finite width electrodes
($\mbox{\boldmath$H$}_{\mbox{\tiny eltd}}$) and electrode-ladder couplings
($\mbox{\boldmath$H$}_{\mbox{\tiny tun}}$) are also expressed quite similar
to Eq.~\ref{hladder}, where the electrodes are described with on-site
energies $\epsilon_0$ and nearest-neighbor interactions $t_0$. Two atomic
sites of the ladder are directly coupled to the atomic sites of the source
electrode where the coupling strength is described by $\tau_S$ and it is
$\tau_D$ for the drain-ladder coupling.
\subsection{Theoretical methods: Green's function approach}
To find electronic transmission coefficient for this bridge setup, we use
Green's function formalism~\cite{fisher,datta1,datta2} which is quite robust
compared to other existing theories available in literature. In this
approach an infinite dimensional system (since electrodes are semi-infinite)
gets effectively reduced to the dimension of a finite size conductor clamped
between two electrodes.
Using Fisher-Lee relation~\cite{fisher}, the two-terminal transmission
probability can be written as,
\begin{eqnarray}
T= {\mbox{Tr}} \left[ \mbox{\boldmath$\Gamma$}_S
\mbox{\boldmath$G$}^r_{\mbox{\tiny lad}} \mbox{\boldmath$\Gamma$}_D
\mbox{\boldmath$G$}^a_{\mbox{\tiny lad}} \right]
\label{transm}
\end{eqnarray}
where, $\mbox{\boldmath$\Gamma$}_S$ and $\mbox{\boldmath$\Gamma$}_D$ are
the coupling matrices and $\mbox{\boldmath$G$}^r_{\mbox{\tiny lad}}$ and
$\mbox{\boldmath$G$}^a_{\mbox{\tiny lad}}$ are the retarded and advanced
Green's functions of the ladder network, respectively. The single particle
Green's function operator representing the complete system i.e., ladder
including source and drain electrodes, for an electron having energy $E$
is defined as,
\begin{equation}
\mbox{\boldmath$G$}=\left[ (E+i \eta) \mbox{\boldmath$I$}- \mbox{\boldmath$H$}
\right]^{-1}
\end{equation}
where, $\eta \rightarrow 0^+$. $\mbox{\boldmath$H$}$ is the Hamiltonian of
the full system and $\mbox{\boldmath$I$}$ denotes the identity matrix.
Introducing the concept of self-energies due to source and drain, the
problem of finding $\mbox{\boldmath$G$}$ in the full Hilbert space of
$\mbox{\boldmath$H$}$ can eventually be mapped to a reduced Hilbert space
of the finite ladder, and the effective Green's function
$\mbox{\boldmath$G$}_{\mbox{\tiny lad}}^r$ looks like,
\begin{eqnarray}
\mbox{\boldmath$G$}^r_{\mbox{\tiny lad}}=\left[(E+i\eta)\mbox{\boldmath$I$}
-\mbox{\boldmath$H$}_{\mbox{\tiny lad}}-\mbox{\boldmath$\Sigma$}_S^r
- \mbox{\boldmath$\Sigma$}_D^r \right]^{-1}.
\label{greendevice}
\end{eqnarray}
where, $\mbox{\boldmath$\Sigma$}_S$ and $\mbox{\boldmath$\Sigma$}_D$ are the
contact self-energies used to describe the effect of couplings of the ladder
to the source and drain, respectively. A detailed derivations of these
self-energies are available in Refs.~\cite{datta1,datta2,nikolic,we1}.
The coupling matrices $\mbox{\boldmath$\Gamma$}_S$ and
$\mbox{\boldmath$\Gamma$}_D$ corresponding to the couplings of the ladder
to the source and drain electrodes are directly associated with the
self-energies and they are,
\begin{eqnarray}
\mbox{\boldmath$\Gamma$}_{S(D)}&=&i \left[ \mbox{\boldmath$\Sigma$}^r_{S(D)}
-\mbox{\boldmath$\Sigma$}^a_{S(D)}\right] \nonumber\\
&=& -2~ {\mbox{Im}}\left( \mbox{\boldmath$\Sigma$}^r_{S(D)}\right).
\end{eqnarray}
$\mbox{\boldmath$\Sigma$}^r_{S(D)}$, on the other hand, is the sum
$\mbox{\boldmath$\Sigma$}^r_{S(D)}=\mbox{\boldmath$\Lambda$}_{S(D)}
+i \mbox{\boldmath$\Delta$}_{S(D)}$, where the real part
$\mbox{\boldmath$\Lambda$}_{S(D)}$ corresponds to the shift of energy
levels of the ladder, and the imaginary part
($\mbox{\boldmath$\Delta$}_{S(D)}$) is responsible for the broadening
of these levels.
Finally, the average density of states (ADOS) of the ladder is determined
from the relation,
\begin{equation}
\rho(E)=-\frac{1}{\pi Ns} {\mbox{Im}} \left[{\mbox{Tr}} \,
\mbox{\boldmath$G$}^r_{\mbox{\tiny lad}}\right]
\end{equation}
where, $N_s$ is the total number of atoms in the ladder network.
\section{Numerical Results and Discussion}
In this section, we describe numerical results obtained from the above
theoretical prescription given in Section II to investigate localization
\begin{figure}[ht]
{\centering \resizebox*{8cm}{13cm}{\includegraphics{figVT.eps}}\par}
\caption{(Color online). Two-terminal transmission probability $T$
(orange line) as a function of electron energy $E$ for a perfect ladder,
in absence of backbone sites, considering the linear potential drop as
shown by the red curve in Fig.~\ref{profile}, where (a), (b) and (c)
correspond to $V=0$, $0.2$ and $0.4$, respectively. The ADOS (green line)
is superimposed in each spectrum. Here we set $N_r=80$.}
\label{figvt}
\end{figure}
phenomena in the ladder network subjected to an external bias. Throughout
the analysis we fix the electronic temperature at zero and choose $c=e=h=1$
for simplification. All the energies are scaled in unit of the hopping
integral $t$.
As stated earlier, two different cases are analyzed depending on the
superlattice configurations in the strands and backbones. In one setup,
the strands are arranged with superlattice sites considering identical
backbone sites, while in the other setup its opposite consequence is taken
into account.
\begin{figure*}[t]
{\centering \resizebox*{17cm}{15cm}{\includegraphics{fig2ordered.eps}}\par}
\caption{(Color online). Two-terminal transmission probability $T$ together
with average density of states $\rho$ as a function of energy $E$ for a
$150$-rung ladder network, considering of a linear bias drop along the
ladder, where the upper, middle and lower rows correspond to $V=0$, $0.2$
and $0.4$, respectively. Two columns represent two different arrangements
of $\alpha$ and $\beta$ sites. In the left column, the results are computed
for the network where we choose $p=3$ and $q=2$ in the upper strand, and
$p=q=1$ in the lower strand. While, in the other column we set $p=3$ and
$q=2$ in both strands. All these results are done for $t^b=0$. The orange
and green lines correspond to the identical meaning as stated in
Fig.~\ref{figvt}.}
\label{figord}
\end{figure*}
\begin{figure*}[t]
{\centering \resizebox*{17cm}{11cm}{\includegraphics{figtb.eps}}\par}
\caption{(Color online). Two-terminal transmission probability and
average density of states as a function of energy $E$ for a $150$-rung
ladder network considering $t^b=0.4$, where the orange and green curves
represent the similar meaning as stated earlier. The results shown in (a)
and (b) are computed for a linear potential profile, whereas (c) and (d)
are done for the non-linear potential profiles given by the blue and
pink lines in Fig.~\ref{profile}, respectively. In (a) we set $V=0$,
while in the other three spectra (c-d) we choose $V=0.2$. All these results
are performed setting $p=3$ and $q=2$ in the upper and lower strands of the
ladder network.}
\label{figtb}
\end{figure*}
Before addressing these central cases, let us first discuss the effect
of voltage bias on electron transport in a simple system which is a perfect
ladder without any backbone sites. The results of such a simple model is
presented in Fig.~\ref{figvt}, where we choose $N_r=80$. For this ladder,
the voltage independent sites energies ($\epsilon_{i,j}(0)$'s) are identical,
and therefore, they can be fixed at zero, without any loss of generality.
The other parameters used here are as follows. The nearest-neighbor hopping
integrals $t$ and $v$ are fixed at $1$, and the ladder-electrode coupling
constants are taken to be $\tau_S=\tau_D=1$. In side-attached electrodes,
the on-site potentials $\epsilon_0$ and hopping integrals $t_0$ are set at
$0$ and $1$, respectively. Depending on the potential drop between the
source and drain electrodes, three different cases are taken into account
those are presented in Figs.~\ref{figvt}(a), (b) and (c) where we select
$V=0$, $0.2$ and $0.4$, respectively. In each of these three spectra
two-terminal transmission probability together with average density of
states are shown. All these results are computed considering a linear
potential profile along the ladder. In the absence of any electric field
associated with the voltage bias, electrons can conduct through the ladder
for all possible allowed energies corresponding to the energy eigenvalues
of the ladder which can be clearly noticed from the upper panel of
Fig.~\ref{figvt}(a). Under this condition i.e., $V=0$, on-site potentials
are no longer affected and it results extended energy levels throughout
the energy band window of the ordered ladder. The situation becomes
somewhat interesting when a finite potential drop appears along the ladder
(Figs.~\ref{figvt}(b) and (c)). It is observed that several energy levels
along the energy band edges do not contribute anything to the electron
conduction i.e., transmission coefficient drops to zero for these energies.
Obviously, if the Fermi energy of the ladder lies within these regions,
no electron transmission takes place and the insulating phase is obtained.
The finite transmission will take place only when the energy moves towards
the band centre. Thus, a sharp transition between a fully insulating zone
to a conducting zone is obtained which gives rise to the phenomenon of
mobility edge in presence of external electric field. In presence of a
non-zero bias, site energies are voltage dependent (see Eq.~\ref{ss}),
those are also not identical to each other, which are responsible for
generating localized energy levels in the energy spectrum. Certainly, more
such localized energy levels will be available for higher potential drop.
This is exactly shown in Fig.~\ref{figvt}(c), considering a finite bias
$V=0.4$. It should be noted that, like a conventional disordered material,
the localization of energy levels always starts from the band edges, keeping
the extended states towards the band centre. This phenomenon has been
revisited in our recent work in the context of studying integer quantum Hall
effect~\cite{hall} within a tight-binding framework. It reveals that, for
large enough electric field almost all the energy levels get localized, and
therefore, no such crossover between a conducting zone and an insulating
phase is observed.
Based on the above analysis of electronic localization in a perfect ladder
in presence of external electric field now we concentrate on the central
results i.e., the interplay of superlattice configuration in parent strands
as well as in backbones of the ladder network and the external electric
field associated with the voltage bias. In what follows, we present our
results in two sub-sections for two different setups.
\subsection{The TB ladder network with superlattice chains and identical
backbone sites}
Here we present our numerical results for the ladder network considering
superlattice sites in the strands, setting identical lattice sites in
backbones. Two different types of atomic sites, labeled as $\alpha$ and
$\beta$, are taken into account and depending on their arrangements in
the strands several cases are analyzed. These arrangements are specified
by the rule $\alpha^p\beta^q$, as stated earlier, where $p$ and $q$ are
variables. The parameters used here for the calculations, unless stated
otherwise, are $\epsilon_{\alpha}(0)=-\epsilon_{\beta}(0)=1$,
$\epsilon_{\alpha}^b(0)=\epsilon_{\beta}^b(0)=0$, $t=v=1$, $\epsilon_0=0$,
$t_0=1$ and $\tau_S=\tau_D=1$. The hopping integral $t^b$ is mentioned
in the figure captions.
As illustrative examples, in Fig.~\ref{figord} we present two-terminal
transmission probability together with average density of states of
\begin{figure*}[t]
{\centering\resizebox*{17cm}{15cm}{\includegraphics{backbone.eps}}\par}
\caption{(Color online). Two-terminal transmission probability $T$ and
average density of states $\rho$ as a function of energy $E$ for a ladder
network in which strands are arranged with identical lattice sites and the
backbones are configured with superlattice sites. In the left column,
backbones are arranged with $\alpha$ and $\beta$ sites where we set $p=5$
and $q=2$, while in the right column they are arranged with $\alpha$,
$\beta$ and $\gamma$ sites considering $p=5$, $q=3$ and $r=2$. The results
given in the 1st row are done with $t^b=0.4$, while all the other results
are computed when $t^b$ is fixed at $0.8$. The voltage $V$ is taken to be
zero in the 1st and 2nd rows, whereas it is $0.2$ in the 3rd row. All these
results are performed considering a linear electrostatic potential profile.
Moreover, we choose total number of rungs $N_r=140$ for the left column,
while it is $150$ in the right column. Two different colors correspond to
the identical meaning as mentioned in the above figures.}
\label{backbone}
\end{figure*}
a $150$-rung ladder network, considering $t^b=0$ i.e., effect of backbone
sites is ignored, for different configurations of $\alpha$ and $\beta$ sites
in two strands. For example, in the left column the results are shown when
we choose $p=3$ and $q=2$ in the upper strand and these parameters are set
equal to $1$ for the other strand. On the other hand, in the right column
we select $p=3$ and $q=2$ for both these two strands. All these results are
computed taking a linear potential profile along the ladder. Quite
interestingly we see that, in presence of superlattice sites multiple
energy bands, separated by finite gaps, are appeared in the energy spectra.
Depending on the arrangements of superlattice sites, i.e., the choices of
unit cells in two different strands, total number of such energy bands
varies. Under this situation when we consider the effect of external
electric field associated with the voltage bias, we get multiple localized
regions, generated from the band edges, those are separated from the extended
energy zones (see the spectra in Fig.~\ref{figord}). Therefore, setting
the Fermi energy in suitable energy regions the ladder network can be
used to transmit electrons or not from the source to drain electrode, and
thus, support in fabricating mesoscopic switching devices at multiple
energies. This essentially leads to the phenomenon of getting multiple
mobility edges.
The appearance of energy bands, separated by finite energy gaps, in presence
of superlattice sites in the strands is strongly affected by the backbone
sites, and the localization properties for a particular configuration, on
the other hand, are highly influenced by the nature of electrostatic
potential profiles. These issues are addressed in Fig.~\ref{figtb}, where
we choose $N_r=180$ and set $p=3$ and $q=2$ in both strands of the ladder
network. In the upper panel the results are calculated considering a
linear bias drop along the sample, while in the lower panel they are
computed for non-linear potential profiles. Comparing the spectra shown
in Figs.~\ref{figord}(d) and \ref{figtb}(a) (since in these two cases
all other parameters, except $t^b$, are identical) we emphasize that the
ADOS spectra can be controlled in a significant way by means of the
backbone environment. This change in the density profile becomes much
pronounced when the difference between the parameters $p$ and $q$ is
quite large, which we justify through our extensive numerical analysis.
The effects of backbone sites is elaborately discussed in the following
sub-section.
Keeping in mind the possibilities of having much more electron screening
now we present the results (Figs.~\ref{figtb}(c) and (d)) where
non-linear potential profiles are considered. With increasing the flatness
of the electrostatic potential profile, the localized regions get
decreased. Eventually, for the extreme condition i.e., when the bias
drop takes place only at the ladder-to-electrode interfaces no such
localized regions will be obtained, and certainly, metal-to-insulator
transition cannot be obtained.
So far the results discussed are for the systems where strands are arranged
with superlattice configuration. Now the question is, can one get multiple
mobility edges if the backbone sites are modulated with superlattice sites
instead of the strands? If yes, then few more advantages are there than the
cases studied above. First, one can easily tune backbone site energies by
placing the sample in the vicinity of external gate electrodes which can
control the band spectra, and thus, results a selective switching action
for a fixed Fermi energy in presence of external electric field. Second,
the environmental effect on electron transport as present in natural DNA
molecules can also be explained through our model quantum systems. Below,
these issues are discussed with some typical examples.
\subsection{The TB ladder network with identical lattices in strands and
superlattice sites in backbones}
Finally, we consider the system where strands are arranged with identical
atoms and backbones are arranged with superlattice sites. Some typical
cases are analyzed depending on the choices of superlattice sites and the
results are presented in Fig.~\ref{backbone}. For example, in the left column
of Fig.~\ref{backbone} the backbone sites are arranged with two different
atomic sites $\alpha$ and $\beta$ obeying the sequence $\alpha^p\beta^q$,
while in the right column three atomic sites $\alpha$, $\beta$ and $\gamma$
are used to construct the backbones following the relation
$\alpha^p\beta^q\gamma^r$. The parameters considered for these calculations
are: $\epsilon_{\alpha}(0)=\epsilon_{\beta}(0)=\epsilon_{\gamma}(0)=0$,
$\epsilon_{\alpha}^b(0)=-2$, $\epsilon_{\beta}^b(0)=0$,
$\epsilon_{\gamma}^b(0)=2$, $t=v=1$, $\epsilon_0=0$, $t_0=1$ and
$\tau_S=\tau_D=1$. The hopping strength $t^b$ is given in the figure
caption. From the spectra (Fig.~\ref{backbone}), it is interesting to
note that multiple energy bands separated by finite gaps are also
obtained here like the cases for the ladders with superlattice strands.
At the same time, the gap between the energy bands can be tuned by
regulating the coupling strength $t^b$ which may provide additional
control parameter for getting selective switching effect at the meso-scale
level for multiple energies.
\section{Concluding Remarks}
In the present communication, we address electric field effect associated
with the applied voltage bias on localization phenomena in a tight-binding
ladder network with superlattice configuration. Using Green's function
approach we compute two-terminal transmission probability along with average
density of states for different arrangements of superlattice sites in the
strands as well as in the backbones. Our numerical results exhibit several
interesting features which essentially lead to a possibility of using the
ladder network as a switching device at multiple energies. This switching
action is significantly controlled by many factors like, the nature of
electrostatic potential profile, presence of backbone sites, coupling
between the backbone sites and the sites in the parent strands, etc.
With increasing the flatness of the potential profile, related to screening
effects, localized energy regions across the band edges gradually decreases,
and, for the typical case when potential drop takes place only at the
electrode-ladder interfaces, localized regions disappear, and thus,
metal-to-insulator transition cannot be observed upon changing the
system Fermi energy. Similarly, the backbone sites are also responsible
for controlling the switching action which practically suggest us an
experiment towards this direction.
Finally, it is important to state that, although the electronic temperature
is fixed at zero throughout the analysis, all these results are still valid
in the low-temperature limit. This is due to the fact that the thermal
broadening of energy levels is much weaker than the broadening caused by
ladder-to-electrode coupling~\cite{datta1,datta2,we2,we3}.
\section*{Acknowledgment}
Second author is thankful to Prof. A. Nitzan for useful conversations.
|
\section{Concurrent Reaction}
In classical high-temperature kinetics we need to consider complex reactions involving numerous reactants in order to obtain oscillating or irregular dynamics. However, we will see that when we add to the ultracold diatomic molecule formation a simple zeroth order reaction\blue{,} we already encounter Hamiltonian chaos in the mean field regime. For this purpose consider the chemical reaction
\begin{eqnarray}
A+A \overset{k_1}{\rightleftarrows} A_2 \\
\text{bath} \overset{k_2}{\rightleftarrows} A
\label{eq:concurrent}
\end{eqnarray}
Applying the proposed rules and assuming that our particles are trapped in the ground state of some harmonic potential, we obtain the following Hamiltonian:
\begin{equation}
\hat{H}=E_A\hat{n}_A +E_{A_2}\hat{n}_{A_2} + k_1(\hat{a}_A^\dag\hat{a}_A^\dag\hat{a}_{A_2} +\hat{a}_A\hat{a}_A\hat{a}_{A_2}^\dag)+k_2(\hat{a}^\dag_A+\hat{a}_A),
\label{eq:hamilton_expanded}
\end{equation}
where $E_A$ and $E_{A_2}$ denote the ground state energy of the respective molecular or atomic species. The concurrent zeroth order reaction \eref{eq:concurrent} breaks the conservation of the overall particle number $\hat{N}_{tot}$ and therefore impedes a full quantum mechanical treatment. Consequently, we investigate the dynamics of the system in the mean field approximation by replacing the creation and annihilation operators in \eref{eq:hamilton_expanded} with complex numbers $(\alpha_A,\alpha_{A_2})$ labeling coherent states. Note that within this approximation we obtain the average particle number of a certain species $\langle \hat{N_i}\rangle$ by considering the square of the modulus of the respective complex number $|\alpha_i|^2$.
In the previous section we have seen that the diatomic molecule formation amounts to deviations from mean field dynamics because of the occurrence of quantum effects. Therefore, we need to keep the coupling parameter $k_1$ as small as possible compared to some relevant energies to consider the mean field limit as an appropriate description of the actual dynamics. Keeping this in mind, we obtain the equations of motion from the variational principle:
\begin{eqnarray}\eqalign{
i\dot{\alpha}_A&=E_A\alpha_A + 2k_1\overline{\alpha}_A\alpha_{A_2}+k_2\\
i\dot{\alpha}_{A_2}&=E_{A_2}\alpha_{A_2} + k_1\alpha_A^2}.
\end{eqnarray}
To reduce the number of parameters in the system, we remove unnecessary degrees of freedom by replacing the dynamical variables with nondimensionalized quantities
\begin{eqnarray}
\tilde{\alpha}_A&=\frac{\alpha_A}{\alpha_{0}} , \qquad \tilde{\alpha}_ {A_2}=\frac{\alpha_{A_2}}{\alpha_{0}},\qquad \tau=\frac{t}{t_0}.
\end{eqnarray}
This amounts to the coupled equations
\begin{eqnarray}\eqalign{
i\frac{d\tilde{\alpha}_A}{d\tau} - \tilde{\alpha}_A = 1 + c_12\overline{\tilde{\alpha}}_A\tilde{\alpha}_{A_2}\\
i\frac{d\tilde{\alpha}_{A_2}}{d\tau} - c_2\tilde{\alpha}_{A_2} =c_1\tilde{\alpha}_A^2,}
\label{eq:coupled}
\end{eqnarray}
and energy
\begin{equation}
\tilde{H}=|\tilde{\alpha}_A|^2+c_2|\alpha_{A_2}|^2+c_1\left(\overline{\tilde{\alpha}}_A\overline{\tilde{\alpha}}_A\tilde{\alpha}_{A^2}+c.c.\right)+(\overline{\tilde{\alpha}}_A+c.c.),
\label{eq:nondimham}
\end{equation}
with parameters
\begin{eqnarray}
t_0&=\frac{1}{E_A},\ \alpha_{0}=\frac{k_2}{E_A},\ c_1=\frac{k_1k_2}{E_A^2},\ c_2=\frac{E_{A_2}}{E_A}.
\end{eqnarray}
Therefore, the dynamics of the system is completely determined by the choice of the two parameters $c_1,c_2\in\mathbb{R}$ and its initial conditions. Note that due to our choice of parameters $c_1=0$ implies $k_1=0$, i.e.\ no molecule formation, and arbitrary $k_2$. Moreover, the constraint stemming from the validity of the mean field approximation can now precisely be expressed as $c_1\ll 1$ which translates the constraint on the molecular reaction constant to $k_1\ll\frac{E_A^2}{k_2}$.
What is the physical meaning of those two parameters? The ODE system \eref{eq:coupled} is equivalent to a pair of nonlinearly coupled harmonic oscillators. Whereas the oscillator describing the atoms has eigenfrequency one, $c_2$ determines the eigenfrequency of the molecular oscillator. Moreover, $c_1$ is the coupling strength between the two oscillators and is the only nonlinear term in the system. We therefore expect a regular behaviour for small values of $c_1$. Moreover, we see from \eref{eq:nondimham} that if $c_1\gg1$ the system is also integrable due to the conservation of the overall particle number $N_{tot}=|\tilde{\alpha}_A|^2+2|\tilde{\alpha}_{A_2}|^2$. In what follows, we investigate the different dynamical regimes determined by the choice of $c_1$ and fix $c_2$ to an experimentally realistic value.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{merged.png}
\caption{Time evolution of the (nondimensionalized) average number of atoms and molecules in mean field approximation. The relative energy of the ground states is $c_2=1.1$ and the initial number of atoms is $40000$. The oscillation of becomes more and more modulated with increasing molecule formation $c_1$. Perturbation theory determines an amplitude $A_{mod}\approx111$ atoms for $c_1=8\times10^{-5}$ and $A_{mod}\approx320$ atoms for $c_1=13\times10^{-5}$ (see \eref{eq:ammod}).}
\label{fig:merged}
\end{figure}
Applying perturbation theory for anharmonic oscillators \cite{hinch1991perturbation} provides us analytic solutions of the occurring dynamical phenomena for sufficiently small coupling constants $c_1$: The trajectories depicted in figure \ref{fig:merged} show the expectation value of atoms $|\tilde{\alpha}_A|^2$ and molecules $|\tilde{\alpha}_{A_2}|^2$ for an initially depleted molecular mode and $40000$ atoms. In case of no coupling at all, i.e., $c_1=0$, the molecular mode remains completely depleted, whereas the atomic mode oscillates due to the coupling to the bath. However, as we increase $c_1$ the system progressively enters a \emph{modulational regime}, in which the molecular site regularly oscillates and the amplitude of the free oscillation on the atomic site is modulated. Let $A_0=\tilde{\alpha}_A(0)$ denote the square root of the initial nondimensionalized number of atoms, then we obtain the following analytical expressions for the amplitude $A_{mod}$ and frequency $\omega_{mod}$ of this modulation:
\begin{eqnarray}\eqalign{
A_{mod}&=\frac{4 (A_0+1) A_0^3 c_1 ^2}{(c_2 -2)^2}+\mathcal{O}\left(c_1^4\right), \\
\omega_{mod}&=c_2-1+\mathcal{O}\left(A_0^2c_1^2\right).}
\label{eq:ammod}
\end{eqnarray}
This means increasing $c_1$ causes an quadratic increase of $A_{mod}$ whereas the frequency of modulation $\omega_{mod}$ remains approximately unchanged. The restriction for the parameters for perturbation theory to be valid are $c_1^2A_0^2\ll 1\ll A_0$ and $c_2\in\left(1,2\right)$.
What happens to the system, if we increase $c_1$ beyond the regime of perturbation theory? We already mentioned that $c_1$ interpolates between integrable systems. But does the system remain integrable for all choices of $c_1$? A well-known tool to characterize irregular behaviour of a system is to consider its Poincar\'e sections \cite{ott2002chaos}. In our case, we choose the quadratures $X_i=\frac{1}{2}(\alpha_i+\overline{\alpha_i})$ and $P_i=\frac{1}{2i}(\alpha_i-\overline{\alpha_i})$ with $i\in\{A,A_2\}$ as dynamical variables and the surface $X_{A_2}=0$ as intersection surface. Illustrative examples of these sections for different $c_1$ are shown in figure \ref{fig:poincare}. We find that for a certain range of $c_1$ the system shows behaviour, which is typical for Hamiltonian chaos: As long as the system remains integrable, the Poincar\'e section consists of closed curves corresponding to sections of two-dimensional tori. However, increasing $c_1$ deforms and finally destroys some of the closed curves. Some of the sampled trajectories start to densely fill out parts of the energy hypersurface. We call this the \emph{chaotic regime} of the reaction. Finally, further increase of $c_1$ leads to deformation of the energy hypersurface and eventually restores the integrability of the system.
\begin{figure}[htbp]%
\centering
\subfloat[][]{%
\label{fig:ex3-a}%
\includegraphics[width=0.25\textwidth]{plot1.png}}%
\hspace{8pt}%
\subfloat[][]{%
\label{fig:ex3-b}%
\includegraphics[width=0.25\textwidth]{plot2.png}}\\
\subfloat[][]{%
\label{fig:ex3-c}%
\includegraphics[width=0.25\textwidth]{plot3.png}}%
\hspace{8pt}%
\subfloat[][]{%
\label{fig:ex3-d}%
\includegraphics[width=0.25\textwidth]{plot4.png}}\\
\subfloat[][]{%
\label{fig:ex3-e}%
\includegraphics[width=0.25\textwidth]{plot5.png}}%
\hspace{8pt
\subfloat[][]{%
\label{fig:ex3-f}%
\includegraphics[width=0.25\textwidth]{plot6.png}}%
\caption[Poincare Sections]{Poincar\'e sections with $X_A$ and $P_A$ on the (x,y)-axes. We set $c_2=1.1$ and the energy $E=100$. Note that in contrast to usual Poincar\'e sections in the literature, we plotted $P_{A_2}$ on the z-axis to get an better impression of the projected energy hypersurface. The sections are plotted for 25 long-time trajectories with arbitrary initial conditions: For a small perturbation pictured in \subref{fig:ex3-a} the system remains integrable. Increasing the perturbation in \subref{fig:ex3-b} leads to splitting of the first orbits into little islands according to the Poincar\'e-Birkhoff theorem \cite{poincare}. In \subref{fig:ex3-c} the irregular trajectories begin to spread out and densely fill out the energy hypersurface. In \subref{fig:ex3-d} most of the energy hypersurface is covered by irregular trajectories. Figure \subref{fig:ex3-e} shows a significant deformation of the energy hypersurface. In \subref{fig:ex3-f} all of the sampled trajectories are again on integrable curves.}%
\label{fig:poincare}%
\end{figure}
\section{Introduction}
A chemical reaction normally occurs at a few hundred kelvin between reagents involving large numbers of particles ($\sim 10^{23}$). This is because reactions are usually activated by \emph{thermal fluctuations} which are only significant for large concentrations of particles with high momenta. Therefore, the possibility of a chemical reaction taking place in the dilute and ultracold ($T<1\, \mu $K) regime is somewhat counterintuitive.
However, the formation of ultracold molecules from atoms in a Bose-Einstein condensate (BEC) was observed almost 15 years ago \cite{wynar2000molecules}.
This interaction of atoms and molecules close to the absolute zero of temperature has been referred to as \emph{ultracold chemistry} \cite{hutson2010ultracold}.
A variety of experimental techniques such as the coupling of atoms and molecules via magnetic Feshbach resonance \cite{donley2002atom,regal2003creation} have been successfully employed to achieve chemical bonding in an ultracold environment. For example, ultracold Potassium-Rubidium molecules \cite{ospelkaus2010quantum} have been investigated to analyse the quantum mechanical effects of particle statistics on molecular reactivity. The now ready experimental accessibility of chemical processes in the dilute ultracold regime strongly motivates us to develop a general physical understanding of their \emph{reaction kinetics}.
A quantized description is required in order to study the dynamics of ultracold reactions, in order to fully account for the effects of quantum fluctuations and entanglement. Here we should replace the classical notion of a temperature-dependent reaction with a coherent reversible Hamiltonian evolution. The first phenomenological steps toward such a description were taken in \cite{heinzen2000superchemistry}, where a mean field ansatz was exploited to describe the coherent formation of diatomic molecules in a BEC. Since then a variety of extensions to this model, mainly focussed on adding quantum corrections to the original mean field ansatz, have been studied \cite{goral2001multimode,Santos2005classical}. Concomitantly, theoretical explanations, via two-body scattering processes, for the ultracold chemical reaction rates observed in recent experiments have been proposed \cite{carr2009cold}. In the case of fermionic particles with several internal states the observed reaction rates may be understood in terms of a simple quantum threshold model \cite{quemener2010strong}. When there are different particle types, e.g.\ bosons and fermions, more sophisticated modelling in terms of multichannel quantum defect theory \cite{idziaszek2010universal} is required. So far, a general and systematic investigation of the dynamics of ultracold reactions, analogous to the study of reaction kinetics for classical thermal reactions, has not yet been undertaken.
Such an approach seems to be indispensable if one wants to study the role of quantum coherence and the production of quantum entanglement in ultracold chemical systems.
In this paper we propose a general scheme to study the kinetics of ultracold chemical reactions. Investigating the predicted dynamics, we find that entanglement replaces the role of thermal fluctuations in activating ultracold chemical reactions. This allows us to draw a parallel between ultracold chemical reactions and quenched dynamics. Thanks to the generality of our formulation it is possible to consider complex chemical reactions involving many reagents. We exemplify our approach by studying an experimentally accessible example of an ultracold reaction exhibiting rich quantum phenomena leading to a dynamical relaxation to local equilibrium.
\section{Classical Reaction Kintetics}
The subject of reaction kinetics in classical chemistry is concerned with the temporal dynamics of a chemical reaction and its reaction rate. One usually studies a general reaction
\begin{equation}
\sum_{i} \mu_i A_{i} \overset{k_{ab}}{\underset{k_{ba}}{\rightleftarrows}} \sum_{j} \nu_j B_{j},
\label{eq:reaction_1}
\end{equation}
where the numbers $\mu_i$ and $\nu_j$ are referred to as \emph{stoichiometric coefficients} and $k_{ab}$ and $k_{ba}$ are called \emph{reaction constants}. If the reaction is elementary, i.e.\ there are no catalytic, intermediate, or concurrent reactions, the rate equation for the concentration of species $A_i$, denoted as $[A_i]$, can be inferred straightforwardly \cite{connors1990chemical}:
\begin{equation}
\frac{1}{\mu_i}\frac{d[A_i]}{dt} =-k_{ab}\prod_j [A_j]^{\mu_j}+k_{ba}\prod_k [B_k]^{\nu_k}.
\label{eq:reaction_2}
\end{equation}
In the terminology of irreversible reactions, i.e.\ when either $k_{ab}$ or $k_{ba}$ equals zero, the number $\max\{M,N\}$, where $M:=\sum_j \mu_j$ and $N:=\sum_k \nu_k$, is called the \emph{order} of the reaction. Systems of parallel reactions are treated in an analogous manner. An important feature of classical reaction kinetics is that the temperature dependence of the reaction constants is described by Arrhenius' law:
\begin{equation}
k(T)\propto\exp\left(\frac{-E_A}{\kappa T}\right),
\end{equation}
where $E_A$ represents the \emph{activation energy} of the corresponding reaction. Microscopically this means the statistical fluctuations drive the reaction.
This model of classical chemical reaction kinetics has been validated by its long term success in describing a wide range of astonishing phenomena from the oscillating Belousov-Zhabotinskii reaction \cite{belousov1959periodic,zhabotinsky1964periodical} to deterministic chaos \cite{epstein1996nonlinear}. However, classical reaction kinetics is still an active field, e.g., the question of the existence of an equilibrium was only recently answered in \cite{baez2012course}.
\section{Proposed Framework}
The description of the dynamics of chemical reactions at ultracold temperatures falls into the framework of quantum mechanics as the basic assumption of Arrhenuis' Law, namely a disordered movement of many high-momentum particles, is no longer satisfied in the low temperature regime. Hence we must replace Arrhenius' law with a new concept. Here we propose that \emph{quantum entanglement} provides the driving source of fluctuations in the ultracold regime. This is because when the system is in a non-equilibrium pure state quantum correlations are generically induced by unitary dynamics \footnote{Unless the dynamics is trivial, i.e., noninteracting, or the initial state is an equilibrium pure state.}. An analogous situation occurs in the study of \emph{quench dynamics} for quantum systems \cite{cramer2008exact,cramer2010quantum}, i.e.\ large many-body systems where the interaction undergoes a sudden change at some fixed time. Based on this analogy, we expect the dynamics of ultracold reactions to be dominated by entanglement-induced phenomena, such as the relaxation of subsystems to a local equilibrium.
A chemical reaction is inherently a many particle problem so that second quantization is the natural framework. Hence we describe every reacting species $A_i$ with a corresponding quantum field $\hat{\psi}_{A_i}$. The underlying Hilbert space is a tensor product of the single-species Fock spaces. The role of the classical particle concentration $[A]$ is replaced by the particle density operator $\hat{n}_A=\hat{\psi}_A^\dag\hat{\psi}_A$. The dynamics of the system is induced by its Hamiltonian $\hat{H}=\hat{H}_0+\hat{H}_{\text{int}}$, where $\hat{H}_0=\hat{H}_{\text{kin}}+\hat{H}_{\text{pot}}$ represents the standard second-quantised kinetic and potential terms, respectively. The crucial part of our proposal is now the choice of a proper interaction term $H_{\text{int}}$ to describe particle conversion, such as molecule formation. Taking guidance from the classical setting we propose the following $\hat{H}_{\text{int}}$ for the elementary reaction Eq.~\Eref{eq:reaction_1}:
\begin{equation}
\hat{H}_{\text{int}}=k\int{dx}\prod_{i,j}\left(\hat{\psi}_{A_i}^\dag(x)\right)^{\mu_i}\left(\hat{\psi}_{B_j}(x)\right)^{\nu_i} + \text{h.c.},
\end{equation}
where $\hat{\psi}_{A_i}$ and $\hat{\psi}_{B_j}$ obey the canonical commutation relations (CCR) or canonical anti-commutation relations (CAR) according to particle type of the species. If there are multiple concurrent reactions one should take the sum of each of the interaction Hamiltonians $\hat{H}_{\text{int}}=\sum_i\hat{H}_{\text{int},i}$. Notice that within this framework we are restricted to the description of reversible reactions with a single reaction rate $k\equiv k_{ab}=k_{ba}$. However, with a little work, the interaction Hamiltonian can be naturally extended via the inclusion of ancillary baths to incorporate more general effects such as a self-interaction of particles, particle loss, and dissipation. For the rest of the paper, we neglect those effects. We make an additional crucial assumption, namely that the positional degrees of freedom are not excited during the reaction. Therefore, we perform a cutoff after the first term in the operator mode expansion, i.e.\ we replace field operators with annihilation operators according to $\hat{\psi}_{A_i}\rightarrow \phi^{A_i}_{0}(x) \hat{a}_{A_i}$, where $\phi^{A_i}_0(x)$ denotes the single-particle ground state wave-function of the respective species. Consequently, the density observable is replaced by the number operator $\hat{n}_{A_i}=\hat{a}^\dag_{A_i}\hat{a}_{A_i}$ and $\hat{H}_0$ simplifies to a harmonic oscillator: $\hat{H}_0=E_{A_i}\hat{n}_{A_i}$. For simplicity we also restrict ourselves to the bosonic systems although the addition of fermionic particles is straightforward. Keeping these approximations in mind we summarise in Table \ref{tab:interaction} the proposed interaction Hamiltonians for the first few elementary low-order ultracold chemical reactions.
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c}
Order&Reaction & Proposed $\hat{H}_{\text{int}}$ \\ \hline
0.&$\text{bath} \overset{k}{\rightleftarrows} A$ & $k(\hat{a}_A^\dag+ \text{h.c.})$ \\
1.&$A \overset{k}{\rightleftarrows} B$ & $k(\hat{a}^\dag_{A} \hat{a}_{B} + \text{h.c.})$ \\
2.&$A + B \overset{k}{\rightleftarrows} C$ & $k(\hat{a}_{A}^\dag\hat{a}_{B}^\dag\hat{a}_{C} + \text{h.c.})$ \\
3.&$A + B \overset{k}{\rightleftarrows} C + D$ & $k(\hat{a}_{A}^\dag\hat{a}_{B}^\dag\hat{a}_{C} \hat{a}_{D} + \text{h.c.})$
\end{tabular}
\vspace{3mm}
\caption{The proposed interaction Hamiltonians for low-order bosonic reactions.}
\label{tab:interaction}
\end{table}
To complete our proposal, we need to consider suitable initial conditions for our dynamical system. It does not make sense to initialize the system in an equilibrium state, i.e.\ an eigenstate or thermal state of $\hat{H}_0+\hat{H}_{\text{int}}$, if one is interested in dynamical effects. Rather, we assume that the system is in an eigenstate of the non-interacting part of the Hamiltionian, which corresponds to the dynamical isolation of the reacting species before the chemical reaction starts. In order to make the comparison to classical reaction kinetics as meaningful as possible, we consider the system to be initialized in a product of coherent states~\footnote{To avoid experimentally unrealistic situtations, we also restrict our considerations to states with finite particle number.}, although the generalization to different initial conditions is obvious.
Considering the scheme outlined in table \ref{tab:interaction}, we see that a zeroth-order reaction is not a chemical reaction per se, due to the absence of any interaction of different species. It can be rather understood as a connection of the considered species to some reservoir. Neglecting $H_0$, this coupling $\hat{H}_{\text{int}}=k(\hat{a}_A^\dag+\hat{a}_A)$ results in a quadratic scaling of the average particle number of the coupled species $\left<\hat{n}_A(t)\right>\propto t^2$. Notice, by way of contrast, that in the classical case we obtain a linear growth or decay of the particle concentration $\left[A\right]\propto\pm t$ according to the choice of the sign of the coupling constant.
There are two types of elementary first-order reaction. The first describes the simplest of all ultracold chemical reactions, namely a simple particle conversion between two species. The Hamiltonian $H_{\text{int}}=k(\hat{a}^\dag_{A} \hat{a}_{B} + \text{h.c.})$ models a linear interaction between two quantum fields (e.g.\ a beam-splitter in quantum optics \cite{Mar-Sarao_08}). It is exactly solvable in the sense that it can be linearly transformed into decoupled harmonic oscillators. The average particle number of each species therefore periodically oscillates, in contrast to the classical first-order reaction, which relaxes to a stationary state. The other type of first-order reaction is modelled by $H_{\text{int}}=k(\hat{a}^\dag_{A} \hat{a}_{B}^\dag + \text{h.c.})$ and describes the production of ``pairs'' from a bath. This interaction is familiar in quantum optics where it models two-mode squeezing.
In order to describe more complex chemical reactions, e.g.\ the formation of a molecule from two particles, we have to go beyond first-order reactions. The corresponding interaction Hamiltonians describe non-quadratically interacting field theories. These models are not generally exactly solvable and we must employ approximations, heuristics, and numerical methods to study their dynamics.
\section{Diatomic Molecule Formation}
Here we study to the most elementary second-order reaction, \emph{diatomic molecule formation} \cite{yurovsky2006formation}:
\begin{equation}
A + A \overset{k}{\rightleftarrows} A_2
\label{eq:molecule}
\end{equation}
The stoichiometric coefficient for the atomic species $\mu_A$ equals two and the molecular coefficient $\nu_{A_2}$ is one. According to our proposed framework this reaction is modelled by the Hamiltonian:
\begin{equation}
\hat{H} =E_{A}\hat{n}_A +E_{A_2}\hat{n}_{A_2} + k(\hat{a}_A^\dag\hat{a}_A^\dag\hat{a}_{A_2} + h.c.),
\label{eq:quantum_hamiltonian}
\end{equation}
where $E_A$ and $E_{A_2}$ label the corresponding ground-state energies. This Hamiltonian has been the subject of much recent research as it also models second harmonic generation \cite{walls1972nonlinear} or two-photon down conversion \cite{hillery1990photon}, as well as molecule formation via coherent photoassociation in an atomic BEC \cite{javanainen1999coherent,javainen2,santos_links}. Nonetheless, a comprehensive investigation of its dynamical regimes relevant for the reaction kinetics of ultracold chemistry is missing.
We are interested in the kinetics of the reaction \eref{eq:molecule}, and in particular the role of entanglement, so we restrict our analysis to the interacting part of the Hamiltonian and assume the reaction rate to be much larger than the ground state energies, i.e.\ $k \gg |E_A| + |E_{A_2}|$.
Additionally, the overall particle number operator $\hat{N}_{tot}=\hat{N}_A+2\hat{N}_{A_2}$ is a conserved quantity for the dynamics, hence we can restrict our study to the reduced dynamics of the atomic mode, writing $N \equiv N_A$. The behaviour of the molecular mode can be deduced straightforwardly.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{Approximation.pdf}
\caption{The dynamics of the atomic occupation number expectation value $\langle \hat{N} \rangle$ with respect to the
rescaled time $\tau=tk/\hbar$ for the diatomic molecular reaction $A + A \rightleftarrows A_2$, with $N=500$ particles. The quantum and mean field trajectories coincide until they separate at $\tau_{\rm MF}$. From there on the quantum trajectory approaches a stationary value $\overline{N}$ via entanglement-induced damped oscillations.}
\label{fig:1}
\end{figure}
In Fig~\ref{fig:1} we have plotted the full quantum time evolution of the expectation value of $\hat{N}$ (obtained via exact diagonalisation) together with the mean field prediction for the dynamics of the reaction. The system is initially in a product state of a coherent atomic state and a completely depleted molecular state. Mean field theory predicts a complete inversion of the population, where the system is driven to an unstable fixed point \cite{vardi2001quantum}. However, considering the full quantum solution, we identify three different dynamical regimes: First, a \textit{semi-classical} regime, where the quantum and mean field dynamics coincide. At the \emph{breakdown time} $\tau_{\rm MF}$ the full solution drifts away from the mean field approximation \cite{vardi2001bose} and the semi-classical regime transitions to an intermediate \textit{evanescent regime}, where the quantum trajectory oscillates with an increasingly damped amplitude. Eventually, the system reaches the \textit{asymptotic} regime, where the expectation value of the population imbalance relaxes to a stationary value $\overline{N}$.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{Entropy_plot.pdf}
\caption{Evolution of the quantum entanglement (in terms of the von Neumann entropy of the atomic reduced density operator $\rho_A$) between the atomic and molecular modes for the reaction Eq.~(\ref{eq:molecule}) with respect to rescaled time $\tau=tk/\hbar$. For an initially coherent atomic state the amount of entanglement is close to zero for short times. In the vicinity of the breakdown time $t_{MF}$ the entanglement rapidly increases and stays roughly at a constant level for later times.}
\label{fig:2}
\end{figure}
We can understand the three different dynamical regimes by studying the time evolution of the quantum entanglement between the atomic and molecular modes. These results are shown in Fig.~\ref{fig:2}. In the semi-classical regime we see a rapid increase of the entanglement at the beginning of the reaction, which is necessary for the formation of molecules. It is initially rather surprising that the mean field approximation works as well as it does in the semi-classical regime given that the state rapidly becomes entangled and is not well-modelled via a product ansatz. A possible explanation is that the entanglement evolution in the semi-classical regime is typical of that produced by \emph{integrable interactions} \cite{hines2003entanglement}, at least until the breakdown time $\tau_{\rm MF}$. After the breakdown time, in the evanescent regime, the system rapidly reaches the maximum available entanglement and begins to explore the full Hilbert space. Soon after, it enters the asymptotic regime where it ergodically evolves through highly entangled states. It remains in the asymptotic regime until it experiences a quantum revival.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{CompareData.pdf}
\caption{The time-averaged relative mean value of atoms and the fluctuations around it against the overall particle number for the reaction Eq.~(\ref{eq:molecule}). As the particle number is increased the fraction of atoms in the asymptotic regime and the relative temporal fluctuations decrease.}
\label{fig:3}
\end{figure}
The dynamical behavior exhibited by the reaction \Eref{eq:molecule} is reminiscent of the local relaxation observed in quenched many particle quantum systems \cite{cramer2008exploring}. This hypothesis is supported by studying the the time-averaged fluctuations $\Delta N^2\!=\!\lim_{\tau \rightarrow \infty} \frac{1}{\tau}\int_0^\tau dt\left(\langle{\hat{N}(t)}\rangle-\overline{N}\right)^2$ relative to the mean value $\overline{N}\!=\!\lim_{\tau \rightarrow \infty} \frac{1}{\tau}\int_0^\tau dt\langle\hat{N}(t)\rangle$, plotted in Fig.~\ref{fig:3}: we find that the fluctuations decrease as the particle number $N$ is increased. However, the mechanism leading to the local relaxation observed in quenched dynamics is slightly different to that found here. In quenched many particle systems the incoherent interference of localised excitations travelling at different velocities leads to a cumulative effect of relaxation. However, in our case, we have an interaction between just two modes and the relaxation we observe here is directly related to the growth of entanglement between them, namely the loss of coherence, or purity, of the reduced density operators.
Finally, we point out that the relaxation behavior in the asymptotic regime is remarkably similar to the classical high temperature kinetics of \Eref{eq:molecule}, which relaxes to the fixed point $\left[A \right]^2=\left[A_2 \right]$, even though our system is always in a global pure state.
We obtained $\overline N$ and $\Delta N^2$ via exact diagonalisation. Although the full Hamiltonian has degenerate eigenvalues, the dynamical problem, due to the conservation of $\hat{N}_{\text{tot}}$, can be separated into finite-dimensional problems with no degeneracy. Exploiting this we find that the time-averaged expectation value of the atoms and the molecules coincides with the predictions given by the diagonal ensemble $\overline{N}_{\text{ens}}\!:=\!\sum_\alpha|c_{\alpha}|^2 N_{\alpha,\alpha}$, where $c_\alpha=\bra{\psi_{\text{in}}}\alpha\rangle$ and $N_{\alpha,\beta}=\bra{\alpha}\hat{N}\ket{\beta}$ \cite{neumann1,rigol2008thermalization} are the coefficients in the complete energy eigenbasis. Moreover, the time-averaged fluctuations around this mean value can be obtained via $\Delta N_{t}^2\!=\!\sum_{\alpha\neq\beta}|c_{\alpha}|^2|c_{\beta}|^2 |N_{\alpha,\beta}|^2$ \cite{Srednicki}. However, the system \emph{does not thermalize} as the predicted expectation values do not coincide with those of the microcanonical ensemble.
For an experimental implementation of \Eref{eq:molecule} we consider ultracold Cesium atoms:
\begin{equation}
\mathrm{Cs} + \mathrm{Cs} \overset{k}{\rightleftarrows} \mathrm{Cs}_{2}
\end{equation}
with $\mathrm{Cs_{2}} = |^{3}\Sigma^{+}_{u}, v = 0, J = 0 \rangle$ and $\mathrm{Cs} = |F = 4, m_{F} = 4 \rangle$. The chosen atomic state has pure triplet spin character, and we neglect spin mixing effects due to Spin-Orbit coupling, therefore excluding any vibrational relaxation to the singlet ground state. Potential energy curves of the states under consideration are available through \cite{xie2009experimental,krauss1990effective,allouche2012transition}. $E_{A}$ and $E_{A_2}$ are given by an optical trap confinement and will be different and tunable due to the different dynamic polarizabilities of the molecular and atomic state \cite{vexiau_optimal_2011}. For our calculation, $E_{M} = 1.05, E_{A} = 1.05 (\frac{1}{2}m \omega_{A}^{2})$ with $ \omega_{A} = 50\mathrm{Hz}.$ The reaction is realized through a 2-color photoassociation in continuous Raman configuration. For the molecule conversion, one has $k = \frac{k_{1} k_{2}}{2\Delta}$, where $k_{i}$ (i = 1,2) are the individual Rabi frequencies and $\Delta$ the one-photon detuning. We evaluate the corresponding wavefunctions and transition dipole matrix elements to and from the intermediate states $|b^{3}\Pi_{u}, v', J'=0 \rangle$ and $|c^{3}\Sigma_{g}^{+}, v', J'=0 \rangle $. As an example, for a transition to the $|b^{3}\Pi_{u}, v', J'=0 \rangle$ manifold via lasers operating around 1275nm and 1232nm respectively, our calculation yields Rabi frequencies $2\pi \times 1.17\mathrm{MHz} [0.4 \mathrm{kHz}] \sqrt{I/\mathrm{mW}/\mathrm{cm}^{2}}$ for the molecular [atomic] coupling respectively. Therefore for a typical Raman setup ($\Delta=750 \mathrm{MHz}, I_{1} = I_{2} = (10\mathrm{mW}/0.25 \mathrm{mm}^{2})$, hence $k = 1.25 \mathrm{kHz}$), $k > |E_A| + |E_{A_2}|$ is safely fulfilled.
\input{example}
\section{Summary and Outlook}
We have presented a proposal to systematically investigate the kinetics of ultracold chemical reactions. The dynamical analysis for the most elementary bosonic examples already implies that entanglement plays a major role in the formation of ultracold molecules. This leads to a relaxation of the atomic and molecular subsystems resulting in the stationarity of local observables similar to quenched systems. Moreover, considering two concurrent elementary reactions amounts to complex dynamical behaviour like Hamiltonian chaos in the mean field approximation. Extending our analysis to fermionic or mixed systems and validating the predicted phenomena against experimental data are the next natural steps to be taken.
\section*{Acknowledgments} Helpful discussions with J.\ Eisert, R.\ Werner, E.\ Schreiber, S.\ Schmidt, L.\ C.\ Venuti, B.\ Neukirchen, A.\ Werner and A.\ Milsted are most gratefully acknowledged. This work was supported by the ERC grants QFTCMPS, SIQS, and POLAR and by the cluster of excellence EXC 201 Quantum Engineering and Space-Time Research.
\section*{References}
|
\section{Introduction}
\label{sec:intro}
In recent years, Twitter has established itself as a popular medium for public mass-personal communication \cite{Wu2011}.
Our community has studied different aspects of Twitter communication, such as network structures (e.g., \citeauthor{Juergens2011} \citeyear{Juergens2011}; \citeauthor{Larsson2012} \citeyear{Larsson2012}), conversational practices (e.g., \citeauthor{Honeycutt2009} \citeyear{Honeycutt2009}; \citeauthor{Huang2010} \citeyear{Huang2010}), or analyses of dynamics (e.g., \citeauthor{Becker2011} \citeyear{Becker2011}; \citeauthor{2011ausvotes} \citeyear{2011ausvotes}).
While this work has advanced our understanding about communication on Twitter, we know little about how these different perspectives can be integrated and quantified, and how they relate to conversational practices in a political context. Future computational social scientists would certainly benefit from computational tools and instruments that translate theoretical constructs from sociology to quantifiable measures that are amenable to computation and therefore can be applied on larger scales.
For example, it can be expected that in online conversations, political parties attempt to define themselves by adopting unique topical foci and conversational styles that are inherently difficult to assess. Having computational methods available to track and assess these styles would greatly increase our sociological understanding of such processes. In Figure \ref{networks}, we present preliminary evidence for this idea. The three networks depict the static outcome of communication among politicians on Twitter dissected according to politicians' following, retweeting, and mentioning practices. Table \ref{statistics} provides accompanying statistics about these networks. Interesting differences are present. For example, homophily effects are weakest for mentioning practices and the variance of homophily is also largest for this practice (cf. Table \ref{statistics}). This indicates that all parties tend to follow and retweet members of their own party but may use mentioning for inter-party debate. This finding is in line with previous results from a study on Twitter use before the U.S. midterm election of 2010 \cite{Truthy_icwsm2011politics}. Sound computational methods that would enable such qualitative insights into online conversational practices on a wide scale would represent a useful addition to the arsenal of social science methods.
\textbf{Research questions:} These interesting yet preliminary differences in the static outcomes of conversational practices lead us to expect to see differences in the socio-cultural processes that produce these outcomes as well. While our observations above suggest that following, retweeting, and mentioning exhibit unique traits, we know little about the particularities and idiosyncrasies of these practices and their party-specific adoption.
For example, what are the different purposes that tagging, retweeting, and mentioning serve in online political conversations? How do they differ from each other? How consistently are different practices used across different parties? Moreover, one would also expect that these different practices are effected to varying extents by external events or factors. For example, how would a (TV) debate or the day of the election itself effect the online conversational practices of parties in general or individual parties specifically?
\textbf{Approach:} In this paper, we set out to answer these and related questions by presenting and applying a computational approach to assessing the socio-cultural dynamics of online conversational practices \emph{over time}.
Our work is rooted in relational sociology, specifically in theoretical work that considers episodes of stability and change in practice \cite{Mohr2008,White2008,Fuhse2009,Padgett2012}.
We are interested in making different aspects of online conversational practices, in particular cultural focus, - similarity, and - reproduction as well as institutions and punctuations, amenable to quantitative measurements.
In doing so, we follow a \emph{deductive style of research}, deriving measures from a theoretical discussion of sociological constructs. While this enables us to root our measures in theory, it makes validation a challenging endeavour. To evaluate our approach nonetheless, we choose to apply it to a particular case, i.e., to conversational practices of political parties on Twitter before, during, and after the German federal election of September 22nd, 2013. This enables us to generate insights into the practical utility of our deductive approach in a real world scenario, as well as into the conversational practices of the case itself.
\textbf{Contributions:} The contributions of our work are threefold:
First, we present and discuss several sociological constructs related to conversational practices on a theoretical level.
Second, we present a computational approach that deduces measures for each of the sociological constructs of interest. While the constructs are grounded in sociological theory, the proposed measures stem from computer science, social science, information science, and related fields.
Third, to demonstrate the utility of our computational approach, we conduct a case study on the German federal election 2013 and present empirical insights into the conversational practices of German politicians during the course of this event.
The paper is structured as follows: After related work we introduce the sociological background and constructs which form the theoretical foundation of our computational approach which we present subsequently. Finally, we describe our empirical study which demonstrates the utility of our approach, discuss our empirical insights, and conclude our work.
\section{Related Work}
\label{sec:relWork}
Previous work which is relevant for our research focused either on analyzing the hashtagging, mentioning, and retweeting behavior of Twitter users in general or on analyzing the role of Twitter in a political context. To our best knowledge, ours is the first work which coherently studies the tagging, mentioning, and retweeting behavior of politicians as socio-cultural processes, and presents comparative analysis that yield unique insights into online conversational practices of politicians and political parties on Twitter.
\smallskip
\textbf{Hashtags, Mentions, and Retweets:} Twitter is used for many purposes, including the reporting of daily activities, communicating with other users, sharing information, and reporting, or commenting on, news \cite{Java2007}. As such it enables several \emph{conversational practices}.
Hashtags are primarily used to describe news or communications with others and to find other users’ tweets about certain topics. Since tagging behavior is inspired by the observed use of hashtags in a users’ network \cite{Huang2010}, coherent semantic structures emerge from hashtag streams \cite{Wagner2010}.
Retweeting is the forwarding of other users’ tweets. By 2010, conventions as to how, why, and what users retweet had emerged, but the practice had not yet stabilized \cite{Boyd2010}. The retweetability of a Twitter message is related to its informational content and value and the embeddedness of its sender in following networks \cite{Suh2010}.
Mentions, sometimes called @mentions or replies, emerged to be Twitter’s convention for the interactive use of addressing others, although it is also being used for other purposes like referencing. In 2009, most tweets without a mention reported daily activities while tweets with @ signs exhibited much higher variance in terms of topics and types of content \cite{Honeycutt2009}.
\smallskip
\textbf{Elections on Twitter:} With the rise of social media many researchers got interested in exploring the role of Twitter for politics.
Scientists from different backgrounds started exploring to what extent Twitter can predict election results with contradicting conclusions.
Some research suggests that election results can be predicted by analyzing Twitter (e.g., \citeauthor{livne2011party} \citeyear{livne2011party}), while \citeauthor{Jungherr2013votes} (\citeyear{Jungherr2013votes}) shows that, at least for the German multi-party system, election result can not be predicted using Twitter. \citeauthor{MetaxasMG11} (\citeyear{MetaxasMG11}) conduct a meta study and conclude that electoral predictions using the published research methods on Twitter data are not better than chance.
\textbf{Political discourse on Twitter:} Besides the interest in the voting behavior of persons and its reflection on Twitter, researchers also got interested in studying political discourse on Twitter. For example, \citeauthor{Truthy_icwsm2011politics} (\citeyear{Truthy_icwsm2011politics}) analyze the retweet and mention networks from 6 weeks leading up to the 2010 U.S. midterm election. Interestingly, the authors find extremely limited connectivity between right- and left-leaning users in the retweet network, but not in the mention network. This indicates that retweets and mentions are different conversational practices.
The work of \citeauthor{Juergens2011} (\citeyear{Juergens2011}) shows that on Twitter new gatekeepers and ordinary users tend to filter political content based on their personal preferences. Therefore, political communication on Twitter might be highly dependent on a small number of users, critically positioned in the structure of the network.
\textbf{Politicians on Twitter:} In previous research \citeauthor{thammbleier2013} (\citeyear{thammbleier2013}) found that retweets are motivated by professional uses while replies are used mainly for personal communication.
\citeauthor{Schweitzer2011} (\citeyear{Schweitzer2011}) explored political e-campaigns in Germany and concluded that they increasingly reflect those patterns of traditional election coverage that have been held accountable for rising political alienation among the public, i.e., strategic news and extensive negativism.
Following the suggestion of \citeauthor{Murthy2012} (\citeyear{Murthy2012}) who emphasizes the importance of a sociological understanding of Twitter, we now present an approach based on sociological constructs that allows to assess political online conversations.
\section{Theoretical Constructs}
\label{background}
In this section, we elaborate the theoretical sociological constructs behind conversational practices which build the basis of our computational approach.
In doing so we start from sociological network theory known as relational sociology \cite{Carley1991,Mohr2008,White2008,Fuhse2009,Padgett2012}.
The central premise is that social life is complex and stochastic and that \emph{identities}, which can be persons, groups, or higher-level agents, try to gain \emph{control} over the ensuing uncertainty through regularities. In many contexts, control is gained by collectively forming a densely clustered and culturally coherent community, triadic closure and homophily being the mechanisms \cite{mcpherson2001birds,Kossinets2009}.
We first discuss an accessible understanding of culture and then offer an idealtype of how culture is reproduced by styles of practice. Because of the stochastic nature of social formations, we finally deal with system perturbation.
\smallskip
\textbf{Cultural Facts and Homophily:}
Gaining control is a project in the socio-cultural space-time of simultaneous emergence of culture and feedback on structure: As agents act, they are not only engaged in transactions with other agents, they also make reference to pieces of information or \emph{facts} such as ideas, beliefs, concepts, symbols, knowledge, etc. From this practice, a culture emerges which can be understood as a distribution of referenced facts \cite{Carley1991}. But as agents in transactions are confronted with facts, existing culture positively or negatively feeds back onto agent behavior.
Computer simulations are able to produce the densely clustered and culturally coherent communities which are ubiquitously found if cultural similarity breeds social connection \cite{Carley1991,Axelrod1997}.
\smallskip
\textbf{Styles, Institutions, and Reproduction:}
If an identity has gained control and maintains it over time it has a \emph{style}. Styles are inert mechanisms of reproduction. It is due to the existence of such dynamical regularity that identities are predictable \cite{Kosinski2013,schoen2013power}.
Cycles of reproduction (styles) simultaneously determine practices of individuals and groups \cite{Mohr2008,White2008}.
Distributions of facts like words, religious beliefs, surnames, and citations are fat-tailed \cite{Clauset2009}. \emph{Institutions} are cultural facts that are stably reproduced over time. Thereby, they become popular and are consequently found in the fat tail of probability distributions.
Mechanisms of cumulative advantage \cite{Simon1955,Dellschaft2008,Papadopoulos2012} are theoretically compatible and have been shown to be capable of producing these skewed distributions.
However, not all facts in the tail are necessarily institutions, since facts can also become popular due to short activity bursts rather than prolonged activity.
\smallskip
\textbf{Punctuations:}
Continuity and normality is just one side of social life.
Only in equilibrium persons and groups can reproduce uninterruptedly. In reality, social life is \emph{punctuated} by minor and major events that interrupt the normal flow of reproduction. Such punctuations may reflect scientific \cite{Kuhn1962}, political \cite{Brunk2001}, or economic \cite[Ch. 6]{Padgett2012} innovations or generally perturbations originating from inside or outside the observed system. Any account of socio-cultural processes must account for both normality and change \cite{Padgett2012}.
It is both reproduction and punctuation that we study in this paper.
\footnotesize
\begin{table*}[h!b!]
\centering
\caption{\textbf{Operationalization of sociological constructs}.
}
\begin{tabularx}{\textwidth}{m{0.2\textwidth}|m{0.35\textwidth}|m{0.35\textwidth}}
\hline
\bf{Theoretical Construct} & \bf{Measure} & \bf{Description} \\ \hline
Cultural Focus $F$ & Shannon Entropy \cite{Shannon2001}& \emph{How strongly does a party focus on cultural facts?} \\ \hline
Cultural Similarity $S$ & Cosine Similarity \cite{Baeza-Yates1999} & \emph{How similar are parties in terms of their culture vectors?} \\ \hline
Cultural Reproduction $R$ & Rank Biased Overlap \cite{Wagner2014}& \emph{How stable is a party's culture vector over time?} \\ \hline
Institutionness $I$ & Hirsch Index \cite{Hirsch2005} & \emph{How many weeks is a fact referenced that many times?} \\ \hline
Burstiness $B$ & Kleinberg's Burst Weight \cite{Kleinberg2003} & \emph{How popular is a fact in a given period relative to other facts?} \\ \hline
\end{tabularx}
\label{operationalization}
\end{table*}
\normalsize
\begin{figure*}[ht!]
\centering
\subfigure[Tagging Focus]{\includegraphics[width=0.23\textwidth]{TagEntr.pdf}
\label{TagEntr}}
\subfigure[Retweeting Focus]{\includegraphics[width=0.23\textwidth]{RTEntr.pdf}
\label{RTEntr}}
\subfigure[Mentioning Focus]{\includegraphics[width=0.23\textwidth]{AtEntr.pdf}
\label{@Entr}} \\
\subfigure[Tagging Similarity]{\includegraphics[width=0.23\textwidth]{TagCos.pdf}
\label{TagCos}}
\subfigure[Retweeting Similarity]{\includegraphics[width=0.23\textwidth]{RTCos.pdf}
\label{RTCos}}
\subfigure[Mentioning Similarity]{\includegraphics[width=0.23\textwidth]{AtCos.pdf}
\label{@Cos}} \\
\subfigure[Tagging Reproduction]{\includegraphics[width=0.23\textwidth]{TagStab.pdf}
\label{TagStab}}
\subfigure[Retweeting Reproduction]{\includegraphics[width=0.23\textwidth]{RTStab.pdf}
\label{RTStab}}
\subfigure[Mentioning Reproduction]{\includegraphics[width=0.23\textwidth]{AtStab.pdf}
\label{@Stab}} \\
\subfigure[Tagging Frequency]{\includegraphics[width=0.23\textwidth]{TagFreq.pdf}
\label{TagFreq}}
\subfigure[Retweeting Frequency]{\includegraphics[width=0.23\textwidth]{RTFreq.pdf}
\label{RTFreq}}
\subfigure[Mentioning Frequency]{\includegraphics[width=0.23\textwidth]{AtFreq.pdf}
\label{@Freq}}
\caption{\textbf{Online conversational practices of the six main political parties during the German federal election 2013:} Tagging, retweeting, and mentioning practices are shown over a period of 13 weeks. The three conversational practices (columns) are assessed using measures (rows) summed up in Table \ref{operationalization}.
For the first three rows, curves for an average party are shown, obtained by averaging the six party scores. Shaded regions show standard deviations. The weeks of the election (straight line) and the TV election debate (broken line) clearly mark episodes of irregular activity. The last row shows the total number of references made to facts and is included as reference.}
\label{dynamics}
\end{figure*}
\section{Operationalization of Constructs}
\label{research_design}
Next, we describe how we operationalize the sociological constructs described in the previous section for assessing online conversational practices of politicians and political parties on Twitter.
\textbf{Approach:} The central idea behind conversational practices as socio-cultural processes is that a transaction -- a user using a hashtag, a user following another user, a user retweeting another user, or a user mentioning another user -- involves a social subject (the referencing user) and a cultural object (a referenced hashtag, followee, retweetee, or mentionee).
In the case of following, retweeting, and mentioning, subjects and objects are of the same type, exemplifying the duality of the social and the cultural.
In the empirical study which we will present later we are interested in how politicians practice online conversations on Twitter, concretely we study their mentioning, retweeting, and hashtagging practices. We use aggregations of users (parties) as \emph{objects of study} and referenced facts (i.e., user handles and hashtags) as \emph{units of analysis}.
For analyzing dynamics, tweets are batched into bins of one week each.
Though our operationalization is specific to the context of our empirical study (politics on Twitter), our computational approach for assessing online conversational practices is general and can be applied to study the practices of other groups of users on Twitter (or even on other stream-based social systems which provide support for conversations and allow to observe the referencing of facts over time).
\subsection{Cultural Focus and Similarities}
In the following, we interpret the observable cultural objects (i.e., hashtags and user handles) of conversational practices (i.e., retweeting, hashtagging, and mentioning) of Twitter users as cultural \emph{facts}.
The culture of a group of users (which corresponds to a party $i$ in our case) is represented by a vector $\sigma_{i}=(a_1, a_2, ..., a_n)$ where the elements are the frequencies of (a) the hashtags that members of party $i$ used, (b) the users that members of party $i$ retweeted, or (c) the users which members of party $i$ mentioned within the observation period.
\smallskip
\textbf{Cultural Focus:}
To quantify the extent to which a party $i$ reveals a \emph{cultural focus} $F$ on few selected hashtags or users (facts), the normalized Shannon entropy \cite{Shannon2001} -- a measure of disorder -- of the party's fact vector is used:
\begin{equation}
F(\sigma_i) = 1-\frac{-\sum_{j=1}^{n} p(a_j)*\log_2 p_(a_j)}{\log_2(n)}
\end{equation}
Here, $p(a_j)$ corresponds to the frequency of a cultural fact $a_j$ for party $i$ divided by the frequency of all other facts of that party.
Because the denominator holds the maximum entropy and normalized entropy is subtracted from 1, $F$ falls in the range $[0,1]$ where 0 means that there is no focus (highest entropy) and 1 that the focus is highest (no entropy).
\smallskip
\textbf{Cultural Similarity:}
Culturally similar agents are more probable to interact with each other. To reveal potential homophily effects, cultural similarities of parties are studied over time.
We propose to measure the \emph{cultural similarity} $S$ of two parties $i$ and $j$ using the cosine similarity \cite{Baeza-Yates1999} of their fact vectors:
\begin{equation}
S(\sigma_i, \sigma_j) = \frac{\sigma_{i} \cdot \sigma_{j}}{\parallel \sigma_{i}\parallel \parallel \sigma_{j} \parallel}
\end{equation}
Because facts cannot have negative frequencies, similarities are in the range $[0,1]$ where 0 indicates no similarity and 1 highest similarity. Cosine measures the similarity for individual party pairs. To obtain a score for a single party $i$, we compute its average cosine similarity to all other parties $j$. We consciously do not weight scores by user size or tweet volume because we operate on the level of parties as objects of study.
\subsection{Styles, Institutions and Reproduction}
\smallskip
\textbf{Cultural Reproduction:}
Styles are mechanisms of reproduction of focus.
To study and quantify styles, we proceed in two steps. First, we measure the reproduction of cultural focus from week to week. In a second step, the stability of reproduction over time is studied to generate a quantitative judgement about styles.
Reproduction is operationalized by an extended version of the Rank Biased Overlap (RBO) metric \cite{Wagner2014} which allows to measure stability in social streams. The \emph{cultural reproduction} $R$ of a party $i$ is defined as follows:
\begin{equation}
R_{i}(\sigma_{i}, p) = (1-p) \sum_{d=1}^{\infty} \frac{2 \cdot \sigma_{i}^{1:d}(t1) \cap \sigma_{i}^{1:d}(t2) }{|\sigma_{i}^{1:d}(t1) + \sigma_{i}^{1:d}(t2)|} p^{(d-1)}
\end{equation}
The cultural fact vectors $\sigma_{i}(t1)$ and $\sigma_{i}(t2)$ of party $i$ hold the frequencies of the facts which have been observed
at $t1$ and $t2$, respectively.
These vectors are not necessarily conjoint since new facts can be introduced at any point in time.
Let $\sigma_{i}^{1:d}(t1)$ and $\sigma_{i}^{1:d}(t2)$ be the ranked vectors at depth $d$.
The scores fall in the range $[0,1]$ where 0 means that the culture of a party $i$ at $t1$ is completely different from its culture at $t2$ (culture is unstable, turnover is maximum), and 1 means that it is identical (i.e., does not change over time and is therefore perfectly stable).
The parameter $p$ ($0 \leq p < 1$) determines how steep the decline in weights is.
The smaller $p$ is, the more tail-weighted the metric is, the tail being the high frequency part of the distribution.
We chose $p=0.9$ which means that facts in the fat tail are most relevant for assessing the cultural stability of a party compared to changes in the head of the distribution (86\% of the weight is given to the 10 highest ranked cultural facts if $p=0.9$).
This makes intuitively sense because, if a party changes a cultural fact which is core to them, they will become very unstable according to our measures, while, if they change a fact which is just one among many with low frequencies, this will only have a small impact on their stability.
Mathematically, a style is present when $\delta R_{i}/\delta t \geq 0$. If $R$ decreases over time, the party will dissolve into chaos as less and less facts are reproduced. If $R$ increases, the party will eventually freeze as reproduction reaches 1 and cultural facts will forever be the same. From a systems perspective, a constant turnover of facts is desired.
\smallskip
\textbf{Institutionness:}
Cultural reproduction is strongly related to the construct of stable facts called institutions which are expected to be found in the fat tail of the probability distribution of cultural facts.
Two notions resonate with the idea of stable facts: that they are (a) highly referenced and (b) continuously highly referenced.
It seems straightforward to divert the Hirsch index \cite{Hirsch2005} from its intended use.
Originally proposed to quantify an individual's scientific research output, a scientist has an index $h$ if $h$ of his papers have at least $h$ citations each, and all but $h$ papers have no more than $h$ citations each.
Applied to our temporal case, a fact has an index $h$ if for $h$ weeks it is referenced at least $h$ times, and in all but $h$ weeks no more than $h$ times.
A normalization -- one part of which is a division of a paper's citation rate by the citation rate of an average paper in the subject area -- has been proposed to make the Hirsch index comparable across scientific disciplines with different publication and citation characteristics \cite{radicchi2008universality}.
In analogy, a fact $a$ has index $h$ if for $h$ weeks it is referenced at least $h/h_{0}$ times, and in all but $h$ weeks no more than $h/h_{0}$ times, where $h_{0}$ is the week-specific reference rate of an average fact.
Let's call this index the \emph{institutionness} $I$ of fact $a$.
Since data ranges over 13 weeks, $I$ is in the range $[0,13]$.
\subsection{Punctuations}
\textbf{Burstiness:}
Having a style means being predictable. Punctuations interrupting the normal flow of reproduction leave traces in Twitter in the form of activity bursts.
To operationalize, we refer to \citeauthor{Kleinberg2003}'s (\citeyear{Kleinberg2003}) observation that social streams are ``punctuated by the sharp and sudden onset of particular episodes''.
A fact is considered to burst if it leaves a period where its popularity relative to other facts is small and enters a period in which its relative popularity is reasonably large.
Suppose there are $n$ batches of transactions made by party $i$ (in our case $n=13$ since we use weeks as batches). The $t$-th batch contains $r_t$ transactions with references to fact $a$
out of a total of $d_t$ transactions.
Let $R=\sum_{t=1}^{n}r_t$ and $D=\sum_{t=1}^{n}d_t$ and $p_{s}=\frac{R}{D}2^{s}$.
To reveal the burstiness of a fact $a$, we compute for each week $t$ the party-specific cost of that fact to be transformed in a burst state $s$:
\begin{equation}
\gamma(s, r_t, d_t) = -\ln\left(\binom{d_t}{r_t}p_{s}^{r_t}(1-p_{s})^{d_t-r_t}\right)
\end{equation}
The weight of a fact's burst -- its \emph{burstiness} $B$ -- in the period $[t1,t2]$ is the improvement in cost by being in state $s=1$ for the period $[t1,t2]$:
\begin{equation}
B(a,t1,t2) = \sum_{t=t1}^{t2}\left(\gamma(0, r_t, d_t)-\gamma(1, r_t, d_t)\right)
\end{equation}
In words, for each bursting fact $a$ its burstiness in a period $t1$ (onset) to $t2$ is obtained. Facts can burst multiple times.
There is no upper bound for a fact's burst weight. The final burstiness measure is normalized by the weight of a party's strongest burst and is, therefore, also in the range $[0,1]$.
\tiny
\begin{table*}[h!b!]
\caption{\textbf{Institutionness $I$ and burstiness $B$ of selected facts:} The election received continuous attention by all parties (high $I$ scores) but bursted only for some parties (it was the Left's strongest burst). The short-lived TV debate, on the other hand, bursted strongly for all parties. In the process, the one debater's hashtag \#merkel and the other debater's username @peersteinbrueck also bursted for almost all parties. Square brackets give onset and end of burst episodes.}
\tabcolsep=0.11cm
\begin{tabularx}{\textwidth}{m{2.7cm}|m{3.5cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}|>{\centering\arraybackslash}m{0.4cm}>{\centering\arraybackslash}m{1cm}} \hline
& & \multicolumn{2}{c| }{{\bf{CDU/CSU}}} & \multicolumn{2}{c| }{\bf{SPD}} & \multicolumn{2}{c| }{\bf{FDP}} & \multicolumn{2}{c| }{\bf{Greens}} & \multicolumn{2}{c| }{{\bf{Left}}} & \multicolumn{2}{c }{\bf{Pirates}} \\
\bf{Fact} & \bf{Description} & \bf{\textit{I}} & \bf{\textit{B}} & \bf{\textit{I}} & \bf{\textit{B}} & \bf{\textit{I}} & \bf{\textit{B}} & \bf{\textit{I}} & \bf{\textit{B}} & \bf{\textit{I}} & \bf{\textit{B}} & \bf{\textit{I}} & \bf{\textit{B}} \\ \hline
\#btw13 & Federal election (Sep 22) & 13 & 0 & 11 & 0 & 10 & 0.69 [8,10] & 13 & 0.18 [10,10] & 11 & 1.00 [8,10] & 13 & 0.46 [10,10] \\ \hline
\#tvduell & TV election debate (Sep 1) & 3 & 1.00 [7,7] & 3 & 0.94 [7,7] & 1 & 1.00 [7,7] & 3 & 0.66 [7,7] & 1 & 0.75 [7,7] & 2 & 0.85 [7,7] \\ \hline
\#merkel & CDU/CSU, chancellor & 11 & 0.10 [7,7] & 11 & 0.25 [7,7] & 4 & 0.25 [7,7] & 13 & 0.18 [7,7] & 8 & 0.39 [6,7] & 13 & 0.10 [7,7] \\ \hline
RT sigmargabriel & SPD, chairman & 1 & 0 & 3 & 1.00 [7,8] & 0 & 0 & 2 & 0.26 [5,8] & 0 & 0 & 0 & 0 \\ \hline
RT Volker\_Beck & Greens, parliamentary managing director & 2 & 0 & 3 & 0 & 1 & 0 & 8 & 1.00 [7,7] & 1 & 0 & 1 & 0.26 [6,7] \\ \hline
@peersteinbrueck & SPD, candidate for chancellor & 5 & 1.00 [7,8] & 8 & 1.00 [6,8] & 5 & 0 & 6 & 0.78 [7,8] & 2 & 0 & 2 & 0.24 [7,7] \\ \hline
\end{tabularx}
\label{facts}
\end{table*}
\normalsize
\section{The German Federal Election 2013}
In this section we describe the dataset which we used to conduct our empirical study of three conversational practices of politicians before, during, and after the German federal election 2013. Then, we present our results. In particular, we present our findings on hashtagging, mentioning, and retweeting practices of German politicians during the 2013 elections. While our approach would also be applicable to analyze following practices (as introduced in Figure \ref{hashtagNetwork}), we could not perform this analysis since the fine-grained temporal data necessary to study the evolution of follower networks on Twitter is impossible to collect via the public Twitter API.
\subsection{Dataset}
The dataset consists of German politicians and their communicative transactions on the microblogging platform Twitter between July 20th and October 18th, 2013. This corresponds to about 9 weeks before and 4 weeks after the federal election of September 22nd. A preliminary version of the ``Twitter corpus of candidates'' described by \citeauthor{Kaczmirek2013} (\citeyear{Kaczmirek2013}) is used. Our dataset contains 1,031 user account handles, 98\% of which were the ``most relevant'' candidates for the German parliament in early 2013, and their tweets.
983 politicians were active in the sense that they either tagged a tweet or followed, retweeted, or mentioned another politician during our window of observation.
In total,
there are 123,819 tweets containing at least one hashtag, 16,292 tweets retweeting at least one other politician (identified by ``RT username''), and 25,778 tweets mentioning at least one other politician (identified by ``@ username'').
Each politician belongs to one political party. The Social Democratic party SPD is the oldest (150yrs), followed by the Conservatives CDU/CSU (68yrs), the Liberals FDP (65yrs), the Greens (Die Grünen, 34yrs), the Left (Die Linke, 24yrs), and the new, internet-affine Pirates (Piratenpartei, 7yrs).
\subsection{Results}
Before going into the details of the dynamics that created the networks shown in Figure \ref{retweetNetwork} and \ref{mentionNetwork}, we first address the two episodes that left spikes in the curves of Figure \ref{dynamics} which visualizes the main dynamics results.
Both relate to offline events.
The first is the election of September 22nd in week 10. Message volumes, and average tagging focus and retweeting reproduction rate increase towards the events and decrease afterwards.
Though the election day was early in week 10, we can see that most of the communication occurred in week 9. Therefore, the horizontal line marking the event in Figure \ref{dynamics} is in week 9.
The second is the TV election debate between the two candidates for chancellor, A. Merkel (CDU) and P. Steinbrück (SPD), of September 1st, early in week 7.
One can tell because the corresponding hashtag \#tvduell is the strongest burst for the CDU/CSU and FDP, the second strongest for the SPD and the Greens, the third strongest for the Left, and the fourth strongest for the Pirates, restricted to week 7 in each case.
See Table \ref{facts} for the burstiness of selected facts.
Despite the election being the main event in this stream, the debate marks the maximum in terms of total message frequency (cf. Figure \ref{TagFreq}) and the average tagging focus and similarity (cf. Figure \ref{TagEntr} and \ref{TagCos}). It also left spikes in the average retweeting focus and mentioning similarity (cf. Figure \ref{RTEntr} and \ref{@Cos}).
In the following, we will discuss these events, average party dynamics, as well as the peculiarities of single parties.
\smallskip
\textbf{Cultural Focus:}
The two major events in the stream have different impacts on the three practices. Figure \ref{TagEntr} shows that all parties almost monotonously concentrate their focus on hashtags towards the election (average increases from 0.13 in week 1 to 0.27 in week 10) and drastically lose focus afterwards (down to 0.11).
The TV debate increased the tagging focus of the CDU/CSU, SPD, FDP, and the Greens, but did not gain enough attention by the Left and the Pirates to do so for them.
The latter were the two parties which could not hope for being a junior partner in the coming government coalition.
The two events hardly caused changes of focus in the retweeting practice (cf. Figure \ref{RTEntr}). One can further see that the reweeting as well as the mentioning practice have a very low focus (high entropy) compared to the tagging practice and don't seem to be impacted by external events.
For the retweeting practice we observe a slight increase of focus during the TV debate. A closer look into our data reveals that the debate caused the SPD and Greens to retweet what their most vocal members had to say. The SPD's chairman S. Gabriel and the Green's parliamentary managing director V. Beck were even subject of their parties' strongest retweet bursts.
These two parties' activities contributed most to slightly increasing the retweet focus on average.
The absence of any change in the mentioning practice (cf. Figure \ref{@Entr}) reveals that, even if real world events like the debate or the election itself caused the average party to sharpen its thematic focus, it did so without changing its focus of who to talk to.
\smallskip
\textbf{Cultural Similarity:}
Even though parties increase their individual thematic focuses towards the election, they don't become more similar to each other until the election debate causes them all to increasingly focus on overlapping hashtags (cf. Figure \ref{TagCos}).
This general tagging behavior suggests that, absent punctuations, parties try to maintain distinctiveness by finding a topical niche and focusing on it.
This time, the election debate also impacts the Left and Pirates and causes them to become more similar to other parties, though they remain least similar to other parties over time.
This suggests that all parties included topics which emerged during the TV debate into their tagging practice.
Interestingly, neither of the events had an impact on the similarities between parties according to their retweeting practices, but both impacted their mentioning similarities (cf. Figure \ref{RTCos} and \ref{@Cos}).
One possible explanation for this observation is that the debate narrowed down the pool of politicians to those which were subject of the debate.
Evidence for this explanation can be found by taking a closer look at the data: both parties whose candidates were debating (CDU/CSU and SPD) experienced their strongest bursts as their members started mentioning the SPD's candidate P. Steinbrück who also bursted strongly for the Greens and received most mentions that week from the FDP.
The CDU/CSU candidate A. Merkel, who did not maintain a Twitter account, was tagged instead, her hashtag \#merkel bursting in week 7 for all parties (see Table \ref{facts} for details). This indicates that the TV debate caused changes in the mentioning and tagging practices of most parties, while neither the debate nor the election caused any party to notably start retweeting another party's voices.
\smallskip
\textbf{Cultural Reproduction:}
So far, we have shown that parties increase their thematic focus towards the election and become more similar according to their tagging practices as the election comes closer and the debate bursts.
However, it remains unclear how stably the parties reproduce their focuses.
Recall that a style is absent if cultural reproduction decreases.
Average tagging reproduction -- the stability of practice -- increases from 0.49 to 0.63 in election week 9 where it peaks and subsequently drops (cf. Figure \ref{TagStab}). The average party's stability increases, focuses are increasingly reproduced, and turnover is reduced.
But the style is strongly perturbed by the TV debate.
The sudden drop in reproduction means that the punctuation keeps parties, on average, from increasingly narrowing their focuses. They lose stability.
Again, the Left and the Pirates are not affected by this event. But least effected is the SPD.
This is because the SPD had the largest turnover of hashtags in the two weeks before the debate (i.e., the SPD was least stable in weeks 5 and 6, as one can see in Figure \ref{TagStab}). Consequently, the event did not punctuate their style.
The CDU/CSU, on the other hand, had a strong reproduction of focus and was consequently negatively affected by the event.
The debate does not have an impact on reproduction in retweeting (cf. Figure \ref{RTStab}) and mentioning (cf. Figure \ref{@Stab}).
This is expected because focus and similarity were not, or only weakly, affected by the event.
There is a slight trend in the retweet practice for all parties, except the FDP (cf. Figure \ref{RTStab}).
Parties increasingly reproducing their retweeting focus, while not becoming more similar to each other (cf. Figure \ref{RTCos}), suggests that they maintain distinctiveness in terms of who their spokespersons are -- the homophily principle at work.
But all growth of reproduction must come to an end and actually does after the election, as one can see in Figures \ref{TagStab}, \ref{RTStab}, and \ref{@Stab}. Election styles terminate and post-election styles start, from lower levels of reproduction.
Average mentioning reproduction is noisy but rather constant over time, indicating that the debate about which politician is worth commenting on is progressing with smooth turnover.
\smallskip
\textbf{Institutionness:}
Institutionness $I$ captures how long a fact is getting referenced often. Since election styles had a duration of up to 10 weeks, any fact with $I>10$ can be regarded to represent a defining factor for a party's identity.
Indeed, among the facts with $I>10$, we find environment-related hashtags only for the Greens and internet and surveillance-related tags mostly for the Pirates.
Unsurprisingly, each party's name hashtag (\#cdu, \#csu, \#spd, etc) has a score $I=13$.
Table \ref{facts} shows that the election hashtag \#btw13 has the same score $I=13$ for all parties but the FDP ($I=10$).
This points at the fundamental difference between the two events discussed above. The election is a long-term topic that endures and is repeatedly reproduced by long-term styles. The TV debate is a bursty punctuation interrupting these styles (see low institutionness and high burstiness scores for all parties in Table \ref{facts}).
Politicians getting retweeted or mentioned for more than 10 weeks are all top-runners and authorities with functions like minister, vice chairman, or political managing director.
\smallskip
\textbf{Burstiness:}
So far, there is no evidence that hashtags that were important online in the TV debate episode were actually new.
But in fact, the average proportion of hashtags not referenced before week 7 is 65\% across parties, a percentage not reached in any other week.
The increase of focus and similarity, and the decrease in reproduction in tagging are likely due to this renewal.
Of all hashtags having its burst onset in week 7, due to the selection of the sample, only a handful used by the Pirates is devoid of political relevance or meaning.
Table \ref{facts} shows examples of selected facts and their burstiness scores.
\section{Discussion}
\label{discussion}
In our empirical study, we found that the focus on hashtags is more concentrated and similarities between parties through common hashtag usage are much more pronounced than foci on, and similarities through, retweeted or mentioned politicians.
Tagging is also more prone to perturbation by offline events than retweeting or mentioning.
Retweet practices are much more stable because the users who enjoy attention tend to remain the same even as bursts of topical activity propagate through the system.
Even though studying each practice is diagnostic of socio-cultural process, studying tagging is most indicative of culture as the use of language is the ultimate form on symbolic communication \cite[Ch. 4]{Padgett2012}.
The notion of ``social'' is stronger in retweeting and mentioning, where the facts and the users referencing the facts are identical.
Furthermore, we used our approach to study the effects of a major offline event, the TV debate of the two top candidates three weeks prior to the election. It left a distinct mark in the tagging practice of politicians but only impacted some aspects of their retweeting an mentioning practice.
In addition to revealing differences between distinct practices we also found interesting differences between individual parties.
For example, while most parties increased their tagging focused during the week of the TV election debate, the focus of the Pirates and the Left was not impacted by this event.
This points at differences in what parties were trying to control: All but the latter were campaigning for becoming a part of the new government -- the latter never had a chance to be more than part of the opposition.
The results of our empirical study can be seen as anecdotal evidence for the fact that our approach is suitable for detecting differences in the conversational practices of one or several groups of users.
Regarding the objects of study, a typical dynamic for an average party was described: In the run-up to the election a party has a style of practice geared toward controlling uncertainties in communication. Certain topics are focused on and reproduced. Members of other parties are not retweeted but mentioned if an event -- in out case the TV election debate -- creates a need to do so.
After the election, attention and control is directed in another direction. Style is altered but certain institutions -- well-established (self-)references in topics and users -- remain at the core of how parties define themselves.
An interesting and plausible extension of this work would be the study of inter-party similarities rather than average similarities over longer periods of time.
Finally, our computational approach in general can also be applied to understand how non-politicians participate in political conversations as well.
\section{Conclusions}
\label{conclusion}
We have presented a computational approach to assessing online conversational practices of political parties on Twitter. Our approach is rooted in, and informed by, theoretical constructs from relational sociology, in particular cultural focus, - similarity, and - reproduction of agents as well as institutions and punctuations. We devised measures for each of these dimensions to enable the computational assessment of socio-cultural structure and dynamics of online conversational practices. We presented our approach and demonstrated its usefulness in a study on the German federal election 2013.
Overall, we find that several online conversational practices differ significantly. We highlight these differences along with interesting commonalities among political parties by studying political communication on Twitter.
While our computational approach was applied to a single case study, the approach is not limited to a single case. Rather it is general enough that it can be applied to other contexts in which different groups of agents
communicate and to other social media systems similar to Twitter.
We hope that our work equips future computational social scientists with an improved instrument to assess and study online conversational practices in social media over time.
\balance
\footnotesize
\bibliographystyle{aaai}
|
\section{Introduction}
Let $G$ be an undirected graph that is simple and finite. A path in $G$ is \emph{rainbow} if no two edges of it are colored the same. The graph $G$ is \emph{rainbow colored} if there is a rainbow path between every pair of vertices. If there is a rainbow shortest path between every pair of vertices, the graph $G$ is \emph{strong rainbow colored}. The minimum number of colors needed to make $G$ rainbow colored is known as the \emph{rainbow connection number} and is denoted by $\rc(G)$. Likewise, the minimum number of colors needed to make $G$ strong rainbow colored is known as the \emph{strong rainbow connection number} and is denoted by $\src(G)$.
Rainbow connectivity was introduced by Chartrand, Johns, McKeon, and Zhang~\cite{Chartrand2008} in 2008. While being a theoretically interesting way of strengthening connectivity, rainbow connectivity also has applications in data transfer and networking~\cite{Li2012}. The study of rainbow colorings and several of its variants have recently attracted increasing attention in the research community. For a comprehensive treatment, we refer the reader to the books~\cite{Chartrand2008b, Li2012b}, or the recent survey~\cite{Li2012}.
Not surprisingly, determining the rainbow connection numbers is computationally hard. Chakraborty {\em et al.}~\cite{Chakraborty2009} proved that given a graph $G$, it is $\NP$-complete to decide if $\text{rc}(G)=2$, and that computing $\rc(G)$ is $\NP$-hard. Ananth {\em et al.}~\cite{Ananth2011} further showed that for every $k \geq 3$, deciding whether $\rc(G) \leq k$ is $\NP$-complete. The hardness of computing the strong rainbow connection number was shown by Ananth {\em et al.}~\cite{Ananth2011} as well. They proved that for every $k \geq 3$, deciding whether $\src(G) \leq k$ is $\NP$-hard, even when $G$ is bipartite.
Because rainbow coloring is hard in general, there has been interest in approximation algorithms and easier special cases. The \emph{shortest path distance} $d(s,t)$ from $s$ to $t$ is the minimum number of edges in any path from vertex $s$ to vertex $t$. If $s$ and $t$ are disconnected, $d(s,t) = \infty$. A path of length $d(s,t)$ from $s$ to $t$ is a \emph{shortest path} from $s$ to $t$. The \emph{eccentricity} of a vertex $v$ is the maximum shortest path distance between $v$ and any other vertex $u$. The \emph{radius} of a graph $G$ is the minimum eccentricity of the vertices. The \emph{diameter} of a graph $G$, denoted by $\diam(G)$, is the maximum eccentricity of the vertices. Basavaraju {\em et al.}~\cite{Basavaraju2012} presented approximation algorithms for computing the rainbow connection number with factors $(r+3)$ and $(d+3)$ respectively, where $r$ is the radius and $d$ the diameter of the input graph. Chandran and Rajendraprasad~\cite{Chandran2013} proved there is no polynomial time algorithm to rainbow color graphs with less than twice the optimum number of colors, unless $\P = \NP$. Ananth {\em et al.}~\cite{Ananth2011} showed there is no polynomial time algorithm for approximating the strong rainbow connection number of an $n$-vertex graph with a factor of $n^{1/2-\epsilon}$, where $\epsilon > 0$ unless $\NP = \ZPP$.
A graph is \emph{chordal} if it contains no induced cycle of length 4 or more. It is $\NP$-complete to decide whether $\rc(G) \leq k$ for every $k \geq 3$ even when $G$ is chordal~\cite{Chandran2012}. Furthermore, the rainbow connection number of a chordal graph can not be approximated to a factor less than $5/4$ unless $\P = \NP$~\cite{Chandran2013}. To the best of our knowledge, the complexity of deciding whether $\src(G) \leq k$ is an open problem even for chordal graphs. A \emph{split graph} is a chordal graph whose vertices can be partitioned into a clique and an independent set. Chandran {\em et al.}~\cite{Chandran2014} showed that for split graphs, the problem of deciding if $\rc(G) = k$ is $\NP$-complete for $k \in \{2,3\}$, and in $\P$ for all other values of $k$. Chandran and Rajendraprasad~\cite{Chandran2012} showed split graphs can be rainbow colored in linear time using at most one more color than the optimum. In the same paper, the authors also gave a linear time algorithm for finding a minimum rainbow coloring of a threshold graph. Furthermore, they noted that their result is apparently the first efficient algorithm for optimally rainbow coloring any non-trivial subclass of graphs. Similarly, we are not aware of any efficient exact algorithms for computing the strong rainbow connection number for any non-trivial subclass of graphs.
A \emph{cut vertex} is a vertex whose removal will disconnect the graph. A \emph{biconnected graph} is a connected graph having no cut vertices. A \emph{block graph} is an undirected graph where every maximal biconnected component, known as a \emph{block}, is a clique. In a block graph $G$, different blocks intersect in at most one vertex, which is a cut vertex of $G$. In other words, every edge of $G$ lies in a unique block, and $G$ is the union of its blocks. It is easy to see that a block graph is chordal.
In this paper, we investigate the rainbow and the strong rainbow connection number of block graphs. We extend the known hardness result for computing the strong rainbow connection number by showing it is $\NP$-complete to decide if a given split graph can be strong rainbow colored using $k$ colors, where $k \geq 3$. Furthermore, we obtain that there exists no polynomial time algorithm for approximating the strong rainbow connection number of an $n$-vertex split graph with a factor of $n^{1/2-\epsilon}$ for any $\epsilon > 0$ unless $\P = \NP$. These negative results further motivate the investigation of tractable special cases. Indeed, we determine the strong rainbow connection number of block graphs, and show any block graph can be strong rainbow colored optimally in linear time. We then turn to the rainbow connection number, and derive a tight upper bound of $|S|+2$ on $\rc(G)$ of a block graph $G$, where $S$ is the set of minimal separators. We also characterize the bridgeless block graphs with rainbow connection number 2, 3, or 4. Throughout the rest of the paper, we let $n$ denote the number of vertices and $m$ the number of edges of the graph in question.
\section{Hardness of strong rainbow coloring split graphs}
\label{sec:hardness_src}
In this section, we show that deciding if a split graph can be strong rainbow colored with $k \geq 3$ colors is $\NP$-complete. We remark that to the best of our knowledge, the complexity of the problem for $k \geq 3$ has been open even for chordal graphs.
In the \emph{$k$-subset strong rainbow connectivity problem} ($\subsetsrc$), we are given a graph $G$, a set of pairs $P \subseteq V(G) \times V(G)$, and an integer $k$. The goal is to decide if $G$ can be colored with $k$ colors such that each pair of vertices in $P$ is connected by a rainbow shortest path. The problem was shown to be $\NP$-complete by Ananth {\em et al.}~\cite{Ananth2011} even when the graph $G$ is a star.
\begin{lemma}[\hspace{1sp}{\cite{Ananth2011}}]
For every $k \geq 3$, the $\subsetsrc$ problem is $\NP$-complete when the graph $G$ is a star.
\end{lemma}
We reduce from this problem, and make use of some ideas of~\cite{Chakraborty2009} in the following.
\begin{theorem}
\label{thm_src_chordal_hardness}
For every integer $k \geq 3$, it is $\NP$-complete to decide if $\src(G) \leq k$, where $G$ is a split graph.
\end{theorem}
\begin{proof}
Let $\langle S=(V,E), P, k \rangle$ be an instance of the $\subsetsrc$ problem, where $S$ is a star, both $p$ and $q$ in each $(p,q) \in P$ are leaves of $S$, and $k \geq 3$ is an integer. We will construct a split graph $G'=(V',E')$ such that $G'$ is strong rainbow colorable with $k$ colors if and only if $\langle S, P \rangle$ is $k$-subset strong rainbow connected.
Let $a$ be the central vertex of $S$. For every vertex $v \in V \setminus \{a\}$, we add a new vertex $x_v$, and for every pair of leaves $(u,v) \in (V \times V) \setminus P$, we add a new vertex $x_{(u,v)}$. Formally, we construct $G'=(V',E')$ such that
\begin{itemize}
\item $V' = V \cup \{ x_v \mid v \in V \setminus \{a\}\} \cup \{ x_{(u,v)} \mid (u,v) \in (V \times V) \setminus P\}$,
\item $E' = E \cup E_1 \cup E_2 \cup E_3$,
\item $E_1 = \{ (v,x_v),(a,x_v) \mid v \in V \setminus \{a\} \}$,
\item $E_2 = \{ (u,x_{(u,v)}), (v, x_{(u,v)}), (a,x_{(u,v)}) \mid (u,v) \in (V \times V) \setminus P \}$, and
\item $E_3 = \{ (x,x') \mid x,x' \in V' \setminus V \}$.
\end{itemize}
Let us then verify $G'$ is a split graph. Observe the leaves of $S$ form an independent set in $G'$. The remaining vertices $\{a\} \cup (V' \setminus V)$ form a clique, proving $G'$ is split. An example illustrating the construction is given in Figure~\ref{fig:reduction_chordal}.
\begin{figure}[t]
\centering
\includegraphics[scale=1]{reductionfig}
\caption{A star graph $S$ on the vertex set $\{a,1,2,3\}$ transformed to a split graph $G'$ with $P = \{ (1,2),(2,3) \}$. The white vertices form an independent set while the black vertices form a clique. The symbol ?\ marks an edge-coloring $\chi$ of $S$ with $k$ colors under which the pairs in $P$ are connected by a rainbow path.}
\vspace{-0.5cm}
\label{fig:reduction_chordal}
\end{figure}
We will now prove $G'$ is strong rainbow colorable with $k$ colors if and only if $\langle S, P \rangle$ is $k$-subset strong rainbow connected. First, suppose $\langle S, P \rangle$ is not $k$-subset strong rainbow connected; we will show $G'$ is not strong rainbow colorable with $k$ colors. Observe that for each $(p,q) \in P$, there is a unique shortest path between $p$ and $q$ in $S$. Moreover, the same holds for $G'$. Therefore, any strong rainbow coloring using $k$ colors must make this path strong rainbow colored in $G'$. But because the pairs in $P$ cannot be strong rainbow connected with $k$ colors in $S$, the graph $G'$ cannot be strong rainbow colored with $k$ colors.
Finally, suppose $\langle S, P \rangle$ is $k$-subset strong rainbow connected under some edge-coloring $\chi : E \to \{c_1,\ldots,c_k\}$. We will describe an edge-coloring $\chi'$ given to $G'$ by extending $\chi$. We retain the original coloring on the edges of $S$, that is, $\chi'(e) = \chi(e)$, for every $e \in E$. The rest of the edges are colored as follows:
\begin{itemize}
\item $\chi'(e) = c_1$, for all $e \in E_1$,
\item $\chi'(e) = c_2$, for all $e \in E_3$, and
\item $\chi'((u,x_{(u,v)}))=c_1$, $\chi'((v,x_{(u,v)}) = c_2$, and $\chi'((a,x_{(u,v)}) = c_2$ for all $(u,v) \notin P$.
\end{itemize}
We can then verify $G'$ is indeed strong rainbow connected under $\chi'$, completing the proof.
\end{proof}
Ananth {\em et al.}~\cite{Ananth2011} reduced the chromatic number problem to the $\subsetsrc$ problem. The latter problem was then reduced to the problem of computing the strong rainbow connection number of an bipartite graph. The size of none of the reductions depend on the number of colors $k$. Thus, it can be shown that an $\alpha$-approximation algorithm for computing the strong rainbow connection number implies an $f(\alpha)$-approximation algorithm for the chromatic number problem, where $f$ is some function. But since the chromatic number of an $n$-vertex graph cannot be approximated with a factor of $n^{1-\epsilon}$ for any $\epsilon > 0$ unless $\P = \NP$~\cite{Zuckerman2006}, we obtain the following.
\begin{corollary}
There is no polynomial time algorithm that approximates the strong rainbow connection number of an $n$-vertex split graph with a factor of $n^{1/2-\epsilon}$ for any $\epsilon > 0$, unless $\P = \NP$.
\end{corollary}
\begin{proof}
It is enough to show the size of the constructed split graph $G'$ in Theorem~\ref{thm_src_chordal_hardness} does not depend on the number of colors $k$. Indeed, for every vertex of the star $S$, we add at most 1 vertex to $G'$. For each of the possible $V \choose 2$ pairs in $P$, we add exactly 1 vertex. Thus, the claim follows.
\end{proof}
\section{Clique trees of chordal graphs}
\label{sec:chordal_theory}
In this section, we apply tools from the theory of general chordal graphs to block graphs. We will show that when restricted to block graphs, a particular compact tree representation preserves enough structural information about the graph for our purposes. This fact enables us to design a linear time algorithm for optimally strong rainbow coloring block graphs later on.
A \emph{clique tree} of a connected chordal graph $G$ is any tree $T$ whose vertices are the maximal cliques of $G$ such that for every two maximal cliques $C_i,C_j$, each clique on the path from $C_i$ to $C_j$ in $T$ contains $C_i \cap C_j$. Chordal graphs are precisely the class of graphs that admit a clique tree representation~\cite{Gavril1974}. In general, a connected chordal graph does not have a unique clique tree.
\begin{lemma}[\hspace{1sp}{\cite{Galinier1995}}]
\label{lemma_clique_tree_computed_in_linear_time}
A clique tree of a chordal graph $G$ can be computed in $O(n+m)$ time.
\end{lemma}
A set $S \subseteq V(G)$ disconnects a vertex $a$ from vertex $b$ in a graph $G$ if every path of $G$ between $a$ and $b$ contains a vertex from $S$. A non-empty set $S \subseteq V(G)$ is a \emph{minimal separator} of $G$ if there exists $a$ and $b$ such that $S$ disconnects $a$ from $b$ in $G$, and no proper subset of $S$ disconnects $a$ from $b$ in $G$. If we want to identify the vertices that $S$ disconnects, we may also refer to $S$ as a \emph{minimal $a$-$b$ separator}.
The \emph{reduced clique graph}~\cite{Habib2012} of a chordal graph $G$ captures all possible clique tree representations of $G$. It is obtained by taking the maximal cliques of $G$ as vertices, and by putting edges between those vertices for which the corresponding cliques intersect in a minimal separator that separates them. The reduced clique graph of $G$ is denoted by $\mathcal{C}_r(G)$. In other words, the reduced clique graph $\mathcal{C}_r(G)$ is the union of all clique trees of $G$~\cite{Habib2012}. An example of a chordal graph, a corresponding clique tree, and the corresponding reduced clique graph are given in Figure~\ref{fig:ldeg_example}.
We may label each edge in $\mathcal{C}_r(G)$ by the minimal separator that separates its endpoints. Let $C$ be a vertex in $\mathcal{C}_r(G)$. For each edge in $\mathcal{C}_r(G)$ incident to $C$, consider its label. The \emph{labeled degree} of the vertex $C$, denoted by $\lambda_{\text{deg}}(C)$, is the number of edges incident to $C$ with distinct labels. Notice that the labeled degree of a vertex is different from the degree of a vertex. Consider the following example illustrated in Figure~\ref{fig:ldeg_example}. Let $C_1 = \{ b,c,e \}$, $C_2 = \{ e,g \}$ and $C_3 = \{ c,d,e \}$ be vertices in $\mathcal{C}_r(G)$. Notice that $S_{1,2} = \{ e \}$, and so the label on the edge $(C_1,C_2)$ is $\{ e \}$. Also, $S_{1,3} = \{ c,e \}$, thus the label on the edge $(C_1,C_3)$ is $\{c,e\}$. We have that $\text{deg}(C_1) = 4$, and $\lambda_{\text{deg}}(C_1) = 4$. However, $\text{deg}(C_2) = 3$, but $\lambda_{\text{deg}}(C_2) = 1$.
\begin{figure}[t]
\bsubfloat[]{%
\includegraphics[scale=1]{chordal1}%
}
\bsubfloat[]{%
\includegraphics[scale=1]{chordal2}%
}
\bsubfloat[]{%
\includegraphics[scale=1]{chordal3}%
}
\bsubfloati\qquad\bsubfloatii\qquad\bsubfloatiii
\caption{\textbf{(a)} A chordal graph $G$, \textbf{(b)} a clique tree of $G$, and \textbf{(c)} the reduced clique graph $\mathcal{C}_r(G)$.}
\vspace{-0.5cm}
\label{fig:ldeg_example}
\end{figure}
If a graph $G$ has exactly one shortest path between any pair of vertices, $G$ is said to be \emph{geodetic}. It was shown by Stemple and Watkins~\cite{Stemple1968} that a connected graph $G$ is geodetic if and only if every block of $G$ is geodetic. By observing that a clique is geodetic, we get the following result that is later exploited by our algorithms.
\begin{theorem}
\label{thm_block_graph_is_geodetic}
Every block graph is geodetic.
\end{theorem}
The reduced clique graph $\mathcal{C}_r(G)$ is a useful tool in reasoning about a chordal graph $G$. However, it is not a linear representation of $G$. For instance, the reduced clique graph of an $n$-vertex star graph uses $\Theta(n^2)$ space. Since a chordal graph on $n$ vertices admits at most $n$ maximal cliques, the size of a clique tree is always bounded from above by $n$. To save space and simplify our algorithms, we will show that we do not need to explicitly compute $\mathcal{C}_r(G)$, but instead that any clique tree of $G$ will do. More specifically, we will show that no matter what clique tree $T$ of $\mathcal{C}_r(G)$ we use, the labeled degree of a vertex in $\mathcal{C}_r(G)$ is preserved in $T$. We first present two results due to Galinier {\em et al.}~\cite{Galinier1995}.
\begin{lemma}[Triangle Lemma, {\cite{Galinier1995}}]
Let $[C_1, C_2, C_3]$ be a triangle in $\mathcal{C}_r(G)$ and let $S_{1,2},S_{1,3},S_{2,3}$ be the associated minimal separators of $G$. Then~2 of these~3 minimal separators are equal and included in the third.
\end{lemma}
\begin{lemma}[Weak Triangulation Lemma, {\cite{Galinier1995}}]
\label{lemma_weak_triangulation}
Let $[C_1,\ldots,C_k]$, $k \geq 4$, be a path in a clique tree $T$ of a chordal graph $G$. If $(C_1,C_k)$ is an edge of $\mathcal{C}_r(G)$, then either $(C_2,C_k)$ or $(C_1,C_{k-1})$ is an edge of $\mathcal{C}_r(G)$.
\end{lemma}
Recall that in a block graph $G$, the blocks intersect in at most one vertex, which is a cut vertex of $G$. This cut vertex is a minimal separator, so in a block graph, the size of every minimal separator is 1. We claim that for block graphs, the triangle lemma implies that $S_{1,2} = S_{1,3} = S_{2,3}$. If this was not the case, for example, if $S_{1,2} = s$, $S_{1,3} = t$ and $s \neq t$, then $S_{2,3}$ would have to be $s$ or $t$. Without loss we may assume $S_{2,3} = s$. Then the separator $s$ would have to be included in $S_{1,3}$. In other words, $S_{1,3} = \{s,t\}$, and now $S_{1,3}$ has size~2. Thus we have the following lemma.
\begin{lemma}
\label{lemma_eq_seps}
If $[C_1,C_2,C_3]$ is a triangle in the reduced clique graph of a block graph, then $S_{1,2} = S_{1,3} = S_{2,3}$.
\end{lemma}
We are then ready to proceed with the main result of the section.
\begin{theorem}
\label{thm_labeled_edges}
Let $T$ be a clique tree of a block graph $G$, let $\mathcal{C}_r(G)$ be the corresponding reduced clique graph, and let $C_1$ be the same vertex in each. If $e_1,e_2,\ldots,e_l$ are all labeled edges in $\mathcal{C}_r(G)$ incident to $C_1$ with the label $s$, then at least one of these edges must be in $T$.
\end{theorem}
\begin{proof}
Suppose not. Let $C_1$ and $C_k$ be adjacent vertices in $\mathcal{C}_r(G)$ with minimal separator $s$, so that $(C_1,C_k)$ is not an edge in $T$. Let $C_2$ be another vertex in $\mathcal{C}_r(G)$ adjacent to $C_1$, and suppose $(C_1,C_2)$ is an edge of $T$. Then this edge is not labeled $s$, so let $t$ denote the label of this edge. However, because $T$ is a spanning tree, $C_2$ and $C_k$ are connected. Let $[C_2,C_3,\ldots,C_k]$ be the path in $T$ from $C_2$ to $C_k$. If this path is simply an edge, then we have a triangle $[C_1,C_2,C_k]$ of $\mathcal{C}_r(G)$. Thus, $S_{1,2} = S_{2,k} = S_{1,k}$. However, $S_{1,2} = t$ and $S_{1,k} = s$, so this is a contradiction to Lemma~\ref{lemma_eq_seps}. Thus there are at least~3 vertices on this path. Then the path $[C_1,C_2,\ldots,C_k]$ has at least 4 vertices. Consider this path. Because $(C_1,C_k)$ is an edge of $\mathcal{C}_r(G)$, then by Lemma~\ref{lemma_weak_triangulation}, either $(C_2,C_k)$ or $(C_1,C_{k-1})$ is an edge of $\mathcal{C}_r(G)$. We have already shown $(C_2,C_k)$ is not an edge, so $(C_1,C_{k-1})$ is an edge of $\mathcal{C}_r(G)$. Because $[C_1,C_k,C_{k-1}]$ is a triangle and $S_{1,k} = s$, we have $S_{1,k-1} = s$.
Now consider the path $[C_1,C_2,\ldots,C_{k-1}]$. We assume this path has at least 4 vertices, for if it was a triangle, we would have a contradiction to $S_{1,k-1} = s$ and $S_{1,2} = t$. Because $(C_1,C_{k-1})$ is an edge of $\mathcal{C}_r(G)$, then by Lemma~\ref{lemma_weak_triangulation}, either $(C_2,C_{k-1})$ or $(C_1,C_{k-2})$ is an edge of $\mathcal{C}_r(G)$. If $(C_2,C_{k-1})$ was an edge, then $[C_1,C_2,C_{k-1}]$ would be a triangle with $S_{1,2}=t$ and $S_{1,k-1}=s$. This contradicts Lemma~\ref{lemma_eq_seps}, so $(C_2,C_{k-1})$ is not an edge. Thus $(C_1,C_{k-2})$ is an edge of $\mathcal{C}_r(G)$, and we have the triangle $[C_1,C_{k-1},C_{k-2}]$. Because $S_{1,k-1} = s$, it follows that $S_{1,k-2} = s$.
We can continue this process, showing that all of the edges $$(C_1,C_{k-1}), (C_1,C_{k-2}), (C_1,C_{k-3}), \ldots, (C_1,C_3)$$ are in $\mathcal{C}_r(G)$ and have label $s$. But now we have the triangle $[C_1,C_2,C_3]$, and since $S_{1,2} = t$ and $S_{1,3} = s$, we have a contradiction to Lemma~\ref{lemma_eq_seps}. Thus, $T$ has at least one edge incident to $C_1$ with the label $s$.
\end{proof}
\begin{corollary}
\label{corollary_ldeg_preserved}
Let $G$ be a block graph. Then any pair of clique trees $T_1$ and $T_2$ of $G$ has the property that every vertex in $\mathcal{C}_r(G)$ has the same labeled degree in $T_1$ as it does in $T_2$.
\end{corollary}
\begin{proof}
Let $C$ be a vertex in $\mathcal{C}_r(G)$. Denote the set of edges incident to $C$ with the label $x$ as $I_x$. Then if $C$ has $l$ distinct minimal separators, $C$ has the following incident edges: $I_1,I_2,\ldots,I_l$. By Theorem \ref{thm_labeled_edges}, any clique tree of $G$ contains at least one edge from each of $I_1,I_2,\ldots,I_l$. Thus $C$ has labeled degree $l$ in any clique tree of $G$.
\end{proof}
\section{Strong rainbow coloring block graphs in linear time}
\label{sec:algorithm}
In this section, we determine exactly the strong rainbow connection number of block graphs. We present an exact linear time algorithm for constructing a strong rainbow coloring using $\src(G)$ colors for a given block graph $G$.
Let $C$ be a block in a block graph $G$ whose edges are colored by using colors from the set $R = \{c_1,\ldots,c_r\}$. Then we say that $C$ is \emph{colored} and $C$ is \emph{associated} with each color $c_1,\ldots,c_r$. Furthermore, any color from $R$ can be used as a representative for the color of $C$. Thus we may say that $C$ has been colored $c_i$ for any $i \in \{ 1,\ldots,r \}$.
\begin{lemma}
\label{lemma_coloring}
Let $G$ be a block graph, let $T$ be a clique tree of $G$, let $C$ be a vertex of $T$ that is associated with the color $c$, let $(u,v)$ be an edge in $G$ such that $u,v \notin C$, and let $y$ be the minimal $a$-$b$ separator for any $a \in C \setminus \{y\}$ and $b \in \{u,v\}$. If no shortest $y$-$u$ path or shortest $y$-$v$ path contains $(u,v)$, then by coloring $(u,v)$ with the color $c$, any shortest path between $u$ or $v$ and $w \in C$ contains at most one edge of color $c$.
\end{lemma}
\begin{proof}
Any shortest path between $u$ or $v$ and $y$ does not contain the edge $(u,v)$, and does not contain any edges in $C$, so these paths do not have any edges of color $c$. Any shortest path between $y$ and $w$ is just an edge of color $c$.
\end{proof}
The algorithm for strong rainbow coloring a block graph is presented in Algorithm~\ref{alg:src_block}. Given a block graph $G$, the algorithm first computes a clique tree $T$ of $G$. Next, it partitions the vertices of $T$ into two sets $V_{<3}$ and $V_{\geq 3}$ based on their labeled degree. If the labeled degree of a vertex is less than~3, it is added to $V_{<3}$. Otherwise, it is added to $V_{\geq 3}$. Then, for each vertex in $V_{<3}$, a distinct color is used to color the edges of the block the vertex corresponds to in $G$. At the final step the algorithm goes through every vertex $C_j \in V_{\geq 3}$. Let $N_\lambda(C_j)$ denote the set of vertices adjacent to $C_j$ via distinct labels. Fix~3 distinct vertices $C_1$, $C_2$, and $C_3$ in $N_\lambda(C_j)$. Observe that in $T \setminus C_j$, we would have at least 3 connected components, and $C_1$, $C_2$, and $C_3$ would be in different connected components. Suppose $C_j$ was removed, and from each connected component $C_1$, $C_2$, and $C_3$ is in, find a vertex in $V_{<3}$. The picked three vertices are each associated with a distinct color. These colors are used to color the edges of the block $C_j$ corresponds to.
\begin{algorithm}[t]
\caption{Algorithm for strong rainbow coloring a block graph}
\label{alg:src_block}
\begin{algorithmic}[1]
\Require A block graph $G$
\Ensure A strong rainbow coloring of $G$
\State $T := \text{a clique tree of $G$}$
\State $V_{<3} := \{ U \mid U \in V(T) \wedge \lambda_{\text{deg}}(U) < 3 \}$
\State $V_{\geq 3} := V(T) \setminus V_{<3}$
\ForAll{$U \in V_{<3}$}
\State Color edges in $U$ with a fresh distinct color
\EndFor
\ForAll{$C_j \in V_{\geq 3}$}
\State Let $C_1$, $C_2$, $C_3$ be distinct vertices in $N_\lambda(C_j)$
\State Let $S_{j,1} = x_1$, $S_{j,2} = x_2$, $S_{j,3} = x_3$ be the corresponding minimal separators
\State Assume $C_j$ is removed
\State From each connected component $C_1$, $C_2$, $C_3$ is in, find a vertex in $V_{<3}$
\State Let $c_1,c_2,c_3$ be the respective colors associated with the found vertices
\State Color all edges not incident to $x_1$ with color $c_1$
\State Color all edges incident to $x_1$, except $(x_1,x_2)$, with color $c_2$
\State Color the edge $(x_1,x_2)$ with color $c_3$
\EndFor
\end{algorithmic}
\end{algorithm}
The correctness of Algorithm~\ref{alg:src_block} is established by an invariant, which says that we always maintain the property that if the shortest path between two vertices is colored, then it is rainbow colored. We refer to this property as the \emph{shortest rainbow path} property.
\begin{theorem}
\label{thm_alg_correctness}
At every step, Algorithm~\ref{alg:src_block} maintains the shortest rainbow path property.
\end{theorem}
\begin{proof}
Before the execution of the first loop, nothing is colored so the claim is trivially true. Furthermore, the first loop obviously maintains the property. To see this, consider any shortest path of length $l \geq 1$ at any step. The path consists of $l$ edges that are in $l$ distinct blocks. Since each colored block has received a distinct color, the shortest path is rainbow colored. This establishes the base step for the correctness of the second loop.
Assume after iteration $i-1$ of the second loop, if the shortest path between any two vertices is colored, then it is rainbow colored. We show that this property is maintained after iteration $i$ of the second loop. Consider any edge $(u,v)$ in $C_j$ not incident to $x_1$, and let $y \in C_1$ be the minimal $a$-$b$ separator for any $a \in C_1 \setminus \{y\}$ and $b \in \{u,v\}$. The algorithm states that $(u,v)$ will be colored with color $c_1$. Because $u$ and $v$ are both at a distance 1 from $x_1$, it follows that neither shortest path $y$-$u$ or $y$-$v$ contains $(u,v)$. Thus by Lemma~\ref{lemma_coloring}, if the shortest $w$-$u$ path, for $w \in C_1$ is colored, then it is rainbow colored. (The same is true for the shortest $w$-$v$ path). Therefore, by coloring $(u,v)$ with color $c_1$, the shortest rainbow path property is maintained.
Consider any edge $(u,v)$ in $C_j$ not incident to $x_2$, and let $y \in C_2$ be the minimal $a$-$b$ separator for any $a \in C_2 \setminus \{y\}$ and $b \in (u,v)$. By Lemma~\ref{lemma_coloring}, this edge can be colored with $c_2$ to maintain the shortest rainbow path property. Notice that $u$ and $v$ are both at a distance 1 from $x_2$, so it follows that $x_1$ must be one of these vertices (i.e.\ either $u=x_1$ or $v=x_1$). So we conclude that every edge incident to $x_1$, except $(x_1,x_2)$, can be colored with $c_2$ to maintain the shortest rainbow path property.
Now the only uncolored edge in $C_j$ is the edge $(x_1,x_2)$. Because $x_1$ and $x_2$ are both at a distance 1 from $x_3$, Lemma~\ref{lemma_coloring} assures us that by coloring $(x_1,x_2)$ with color $c_3$, the shortest rainbow path property is maintained.
\end{proof}
We will then consider the complexity of Algorithm~\ref{alg:src_block}. It is an easy observation that lines 1 to 6 take linear time. Observe that on lines 8 to 11, we essentially perform reachability queries of the form \emph{given a vertex $v \in V(T)$, return any vertex of degree less than 3 of $T$ that is reachable from $v$ with a path including a given edge $(u,v)$, and no other edges incident to $v$}. In our context, $v$ is $C_j$, and $u$ is $C_i$, where $C_i \in \{C_1,C_2,C_3\}$. The naive way of answering such queries is to start a depth-first search (DFS) from each $C_i$, and halt when a suitable vertex is found. However, such implementation requires $O(d)$ time, where $d$ is the diameter of the input graph $G$. Using elementary techniques, we can preprocess the clique tree $T$ after line 1 using linear time to answer such queries in $O(1)$ time. Thus, the total runtime will be linear as the for-loop on line 7 loops $O(n)$ times.
\begin{theorem}
Algorithm~\ref{alg:src_block} constructs a strong rainbow coloring in $O(n+m)$ time.
\end{theorem}
We will now show that the strong rainbow coloring produced by Algorithm~\ref{alg:src_block} is optimal. This is done by first showing that we need at least $k$ colors, where $k$ is the number of vertices with labeled degree less than~3 in any clique tree $T$ of $G$. This is then shown to be sufficient as well by a matching upper bound. Recall from Corollary~\ref{corollary_ldeg_preserved} that the labeled degree of a vertex of $\mathcal{C}_r(G)$ is preserved in any clique tree $T$ of $G$.
\begin{theorem}
\label{thm_src_lower_bound}
Let $G$ be a block graph, and let $k$ be the number of vertices with labeled degree less than~3 in any clique tree $T$ of $G$. Then $\text{src}(G) \geq k$.
\end{theorem}
\begin{proof}
Let $E$ be a set of $k$ edges in $G$, one from each block with labeled degree less than~3, selected as follows. For each vertex $C \in T$, if $\lambda_{\text{deg}}(C)=1$, pick an edge incident to the minimal separator. If $\lambda_{\text{deg}}(C)=2$, pick the edge connecting the 2 minimal separators. We claim that if we are to strong rainbow color $G$, then the edges in $E$ must all receive distinct colors.
Suppose there are 2 edges in $E$ that are of the same color, say $(u,x) \in C_i$ and $(v,y) \in C_j$. Without loss, we may assume that $u$ and $v$ are the minimal separators of $C_i$ and $C_j$, respectively, such that $d(u,v)$ is minimized. Then the shortest $x$-$y$ path is unique by Theorem~\ref{thm_block_graph_is_geodetic}, and it contains two edges of the same color.
\end{proof}
\begin{theorem}
\label{thm_src_equals_k}
Let $G$ be a block graph, and let $k$ be the number of vertices with labeled degree less than~3 in any clique tree $T$ of $G$. Then $\text{src}(G) = k$.
\end{theorem}
\begin{proof}
It is shown in the proof of Theorem~\ref{thm_alg_correctness} that any vertex $C \in V(T)$ with labeled degree at least~3 can be colored using the colors associated with vertices of labeled degree less than~3. Thus we need at most $k$ colors to color $G$. This establishes a matching upper bound for Theorem~\ref{thm_src_lower_bound}, so it follows that $\text{src}(G) = k$.
\end{proof}
To summarize, Theorems~\ref{thm_alg_correctness} and~\ref{thm_src_equals_k} show that Algorithm~\ref{alg:src_block} is correct, and always finds an optimal solution.
If an explicit coloring is not required, then it is easy to see that there is a linear time algorithm for computing $\src(G)$, where $G$ is a block graph. This is obtained by computing a clique tree $T$ of $G$, and counting the number of vertices with labeled degree less than~3 in $T$.
\begin{corollary}
There is an algorithm such that given a block graph $G$, it computes $\src(G)$ in $O(n+m)$ time.
\end{corollary}
\section{The rainbow connection number of block graphs}
\label{sec:block_rc}
In this section, we consider the rainbow connection number of block graphs. Using known results, we begin by deriving a tight linear time computable upper bound on the rainbow connection number. As a main result of the section, we prove a polynomial-time characterization of bridgeless block graphs with rainbow connection number at most 4.
A \emph{peripheral vertex} is a vertex of maximum eccentricity. A \emph{peripheral block} is a block that contains a peripheral vertex. The following lower bound helps us demonstrate an upper bound we derive later is tight.
\begin{theorem}
\label{thm_sep_touching_two_pvertices}
Let $G$ be a block graph with at least 3 blocks, and let $x$ and $y$ be two peripheral vertices in distinct blocks. If $G$ has a minimal separator $s$ adjacent to $x$ and $y$, then $\rc(G) > \diam(G)$.
\end{theorem}
\begin{proof}
Suppose not. That is, assume $\rc(G) = \diam(G)$. Let $C_x$ and $C_y$ be the two distinct blocks $x$ and $y$ are in, respectively. Choose a vertex $z \in C_z$ such that $d(x,z) = \diam(G)$, where $C_z$ is a block different from $C_x$ and $C_y$. Rainbow color the shortest $x$-$z$ path. Without loss, suppose the edge $(x,s)$ was colored with the color $c_1$. Then consider each uncolored edge incident to $s$ in $C_x$. Notice we must color each such edge with the color $c_1$, otherwise $G$ would not be rainbow connected. Finally, consider the edges incident to $s$ in $C_y$. Again, each such edge must receive the color $c_1$. But now $x$ and $y$ are not rainbow connected, thus $\rc(G) > \diam(G)$.
\end{proof}
Figure~\ref{fig:tight_approx_example} (a) illustrates the previous theorem: the block graph $G$ has two peripheral vertices adjacent to a minimal separator $s$. Both the edges $(x,s)$ and $(y,s)$ would have to receive the same color in a rainbow coloring of $G$ using $\diam(G)$ colors, but then there is no way to rainbow connect $x$ and $y$ without introducing new colors.
We next give an upper bound on the rainbow connection number of block graphs using a technique of~\cite{Chandran2011}. Given a graph $G=(V,E)$, a subset $S$ of $V$ is called a \emph{dominating set} if every vertex in $V \setminus S$ is adjacent to some vertex in $S$. The \emph{domination number} $\gamma(G)$ is the size of the smallest dominating set for the graph $G$. A dominating set $S$ is called a \emph{connected dominating set} if the graph induced by $S$ is connected. The \emph{connected domination number} $\gamma_c(G)$ is the size of the smallest connected dominating set of the graph $G$. We have the following.
\begin{theorem}[\hspace{1sp}{\cite{Chandran2011}}]
For every connected graph $G$, with $\delta(G) \geq~2$,
\begin{equation*}
\rc(G) \leq \gamma_c(G) + 2.
\end{equation*}
\end{theorem}
Further, the following has been determined.
\begin{theorem}[\hspace{1sp}{\cite{Chen2004}}]
Let $G$ be a connected block graph, $S$ the set of minimal separators of~$G$, and~$l$ the number of blocks in $G$. Then
\[
\gamma_c(G) =
\begin{cases}
1 & \text{for}\ l = 1, \\
|S| & \text{for}\ l \geq 2.
\end{cases}
\]
\end{theorem}
Combining the two previous theorems, we get the following.
\begin{theorem}
\label{thm_rc_upper_bound}
Let $G$ be a connected block graph with at least two blocks and $\delta(G) \geq~2$. Then $\rc(G) \leq |S|+2$, where $S$ is the set of minimal separators of $G$.
\end{theorem}
This bound is also tight as demonstrated by graph $G$ in Figure~\ref{fig:tight_approx_example} (a). Using the linear time algorithm of~\cite{Chandran2011} for enumerating the minimal separators of a chordal graph, we get a linear time algorithm for computing this upper bound for a given block graph.
\begin{figure}[t]
\bsubfloat[]{%
\includegraphics[scale=1]{block1}%
}
\bsubfloat[]{%
\includegraphics[scale=1]{block2}%
}
\bsubfloati\qquad\bsubfloatii
\caption{\textbf{(a)} A block graph $G$ with a minimal separator $s$ adjacent to two peripheral vertices $x$ and $y$ in distinct peripheral blocks. \textbf{(b)} A $K_n$ with $n$ triangles glued to it for $n = 5$.}
\vspace{-0.5cm}
\label{fig:tight_approx_example}
\end{figure}
We will then characterize the bridgeless block graphs having a rainbow connection number 2, 3, or 4. The following also determines exactly the rainbow connection number of the \emph{windmill graph} $K^{(m)}_n$ ($n > 3$), which consists of $m$ copies of $K_n$ with one vertex in common.
\begin{theorem}
\label{thm_rc_cases}
Let $G$ be a bridgeless block graph, and let $k$ a positive integer such that $k \leq 4$. Deciding whether $\rc(G) = k$ is in $\P$.
\end{theorem}
\begin{proof}
It is enough to consider bridgeless block graphs with diameter at most 4. For every value of $d = \diam(G) \leq 4$, we will give an efficient algorithm for optimally rainbow coloring the given block graph $G$.
\begin{itemize}
\item Case $d = 1$. Trivial.
\item Case $d = 2$. If $G$ has exactly 2 blocks, it is easy to see that $\rc(G) = 2$. Moreover, if the graph has $\rc(G) = 2$, it must have exactly 2 blocks. Suppose this is was not the case, i.e.\ $G$ has at least 3 blocks and $\rc(G) = 2$. By an argument similar to Theorem~\ref{thm_sep_touching_two_pvertices}, this leads to a contradiction. Thus, $\rc(G) = 2$ if and only if $G$ has exactly 2 blocks. When $G$ consists of 3 or more blocks, we will show that $\rc(G) = 3$. Let $\mathcal{K}$ be the set of all blocks of $G$, and let $a$ be the unique central vertex of $G$. For each $K \in \mathcal{K}$, color one edge incident to $a$ with the color $c_1$, and every other incident edge with the color $c_2$. Then color every uncolored edge of $G$ with the color $c_3$. To see this is a rainbow coloring of $G$, observe there is a rainbow path from any vertex to the central vertex $a$ avoiding a particular color in $\{c_1,c_2,c_3\}$.
\item Case $d = 3$. The graph $G$ consists of a unique central clique, and at least 2 other blocks. If $G$ has altogether 3 blocks, then $\rc(G) = \src(G) = 3$. If $G$ has 4 blocks, there are two cases: either $G$ has a cut vertex adjacent to two peripheral vertices in distinct blocks (then $\rc(G) \geq 4$ by Theorem~\ref{thm_sep_touching_two_pvertices}) or it does not (then $\rc(G) = \src(G) = 3$). Otherwise, $G$ has at least 5 blocks, and by an argument similar to Theorem~\ref{thm_sep_touching_two_pvertices}, we have that $\rc(G) \geq 4$. We will then color every block that is not the central clique with 3 colors exactly as in the case $d = 2$, and color every edge of the central clique with a fresh distinct color $c_4$ proving $\rc(G) = 4$.
\item Case $d = 4$. Let us call the set of blocks which contain the central vertex $a$ the \emph{core} of the graph $G$. The set of blocks not in the core is the \emph{outer layer}. First, suppose the core contains exactly 2 blocks, and the outer layer at most 4 blocks. Furthermore, suppose the condition of Theorem~\ref{thm_sep_touching_two_pvertices} does not hold (otherwise we would have $\rc(G) > 4$ immediately). When the outer layer contains 2 or 3 blocks, we have that $\rc(G) = \src(G) = 4$. Suppose the outer layer contains exactly 4 blocks. First, consider the case where a core block is adjacent to 3 blocks in the outer layer. Because the condition of Theorem~\ref{thm_sep_touching_two_pvertices} does not hold, it must be the case that at least one of the core blocks is not a $K_3$. Clearly, every two vertices $x$ and $y$, such that $d(x,y) = \diam(G)$, have to be connected by a rainbow shortest path. By an argument similar to Theorem~\ref{thm_src_lower_bound}, we have that $\rc(G) > 4$. Otherwise, when a core block is not adjacent to 3 blocks in the outer layer, $\rc(G) = \src(G) = 4$. Now suppose the outer layer has at least 5 blocks. As above, by an argument similar to Theorem~\ref{thm_src_lower_bound}, we have that $\rc(G) > 4$. Finally, suppose the core has 3 or more blocks. We argue that in this case, $\rc(G) = 4$ if and only if the outer layer contains exactly 2 blocks. For the sake of contradiction, suppose $\rc(G) = 4$, and that the outer layer has 3 or more blocks. If the condition of Theorem~\ref{thm_sep_touching_two_pvertices} holds, we have an immediate contradiction. Otherwise, by an argument similar to Theorem~\ref{thm_sep_touching_two_pvertices}, we arrive at a contradiction. When the outer layer contains exactly 2 blocks, we will show $\rc(G) = 4$. Let $B_1$ and $B_2$ be the blocks in the outer layer. We color every edge of $B_1$ with the color $c_1$, and every edge of $B_2$ with the color $c_4$. Then color $(b_1,a)$ with $c_2$, and $(a,b_2)$ with $c_3$, where $a$ is the central vertex of $G$, and $b_1$ and $b_2$ are the cut vertices in $B_1$ and $B_2$, respectively. For every block $B_i$ in the core, let $Q_i$ denote the set of edges in $B_i$ incident to $a$. Color the uncolored edges of $Q_i$ with either $c_2$ or $c_3$, such that both colors appear at least once in $Q_i$. Then, color every uncolored edge of the block that contains both $a$ and $b_2$ with the color $c_1$. Every other uncolored edge of $G$ receives the color $c_4$. We can now verify $G$ is indeed rainbow connected under the given coloring.
\end{itemize}
\end{proof}
Given that the strong rainbow connection number of a block graph $G$ can be efficiently computed, it is interesting to ask when $\rc(G) = \src(G)$, or if the difference between $\src(G)$ and $\rc(G)$ would always be small. Because $\diam(G) \leq \rc(G)$ for any connected graph $G$, the following is easy to see.
\begin{corollary}
Let $G$ be a block graph, and let $k$ be the number of vertices with labeled degree less than 3 in any clique tree $T$ of $G$. If $k = \diam(G)$, then $\rc(G) = \src(G)$.
\end{corollary}
However, the difference between $\src(G)$ and $\rc(G)$ can be made arbitrarily large: attach $n$ triangles to a $K_n$, one to each vertex of the $K_n$ (see Figure~\ref{fig:tight_approx_example} (b) for an illustration). As $n$ increases, the rainbow connection number remains 4 by Theorem~\ref{thm_rc_cases}, while the strong rainbow connection number increases by Theorem~\ref{thm_src_equals_k}. This example also shows the difference between the upper bound of Theorem~\ref{thm_rc_upper_bound} and $\rc(G)$ can be arbitrarily large.
\section{Concluding remarks}
We investigated the complexity of computing the rainbow and strong rainbow connection numbers of subclasses of chordal graphs, namely split graphs and block graphs. In particular, Theorem~\ref{thm_src_chordal_hardness} shows the strong rainbow connection number is significantly harder to approximate than the rainbow connection number, even on very restricted graph classes. Indeed, the result should be contrasted with the fact that any split graph can be rainbow colored in linear time using at most one color more than the optimum~\citep{Chandran2012}.
We believe our results for rainbow and strong rainbow coloring block graphs can serve as a starting point for an even more systematic study of strong rainbow coloring more general graph classes -- a topic which has received quite little attention despite the interest. Indeed, the investigation of the strong rainbow connection number has been deemed ``much harder than that of rainbow connection number''~\cite{Li2012}. To avoid confusion, we note that similar problems have been considered in e.g.\ \citep{Uchizawa2013,Lauri15}: given an edge-colored graph $G$, decide if $G$ is (strong) rainbow connected. We stress that known hardness results for these problems do not imply hardness results for finding optimal rainbow colorings. Indeed, the problems are strictly different.
For example, complete graphs are exactly the graphs that can be (strong) rainbow colored with a single color. Informally, \emph{clique-width} is a measure of how close a graph is to being a clique. Complete graphs, complete bipartite graphs, and block graphs have bounded clique-width~\cite{Golumbic2004}, and for each the strong rainbow connection number can be computed efficiently. How does the complexity of computing the strong rainbow connection number behave on bounded clique-width graphs in general? This is particularly interesting, as there is no easy way of describing the property of being strong rainbow colored using $k$ colors in MSO$_1$ (for an introduction, see e.g.\ \cite{Downey2013}). Theorem~\ref{thm_src_chordal_hardness} does not settle the question, as the constructed graph $G'$ appears to have unbounded clique-width.
Theorem~\ref{thm_src_chordal_hardness} shows that there are graph classes for which computing the rainbow connection number is easy, while computing the strong rainbow connection number is hard. Block graphs seem like a good candidate for the opposite. We leave open the question of the complexity of rainbow coloring block graphs. It is arguably often the case that ``edge problems'' are hard for bounded clique-width. Indeed, we conjecture it is $\NP$-hard to rainbow color block graphs using the minimum number of colors.
\bibliographystyle{abbrv}
|
\section{Main result and discussion}
We continue the analysis of closed algebraic embeddings of $\C^*=\C^1\setminus \{0\}$ into the complex affine plane $\C^2$ initiated in \cite{CKR-Cstar_good_asymptote}, where embeddings admitting a good asymptote have been classified.
\begin{definition}\label{asymptote} Let $U\subset \C^2$ be a closed curve isomorphic to $\C^*$. A curve $A\subset \C^2$ isomorphic to the affine line $\C^1$ is called a \textit{good asymptote} of $U$ if and only if $A\cdot U\leq 1$. If $U\subset \C^2$ does not admit a good asymptote we call the embedding \emph{sporadic}.\end{definition}
Note that the intersection is taken in $\C^2$, so the definition is independent of the choice of coordinates, i.e.\ of a choice of generators of the algebra of regular functions on $\C^2$. Surprisingly, although the defining condition of sporadic embeddings seems to be weak, up to now we know only very few of them: one discreet family with no deformations and one more embedding, see \cite[Main Theorem (s),(t)]{BoZo-annuli} (note that the list in loc.\ cit.\ is produced assuming strong 'regularity condition'). The goal of this article is to establish geometric and algebraic machinery which allows to prove strong restrictions on sporadic $\C^*$-embedding in terms of the resolution of singularities of their closures on $\PP^2$. With these tools in hand we are going to obtain the full classification in a forthcoming paper.
We introduce the following numbers characterizing the embedding. Let $(\lambda,P)$ be an analytically irreducible germ of a planar curve and let $L$ be a curve smooth at $P$ which does not cross $\lambda$ normally (i.e.\ which is tangent to $\lambda$ at $P$ in case $\lambda$ is smooth). The \emph{jumping number} $j(\lambda, L)$ of $\lambda$ \emph{with respect to L} is the maximal number of blowups on the proper transform of $\lambda$ after which $\lambda$ meets the total transform of $L$ not in a node. In particular, $j(\lambda, L)=0$ if and only if $L$ is tangent to $\lambda$. If $\lambda$ is singular and $(\eta,P)$ is a smooth germ maximally tangent to $\lambda$, i.e., such that $\lambda\cdot \eta$ is maximal for intersections of $\lambda$ with smooth germs, then $j(\lambda, L)$ is the integral part of $\lambda\cdot \eta$/$\lambda\cdot L$. We write $\C^2=\PP^2\setminus L_\8$, where $L_\8$ is a line (degree $1$ curve) in $\PP^2$, called the \emph{line at infinity}.
\begin{definition}\label{def:type} Let $U\subset \C^2=\PP^2\setminus L_\8$ be a closed curve and let $\lambda_1, \lambda_2,...,\lambda_k$ be the germs at
infinity of the closure $\bar U$ of $U$ in $\PP^2$ ordered so that the jumping numbers $j_i=j(\lambda_i,L_\8)$ do not decrease with $i$. We then call
$(j_1,j_2,...,j_k)$ the \textit{type at infinity of} $U$.
\end{definition}
Note that the identification $\C^2=\PP^2\setminus L_\8$ and the type of a curve at infinity depend on a choice of coordinates. Another number associated to an embedding $U\subset \C^2=\PP^2\setminus L_\8$ is the self-intersection of the proper transform of the closure $\bar U$ of $U$ under the minimal log resolution of singularities of $(\PP^2,\bar U +L_\8)$, i.e., after the minimal number of blowups so that the total transform of $\bar U +L_\8$ is a simple normal crossing divisor. For $U\cong \C^*$ our main result proves the existence of special coordinates with respect to which these numbers are sharply limited.
\begin{theorem}\label{thm:main_result} For every sporadic $\C^*$-embedding $U\subset \C^2$ one can choose coordinates on $\C^2$ so that:
\begin{enumerate}
\item the branches at infinity of $U$ are disjoint and the type at infinity of $U$ is $(1,\ti j)$ for some $\ti j\in\{2,3,4,5,6\}$ and
\item if $\bar U$ denotes the closure of $U$ on $\PP^2=\C^2\cup L_\8$ then the proper transform of $\bar U$ under the minimal log resolution of $(\PP^2,\bar U+L_\8)$ has self-intersection between $-2$ and $-5$.
\end{enumerate}
\end{theorem}
\bsrem In elementary planar geometry, a choice of coordinates on $\C^2$ having been made, an \textit{asymptote} of a closed curve $U$ is a straight line tangent to a branch of $U$ at infinity. If $A$ is a good asymptote in the sense of \ref{asymptote} then, since $A$ is an affine line, we can by the Abhyankar-Moh-Suzuki theorem choose coordinates so that $A$ is a straight line. Then, assuming $U$ is irreducible and of degree at least three, $A$ is tangent to $U$ at infinity. It follows that a good asymptote of $U\cong \C^*$ is an asymptote in suitable coordinates. Let $U\subseteq \C^2$ be a sporadic $\C^*$-embedding. Interestingly, although $U$ does not have a good asymptote, by \ref{thm:main_result}(1) both jumping numbers are positive, so the branches at infinity of $U$ are not tangent to the line at infinity. Therefore, $U$ has an asymptote at each of its two points at infinity, each of them meets $U$ at lest twice on $\C^2$, see \ref{asymptotes}.
\esrem
We now discuss our approach. The first step to understand the geometry of sporadic $\C^*$-embeddings was made in \cite{KR-Some_properties_of_Cstar}, where it has been proved that one can choose coordinates on $\C^2$ so that the closure of $U$ in $\PP^2$ meets the line at infinity in exactly two points, i.e., so that the two branches at infinity of $\C^*$ are separated.\footnote{We recall that in contrast there are $\C^*$-embeddings admitting a good asymptote for which the branches at infinity meet for every choice of coordinates, \cite[6.8.1]{CKR-Cstar_good_asymptote}.} Remembering that the known sporadic embeddings are given by very special equations (or parametrizations) which we need to somehow recover, from the very beginning we need to have a precise control over the singularities and their behaviour under subsequent steps of the log resolution $\Phi\: (\ov S',D'+E')\to (\PP^2,L_\8+\bar U)$, where $E'$ and $D'$ are respectively the proper transform of $\bar U$ and the reduced total transform of $L_\8$. This is achieved using Hamburger Noether pairs of the resolution (see sec. \ref{ssec:HN}) and two fundamental equations \eqref{eq:1} and \eqref{eq:2} relating them with properties of $E'$. Since the proper transform of $L_\8$ on $\ov S'$ may be a non-branching $(-1)$-curve, we need to consider an snc-minimalization $\Psi\:(\ov S',D'+E')\to (\ov S,D)$. The basic characteristic numbers are $\varepsilon=2-(K_{\ov S}+D)^2$ and $\gamma=-E^2$, where $E$ is the proper transform of $U$ on $\ov S$. They are bounded by $\gamma\geq 1$ and $\varepsilon\geq 0$ (\ref{lem:eps_and_gamma}). The nonexistence of a good asymptote turns out to have strong consequences: we obtain $2\varepsilon+\gamma\leq 9$ (\ref{prop:basic_inequality}) and we show that the pair $(\ov S,D)$ is almost minimal and of log general type (\ref{lem:eps_and_gamma}(iv), \ref{lem:S-E_has_k=2}). Further bounds are obtained in section \ref{sec:surgeries} by studying the geometry of the pair $(\ov S,Q=D+E-C-\ti C)$, where $C$ and $\ti C$ are the last curves produced by the log resolution over each branch at infinity. Importantly, the surface $Y=\ov S\setminus Q$ has negative Euler characteristic and the pair $(\ov S,Q)$ turns out to be almost minimal. These facts constitute a basis of the improved bound $\gamma\leq 5$ (see \ref{prop:gamma<=5}). In many proofs we heavily rely on the logarithmic version of the Bogomolov-Miyaoka-Yau inequality \ref{lem:BMY} and on structure theorems for quasi-projective surfaces of non-general type. The nonexistence of a good asymptote is used again via \ref{lem:c1B-p1>=2} in sections $5$ and $6$, where we bound the possible types of $U$ at infinity.
\tableofcontents
\section{Preliminaries}
\subsection{Surfaces, divisors and minimal models}\label{ssec:generalities} We recall some notions and results from the theory of open algebraic surfaces. We refer the reader to \cite[\S2]{Miyan-OpenSurf} and \cite[\S3]{Fujita-noncomplete_surfaces} for details.
Let $X$ be a normal projective surface and $T=\sum_{i=1}^n m_iT_i$ a divisor contained in the smooth part of $X$ with $T_1,\ldots, T_n$ distinct, irreducible curves. We say that the pair $(X,T)$ is \emph{smooth} if $X$ is smooth and $T$ is a reduced simple normal crossing (snc) divisor. A $(b)$-curve on $X$ is a curve $L\cong \PP^1$ with $L^2=b$. An snc divisor $T$ is snc-minimal if a contraction of any $(-1)$-curve contained in $T$ leads to a divisor which is not snc. We call $Q(T)=(T_i\cdot T_j)_{1\leq i,j\leq n }$ the \emph{intersection matrix} of $T$ and we define the discriminant of $T$ by $d(T)=\det(-Q(T))$. We put $d(T)=1$ if $T=0$. If $T=T_1+\ldots+T_n$ is an ordered chain of rational curves (by definition the components of a chain are smooth) we write $T=[-T_1^2,\ldots,-T_n^2]$. We have, by elementary linear algebra, \begin{equation}\label{eq:d(T)_recurrence} d(T)=(-T_1^2)d(T_2+\ldots+T_n)-d(T_3+\ldots+T_n). \end{equation} The formula implies in particular that for \emph{admissible} chains, i.e.\ the ones with $T_i^2\leq -2$ for all $i$, one has $d(T-T_1)<d(T)$. A chain of $(-2)$-curves of length $m$ is denoted by $[(2)_m]$.
The linear equivalence of integral divisors and the numerical equivalence of divisors are denoted by $\sim$ and $\equiv$ respectively. If $\sigma\:X'\to X$ is a blowup then we say that it is \emph{inner} (\emph{outer}) for $T$ if the center belongs to exactly two (resp. exactly one) component of $T$.
\label{srem:weighted graph}Assume $T$ is a reduced snc divisor with connected support and a negative definite intersection matrix. We denote the local fundamental group of the analytic singularity arising from the contraction of $T$ by $\Gamma(T)$. If the singularity is of quotient type, i.e.\ analytically isomorphic to $0\in \C^2//G$ for some finite group $G<GL(2,\C)$, then we say that $T$ {\it is of quotient type}. In the latter case the singularity is algebraic (in fact rational). We allow the possibility that $G\cong \{1\}$, in which case the corresponding point is smooth. If $Q(T)$ is not negative definite we put formally $|\Gamma(T)|=\infty$. It is known (see \cite{Mumford}) that if $T$ has rational components then $\Gamma(T)$ depends only on the weighted dual graph of $T$. Also, $d(T)$ is the order of the first local integral homology group, which is the abelianization of $\Gamma(T)$. Moreover, $\Gamma(T)$ is finite if and only if the corresponding singularity is of quotient type.
Assume $T$ is of quotient type and snc-minimal. Then $T$ is a rational tree and contains at most one branching component (see \cite{Brieskorn}). If it does contain one (i.e.\ $T$ is a fork) then the corresponding singularity is non-cyclic, $T$ contains a unique branching component, which has self-intersection $b\leq -2$, and three rational admissible chains attached to it with discriminants $(d_1,d_2,d_3)=(2,3,3), (2,3,4), (2,3,5)$ or $(2,2,n)$ for some $n\geq 2$. The sequence $(-b;d_1,d_2,d_3)$ (and also the triple $(d_1,d_2,d_3)$), is called the \emph{type} of the fork $T$. If $T$ is a chain then either $T=[1]$ (if $\Gamma(T)=\{\id\}$) or $T$ is admissible and the corresponding singularity is cyclic.
We denote the Neron-Severi group of $X$ by $\NS(X)$ and its rank by $\rho(X)$. The number of irreducible components of a divisor $D$ is denoted by $\#D$. The following lemma is due to Fujita \cite[4.16]{Fujita-noncomplete_surfaces}.
\begin{lemma}\label{lem:rulings}
Assume $(\ov X,D)$ is a smooth pair and $\pi:\ov X\to B$ a $\PP^1$-fibration onto a smooth curve. Put $X=\ov X\setminus D$. Let $h$ be the number of horizontal components of $D$ and $\nu$ the number of fibers contained in $D$. Put $\Sigma_X=\sum_{F\not\subseteq D}(\sigma(F)-1)$, where $\sigma(F)$ is the number components of a fiber $F$ not contained in $D$ and where the sum is taken over all fibers not contained in $D$. Then the following relation holds $$\Sigma_X=h+\nu-2+\rho(X)-\#D.$$ \end{lemma}
As a consequence of the Hodge index theorem we have the following lemma.
\begin{lemma}\label{lem:Hodge lemma}Let $C_1,\ldots, C_r$ be distinct irreducible curves on a smooth projective surface $X$. If the matrix $(C_i\cdot C_j)_{i,j\leq r}$ is negative definite then $r< \rho(X)$.
\end{lemma}
Recall that in dimension $2$ running the log Minimal Model Program for a smooth pair $(X,D)$ results with a morphism onto a minimal model $\alpha_m\:(X,D)\to (X_m,D_m)$ for which the log surface $(X_m,D_m)$ is log terminal, hence $X_m$ has only quotient singularities. If $\tau\:(X_a,D_a)\to (X_m,D_m)$ is the minimal log resolution then there is a lift $\alpha_a\:(X,D)\to (X_a,D_a)$ of $\alpha_m$. The (smooth) pair $(X_a,D_a)$ is called an \emph{almost minimal model} of $(X,D)$ and the morphism $\alpha_a$ is well described (see \cite[2.3.11]{Miyan-OpenSurf}). If $D$ is snc-minimal and $X\setminus D$ is affine then $X_a\setminus D_a$ is an open subset of $X\setminus D$ with the complement being a disjoint sum of a finite number of curves isomorphic to $\C^1$ and $\tau$ contracts exactly the maximal admissible rational twigs of $D_a$. If $\kappa(K_X+D)\geq 0$ then the divisor $\tau^*(K_{X_m}+D_m)$ is the positive part of the Zariski-Fujita decomposition of $K_{X_a}+D_a$. The negative part can be described very explicitly in terms of a \emph{bark} of a divisor as follows (for details see \cite[\S3]{Fujita-noncomplete_surfaces} and \cite[\S2, section 3]{Miyan-OpenSurf}).
If $R=R_1+\ldots+R_s$ is a linearly ordered admissible chain we put $$e(R)=d(R_2+\ldots +R_s)/d(R).$$ Let $D$ be a reduced snc-minimal divisor with connected support. First assume $D$ is not of quotient type. Let $T_1,\dots, T_s$ be the maximal admissible twigs of $D$. By convention the components of $T_i$ are ordered linearly so that the tip of $T_i$ is its first component (the tip is the component meeting only one other component of $D$). In this case we put $e(D)=\sum_{i=1}^s e(T_i)$ and we define the \emph{bark of $D$} as the unique $\Q$-divisor supported on $\bigcup_i \Supp T_i$ and satisfying $$(K+D-\Bk D)\cdot Z=0$$ for every component $Z$ of $T_1+\ldots+T_s$. Equivalently, $Z\cdot \Bk D$ is $-1$ if $Z$ is a tip of $D$ and $0$ otherwise. One checks that $$\Bk^2 D=-e(D).$$ Now assume $D$ is of quotient type. In this case we define \emph{bark of $D$} as the unique $\Q$-divisor supported on $D$ satisfying $(K+D-\Bk D)\cdot Z=0$ for every component $Z$ of $D$. If $D=D_1+D_2+\ldots+ D_n$ is a chain then $$-\Bk^2(D)=e(D_1+D_2+\ldots+ D_n)+e(D_n+D_{n-1}+\ldots+ D_1).$$ If $D$ is a fork then the formula is a bit more complicated but we will not need it. For a general (reduced snc-minimal) $D$ one simply defines $\Bk D$ by summing barks of connected components of $D$. One shows that $\Bk D$ is an effective $\Q$-divisor with proper fractional coefficients and support equal to the sum of connected components of quotient type and the sum of maximal admissible twigs of the remaining connected components. What is most important for us is that (see (\cite[\S2, section 3]{Miyan-OpenSurf}) if $(X,D)$ is an almost minimal smooth pair with $\kappa(K_X+D)\geq 0$ then $$(K_X+D)^-=\Bk D.$$
The following version of the logarithmic Bogomolov-Miyaoka-Yau inequality follows from \cite{Langer} (for how it follows see \cite[5.2]{Palka-exceptional}). The original, weaker version was proved by Kobayashi-Nakamura-Sakai. The Euler characteristic of a topological space $Z$ will be denoted by $\chi(Z)$.
\begin{lemma}[The log BMY inequality]\label{lem:BMY}
Let $(X,D)$ be a smooth pair and let $D_1,\dots, D_k$ be the connected components of $D$ which are of quotient type. If $(X,D)$ is almost minimal and $\kappa(X\setminus D)\geq 0$ then \[\frac{1}{3}((K_{X}+D)^+)^2\leq \chi(X\setminus D)+\sum_{i=1}^k\frac{1}{|\Gamma(D_i)|},\] \end{lemma}
\subsection{The log resolution}\label{ssec:resol}
\begin{notation}\label{not:setup} Let $S=\C^2=\Spec \C[x,y]$. Let $U$ be a $\C^*$ embedded as a closed subset of $S$. Having coordinates $(x,y)$ on $\C^2$ we have an identification $\C^2=\PP^2\setminus L_\8$, where $L_\8$ is a line on $\PP^2$. We denote by $\bar U$ the closure of $U$ in $\PP^2$. Let $\lambda$ and $\tilde \lambda$ denote the branches of $\bar U$ at infinity i.e.\ the germs of $\bar U$ at $\bar U\cap L_\infty$. Note that an automorphism $\alpha$ of $\C^2$ gives rise to new coordinates $(\alpha^*x,\alpha^*y)$ on $\C^2$. It is proved in \cite{KR-Some_properties_of_Cstar} that if $U$ does not admit a good asymptote then there is a choice of coordinates on $\mathbb{C}^2$, such that $\lambda$ and $\tilde\lambda$ are disjoint. From now on throughout the paper we assume that it is the case.
Let $$\Phi\colon (\ov{S}',D'+E')\rightarrow (\mathbb{P}^2,L_\8+\bar U),$$ where $D'$ is the reduced total transform of $L_{\infty}$ and $E'$ is the proper transform of $\bar U$, be the minimal log resolution of singularities. By definition $\Phi^{-1}$ is the minimal sequence of blow-ups such that $D'+E'$ is an snc divisor. Let $L_\8'\subseteq D'$ be the proper transform of $L_\8$ in $\ov S'$ (see Fig. \ref{fig:Sbarprim}).
\begin{figure}[h] \includegraphics[scale=0.53]{Sbarprim}\caption{The log resolution $\Phi\: (\ov S',D'+E')\to (\PP^2,L_\8+\bar U)$.} \label{fig:Sbarprim}\end{figure}
\noin It may happen that $L_\8'$ is a $(-1)$-curve. Let $$\Psi\: (\ov{S}',D'+E')\to (\ov{S},D+E),$$ where $D=\Psi_*D'$ and $E=\Psi_*E'$, be the {\it snc minimalization of the divisor $D'$ with respect to $E'$}, i.e., $\Psi$ is the identity if $L_{\infty}'$ is not a $(-1)$-curve and otherwise it is the composition of successive contractions of $L_{\infty}'$ and then possibly of other $(-1)$-curves in the successive images of $D'$, such that $D+E$ is an snc divisor and each $(-1)$-curve of $D$ is a branching component of $D+E$. Now the only case when $D+E$ is not snc-minimal is when $E^2=-1$. Of course, $\ov{S}\setminus D=\ov S'\setminus D'=S=\C^2$ and $E'\cdot D'=E\cdot D=2$. Put $\gamma=-E^2$, $\gamma'=-(E')^2$ and define $\varepsilon$ by the equality $$(\ks+D+E)^2=2-\varepsilon.$$
\end{notation}
The assumption that a good asymptote for $U\subseteq S$ does not exist can be restated as follows.
\begin{lemma}\label{lem:no asymptote}
There is no curve $L\subset \ov S$ such that $L\cap S\cong \C^1$ and $L\cdot E\leq 1$. \end{lemma}
\subsection{The Hamburger-Noether pairs}\label{ssec:HN}
We write $D'=L'_\infty+F+\tilde F$ where $F=\Phi^{-1}(\lambda\cap L_\infty)_{red}$ and $\tilde F=\Phi^{-1}(\tilde\lambda\cap L_\infty)_{red}$. Let $C',\ti C'$ be the components of $D'$ meeting $E'$ contained in $L_\8'\cup F$ and $L_\8'\cup \tilde F$ respectively. Note that it may happen that $F=0$ (or $\ti F=0$). This means that the branch $\lambda$ crosses $L_\8$ normally, i.e.\ it is smooth and transversal to $L_\8$. We call such a branch \emph{simple}. The resolution process $\Phi$, see \ref{not:setup}, can be described in terms of Hamburger-Noether (HN-) pairs. For details we refer to \cite[Appendix]{KR-C*_actions_on_C3} or \cite{Russell_HN_pairs}. By $(T\cdot Z)_p$ we denote the local intersection index of two curves $T,Z$ at a point $p$.
We describe the process for $\lambda$, for $\ti \lambda$ it is analogous. As an input data we have a locally analytically irreducible branch $(\lambda_1,q_1)=(\lambda,\lambda\cap \bar U)$ and a curve $L_\8$ smooth at $q_1$. Let $x_1$ be a coordinate defining $L_{x_1}=L_\8$ at $q_1$. Put $$c_1=(\lambda_1\cdot L_{x_1})_{q_1}.$$ If $c_1=1$ (equivalently, if $\lambda$ and $L_\8$ cross normally at $q_1$) we put $p_1=0$, $h=0$ and we do nothing. Assume $c_1>1$. We then pick $y_1$ so that $(x_1, y_1)$ is a system of parameters at $q_1$ and $$p_1 = (\lambda_1\cdot L_{y_1})_{q_1}$$ is the multiplicity of $\lambda_1 \ \ \text {at} \ \ q_1.$ This forces $c_1\geq p_1$. We blow up successively over $q_1$ until the proper transform $\lambda_2 $ of $\lambda_1$ meets the inverse image of the divisor $L_{x_1}+L_{y_1}$ not in a node. The exceptional curves form a chain called {\it {the chain produced by the pair}} $\binom{c_1}{p_1}$. Let $C_1$ be the last exceptional curve. We then say that $C_1$ is the {\it {exceptional curve produced by the pair}} $\binom{c_1}{p_1}$. Since $\lambda_1$ is locally analytically irreducible, $\lambda_2$ meets $C_1$ in a unique point $q_2$ and does not meet any other component of the inverse image of $L_{x_1}+L_{y_1}$. We choose for $x_2$ a local coordinate for $C_1$ at $q_2$ and continue the process, noting that $c_2= \lambda_2 \cdot C_1 =\gcd(c_1, p_1)$. We continue this process until the proper transform of $\lambda_1$ meets the last exceptional curve normally. This describes the log resolution of $(\PP^2,\lambda_1 \cup L_{x_1})$ at $q_1$ and $$\binom{c_1}{p_1},\binom{c_2}{p_2},\dots ,\binom{c_h}{p_h}$$ are called the \emph{HN-pairs of $(\lambda_1,L_{x_1})$}. We note that $$c_{i+1}=\gcd(c_i,p_i)\text{\ for\ } i=1,\ldots,h-1 \text{\ and \ } \gcd(c_h,p_h)=1.$$
After a slight change of numbering we may write the sequences of HN-pairs of $\lambda$ and $\tilde\lambda$ as
\begin{equation}\binom{c_1}{c_1}_j,\binom{c_1}{p_1},\dots,\binom{c_h}{p_h}\text{\ \ and\ \ } \binom{\ti c_1}{\ti c_1}_{\ti j},\binom{\tilde c_1}{\tilde p_1},\dots, \binom{\tilde c_{\tilde h}}{\tilde p_{\tilde h}},\label{eq:def_of_pairs}\end{equation} where $p_1<c_1$ and $\ti p_1<\ti c_1$ and by $\binom{c}{c}_j$ we mean a sequence of pairs $\binom{c}{c},\ldots,\binom{c}{c}$ of length $j$. Moreover, interchanging the names $\lambda$ and $\ti \lambda$ if necessary, we may, and shall, assume that $\ti j\geq j$. As a consequence of the minimality of the resolution we have $p_h<c_h$ and $\ti p_{\ti h}<\ti c_{\ti h}$.
\bsrem The HN-pairs for $U\subseteq \C^2$ depend in general on the choice of coordinates on $\C^2$. We remark that to each HN-pair there is tacitly associated a complex number that determines the location of the branch on the last exceptional curve produced by the blowups prescribed by the pair; we will not make use of it.\esrem
Let $\mu_1,\mu_2,\ldots$ (resp. $\tilde \mu_1,\tilde \mu_2,\ldots$) be the sequence of multiplicities of all singular
points of $\lambda $ infinitely near $\lambda\cap L_\infty$ (resp. of $\tilde\lambda$ infinitely near $\tilde\lambda\cap L_\infty$). Then (\cite[Appendix]{KR-C*_actions_on_C3})
\begin{align*} \sum_{i \geq 1}\mu_i&=(j+1) c_1+p_1+p_2+\cdots +p_h-1.\\
\sum_{i \geq 1}\mu_i^2&=jc_1^2+c_1p_1+c_2p_2+\cdots +c_hp_h.\\
\sum_{i \geq 1}\tilde \mu_i&=(\ti j+1)\tilde c_1+\tilde p_1+\tilde p_2+\cdots +\tilde p_{\tilde h}-1.\\
\sum_{i \geq 1}\tilde\mu_i^2&=\ti j \ti c_1^2+\tilde c_1\tilde p_1+\tilde c_2\tilde p_2+\cdots + \tilde c_{\tilde h}\tilde p_{\tilde h}.\end{align*}
\begin{lemma}Let $d$ be the degree of $\bar U\subseteq \PP^2$. The following equations hold:
\begin{align} d&=c_1+\ti c_1,\\
2d+\gamma'&=jc_1+\sum_{i=1}^h p_i+\ti j \ti c_1+\sum_{i=1}^{\ti h} \ti p_i,\label{eq:1}\\
d^2+\gamma'&=jc_1^2+\sum_{i=1}^h c_ip_i+\ti j \ti c_1^2+\sum_{i=1}^{\ti h} \ti c_i\ti p_i,\label{eq:2}\\
h+j+\ti h+\ti j&=\gamma'+\varepsilon+2.\label{eq:h+hB}\end{align}
\end{lemma}
\begin{proof} By the definition of $c_1, \ti c_1$ we have $d=L_\8\cdot \bar U=L_\8\cdot(\lambda+\ti \lambda)=c_1+\ti c_1$. Tracking the self-intersection and intersections of the proper transforms of $\bar U$ with canonical divisors under blowups constituting the resolution $\Phi$ we get $K_{\ov S'}\cdot E'-K_{\PP^2}\cdot \bar U=\ds\sum_{i \geq 1}\mu_i+\ds\sum_{i \geq 1}\ti \mu_i$ and $\bar U^2-(E')^2=\ds\sum_{i \geq 1}\mu_i^2+\ds\sum_{i \geq 1}\ti \mu_i^2$. We compute $K_{\ov S'}\cdot E'-K_{\PP^2}\cdot \bar U=\gamma'-2+3d$ and $\bar U^2-(E')^2=d^2+\gamma'$, where $\gamma'=-(E')^2$. Using the first equation and the equations preceding the lemma we obtain the second and third equation.
The contractions in $\Psi$ are inner for $D'+E'$, so $\ks\cdot(\ks+D+E)=K_{\ov S'}\cdot (K_{\ov S'}+D'+E')=K_{\ov S'}\cdot E'+K_{\PP^2}\cdot(K_{\PP^2}+L_\8)-(h+j+\ti h+\ti j)=\gamma'+4-(h+j+\ti h+\ti j)$. On the other hand, since the arithmetic genus of $D+E$ vanishes, we have $\ks\cdot(\ks+D+E)=(\ks +D+E)^2=2-\varepsilon$. This gives the fourth equation.
\end{proof}
Recall that we assume that the branches at infinity of the closure of $U$ are separated. Note that if we blow according to a HN-pair $\binom{c}{p}$ then after making the first blowup, the branch either 'stays' on a given irreducible component of the boundary in case $c=p$, i.e.\ its proper transform meets the proper transform of the component, or it 'jumps', i.e.\ it separates from it in case $c>p$. Hence the following holds.
\blem The pair $(j,\ti j)$ defined above is the type of $U$ at infinity in the sense of \ref{def:type}.
\elem
\brem\label{rem:capacity_using_pairs} It is an elementary exercise to show that the maximal twig of $D+E'$ created by a pair $\binom{c_i}{p_i}$ has discriminant $c_i/\gcd(c_i,p_i)$ and its contribution to $e(D+E')$ is $(c_i-p_i)/c_i$.
\erem
\section{Basic results}
\subsection{Basic inequalities} We use the notation from the previous section.
\begin{lemma}\label{lem:eps_and_gamma} We have:
\begin{enumerate}[(i)]
\item $\ks\cdot (\ks+D+E)=2-\varepsilon,$
\item $\varepsilon\geq 0$,
\item $\gamma\geq 1$.
\item If $\gamma\neq 1$ then $(\ov S,D+E)$ is almost minimal and $(\ks+D+E)^-=\Bk {(D+E)}$.
\item $-\Bk^2(D+E)=e(D+E)\leq 1+\varepsilon$.
\end{enumerate}
\end{lemma}
\begin{proof} (i) $\ks\cdot (\ks+D+E)=(\ks+D+E)^2-2(p_a(D+E)-1)=(\ks+D+E)^2$.
(ii) Note that the support of $D+E$ contains a loop, so $|\ks+D+E|\neq \emptyset$ and hence $\kappa(\ks+D+E)\geq 0$ (see \cite[1.2.9]{Miyan-OpenSurf}). Let $\cal N$ be the negative part in the Zariski decomposition of the divisor $\ks+D+E$. By \cite{Miyaoka} (or by \cite[Corollary 5.2]{Langer}), $(\ks +D+E)^2\leq 3\chi(S\setminus E)+\frac{1}{4}\cal N^2=3+\frac{1}{4}\cal N^2$, so $\varepsilon\geq -1-\frac{1}{4}\cal N^2$. Since $U$ has no good asymptote, $\bar U$ is not a conic, hence $\Phi\neq \id$. The resolution $\Phi\colon (\ov{S}',D'+E')\rightarrow (\mathbb{P}^2,L_\8+\bar U)$ is minimal, so the last $(-1)$-curve produced by $\Phi$ is not a tip of $D'+E'$, hence it meets some twig of $D'+E'$. Let $W$ be a tip of one of these twigs. $W$ is not touched by $\Psi$ so, because $W^2<0$, it is contained in $\Supp \cal N$. Thus $\cal N\neq 0$ and we get $\varepsilon>-1$.
(iii) Suppose $\gamma\leq 0$. After blowing up over one of the points in $E\cap D$ we may assume that $E^2=0$. Then $U$ is a fiber of a $\C^*$-fibration of $\C^2$. The fibration is trivial over some Zariski open subset of the base, and hence the Euler characteristic of the total space over this subset vanishes. Thus, if $F\subseteq \C^2$ is the sum of the remaining fibers then $\chi(F)=\chi(\C^2)=1$. It follows that $F$ contains an irreducible (smooth) component with positive Euler characteristic. Since $\C^2$ contains no complete curves, it is necessarily $\C^1$. This is a good asymptote of $U$ and we reach a contradiction.
(iv) If $\gamma\neq 1$ then $D+E$ is snc-minimal. If $(\ov S,D+E)$ is not almost minimal then, since $D+E$ is connected and not negative definite, \cite[\S2.3]{Miyan-OpenSurf} implies that there exists a $\C^1$ contained in $S\setminus E$ witnessing the non-minimality. The latter is impossible by \ref{lem:no asymptote}.
(v) Suppose $\gamma\neq 1$. Let $\cal P$ be the positive part of the Zariski decomposition of $\ks+D+E$. By (iv) we compute $\cal P^2=2+e(D+E)-\varepsilon$. Since $\chi(S\setminus E)=1$, \ref{lem:BMY} gives (v). In case $\gamma=1$ we note that the snc-minimalization of $D+E$ does not touch the maximal twigs of $D+E$, so we get (v) by applying \ref{lem:BMY} to the resulting minimal model.
\end{proof}
Define $t_\lambda\in\{0,1\}$ by $t_{\lambda}=1$ if $p_h=1$ and $h>0$ and $t_{\lambda}=0$ otherwise. Define $t_{\ti \lambda}\in \{0,1\}$ analogously for $\ti \lambda$. Let \begin{equation}t=t_\lambda+t_{\ti \lambda}\in \{0,1,2\}\label{eq:def_of_t}.\end{equation}
We see easily that $t_{\lambda}=1$ if and only if $C$ is a $(-1)$-curve and $C$ together with some $(-2)$-twig of $D$ is contained in a twig of $D$. An analogous statement holds for $t_{\ti \lambda}$. The following inequality is proved in \cite[2.5]{KR-Some_properties_of_Cstar} as a consequence of the non-existence of a good asymptote.
\begin{proposition}\label{prop:basic_inequality}
$$2\varepsilon +\gamma\leq 7+t$$ \end{proposition}
\begin{corollary}\label{cor:gamma<9} $\gamma\leq 8$ and $h+j+\ti h+\ti j\leq 9+t-\varepsilon\leq 11$. \end{corollary}
\begin{proof}
Suppose that $\gamma\geq 9$. By \ref{prop:basic_inequality} $\gamma=9$, $\varepsilon=0$ and $t=2$. By the above remark $D+E$ has at least two $(-2)$-twigs. It follows also that the branches $\lambda$, $\tilde\lambda$ are not simple, because otherwise $t\leq 1$. A maximal twig $[(2)_k]$ contributes to $e(D+E)$ by $\frac{k}{k+1}\geq\frac{1}{2}.$
By \ref{lem:eps_and_gamma}(v) $e(D+E)\leq 1$, so $D+E$ has two maximal twigs and they are both $(-2)$-curves. In particular $D$, whence $D'$, is a chain. The sequences of characteristic pairs for $\lambda$ and $\ti \lambda$ are $\binom{2}{2}_j,\binom{2}{1}$ and $\binom{2}{2}_{\ti j},\binom{2}{1}$ respectively. Then the sum $jc_1+\sum p_i+\ti j \ti c_1+\sum\ti p_i=2j+2\ti j+2$ is even, which contradicts \eqref{eq:1}, because $\gamma=9$. The second inequality follows directly from \ref{prop:basic_inequality} and \eqref{eq:h+hB}.
\end{proof}
\subsection{Branches are not simple} Recall that a branch $\lambda$ or $\ti \lambda$ is \emph{simple} if and only it meets $L_\8$ normally.
\bprop\label{prop:no_simple_branch} One can choose coordinates on $\C^2$ so that the branches $\lambda$, $\ti \lambda$ are separated and not simple. In particular, in these coordinates $c_1, \ti c_1>1$ and $h,\ti h\geq 1$.\eprop
\begin{proof} Suppose one of the branches is simple. We may assume it is $\ti \lambda$. Then $\lambda $ is not simple, because otherwise $\bar U$ is a conic, which clearly has a good asymptote.
Suppose $j>0$. After the first blowing-up over $\lambda\cap L_\infty$ the proper transform of $L_\infty$ becomes a $(0)$-curve and it meets the proper transform of $\bar U$ once. Let $L$ be the proper transform of a general member of the linear system of this $(0)$-curve on $\ov S$. Then $L\cdot D=1$ and $L\cdot E=1$, so $L\cap S$ is a good asymptote of $U$; a contradiction.
Thus $j=0$. The morphism $\Psi\colon\ov S'\rightarrow\ov S$ is a composition of blowdowns starting from the contraction of $L'_\infty$ if $L'_\infty$ is a $(-1)$-curve (otherwise $\Psi=\id$). Let $u$ be the number of these blowdowns. Then $-\gamma=-\gamma'+u$. The formulas \eqref{eq:1} and \eqref{eq:2} take the form
\begin{align*}
\gamma+u+2c_1+2&=p_1+\dots +p_h\\
\gamma+u+(c_1+1)^2&=p_1c_1+\dots+p_hc_h.
\end{align*}
Let $c_1=kc_2$, $p_1=k'c_2$. Let $c_1-p_1=\beta c_2$ i.e.\ $\beta=k-k'$. The numbers $k$ and $k'$ are relatively prime. We rewrite the formulas in the following form
\begin{align}
\gamma+u+2+(\beta+k)c_2&=p_2+\dots+p_h,\label{eq:1_simple_branch}\\
\gamma+u+1+\beta kc_2^2+2kc_2&=p_2c_2+\dots+p_hc_h.\label{2_simple_branch}
\end{align}
Multiply \eqref{eq:1_simple_branch} by $c_2$ and subtract \eqref{2_simple_branch}. We get \begin{equation} (\gamma+u+2-2k)c_2-\gamma-1-u+(k+\beta-\beta k)c_2^2=\sum\limits_{i\geq 2}p_i(c_2-c_i)\geq 0,\label{eq:3_simple_branch}\end{equation} hence, because $\gamma\geq 0$ by \ref{lem:eps_and_gamma}(iii), \begin{equation}(\gamma+u+2-2k)c_2 >((k-1)(\beta-1)-1)c_2^2\label{eq:4_simple_branch}.\end{equation} Then $\gamma+u+1-2k\geq ((k-1)(\beta-1)-1)c_2.$
Suppose that $\beta\geq 2$. Then $k=k'+\beta\geq 3$ and the chain produced by $\binom{c_1}{p_1}$ which is contained between $L_\8'$ and the $(-1)$-curve created by the pair does not consist of only $(-2)$-curves. Hence it starts with $u-1$ $(-2)$-curves and then comes a $(\leq -3)$-curve. So the determinant of that chain, which is equal to $k'$, is at least $2u+1$. We obtain $2u\leq k'-1=k-\beta-1\leq k-3$, which gives $$\gamma> ((k-1)(\beta-1)-1)c_2+\frac{3k+1}{2}.$$ By \ref{cor:gamma<9} it follows that $8\geq \gamma> (k-2)c_2+\frac{1}{2}(3k+1)$, hence $k\leq 3$. Then $k=3$, so $c_2\leq 2$, $u=0$, $\beta=2$ and $k'=1$. It follows that $c_i=c_2=2$ for $i\geq 2$ or $h=1$. From \eqref{eq:3_simple_branch} we get $(\gamma-4)c_2=\gamma+1+c_2^2$, which has no solution for $c_2\leq 2$ and $\gamma\leq 8$; a contradiction.
Thus $\beta=1$. Then the chain produced by $\binom{c_1}{p_1}$ which is contained between $L_\8'$ and the $(-1)$-curve created by the pair consists of $k'-1$ $(-2)$-curves. $\Psi$ contracts $L_\infty'$ and that subchain, so $u=k'=k-1$. Let $r$ be the number of pairs equal to $\binom{c_2}{c_2}$. Rewrite \eqref{eq:3_simple_branch} as:
$$c_2(\gamma-k+1)-\gamma-k+c_2^2=\sum\limits_{i\geq r+3}p_i(c_2-c_i).$$
Because $c_i<c_2$ for $i\geq r+3$, we have $c_2-c_i\geq\frac{c_2}{2}$. Then $c_2(\gamma-k+1)-\gamma-k+c_2^2\geq\frac{c_2}{2}\sum\limits_{i\geq r+3}p_i$. We get $\sum\limits_{i\geq r+3}p_i< 2(\gamma-k+1)+2c_2 $. From \eqref{eq:1_simple_branch} we get $\gamma+k+(k+1)c_2+1=rc_2+p_{r+2}+\sum\limits_{i\geq r+3}p_i< rc_2+p_{r+2}+2\gamma-2k+2+2c_2$. Hence $$c_2(k-r-1)\leq p_{r+2}+\gamma-3k.$$
If $r\geq k'$ then, after performing $k'$ blowing ups of type $\binom{c_2}{c_2}$ and successive contractions starting from $L_\infty'$ we get new coordinates on $\C^2$ in which the branches of $\bar U$ are separated and none of them is simple. So we may assume that $r\leq k'-1=k-2$. By \eqref{eq:h+hB} $h=\gamma+k'+\varepsilon+2\geq \gamma+\varepsilon+r+3$, so since $\gamma\geq 0$, we have $h\geq r+3$. In particular, $c_{r+3}\geq 2$. The above inequality reads as $$c_2-p_{r+2}\leq\gamma-3k,$$ which gives $2\leq c_{r+3}\leq \gamma-3k\leq \gamma-6$. By \ref{prop:basic_inequality} it follows that $\gamma=8$ and $\varepsilon=0$, hence $k=2$, $k'=1$ and $c_i=2$ for $i\geq r+3$. Since $r\leq k-2$, we get also $r=0$, so $c_2>p_2$, hence $D+E$ has at least three tips. Since $k=2$ and $p_h=1$, at leat two of them are $(-2)$-tips, so $e(D+E)>1$. But $\varepsilon=0$, so we get a contradiction with \ref{lem:eps_and_gamma}(v).
\end{proof}
\subsection{Properties of the snc-minimalization $\Psi$} From now on we may, and shall, assume that both branches $\lambda$ and $\ti \lambda$ are not simple.
\begin{lemma}\label{lem:S-E_has_k=2} The surface $\ov S\setminus(D+E)=S\setminus U$ is of log general type. \end{lemma}
\begin{proof} The surface $S\setminus E$ has Euler characteristic $1$. Suppose it is $\C^1$- or $\C^*$-fibered. Then there is a line contained in some fiber (for a $\C^*$-fibration we argue as in \ref{lem:eps_and_gamma}(iii)). We may assume, blowing on $D$ if necessary, that the fibration extends to a $\PP^1$-fibration of $\ov S$. Let $L$ be the closure of the line and $F$ the fiber of the extension containing it. We have $E\cdot L\leq E\cdot F$. Since $S$ contains no completes curves, $E$ is not a $2$-section of the fibration. It follows that $E\cdot F\leq 1$, so $L$ is a good asymptote of $U$; a contradiction. Thus $S\setminus E$ is neither $\C^1$- nor $\C^*$-ruled. Because $S\setminus E$ contains no lines, the pair $(\ov S,D+E)$ becomes almost minimal after the snc-minimalization of $D+E$. By structure theorems for affine surfaces, if $S\setminus E$ is not of general type then $\kappa(S\setminus E)=0$ and by \cite[8.8]{Fujita-noncomplete_surfaces} the image of $D+E$, having arithmetic genus one, is a cycle of rational curves. Then $D+E$ is a cycle of rational curves. But, since both branches of $\bar U$ at infinity are not simple, $D+E$ has at least two tips. We reach a contradiction.
\end{proof}
\bcor\label{cor:eps=3-h0} $\varepsilon=3-h^0(2\ks+D+E)$.\ecor
\begin{proof} The Riemann-Roch theorem gives $\chi(\cal O_{\ov S}(2\ks+D+E))=\ks\cdot(\ks+D+E)+p_a(D+E)=3-\varepsilon$. Because $(K+D+E)^+$ is nef and (by \ref{lem:S-E_has_k=2}) big and because $(K+D+E)^-$ has proper fractional coefficients, the Kawamata-Viehweg vanishing theorem gives $\chi(\cal O_{\ov S}(2\ks+D+E))=h^0(2\ks+D+E)$.
\end{proof}
\bcor\label{lem:j<=1} $j\leq 1$. \ecor
\begin{proof} Suppose $j\geq 2$. Then $\ti j\geq 2$ and after blowing twice over each point of intersection of $\bar U$ with $L_\8$ the proper transform of $2L_\8$ together with the exceptional curves meeting it constitute a divisor $F_\8$ of type $[2,1,2]$ disjoint from the proper transform of $\bar U$. The linear system $|F_\8|$ gives a $\C^*$-fibration of $S\setminus U$, which contradicts \ref{lem:S-E_has_k=2}.
\end{proof}
If $j\neq 0$ then the line on $\PP^2$ tangent to $\lambda$ is different than $L_\8$. We denote its proper transform on $\ov S$ by $L_F$. We define $L_{\ti F}$ analogously.
\blem\label{lem:tangents} If $j\neq 0$ then $L_F$ meets the tip of $D'$ created by $\binom{c_1}{p_1}$. The tip and $L_F$ are not touched by $\Psi$. Moreover, $L_F\cdot E\geq 2$. \elem
\begin{proof} By \ref{lem:j<=1} $j=1$. Since the tangent to $\lambda$ on $\PP^2$ is a smooth curve, after blowing up twice over $\lambda\cap L_\8$ it separates from $\lambda$, which gives the first claim. The second claim follows from \ref{lem:Psi_properties}(a) below. Since $L_F\cap S\cong \C^1$, $L_F\cdot E\geq 2$ by \ref{lem:no asymptote}.
\end{proof}
\begin{notation}\label{not:C_and_Q} Recall that the curves $C'$ and $\ti C'$ (defined in \ref{ssec:HN}) are different. We put $C=\Psi_*(C')$, $\ti C=\Psi_*(\ti C')$, where $\Psi\: (\ov{S}',D'+E')\to (\ov{S},D+E),$ is as in \ref{not:setup}, and we define $Q_1$ (respectively $\ti Q_1$) as the connected component of $D-C$ (respectively of $D-\ti C$) which does not contain $\ti C$ (respectively $C$). Put $Q_0=D-C-\ti C-Q_1-\ti Q_1$. It follows that $Q_1$ and $\ti Q_1$ are chains. They are the maximal twigs of $D+E$ created by the pairs $\binom{c_h}{p_h}$, $\binom{\tilde c_h}{\tilde p_h}$ (nonzero, because the log resolution $\Phi$ is minimal). We denote by $G$ and $\ti G$ the components of $Q_0$ meeting $C$ and $\ti C$ respectively (see Fig. \ref{fig:Sbarprim}.) Let $T_1'$ (resp. $\tilde T_1'$) be the branching component of $D'+E'$ contained in $F$ (resp. $\tilde F$) nearest to $L'_\infty$. These are the last curves produced by the pairs $\binom{c_{1}}{p_{1}}$ and $\binom{\ti c_{1}}{\ti p_{1}}$ respectively (note in case $h=1$ we have $T_1'=C_1'$; similarly for $\ti T_1'$.) Put $T_1=\Psi_*T_1'$ and $\ti T_1=\Psi_*\ti T_1'$.
\end{notation}
\begin{lemma}\label{lem:Psi_properties}\
\begin{enumerate}[(a)]
\item $\Psi$ involves only contractions which are inner for $D'+E'$ and it does not touch maximal twigs of $D'+E'$.
\item $C'$, $\ti C'$ and $E'$ are not touched by $\Psi\colon\ov S'\To\ov S$. In particular, $\gamma'=\gamma$.
\item $C$ and $\wt C$ meet different components of $Q_0$, i.e.\ $G\neq \wt G$.
\item $Q_0$ is snc-minimal and contains only curves of negative self-intersection.
\end{enumerate}
\end{lemma}
\begin{proof}
(a) There is a unique chain in $D'$ meeting both $T_1'$ and $\tilde T'_1$ and not containing them. It contains $L'_\infty$ and its components are non-branching in $D'+E'$. All contractions in $\Psi$ take place inside this chain. The statement follows.
(b) Since neither $\lambda$ nor $\tilde \lambda$ is simple, $E'$ is disjoint from $L'_\infty$. Suppose $C'+\ti C'$ is touched by $\Psi$ at some stage (this may happen only if $h=1$ or $\ti h=1$). We have $\Psi\neq \id$, so $L_\8'$ is a $(-1)$-curve. Since the branches are not simple, this implies that $j>0$, hece $j=1$ by \ref{lem:j<=1}. At some stage of $\Psi$ the proper transform of $C'$ or of $\ti C'$ becomes a $0$-curve and its total transform on $\ov S'$ induces a $\PP^1$-fibration $\pi$ of $\ov S'$ with $E'$ as a $1$-section. Then $L_F$ is contained in a fiber, so it is met by $E$ at most once, hence it is a good asymptote; a contradiction. Therefore, $C'+\ti C'$ is not touched by $\Psi$. It follows that $E'$ is not touched.
(c) Suppose $G=\wt G$. Since by (b) $C'+\ti C$ is not touched, the proper transform of $G$ in $\ov S'$ is a component of $D'$ meeting both $C'$ and $\ti C'$. It must be $L'_\infty$, so $h=\ti h=1$ and $p_1=\ti p_1=1$. But then $\ov U$ is smooth, so it is a conic. We reach a contradiction, because in the latter case there exists a good asymptote of $U$.
(d) Suppose $Q_0$ contains a non-branching $(-1)$-curve $D_0$. Because, possibly with the exception of $E$, the divisor $D+E$ contains no non-branching $(-1)$-curves, $D_0$ is a branching component of $D+E$. Thus $D_0$ meets $C$ or $\ti C$, say $C$. We have $C\cdot \ti C=0$, so $D_0\neq \ti C$. Also, $\Psi\neq \id$, so $j\neq 0$. By (b) $C$ and $\ti C$ are $(-1)$-curves. By \ref{lem:tangents} $L_F\cdot (C+D_0)=0$ and $L_F\cdot E\geq 2$. The former equality implies that $L_F$ is contained in a fiber of the $\PP^1$-fibration of $\ov S$ given by the linear system $|C+D_0|$. Then $L_F\cdot E\leq (C+D_0)\cdot E=C\cdot E=1$; a contradiction.
Suppose $Q_0$ contains a component $D_0$ of non-negative self-intersection. Then $D_0$ is, by the definition of $D+E$, contained in the chain $T_1+R+\ti T_1$, where $T_1=\Psi_*T_1'$, $\ti T_1=\Psi_*\ti T_1'$ and $R$ is the chain in $D$ between $T_1$ and $\ti T_1$. Also, $\Psi\neq \id$, so $j>0$, hence $L_F$ is contained in a member of the linear system $|D_0|$. Then $L_F\cdot E\leq D_0\cdot E\leq 1$; a contradiction with \ref{lem:tangents}.
\end{proof}
\blem \label{prop:gamma>1} $\gamma\geq 2$.
\elem
\begin{proof}Suppose that $\gamma=1$, i.e.\ $E^2=-1$. Let $T_1$ and $\ti T_1$ be as in \ref{not:C_and_Q}. The divisor $F=C+E$ induces a $\PP^1$-fibration of $\ov S$ for which $D$ has three horizontal components: $G$, $\ti C$ and $Q_1^+\subseteq Q_1$. Suppose that $A$ is a $(-1)$-curve in $D-C-\ti C$. Then $A$ is contained in $Q_0$. By \ref{lem:Psi_properties}(d) $A$ is a branching component of $Q_0$. Also, $\Psi\neq \id$, so $j\neq 0$. If $A\cdot C=1$ then the linear system $|C+A|$ induces a $\PP^1$-fibration of $\ov S$ for which $E$ is a section, so $L_F$ if vertical and $L_F\cdot E\leq 1$, which contradicts \ref{lem:no asymptote}. Therefore, $A\cdot C=0$. Let $F_A$ be the fiber containing $A$. Since fibers of $\PP^1$-fibrations do not contain branching (in fiber) $(-1)$-curves, $C$ meets one of the adjacent components of $A$ in $D$, which is therefore a section of the fibration. It follows that multiplicity of $A$ in $F_A$ equals $1$. But the remaining two adjacent components meet $A$ and are contained in $F_A$, hence the multiplicity of $A$ is greater than $1$; a contradiction.
Thus $D-C-\ti C$ contains no $(-1)$-curves, so the only vertical $(-1)$-curve in $D$ is $C$. It follows that there is no fiber contained in $D$. Indeed, such a fiber would have to be smooth, and by \ref{lem:Psi_properties}(d) we know that $D$ contains no $(0)$-curves. Let $F_1$ be the fiber containing $\ti G$ (note $G\neq \ti G$ by \ref{lem:Psi_properties}(c)). By \ref{lem:rulings} $\Sigma_S=1$, i.e., there exists exactly one fiber $F_2$ which contains more than one, and in fact two, components not contained in $D$. Any other fiber contains a unique component not contained in $D$. It may happen that $F_1=F_2$. Suppose there is a singular fiber $F_0$ other than $E+C$, $F_1$ and $F_2$. The unique component $L_0\subseteq F_0$ not contained in $D$ is also the unique $(-1)$-curve in $F_0$. Since $L_0$ has multiplicity bigger than $1$ and since $F_0-L_0$ has at most two connected components (both contained in $D$), two of the sections contained in $D$ meet a common connected component of $F_0-L_0$. Since $C$ is not a component of $F_0$, the sections are necessarily $G$ and $\ti C$, hence $F_0$ contains $\ti G$; a contradiction. Since the three sections contained in $D$ are disjoint, it is easy to see that we may contract successively all $(-1)$-curves in $E+C+F_1+F_2$ in such a way that the images of these sections, call them $H_1$, $H_2$, $H_3$, remain disjoint. Because there are no other singular fibers, this results with a morphism $\ov S\to \F$ onto a Hirzebruch surface $\F$. Since $H_2-H_3$ intersects trivially with $H_1$ and with a fiber, it is numerically trivial, hence $H_2^2=H_2\cdot H_3=0$. Then $\F=\PP^1\times \PP^1$, so $S\setminus (E+C+F_1+F_2)$ is isomorphic to $\PP^1\times \PP^1$ with three fibers and two disjoint sections (images of $G$ and $Q_1^+$) removed, i.e.\ to $\C^{**}\times \C^*$. But this means that $S\setminus E=S\setminus U$ contains an open subset with Kodaira dimension equal to $\kappa(\C^{**}\times \C^*)=1$, which contradicts \ref{lem:S-E_has_k=2}.
\end{proof}
\section{\texorpdfstring{Surgeries on $(\C^2,U)$}{Surgeries on (C2,U)}}\label{sec:surgeries}
In this section we analyse surfaces resulting from surgeries on $S=\C^2$. We cut out $U$ and we glue in $C$, $\ti C$ or both. Studying the geometry of resulting surfaces we obtain a lot of information on the $\C^*$-embedding $U\mono S=\C^2$. In particular, we improve the upper bound on $\gamma$ to $\gamma\leq 5$.
\subsection{Double sided surgery - the surface $Y$}
By \ref{lem:Psi_properties}(a) $\Psi$ does not touch maximal twigs of $D'+E'$. Recall that $Q_1$ and $\wt Q_1$ are the maximal twigs of $D+E$ which are images of maximal twigs of $D'+E'$ produced by the pairs $\binom{c_h}{p_h}$ and $\binom{\ti c_{\ti h}}{\ti p_{\ti h}}$ respectively (see \ref{not:C_and_Q}). We have $D=Q_1+C+Q_0+\ti C+\ti Q_1$. Put $Q=Q_1+Q_0+\ti Q_1+E$ and $Y=\ov S\setminus Q.$ This surface is obtained from $S=\C^2$ by cutting out $E$ and gluing in $C+\ti C$. We have $\chi(Y)= -1$.
\begin{lemma}\label{lem:k(Y)} If $\gamma=2$ then $\kappa(Y)=-\8$. If $\gamma+t\geq 6$ (or, more generally, if $\varepsilon\leq t$) then $2\ks+Q\geq 0$, so $\kappa(Y)\geq 0$. \end{lemma}
\begin{proof} If $\gamma=2$ then $nE$ is in the fixed part of a divisor $n(\ks+Q)$, so $\kappa(Y)=\kappa(\ks+Q_0+Q_1+\ti Q_1)=\kappa(\ks+D-C-\wt C)\leq \kappa(\ks+D)=-\8$. Note that if $\gamma+t\geq 6$ then \ref{prop:basic_inequality} gives $\varepsilon\leq t+\frac{1}{2}$. Thus for the proof of the second part of the lemma we may assume $\varepsilon\leq t$. By \ref{lem:Psi_properties}(b) $C$ and $\wt C$ are $(-1)$-curves, so $$\ks\cdot(\ks+Q)=\ks\cdot(\ks+D+E)-\ks\cdot C-\ks\cdot \ti C=4-\varepsilon.$$ We have $\varepsilon\leq 2$, so by \ref{cor:eps=3-h0} $2\ks+D+E\geq 0$. Suppose $h^0(2\ks+Q)=0$. If $\varepsilon=0$ then the Riemann-Roch theorem gives $h^0(2\ks+Q)+h^0(-\ks-Q)\geq \ks\cdot(\ks+Q)+p_a(Q)=1,$ so $\ks+C+\wt C=(2\ks+D+E)+(-\ks-Q)\geq 0$ and hence $\ks\geq 0$, which is impossible.
Suppose $\varepsilon=1$. Then $\ks\cdot(\ks+Q)=3$ and $t\geq 1$, so $Q_1$ or $\ti Q_1$, say $\ti Q_1$, consists of $(-2)$-curves. The Riemann-Roch theorem gives $h^0(-\ks-Q_0-Q_1-E)+h^0(2\ks+Q_0+Q_1+E)> 0,$ so $\ks+C+\ti C+\ti Q_1=2\ks+D+E+(-\ks-Q_0-Q_1-E)\geq 0,$ hence $\ks\geq 0$; a contradiction.
Thus $\varepsilon=2$. Then $\ks\cdot(\ks+Q)=2$ and $t=2$, so $Q_1$ and $\ti Q_1$ consist of $(-2)$-curves. Now $h^0(-\ks-Q_0-E)+h^0(2\ks+Q_0+E)>0,$ so $\ks+C+\ti C+\ti Q_1+Q_1=2\ks+D+E+(-\ks-Q_0-E)\geq 0,$ hence $\ks\geq 0$; a contradiction.\end{proof}
\begin{lemma}\label{lem:log-exc_curves_on_Y} Assume $Y$ has no $\C^1$-fibration. Then there is no $(-1)$-curve $L$ on $\ov S$ such that $L\cdot Q_0=0$, $L$ meets two connected components of $Q_1+E+\ti Q_1$ and together with these components contracts to a quotient singularity. \end{lemma}
\begin{proof} Suppose that such an $L$ exists. Let $\pi\colon \ov S\To \ov X$ be the contraction of $L$ and the two connected components of $Q_1+E+\ti Q_1$ to the point $q_1\in \ov X$ (the possibility that $q_1$ is smooth is not excluded). Let $Q_2$ be the third connected component of $Q_1+E+\ti Q_1$. The surface $X=\ov S\setminus Q_0$ is a sum of the open topological subspaces $S$ and the sum of tubular neighbourhoods of $C+Q_1$ and $\ti C+\ti Q_1$, which are all simply connected. By the van Kampen theorem $X$ is simply connected. Let $\ov X\To \ov X'$ be the contraction of $Q_2$ to a cyclic singular point $q_2\in X'$. Then $X'=\ov X'\setminus Q_0$ is simply connected and we have $\rho(\ov X')=\#Q_0$. Since the components of $D$ are independent in $\NS(\ov S)$, the components of $Q_0$ are independent in $\NS(\ov S)$, hence the components of $Q_0$ are independent in $\NS(\ov X')$. It follows that they generate $\NS(\ov X')$, so $X'$ is affine by an argument by Fujita \cite[2.4(3)]{Fujita-noncomplete_surfaces}. It follows that $b_3(X')=b_4(X')=0$ and $H_2(X',\Z)$ has no torsion. Since $\chi(X')=1$, we have $b_2(X')=b_1(X')=0$, so $X'$ is simply connected and $\Z$-acyclic, hence contractible. Moreover $\kappa(X')=\kappa(\ov S\setminus Q_0)\leq \kappa(S)=-\8$. Because $q_2\in X'$ is a non-trivial cyclic singularity, by \cite[1.1, 3.1]{KR-Contractible_surfaces} $X'\setminus \Sing X'$ has a $\C^1$-fibration. Then $Y$ has a $\C^1$-fibration; a contradiction.
\end{proof}
\bprop\label{lem:Y minimal} If $Y$ has no $\C^1$-fibration then the pair $(\ov S, Q)$ is almost minimal.\eprop
\begin{proof} By \ref{prop:gamma>1} and \ref{lem:Psi_properties} $Q$ snc-minimal. Let $(\ov{Y}',T')$ be an almost minimal model of $(\ov S,Q)$. $\ov{Y}'$ is obtained from $\ov{S}$ by a sequence of birational morphisms $$\pi_i : \ol{Y}_i \to \ol{Y}_{i+1},\ \ \ol{S}=\ol{Y}_0 \to \ol{Y}_1 \to \cdots \to \ol{Y}_\ell=\ol{Y}'.$$ Let $T_i=(\pi_{i-1})_*(T_{i-1})$, $T_0=Q$, $T'=T_\ell$. Let $Y_i=\ol{Y}_i\setminus T_i$. For every $i$ there exists a $(-1)$-curve $C_i \nsubseteq T_i$ such that $\pi_i : \ol{Y}_i \to \ol{Y}_{i+1}$ is the {\em snc-minimalization} of $C_i+T_i$. Finally, for the almost minimal model $(\ol{Y}',T')$, the negative part $(K_{\ol{Y}'}+T')^-$ coincides with the bark $\Bk T'$ if $\kappa(Y)\geq 0$. The contractions in this process involve only curves (or their images) contained in the support of $(\ks+Q)^-$. We put
$$e(\ol{Y_i},T_i)=\chi(\ol{Y_i}\setminus T_i)+\#\{\mbox{connected components of $T_i$}\}.$$ Relying on the theory of peeling \cite[2.3.6]{Miyan-OpenSurf}, which gives a description of curves $C_i$, we find that $e(\ov Y_{i+1},T_{i+1})=e(\ov Y_i,T_i)-1.$ Hence $e(\ov Y',T')=e(\ov S,Q)-\ell=3-\ell.$
Suppose that $(\ov S, Q)$ is not almost minimal, i.e.\ $\ell\geq 1$. If all connected components of $T'$ are of quotient type then the intersection matrix of $Q+C_0$ is negative definite and has rank $\#Q+1=\rho(\ov Y')$, which contradicts \ref{lem:Hodge lemma}. Thus $T'$ contains a connected component which is not of quotient type.
Suppose $\kappa(Y)=-\8$. There does not exist a $\PP^1$-fibration of $Y$, because then $E$ would be a smooth fiber, and this is impossible by \ref{lem:eps_and_gamma}(iii). Because $Y$ has no $\C^1$-fibration, by \cite{MiTs-PlatFibr} $T'$ consists of two disjoint forks, exactly one of which is of quotient type and $Y'\cong(\C^2-\{0\})/ G$, where $G$ is a finite group. Hence $\chi(Y')=0$ and therefore $e(\ov Y',T')=2$. It follows that $\ell=1$. But then two connected components of $Q$ do not meet $C_0$, hence one of the connected components of $T'$ is a chain; a contradiction.
Thus $\kappa(Y)\geq 0$. Let $k$ be the number of connected components of $T'$ which are of quotient type and let $G_j$ be the local fundamental groups. The resulting quotient points are singular because $T'$ is snc-minimal, hence $|G_j|\geq 2$. Let $u$ be the number of connected components of $T'$. We have $\ell\geq 1$ and, by the argument above, $k\leq u-1$. By \ref{lem:BMY} $$\chi(Y')+\frac{k}{2}\geq \chi(Y')+\sum\limits_{i=1}^k \frac{1}{|G_j|}\geq 0.$$ We have $\chi(Y')=e(\ov Y',T')-u=3-\ell-u$. We obtain $3-\ell -u +\frac{k}{2}\geq 0,$ so $3\geq \ell+u-\frac{k}{2}\geq \ell +1+\frac{k}{2}$, hence $k\leq 2$ and $u\leq 3$. Suppose $\chi(Y_{i+1})>\chi(Y_i)$ for some $i$. This is possible only if $C_i$ meets two connected components of $T_i$ and contracts to a smooth point together with these connected components. By \ref{lem:log-exc_curves_on_Y} one of these connected components contains the image of $Q_0$. But then all connected components of $T'$ are of quotient type; a contradiction.
It follows that $\chi(Y')\leq \chi(Y)=-1$, hence $k=2$. Then $|G_1|=|G_2|=2$, $\ell=1$ and $u=3$. Also, $\chi(Y')=-1=\chi(Y)$. The latter implies that $C_0$ meets two connected components of $Q$. Suppose that $C_0\cdot Q_0=0$. Then, since $Q_1+E+\ti Q_1+C_0$ is contained in $\Supp (\ks+Q)^-$, the intersection matrix of $Q_1+E+\ti Q_1+C_0$ is negative definite, which implies that the intersection matrix of $Q_0$ is not negative definite. Hence $Q_0$ is not a contractible connected component of $T'$. Since $k=2$, $C_0$ together with two connected components of $Q_1+E+\ti Q_1$ contracts to a quotient singularity; it is impossible by \ref{lem:log-exc_curves_on_Y}. Thus $C_0$ meets $Q_0$ and one of connected components of $Q_1+E+\ti Q_1$. Since $|G_1|=|G_2|=2$, the remaining two connected components are $(-2)$-curves. By \ref{lem:BMY} $((K_{\ov Y'}+T')^+)^2=0$, so $\kappa(Y)\leq 1$. We have $\gamma\neq 2$, otherwise $\kappa(Y)=-\8$ by \ref{lem:k(Y)}.
Thus $C_0$ meets $Q_0$ and $E$ and we have $d(Q_1)=c_h=2$, $d(\ti Q_1)=\ti c_{\ti h} =2$. By \eqref{eq:1} and \eqref{eq:2} $4$ divides $d^2-\gamma$ and $2$ divides $d$, hence $4$ divides $\gamma$. Suppose that $\gamma=8$. Then $\varepsilon=0$ and by \ref{lem:eps_and_gamma}(v) $D+E$ has two $(-2)$-tips $Q_1, \ti Q_1$ and no other tip. Hence $D$ is a chain so $Q_0$ is a chain. Then $Q_0$ is contained in $\Supp(\ks+Q)^-$ and $Q+C_0$ is contained in $\Supp(\ks+Q)^-$, hence the intersection matrix of $Q+C_0$ is negative definite, which again contradicts \ref{lem:Hodge lemma}.
Thus $\gamma=4$. By \ref{prop:basic_inequality} $\varepsilon\leq 2$. We have $t=2$, so by \ref{lem:k(Y)} $2\ks+Q\geq 0$. Also, $Q_0+C_0+E$ is not of quotient type and $\kappa(Y)=0,1$. We write $T'=Z+Q_1+\ti Q_1$ where $Z$ is the image of $Q_0+C_0+E$ in $\ov Y'$. Note that since $Z$ is not negative definite, it cannot be a chain. Indeed, otherwise it contains a component of non-negative self-intersection, which would imply that $Y'$, and hence $Y$, is $\PP^1$- or $\C^1$-fibered. Let $\cal P=(K_{\ov Y'}+T')^+$. We have $2K_{\ov Y'}+T'\geq 0$ and $\cal P\cdot (K_{\ov Y'}+T')=\cal P^2=0$. Since $2K_{\ov Y'}+2T'\geq T'$, we obtain $0\geq \cal P\cdot T'$, i.e.\ $\cal P\cdot V=0$ for every irreducible component $V$ of $T'$. By \cite[8.8]{Fujita-noncomplete_surfaces} either $Z$ is a fork $H+R_1+R_2+R_3$ where $H$ is a branching component and $R_1,R_2, R_3$ are the maximal twigs or $Z$ is a rational tree which consists of a nonzero chain and two $(-2)$-curves attached to each tip of the chain (four $(-2)$-curves in total). We have $K_{\ov Y'}\cdot (K_{\ov Y'}+T')=\ks\cdot(\ks+Q+C_0)+m=3-\varepsilon+m\geq m+1$, where $m$ is the number of outer (with respect to $Q_0+C_0+E$) blow-downs in $\ov S\rightarrow\ov Y'$. From this $-4+\Bk^2 Z=\Bk^2 T'=(K_{\ov Y'}+T')^2=m-\varepsilon-3\geq m-5$, hence $\Bk^2 Z\geq m-1$. Then $m=0$ and $0>\Bk^2 Z=1-\varepsilon$, hence $\varepsilon=2$ and $\Bk^2 Z=-1$. But then $Z$ cannot have four $(-2)$-tips, hence $Z$ is a fork. From the description of the fork in loc. cit. we get $\frac{1}{d(R_1)}+\frac{1}{d(R_2)}+\frac{1}{d(R_3)}=1$, so $(d(R_1), d(R_2), d(R_3))=(3,3,3)$ or $(2,4,4)$ or $(2,3,6)$. Since $\Bk^2 Z=-1$ one checks easily that the maximal twigs of $Z$ are tips, so $\#Z=4$. We have $\rho(\ov S)=\#(Q_0+C_0)$, so $\rho(\ov Y')=\#T'=6$, hence $K_{\ov Y'}^2=4$ by the Noether formula. Because $K_{\ov Y'}\cdot (K_{\ov Y'}+T')=1$, we get $K_{\ov Y'}\cdot Z=K_{\ov Y'}\cdot T'=-3.$ Since $\sum\limits_{i=1}^3K_{\ov Y'}\cdot R_i>0$ we get $K_{\ov Y'}\cdot H\leq -4$, hence $H^2\geq 2$. Now since $m=0$, the blowing ups in the reverse of $\ov S\rightarrow \ov Y'$ are inner, i.e.\ we blow up over $H\cap R_i$ for some $i$. Moreover, we blow up only once on $H$, because the connected component of $Q$ containing $E$ consists of $E$ only. So the proper transform of $H$ in $\ov S$ is a positive curve contained in $D$; a contradiction by \ref{lem:Psi_properties}(d).
\end{proof}
\blem\label{lem:Y_geometry} If $\kappa(Y)=-\8$ then $Q_0$ is of quotient type or $Y$ has a $\C^1$-ruling with no base points on $\ov S$, $Q_0$ is branched and has a maximal twig of type $[(2)_{\gamma-1}]$. If $\kappa(Y)\geq 0$ then $$\frac{1}{\ti c_h}+\frac{1}{\ti c_{\ti h}}+\frac{1}{\gamma}+\frac{1}{\Gamma(Q_0)}\geq 1.$$ \elem
\begin{proof} If $\kappa(Y)\geq 0$ then $((K_{\ov S}+Q)^+)^2\geq 0$, so the above inequality follows from \ref{lem:BMY}. Assume $\kappa(Y)=-\8$ and $Y$ is not $\C^1$-ruled. If $Y$ is $\PP^1$-ruled then by \ref{lem:rulings} $\nu=\Sigma_Y+1\geq 1$, so $Q$, being snc-minimal, contains a $0$-curve. But the latter is impossible by \ref{lem:Psi_properties}(d), hence $Y$ is not $\PP^1$-ruled. By \ref{lem:Y minimal} $(\ov S,Q)$ is almost minimal, so after the contraction of connected components of $Q$ which are of quotient type it becomes a log del Pezzo surface of rank one. Since $Q$ has more than two connected components, by \cite[2.5.1]{Miyan-OpenSurf} the resulting del Pezzo is closed, i.e.\ all connected components of $Q$ are of quotient type. Assume $p\:Y\to B$ is a $\C^1$-fibration. Again, $\nu=0$, otherwise $Q$ would contain a $0$-curve. Let $(\ov S,\ti Q)\to (\ti S,Q)$ be a minimal modification over $Q$ such that $p$ has no base points of $\ov S$. We have $\Sigma_Y=\nu=0$, so every singular fiber has a unique component not contained in $\ti Q$, and hence, by the minimality, a unique $(-1)$-curve. The complement of this $(-1)$-curve in the fiber has at most two connected components, one of which meets the section contained in $\ti Q$. Since $\ti Q$ has four connected components, there are at least three singular fibers. This is possible only if $\ti S=S$ and the section contained in $Q$ is in fact contained in $Q_0$. In particular, $p$ has no base points on $\ov S$. Let $F_E$ be the fiber containing $E$. Since $E$ is irreducible, we have necessarily $F_E=[(2)_{\gamma-1},1,\gamma]$, so $Q_0$ contains a maximal twig of type $[(2)_{\gamma-1}]$.
\end{proof}
While in principle the divisor $Q_0$ my be complicated, this is not so if $\gamma\geq 6$.
\bcor \label{lem:Q0_quotient_for_gamma>=6} If $\gamma\geq 6$ then $Q_0$ is of quotient type. \ecor
\begin{proof} Suppose $Q_0$ is not of quotient type. From \ref{lem:k(Y)} we get that $\ovk(Y)\geq 0.$ By \ref{lem:Y minimal} $(\ov S, Q)$ is almost minimal. The log BMY inequality gives $1\leq \frac{1}{c_h}+\frac{1}{\ti c_h}+\frac{1}{\gamma}$ and the inequality is strict if $\ovk(Y)=2$. Suppose $c_h=\ti c_{\ti h}=2$. Equations \eqref{eq:1} and \eqref{eq:2} imply that $4$ divides $\gamma$, hence $\gamma=8$. Then $\varepsilon=0$ and $e(D+E)\leq 1$ by \ref{lem:eps_and_gamma}(v). Since $Q_1$ and $\ti Q_1$ are $(-2)$-tips of $D+E$, $D+E$ has no more tips, so $D$ is a chain. But then $Q_0$ is a chain, hence of quotient type; a contradiction. Therefore $\max\{c_h,\ti c_{\ti h}\}>2$. Since $\gamma\geq 6$, the log BMY is an equality, so $\{c_h, \ti c_h\}=\{2,3\}$, $\gamma=6$ and $\cal P^2=0$. In particular, $\ovk(Y)=0,1$. Say $\ti c_2=2$. Then $\Bk^2 \ti Q_1=-2$ and $d(Q_1)=c_h=3$.
By \ref{lem:k(Y)} $2\ks+Q\geq 0$. We put $\cal{P}=(\ks+Q)^+$ and we argue as in the proof of \ref{lem:Y minimal} that $\cal{P}\cdot Q=0$ and hence that $Q_0$ is a fork with $\delta=1$ or a tree with exactly four $(-2)$-tips. Assume the latter case. Then $\Bk^2 Q_0=-2$. We have $-4-\varepsilon=(\ks+Q)^2=\Bk^2 Q=-4-\frac{4}{6}+\Bk^2 Q_1$, i.e.\ $\varepsilon=-\Bk^2 Q_1+\frac{2}{3}.$ Now, since $d(Q_1)=3$, $Q_1$ is either a $(-3)$-curve or a chain of two $(-2)$-curves, so $\varepsilon=2$ or $\varepsilon=2+\frac{2}{3}$. But if $\varepsilon=2$ then $\gamma\leq 5$ by \ref{prop:basic_inequality}; a contradiction.
Thus, $Q_0$ is a fork with a branching component $H$ and three maximal twigs $R_1$, $R_2$, $R_3$, such that $\delta(Q_0)=\frac{1}{d(R_1)}+\frac{1}{d(R_2)}+\frac{1}{d(R_3)}=1$. We have $-4-\varepsilon=(\ks+Q)^2=\Bk^2 Q= \Bk^2 Q_1-2-\frac{2}{3}-e(Q_0)$. By \ref{prop:basic_inequality} $\varepsilon=1$, so $1\leq \delta(Q_0)\leq e(Q_0)=\frac{7}{3}-e(Q_1)$. If $Q_1=[2,2]$ then $e(Q_1)=2$ and the latter inequality fails. Therefore, $Q_1$ is a $(-3)$-curve, hence $e(Q_1)=\frac{4}{3}$ and $\delta(Q_0)=e(Q_0)=1$. It follows that the twigs of $Q_0$ are irreducible. Then $b_2(\ov S)=\#Q+1= 8$, so $\ks^2=2$ by the Noether formula. Since $\ks\cdot (\ks+Q)=4-\varepsilon=3$ we get $\ks\cdot Q=1$. We compute $\ks\cdot H=1-\ks\cdot \sum R_i-\ks\cdot Q_1-\ks\cdot \ti Q_1=-\ks\cdot \sum R_i$. But $\ks\cdot\sum R_i\geq 3$, so $\ks\cdot H\leq -3$, i.e.\ $H^2\geq 1$; a contradiction in view of \ref{lem:Psi_properties}(d).
\end{proof}
\subsection{One-sided surgeries - surfaces $Y_C$ and $Y_{\ti C}$}
By \ref{cor:gamma<9} $\gamma\leq 8$. The goal of this subsection is to improve this bound to $\gamma\leq 5$. We suppose, for a contradiction, that $\gamma\geq 6$. By \ref{prop:basic_inequality} $\varepsilon \leq 1$. We introduce some notation. By $ Y_C$ (resp. $ Y_{\tilde C}$) we denote the surface $\ov S\setminus (D+E-C)$ (resp. $\ov S\setminus (D+E-\tilde C)$). So $Y_C\setminus C=Y_{\ti C}\setminus {\tilde C}=Y$. The boundary divisor of $Y_C$ equals $Q+\tilde C$ and the boundary divisor of $Y_{\tilde C}$ equals $Q+\tilde C$. Let $Y_C'$ (resp. $Y_{\ti C}'$) be the surface obtained from $Y_C$ (resp. $Y_{\ti C}$) by contracting $Q_1$ (resp. $\ti Q_1$) to a (quotient) singular point $z$ (resp. $z'$). Recall that $d(Q_1)=c_h$ and $d(\ti Q_1)=\ti c_{\ti h}$, see \ref{rem:capacity_using_pairs}.
\blem\label{lem:no_simple_fibers_in_D}\ \benum[(a)]
\item $D$ does not contain a component $D_0$ for which $D_0^2\geq 0$.
\item $D$ does not contain components $D_1$, $D_2$ for which $D_1^2=D_2^2=-1$ and $D_1\cdot D_2=1$.
\eenum\elem
\begin{proof} If $j=0$ then $D'=D$ is snc-minimal and both statements are true. So by \ref{lem:j<=1} we may assume $j=1$. Suppose (a) or (b) fails. In case (a) blow up on $D_0$ until it becomes a $0$-curve; denote it by $F_\8$. In case (b) put $F_\8=D_1+D_2$. In both cases $L_F\cdot F_\8=0$ and $E\cdot F_\8\leq 1$ (note that $E$ meets $D$ only in $C$ or $\ti C$ and we have $C\cdot \ti C=0$). Thus $L_F$ is contained in a fiber of the $\PP^1$-fibration given by $|F_\8|$ and $E\cdot L_F\leq E\cdot F_\8\leq 1$. By \ref{lem:tangents} this is a contradiction.
\end{proof}
\blem With the above notation we have:\benum[(i)]
\item $\kappa(Y_C'\setminus\{z\}) \geq 0$.
\item The pair $(\ov S, Q+\tilde C)$ is almost minimal.
\eenum Analogous statements hold for $Y_{\tilde C}$ and $(\ov S,Q+C)$.
\elem
\begin{proof} (i) Suppose that $\kappa(Y'_C\setminus\{z\})=-\infty$. First of all note that $\ti C$ is a branching $(-1)$-curve in $Q_0+\ti C+\ti Q_1+E$, which implies that the latter divisor cannot be vertical for a $\PP^1$-fibration of $\ov S$. It follows that $Y'_C\setminus\{z\}$ is not $\PP^1$-ruled. Suppose it is also not $\C^1$-ruled. Since $\chi(Y_C')=1$, by the result of Miyanishi-Tsunoda \cite{MiTs-PlatFibr} $Y_C'$ is isomorphic to $\C^2// G$, where $G$ is a small non-abelian subgroup in $GL(2,\C)$. But $z$ is a cyclic quotient singularity, so we come to a contradiction. So there is a $\C^1$-fibration $g\: Y_C\setminus Q_1\To \PP^1$. Let $X\to \ov S$ be a minimal modification over $Q+\ti C$, such that $g$ extends to a $\PP^1$-fibration $G\: X\To \PP^1$. From Lemma \ref{lem:rulings} we get $\nu>0$. Let $G_0$ be a fiber contained in $X\setminus Y_C$. If $G_0$ is irreducible then the image of $G_0$ in $Q+\ti C$ is a nonnegative curve; a contradiction by \ref{lem:no_simple_fibers_in_D}. Hence $G_0$ contains a $(-1)$-curve $B'$, which, by the minimality of the modification, is a branching component in $X\setminus Y_C$. Since $B'$ is not a branching component of $G_0$, it meets a horizontal component of the boundary. But the latter is a section, so the multiplicity of $B'$ in $G_0$ equals $1$ and hence $B'$ meets only one component of $G_0$. Then $B'$ is not a branching component in $X\setminus Y_C$; a contradiction.
(ii) Put $Q_2=Q_0+\ti C+\ti Q_1+E=Q+\ti C-Q_1$. Since $\#(Q_1+Q_2)=\rho(\ov S)$, \ref{lem:Hodge lemma} implies that the intersection matrix of $Q_2$ is not negative definite. By \ref{lem:Psi_properties}(d) $Q$ is snc-minimal, so $Q_1+Q_2$ is snc-minimal. By (i), $\kappa(Y'_C\setminus\{z\})\geq 0$. Suppose $(\ov S,Q_1+Q_2)$ is not almost minimal. We consider the proces of obtaining the almost minimal model $\pi_{\ell}\circ\ldots\circ \pi_1\:(\ov S,Q_1+Q_2)\to (\ov Z,T)$ as in the proof of \ref{lem:Y minimal}. The divisors $Q_1$ and $Q_2$ are the connected components of the boundary of $Y'_C\setminus\{z\}$, so because $Q_2$ cannot be contained in a divisor of quotient type, $T$ has at most one connected component of quotient type and we see that the Euler characteristic of the open part does not increase in this process. Since $\chi(\ov S\setminus(Q_1+Q_2))=0$, the log BMY inequality implies that $\chi(\ov Z\setminus T)\geq 0$, so in fact the Euler characteristic of the open part does not change in the process. Since $e(\ov Y_i,T_i)$ drops, it follows that $T$ is connected and $\ell=1$, i.e.\ the process has exactly one step. Because $T$ is not of quotient type, the log BMY inequality implies also that for $\cal P=(K_{\ov Z}+T)^+$ we have $\cal P^2=0$. In particular, $Y'_C\setminus\{z\}$ is not of log general type.
Let $L\not\subset Q_1+Q_2$ be the $(-1)$-curve on $\ov S$ witnessing the non almost-minimality of $(\ov S,Q_1+Q_2)$. Since $T$ is connected and not of quotient type, we have either $L\cdot Q_2=L\cdot Q_1=1$ or $L\cdot (Q+\ti C-Q_1)=0$, $L\cdot Q_1=1$ and $L+Q_1$ contracts to a smooth point. We have $\ks\cdot (\ks+Q_1+Q_2+L)=\ks\cdot(\ks+D+E)=2-\varepsilon$, so $K_{\ov Z}\cdot (K_{\ov Z}+T)\geq 2-\varepsilon$ and $(K_{\ov Z}+T)^2\geq -\varepsilon\geq -1$. We obtain $\Bk^2 T=(K_{\ov Z}+T)^2\geq -1$. But $T$ is a non-empty rational tree, so $\Bk T\neq 0$, hence $\Bk^2 T$ is a negative integer, i.e.\ $\Bk^2 T=-1$. Then $K_{\ov Z}\cdot (K_{\ov Z}+T)=1$ and $\varepsilon=1$. From Riemann-Roch $2K_{\ov Z}+T\geq 0$. We argue as in the proof of \ref{lem:Y minimal}, case $\gamma=4$, that $T$ is not a chain and $\cal P\cdot T=0$. Hence again by Fujita's classification $T$ is a fork with a branching component $H$. However, then the intersection matrix of $T-H$ is negative definite, so because $\#(T-H)=\rho(\ov Z)$, we get a contradiction with \ref{lem:Hodge lemma}.
\end{proof}
\blem\label{lem:delta-s-inequality} Let $(\ov X,T+R)$ be a smooth almost minimal pair, such that $T$ is a (connected) rational tree of non-quotient type and $R$ is a rational chain disjoint from $T$. Let $T_1,\dots, T_s$ be the maximal twigs of $T$. Assume $2K_{\ov X}+T+R\geq 0$ and put $\cal P=(K_{\ov X}+T+R)^+$ and $\delta(T)=\sum\limits_{i=1}^s\frac{1}{d(T_i)}$. Then $2\cal P^2+\delta(T)\geq s-2$.
\elem
\begin{proof} Put $\cal P=(K_{\ov X}+T+R)^+=K_{\ov X}+T+R-\Bk T-\Bk R$. Since $\ovk(\ks +T+R)\geq 0$, $R$ and $T_i$'s are admissible, so are contained in the support of $\Bk (T+R)$. Since $\cal P$ is nef and $2K_{\ov X}+T+R$ is effective, we get $$\cal P\cdot T\leq \cal P\cdot (2K_{\ov X}+R+T)+\cal P\cdot T=2\cal P\cdot (K_{\ov X}+T+R)-\cal P\cdot R=2\cal P^2.$$ By the properties of barks (see the formula for the coefficients of components in $\Bk$ in \cite[??]{Miyan-OpenSurf}) and of the Zariski decomposition we have $\cal P\cdot R=0$ and $$\cal P\cdot T=(K_{\ov X}+T+R)\cdot T-\Bk T\cdot T=-2-\sum_{i=1}^s(\frac{1}{d(T_i)}-1)=s-2-\delta(T).$$
\end{proof}
\blem\label{lem:delta-eps-inequality} If $\varepsilon\leq 1$ or $\gamma\geq 5$ then the following inequalities hold: \benum[(i)]
\item $s-2-\frac{6}{c_h}\leq \delta(Q-Q_1+\ti C)\leq e(Q-Q_1+\ti C)\leq 1+\varepsilon -\frac{d'(Q_1)+d''(Q_1)-1}{c_h},$
\item $s-3-\varepsilon\leq-\frac{d'(Q_1)+d''(Q_1)-7}{c_h}$,
\eenum where $d'(Q_1)$, $d''(Q_1)$ denote determinants of the chain $Q_1$ with tips removed, $\delta$ and $s$ are defined as in \ref{lem:delta-s-inequality} with $T=Q+\ti C$. Analogous statements hold for the surface $Y_{\ti C}$.
\elem
\begin{proof} We have $\ks\cdot (\ks+Q+\ti C)=3-\varepsilon$ and $p_a(Q+\ti C)=-1$, so by Riemann-Roch $h^0(2\ks+Q+\ti C)\geq 2-\varepsilon$. So if $\varepsilon\leq 2$ then $2K_X+Q+\ti C\geq 0$. On the other hand, if $\gamma\geq 5$ and $\varepsilon\geq 2$ then $\varepsilon=t=2$ by \ref{prop:basic_inequality}, so by \ref{lem:k(Y)} $2K_X+Q+\ti C\geq 2K_X+Q\geq 0$. Applying \ref{lem:delta-s-inequality} to $(\ov X,T+R)=(\ov S,Q+\ti C)$ we have $s-2-2\cal P^2\leq \delta$. Since $\chi(Y_C\setminus\{z\})=0$, \ref{lem:BMY} gives $\cal P^2\leq \frac{3}{d(Q_1)}=\frac{3}{c_h}$. We have $-1-\varepsilon=(\ks+Q+\ti C)^2=\cal P^2+\Bk^2 (Q-Q_1+\ti C)+\Bk^2 Q_1=\cal P^2-e(Q-Q_1+\ti C)-e(Q_1)$ and, since $Q_1$ is an admissible chain, $e(Q_1)=\frac{1}{d(Q_1)}(d'(Q_1)+d''(Q_1)+2)$. Then $e(Q-Q_1+\ti C)\leq 1+\varepsilon-\frac{1}{c_h}(d'(Q_1)+d''(Q_1)-1)$. This gives (i) and hence (ii).
\end{proof}
\blem\label{lem:c1B-p1>=2} If $\ti j\geq 1$ then $c_1-\ti p_1\geq 2$. If $\ti j>1$ then $c_1-\ti c_1\geq 2$. \elem
\begin{proof}
After blowing twice on $\lambda$, the intersection of proper transforms of $\bar U$ and the line tangent to $\ti \lambda$ is $d-\ti c_1-\ti p_1=c_1-\ti p_1$ in case $\ti j=1$ and $d-\ti c_1-\ti c_1=c_1-\ti c_1$. Because $\bar U$ does not admit a good asymptote, the intersection is bigger than $1$.
\end{proof}
\blem\label{lem:Q0_chain_for_gamma>5} If $\gamma\geq 6$ then $Q_0$ is a chain. \elem
\begin{proof} By \ref{lem:k(Y)} $\kappa(Y)\geq 0$. By \ref{prop:basic_inequality} $\varepsilon\leq 1$. It is now more convenient to treat $\lambda$ and $\ti \lambda$ in a symmetric way, so for the needs of this proof we temporarily cancel the assumption that $j\leq \ti j$. Suppose that $Q_0$ is not a chain.
\bcl $\{c_h,\ti c_{\ti h}\}=\{2,3\}$ or $\{2,4\}$. \ecl
\begin{proof} By \ref{lem:Q0_quotient_for_gamma>=6} $Q_0$ is of quotient type hence it is a contractible fork. The local fundamental group $\Gamma(Q_0)$ of the associated singular point is small and non-abelian, hence its order is at least $8$. We have $d(Q_1)=c_h$ and $d(\ti Q_1)=\ti c_{\ti h}$. By \ref{lem:BMY} $\frac{1}{c_h}+\frac{1}{\ti c_{\ti h}}\geq 1-\frac{1}{6}-\frac{1}{8}=\frac{17}{24}$, so $\min\{c_h,\ti c_{\ti h}\}=2$ and $\max\{c_h,\ti c_{\ti h}\}\leq 4$. Suppose $c_h=\ti c_{\ti h}=2$. Then equations \eqref{eq:1} and \eqref{eq:2} imply that $4$ divides $\gamma$, hence $\gamma=8$. Then $\varepsilon=0$ and $e(D+E)\leq 1$ by \ref{lem:eps_and_gamma}(v). Since $Q_1$ and $\ti Q_1$ are $(-2)$-tips of $D+E$, $D+E$ has no more tips, so $D$ is a chain. But then $Q_0$ is a chain; a contradiction. Thus $\{c_h,\ti c_{\ti h}\}=\{2,3\}$ or $\{2,4\}$.
\end{proof}
Because $Q_0$ is a fork, say, $\ti T_1$ is a branching component of $Q_0$ and $T_1$ is not. We write $Q_0=\ti T_1+R_1+R_2+R_3$, where $R_1$ is the twig of $Q_0$ containing the image of $L'_\infty$ and $R_2$ is the twig meeting $\ti C$. We have either $h=1$ or $h=2$ and $p_2=1$.
\bcl $\ti h\leq 3$. \ecl
\begin{proof} Suppose $\ti h\geq 4$. Then $R_2$ has at least three irreducible components, among them a $\leq (-3)$-curve. Then $d(R_2)>5$, which implies that $Q_0$ is a fork of type $(2,2,n)$, hence $R_1=[2]$ and $R_3=[2]$. A determinant of a $(2,2,n)$-fork is $4(n(b-1)-d'(R_2))$, where $-b\leq -2$ is the self-intersection of the branching component and $d'(R_2)$ is the discriminant of $R_2$ with the last component removed. Since $Q_0$ is not a $(-2)$-fork, $d(Q_0)\geq 8$, hence $|\Gamma(Q_0)|\geq 16$ and the BMY inequality gives $\{c_h,\ti c_{\ti h}\}=\{2,3\}$. Since $R_1$ is irreducible, we have $h=1$ and $p_1=1$. In particular, $c_1\leq 3$. By \ref{lem:c1B-p1>=2} $\ti j=0$, because otherwise $c_1\geq \ti c_2+2\geq 4$. Then $\Psi=\id$, so since $R_1$ is irreducible, $j=0$, hence $R_1^2\leq -3$; a contradiction.
\end{proof}
\bcl $j,\ti j\geq 1$. \ecl
\begin{proof}
We have $j+\ti j =2+\varepsilon+\gamma-(h+\ti h)\geq 8-5=3$. Suppose that $\ti j=0$ or $j=0$. Then $\Psi=\id$ and since $j+\ti j\geq 3$, $R_1$ has at least four components. But $R_1$ contains a $\leq (-3)$-curve, so $d(R_1)>5$ and hence $Q_0$ is of type $(2,2,n)$. Then $R_2=[2]$, so $\ti h=2$ and $\ti p_2=2$. Then $\ti c_2\geq \ti p_2+1=3$, so $\ti c_2\in \{3,4\}$ and $c_h=2$. Because $\ti c_2$ and $\ti p_2$ are coprime, $\ti c_2=3$. We have $d(R_3)=\frac{\ti c_1}{\ti c_2}$, so since $R_3=[2]$, $\ti c_1=6$ and hence $\ti p_1=\ti c_2=3$. Since $c_h=2$ and $R_3=[2]$, $e(D+E)>1$, so $\varepsilon=1$ and hence $\gamma\leq 5+t=6$, i.e.\ $\gamma=6$. The total contribution of $E$, $Q_1$, $R_2$ and $R_3$ to $e(Q-\ti Q_1+C)$ is $\frac{1}{6}+\frac{1}{2}+\frac{1}{2}+\frac{1}{2}=\frac{5}{3}$, so since \ref{lem:delta-eps-inequality}(i) gives $e(Q-\ti Q_1+C)\leq \frac{5}{3}$, we have $h=1$. Then $(c_1,p_1)=(1,2)$, so \eqref{eq:1} gives $j+3\ti j=8$. But $j+\ti j=6$, so both $j$ and $\ti j$ are nonzero; a contradiction.
\end{proof}
\bcl $(h,\ti h)\neq (2,2)$. \ecl
\begin{proof}
Suppose $(h,\ti h)=(2,2)$. Then $j+\ti j=\varepsilon+\gamma-2$ and $t\leq 1$. By \ref{lem:j<=1} and by the previous claim $\min\{j,\ti j\}=1$.
Suppose $j=1$. Then $\ti j\geq 3+\varepsilon$. Let $A$ be the component of $D'$ produced by the pair $\binom{c_1}{c_1} $. We have at least $3+\varepsilon$ successive contractions in $\Psi$ affecting $A$. Hence $A^2\leq -4-\varepsilon$, i.e.\ we blow up at least $3+\varepsilon$ times on $A$, otherwise $A$ becomes a nonnegative curve in $D$, which is impossible by \ref{lem:Psi_properties}(d). It follows that $R_1$ begins with $1+\varepsilon$ $(-2)$-curves and contains at least $2$ other components ($\Psi$ does not contract $T_1$ and the twig produced by $\binom{c_1}{p_1}$), hence its contribution to $e(D+E)$ is bigger than $\frac{1+\varepsilon}{2+\varepsilon}$. Since $p_2=1$, $Q_1$ consists of $(-2)$-curves, hence $e(D+E)>1$. It follows from \ref{lem:eps_and_gamma}(v) that $\varepsilon=1$. If $d(R_1)>5$ then $Q_0$ is of type $(2,2,n)$, so $R_2=[2]$ and hence $\ti p_2=2$. But $\{c_h,\ti c_{\ti h}\}=\{2,3\}$ or $\{2,4\}$, so in the latter case $\ti c_2=3$ and $c_2=2$ and we get $e(D+E)>\frac{2}{3}+\frac{1}{2}+\frac{1}{2}+\frac{1}{3}=2$, which contradicts \ref{lem:eps_and_gamma}(v). Thus $d(R_1)\leq 5$, hence $R_1=[2,2,2,2]$. Also, $Q_0$ is a fork of type $(2,3,5)$ or $(2,2,n)$, so $\{d(R_2), d(R_3)\}=\{2,3\}$ or $\{2,2\}$. It follows that $e(Q-\ti Q_1+C)\geq \frac{2}{3}+\frac{1}{2}+\frac{1}{d(R_2)}+\frac{1}{d(R_3)}+\frac{1}{\gamma}>2$; a contradiction by \ref{lem:delta-eps-inequality}(i).
Thus $\ti j=1$. Then $ j\geq 3+\varepsilon$. Let $\ti A$ be the component of $D'$ produced by the pair $\binom{\ti c_1}{\ti c_1} $. By the argument above $\ti A^2\leq -4-\varepsilon$, so we blow up at least $3+\varepsilon$ times on $\ti A$. In particular, $R_3$ begins with $1+\varepsilon$ $(-2)$-curves and it has at least $1$ component more. As above, we infer that $\varepsilon=1$, $\gamma=6$ and $j\geq 4$. Since $h>1$, $R_1$ has at least $2$ components ($\Psi$ does not contract $T_1$), so it is not a $(-2)$-curve. Since $Q_0$ is of quotient type, it follows that $d(R_3)\leq 5$, hence $R_3=[2,2,2]$ or $R_3=[2,2,2,2]$. Then $Q_0$ is a fork of type $(2,3,4)$ or $(2,3,5)$, so $d(R_2)\leq 3$. It follows that $e(Q-\ti Q_1+C)\geq \frac{1}{2}+\frac{1}{2}+\frac{1}{d(R_2)}+\frac{3}{4}+\frac{1}{\gamma}>2$; a contradiction by \ref{lem:delta-eps-inequality}(i).
\end{proof}
\bcl $h=1$. \ecl
\begin{proof}
Suppose $h\neq 1$. Then, by the previous claims $h=2$ and $\ti h=3$. Suppose that $\ti c_2>\ti p_2$. Then $R_2$ has at least two components and a $\leq(-3)$-curve between them, so $d(R_2)\geq 5$. Since $\#R_1\geq 2$, $Q_0$ is a fork of type $(2,3,5)$, hence $R_2=[2,3]$, $R_1=[2,2]$ and $R_3=[2]$. It follows that $Q-\ti Q_1+C$ has at least four $(-2)$-tips, hence $e(Q-\ti Q_1+C)\geq 2$; a contradiction with \ref{lem:delta-eps-inequality}(i).
Therefore, $\ti c_2=\ti p_2$. Then $R_2$ contains a $\leq (-3)$-curve. Since $\#R_1>1$, $R_3=[2]$, so $\frac{\ti c_1}{\ti c_2}=2$. Also, denoting by $\ti A$ be the curve in $D'$ produced by the $\ti j$-th pair $\binom{\ti c_1}{\ti c_1}$ we have $\ti A^2=-3$. Since $\ti c_2=\ti p_2$, $\ti T_1'$ is a $(-2)$-curve, so because $\ti T_1^2\leq -2$, $\Psi$ does not contract $\ti A$. It follows that $\ti j\geq 2$, otherwise $j=\gamma+\varepsilon-3-\ti j\geq 2$ and $\Psi$ contracts $\ti A$. Then $j=1$. Moreover, if $\Psi$ contracts the curve $A$ produced by the pair $\binom{c_1}{c_1}$ then it also contracts $\ti A$. Thus $A$ is not contracted by $\Psi$. We get $\#R_1\geq 3$. Since $R_2$ is not a $(-2)$-curve, $Q_0$ is of type $(2,3,4)$ or $(2,3,5)$. Then $R_1=[2,2,2]$ or $[2,2,2,2]$ and $d(R_2)=3$. The latter gives $R_2=[3]$, so $\ti c_2=2$. The divisor $Q-\ti Q_1+C$ has at least three $(-2)$-tips, so $e(Q-\ti Q_1+C)>\frac{3}{2}$. But by \ref{lem:delta-eps-inequality}(i) $e(Q-\ti Q_1+C)\leq \frac{3}{2}$; a contradiction.
\end{proof}
Since $h=1$, Claim 1 gives $c_1\leq 4$. Since $\ti j\geq 1$, by \ref{lem:c1B-p1>=2} we have $c_1\geq \ti p_1+2\geq \ti c_2+2\geq 4$, so Claim 1 gives $c_1=4$ and $\ti c_{\ti h}=2$. Then $\ti p_1=\ti c_2=2$ and by \ref{lem:c1B-p1>=2} $\ti j=1$, hence $j=\gamma+\varepsilon-\ti h\geq 6-\ti h+\varepsilon$. Because $R_2\neq 0$, we have $\ti h\geq 3$. Thus $\ti h=3$ and $(\ti c_2,\ti p_2,\ti c_3,\ti p_3)=(2,2,2,1)$. Put $\ti k=\ti c_1/\ti c_2=\ti c_1/2$. Clearly, $\ti k\geq 2$. We have $e(Q-\ti Q_1+C)=\frac{4-p_1}{4}+\frac{1}{\gamma}+(1-\frac{1}{\ti k})+\frac{1}{3}$. By \ref{lem:delta-eps-inequality}(i) $e(Q-\ti Q_1+C)\leq \varepsilon+\frac{1}{2}$, so $\varepsilon=1$ and $(\frac{4-p_1}{c_1}+\frac{1}{\gamma}\leq \frac{1}{\ti k}+\frac{1}{6})$. But since $\varepsilon=1$, \ref{prop:basic_inequality} gives $\gamma\leq 5+t\leq 6$, so $\gamma=6$. Then $p_1=3$ and $\ti k\leq 4$. Also, $j=4$. Then \eqref{eq:1} gives $\ti k=8$; a contradiction.
\end{proof}
Since by \ref{prop:gamma>1} $\gamma\geq 2$, the following proposition completes the proof of \ref{thm:main_result}(2).
\bprop\label{prop:gamma<=5} $\gamma\leq 5.$ \eprop
\begin{proof} As in the proof of \ref{lem:Q0_chain_for_gamma>5} we temporarily cancel the assumption that $j\leq \ti j$. Instead we may, and shall, assume $h\leq \ti h$. Suppose that $\gamma\geq 6$. By \ref{lem:Q0_chain_for_gamma>5} $Q_0$ is a chain, so $\ti h\leq 2$. By \ref{prop:basic_inequality} $\varepsilon\leq 1$.
Suppose that $h=2$. Then $\ti h=2$ and $t=2$, so $e(D+E)>1$, hence $\varepsilon=1$ and $\gamma\leq 7$. We have $d'(Q_1)+d''(Q_1)\leq 7$ by \ref{lem:delta-eps-inequality}(ii). Hence $Q_1$, and similarly $\ti Q_1$, consist of at most three $(-2)$-curves. If, say, $Q_1=[2,2,2]$ then $\frac{d'(Q_1)+d''(Q_1)-1}{c_h}=\frac{5}{4}$ and we have a contradiction with \ref{lem:delta-eps-inequality}(i), because $e(Q-Q_1+\ti C)>(1-\frac{1}{\ti c_{\ti h}}) +\frac{1}{\gamma}\geq \frac{5}{6}$. So $Q_1$ and $\ti Q_1$ consist of at most two $(-2)$-curves. If $Q_1=[2]$ and $\ti Q_1=[2]$ then from \eqref{eq:1} and \eqref{eq:2} we get that $4$ divides $\gamma$, which is impossible. Similarly, it is not possible that $Q_1=[2,2]$ and $Q_2=[2,2]$, because otherwise $c_2=\ti c_2=3$ and then we get that $3$ divides $\gamma-2$. Thus, say, $Q_1=[2,2]$ and $\ti Q_1=[2]$. If one of the two tips of $Q_0$ is a $\geq (-3)$-curve then $e(Q-Q_1+\ti C)>\frac{1}{2}+\frac{1}{6}+\frac{1}{3}=1$, while $\frac{d'(Q_1)+d''(Q_1)-1}{d(Q_1)}=1$ and we have a contradiction with \ref{lem:delta-eps-inequality}(i).
Thus both tips of $Q_0$, call them $B$ and $\ti B$, are $\leq(-4)$-curves. Let $\alpha\: \ov S\to \ov N$ be the contraction of $C+Q_1$ and $\ti C+\ti Q_1$. Put $E_0=\alpha(E)$, $Z=\alpha(D)$. It does not touch $B+\ti B$. We have $E_0^2=-\gamma+5$, $K_{\ov N}\cdot(K_{\ov N}+Z)=\ks\cdot(\ks+D)+2=1-\ks\cdot E+2=5-\gamma$, $E_0\cdot(K_{\ov N}+Z)=E_0\cdot K_{\ov N}+5=\gamma-2$. Since $E_0\cdot Z>1$, the divisor $E_0+K_{\ov N}+Z$ is effective. Because $\kappa(K_{\ov N}+Z)=-\8$, we can find a maximal $m\in \mathbb{N}_+$ such that $|E_0+m(K_{\ov N}+Z)|\neq\emptyset$. Write $$E_0+m(K_{\ov N}+Z)=\sum A_i,$$ where $A_i$'s are irreducible. By the maximality of $m$ we have $|A_i+K_{\ov N}+Z|=\emptyset$, which gives $A_i\cong \PP^1$ and $A_i\cdot Z\leq 1$ for every $i$. We have $\rho(\ov N)>2$, because $\ov N$ contains negative curves $B$ and $\ti B$. Hence replacing successively $A_i$'s having non-negative self-intersection by singular members of their linear systems we may assume that $A_i^2<0$ for every $i$. We find $$(K_{\ov N}+Z)\cdot (E_0+B+\ti B+2K_{\ov N})=E_0\cdot(K_{\ov N}+Z)-2+2K_{\ov N}\cdot(K_{\ov N}+Z)=6-\gamma\leq 0$$ and $E_0\cdot(E_0+B+\ti B+2K_{\ov N})=\gamma-9<0$. Therefore, there exists $A_{i_0}$ for which $A_{i_0}\cdot (E_0+B+\ti B+2K_{\ov N})<0$. The intersection of $B$ (respectively $\ti B$) with $E_0+B+\ti B+2K_{\ov N}$ equals $-2+B\cdot K_{\ov N}\geq 0$ (respectively $-2+B\cdot K_{\ov N}\geq 0$), so $A_{i_0}\neq B, \ti B$. Also, $A_{i_0}\neq E_0$, because $|E_0+K_{\ov N}+Z|\neq \emptyset$. Thus $A_{i_0}$ is a $(-1)$-curve and $A_{i_0}\cdot E_0<2$. Since by the definition of $\alpha$ the curve $E_0$ meets every $(-1)$-curve in $Z$ at least twice, $A_{i_0}$ is not a component of $Z$. Then $A_{i_0}\setminus Z\subset S$ is a good asymptote of $U$; a contradiction.
Thus $h=1$. Suppose $\ti h=2$. Then $d(Q_1)=c_1$, $d(\ti Q_1)=\ti c_2$ and $\ti Q_1$ is a $(-2)$-chain. Now \ref{lem:BMY} gives $\frac{1}{c_1}+\frac{1}{\ti c_2}\geq \frac{5}{6}-\frac{1}{d(Q_0)}$. By \ref{lem:j<=1} $\min\{j,\ti j\}\leq 1$.
Suppose that $j=1$. Let $A$ be the curve in $D'$ produced by the pair $\binom{c_1}{c_1}$. By \eqref{eq:h+hB} $\ti j=\gamma+\varepsilon-2\geq 4+\varepsilon$, so $L_\8'$ is a $(-1)$-curve and we have at least $4+\varepsilon$ contractions in $\Psi$ affecting $A$. By \ref{lem:Psi_properties}(d) we get $A^2\leq -5-\varepsilon$. It follows that we blow up at least $4+\varepsilon$ times on the proper transform of $A$ in the pair $\binom{c_1}{p_1}$, so $Q_1$ has at least $3+\varepsilon$ components. But \ref{lem:delta-eps-inequality}(ii) gives $1+\frac{d'(Q_1)+d''(Q_1)-7}{d(Q_1)}\leq \varepsilon$. For $\varepsilon=0$ we get $d(Q_1)+d'(Q_1)+d''(Q_1)\leq 7$, which is impossible, because $Q_1$ has at least three components. For $\varepsilon=1$ we get $d'(Q_1)+d''(Q_1)\leq 7$, which is also impossible, because now $Q_1$ has at least four components.
Suppose $\ti j=1$. Let $\ti A$ be the curve in $D'$ produced by the pair $\binom{\ti c_1}{\ti c_1}$. We have $j=\gamma+\varepsilon-2\geq 4+\varepsilon$, so $L_\8'$ is a $(-1)$-curve and we have at least $4+\varepsilon$ contractions in $\Psi$ affecting $\ti A$. By \ref{lem:Psi_properties}(d) we get $\ti A^2\leq -5-\varepsilon$. It follows that we blow up at least $4+\varepsilon$ times on the proper transform of $\ti A$ in the pair $\binom{\ti c_1}{\ti p_1}$, so $\ti R$, the twig of $D+E$ produced by the pair $\binom{\ti c_1}{\ti p_1}$ has at least $3+\varepsilon$ components and begins with at least $2+\varepsilon$ $(-2)$-curves. Therefore, its contribution to $e(D+E)$ (and to $e(Q-Q_1+\ti C)$), is bigger than $\frac{2+\varepsilon}{3+\varepsilon}$. Since $\ti Q_1$ is a $(-2)$-chain we get $e(D+E)>1$, so $\varepsilon=1$ by \ref{lem:eps_and_gamma}(v). By \ref{prop:basic_inequality} $\gamma\leq 5+t\leq 7$. We have $\#Q_0\geq \#\ti R+2\geq 6$, hence $d(Q_0)\geq 7$. Now \ref{lem:BMY} gives $\frac{1}{c_1}+\frac{1}{\ti c_2}\geq 1-\frac{1}{\gamma}-\frac{1}{d(Q_0)}>\frac{2}{3}$, so $\min\{c_1,\ti c_2\}=2$ and $\max\{c_1,\ti c_2\}\leq 5$. But since $\ti j=1$, \ref{lem:c1B-p1>=2} gives $c_1\geq \ti p_1+2\geq \ti c_2+2\geq 4$, so $\ti c_2=2$ and $c_1\in\{4,5\}$. Denoting by $e_1$ the contribution of the twig of $Q_0$ meeting $C$ to $e(Q-Q_1+\ti C)$ we get from \ref{lem:delta-eps-inequality}(i) that $e_1+\frac{d'(Q_1)+d''(Q_1)-1}{c_1}<2-\frac{1}{2}-\frac{3}{4}-\frac{1}{7}=\frac{17}{28}<\frac{2}{3}$. Since $c_1=4,5$, the latter inequality implies that $Q_1$ is irreducible. Then the tip of $Q_0$ meeting $C$ is a $(-2)$-curve. Indeed, it is not contracted by $\Psi$, its proper transform on $\ov S'$ is a $(-2)$-curve and $Q_0$ contains no curves of non-negative self-intersection, hence the tip is not touched by $\Psi$. We obtain $e_1\geq\frac{1}{2}$, hence $\frac{1}{5}\leq\frac{1}{c_1}<\frac{2}{3}-\frac{1}{2}=\frac{1}{6}$; a contradiction.
Thus, we have $\min\{j,\ti j\}=0$. Suppose $j=0$. Then $\Psi=\id$ and $\ti j=\varepsilon +\gamma-1\geq 5$. It follows that $\# Q_0\geq 8$. The component produced by the last pair $\binom{\ti c_1}{\ti c_1}$ is a $(\leq -3)$-curve. Also, $(\ti T_1)^2\leq -3$. We find $d(Q_0)\geq 53$. From \ref{lem:BMY} we get that $c_1\leq 3$. But $c_1\geq \ti p_1+2\geq \ti c_2+2\geq 4$; a contradiction.
Now suppose that $\ti j=0$. We find that $\#Q_0\geq 8$ and (because $\ti T_1$ and the tip contained in $F$ are $\leq (-3)$-curves) that $d(Q_0)\geq 43$. If $c_1=\ti c_2=2$ then $e(D+E)>1$, so $\varepsilon\geq 1$ and hence $\gamma\leq 7$. But in the latter case \eqref{eq:1} and \eqref{eq:2} give $4|\gamma$. Thus $\max\{c_1,\ti c_2\}\geq 3$. Now \ref{lem:BMY} gives $\{c_1,\ti c_2\}=\{2,3\}$ and $\gamma=6$. Then $j=5+\varepsilon$. Write $\ti c_1=\ti k\ti c_2$ and $\ti p_1=\ti l\ti c_2$. Multiplying \eqref{eq:1} by $c_1$ and subtracting \eqref{eq:2} we get $$\ti c_2^2\ti k(\ti k-\ti l)+c_1\ti c_2\ti l=6(c_1-1)+c_1^2-c_1+\ti c_2.$$ If $(c_1,\ti c_2)=(3,2)$ we get $2\ti k(\ti k-\ti l)+3\ti l=10$, so $\ti l$ is even and hence $10\geq 2(\ti l+1)+3\ti l\geq 6+6$; a contradiction. Thus $(c_1,\ti c_2)=(2,3)$. We get $3(3\ti k(\ti k-\ti l)+2\ti l)=11$; a contradiction.
We are left with the case $h=\ti h=1$. We have now $j+\ti j=\varepsilon+\gamma\geq 6$. Suppose that $d(Q_0)>6$. Then \ref{lem:BMY} gives $\frac{1}{c_1}+\frac{1}{\ti c_1}\geq 1-\frac{1}{\gamma}-\frac{1}{d(Q_0)}>\frac{2}{3},$ so $\min\{c_1, \ti c_1\}=2$. By symmetry, we may assume that $\ti c_1=2$. Then $j=0$, because otherwise $\ti c_1\geq p_1+2$. It follows that $\#Q_0\geq 7$. Since the curve produced by the last pair $\binom{\ti c_1}{\ti c_1}$ is a $\leq (-3)$-curve, $d(Q_0)\geq 15$. Then $\frac{1}{c_1}\geq 1-\frac{1}{6}-\frac{1}{15}-\frac{1}{2}=\frac{4}{15}$, so $c_1\leq 3$. But by \ref{lem:c1B-p1>=2} we have $c_1\geq \ti c_1+2$, because $\ti j>1$; a contradiction.
Thus $d(Q_0)\leq 6.$ Then $j>0$ and $\ti j>0$, because otherwise $\#Q_0\geq j+\ti j+1\geq 7$ and $d(Q_0)\geq 8$. We may assume $c_1\geq \ti c_1$. Then \ref{lem:c1B-p1>=2} implies that $j=1$. We obtain $\ti j=\gamma+\varepsilon-1\geq 5$. Again by \ref{lem:c1B-p1>=2} $\ti c_1\geq p_1+2\geq 3$ and $c_1\geq \ti c_1+2$, so $\ti c_1\geq 3$ and $c_1\geq 5$. Then \ref{lem:BMY} gives $\ti c_1=3$ and $d(Q_0)\leq 3$. In particular, $p_1=1$. Since $C'+\ti C'$ is not touched by $\Psi$, we have $\#Q_0\geq 2$, hence $Q_0=[2,2]$. Then $\frac{1}{c_1}\geq \frac{1}{3}-\frac{1}{\gamma}$, so $c_1\leq 6$. The formulas \eqref{eq:1} and \eqref{eq:2} read as
\begin{align*}c_1+\gamma+5&=3\ti j+\ti p_1,\\
5c_1+\gamma+9&=9\ti j+3\ti p_1.
\end{align*}
We obtain $3(c_1+\gamma+5)=5c_1+\gamma+9$, so $c_1=\gamma+3\geq 9$; a contradiction.
\end{proof}
\section{\texorpdfstring{Type $(0,\ti j)$}{Type (0,j)}}
In the following three subsections we prove the following lemma.
\blem\label{prop:type_(0,0-1-2)} If $U$ is of type $(0,\ti j)$ for some $\ti j\leq 2$ then we can make a change of coordinates on $S$, after which $U$ still has separate non-simple branches at infinity, it is of type $(0,\ti m)$ for some $\ti m\in \Z$ and the degree of $U$ drops.\elem
\subsection{\texorpdfstring{Type $(0,0)$}{Type (0,0)}}Here we prove \ref{prop:type_(0,0-1-2)} in case $j=\ti j=0$.
\begin{proof}[Proof of \ref{prop:type_(0,0-1-2)}] Let $c_1=kc_2$ and $\ti c_1=\ti k\ti c_2$. Let $c_1-p_1=\beta c_2$, $\ti c_1-\ti p_1=\ti \beta \ti c_2$. Then $\beta, \ti \beta \geq 1$. We have $\gamma=\gamma'$ by \ref{lem:Psi_properties}(b), so the formulas \eqref{eq:1} and \eqref{eq:2} take the form
\begin{align}
\gamma+(\beta +k)c_2+(\ti \beta +\ti k)\ti c_2&=p_2+\dots+p_h+\ti p_2+\dots+\ti p_h,\label{eq:1_case00}\\
\gamma+\beta kc_2^2+\ti \beta \ti k\ti c_2^2+2k\ti kc_2\ti c_2&=p_2c_2+\dots+p_hc_h+\ti p_2\ti c_2+\dots+\ti p_h\ti c_h.\label{eq:2_case00}
\end{align}
We may assume that $c_2\geq \ti c_2$. We have $\gamma\leq 5$, so \ref{prop:basic_inequality} gives $h+\ti h=\gamma+\varepsilon+2\leq 9$. We have $2k\ti kc_2\ti c_2\geq 8\ti c_2^2\geq\sum\limits_{i\geq 2}\ti p_i\ti c_i$. It follows from \eqref{eq:2_case00} that \begin{equation} \beta kc_2^2<\sum\limits_{i\geq 2}p_i c_i \label{eq:3_case00}.\end{equation} We obtain $\beta kc_2^2<(h-1)c_2^2$ so $\beta k\leq h-2\leq 6$. Let $r$ be the number of pairs $\binom{c_2}{c_2}$, i.e.\ $r$ is such that $p_{r+2}$ is the first $p_i$ smaller than $c_2$. We have $$\sum\limits_{i\geq 2} p_i c_i\leq rc_2^2+c_2p_{r+2}+\sum_{i\geq r+3}c_ip_i\leq rc_2^2+c_2(c_2-c_{r+3})+\sum_{i\geq r+3}c_ip_i\leq (r+1)c_2^2-2c_{r+3}^2+(h-r-2)c_{r+3}^2,$$ hence \begin{equation}
\sum\limits_{i\geq 2} p_i c_i\leq (r+1)c_2^2+(h-r-4)c_{r+3}^2\label{eq:3p_case00}.
\end{equation}
Suppose $\beta=2$. Since $\beta k\leq h-2\leq 6$, we get $\beta=2$, $k=3$, $h=8$ and then consequently $\ti h=1$, $\gamma=5$, $\varepsilon=2$ and by \ref{prop:basic_inequality} $t=2$, hence $\ti p_1=p_8=1$. The equations \eqref{eq:3_case00} and \eqref{eq:3p_case00} give $6<r+1+\frac{4-r}{4}$, so $r\geq 6$. Because $h=8$ we get $r=6$ and $(c_h,p_h)=(c_2,1)$. Then \eqref{eq:2_case00} reads as $5+\ti c_1+c_2(6\ti c_1-1)=0$; a contradiction. If $r\leq k-2$ then \eqref{eq:3p_case00} and \eqref{eq:3_case00} give $4c_{r+3}^2\leq c_2^2<(h-r-4)c_{r+3}^2$ and hence $r+9\leq h\leq 9-\ti h\leq 8$, which is impossible.
Thus $\beta=1$ and $r\geq k-1$. Suppose $\ti c_1\geq c_2$. Using \eqref{eq:2_case00} we obtain that $$7c_2^2\geq(h+\ti h-2)c_2^2\geq \sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2}^{\ti h} \ti p_i \ti c_i>kc_2^2+2kc_2\ti c_1\geq kc_2^2+2k c_2^2=3kc_2^2,$$ hence $k=2$. We have also $\gamma+\epsilon=h+\ti h-2=7$. By \ref{prop:basic_inequality} and \ref{prop:gamma<=5} $\gamma=5$, $\varepsilon=2$ and $t=2$. From \eqref{eq:3_case00} it follows that $h\geq 2$. If $\ti h\geq 2$ then $$3kc_2^2=6c_2^2<\sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2}^{\ti h} \ti p_i \ti c_i\leq 5c_2^2+c_h+\ti c_h\leq 5c_2^2+2c_2,$$ which is impossible, because $c_2\geq 2$. Therefore, $h=8$ and $\ti h=1$. Now $\sum\limits_{i\geq 2} p_i c_i\leq 6c_2^2+c_2$, which together with \eqref{eq:2_case00} gives $2c_2^2+4c_2\ti c_1<6c_2^2+c_2$. From this we get $c_2\geq \ti c_1$. Hence $c_2=\ti c_1$. Because $h=8$ and $\ti h=1$, the formulas \eqref{eq:1_case00} and \eqref{eq:1_case00} give $\gamma-\ti p_1\equiv p_8 \mod c_8$ and $\gamma\equiv 0\mod c_8$, so $c_8$ divides $\gamma=5$ and $p_8+\ti p_1$. But $t=2$, so $p_8+\ti p_1=2$; a contradiction.
Therefore, $\ti c_1<c_2$. Recall that $r-k+1\geq 0$. After blowing up on $\lambda$ according to $\binom{c_1}{p_1}\binom{c_2}{c_2}_{k-1}$ and then successively contracting $(-1)$-curves starting from $L_\infty'$ we get new coordinates on $S$ in which the type of $U$ at infinity is $(0,r-k+1)$, the branches at infinity are separated and not simple ($c_2>1$). The degree of $U$ in these new coordinates is $k\ti c_1+c_2$ and, since $\ti c_1<c_2$, it is smaller than the original degree $c_1+\ti c_1$. Indeed, $k\ti c_1+c_2<kc_2+\ti c_1\iff (k-1)\ti c_1<(k-1)c_2$.
\end{proof}
\subsection{\texorpdfstring{Type $(0,1)$}{Type (0,1)}}We prove Proposition \ref{prop:type_(0,0-1-2)} for $(j,\ti j)=(0,1)$. Put $P=\sum\limits_{i\geq 2}p_i$ and $\ti P=\sum\limits_{i\geq 2}\ti p_i$.
\begin{proof}[Proof of \ref{prop:type_(0,0-1-2)}] The formulas \eqref{eq:1} and \eqref{eq:2} read as
\begin{align}
\gamma+2c_1+\ti c_1&=p_1+\ti p_1+P+\ti P.\label{eq:1_case01}\\
\gamma+c_1^2+2c_1\ti c_1&=p_1c_1+\ti p_1\ti c_1+\sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2}\ti p_i \ti c_i.\label{eq:2_case01}
\end{align}
Write $c_1=\ti p_1+\theta$. By \ref{lem:c1B-p1>=2} $\theta\geq 2$. We keep the notation from the previous section. We rewrite the formulas:
\begin{align}
\gamma+\beta c_2+\theta+\ti c_1&=P+\ti P,\label{eq:3_case01}\\
\gamma+\beta c_2c_1+c_1\ti c_1+\theta \ti c_1&= \sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2} \ti p_i \ti c_i.\label{eq:4_case01}
\end{align}
Suppose that $\ti c_2\geq c_2$. Then $\ti c_2(P+\ti P)\geq \sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2} \ti p_i \ti c_i$, so multiplying \eqref{eq:3_case01} by $\ti c_2$ and subtracting \eqref{eq:4_case01} we get $$\gamma\ti c_2+\beta c_2\ti c_2+\theta\ti c_2+\ti c_1\ti c_2 \geq \gamma+\beta c_2c_1+c_1\ti c_1+\theta\ti c_1,$$ and then subsequently
\begin{align*}
\gamma\ti c_2+\beta c_2\ti c_2 &>\beta c_2c_1+\ti c_1(c_1-\ti c_2)+\theta(\ti c_1-\ti c_2),\\
\gamma\ti c_2+\beta c_2\ti c_2 &>\beta c_2c_1+\ti c_1\theta+\theta(\ti k-1)\ti c_2,\\
\beta c_2(\ti c_2-c_1) &>\ti c_2(\ti k\theta-\gamma+\theta(\ti k-1)).
\end{align*}
Because $c_1>\ti p_1\geq \ti c_2$, we infer $\ti k\theta -\gamma+\theta(\ti k-1)<0$. It follows that $\theta(2\ti k-1)<\gamma\leq 5$, so $\ti k<2$; a contradiction.
Thus $c_2>\ti c_2$. We now show that \begin{equation}\beta c_2 c_1=k\beta c_2^2<\sum\limits_{i\geq 2} p_i c_i.\label{eq:5_case01} \end{equation} Suppose \eqref{eq:5_case01} fails. By \eqref{eq:4_case01} $c_1\ti c_1+\theta\ti c_1<\sum\limits_{i\geq 2}\ti p_i\ti c_i\leq \ti P\ti c_2$, so by \eqref{eq:3_case01} $\ti k c_1+\ti k\theta\leq \ti P-1\leq \gamma-1+\beta c_2+\ti k(c_2-1)+\theta$. Then $0\leq 2+3-\gamma\leq \theta(\ti k-1)+\ti k+1-\gamma\leq c_2(\beta+\ti k-\ti k k)$, so $\ti k(k-1)\leq \beta\leq k-1$. But $\ti k\geq 2$; a contradiction.
We have $h+\ti h=\gamma+\varepsilon+1\leq 8$, so $(h-1)+(\ti h-1)\leq 6$ and \eqref{eq:5_case01} gives now $(\beta+1)\beta c_2^2<\sum\limits_{i\geq 2} p_i c_i\leq 6c_2^2$, so $\beta =1$.
Suppose $\ti c_1\geq c_2$. We multiply \eqref{eq:3_case01} by $c_2$. Since $c_2>\ti c_2$ we obtain that $\gamma c_2+c_2^2+\theta c_2+c_2\ti c_1\geq \gamma +c_2c_1+c_1\ti c_1+\theta\ti c_1.$ We get $$\gamma c_2> c_2^2(k-1)+\theta(\ti c_1-c_2)+\ti c_1c_2(k-1)\geq c_2^2(k-1)+\ti c_1c_2(k-1).$$ Because $c_2\geq \ti c_2+1\geq 2$, the above inequality gives $$4\geq \gamma-1\geq (k-1)(c_2+\ti c_1)\geq 4(k-1),$$ hence $k=2$, $c_2=\ti c_1=2$ and $c_1=4.$ Then $\ti p_1=1$, which gives $\theta=c_1-\ti p_1=3$. We have also $\ti h=1$. Now \eqref{eq:3_case01} and \eqref{eq:4_case01} give $\gamma+7=P=2(h-2)+1$ and $\gamma+22=4(h-2)+2$, hence $\gamma=8, h=9$; a contradiction since $h+\ti h\leq 8$.
Thus $\ti c_1<c_2$. The inequality \eqref{eq:3p_case00} holds, so we prove as in the previous section that $r\geq k-1$ and then that we may change coordinates on $S$ so that in the new coordinates $U$ the branches at infinity are non-simple and separated, the type at infinity of $U$ is $(0,r-k+1)$ and its degree equals $k\ti c_1+c_2$. Since $\ti c_1<c_2$, we get $k\ti c_1+c_2<kc_2+\ti c_1=c_1+\ti c_1$, so the degree drops.
\end{proof}
\subsection{\texorpdfstring{Type $(0,2)$}{Type (0,2)}}Here we prove Proposition \ref{prop:type_(0,0-1-2)} in the remaining case $(j,\ti j)=(0,2)$.
\begin{proof}[Proof of \ref{prop:type_(0,0-1-2)}] The formulas \eqref{eq:1} and \eqref{eq:2} read as
\begin{align}
\gamma+2c_1&=p_1+\ti p_1+P+\ti P.\label{eq:1_case02}\\
\gamma+c_1^2+2c_1\ti c_1&=\ti c_1^2+c_1p_1+\ti c_1\ti p_1+\sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2} \ti p_i \ti c_i.\label{eq:2_case02}
\end{align}
Put $a=c_1-\ti c_1$. By \ref{lem:c1B-p1>=2} $a\geq 2$. In particular, $c_1>\ti c_1$. We rewrite the formulas as:
\begin{align}
\gamma+\beta c_2+c_1-\ti p_1&=P+\ti P.\label{eq:3_case02}\\
\gamma+\beta c_1c_2+c_1\ti c_1+\ti c_1(c_1-\ti p_1-\ti c_1)&=\sum\limits_{i\geq 2} p_i c_i+\sum\limits_{i\geq 2} \ti p_i \ti c_i.\label{eq:4_case02}
\end{align}
Suppose $c_1-\ti p_1-\ti c_1<0$. Multiplying \eqref{eq:1_case02} by $c_1$ and subtracting \eqref{eq:2_case02} we get $$\gamma(c_1-1)+c_1^2-2c_1\ti c_1=\ti p_1(c_1-\ti c_1)-\ti c_1^2+\sum\limits_{i\geq 2}p_i(c_1-c_i)+\sum\limits_{i\geq 2}\ti p_i(c_1-\ti c_i),$$ hence \begin{equation}\gamma(c_1-1)+(c_1-\ti c_1)(c_1-\ti c_1-\ti p_1)=\sum\limits_{i\geq 2}p_i(c_1-c_i)+\sum\limits_{i\geq 2}\ti p_i(c_1-\ti c_i).\label{eq:6_case02}\end{equation} Because $\frac{1}{2}c_1>\frac{1}{2}\ti c_1\geq \ti c_2$, we have $c_1-\ti c_i\geq \frac{1}{2}c_1$ for $i\geq 2$, so \eqref{eq:6_case02} gives $\gamma>\frac{1}{2}(P+\ti P)$. From \eqref{eq:3_case02} we infer that $$\beta c_2+c_1-\ti p_1=P+\ti P-\gamma\leq\gamma-1\leq 4.$$ Since $c_1-\ti p_1\geq c_1-\ti c_1+1\geq 3$, we obtain that $\beta=c_2=1$, so $h=1$. Also $\gamma=5$, $c_1-\ti p_1=3$ and $\ti c_1-\ti p_1=1$. Thus $\ti c_2=1$ and $\ti h=1$. Then $\varepsilon+\gamma=h+\ti h=2$, so $\varepsilon=0$ and $\gamma=2$. We get $\ks\cdot(\ks+D)=2-\varepsilon-\ks\cdot E=2$. The Riemann-Roch theorem gives $-\ks-D\geq 0$. Because $2\ks+D+E\geq 0$ by \ref{cor:eps=3-h0}, we have $\ks+E\geq 0$. This implies that $\ks\geq 0$; a contradiction.
Thus $c_1-\ti p_1-\ti c_1\geq 0$. If $\ti c_2\geq c_2$ then multiplying \eqref{eq:3_case02} by $\ti c_2$ and subtracting \eqref{eq:4_case02} we get \begin{equation}\gamma\ti c_2+\beta c_2\ti c_2+\ti c_2(c_1-\ti p_1)>\beta c_1c_2+c_1\ti c_1+\ti c_1(c_1-\ti p_1-\ti c_1)> (\beta c_2+c_1) \ti c_1\label{eq:5_case02},\end{equation} hence $\gamma+\beta c_2+c_1-\ti p_1>(\beta c_2+c_1)\ti k$. But then $\gamma>(\beta c_2+c_1)(\ti k-1)+\ti p_1\geq c_1+1\geq \ti c_1+3\geq 5$, which is impossible. Thus $c_2>\ti c_2$. By \eqref{eq:4_case02} we have $$\beta k c_2^2<\gamma+\beta c_1c_2+c_1\ti c_1\leq \sum\limits_{i\geq 2}p_i c_i+\sum\limits_{i\geq 2} \ti p_i \ti c_i\leq (h+\ti h-2)c_2^2=(\gamma+\varepsilon-2)c_2^2\leq 5c_2^2,$$ so $\beta=1$. We write the formulas in the following form
\begin{align*}\gamma +c_2+\ti c_1 +(c_1-\ti p_1-\ti c_1) &=P+\ti P \\
\gamma+kc_2^2+kc_2\ti c_1+\ti c_1(c_1-\ti p_1-\ti c_1)&=\sum\limits_{i\geq 2} p_ic_i+\sum\limits_{i\geq 2} \ti p_i\ti c_i.
\end{align*}
Multiplying the first one by $c_2$ and subtracting the second one we get $$\gamma(c_2-1)\geq c_2^2(k-1)+\ti c_1c_2(k-1)+(c_1-\ti p_1-\ti c_1)(\ti c_1-c_2).$$ In case $\ti c_1\geq c_2$ we get $$4\geq \gamma-1\geq (k-1)(c_2 +\ti c_1)\geq (k-1)(\ti c_2+1+\ti c_1)\geq 4(k-1),$$ so $k=2$, $c_2=2$, $\ti c_1=2$ and $\gamma=5$. But in the latter case the previous inequality gives $5\geq 4+4$.
Thus $\ti c_1<c_2$. It follows that $\ti c_2<\frac{1}{2}c_2$. The inequality \eqref{eq:3p_case00} holds, so by \eqref{eq:4_case02}
\begin{equation*}
kc_2^2<\sum \limits_{i\geq 2} p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i\leq \sum\limits_{i\geq 2} p_i c_i+(\ti h-1)\ti c_2^2\leq (r+1)c_2^2+(h-r-4)c_{r+3}^2 +(\ti h-1)\frac{c_2^2}{4}.
\end{equation*}
Note that $c_2\geq \ti c_1+1>1$, so $h\geq 2$ and hence $\ti h=\gamma+\varepsilon-h\leq 7-h\leq 5$.
We have $r\geq k-1$, otherwise the above inequality gives $$0\leq (1-\frac{\ti h-1}{4})c_2^2<(h-r-4)c_{r+3}^2\leq \frac{c_2^2}{4}(h-r-4),$$ hence $9+r<h+\ti h$, which is impossible, because $h+\ti h\leq 7$. As in the previous two subsections, we can now change coordinates on $S$ so that $U$ has non-simple, separated branches at infinity, it is of type $(0,\ti m)$ for some $\ti m\geq 0$ and its degree is $k\ti c_1+c_2$. Since $\ti c_1<c_2$, we have $k\ti c_1+c_2<kc_2+\ti c_1=c_1+\ti c_1$, so the degree drops.
\end{proof}
\subsection{\texorpdfstring{Types $(0,\ti j)$ for $\ti j\geq 3$}{Types (0,j) for j>=3}} In this section we show the following proposition. Recall that we assume the coordinates on $S=\C^2$ are chosen so that branches of $U$ at infinity are disjoint.
\bprop\label{prop:j>0} $U$ is of type $(1,\ti j)$ for some $\ti j\geq 1$. \eprop
\begin{proof}
Note that by \ref{lem:j<=1} $j\leq 1$. Suppose $j=0$. By \ref{prop:type_(0,0-1-2)} and by induction on the degree of $\bar U$ we may assume that $\ti j\geq 3$. We are going to show this is impossible. Let $A$ be a conic in $\PP^2$ which meets $\lambda$ and which follows $\ti \lambda$ during the first three blowing ups. Blow up once over $\lambda$ and four times over $\ti \lambda$. If $\ti j=3$ then the intersection of the proper transforms of $A$ and $\bar U$ at this stage is $2d-p_1-3\ti c_1-\ti p_1=2c_1-p_1-\ti c_1-\ti p_1$ and for $\ti j>3$ it is $2d-p_1-4\ti c_1=2c_1-p_1-2\ti c_1$. It follows that
\begin{equation}\label{eq:noasy_case03}
\begin{aligned}
2c_1-\ti c_1&\geq \ti p_1+p_1 &\text{if } \ti j=3,\\
2(c_1-\ti c_1)& \geq p_1 &\text{\ if } \ti j>3.
\end{aligned}
\end{equation}
Put $a=c_1-\ti c_1$. Again, by \ref{lem:c1B-p1>=2} $a\geq 2$. We keep the notation $P=\sum\limits_{i\geq 2}p_i$ and $\ti P=\sum\limits_{i\geq 2}\ti p_i$. The formulas \eqref{eq:1} and \eqref{eq:2} read as
\begin{align}
\gamma+2c_1 &=p_1+(\ti j-2)\ti c_1+\ti p_1+P+\ti P.\label{eq:1_case03}\\
\gamma+c_1^2+2c_1\ti c_1 &=c_1p_1+(\ti j-1)\ti c_1^2+\ti c_1\ti p_1+\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i.\label{eq:2_case03}
\end{align}
Put $x=2c_1-p_1-(\ti j-2)\ti c_1-\ti p_1=\beta+a-(\ti j-3)\ti c_1-\ti p_1$. We rewrite the equations in the following form
\begin{align}
\gamma+x &=P+\ti P.\label{eq:3_case03}\\
\gamma+a(c_1-p_1+\ti c_1)+x\ti c_1 &= \sum\limits_{i\geq 2}p_ic_i+\sum \limits_{i\geq 2}\ti p_i\ti c_i.\label{eq:4_case03}
\end{align}
We divide the proof into two cases.
\bca $x\geq 0$. \eca
Suppose that $\ti c_2\geq c_2$. Then multiplying \eqref{eq:3_case03} by $\ti c_2$ and subtracting \eqref{eq:4_case03} we get
$$\gamma(\ti c_2-1)\geq a(c_1-p_1+\ti c_1)+(\ti c_1-\ti c_2)x,$$ so $\gamma-1\geq a\ti k+x$. From this we obtain that $c_1-\ti c_1=a=2$, $\ti c_1=2\ti c_2$, $\gamma=5$, and $2c_1-\ti c_1-p_1-\ti p_1=x=0$.
From \eqref{eq:3_case03} we get $P+\ti P= 5$, which is possible only if $(h-1)+(\ti h-1)\leq 3$. From \eqref{eq:4_case03} we obtain that $5+2\beta c_2+4\ti c_2\leq Pc_2+\ti P\ti c_2$, hence $(2\beta+\ti P-5)c_2<(\ti P-4)\ti c_2$. If $\ti P-4\leq 0$ then $(\ti P-4)\ti c_2\leq (\ti P-4)c_2$, hence $2\beta+\ti P-5<\ti P-4$, so $2\beta<1$, which is impossible. Thus $\ti P=5$ and hence $P=0$. Because $x=0$ and $\ti j\geq 3$, we have $\beta c_2\geq \beta\geq \ti p_1-a=\ti p_1-2$, so the inequality gives $1+2\ti p_1\leq \ti c_2$; a contradiction.
Therefore, $c_2>\ti c_2$. In particular, $h\geq 2$. Multiplying \eqref{eq:3_case03} by $c_2$ and subtracting \eqref{eq:3_case03} we get $$\gamma (c_2-1)\geq a(\beta c_2+\ti c_1)+x(\ti c_1-c_2).$$ Suppose $\ti c_1\geq c_2$. Then $\gamma-1\geq a(\beta+1)$, so $\gamma=5$, $a=2$ and $\beta=1$. Then $\ti c_1=c_1-a=kc_2-2$, so the inequality gives $5c_2-5\geq 2(c_2+kc_2-2)$, hence $3>2k$; a contradiction.
Thus $c_2>\ti c_1$. We rewrite again \eqref{eq:1_case03} and \eqref{eq:2_case03}:
\begin{align}
\gamma+(k+\beta )c_2 &=(\ti j-2)\ti c_1+\ti p_1+P+\ti P.\label{eq:7n_case03}\\
\gamma +\beta kc_2^2+2kc_2\ti c_1 &=(\ti j-1)\ti c_1^2+\ti c_1\ti p_1+\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i.\label{eq:8n_case03}
\end{align}
Multiplying \eqref{eq:7n_case03} by $c_2$ and subtracting \eqref{eq:8n_case03} we get
\begin{equation*}\gamma(c_2-1)\geq ((\beta -1)(k-1)-1)c_2^2+(c_2-\ti c_1)((\ti j-2)\ti c_1+\ti p_1)+\ti c_1(c_1+a).\end{equation*}
It follows that $\beta=1$. Indeed, if $\beta\geq 2$ then the above inequality gives $\gamma c_2>\ti c_1 (c_1+a)=\ti c_1(2c_1-\ti c_1)\geq 2(2c_1-\ti c_1)>2(2c_1-c_2)\geq 6c_2$, which is impossible, because $\gamma\leq 5$. Note that $c_2\geq \ti c_1+1\geq 3$, so the inequality implies also that $c_2\geq 4$. Indeed, for $c_2=3$ we get $\ti c_1=2$, so $2\gamma+9\geq \ti c_1(2c_1-\ti c_1)\geq 2(4c_2-2)=20$, which is impossible. Put $f=c_2-\ti c_1$. We have $f\geq 1$ and $a=(k-1)c_2+f$, so we can rewrite the above inequality as
\begin{equation}c_2(\gamma+c_2)-\gamma\geq f((\ti j-2)\ti c_1+\ti p_1)+\ti c_1((2k-1)c_2+f)\label{eq:9n_case03}.\end{equation}
It follows that \begin{equation}\gamma+c_2>f(\ti j-2)\frac{\ti c_1}{c_2}+\ti c_1(2k-1).\label{eq:9f_case03}\end{equation}
Consider the case $\ti j\geq 6$. Suppose $\ti c_1\geq \frac{1}{2}c_2$. Then \eqref{eq:9f_case03} gives $2\gamma-1\geq f(\ti j-2)+c_2(2k-3)$, so since $c_2\geq 4$, we get $\gamma=5$, $k=2$ and $f=1$. In particular, $\ti c_1=c_2-1$. The inequality \eqref{eq:9n_case03} gives $c_2(5+c_2)-5\geq 4(c_2-1)+1+(c_2-1)(3c_2+1)$, hence $c_2(3-2c_2)\geq 1$; a contradiction. Thus $c_2>2\ti c_1$. We have $(h-1)+(\ti h-1)=\gamma+\varepsilon-\ti j\leq 7-\ti j\leq 1$ and $2kc_2\ti c_1\geq 4c_2\ti c_1\geq 8\ti c_1^2$. It follows from \eqref{eq:8n_case03} that $\gamma+2c_2^2+8\ti c_1^2\leq \ti j\ti c_1^2+(h-1)c_2^2+(\ti h-1)\ti c_2^2\leq \ti j\ti c_1^2+c_2^2+\ti c_1^2$, hence $c_2^2<(\ti j-7)\ti c_1^2$, so $\ti j\geq 8$. But then $h+\ti h\leq 1$; a contradiction.
Thus $\ti j\leq 5$. Suppose $\ti c_1\geq \frac{2}{3} c_2$. Then \eqref{eq:9f_case03} gives $\gamma+c_2>\frac{2}{3}f(\ti j-2)+\frac{2}{3}(2k-1)c_2,$ so $14\geq 3\gamma-1\geq 2f(\ti j-2)+(4k-5)c_2$. Since $c_2\geq 4$, we get $k=2$, $f=1$, $\ti j=3$, $\gamma=5$ and $c_2=4$. But then $\ti c_1=3$ and \eqref{eq:9f_case03} fails; a contradiction.
Thus $c_2>\frac{3}{2}\ti c_1$. Since $(h-1)+(\ti h-1)=\gamma+\varepsilon-\ti j\leq 4$, we have $\sum\limits_{i\geq 2}\ti p_i\ti c_i\leq 4\ti c_2^2\leq \ti c_1^2$. Hence $(\ti j-1)\ti c_1^2+\ti c_1\ti p_1+\sum\limits_{i\geq 2}\ti p_i\ti c_i<6\ti c_1^2\leq 4c_2\ti c_1\leq 2kc_2\ti c_1$. Then \eqref{eq:8n_case03} gives
\begin{equation}kc_2^2<\sum\limits_{i\geq 2}p_ic_i\leq (h-1)c_2^2\leq 4c_2^2,\label{eq:10n_case03}\end{equation}
so $k\leq 3$ and $h-1\geq k+1\geq 3$. If $k=3$ then $h-1=4$, $\ti h=1$, $\ti j=3$ and $\varepsilon +\gamma=7$, so $\gamma=5$, $\varepsilon=2$ and $t=2$ by \ref{prop:basic_inequality}. But in the latter case from \eqref{eq:8n_case03} we get $3c_2^2+6c_2\ti c_1<2\ti c_1^2+\ti c_1\ti p_1+3c_2^2+c_2<3\ti c_1^2+3c_2^2+c_2$, which is impossible, because $c_2>\ti c_1$. Hence $k=2$, i.e.\ $c_1=2c_2$. We claim that $c_2=p_2$ and $c_3=p_3$. Indeed, if $c_2>p_2$ then $\sum\limits_{i\geq 2}p_ic_i\leq (c_2-c_3)c_2+3c_3^2\leq c_2^2-c_3\cdot 2c_3+3c_3^2=c_2^2+c_3^2\leq \frac{5}{4}c_2^2$ and if $c_3>p_3$ then $\sum\limits_{i\geq 2}p_ic_i=p_2c_2+p_3c_3+p_4c_4\leq c_2^2+(c_2-c_4)c_2+c_4^2< 2c_2^2$, and both inequalities contradict \eqref{eq:10n_case03}. We obtain $h-1\geq 3$.
Suppose that $h-1=3$ and $\ti h-1=0$. Then $Q_0$ is a fork. Because $j=0$ and $\ti j\geq 3$, we have $L_\8'\leq -2$ and $Q_0$ is not of quotient type. $Q_0$ has only one $(-2)$-twig (produced by $\binom{p_1}{c_1}$) and since $k=2$, the twig is irreducible. Note that $\gamma=\ti j+3-\varepsilon\geq \ti j\geq 3$. By \ref{lem:Y_geometry} $Y$ is not ruled and the inequality $\frac{1}{c_4}+\frac{1}{\ti c_1}+\frac{1}{\gamma}\geq 1$ holds. Since $c_4=c_2\geq 4$, we get $\ti c_1=2$. Then $\ti p_1=1$ and \eqref{eq:7n_case03} gives $\gamma=3+p_4-c_2<3$; a contradiction.
Hence $(h-1)+(\ti h-1)=4$. Then $\gamma=5$, $\varepsilon=2$, $\ti j=3$ and $t=2$. Also, $Q_0$ is not of quotient type, so by \ref{lem:k(Y)} $\kappa(Y)\geq 0$. Then \ref{lem:Y_geometry} gives $\frac{1}{c_h}+\frac{1}{\ti c_{\ti h}}\geq \frac{4}{5}$, hence $\{c_h,\ti c_{\ti h}\}=\{2,3\}$ or $\{2,2\}$. Because $c_4=c_2\geq 4$, we have $h\geq 5$, so $h=5$ and $\ti h=1$. From \eqref{eq:7n_case03} we get $3+c_2=\ti c_1+p_4$. Since $\ti c_1\leq 3$ and $p_4\leq c_2$, we get $\ti c_1=3$ and $p_4=c_2$. Hence $p_4=c_4$, so $c_2=c_5\leq 3$; a contradiction, because $c_2\geq 4$.
\bca $x<0$. \eca
We have $2c_1\geq \ti c_1+\ti p_1+p_1$. Since $x=2c_1-p_1-\ti p_1-\ti c_1-\ti c_1(\ti j-3)$, $x<0$ gives $\ti j\geq 4$. We have $$(h-1)+(\ti h-1)=\gamma+\varepsilon-\ti j\leq 7-\ti j\leq 3.$$ We rewrite the formulas \eqref{eq:3_case03}, \eqref{eq:4_case03}:
\begin{align}
\gamma+x &=P+\ti P,\label{eq:5_case03}\\
\gamma+\beta a c_2 +a\ti c_1 &=-x\ti c_1 +\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i.\label{eq:6_case03}
\end{align}
Note that, because $x<0$, \eqref{eq:5_case03} gives $P+\ti P\leq \gamma-1\leq 4$.
Consider the case $(h-1)+(\ti h-1)=3$. Then $P+\ti P\geq 4$ hence, by \eqref{eq:5_case03}, $\gamma=5$, $x=-1$ and $P+\ti P=4$. Also $\varepsilon=2$, $t=2$ and $\ti j=4$. In particular, $p_h=\ti p_h=1$. Since $P+\ti P=4$, it cannot happen that $h-1=3$ or $\ti h-1=3$. If $h-1=2$ and $\ti h-1=1$ then $p_2=c_3=2$ and \eqref{eq:6_case03} gives $5+2c_2+2\ti c_1\leq 5+\beta a c_2+a \ti c_1=\ti c_1+2c_2+2+\ti c_2$, so $3+\ti c_1\leq \ti c_2$; which is impossible. Similarly, if $h-1=1$ and $\ti h-1=2$ then $\ti p_2=\ti c_3=2$ and then \eqref{eq:6_case03} gives $5+2c_2+2\ti c_1\leq 5+\beta a c_2+a\ti c_1=\ti c_1+c_2+2\ti c_2+2$, so $3+c_2+\ti c_1\leq 2\ti c_2\leq \ti c_1$; again a contradiction.
Consider the case $(h-1,\ti h-1)=(2,0)$. Then $\gamma+\varepsilon=\ti j+2$ and $\ti P=0$. We have $P=p_2+p_3\geq 3$. Suppose that $p_2=2$. Then $p_3=1$ and $c_3=2$, so $P=3$ and \eqref{eq:5_case03} gives $x=3-\gamma$. From \eqref{eq:6_case03} we get $\gamma+\beta a c_2+a \ti c_1= (\gamma-3)\ti c_1+2c_2+2\leq 2\ti c_1+2c_2+2$, this is a contradiction. Suppose that $p_2=3$. Then $P=4$, $p_3=1$, $c_3=3$, $x=-1$, $\gamma=5$ and \eqref{eq:6_case03} gives $5+\beta a c_2+a \ti c_1=\ti c_1+3c_2+3.$ It follows that $\beta =1$, $a=2$ and $c_2=\ti c_1+2$. We get $c_1=\ti c_1+2=c_2$; a contradiction.
Consider the case $(h-1,\ti h-1)=(0,2)$. Now $\gamma+\varepsilon=\ti j+2$ and $P=0$. As above we have $\ti p_3=1$ and $\ti p_2=2$ or $3$. Suppose that $\ti p_2=3$. Then $\ti P=4$, $\ti p_3=1$, $\ti c_3=3$, $\gamma=5$ and $x=-1$. From \eqref{eq:6_case03} we get $5+\beta a +a\ti c_1=\ti c_1+3\ti c_2+3$. From that we obtain $a=2$, $\ti k=2$ and $4+2\beta=\ti c_2$. Then $\ti c_2$ is even and divisible by $\ti c_3=3$, hence $\ti p_1\geq \ti c_2\geq 6$. But then $x=4-(\ti j-4)\ti c_1-p_1-\ti p_1\leq -2$; a contradiction. Thus $\ti p_2=2$. Then $\ti p_3=1$, $\ti c_3=2$, $\ti P=3$ and $x=3-\gamma$, so \eqref{eq:6_case03} gives $\gamma+\beta a+a\ti c_1=(\gamma-3)\ti c_1+2\ti c_2+2\leq (\gamma-2)\ti c_1^2+2$. It follows that $a=2$ and $\gamma=5$. But then the equation implies that $\gamma$ is even; a contradiction.
Consider the case $(h-1,\ti h-1)=(1,1)$. We still have $\gamma+\varepsilon=\ti j+2$. Suppose that $p_2+\ti p_2=4$. Then $\gamma=5$, $x=-1$ and $p_2, \ti p_2\leq 3$. If $a\geq 3$ then $\beta a c_2\geq p_2c_2$ and $a \ti c_1\geq \ti c_1+\ti p_2\ti c_2$, which is inconsistent with \eqref{eq:6_case03}. So $a=2$ and we have $-1=x=4-(\ti j-4)\ti c_1-p_1-\ti p_1$, i.e.\ $5=(\ti j-4)\ti c_1+p_1+\ti p_1$. We obtain $5\geq p_1+\ti p_1\geq c_2+\ti c_2\geq p_2+1+\ti p_2+1=6$; a contradiction. Suppose that $p_2+\ti p_2=3$. If $x=-1$ then \eqref{eq:6_case03} gives $\gamma +\beta a c_2+a \ti c_1=\ti c_1+p_2c_2+\ti p_2\ti c_2\leq \ti c_1+2c_2+2\ti c_2$, which is impossible. So $x=-2$ and $\gamma=5$. We have $5+\beta a c_2+a \ti c_1=2\ti c_1+p_2c_2+\ti p_2\ti c_2$, so again $a=2$. Then the equality $x=-2$ gives $6-(\ti j-4)\ti c_1=p_1+\ti p_1\geq c_2+\ti c_2\geq p_2+\ti p_2+2=5$, so $\ti j=4$, $p_1=c_2$, $\ti p_1=\ti c_2$ and $c_2+\ti c_2=6$. Now \eqref{eq:6_case03} gives $5+(2\beta +\ti p_2-p_2)c_2=6\ti p_2$. We have $(\ti p_2,p_2)\neq (1,2)$, because otherwise $5+(2\beta -1)c_2=6$, which is impossible, because $c_2>p_2=2$. So $\ti p_2=2$ and $p_2=1$ and then $(2\beta +1)c_2=7$, which is again a contradiction, because $c_2>1$.
Thus we may assume that $p_2=\ti p_2=1$. In particular, $c_2, \ti c_2\geq 2$. We have $x=2a-(\ti j-4)\ti c_1-p_1-\ti p_1$, so $2a-x\geq p_1+\ti p_1$. From \eqref{eq:6_case03} we get $\gamma+(a+x)\ti c_1+(\beta a -1)c_2=\ti c_2$, hence $a+x\leq 0$. Suppose that $a+x= 0$. Then $\ti c_2= \gamma+c_2(\beta a-1)$. We obtain that $2a-x\geq p_1+\ti p_1\geq c_2+\ti c_2= c_2+\gamma+c_2(\beta a -1)$, so $\beta a c_2\leq 2a-x-\gamma=2a-2<2a$, which gives $c_2<2$; a contradiction. We infer that $2-\gamma=x\leq -a-1\leq -3$, so $x=-3$, $\gamma=5$ and $a=2$. From \eqref{eq:6_case03} we get $5+2\beta c_2=\ti c_1+c_2+\ti c_2$. Because $x=-3$, we have $7=(\ti j-4)\ti c_1+p_1+\ti p_1$. If $\ti j>4$ then $7\geq \ti c_1+p_1+\ti p_1\geq \ti c_1+c_2+\ti c_2=5+2\beta c_2\geq 9$, which is impossible. Therefore, $\ti j=4$ and $p_1+\ti p_1=7$. From \eqref{eq:6_case03} we get $5+2\beta c_2=\ti c_1+c_2+\ti c_2=c_1-2+c_2+\ti c_2$ i.e.\ $7-(c_2+\ti c_2)=(k-2\beta)c_2$. Clearly, if $c_2$ is even then $\ti c_2$ is odd. In particular, $c_2+\ti c_2\geq 5$. Because $c_2\geq 2$ divides $7-(c_2+\ti c_2)$, we get $c_2+\ti c_2=7$ and $k=2\beta$. But $\gcd(k,\beta)=1$, so $\beta=1$ and $k=2$. We have $7-c_2=\ti c_2=\ti p_1<\ti c_1=c_1-2=2c_2-2$, so $c_2>3$. By \eqref{eq:noasy_case03} $c_2=p_1\leq 2a=4$, so $c_2=4$ and hence $\ti c_2=3$. By \ref{lem:k(Y)} $\kappa(Y)\geq 0$, so \ref{lem:Y_geometry} gives $\frac{1}{3}+\frac{1}{4}+\frac{1}{5} \geq 1-\frac{1}{|\Gamma(Q_0)|}$. We obtain $d(Q_0)\leq |\Gamma(Q_0)|\leq 4$. Since $L_\8'$ is not a $(-1)$-curve, we check easily that $d(Q_0)>4$; a contradiction.
Consider the case $(h-1,\ti h-1)=(1,0)$. We have now $c_2\geq 2$. Then \eqref{eq:5_case03} gives $x=p_2-\gamma\geq -4$ and \eqref{eq:6_case03} takes the form
\begin{equation} \gamma+\beta a c_2+a\ti c_1 =-x\ti c_1+p_2c_2.\label{eq:6p_case03} \end{equation}
From this we get
\begin{equation}\gamma+(\beta a-\gamma-x)c_2+(a +x)\ti c_1=0.\label{eq:7_case03}\end{equation} which, because $\ti c_1=c_1-a=kc_2-a$, gives
\begin{equation}c_2(\gamma+x-k(a+x)-\beta a)=\gamma-a(a+x).\label{eq:7m_case03}\end{equation}
In particular, by \eqref{eq:7_case03} we have $a\beta <\gamma+x$ or $x<-a$. Hence $a\beta<5+x\leq 4$ or $a<-x\leq 4$, so $a\leq 3$. Suppose $a=3$. If $3\beta<\gamma+x$ then $\beta=1$, $\gamma=5$ and $x=-1$, so $\gamma-a(a+x)=-1$ and $c_2|1$ by \eqref{eq:7m_case03}, which is impossible. Thus $x<-a$, i.e.\ $x=-4$. Then $\gamma=5$ and \eqref{eq:7m_case03} gives $c_2(k+1-3\beta)=8$, hence $p_1=(k-\beta)c_2=8+(2\beta-1)c_2\geq 10$. But by the definition of $x$ and $a$ we have $p_1<p_1+\ti p_1+(\ti j-4)\ti c_1=2a-x=10$; a contradiction.
\noin Therefore, $a=2$. For $x=-1$ \eqref{eq:7m_case03} gives $c_2(\gamma-1-k-2\beta)=\gamma$, so $\gamma>1+k+2\beta\geq 5$, which is impossible. For $x=-2$ it gives $c_2(\gamma-2-2\beta)=\gamma$, so $\gamma\geq 3+2\beta$ and consequently $\gamma=5$, $\beta=1$ and $c_2=5$. But in the latter case $p_1\geq c_2>4=2a$, which contradicts \eqref{eq:noasy_case03}. Suppose $x=-3$. Then \eqref{eq:7m_case03} gives $c_2(\gamma-3+k-2\beta)=\gamma+2$. If $\gamma=5$ then $c_2|7$, hence $p_1\geq c_2\geq 7$. But by the definition of $x$ we have $p_1<p_1+\ti p_1+(\ti j-4)\ti c_1=2a-x=7$; a contradiction. Thus $x=-4$. Then $\gamma=5$, $p_2=1$ and $c_2(2(k-\beta)+1)=9$. Since $k-\beta\geq 1$, it follows that $c_2=3$ and $k=\beta+1$. In particular, $p_1=c_2=3$. We have $\ti c_1=c_1-2=3k-2$. By the definition of $x$ we have $\ti p_1+(\ti j-4)(3k-2)=5$, hence $\ti j\leq 5$. By \ref{lem:k(Y)} $\kappa(Y)\geq 0$, so by \ref{lem:Y_geometry} $\frac{1}{3}+\frac{1}{3k-2}+\frac{1}{5}\geq 1-\frac{1}{|\Gamma(Q_0)|}$, hence $|d(Q_0)|\leq |\Gamma(Q_0)|\leq 4$. We check easily that this is not the case (note $L_\8'$ is a $(-3)$-curve); a contradiction.
Consider the case $(h-1,\ti h-1)=(0,1)$.
\noin We have $\gamma+x=\ti p_2$ and we can write the equation \eqref{eq:6_case03} in the form
\begin{equation} \gamma+\beta a+(a+x)\ti c_1 =\ti p_2\ti c_2.\label{eq:10m_case03} \end{equation}
Because $\beta=c_1-p_1=\ti c_1+a-p_1$, we can rewrite it as
\begin{equation} (2a+x)\ti c_1=\ti p_2 \ti c_2+a(p_1-a)-\gamma.\label{eq:11m_case03} \end{equation}
By the definition of $x$ we have
\begin{equation}p_1+\ti p_1+(\ti j-4)\ti c_1=2a-x.\label{eq:12m_case03} \end{equation}
We have $\ti p_2\ti c_2=(\gamma+x)\ti c_2\leq \frac{1}{2}(\gamma+x)\ti c_1$, so \eqref{eq:10m_case03} gives $a+x<\frac{1}{2}(\gamma+x)$, hence $a\leq \frac{1}{2}(\gamma-1-x)\leq \frac{1}{2}(\gamma+3)\leq 4$. Suppose $a=4$. Then $\gamma=5$, $x=-4$ and $\ti p_2=1$, so \eqref{eq:10m_case03} gives $5+4\beta=\ti c_2$. But \eqref{eq:12m_case03} gives $\ti c_2\leq \ti p_1<2a-x=12$, so $\beta=1$ and $\ti c_2=9$. By \ref{lem:k(Y)} $\kappa(Y)\geq 0$. We have now $c_1=\ti c_1+4=9\ti k+4\geq 22$, so \ref{lem:BMY} gives $\frac{1}{22}+\frac{1}{9}+\frac{1}{5}\geq \frac{1}{|\Gamma(Q_0)|}$, i.e.\ $|\Gamma(Q_0)|\leq 2$; a contradiction. Thus $a\leq 3$.
Note that by \eqref{eq:12m_case03} $p_1+\ti p_1\leq 2a-x\leq 10$, so $\ti p_2+1\leq \ti c_2\leq \ti p_1\leq 9$ and $p_1\leq 9$. It follows that the expression $M=\ti p_2\ti c_2+a(p_1-a)-\gamma$ is bounded. Suppose $2a+x\neq 0$. Then for each possible choice of $a, x, \gamma, p_1, \ti p_1, \ti c_2, \ti p_2$ and $\gamma$ we can compute $M$, then $\ti c_1=M/(2a+x)$ and then $c_1=\ti c_1+a$. A straightforward verification shows that there are two solutions, both with $\ti j=4$: $(c_1,p_1;\ti c_1,\ti p_1,\ti c_2,\ti p_2)=(6,5;4,2,2,1)$ and $(9,8;6,2,2,1)$. For both \eqref{eq:noasy_case03} fails; a contradiction. Thus $2a+x=0$, which gives $a=2$, $x=-4$ and hence $\gamma=5$ and $\ti p_2=1$. In particular, $Q_0$ is a chain. The equation \eqref{eq:11m_case03} gives $\ti c_2+2p_1=9$. It follows that $\ti c_2$ is odd, so $\ti c_2\geq 3$. By \ref{lem:k(Y)} $\kappa(Y)\geq 0$, so by \ref{lem:Y_geometry} $\frac{1}{c_1}+\frac{1}{\ti c_2}\geq \frac{4}{5}-\frac{1}{|\Gamma(Q_0)|}$. Because $j=0$, we have $(L_\8')^2\leq -3$, so $d(Q_0)\geq 3$. But $c_1=\ti c_1+a\geq 2\ti c_2+2\geq 8$ and we check that the above inequality fails.
We are left with the case $(h-1,\ti h-1)=(0,0)$. We have $x=-\gamma$. The equation \eqref{eq:6_case03} gives $\gamma+\beta a=(\gamma-a)\ti c_1$, hence $a<\gamma$, so $a\leq 4$ and $$(\gamma-2a)\ti c_1=\gamma+a(a-p_1).$$ The definition of $x$ gives $$p_1+\ti p_1+(\ti j-4)\ti c_1=2a+\gamma,$$ so $p_1,\ti p_1\leq 2a+\gamma-1\leq 12$. Assume $\gamma\neq 2a$. For every $a\in\{2,3,4\}$, $\gamma\in\{2,3,4,5\}$ and every $p_1,\ti p_1\leq 2a-\gamma-1$ we computed $\ti c_1=(\gamma+a(a-p_1))/(\gamma-2a)$, $c_1=\ti c_1+a$ and $\ti j=(2a-\gamma-p_1-\ti p_1)/\ti c_1$ and we checked that only three solutions satisfy \eqref{eq:noasy_case03}. These are $(c_1,p_1,\ti c_1,\ti p_1)=(9,1,7,1)$, $(7,2,5,2)$ and $(7,6,4,1)$, all with $\gamma=\ti j=5$. In all cases $\gamma+t\geq 6$, so $\kappa(Y)\geq 0$. Since $d(Q_0)>2$ we check easily that the log BMY inequality (\ref{lem:BMY}) fails; a contradiction. Assume $\gamma=2a$. Then $a=2$ and $\gamma=4$, so the equation gives $p_1=4$. Because $Q_0$ is a chain, we have $\kappa(Y)\geq 0$, so $\frac{1}{\ti c_1+2}+\frac{1}{\ti c_1}\geq 1-\frac{1}{d(Q_0)}$. By the definition of $x$ we get $\ti p_1+(\ti j-4)\ti c_1=4$. If $\ti j>4$ then $\ti j=5$, $\ti p_1=1$ and $\ti c_1=3$, so $d(Q_0)\leq 4$; a contradiction. Thus $\ti j=4$ and $\ti p_1=4$. Then $\ti c_1\geq 5$, so $d(Q_0)\leq 2$; a contradiction.
\end{proof}
\bcor $\ti j\leq 6$. \ecor
\begin{proof} By \ref{lem:Psi_properties}(ii) $\gamma'=\gamma$, so the equation \eqref{eq:h+hB} gives $\ti j+1=\gamma+\varepsilon+2-h-\ti h$. By \ref{prop:no_simple_branch} $h,\ti h\geq 1$, so we get $\ti j+1\leq \gamma+\varepsilon$. Suppose $\ti j\geq 7$. Then $\gamma+\varepsilon\geq 8$, so \ref{prop:gamma<=5} gives $\gamma=5$ and $\varepsilon\geq 3$. We get a contradiction with \ref{prop:basic_inequality}.
\end{proof}
Proposition \ref{prop:j>0} implies that for $\C^*$-embeddings $U\mono \C^2$ which do not admit a good asymptote one can choose coordinates in which the type of $U$ at infinity is $(j,\ti j)$ for some $j,\ti j>0$. But $j>0$ if and only if the line tangent to $\lambda$, which in the spirit of elementary planar geometry, should be called \emph{an asymptote} of $U$, is different than $L_\8$. An analogous remark holds for $\ti j$. In view of results in \cite{CKR-Cstar_good_asymptote} we obtain the following result, which shows that most closed $\C^*$-embeddings are hyperbola-like in suitable coordinates.
\begin{proposition}\label{asymptotes} Let $U\subset \C^2$ be a closed $\C^*$-embedding. Then we can choose coordinates on $\C^2$ with respect to which $U$ has at least one asymptote (in the sense of elementary planar geometry). If the embedding is not as in cases 6.8.1.2(b) and 6.8.1.3 of \cite{CKR-Cstar_good_asymptote}, in particular if it does not admit a good asymptote, then we can choose coordinates with respect to which $U$ has two distinct asymptotes. \end{proposition}
\section{\texorpdfstring{Type $(1,1)$}{Type (1,1)}} We keep the notation from previous sections. In particular, $(j,\ti j)$ is the type at infinity of the $\C^*$-embedding $U\mono S=\C^2$. By \ref{prop:j>0} and \ref{lem:j<=1} we have $j=1$. Here we show that $\ti j\geq 2$.
Assume $U$ is of type $(j,\ti j)=(1,1)$. The formulas \eqref{eq:1} and \eqref{eq:2} read as
\begin{align*}\gamma+c_1+\ti c_1 &=p_1+\ti p_1+P+\ti P.\\
\gamma+2c_1\ti c_1 &=c_1p_1+\ti c_1\ti p_1+\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i.
\end{align*}
By \ref{lem:c1B-p1>=2} (and its analogue for $\ti \lambda$) we have
\begin{equation}\ti c_1-p_1\geq 2 \text{\ \ and\ \ } c_1-\ti p_1\geq 2\label{eq:noasy_case11}.\end{equation}
We may assume that $c_1\geq \ti c_1$. Let $x=c_1+\ti c_1-p_1-\ti p_1$. We rewrite the formulas in the following form. \begin{align}
\gamma+x &=P+\ti P.\label{eq:1_case11}\\
\gamma+c_1x&=(c_1-\ti c_1)(c_1-\ti p_1) +\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2}\ti p_i\ti c_i.\label{eq:2_case11}
\end{align}
Multiplying \eqref{eq:1_case11} by $c_1$ and subtracting \eqref{eq:2_case11} we obtain
\begin{equation}
\gamma(c_1-1)=(c_1-\ti c_1)(c_1-\ti p_1)+\sum\limits_{i\geq 2}p_i(c_1-c_i)+\sum\limits_{i\geq 2}\ti p_i(c_1-\ti c_i).\label{eq:3_case11}
\end{equation}
Since $c_1\geq \ti c_1\geq 2\ti c_2$, we have $c_1-\ti c_i\geq \frac{c_1}{2}$ for $i\geq 2$. Also $c_1-c_i\geq \frac{c_1}{2}$ for $i\geq 2$. We get $$\gamma(c_1-1)\geq (c_1-\ti c_1)(c_1-\ti p_1)+\frac{c_1}{2}(P+\ti P),$$ hence $P+\ti P\leq 2\gamma-1$. Then \eqref{eq:1_case11} gives $x\leq \gamma-1\leq 4$. But we have $x=c_1+\ti c_1-p_1-\ti p_1=c_1-\ti p_1+\ti c_1-p_1\geq 2+2=4$ by \eqref{eq:noasy_case11}, so $x=4$, $\gamma=5$ and $P+\ti P=9$. Also $c_1-\ti p_1=2$ and $\ti c_1-p_1=2$. We have $(h-1)+(\ti h-1)=3+\varepsilon$ by \eqref{eq:h+hB}. Let $c_1-p_1=\beta c_2$, $\ti c_1-\ti p_1=\ti \beta \ti c_2$. We have
\begin{equation}x=c_1-p_1+\ti c_1-\ti p_1=\beta c_2+\ti \beta \ti c_2=4.\label{eq:(i)_case11}
\end{equation}
By \eqref{eq:2_case11}
\begin{equation}5+2c_1+2\ti c_1=\sum\limits_{i\geq 2}p_ic_i+\sum\limits_{i\geq 2} \ti p_i\ti c_i.\label{eq:(ii)_case11}
\end{equation}
If $h>1$ and $\ti h>1$ then from \eqref{eq:(i)_case11} we get $c_2=\ti c_2=2$ and $\beta =\ti \beta =1$, hence $P+\ti P=2(h-2)+1+2(\ti h-2)+1=2h+2\ti h-2$ is even. But we have already shown that $P+\ti P=9$, so $h=1$ or $\ti h=1$. Suppose that $h=1$. Then $\ti h-1=3+\varepsilon\geq 3.$ We have also $\ti P=9$. \eqref{eq:(i)_case11} gives $\beta +\ti \beta \ti c_2=4$. If $\beta =2$ then $\ti c_2=2$, which implies that $\sum\limits_{i\geq 2} \ti p_i\ti c_i$ is even; a contradiction with \eqref{eq:(ii)_case11}. Thus $\beta =1$ and then $\ti \beta =1$, $\ti c_2=3$. We get $9=\ti P=3(\ti h-2)+\ti p_h$. It implies that $3$ divides $p_h$. This is impossible, because $p_{\ti h}<\ti c_h=3$. Thus $\ti h=1$. But then we reach a contradiction the same way as for $h=1$.
\bibliographystyle{amsalpha}
|
\section{ Introduction }
In the collinear factorization {\em the factorization scale} $Q$ limits
transverse phase space of all emitted particles.
Typical practical choices of $Q$ are:
virtuality of the emiter parton at the end of the multiple emission process,
maximum transverse momentum or maximum rapidity
of all emitted partons,
$\mu_F$ of the dimensional regularization,
total energy in the hard process $\sqrt{\hat{s}}$, etc.
Redefinition of the factorization scale may involve factor
$z$ being the relative loss of the energy of the emitter:
$Q\to z^\sigma Q$, $z=x_n/x_{0}$%
\footnote{Variable $x_i$ is
the standard lightcone (Bjorken) variable of the emitter parton after
$i$-th emission, $i=1,2,3,...n$.},
$\sigma=\pm 1, \pm 2$.
Many examples can be found in the literature, for instance:
(i) change from $\mu_F$ to virtuality in the hard process coefficient function
~\cite{Altarelli:1979ub},
(ii) change from time-like to space-like ladder
in the Curci-Furmanski-Petronzio (CFP)
calculation of NLO kernels~\cite{Curci:1980uw},
(iii) change from angular- to kT-ordering
in the modelling of low $x$ structure function
by Catani-Ciafaloni-Fiorani-Marchesini (CCFM)~\cite{CCFM}.
The aim of the present study is to show that:
the redefinition of the factorization scale
$Q\to z^\sigma Q$ in the ladder
can be traded exactly for the NLO correction
to the LO evolution kernel,
$P(z)\to P(z)+\sigma(2C_F \alpha_S/\pi)\Delta(z)$.
Without loss of generality,
in the numerical exercise we shall opt for $\sigma=1$.
As already said,
the above observation was already done/exploited in the literature.
Here, the above mechanism will be demonstrated {\em numerically},
in a form which can be useful
in the construction of the Monte Carlo parton shower
with the built in NLO evolution of the showers~\cite{Jadach:2010aa}.
\section{ Simplified DGLAP evolution in the Markovian Monte Carlo form}
For our numerical exercise we shall use
simplified DGLAP evolution in the Markovian Monte Carlo form.
We consider an incoming quark which emits gluons, before it enter hard process.
Its energy distribution $D(T,x)$
is a function of the evolution time $T=\ln Q^2$.
The DGLAP evolution equation ~\cite{DGLAP} reads%
\footnote{
We are using the following shothand notation
$
\big (f(\cdot)\otimes g(\cdot)\big)(x)
\equiv \int_0^1 dzdy\; f(z)g(y) \delta(x-yz).
$
}:
\begin{equation}
\label{eq:evolution1}
\frac{\partial}{\partial T}D(T,x)=
\frac{2\alpha_s(T)}{\pi}
\int_x^1 \frac{dz}{z}P(z) D\Big(T,\frac{x}{z}\Big)=
\frac{2\alpha_s(T)}{\pi}
\Big[
P(\cdot) \otimes D(T,\cdot)
\Big](x)
\end{equation}
where $x$ is a part of initial energy
(more precisely lightcone variable)
left after the emissions of a gluon from a quark.
The running QCD coupling constant is
$\alpha_s(T)=4\pi/(2\beta_0(T-\ln\Lambda_0))$~\cite{stirling-book}
where $\beta_0$ is that of ref.~\cite{GWP}
and $\Lambda_0$ is the QCD scale parameter.
However, for the sake of simplicity we shall adopt constant $\alpha_s$
in the following numerical exercises.
The evolution kernel $P(z)$ is given by:
\begin{equation}
\label{eq:kernel}
\begin{split}
&P(z)=C_F
\bigg\{
\frac{1+z^2}{2(1-z)_+}
+\frac{3}{4}\delta(1-z)
\bigg\}
= -P^{\delta}(\epsilon)\delta(1-z)
+P^{\theta}(\epsilon,z),
\end{split}
\end{equation}
where:
\begin{equation}
\label{eq:PdelThe}
\begin{split}
&P^{\theta}(\epsilon,z)=
\frac{C_F}{2} \frac{1+z^2}{1-z} \theta(1-z-\epsilon),
\quad
\\&
P^{\delta}(\epsilon)=
\int_0^1dz \;P^{\theta}(\epsilon,z)=
C_F \left[
\ln \left(\frac{1}{\epsilon}\right)
-\frac{3}{4}
\right],
\end{split}
\end{equation}
$\epsilon\rightarrow 0$ is an infrared regulator and $C_F=\frac{3}{4}$ is the
colour-group factor. $P^{\delta}(\epsilon)$ is deliberately chosen to be
positive -- it is uniquely determined from the baryon number
conservation condition,
$
\int_0^1dz\;P(z)=0.
$
\begin{figure}
\begin{center}
\includegraphics[width=90mm]{cDelta.pdf}
\caption{The $\Delta$-function of Curci--Furmanski--Petronzio.}
\label{fig:deltaFig}
\end{center}
\end{figure}
The iteration of the above evolution equation leads to the following solution:
\begin{equation}
\begin{split}
\label{eq:iterative}
&D(T,x)=
e^{-\Phi(T,t_0)}D(t_0,x)+
\\&
+\sum_{n=1}^{\infty} \int_{t_0}^T \prod_{i=1}^n
\bigg[
dt_i \theta(t_i-t_{i-1})
\bigg]
e^{-\Phi(T,t_n)}\times
\\&
\times\prod_{j=1}^n
\bigg[
\frac{2\alpha_s(t_j)}{\pi}\; P^{\theta}(\epsilon,\cdot)
e^{-\Phi(t_j,t_{j-1})}\otimes
\bigg]
D(t_0,\cdot)(x),
\end{split}
\end{equation}
where the Sudakov form-factor $\Phi(T,t_0)$ is given by
$
\Phi(T,t_0)=
\int_{t_0}^T dt' \;
\frac{2\alpha_s(t')}{\pi}\;
P^{\delta}(\epsilon).
$
On the other hand, the exact solution of the evolution equation
for $D(T,x)$ can be obtained
with high numerical precision from the Markovian Monte Carlo program.
The probability distribution
for generating single Markovian step forward,
that is generating the next $(t,x)$ starting
from the previous $(t_0,x_0)$,
is given by:
\begin{equation}
\begin{split}
\label{eq:MarkovProb}
&p(t,x|t_0,x_0)=
\theta(t-t_0)
\frac{2\alpha_s(t)}{\pi}
P^{\theta}
\left(
\epsilon,\frac{x}{x_0}
\right)
e^{-\Phi(t,t_0)},\quad
\\&
\int_{t_0}^\infty dt
\int_0^{x_0} dx\;
p(t,x|t_0,x_0)=1.
\end{split}
\end{equation}
Our toy model Markovian Monte Carlo algorithm works as follows:
\begin{itemize}
\item
$x_0$ is generated according to $D(x_0)=3(1-x_0)^2$,\;
$\int_0^1 dx_0\; D(x_0) =1$.
\item
$t_i=\ln(Q_i)$ and $z_i=\frac{x_i}{x_{i-1}}$
are generated in a loop according to
$p(t_i,x_i|t_{i-1},x_{i-1})$ for $i=1,2,3,...$
\item
Markovian process (loop) is terminated at $i=N$,
when $t_{N+1}>T$ for the first time.
\item
The above procedure is repeated many times
and the resulting distribution of the final $x=x_N$
will be distributed according to $D(T,x)$
being the solution of the evolution equation,
see ref.~\cite{GolecBiernat:2006xw} for more details.
\end{itemize}
\subsection{ The $\Delta$-function of CFP }
In the perturbative QCD
the evolution kernel $P(z)$
is calculable order by order:
\begin{equation}
P(\alpha_s,z)=
P^{(0)}(z)
+\Big(\frac{\alpha_s}{2\pi}\Big)^1 P^{(1)}(z)
+\Big(\frac{\alpha_s}{2\pi}\Big)^2 P^{(2)}(z)+...,
\end{equation}
where $P^{(0)}(z)$,
$P^{(1)}(z)$ and $P^{(2)}(z)$ are the leading (LO),
next-to-leading (NLO) and next-to-next-to-leading order (NNLO)
approximations respectively.
LO kernels are known since DGLAP works~\cite{DGLAP},
while NLO kernels were obtained directly from the Feynman diagrams
in ref.~\cite{Curci:1980uw}.
In the same ref.~\cite{Curci:1980uw} it was noticed that
NLO corrections to the kernels for
the initial state ladder differ from the ones for final state by
$ \big(\frac{\alpha_s}{2\pi}\big) C_F^2 \Delta(z)$
where
\begin{equation}
\label{eq:CFP}
C_F^2 \Delta(z)=
\left[
P^{(0)}(\cdot) \otimes \left(\ln(\cdot) \;P^{(0)}(\cdot)\right)
\right](z)
\end{equation}
and the LO kernel $P^{(0)}(z)=P(z)$ is that of eq.~(\ref{eq:kernel}).
The above $\Delta$-function is easily calculable:
\begin{equation}
\label{eq:nsimCFP}
\begin{split}
&\Delta(z)
= \int_0^1 dx\;
\bigg\{
\frac{\theta(x>z)}{x}
\frac{1+x^2}{2(1-x)}\;
\ln(y) \frac{1+y^2}{2(1-y)} \bigg|_{y=z/x}
\\&~~~~~~~~~~~~~~~~~~~~~~~~~
-\frac{1+x^2}{2(1-x)}\;
\ln(y) \frac{1+y^2}{2(1-y)}\bigg|_{y=z}
\bigg\}
\\&~~~~~~~
=\frac{1+z^2}{2(1-z)} \ln z
\left[
\ln\frac{(1-z)^2}{z}
+\frac{3}{2}
\right]
+\frac{1+z}{8} \ln^2 z
-\frac{1-z}{4} \ln z,
\end{split}
\end{equation}
This function is visualised in {\it Figure \ref{fig:deltaFig}}.
It obeys the sum rule $ \int_0^1 dz\;\Delta(z)=0$
due to $\int_0^1 P^{(0)}(z) dz=0$.
\section{ $\Delta$-function of CFP in the framework of Markovian MC}
In the following we are going to show with the help of the Markovian Monte Carlo
that the change of the time limit from $T$ to $(T+\ln x)$
induces a NLO correction to the
evolution kernel being $C_F^2 \frac{\alpha_s}{\pi}\Delta(z)$.
In the CFP work the $\Delta$-function is generated by the factor $x^{\epsilon}$,
see eq.~(2.61) in \cite{Curci:1980uw}.
Attributing the above factor to rescaling
of the factorization scale $\mu\to \mu/x$ and
defining $T=\ln\mu$, this results in the shift
$T\to T+\ln x$.
In our algorithm this change is realized in a slightly different way:
by means of decreasing the value of the time limit $T$,
step by step, in every iteration of the loop:
after accepting a given step
(by means of checking whether $t_{new}<T$ is satisfied)
we change the value of time limit $T$
at the $i$-the step in the following way:
\begin{equation}
T\rightarrow T+\ln(z_i).
\end{equation}
On the other hand, also within the Markovian MC,
instead of decreasing the time limit $T$,
we add the NLO correction proportional to $\Delta$-function directly
to the evolution kernel.
More precisely it is done by means of correcting MC
events with the following MC weight:
\begin{equation}
\label{eq:weight}
w=\prod_i \frac {P^{(1)}(z_i)} {P^{(0)}(z_i)},
\end{equation}
where $P^{(1)}(z_i)=P^{(0)}(z_i)+\lambda\Delta(z_i)$ and $\lambda = \frac{2C_F \alpha_s}{\pi}=0.100384$. Therefore the weight can be expressed as follows:
\begin{equation}
w=\prod_{i=1}^N
\bigg[
1+ \lambda\;\Delta(z_i) \left( \frac{1+z_i^2}{2(1-z_i)} \right)^{-1}
\bigg],\quad
\end{equation}
where $N$ is a number of emissions before the time limit $T$
is reached and $\Delta(z_i)$ is that of eq.~(\ref{eq:nsimCFP}).
\begin{figure}
\begin{center}
\includegraphics[width=90mm]{mcDx_dec.pdf}
\caption{Energy distributions $D(T,x)$:
(a) the one obtained by using the LO
approximation (blue),
(b) the one obtained by decreasing the evolution time
limit (red),
(c) the one obtained by correcting the LO kernel with the
$\Delta$-function (green) and
(d) the initial energy distribution $D(x_0)$.
They were generated using the following parameters:
$T=9.21034$ and $\epsilon=10^{-4}$. The distributions (b)-(d) coincide.}
\label{fig:DdistFig}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=90mm]{mcDx_dec_log.pdf}
\caption{
The same $xD(x)$ distributions as in Fig.~\protect\ref{fig:DdistFig}
plotted as functions of $log_{10}\;x$.
The parameters and the meaning of colours are also the same.
}
\label{fig:DdistLogFig}
\end{center}
\end{figure}
\subsection{ Numerical results }
Figure \ref{fig:DdistFig} shows various solutions $D(T,x)$ of the
evolution equation.
The blue curve represents the solution
accurate up to LO.
The red curve shows the distribution for the generation with
decreased time limit,
while the green one shows the one obtained by adding the
$\Delta$-correction directly to the kernel
using eq.~(\ref{eq:weight}).
The black curve representing the
initial distribution $D(t_0,x_0)$ is also shown.
It is clearly seen that red and green curves coincide, which confirms the statement
of Curci--Furmanski--Petronzio: decreasing the time limit has the same effect as
correcting the kernel with the $\Delta$-function.
The differences between various curves
are better visible in {\it Figure \ref{fig:DdistLogFig}},
which shows the same
distributions multiplied by $x$
and plotted as a function of $log_{10}\;x$.
\begin{figure}
\begin{center}
{\includegraphics[width=90mm]{mcDx_ratial_dec.pdf}}
\caption{%
Ratios of $D(x)$ distributions:
(b) divided by (a) (red),
(c) divided (a) (green) and
(b) divided by (c) (blue).
The notation (a), (b) and (c) and parameters
are the same as in Figure~\ref{fig:DdistFig}}.
\label{fig:ratials}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
{\includegraphics[width=90mm]{mcDx_ratial_dec_log.pdf}}
\caption{Ratios of $xD(x)$ distributions as functions of $log_{10}x$.
The parameters and the meaning of colors are the same
as in Figure \ref{fig:ratials}}.
\label{fig:ratialsLog}
\end{center}
\end{figure}
In order to see even better the differences between various resulting distribution
we plot in Figures \ref{fig:ratials} and \ref{fig:ratialsLog}
the ratios of the same distributions,
once again as functions of $x$ and $log_{10}\;x$.
Now, the red curve represents the ratio of the solution obtained by
decreasing time limit and the one accurate up to the LO level.
The green curve
shows the ratio of the solution obtained by using the direct $\Delta$-correction
to the kernel and the one obtained by using the LO approximation.
Finally, the
blue curve represents the ratio of the distribution obtained by decreasing the
time limit and the one with the direct $\Delta$-correction to the kernel.
It is seen that the last ratio is close to one.
Once more it indicates clearly our basic result that shifting
the evolution time limit by $\ln z$ (factorization scale by factor $z$)
gives the same result as using the direct
$\Delta$-correction to the kernel
in the way described by Curci--Furmanski--Petronzio~\cite{Curci:1980uw}.
We have checked that the slight
systematic difference between red and green curve
in Figures \ref{fig:ratials} and \ref{fig:ratialsLog}
for small $x$ values results
from the fact that in the MC implementation shortening $T\to T +\ln(1/z_i)$
below the initial $t=0$ cannot be realized%
\footnote{ In the numerical exercise with $T\to T -ln(1/z_i)$
and $\Delta \to -\Delta$ this discrepancy gets reduced.}.
Also, one has to keep in mind, that such a shortening evolution time limit
induces not only ${\cal O}(\alpha_s)$ contribution to the evolution kernel,
but also ${\cal O}(\alpha_s^2)$ term, which is not taken into account
in the present study.
Due to smallness of $\alpha_s$ the corresponding effect seems to be negligible.
Let us finally mention that all
plots and histograms presented in this sections have been obtained
using Monte Carlo software environment MCdevelop~\cite{Slawinska:2010jn}
and ROOT~\cite{Brun:1997pa} package.
\section{Summary}
The most important result presented here
is checking the equivalence of two methods of
implementing the $\Delta$-function of CFP in the Monte Carlo environment.
In the first method the evolution time range
was made shorter, step by step, after each iteration.
In the second method, the evolution time limit was kept fixed, but
the $\Delta$-function was added directly to the LO evolution kernel
as NLO correction,
by means of correcting generated events
with the help of a relevant MC weight.
Both methods have given the same results, within
the statistical error of the MC computations.
The small systematic difference between the results of both methods
is the region of small $x$ values is well understood.
\providecommand{\href}[2]{#2}\begingroup |
\section{Introduction}
\subsection{Context and motivation}
Security protocols are often found to be flawed after their
deployment, which typically requires ``dismissing'' the protocol
and hurrying up with the deployment of a new version hoping to be
faster than those attempting to exploit the discovered flaw. We
present an approach that aims at the neutralization or mitigation
of the attacks to flawed protocols: it avoids the complete
dismissal of the interested protocol and gives honest agents the
chance to continue to use it until a corrected version is released.
The standard attacker model adopted in security protocol analysis
is the one of~\cite{DY:1983}: the \emph{Dolev-Yao (DY) attacker}
can compose, send and intercept messages at will, but, following
the perfect cryptography assumption, he cannot break cryptography.
The DY attacker is thus in complete control of the network --- in
fact, he is often formalized as being the network itself --- and,
with respect to network abilities, he is actually stronger than any
attacker that can be implemented in real-life situations. Hence, if
a protocol is proved to be secure under the DY attacker, it will
also withstand attacks carried out by less powerful attackers;
aside from deviations from the specification (and the consequent
possible novel flaws) introduced in the implementation phase, the
protocol can thus be safely employed in real-life networks, at
least in principle.
A number of tools have been proposed for automated security
protocol analysis (e.g.,
\cite{AVANTSSAR,proverif,Cr2008Scyther,EscobarMM07,Ryan00,avispa02}
to name just a few), all of which follow the classical approach for
security protocol analysis in which there is a finite number of
honest agents and only one DY dishonest agent, given the implicit
assumption that in order to find attacks we can reduce $n$
collaborative DY attackers to $1$ (for a proof of this assumption
see, e.g., \cite{Basin:2011:DTL:1994484.1995278}).
In this paper, we take a quite different approach: we exploit the
fact that if in the network there are \emph{multiple
non-collaborative attackers}, then the interactions between them
make it impossible to reduce their attack ``power'' to that of a
single attacker. This paper is based on the network suitable for
the study of non-collaborative scenarios defined in our previous
works~\cite{FPV:02SECRYPTbook,FPV:SECOTS}, in which we introduced a
protocol-independent model for non-collaboration for the analysis
of security protocols (inspired by the exploratory
works~\cite{Bella03,retaliation} for ``protocol life after
attacks'' and attack retaliation). In this model: (i) a protocol is
run in the presence of multiple attackers, and (ii) attackers
potentially have different capabilities, different knowledge and do
not collaborate but rather may interfere with each other.
Interference between attackers has spawned the definition of an ad
hoc attacker, called \emph{guardian}, as a defense mechanism for
flawed protocols: if two non-collaborative attackers can interfere
with each other, then we can exploit this interference to
neutralize or at least mitigate an ongoing attack (a detailed
cost-effective analysis of this approach is left for future
work).\footnote{It is interesting to note how this idea of ``living
with flaws'' is becoming more and more widespread; see, e.g.,
\cite{mithys} where runtime monitors are employed to warn users of
android applications about ``man in the middle'' attacks on flawed
implementations of SSL. Our approach is also related to
signature-based intrusion detection systems, but we leave the
detailed study of the relations of our approach with runtime
monitors and signature-based intrusion detection systems for future
work.}
There is one fundamental catch, though. We know that a DY attacker
actually cannot exist (e.g., how could he control the whole
network?) but postulating his existence allows us to consider the
worst case analysis so that if we can prove a protocol secure under
such an attacker, then we are guaranteed that the protocol will be
secure also in the presence of weaker, more realistic attackers. A
guardian, however, only makes sense if it really exists, i.e., if
it is implemented to defend flawed protocols for real, but the
attackers and the guardian presented
in~\cite{FPV:02SECRYPTbook,FPV:SECOTS} are modeled in order to
discover interactions between agents in non-collaborative scenarios
rather than pushing for an implementation in the real-world.
\subsection{Contributions}
Since implementing a guardian with the full power of a DY attacker
is impossible, we must investigate ways to make the guardian more
feasible. In order to reduce the complexity of the possible
implementation of such a defense mechanism, in this paper we relax
the notion of guardian and ask him to defend only a subset of the
communication channels of the network, which we put under his
control.
Furthermore, not being obviously able to know where the competitor
is, we investigate where we have to introduce this defense
mechanism in the network from a topological perspective, i.e., how
the guardian can dominate his competitor(s).\footnote{In the
following, we focus on one competitor (i.e., one attacker), but it
is quite straightforwardly possible to extend our work to multiple
competitors.} Modeling the network as a graph, we study how the
topological position of an attacker $E$ and a guardian $G$, with
respect to each other and to honest agents of the protocol, can
influence a protocol attack and, thus, the possible defense against
it. We define six basic topological configurations and study the
outcome of the introduction of a guardian in each specific
position. We also introduce the concept of \emph{topological
advantage}, which guarantees that the guardian has an advantage
with respect to his competitors, and can thus carry out inference
on messages in transit in order to detect an ongoing attack and
eventually mitigate or neutralize it.
The contributions of this paper thus extend, and in a sense are
complementary to, the ones in our previous
works~\cite{FPV:02SECRYPTbook,FPV:SECOTS}. In a nutshell: there we
discussed the \emph{how} we can defend flawed security protocols
and here we discuss the \emph{where}. More specifically, as we will
describe in the following sections,
in~\cite{FPV:02SECRYPTbook,FPV:SECOTS}, we put the basis for the
study of the interaction of two attackers in non-collaborative
scenarios with the goal of understanding and finding the types of
interference the guardian can use, and, in this paper, we give the
means to understand how to exploit the interference from a
topological point of view, thus bringing the guardian close to real
implementation.
\subsection{Organization}
We proceed as follows. In Section~\ref{sec:interference}, we
summarize the main notions of attack interference in
non-collaborative scenarios. In Section~\ref{sec:modeling}, we
formalize the models of the network and of the guardian, with
particular emphasis on the topological advantage that a guardian
must have in order to defend against attacks.
\begin{SPW14}
In Section~\ref{sec:case-studies}, we discuss, as a detailed
proof-of-concept, how we can defend the ISO-SC 27 protocol and
summarize the results we obtained for other case studies.
\end{SPW14}
\begin{LONG}
In Section~\ref{sec:case-studies}, we discuss, as a detailed
proof-of-concept, how we can defend the ISO-SC 27 protocol and
summarize the results we obtained for other case studies, which are
described in more detail in the appendix.
\end{LONG}
In Section~\ref{sec:conclusions}, we briefly summarize our results and discuss future work.
\section{Attack interference in non-collaborative networks}
\label{sec:interference}
\subsection{Network agents}
Let $\mathit{Agents}$ be the set of all the network agents, which comprises
of two disjoint subsets: \begin{itemize} \item the subset $\mathit{Honest}$
of \emph{honest agents} who always follow the steps of the security
protocol they are executing in the hope of achieving the properties
for which the protocol has been designed (such as authentication
and secrecy), and \item the subset $\mathit{Dishonest}$ of \emph{dishonest
agents} (a.k.a. \emph{attackers}) who may eventually not follow the
protocol to attack some (or all) security properties. In addition
to being able to act as legitimate agents of the network, dishonest
agents typically have far more capabilities than honest agents and
follow the model of Dolev-Yao~\cite{DY:1983} that we summarized in
the introduction. \end{itemize}
The \emph{knowledge} of an honest agent $X$ is characterized by a
proprietary dataset $D_X$, which contains all the information that
$X$ acquired during the protocol execution, and is closed under all
cryptographic operations on message terms (e.g., an agent can
decrypt an encrypted message that he knows provided that he knows
also the corresponding decryption key). $D_X$ is monotonic since an
agent does not forget.
\subsection{DY attackers and the network in a non-collaborative scenario}
\begin{table}[!t]
\caption[Attacker model for non-collaboration]{Dolev-Yao attacker model for non-collaborative scenarios: internal operations (synthesis and analysis of messages), network operations (\mathii{spy}, \mathii{inject}, \mathii{erase}) and system configuration ($\mathit{True\text{-}Sender \text{-}ID}${}, \mathii{DecisionalProcess}, $\mathit{NetHandler}${}). $\mathit{NetHandler}${} describes the set of attackers who are allowed to spy by applying one of the \mathii{spy} rules. We omit the usual rules for conjunction.
\label{OurDYmodel}
}
\begin{center}
\scalebox{0.82}{
%
\begin{tabular}{c}
\begin{tabular}{cccc}
\hline
\hline \\
\mathbi{Composition:} &
\mathbi{Encryption: \phantom{\quad}} &
\mathbi{Projection:\phantom{\quad}} &
\mathbi{Decryption:}
\\\\
$\infer[]{(m_{1}, m_{2})\in D^{i}_{E}}{m_{1}\in D^{i}_{E} \quad m_{2}\in D^{i}_{E}}$\quad ~
%
&
$\infer[]{\{m\}_{k}\in D^{i}_{E}}{m\in D^{i}_{E} \quad k\in D^{i}_{E}}$
\quad ~
&
$\infer[]{m_{j}\in D^{i}_{E} \:\: \text{for~} j \in \{1,2\}}{(m_{1}, m_{2})\in D^{i}_{E}}$ \quad ~
&
$\infer[]{m \in D^{i}_{E}}{\{m\}_{k}\in D^{i}_{E} \quad k^{-1} \in D^{i}_{E}}$
\end{tabular}~
\\\\
\hline
\hline \\
\begin{tabular}{cc}
\mathbi{Inflow-Spy}:
&
\mathbi{Outflow-Spy}:
\\~\\
%
$\infer[]{m \in D^{i+1}_{E} \land sender(<X,m,Y>) \in D^{i+1}_{E} }{\mu \in D_{net}^{i} \quad $ \ofinterest{X}$ \quad Y \in D^{i}_{E} \quad \psi} $
\qquad ~
&
$\infer[]{m \in D^{i+1}_{E} \land Y \in D^{i+1}_{E}}{\mu \in D_{net}^{i} \quad {sender}(\mu) \in D^{i}_{E} \quad $\ofinterest{Y}$ \quad \psi }$
\end{tabular}
\\\\
{where $\mu = <X, m, Y>$ \quad and \quad $\psi = E \in \canseeM{<\!X,m,Y\!>, i}$ }\\\\
%
\hline
\\
\begin{tabular}{cc}
\mathbi{Injection}:
&
\mathbi{Erase}:
\\~\\
%
$\infer[]{<E(X), m, Y> \in D_{net}^{i+1}}{m \in D^{i}_{E} \quad X \in D^{i}_{E} \quad Y \in D^{i}_{E}}$
\qquad ~
%
&
%
$\infer[]{<X, m, Y> \notin D_{net}^{i+1}}{<X,m,Y> \in D_{net}^{i} \quad {sender}(<X,m,Y>) \in D^{i}_{E}}$
\end{tabular}
%
\\\\
%
\hline \hline
\\
\mathbi{True-sender-ID}:
\\~\\
$
{sender}(<X, m, Y>)= \begin{cases}
E & \text{~if~ there exists $Z$ such that~} X = E(Z) \\
X & \text{~otherwise}
\end{cases}
$
\\~\\
\mathbi{DecisionalProcess}:
\\~\\
%
\ofinterest{X}
$
=
\begin{cases}
true & \text{if $E$ decides to pay attention to $X$} \\
false & \text{~otherwise}
\end{cases}
$
\\\\
%
\hline
\\
\mathbi{NetHandler}:
\\~\\
$
\canseeM{<\!X,m,Y\!>,i} \: = \
\{Z \in \mathit{Dishonest} \:|\: \text{ $Z$ can spy
<X,m,Y> \text{~on $D_{net}^{i}$} \}
$
\\\\
\hline
\hline
%
\end{tabular}
}
\end{center}
\end{table}
In this paper, we take the non-classical approach that leverages on the
fact that the interactions between multiple non-collaborative attackers
may lead to interference. We base our work on the network suitable for
the study of non-collaborative scenarios defined
in~\cite{FPV:02SECRYPTbook,FPV:SECOTS}, which we now summarize quickly
pointing to these two papers for more details.
Table~\ref{OurDYmodel} shows the model that we adopt to formalize a
DY attacker $E$ in a non-collaborative scenario in which different
attacks may interfere with each other (we restrict the study of
this type of interaction to two active attackers but it can be
generalized to multiple ones). The knowledge base of $E$ is encoded
in the set $D_{E}$, whereas $D_{net}$ is the proprietary dataset
for the network (we will return to the network model below). The
rules in the table describe the operations that an attacker can
perform internally, how he can interact with the network and how
the system (i.e., the network environment) is configured. It is
important to note that the rules in Table~\ref{OurDYmodel} are
transition rules rather than deduction rules, i.e., they describe
knowledge acquisition from a given operation and a particular
configuration rather than the reasoning about ``only'' the
knowledge of the attacker.
As in the classic DY case, an attacker in this model can
\emph{send} and \emph{receive} messages, derive new messages by
\emph{composing, decomposing, modifying, encrypting/decrypting}
known messages (iff he has the right keys), and \emph{intercept} or
\emph{remove} messages from the network. An attacker $E$ may also
masquerade as (i.e., impersonate) another agent $X$, which we
denote by writing $E(X)$.
The most significant features of the attacker abilities are the two
\emph{spy} rules, which formalize the fact that attackers only pay
attention to a selection of the traffic on the network (considering only selected target agents):\footnote{If an attacker were omniscient and omnipotent (i.e., if he were to control the whole network) then there'd actually be no ``space'' for another attacker, and thus there'd be no interference. The more ``adventurous'' reader may want to compare this with the proof of the uniqueness of God by Leibniz, which was based on the arguments started by Anselm of Canterbury and was later further refined by G\"odel.}
\footnote{In this paper
we only use the \emph{inflow-spy} and the \emph{outflow-spy} filters
and not the \emph{restricted-spy} filter used in the previous exploratory
works. This is due to the fact that we can certainly know who we want to
defend, but we cannot know who the attackers are and we want to have
the possibility of intercepting \emph{all} outgoing/incoming messages
which leave/come from/to an agent $X$.}
\begin{itemize}
\item {\small $\mathbi{Inflow-Spy}$}: the attacker pays attention to the
incoming network traffic of a target agent and saves the identifiers of
the sender agents,
\item {\small $\mathbi{Outflow-Spy}$}: the attacker pays attention to
the traffic generated by a target agent.
\end{itemize}
The \emph{target agent} $X$ of the two spy-rules is defined through a
decisional process (the function \ofinterest{X} in
Table~\ref{OurDYmodel}) in which each attacker decides if the traffic
to/from the agent $X$ is worth to be followed. This decision is made at
run-time when a new agent identifier is discovered over the network
(i.e., when a new agent starts sending messages on the channel monitored
by the attacker). In this paper, we do not go into the details of how his decision is actually taken, but different strategies might be devised and we will investigate them in future work.
The network net is also formalized through a dataset, $D_{net}$, which
is changed by the \emph{actions} send, receive, inject and erase a
message. We write $D_{net}^i$ to denote the state of $D_{net}$ after the
$i$-th action. Messages transit on the network in the form of triplets
of the type
\begin{displaymath}
\textit{$\langle \mathit{sender\textrm{-}ID, message, receiver\textrm{-}ID}\rangle${}},
\end{displaymath}
where, as in the classical approaches, both the
attackers and the agents acquire knowledge only from the body of
messages, i.e., $\mathit{sender\textrm{-}ID}$ and
$\mathit{receiver\textrm{-}ID}$ are actually hidden to them and only
used by the network system. As a consequence of message delivery or
deletion, $D_{net}$ is non-monotonic by construction.
In order to regulate the concurrent actions over the network, the model
comprises a $\mathit{NetHandler}${} whose task is to handle the network by
selecting the next action and implementing the dependencies between
selected actions and knowledge available to each attacker. That is,
$\mathit{NetHandler}${}: (i) notifies agents that the state of the network has
changed with newly-inserted messages, (ii) polls agents for their next
intended action, (iii) selects from the set of candidate actions the one
that will be actually carried out, and (iv) informs agents of whether
the computation they performed to propose an action is a consequence of
a message that they did not have access to (i.e., for these agents a
rollback might occur in which all knowledge gained since the last
confirmed action is deleted from the dataset, and internal operations
that have occurred are cancelled).
The outcome of the process governed by the network handler is described
through the function $\mathit{canSee}${}, which returns a subset of dishonest
agents, highlighting the identifier of attackers who can spy ``before''
the message is erased from $D_{net}$. In other words, when a message is
deleted from the network, the network handler, through the function
$\mathit{canSee}${}, can decide if an attacker has spied (and saved in his
dataset) the message or not.
In our previous work we had the possibility of spying a message before its
deletion (in this case, the attacker has to decide if the message has been
received by the honest agent or deleted by another attacker) but in this paper
we relax this assumption and decide that when a message is spied it remains in
the dataset of the attacker.
The function $\mathit{canSee}${} is a configurable parameter of
our network and it corresponds to configuring a particular network
environment in which the agents are immersed: $\mathit{canSee}${} is instantiated
by the security analyst at the beginning of the analysis in order to model
time-dependent accessibility, strategic decision-making and information-sharing,
or to capture a particular network topology (in our framework the function $\mathit{canSee}${} is necessary in order to model the topologies that we will introduce in Section~\ref{sec:networkTA}).
\subsection{Attack interference (in the case of the ISO-SC 27 protocol)}
\begin{table}[t!]
\caption{The ISO-SC 27 protocol and a parallel session attack against it.\label{tab:isoSC27}}
\begin{center}
\scalebox{1}{
\begin{tabular}{cc}
%
%
\begin{tabular}{c} \hline
{\bf ISO-SC 27 protocol} \\
\hline
\\
$
\begin{array}{rll}
(1) & A \to B & : N_A \\
(2) & B \to A & : \{\!| N_A, N_B |\!\}_{K_{AB}}\\
(3) & A \to B & : N_B \\
\end{array}
$
\\ \\
\hline
\end{tabular}
&
$~ \qquad \qquad ~$
\begin{tabular}{c}
\hline
{\bf Attack}\\
\hline
\\
$
\begin{array}{rll}
(1.1) & A \to E(B) & : N_A \\
(2.1) & E(B) \to A & : N_A \\
(2.2) & A \to E(B) & : \{\!| N_A, N'_A |\!\}_{K_{AB}}\\
(1.2) & E(B) \to A & : \{\!| N_A, N'_A |\!\}_{K_{AB}}\\
(1.3) & A \to E(B) & : N'_A \\
(2.3) & E(B) \to A & : N'_A \\
\end{array}
$
\\\\
\hline
\end{tabular}
\end{tabular}
}
\end{center}
\end{table}
As a concrete, albeit simple, example of security protocol,
Table~\ref{tab:isoSC27} shows the ISO-SC 27 protocol \cite{ISOSC27}, which
aims to achieve entity authentication (aliveness) between two honest agents
$A$ and $B$, by exchanging nonces, under the assumption that they already
share a symmetric key $K_{AB}$.
Since in the second message there
is nothing that assures that the message actually comes from $B$, the
protocol is subject to a parallel sessions attack (also shown in the table) in which the attacker
$E$, who does not know $K_{AB}$, uses $A$ as oracle against herself in
order to provide to her a response that he cannot generate by himself:
$E$ masquerades as $B$ intercepting $A$'s first message and sending it
back to her in a parallel session (messages (1.1) and (2.1)). When $A$
receives the first message of the protocol from $E$, she thinks someone
wants to talk with her in another instance of the protocol (she does not
control the nonce), thus she replies to $E$ generating another nonce $N_A'$ and encrypting it
together with $N_A$ (message (2.2)).
Now $E$ has got everything he needs in order to complete the attack to the
protocol (messages (1.2)). The last message is not mandatory as
the session has already been attacked, thus $E$ can omit it (message
(2.3)). At the end of the protocol runs, $A$ is fooled into believing
that $E(B)$ is $B$.
If a protocol is flawed, a single DY attacker will succeed with
certainty. However, if attacks to the same protocol are carried out in a
more complex network environment, then success is not guaranteed since
multiple non-collaborative attackers may interact, and actually
interfere, with each other. The results
of~\cite{FPV:02SECRYPTbook,FPV:SECOTS} show that it is possible, at
least theoretically, to exploit interference between two
non-collaborative attackers to mitigate protocol flaws, thus providing a
form of defense to flawed protocols.
In the case of ISO-SC 27 protocol, which was not studied
in~\cite{FPV:02SECRYPTbook,FPV:SECOTS}\footnote{In~\cite{FPV:02SECRYPTbook,FPV:SECOTS}, we analyzed two protocols: (i) a
key transport protocol described as an example in~\cite{boydMathuria},
which we thus called the Boyd-Mathuria Example (BME), and (ii) the
Shamir-Rivest-Adleman Three-Pass protocol
(SRA3P~\cite{Clark97:asurvey}), which has been proposed to transmit data
securely on insecure channels, bypassing the difficulties connected to
the absence of prior agreements between the agents $A$ and $B$ to
establish a shared key.
},
we can identify six cases for
the possible interaction between two non-collaborative attackers $E_1$
and $E_2$:
\begin{enumerate}\itemsep0em
\item $E_{1}$ and $E_{2}$ know each other as honest.
\item $E_{1}$ and $E_{2}$ know each other as attackers.
\item $E_{1}$ and $E_{2}$ are unaware of each other.
\item $E_{2}$ knows $E_{1}$ as honest.
\item $E_{2}$ knows $E_{1}$ as dishonest.
\item $E_{2}$ knows $E_{1}$, but he is unsure of $E_{1}$'s honesty.
\end{enumerate}
The traces corresponding to the interactions of $E_1$ and $E_2$
attacking the protocol are shown in Table~\ref{tracesBME}. Attack
traces of this type lead to three possible (mutually exclusive)
situations: (i) $E_1$ dominates $E_2$ (i.e., $E_1$'s attack
succeeds while $E_2$'s fails), or (ii) none of their attacks has
success, or (iii) both achieve a situation of uncertainty, i.e.,
they do not know if their attacks have been successful or not.
In order to exploit the interference generated by multiple dishonest
agents attacking the same protocol, we can construct an additional, but
this time non-malicious, attacker: the \emph{guardian} $G$.
To define the guardian as a network agent, we refine the previous
definition of $\mathit{Agents}$ to consider the subset of \emph{benign dishonest
agents}, i.e., $\mathit{BenignDishonest} \subseteq \mathit{Dishonest} \subseteq \mathit{Agents}$, where $X \in
\mathit{BenignDishonest}$ means that $X$ has attacker capabilities and may not follow the
protocol but he ``attacks'' with the goal of ``defending'' the security
properties not of attacking them.
In other words:
\begin{definition}[Guardian]
A \emph{guardian} is a benign dishonest agent of the network,
transparent to the other agents, whose main task is to establish a
partial (or total) defense mechanism in order to mitigate (or
neutralize) protocol attacks at execution time by means of
attack-interference in non-collaborative scenarios.
$G$ is transparent to honest agents during their execution and becomes ``visible'' only in the case he has to report an ongoing attack.
\end{definition}
\begin{table}[t!]
\caption{Traces for non-collaborative attacks against the ISO-SC 27. Traces are exhaustive: $E_1$ and $E_2$ have priority over honest agents. Arrows: relative order between $(2.1')$ and $(2.1'')$ is irrelevant in determining the outcome.}
\label{tracesBME}
\begin{center}
\scalebox{1}{
\begin{tabular}{cc}
\hline
\footnotesize {\bf T1: cases 1, 3, 4}
&
\footnotesize {\bf T2: cases 5}\\
\hline \\
\footnotesize
$
\begin{array}{rll}
(1.1) & A \to E_{1,2}(B) & : N_A \\
(2.1) & E_{1,2}(B) \to A & : N_A \\
(2.2) & A \to E_{1,2}(B) & : \{\!| N_{A}, N'_A |\!\}_{K_{AB}} \\
(1.2) & E_{1,2}(B) \to A & : \{\!| N_{A}, N'_A |\!\}_{K_{AB}} \\
(1.3) & A \to E_{1,2}(B) & : N'_{A} \\
(2.3) & E_{1,2}(B) \to A & : N'_{A}
\end{array}
$
& \qquad
\footnotesize
$
\begin{array}{rll}
(1.1\phantom{''}) & A \to E_{1,2}(B) & : N_A \\
\downarrow(2.1'\phantom{'}) & E_{1}(B) \to E_{2}(A) & : N_A \\
\uparrow(2.1'') & E_{2}(B) \to A & : N_A \\
(2.2\phantom{''}) & A \to E_{2}(B) & : \{\!| N_{A}, N'_A |\!\}_{K_{AB}} \\
(1.2\phantom{''}) & E_{2}(B) \to A & : \{\!| N_{A}, N'_A |\!\}_{K_{AB}} \\
(1.3\phantom{''}) & A \to E_{2}(B) & : N'_{A} \\
(2.3\phantom{''}) & E_{2}(B) \to A & : N'_{A}
\end{array}
$
\\
&\\
\hline
\footnotesize {\bf T3: case 2} &
\footnotesize {\bf T4: case 6} \\
\hline \\
\footnotesize
$
\begin{array}{rll}
(1.1\phantom{''}) & A \to E_{1,2}(B) & : N_A \\
\downarrow(2.1'\phantom{'}) & E_{1}(B) \to E_{2}(A) & : N_A \\
\uparrow(2.1'') & E_{2}(B) \to E_{1}(A) & : N_A \\
\end{array}
$
&
\footnotesize
$
\begin{array}{rll}
(1.1) & A \to E_{1,2}(B) & : N_A \\
(2.1) & E_{1}(B) \to A & : N_A \\
\multicolumn{3}{c}{\text{ + steps of case 5}} \\
\end{array}
$
\\
&\\
\hline
\hline
\end{tabular}
}
\end{center}
\end{table}
\section{Modeling the network and the guardian}
\label{sec:modeling}
In the previous section, we have seen how the interaction between
multiple non-collaborative dishonest agents attacking the same protocol
can interfere with both attacks, thus providing a form of defense.
As we remarked in the introduction, even if the idea of having a
guardian defending honest agents from attacks seems thrilling, the existence
of a guardian agent makes sense only with his implementation in the real
world. In order to reduce the complexity of such an implementation, we
will now investigate where we have to introduce this defense
mechanism in the network from a topological perspective (i.e., how the
guardian can dominate his competitor(s)). Modeling the network as a graph,
we study how the topological position of an attacker $E$ and a guardian
$G$, with respect to each other and to honest agents of the protocol,
can influence a protocol attack and, thus, the possible defense against
it.
We say that the outcome of the introduction of the guardian on the
network for a particular protocol yields a:
\begin{itemize}
\item \emph{false positive} if, for some reason, a normal run of the protocol is considered as an attack,
\item \emph{false negative} if, for some reason, an attack is considered as a normal run of the protocol,
\item \emph{partial defense} iff it admits false negatives,
\item \emph{total defense} iff it does not admit false negatives.
\end{itemize}
Our objective is to realize a defense mechanism that admits as few false
negatives as possible, while limiting also the number of false positives, by investigating the position that gives the guardian a topological advantage (see Definition~\ref{def:def-mechanism} of defense mechanism and the ensuing Theorem~\ref{thm:defense}).
\subsection{A network for topological advantage}
\label{sec:networkTA}
We model the network as a graph (an example is depicted in
Fig.~\ref{fig:network}), where vertices
represent the agents of the network and edges represent communication channels
(we assume no properties of these channels, which are standard insecure
channels over which messages are sent as specified by the security protocols).
Since, as we remarked above, it would be unfeasible for the guardian to
defend the traffic on all network channels, we investigate which of these
channels the guardian should be best positioned on.
\begin{figure*}[!t]
\centering
\subfloat[][\footnotesize An example of network as a graph; vertices represent agents, edges represent communication channels and the bullets $\bullet$ represent the presence of a DY-attacker $E$.]
{
\parbox{0.4\textwidth}
{\centering
\scalebox{0.8}{
\begin{tikzpicture}[font=\small]
\node (A) {$A$};
\node (B) [above=of A,xshift=40pt] {$B$};
\node (S) [below=of A,xshift=60pt,yshift=10pt] {$S$};
\node (D) [right=of A,xshift=75pt] {$D$};
\draw[double,double distance=1.5pt,line width=0.5pt] (A) to node {$\bullet$} (B);
\draw[double,double distance=1.5pt,line width=0.5pt] (A) to node {$\bullet$} (S);
\draw[double,double distance=1.5pt,line width=0.5pt] (A) to node {$\bullet$} (D);
\draw[double,double distance=1.5pt,line width=0.5pt] (B) to node {$\bullet$} (S);
\draw[double,double distance=1.5pt,line width=0.5pt] (B) to node {$\bullet$} (D);
\draw[double,double distance=1.5pt,line width=0.5pt] (S) to node {$\bullet$} (D);
\end{tikzpicture}
}\label{fig:network}
}
}
\qquad
\subfloat[][\footnotesize Two possible allocations, on a channel between $A$ and $B$ that is controlled by an attacker $E$, for the guardian $G$ when he defends an honest agent $A$. For both cases (the above one is implicit), we assume the presence of an authentic and resilient communication channel between $G$ and $A$ (dashed line).]
{
\parbox{0.45\textwidth}
{\centering
\scalebox{0.8}{
\begin{tikzpicture}[font=\small]
\node (A) {$A$};
\node (B) [right=of A,xshift=75] {$B$};
\draw[double,double distance=1.5pt,line width=0.5pt] (A.0) -- (B);
\node[] (G) [right=of A,label=above:$G$] {$\blacksquare$}
\node (E) [right=of G,label=above:$E$] {$\bullet$};
\node (A1) [below=of A]{$A$};
\node (B1) [right=of A1,xshift=75] {$B$};
\draw[double,double distance=1.5pt,line width=0.5pt] (A1.0) -- (B1);
\node[] (E1) [right=of A1,label=above:$E$] {$\bullet$};
\node (G1) [right=of E1,label=above:$G$] {$\blacksquare$}
\draw[->,dashed] (G1) to [bend left=45] (A1);
\end{tikzpicture}
}\label{fig:secChannel}
}
}
\caption{Model of the network and possible allocations of the guardian on a channel.}\label{fig:division}
\end{figure*}
Security protocols typically involve two honest agents $A$ and $B$, who
sometimes enroll also a honest and trusted third party $S$ (we could, of
course, consider protocols with more agents). As depicted in
Fig.~\ref{fig:network}, the DY-attacker $E$ is in control of all the
communication channels of the network, thus, in the case of a ping-pong
protocol between $A$ and $B$, $E$ controls also the communication
channel between $A$ and $B$. If we were to allocate a guardian $G$ on
such a channel in order to defend the honest agent $A$, it could only be
in one of two locations: as shown in Fig.~\ref{fig:secChannel}, either
the guardian $G$ is between the initiator $A$ and the attacker $E$, or
$G$ is between the attacker $E$ and the responder $B$. In the following,
these two cases will be used as a base of network topologies to be
considered during the analysis. We will see in the next section that the
guardian should have the possibility of alerting $A$ of the ongoing
attack without being detected by the attacker; in such a case (especially as
highlighted in the lower topology in Fig.~\ref{fig:secChannel}),
we thus assume the presence of
an authentic and resilient communication channel (confidentiality can be
enforced but it is not mandatory) between $G$ and $A$.\footnote{This
channel could be a digital or a physical channel, say a text message
sent to a mobile (as in some two-factor authentication or e-banking
systems), a phone call (as in burglar alarm systems), or even a flag
raised (as is done on some beaches to signal the presence of sharks). We
do not investigate the features of this channel further but simply
assume, as done in all the above three examples of runtime guarding
(monitoring) systems, that such a channel actually exists.} In the
following, this channel will be omitted from the notation and the
figures for the sake of readability.
If the network topologies for two-agent protocols are simple
(Fig.~\ref{fig:casesA} and \ref{fig:casesB}), for the case where a
trusted third party $S$ (or another agent) is present on the network, we
have to make some assumptions about the position of the attacker $E$
(the attack power of the attacker is never questioned). In this paper,
we consider four main base cases of network topologies for three-agent
protocols, where, for every case, we consider which channel(s) the guardian
is defending:
\begin{itemize}
\item Fig.~\ref{fig:casesC}: the channel between $A$ and $S$ (we assume
that the attacker is not present over these channels\footnote{We do not
make assumptions on the real topology of the network between $A$ and $S$
(i.e., there could be more than one channel) but only consider the fact
that the communications from $E$ are received by $G$.} and the guardian
acts like a proxy),
\item Fig.~\ref{fig:casesD}: the channel between $B$ and $S$ (this is
the specular scenario with respect to the previous case),
\item Fig.~\ref{fig:casesE}: $A$'s communication channel (the guardian
acts as a proxy for $A$), and
\item Fig.~\ref{fig:casesF}: $B$'s communication channel (the guardian
acts as a proxy for $B$).
\end{itemize}
These basic topologies abstract the communication channels of a
complex network (e.g., a LAN) in a way that permits one to reason
about the position of agents without introducing additional
parameters in the process (e.g., additional agents that start the
protocol at the same time, or multiple network paths relaxed in one
link).
\begin{figure}[t]
\centering
\subfloat[][\label{fig:casesA}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (G) [right=of A]{$G$};
\node (E) [right=of G]{$E$};
\node (B) [right=of E]{$B$};
\draw[double,line width=1.5pt] (A) -- (G);
\draw[double,line width=0.5pt] (G) -- (E);
\draw[double,line width=0.5pt] (E) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}
\quad
\subfloat[][\label{fig:casesB}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (E) [right=of A]{$E$};
\node (G) [right=of E]{$G$};
\node (B) [right=of G]{$B$};
\draw[double,line width=0.5pt] (A) -- (E);
\draw[double,line width=0.5pt] (E) -- (G);
\draw[double,line width=1.5pt] (G) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}
\quad
\subfloat[][\label{fig:casesC}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (G) [right=of A]{$G$};
\node (S) [above=of G,xshift=-20pt] {$S$};
\node (E) [right=of G]{$E$};
\node (B) [right=of E]{$B$};
\draw[double,line width=1.5pt] (A) -- (G);
\draw[double,line width=1.5pt] (G) -- (S);
\draw[double,line width=0.5pt] (G) -- (E);
\draw[double,line width=0.5pt] (E) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}
\quad
\subfloat[][\label{fig:casesD}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (E) [right=of A]{$E$};
\node (G) [right=of E]{$G$};
\node (S) [above=of G,xshift=20pt] {$S$};
\node (B) [right=of G]{$B$};
\draw[double,line width=0.5pt] (A) -- (E);
\draw[double,line width=0.5pt] (E) -- (G);
\draw[double,line width=1.5pt] (G) -- (S);
\draw[double,line width=1.5pt] (G) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}
\quad
\subfloat[][\label{fig:casesE}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (G) [right=of A]{$G$};
\node (E) [right=of E]{$E$};
\node (S) [above=of E,xshift=20pt] {$S$};
\node (B) [right=of E]{$B$};
\draw[double,line width=1.5pt] (A) -- (G);
\draw[double,line width=0.5pt] (G) -- (E);
\draw[double,line width=0.5pt] (E) -- (S);
\draw[double,line width=0.5pt] (E) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}
\quad
\subfloat[][\label{fig:casesF}]
{
\scalebox{0.7}{
\begin{tikzpicture}
\node (A) {$A$};
\node (E) [right=of A]{$E$};
\node (S) [above=of E,xshift=-20pt] {$S$};
\node (G) [right=of E]{$G$};
\node (B) [right=of G]{$B$};
\draw[double,line width=0.5pt] (A) -- (E);
\draw[double,line width=0.5pt] (E) -- (S);
\draw[double,line width=0.5pt] (E) -- (G);
\draw[double,line width=1.5pt] (G) -- (B);
\begin{pgfonlayer}{background}
\end{pgfonlayer}
\end{tikzpicture}
}\caption[]{Base cases of network topologies for protocols between two agents (a, b) and three agents (c, d, e, f). We denote with double stretched lines (in boldface) the channels for which we assume that the attacker is not present.}
\label{img:BaseCasesOfNetworkTopologies}
\end{figure}
In general, we cannot state that a base case is the right one or the
wrong one as this actually depends on both the analyzed protocol and the
agent we want to defend.
In order to implement the right guardian, we should consider the protocol
defense possible in each of these cases. We conjecture that all other network
topologies with two or three agents can be reduced to the base cases
introduced above, but leave a formal proof for future work.
\subsection{Network guardian in practice}
Attacks leverage protocol-dependent features, and thus attack traces
always contain particular messages that we can use as signals for
ongoing attacks.
As messages transit continuously through the network, we assume that the
guardian has a way to distinguish them (otherwise, we cannot guarantee any
type of defense).
In order to operate, the network guardian needs to interact with the
messages transiting over the network. The two modules that we define
in the architecture of the guardian are: (i) the \emph{Identification
Module}, and (ii) the \emph{Control Module}. Both modules operate
separately, do not interact with each other (even though they share the
guardian's dataset $D_G$), and are meant to (i) distinguish the
messages that belong to the protocol\footnote{We deliberately wrote
``protocol'' instead of ``protocols'' since, for now, we are not going
to consider multi-protocol attacks or protocol composition, e.g.,
\cite{CiobacaCortier10,CD-fmsd08,ModVig2009}. As future work, we
envision a distinguisher able to to distinguish between messages
belonging to different protocols and thus consider also the attacks that
occur when messages from one protocol may be confused with messages from
another protocol.} that they are defending and (ii) detect ongoing
attacks.
These features are achieved by means of two \emph{distinguishers}
$\Delta_\mathit{Id}$ and $\Delta_C$, two probabilistic polynomial
time algorithms.
$\Delta_\mathit{Id}$ returns $1$ if it believes that a message $m$ belongs to the protocol and $0$ otherwise.
We use the distinguisher $\Delta_C$ in order to detect, from the run of
a security protocol $\mathcal{P}$ (identified by the other module), those
messages $m$ that are considered \emph{critical}, i.e., that can be used to attack $\mathcal{P}$.
For a concrete example of critical message, we can refer to
Table~\ref{tab:isoSC27}. The nonce $N_A$ exchanged in message $(1.1)$ is
the first information that the attacker uses in order to perform the
reply attack against the ISO-SC 27 protocol, so this message must be
considered critical. Even though the nonce is sent as a plaintext, the
use of the distinguisher $\Delta_C$ overcomes the problem with
encrypted messages.
\begin{figure}[t]
\centering
\subfloat[][Identification Module: Assuming that $m$ is the message spied from the spy filter, $\delta$ is the state where the distinguisher is invoked on input $m$, $\phi$ the ``forward state'', $\lambda$ the ``label state''
in which the message $m$ is labeled in the dataset $D_G$ as part of the protocol $\mathcal{P}$. \label{fig:recordModule}]
{
\scalebox{0.75}{
\qqua
\begin{tikzpicture}[auto, node distance=50pt,every transition/.style={->,>=stealth'}]
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (D) [] {$\delta$};
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (F) [above=of D,xshift=40pt,yshift=-20pt]{$\phi$};
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (R) [right=of D,xshift=50pt]{$\lambda$};
\draw[transition,->,bend right] (D) to node [swap,bend right=45] {$0$} (F);
\draw[transition,->] (D) to node [] {$1$} (R);
\draw[transition,->,bend right] (R) to node [swap,bend right=45] {$0/1$} (F);
\draw[transition,->,bend right] (F) to node [swap, bend left=45] {$0/1$} (D);
\begin{pgfonlayer}{background}
\node [draw=black,dotted,fill=gray!10,fit=(D) (R) (F), rounded corners] {};
\end{pgfonlayer}
\end{tikzpicture}\qquad
}
\qquad\qquad
\subfloat[][Control Module: Assuming that $m$ is the message spied from the spy filter, $\delta$ is the state where the distinguisher is invoked on input $m$, $\iota$ the state where the attack invariant is invoked on $m$, $\phi$ represents the ``forward state'', $\rho$ the ``interference state''.
\label{fig:controlModule}]
{
\scalebox{0.75}{
\qqua
\begin{tikzpicture}[auto, node distance=50pt,every transition/.style={->,>=stealth'}]
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (D) [xshift=30pt] {$\delta$};
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10,xshift=50pt] (I) [right=of D] {$\iota$};
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (F) [above=of D,xshift=40pt,yshift=-20pt]{$\phi$};
\node[draw,rounded corners,minimum width=20pt,minimum height=20pt,fill=gray!10] (R) [below=of I,xshift=-40pt,yshift=20pt]{$\rho$};
\draw[transition,->,bend right] (D) to node [swap,bend right=45] {$0$} (F);
\draw[transition,->] (D) to node [] {$1$} (I);
\draw[transition,->,bend right] (I) to node [swap, bend right=45] {$0$} (F);
\draw[transition,->,bend left] (I) to node [bend left=45] {$1$} (R);
\draw[transition,->,bend right] (F) to node [swap, bend left=45] {$0/1$} (D);
\draw[transition,->,bend left] (R) to node [bend left=45] {$0/1$} (D);
\begin{pgfonlayer}{background}
\node [draw=black,dotted,fill=gray!10,fit=(D) (F) (R) (I), rounded corners] {};
\end{pgfonlayer}
\end{tikzpicture}
\qquad
}
\caption[]{Identification and Control Modules implemented in the guardian.}
\label{fig:RecConModule}
\end{figure}
\subsubsection{Identification Module}
Fig.~\ref{fig:recordModule} shows the graphical representation of the
\emph{Identification Module}. The guardian uses this module, together with the
distinguisher $\Delta_\mathit{Id}$, to detect those messages $m$ that
belong to the protocol and label them as part of $\mathcal{P}$ in the dataset $D_G$ in order to do inference subsequently.
We can see the Identification Module as a finite state machine where the transition from state to state depends on the spied messages.
When a message $m$ is spied by the spy filter (see Table~\ref{OurDYmodel} for
the two available spy filters), the Identification Module of the guardian
invokes the distinguisher $\Delta_\mathit{Id}(m)$ to establish whether the message belongs to the protocol or not.
If $\Delta_\mathit{Id}(m) = 0$, the message is not considered useful
and the guardian moves to the forward state $\phi$, which will
let the message go, and subsequently goes back, without checking
any condition, to the initial state $\delta$ in order to wait for
the next message. If $\Delta_\mathit{Id}(m) = 1$, then $m$ belongs to
the protocol and the guardian moves to the ``identification state''
$\lambda$, where the message is labeled in the dataset $D_G$. After
the message has been labeled, the Identification Module goes back
to the initial state $\delta$ in order to wait for the next
message.
From now on, when we do an operation (spy-filters excluded) on the
dataset, we mean (slightly abusing notation) the subset of the
labeled messages.
\subsubsection{Control Module}
Fig.~\ref{fig:controlModule} shows the graphical
representation of the Control Module. The guardian uses this
module, together with the distinguisher $\Delta_C$, in order
to deal with those messages $m$ that he must control in order to be
able to do inference (i.e., check if an attack is ongoing) and
eventually interfere with the attacker; we call these messages
\emph{critical}.
Once the distinguisher, implemented in the Control Module,
believes that $m$ is critical (at time $i$), the \emph{attack invariant} $\mathit{Inv} (m, i)$
is tested to discover (or exclude) an ongoing attack. $\mathit{Inv} (m, i)$ is a
protocol-dependent Boolean condition; formally, it is a first-order logic
formula on a critical message of the protocol (which can be straightforwardly
extended to a set of messages) tested at time $i$ (i.e., after $i$ actions on the dataset $D_{net}$; in order to define more complex functions, more than two parameters can be used):
\begin{displaymath}
\mathit{Inv}(m, i) =
\begin{cases}
1 & \text{if $m$, at time $i$, characterizes an ongoing attack or a false} \\
& \text{positive} \\
0 & \text{if $m$, at time $i$, characterizes a normal run or a false negative}
\end{cases}
\end{displaymath}
If the computation of the invariant returns $1$, then the guardian $G$
carries out the appropriate defense for the attack making the victim
abort the current run of the protocol and, eventually, mislead the
attacker and/or induce him to abort the attack. We give an example of
invariant in Section~\ref{ISOsc27} when we return to our case study.
When a message $m$ is spied by the spy filter, the Control Module is in
the initial state $\delta$, and then the message is passed as input to
the distinguisher $\Delta_C$, whose task is to establish whether
the message is critical or not. If the result of the distinguisher is
$\Delta_C(m) =0$, the message is not considered critical and the
guardian moves to the forward state $\phi$, which will let the
message go, and subsequently goes back, without checking any condition,
to the initial state $\delta$ in order to wait for the next message.
Instead, if $\Delta_C(m) =1$, then a critical message has just
been distinguished from the others; the guardian moves to the invariant
state $\iota$ passing the message as input to the attack invariant
formula $\mathit{Inv}(m, i)$, whose task is to establish whether an attack is
actually ongoing or not (the invariant is computed using the labeled
messages in $D_G$ respecting the temporal constraints). If
$\mathit{Inv}(m, i) = 0$, then either an attack is not ongoing or a false
negative has just happened (i.e., the defense mechanism is partial);
thus, the guardian goes to the forward state $\phi$, which will let
the message go, and subsequently goes back without checking any
condition to the initial state $\delta$. Instead, if $\mathit{Inv}(m, i) =1$
either an attack is ongoing or a false positive has just happened,
independently of the used defense mechanism; thus, the guardian moves
to interference state $\rho$ to carry out the appropriate
countermeasures and subsequently goes back, without checking any
condition, to the initial state.
As the $\Delta_\mathit{Id}$ is needed in order to detect the messages that
belong to the protocol $\mathcal{P}$, we envision $\Delta_C$ to
be useful in the case of protocols with a large number of messages
in order to lighten the computation load of $\mathit{Inv}(m,i)$, i.e., we
compute $\mathit{Inv}(m, i)$ on a subset of the protocol messages:
\[
\mathit{Critical} \quad \subseteq \quad \mathcal{P}_\mathit{labeled} \quad \subseteq \quad \mathit{Messages}
\]
where $\mathit{Messages}$ are all the messages saved in the dataset by
a spy-filter, $\mathcal{P}_\mathit{labeled}$ are the messages that
$\Delta_\mathit{Id}$ labeled as part of the protocol $\mathcal{P}$ and
$\mathit{Critical}$ are the messages that $\Delta_C$ believes
may be used to attack $\mathcal{P}$.
\subsection{Topological advantage}
To defend protocols against attacks, a guardian should be ``near''
one of the agents involved in the protocol executions; otherwise
the guardian could be useless: if he does not see (and thus cannot
control) messages belonging to the protocol in transit from these
agents, then he cannot carry out the interference/defense.
\begin{definition}[Topological Advantage]\label{def:top-advantage}
Let $X\in\mathit{Agents}{}$ be the agent that the guardian $G\in\mathit{BenignDishonest}$ is
defending in a particular protocol (with set $\mathit{Messages}$ of messages), and $Y\in\mathit{Agents}{}$ the other agent.
We say that $G$ is in \emph{topological
advantage} with respect to the attacker $E$ if
\begin{align*}
\forall m\in & \mathit{Messages}.\ \exists i\in \mathbb{N}. \,\\
& G \in \canseeM{<\!X, m, Y\!>, i} \ \lor \ G \in \canseeM{<\!Y, m, X\!>, i} \ \lor\\
& G \in \canseeM{<\!E(X), m, Y\!>, i} \ \lor \ G \in \canseeM{<\!Y, m, E(X)\!>, i}
\end{align*}
\end{definition}
Definition~\ref{def:top-advantage} states that for a guardian to be
in topological advantage, he must be collocated over the network in
one of the configurations of
Fig.~\ref{img:BaseCasesOfNetworkTopologies} so that he can spy (and
eventually modify) all the transiting messages to and/or from the agent that
he is defending, even in the case that they are forged.
In order to define what a defense mechanism is, we have to
formalize how an attack can be formalized based on a parametric
function that the attacker computes during his execution.
Let $E \in \mathit{Dishonest}$, $X \in \mathit{Honest}$, $s$ be the number of steps
composing the attack trace, $m_s$ the message spied over the
network or present in the attacker dataset $D_{E}$ at step $s$,
$\mathit{Func}=\{\mathit{Erase}, \mathit{Injection},
\mathit{Duplicate}, \ldots\}$ a set of functionalities that $E$ can
use on the messages. Note that the names of the functionalities
quite intuitively denote their meaning; not all of the
functionalities are used in this paper and many more could be
defined.
The functionalities in $\mathit{Func}$ have domain in
the messages belonging to a given protocol, whereas the codomain is defined as the union of all the possible transformations
of the messages in the domain that give (i) messages ``acceptable'' by the protocol (i.e., that can be sent/received by the protocol's agents) or (ii) an empty message.
The codomain is thus a set of messages.
We use ${\mathit{func}}_s$ to denote a functionality in $\mathit{Func}$
used at step $s$.
\begin{definition}[Attack Function]\label{def:att-function}
The attack function $f(m, s)$ selects a functionality $\mathit{func}_s$
to be used on the message $m$ at step $s$ and returns the result of the $\mathit{func}_s$ with argument $m$ ($\mathit{func}_s(m)$).
\end{definition}
As a concrete example, the attack function of the attack in Table~\ref{tab:isoSC27} is:
\begin{center}
\begin{tabular}{C{0.6cm}C{2.5cm}C{2.5cm}C{2.5cm}}
\hline
$s$ & $m$ & $\mathit{func}_s$ & $f(s, m)$\\%& Agent\\
\hline
$1$ & $N_A$ & $\mathit{Erase}$ & $\emptyset$ \rule[-2.5 mm]{0mm}{0.65 cm} \\
$2$ & $N_A$ & $\mathit{Injection}$ & $N_A$ \rule[-2.5 mm]{0mm}{0.65 cm} \\
$3$ & ${\{\!|N_A,N'_A|\!\}}_{K_{AB}}$ & $\mathit{Erase}$ & $\emptyset$ \rule[-2.5 mm]{0mm}{0.65 cm}\\
$4$ & ${\{\!|N_A,N'_A|\!\}}_{K_{AB}}$ & $\mathit{Injection}$ & ${\{\!|N_A,N'_A|\!\}}_{K_{AB}}$ \rule[-2.5 mm]{0mm}{0.65 cm}\\
$5$ & $N'_A$ & $\mathit{Erase}$ & $\emptyset$ \rule[-2.5 mm]{0mm}{0.65 cm}\\
$6$ & $N'_A$ & $\mathit{Injection}$ & $N'_A$ \rule[-2.5 mm]{0mm}{0.65 cm} \\
\hline
\end{tabular}
\label{attackFunctionTable}
\end{center}
Of course, more complex attack functions could (and sometimes even
should) be defined, especially for more complex protocols. Since
the attack function is but one parameter, we believe that our
definitions and results are general enough and can be quite easily
adapted to such more complex functions.
Having formalized how an attack can be seen as a parametric
function, we can also assume the existence of an inverse function
$f^{-1}(m, s)$ of the attack function (i.e., the function that from
a message $m$ such that $m=f(m', s)$, and a step $s$, computes $m'$). In this
paper, we will not discuss how to formalize the inverse attack
function; we leave a definition for future work and for now assume
that, during the implementation of the framework, a security
analyst can take care of this matter.
\begin{definition}[Defense Mechanism]\label{def:def-mechanism}
Let $X\in\mathit{Agents}{}$ be the agent that the guardian $G\in\mathit{BenignDishonest}$ is
defending in a particular protocol (with set $\mathit{Critical}$ of
critical messages), let $E\in\mathit{Dishonest}$ be the attacker, and $s$ be the number
of steps composing $E$'s attack trace. We say that $G$ is a \emph{defense mechanism}
if he knows $E$'s attack function $f(m, s)$ and can compute the inverse function $f^{-1}(m,s)$ in order to enforce the following:
\begin{multline*}
\nexists m \in \mathit{Critical}.\,
\forall i\in\mathbb{N}.\ \exists p, j\in\mathbb{N}.\ j > i \ \wedge \ 1\le p \le s \ \wedge \\
m\in D^{i}_{net} \ \wedge
f^{-1}(f(m, p), p)=m \ \wedge \\
(G\notin \canseeM{<\!E, f(m, s), X\!>, j} \lor G\notin \canseeM{<\!E(Y), f(m, s), X\!>, j})
\end{multline*}
\end{definition}
If $G$ can compute the inverse attack function, then $G$ has
knowledge of the possible attacks against the protocol carried out through the attack function and can detect the critical messages even if the attacker modifies/deletes them.
Thus, we can state the following theorem (which can be quite straightforwardly generalized to multiple attackers):
\begin{theorem}\label{thm:defense}
A guardian $G\in\mathit{BenignDishonest}$ is a defense mechanism for an agent $X\in\mathit{Agents}{}$
in a protocol $\mathcal{P}$,
if he is in topological advantage with respect to an attacker $E\in\mathit{Dishonest}$ who is attacking $X$ in $\mathcal{P}$.
\end{theorem}
As a proof sketch, let $X\in\mathit{Agents}{}$ be the agent that $G$ is defending,
$Y\in\mathit{Agents}{}$, $E\in\mathit{Dishonest}$ with attack function $f(m, p)$, $m\in \mathit{Critical}$, $f^{-1}$ known to $G$, $G$ in topological advantage with respect to the attacker $E$, $s$ the number of steps composing $E$'s attack trace, and $1 \le p\le s$. Then, since $f(m, p)\in \mathit{Messages}$, we have that:
$\exists i\in \mathbb{N}. \
G \in \canseeM{<\!X, f(m, p), Y\!>, i} \ \lor \ G \in \canseeM{<\!Y, f(m, p), X\!>, i} \ \lor \
G \in \canseeM{<\!E(X), f(m, p), Y\!>, i} \ \lor \ G \in \canseeM{<\!Y, f(m, p), E(X)\!>, i}
$.
In order to have a defense mechanism, we have to enforce the following:
$\nexists m \in \mathit{Critical}.\
\forall i\in\mathbb{N}.\ \exists p, j\in\mathbb{N}.\ j > i \ \wedge \ 1\le p \le s \ \wedge \
m\in D^{i}_{net} \ \wedge \
G\notin \canseeM{<\!E, f(m, p), X\!>, j} \wedge
f^{-1}(f(m, p), p)=m
$.
Since $f(m, p)\in \mathit{Critical} \subseteq \mathit{Messages}$, only $f^{-1}(f(m, p),p)=m$ must be enforced, but it is known to $G$ by assumption.
\section{Case studies}
\label{sec:case-studies}
\subsection{The ISO-SC 27 protocol}\label{ISOsc27}
Even though the ISO-SC 27 protocol is subject to the parallel sessions
attack shown in Table~\ref{tab:isoSC27}, we can defend it by means of
a guardian $G$.
Since the victim is $A$, for the defense to be possible, it is necessary
that $G$ is in the configuration in Fig.~\ref{fig:casesA},
i.e., between $A$ and the rest of the network agents, so that he can
identify/control all of $A$'s incoming and outgoing messages (by
Definition~\ref{def:top-advantage}, in this configuration the guardian is in topological advantage), whereas in the configuration in Fig.~\ref{fig:casesB}
he can be completely excluded by an attacker $E$.
In the following, we give as an example the successful case and a brief
explanation for the unsuccessful one.
In order to defend the ISO-SC 27 protocol, we have set up the guardian $G$
with the two spy-filters shown in Fig.~\ref{fig:gConf}: an
outflow-spy filter in order to record in his dataset $D_G$ all of $A$'s
outgoing messages, and an inflow-spy filter in order to record and control $A$'s incoming messages.
\begin{figure*}[t]
\centering
\scalebox{0.62}{
\begin{tikzpicture}[node distance=80pt, every transition/.style={->,>=stealth'}]
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (A) {$A$};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (Outflow) [above right=of A,yshift=-50pt]{\emph{outflow-spy}};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (Id1) [right=of Outflow,xshift=65pt]{Identification Module};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (Control) [below right=of A,yshift=50pt]{Control Module};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (Id2) [right=of Control,xshift=-40pt]{Identification Module};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (Inflow) [right=of Id2,xshift=-40pt]{\emph{inflow-spy}};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (NETWORK) [above right=of Inflow,yshift=-50pt]{\bm{$Network$}};
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt] (B) [right=of NETWORK,xshift=-50pt]{$B$};
\draw[transition] (A) to [bend right=-20] (Outflow);
\draw[transition] (Id1) to [bend right=-20] (NETWORK);
\draw[transition,dashed] (Outflow) to [] (Id1);
\draw[transition,dashed] (Inflow) to [] (Id2);
\draw[transition,dashed] (Id2) to [] (Control);
\draw[transition] (Control) to [bend right=-20] (A);
\draw[transition] (Control) to [bend right=-10] (NETWORK);
\draw[transition] (NETWORK) to [bend right=-20] (Inflow);
\draw[transition,<->] (NETWORK) -- (B);
\begin{pgfonlayer}{background}
\node[drop shadow,thick,draw={black!60},fill=white, minimum width=30pt, minimum height = 30pt][fill=gray!40, inner sep=7pt,fit=(Outflow) (Inflow), label=\large{Guardian}] {};
\end{pgfonlayer}
\end{tikzpicture}
}
\caption[]{Guardian configuration for the ISO-SC 27 protocol. With a dashed arrow we describe the fact that the execution flow (not the spied message) continues with the next module. \label{fig:gConf}}
\end{figure*}
Even if $G$ does not know the symmetric key $K_{AB}$, he can become
aware that the protocol has been attacked when he spies via the
\emph{inflow-spy} filter a message of the same form of the message (1) in
Table~\ref{tab:isoSC27} (i.e., $N_A$; the guardian knows that the attacker
will reply the first message because he knows the attack function of
Definition~\ref{def:att-function}) between those that have
previously been identified as such:
if an attack is ongoing, then the message that has been identified by the
Control Module as critical (i.e., is one of the first messages of the protocol) ``has already been seen'' by $G$.
We formalize this concept by means of the invariant $\mathit{Inv}(m, i)$:
\begin{displaymath}
\exists{m'} \in D^{i-1}_G.\ \Delta_C{(m)}=1 ~\wedge~
\Delta_C{(m')}=1 \ ~\wedge~ m = m' .
\end{displaymath}
That is, if an attack is ongoing and $m$ is the message spied by guardian's inflow-spy filter, labeled by the Identification Module, and in the Control Module the distinguisher $\Delta_C$ believes that it is critical, then the guardian's dataset $D_G^{i}$ must
contain another message $m'$ seen before such that $m=m'$ (the implementation of $D_G$ must be done with respect to the temporal constraints of the invariant $\mathit{Inv}{}$, but in this paper we do not discuss the implementation details).
Since the guardian knows that the attacker can use a replay attack, by
Definition~\ref{def:def-mechanism}, he has to define the inverse of the attack
function as the identity function (the use of the identity function is also reflected in the definition of the invariant).\footnote{Formally, for the ISO-SC 27 we have: $f^{-1}(f(N_A,2),2) = f^{-1}(N_A,2) = N_A$ (where $s = 2$ refers to message $(2)$ in Table~\ref{attackFunctionTable} or, equivalently, message $(1.2)$ in Table~\ref{tab:isoSC27}).}
Let us assume, following~\cite{FPV:02SECRYPTbook,FPV:SECOTS}, that each
honest agent defended by the guardian $G$ has a set of flags that $G$
can modify in order to make the agent he is defending abort the protocol. Once
he has detected such an ongoing attack, $G$ can defend it carrying out the
interference. He modifies the content (i.e., he alters the nonce $N_A$) of the
first message in the parallel session (see Table~\ref{isoRET} for the complete
execution trace, and Table~\ref{tbl:isosc27} for the corresponding dataset evolution).
At this point, the guardian already knows that an attack is ongoing, but we
choose to finish the two sessions of the protocol ($G$ changes $A$'s ``abort
flag'' only at the end) in order to show that we can also deliver false
information to the attacker and that the Control Module (shown in
Table~\ref{tbl:isosc27}) checks the invariant only once since the replayed
message in (1.2) is not seen as critical (i.e., it has not the form of the
first message). More specifically, Table~\ref{isoRET} shows the interference
attack that $G$ can use against
the attacker $E$, and Table~\ref{tbl:isosc27} the evolution of the dataset
and the inference during the protocol execution.
\begin{table}[t!]
\caption{Guardian's interference for the ISO-SC 27 protocol. \label{isoRET}}
\begin{center}
\scalebox{1.0}{
\begin{tabular}{c}
\toprule
{\bf Interference}\\
\midrule
$
\begin{array}{rll}
(1.1\phantom{_1}) & A \to E(B) & : N_A \\
(2.1\phantom{_1}) & E(B) \to G(A) & : N_A \\
(2.1_1) & G(B) \to A & : N_\mathit{fake} \\
(2.2\phantom{_1}) & A \to E(B) & : \{\!| N_\mathit{fake}, N'_A |\!\}_{K_{AB}}\\
(1.2\phantom{_1}) & E(B) \to A & : \{\!| N_\mathit{fake}, N'_A |\!\}_{K_{AB}}\\
(2.2\phantom{_1}) & G \text{ raises } A\text{'s flag for abort}
\end{array}
$
\\
\bottomrule
\end{tabular}
}
\end{center}
\end{table}
\begin{table}[t]
\caption[]{Dataset evolution and inference for the ISO-SC 27 protocol.
$\{(x.y)\}$ refers to the message sent in step $(x.y)$ (we omit the repeated messages) and to the configuration in Fig.~\ref{fig:casesA}.}
\label{tbl:isosc27}
\begin{center}
\scalebox{0.83}{
\begin{tabular}{C{.4cm}llC{2.6cm}C{1.3cm}C{1.3cm}}
\toprule
$\bm{i}$\qquad\ & \textbf{Protocol message} & $\bm{D_G^i}$ & \textbf{Identification Module} & \multicolumn{2}{C{2.6cm}}
{\begin{tabular}[x]{@{}c@{}}\textbf{Control}\\\textbf{Module}\end{tabular}}\\%{\textbf{Control Module}}\\
& & & $\Delta_\mathit{Id}(m)$ & ${\Delta_C(m)}$ & ${\mathit{Inv}(m, i)}$\\
\midrule
$0$ & $-$ &
$\{~\}$ & $-$ & $-$ & $-$\\
$1$ & $(1.1\phantom{_1}) \,\, A \to E(B) : N_A$ &
$\{(1.1)\}$\quad & $1$ & $-$ & $-$\\
$2$ & $(2.1\phantom{_1}) \,\, E(B) \to G(A) : N_A$ &
$\{(1.1)\}$ & $1$ & $1$ & $1$\\
$3$ & $(2.1_1) \,\, G(A) \to A : N_{fake}$ &
$\{(1.1), (2.1_1)\} $ & $-$ & $-$ & $-$\\
$4$ & $(2.2\phantom{_1}) \,\, A \to E(B) : \{\!|N_{fake},N_A'|\!\}_{K_{AB}}$ &
$\{(1.1), (2.1_1), (2.2)\}$ & $1$ & $-$ & $-$\\
$5$ & $(1.2\phantom{_1}) \,\, E(B) \to A : \{\!|N_{fake},N_A'|\!\}_{K_{AB}}$ \qquad &
$\{(1.1), (2.1_1), (2.2)\}$ & $1$ & $0$ & $-$\\
$6$ & $(2.2\phantom{_1})\,\, G \text{ raises } A\text{'s flag for abort}$ & $-$ & $-$ & $-$ & $-$\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
To measure the defense mechanism implemented by the guardian for the
parallel sessions attack against the ISO-SC 27 protocol, we consider false
positives and negatives.
\paragraph{False positives:} False positives are possible if, after $A$
completes a protocol run as initiator, $B$ restarts the protocol with $A$ (i.e., they change roles) using (in the first message) a nonce $N_B$ that is already contained in $G$'s dataset. If $N_B$ is represented through a $k$-bit length string, then the probability of this event is equal to the probability of guessing a nonce amongst those belonging to $D^i_G$ (i.e., $G$'s dataset after $i$ actions):
\begin{displaymath}
Pr[N_B \in_R \{0,1\}^k, N_B \in D^i_G] = \frac{|D^i_G|}{2^k}
\end{displaymath}
So, this probability is negligible if $k$ is large enough (e.g., $k =
1024$).
\paragraph{False negatives:} False negatives are not possible, since not
knowing $K_{AB}$ the only way to attack the protocol with the classical
attack (Table~\ref{tab:isoSC27}) is to reflect $A$'s messages in a
parallel session; but if this situation happens, then the guardian has
already seen the message that is coming back to $A$, and thus he can
detect (and afterwards defeat) the ongoing attack. Since $G$ does not
admit false negatives for this scenario, $G$ is a total defense
mechanism when he is in a topological advantage with respect to his
competitor(s), i.e., when he is defending $A$.
Now that we have seen the successful case, let us focus on the
configuration of Fig.~\ref{fig:casesB}. In this configuration, a
guardian would not work because $B$'s participation is not mandatory to
attack the protocol and thus $E$ can easily exclude $G$ from the run of
the protocol; thus there are no false positives and there are only false
negatives. In this case, the presence of the resilient channels does not
help because $G$ is completely excluded from seeing the execution of
the protocol and the attack.
Summing up the analysis of the case study, we have seen how a flawed
protocol as the ISO-SC 27 can be defended through the use of a guardian.
The first step of our analysis was the attack typically found via model
checking and the classical approach. We used the classical attack in
order to select the critical messages that the attacker exploits during
the attacks. Knowing the critical messages allows us to formalize the
invariant, which is also used in order to set up filters and module
configurations in the guardian architecture. Finally, we have
investigated the different outcomes with respect to the position of the
guardian in the network topology.
\subsection{Other protocols}
We have applied our approach also to a number of other security
protocols.
\begin{SPW14}
Table~\ref{tbl:protocols} summarizes our results, while a
more detailed analysis can be found in~\cite{PVZ-TR14}.\fix{Luca}{Complete the bib entry with the full details about the arxiv paper!}
\end{SPW14}
\begin{LONG}
Table~\ref{tbl:protocols} summarizes our results, while a
more detailed analysis can be found in the appendix.
\end{LONG}
For each protocol, in the table we report if the defense is total or
partial, which agent is being defended, and the topologies that
permit the defense.
In Table~\ref{tbl:protocols}, we show only the successful results
for each protocol in the given task (i.e., defending one of the
agents for the corresponding protocol). The outcome of the analysis
of these $7$ ($4$ two-agent and $3$ three-agent) protocols is quite
promising since we have a total defense in $5$ cases and a partial
defense in the remaining $2$ cases.
\begin{table}[t]
\centering
\caption[]{Other case studies. See~\cite{boydMathuria,Clark97:asurvey} for details on the protocols.}
\label{tbl:protocols}
\begin{center}
\scalebox{1.0}{
\begin{tabular}{L{4cm}C{1.8cm}C{2cm}C{2cm}}
\toprule
\textbf{Protocol} & \textbf{Defense} & \textbf{Agent Defended} & \textbf{Topology} \\
\hline
ISO-SC 27 & Total & $A$ & Fig.~\ref{fig:casesA}\\
SRA3P & Total & $A$ & Fig.~\ref{fig:casesA}\\
Andrew Secure RPC & Partial & $A$ & Fig.~\ref{fig:casesA}\\
Otway-Rees & Total & $A$ & Fig.~\ref{fig:casesC},~\ref{fig:casesE}\\
Encrypted Key Exchange
& Total & $A$ & Fig.~\ref{fig:casesA}\\
SPLICE/AS & Total & $A$ & Fig.~\ref{fig:casesC}\\
Modified BME & Partial & $B$ & Fig.~\ref{fig:casesD}\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\section{Conclusions and future work}
\label{sec:conclusions}
Discovering an attack to an already largely deployed security protocol
remains nowadays a difficult problem. Typically, the discovery of an
attack forces us to make a difficult decision: either we accept to use
the protocol even when knowing that every execution can potentially be
attacked and thus the security properties for which the protocol has
been designed can be compromised at any time, or we do not (generating
consequently, kind of a self denial of service). Both choices are
extreme, and typically the classical (and conservative) mindset prefers
to ``dismiss'' the protocol and hurry up with the deployment of a new
version hoping to be faster than those who are attempting to exploit the
discovered flaw.
The above results contribute to showing, we believe, that
non-collaborative attacker scenarios, through the introduction of a guardian,
provide the basis for the active defense of flawed security
protocols rather than discarding them when the attack is found.
Regarding the concrete applicability of this approach to security
protocols, on one hand, we can use our previous work~\cite{FPV:02SECRYPTbook,FPV:SECOTS} as an approach for discovering
how two attackers interact in non-collaborative scenarios and what
type of interference the guardian can use, and, on the other hand, in
this paper we have given the means to understand how to exploit the
interference from a topological point of view, thus bringing the
guardian close to real implementation, which is the main objective of
our current work.
We are also working on a number of relevant issues, such as how the
content of, and the meaning that the honest agents assign to, critical
messages may have an influence on the defense mechanisms enforced by the
guardian, or such as how to define general attack functions and their inverses. We are also
investigating criteria that will allow us to reason about the minimal
and/or optimal configurations for protocol defenses. For instance, to
show that no further configurations are possible (by showing how $m$
possible configurations can be reduced to $n<m$ base ones, such as the
$6$ we considered here) or that the considered configuration is optimal
for the desired defense (and thus for the implementation of the
guardian). It seems obvious, for example, that Fig.~\ref{fig:casesA} is
the optimal configuration for defending the initiator $A$ in the
majority of two-agent protocols. Similarly, our intuition is that a
guardian (with an appropriate defense for a particular protocol) put in
configuration~\ref{fig:casesE} is also valid for the
configuration~\ref{fig:casesC} (and similarly for
configuration~\ref{fig:casesF} with respect to
configuration~\ref{fig:casesD}).
We envision the some general, protocol-independent results might be
possible but that ultimately both the notion (and agents' understanding)
of critical message and that of defense configuration will depend on the
details of the protocol under consideration and of the attack to be
defended against. Our hope is thus to obtain parametric results that can
then be instantiated with the fine details of each protocol and attack.
\bibliographystyle{abbrv}
|
\section{Introduction and summary}
\subsection{AGT relations and ALE spaces}
In this paper we study a new occurrence of the deep relations between the moduli theory of sheaves and the representation theory of affine/vertex algebras.
We are particularly interested in the kind of relations which come
from gauge theory considerations. An important example of these
relations is the AGT correspondence for gauge theories on ${\mathbb{R}}^4$: in
\cite{art:aldaygaiottotachikawa2010} Alday, Gaiotto and Tachikawa
conjectured a relation between the instanton partition functions of
${\mathcal N}=2$ supersymmetric quiver gauge theories on ${\mathbb{R}}^4$ and the
conformal blocks of two-dimensional $A_{r-1}$ Toda conformal field theories (see also
\cite{art:wyllard2009,art:aldaytachikawa2010}); this conjecture has
been explicitly confirmed in some special cases, see
e.g. \cite{art:mironovmorozovshakirov2011,
art:albafateevlitvinovtarnopolskiy2011, art:tan2013,
art:aganagichaouzishakirov2014}. From a mathematical perspective,
this correspondence implies: (1) the existence of a representation of the
W-algebra $\mathcal{W}(\mathfrak{gl}_r)$ on the equivariant cohomology of the
moduli spaces $\mathcal{M}(r,n)$ of framed sheaves on the projective plane ${\mathbb{P}}^2$ of rank $r$
and second Chern class $n$ such that the latter is isomorphic to a Verma module of $\mathcal{W}(\mathfrak{gl}_r)$; (2) the fundamental classes of $\mathcal{M}(r,n)$ give a Whittaker vector of
$\mathcal{W}(\mathfrak{gl}_r)$ (pure gauge theory); (3) the Ext vertex operator is related to a certain ``intertwiner" of $\mathcal{W}(\mathfrak{gl}_r)$ under the isomorphism stated in (1) (quiver gauge theory). The instances (1) and (2) were proved by
Schiffmann and Vasserot \cite{art:schiffmannvasserot2013}, and
independently by Maulik and Okounkov
\cite{art:maulikokounkov2012}. For $r=1$, the moduli space
$\mathcal{M}(1,n)$ is isomorphic to the Hilbert scheme of $n$ points on ${\mathbb{C}}^2$ and $\mathcal{W}(\mathfrak{gl}_1)$ is the W-algebra associated with an infinite-dimensional Heisenberg algebra; the AGT correspondence for pure $U(1)$ gauge theory reduces to the famous result of Nakajima \cite{art:nakajima1997, book:nakajima1999} in the equivariant case \cite{art:vasserot2001, art:liqinwang2004, art:nakajima2014}. Presently, (3) has been proved only in the rank one case \cite{art:carlssonokounkov2012} and in the rank two case \cite{art:carlsson2015,art:negut2015}.
In this paper we are interested in the AGT correspondence for ${\mathcal N}=2$ quiver gauge theories on ALE spaces associated with the Dynkin diagram of type $A_{k-1}$ for $k\geq 2$. The corresponding instanton partition functions are defined in terms of equivariant cohomology classes over Nakajima quiver varieties of type the affine Dynkin diagram $\widehat{A}_{k-1}$. These quiver varieties depend on a real stability parameter $\xi_{\mathbb{R}}$, which lives in an open subset of ${\mathbb{R}}^k$ having a ``chambers" decomposition: if two real stability parameters belong to the same chamber, the corresponding quiver varieties are (equivariantly) isomorphic; otherwise, the corresponding quiver varieties are only ${\mathbb{C}}^\ast$-diffeomorphic. Therefore, the pure gauge theories partition functions should be all nontrivially equivalent, while the partition functions for quiver gauge theories should satisfy ``wall-crossing" formulas (cf.\ \cite{art:belavinbershteinfeiginlitvinovtarnopolsky2011,art:itomaruyoshiokuda2013}).
By looking at instanton partition functions of pure gauge theories associated with moduli spaces of ${\mathbb{Z}}_k$-equivariant framed sheaves on ${\mathbb{P}}^2$ (which are quiver varieties depending on a so-called ``level zero chamber"), the authors of \cite{art:belavinfeigin2011,art:nishiokatachikawa2011, art:belavinbelavinbershtein2011} conjectured an extension of the AGT correspondence in the A-type ALE case as a relation between instanton partition functions of ${\mathcal N}=2$ quiver gauge theories and conformal blocks of Toda-like conformal field theories with ${\mathbb{Z}}_k$ parafermionic symmetry. In particular, the pertinent algebra to consider in this case is the coset
\begin{equation}
{\mathcal A}(r,k):= \frac{\widehat{\mathfrak{gl}}_N}{\widehat{\mathfrak{gl}}_{N-k}}
\end{equation}
acting at level $r$, where $N$ is related to the equivariant parameters. For $r=1$ the algebra ${\mathcal A}(1,k)$ is simply $\widehat{\mathfrak{gl}}_k$ acting at level one. In general, ${\mathcal A}(r,k)$ is isomorphic to the direct sum of the affine Lie algebra $\widehat{\mathfrak{gl}}_k$ acting at level $r$ and the ${\mathbb{Z}}_k$-parafermionic $\mathcal{W}(\mathfrak{gl}_r)$-algebra. Checks of the conjecture has been done \cite{art:wyllard2011, art:ito2011} by using partition functions of pure gauge theories associated with moduli spaces of ${\mathbb{Z}}_k$-equivariant framed sheaves on ${\mathbb{P}}^2$. In \cite{art:bonellimaruyoshitanzini2011, art:bonellimaruyoshitanzini2012} the authors studied in details ${\mathcal N}=2$ quiver gauge theories on the minimal resolution $X_2$ of the Kleinian singularity ${\mathbb{C}}^2/{\mathbb{Z}}_2$ and provided evidences for the conjecture: in this case, the quiver variety depends on a so-called ``level infinity chamber" and corresponds to moduli spaces of framed sheaves on a suitable stacky compactification of $X_2$. In the $k=2$ case, a comparison of these approaches using different stability chambers is done in \cite{art:alfimovbelavintarnopolsky2013}; further speculations in the arbitrary $k$ case are in \cite{art:bonellimaruyoshitanziniyagi2012}.
Mathematically, this correspondence should imply: (1) the existence of a representation of the
coset ${\mathcal A}(r,k)$ on the equivariant cohomology of Nakajima quiver varieties associated with the affine A-type Dynkin diagram such that the latter is isomorphic to a Verma module of ${\mathcal A}(r,k)$; (2) the fundamental classes of the quiver varieties give a Whittaker vector of
${\mathcal A}(r,k)$ (pure gauge theory); (3) the Ext vertex operator is related to a certain ``intertwiner" of ${\mathcal A}(r,k)$ under the isomorphism stated in (1) (quiver gauge theory). As pointed out in \cite{art:alfimovbelavintarnopolsky2013}, different chambers should provide different realizations of the action conjectured in (1). On the other hand, the conjectural wall-crossing behavior of the instanton partition functions for quiver gauge theories \cite{art:itomaruyoshiokuda2013} should be related by a similar behavior of the Ext vertex operators by varying of the stability chambers.
The ALE space we consider in this paper is the minimal resolution $X_k$ of the simple Kleinian singularity ${\mathbb{C}}^2/{\mathbb{Z}}_k$. In \cite{art:bruzzopedrinisalaszabo2013} an orbifold compactification ${\mathscr{X}}_k$ of $X_k$ is constructed by adding a smooth divisor ${\mathscr{D}}_\infty$, which lays the foundations for a new sheaf theory approach to the study of $U(r)$ instantons on $X_k$ (cf.\ \cite{art:eyssidieuxsala2013}). Moduli spaces of sheaves on ${\mathscr{X}}_k$ framed along ${\mathscr{D}}_\infty$ are also constructed in \cite{art:bruzzopedrinisalaszabo2013}; by using these moduli spaces we have a new sheaf theory approach to the study of Nakajima quiver varieties with the stability parameter of $X_k$ and, consequently, of $U(r)$ gauge theories on ALE spaces of type $A_{k-1}$ which are isomorphic to $X_k$. In the present paper we use this new approach to study the AGT correspondence for abelian quiver gauge theories on $X_k$: from a physics point of view we prove the relations between instanton partition functions and conformal blocks and from a mathematical point of view we prove (1), (2) and (3).
\subsection{Summary of results}
Let us now summarize our main results. Recall that the compactification ${\mathscr{X}}_k$ is a two-dimensional projective toric orbifold with Deligne-Mumford torus $T:={\mathbb{C}}^\ast\times {\mathbb{C}}^\ast$; the complement ${\mathscr{X}}_k\setminus X_k$ is a smooth Cartier divisor ${\mathscr{D}}_\infty$ endowed with the structure of a ${\mathbb{Z}}_k$-gerbe. There exist line bundles ${\mathcal O}_{{\mathscr{D}}_\infty}(j)$ on ${\mathscr{D}}_\infty$, for $j=0, 1, \ldots, k-1$, endowed with unitary flat connections associated with the irreducible unitary representations of ${\mathbb{Z}}_k$. Hence by \cite[Theorem 6.9]{art:eyssidieuxsala2013} locally free sheaves on ${\mathscr{X}}_k$ which are isomorphic along ${\mathscr{D}}_\infty$ to ${\mathcal O}_{{\mathscr{D}}_\infty}(j)$ correspond to $U(1)$ instantons on $X_k$ with holonomy at infinity given by the $j$-th irreducible unitary representation of ${\mathbb{Z}}_k$, for $j=0,1, \ldots, k-1$.
Fix $j=0, 1, \ldots, k-1$. A rank one $({\mathscr{D}}_\infty, {\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaf on ${\mathscr{X}}_k$ is a pair $({\mathcal E}, \phi_{{\mathcal E}})$, where ${\mathcal E}$ is a rank one torsion free sheaf on ${\mathscr{X}}_k$, locally free in a neighbourhood of ${\mathscr{D}}_\infty$, and $\phi_{{\mathcal E}}\colon {\mathcal E}\big\vert_{{\mathscr{D}}_\infty} \xrightarrow{\sim} {\mathcal O}_{{\mathscr{D}}_\infty}(j)$ is an isomorphism. Let $\mathcal{M}(\vec{u},n,j)$ be the fine moduli space parameterizing isomorphism classes of rank one $({\mathscr{D}}_\infty, {\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaves on ${\mathscr{X}}_k$, with first Chern class given by $\vec{u}\in {\mathbb{Z}}^{k-1}$ and second Chern class $n$. As explained in Remark \ref{rem:firstchern}, the vector $\vec{u}$ is canonically associated with an element $\gamma_{\vec{u}}+\omega_j\in \mathfrak{Q}+\omega_j$, where $\mathfrak{Q}$ is the root lattice of the Dynkin diagram of type $A_{k-1}$ and $\omega_j$ is the $j$-th fundamental weight of type $A_{k-1}$. We denote by $\mathfrak{U}_j$ the set of vectors $\vec{u}$ associated with $\gamma+\omega_j$ for some $\gamma\in \mathfrak{Q}$.
The moduli space $\mathcal{M}(\vec{u},n,j)$ is a smooth quasi-projective variety of dimension $2n$. On $\mathcal{M}(\vec{u},n;j)$ there is a natural $T$-action induced by the toric structure of ${\mathscr{X}}_k$. Let $\varepsilon_1, \varepsilon_2$ be the generators of the $T$-equivariant cohomology of a point and consider the localized equivariant cohomology
\begin{equation}
{\mathbb{W}}_{\vec{u},j}:=\bigoplus_{n\geq0}\, H^\ast_T\big(\mathcal{M}(\vec{u},n,j)\big)\otimes_{{\mathbb{C}}[\varepsilon_1, \varepsilon_2]}{\mathbb{C}}(\varepsilon_1, \varepsilon_2)\ .
\end{equation}
Define also the total localized equivariant cohomology by summing over all vectors $\vec{u}\in\mathfrak{U}_j$:
\begin{equation}
{\mathbb{W}}_j:=\bigoplus_{\vec{u}\in \mathfrak{U}_j}\, {\mathbb{W}}_{\vec{u},j}\ .
\end{equation}
The affine Lie algebra $\widehat{\mathfrak{gl}}_k$ acts on ${\mathbb{W}}_j$ as follows (see Proposition \ref{prop:representation} and Proposition \ref{prop:Wkweight}).
\begin{proposition*}
There exists a $\widehat{\mathfrak{gl}}_k$-action on ${\mathbb{W}}_j$ under which it is the $j$-th dominant representation of $\widehat{\mathfrak{gl}}_k$ at level one, i.e., the highest weight representation of $\widehat{\mathfrak{gl}}_k$ with fundamental weight $\widehat{\omega}_j$ of type $\widehat{A}_{k-1}$. Moreover, the weight spaces of ${\mathbb{W}}_j$ with respect to the $\widehat{\mathfrak{gl}}_k$-action are the ${\mathbb{W}}_{\vec{u},j}$ with weights $\gamma_{\vec{u}}+\omega_j$.
\end{proposition*}
The vector spaces ${\mathbb{W}}_{\vec{u},j}$ also have a representation theoretic intepretation.
\begin{corollary*}[Corollary \ref{cor:hwrep}]
${\mathbb{W}}_{\vec{u},j}$ is a highest weight representation of the Virasoro algebra associated with $\widehat{\mathfrak{gl}}_k$ of conformal dimension $\Delta_{\vec{u}}:=\frac{1}{2}\, \vec{u}\cdot C^{-1}\vec{u}$, where $C$ is the Cartan matrix of type $A_{k-1}$.
\end{corollary*}
The representation is constructed by using a vertex algebra approach via the Frenkel-Kac construction. A similar construction for the cohomology groups of moduli spaces of rank one torsion free sheaves over smooth projective surfaces is outlined in \cite[Chapter 9]{book:nakajima1999}. In \cite{art:nagao2009}, Nagao analysed vertex algebra realizations of representations of $\widehat{\mathfrak{sl}}_k$ on the equivariant cohomology groups of Nakajima quiver varieties associated with the affine Dynkin diagram $\widehat{A}_{k-1}$, for an integer $k\geq 2$, with dimension vector corresponding to the trivial holonomy at infinity $j=0$; in this case the pertinent representation is the basic representation of $\widehat{\mathfrak{sl}}_k$.
In the following we describe our AGT relations, which connect together ${\mathbb{W}}_j$ for $j=0,1, \ldots, k-1$, the action of $\widehat{\mathfrak{gl}}_k$ on ${\mathbb{W}}_j$ and abelian quiver gauge theories on $X_k$. The first relation we obtain concerns the pure gauge theory. Let ${\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j$ be the instanton partition function for the pure ${\mathcal N}=2$ $U(1)$ gauge
theory on the ALE space $X_k$ with fixed holonomy at infinity given by the $j$-th irreducible representation of ${\mathbb{Z}}_k$ (see Section \ref{sec:N=2Xk}). It has the following representation theoretic characterization.
\begin{theorem*}[AGT relation for pure ${\mathcal N}=2$ $U(1)$ gauge theory]
The Gaiotto state
\begin{equation}
G_j:= \sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{n\geq 0} \,
\big[\mathcal{M}(\vec{u},n,j) \big]_T
\end{equation}
is a Whittaker vector for $\widehat{\mathfrak{gl}}_k$. Moreover, the weighted norm of the weighted Gaiotto state
\begin{equation}
G_j\big(\mathsf{q},\vec{\xi} \ \big) := \sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{
n\geq 0} \, \mathsf{q}^{n+\frac{1}{2}\, \vec{u} \cdot C^{-1}\vec{u}}\
\vec{\xi}^{\ C^{-1}\vec{u}}\ \big[\mathcal{M}(\vec{u},n,j) \big]_T
\end{equation}
is exactly ${\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j$.
\end{theorem*}
We also consider ${\mathcal N}=2$ superconformal quiver gauge theories with gauge group $U(1)^{r+1}$ for some $r\geq0$. By the ADE classification in \cite[Chapter 3]{art:nekrasovpestun2012} the admissible quivers in this case are the linear quivers of the finite-dimensional $A_r$-type Dynkin diagram and the cyclic quivers of the affine $\widehat{A}_{r}$-type extended Dynkin
diagram. In order to state AGT relations in these cases, we introduce Ext vertex operators \cite{art:carlssonokounkov2012,art:carlsson2015, art:negut2015}. Consider the element $\boldsymbol{E}_\mu\in K\big(\mathcal{M}(\vec{u}_1,n_1,j_1)\times\mathcal{M}(\vec{u}_2,n_2,j_2)\big)$ whose fibre over a point $\big([({\mathcal E},\phi_{\mathcal E})]$, $[({\mathcal E}',\phi_{{\mathcal E}'})]\big)$ is
\begin{equation}
\big(\boldsymbol{E}_\mu\big)_{\left([({\mathcal E},\phi_{\mathcal E})]\,,\,[({\mathcal E}',\phi_{{\mathcal E}'})]\right)}=\operatorname{Ext}^1\big({\mathcal E},{\mathcal E}'\otimes
{\mathcal O}_{{\mathscr{X}}_k}(\mu)\otimes{\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty) \big)\ ,
\end{equation}
where ${\mathcal O}_{{\mathscr{X}}_k}(\mu)$ is the trivial line bundle on ${\mathscr{X}}_k$ on which the torus $T_\mu={\mathbb{C}}^\ast$ acts by scaling the fibres with $H_{T_\mu}^\ast({\rm pt};{\mathbb{C}})={\mathbb{C}}[\mu]$. By using the Euler class of $\boldsymbol{E}_\mu$ we define a vertex operator $\operatorname{V}_\mu(\vec x,z)\in\operatorname{End}\big(\bigoplus_{j=0}^{k-1}\,{\mathbb{W}}_j\, \big)[[z^{\pm\, 1},$ $x_1^{\pm\,1},$ $\dots,x_{k-1}^{\pm\,1}]]$ (see Section \ref{se:bifundXk}). Under the decomposition $\widehat{\mathfrak{gl}}_k=\mathfrak{h}\oplus \widehat{\mathfrak{sl}}_k$, we have the following characterization of $\operatorname{V}_\mu(\vec x,z)$ in terms of vertex operators depending respectively on $\mathfrak{h}$ and $\widehat{\mathfrak{sl}}_k$.
\begin{theorem*}[Theorem \ref{thm:virasoroprimary}]
The vertex operator $\operatorname{V}_\mu(\vec x, z)$ can be expressed in the form
\begin{multline}
\operatorname{V}_\mu(\vec x, z)=\operatorname{V}_{-\frac{\mu}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}
}(z) \\ \otimes \ \sum_{j_1,j_2=0}^{k-1} \
\sum_{\vec{u}_1\in\mathfrak{U}_{j_1},\vec{u}_2\in\mathfrak{U}_{j_2}}\
\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)\, z^{\Delta_{\vec u_2}-\Delta_{\vec u_1}}\ \exp\big(\log z\ \mathfrak{c}-\gamma_{21}\,\big)\,\exp\big(\gamma_{21}\,\big)\big\vert_{ {\mathbb{W}}_{\vec{u}_1, j_1}}
\ ,
\end{multline}
where $\operatorname{V}_{\alpha,\beta}(z)$ denotes a generalized bosonic exponential associated with the Heisenberg algebra $\mathfrak{h}$ (see Definition \ref{def:bosonops}), $\exp\big(\log z\ \mathfrak{c}-\gamma_{21}\,\big)\,\exp\big(\gamma_{21}\,\big)$ is the vertex operator on ${\mathbb{W}}_{j_1}$ defined in Equation \eqref{eq:expopWj}, and $\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)$ is the primary field \eqref{eq:barVmudef} of the Virasoro algebra
associated with $\widehat{\mathfrak{sl}}_k$ with conformal
dimension $\Delta_{\vec u_2-\vec u_1}=\frac12\, \vec v_{21}\cdot C\vec v_{21}$, where $\vec v_{21}:= C^{-1}(\vec u_2-\vec u_1)$ for $j_1, j_2=0,1, \ldots, k-1$ and $\vec{u}_1\in\mathfrak{U}_{j_1},\vec{u}_2\in\mathfrak{U}_{j_2}$.
\end{theorem*}
For $j_1,j_2=0,1, \ldots, k-1$ denote by $\operatorname{V}_\mu^{j_1,j_2}(\vec x, z)$ the restriction of the vertex operator
$\operatorname{V}_\mu(\vec x, z)$ to $\operatorname{Hom}({\mathbb{W}}_{j_1}, {\mathbb{W}}_{j_2})[[z^{\pm\,1},x_1^{\pm\,1},\dots,x_{k-1}^{\pm\,1}]]$.
Let ${\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}$ be the instanton partition function for the
${\mathcal N}=2$ superconformal $U(1)^{r+1}$ quiver gauge theory of type $\widehat{A}_{r}$ with holonomy at infinity associated with $\boldsymbol{j}:=(j_0, j_1, \ldots, j_r)$, topological couplings $\mathsf{q}_\upsilon\in{\mathbb{C}}^*$ and $\vec\xi_\upsilon \in ({\mathbb{C}}^*)^{k-1}$ for
$\upsilon=0,1, \ldots, r$, and masses $\boldsymbol{\mu}:=(\mu_0,\mu_1, \ldots, \mu_r)$. We prove the following AGT relation.
\begin{theorem*}[AGT relation for
${\mathcal N}=2$ $U(1)^{r+1}$ quiver gauge theory of type $\widehat{A}_{r}$]
The partition function of the $\widehat{A}_r$-theory on $X_k$ is given
by
\begin{equation}
{\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} = \operatorname{Tr}_{{\mathbb{W}}_{j_0}} \, \mathsf{q}^{L_0}\ \vec\xi{}^{\ C^{-1}\vec h} \ \prod_{\upsilon=0}^r\,
\operatorname{V}_{\mu_\upsilon}^{j_\upsilon,j_{\upsilon+1}}(\vec x_\upsilon,z_\upsilon)\ \delta^{\rm
conf}_{\upsilon,{\upsilon+1}} \ ,
\end{equation}
where
$\mathsf{q}:= \mathsf{q}_0\, \mathsf{q}_1\cdots \mathsf{q}_r$, $( \, \vec\xi\ )_i:=
(\vec\xi_0)_i\, (\vec\xi_1)_i\, \cdots (\vec\xi_r)_i$, $z_\upsilon:=
z_0\, \mathsf{q}_1\cdots \mathsf{q}_\upsilon$, and $(\vec x_\upsilon)_i:= (\vec
x_0)_i\, (\vec\xi_1)_i\cdots$ $ (\vec\xi_\upsilon)_i$ for
$\upsilon=1,\dots,r$ and $i=1,\dots,k-1$. Here $L_0$ is the Virasoro energy operator associated to $\widehat{\mathfrak{gl}}_k$, $\vec h=(h_1,\dots,h_{k-1})$ are the generators of the Cartan subalgebra of $\mathfrak{sl}_k$, and $\delta^{\rm
conf}_{\upsilon,{\upsilon+1}}$ is the conformal restriction operator defined in Equation \eqref{eq:confrestrop}.
\end{theorem*}
We also get a characterization of ${\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}$ in terms of the corresponding partition function on ${\mathbb{C}}^2$ and a part depending only on $\widehat{\mathfrak{sl}}_k$.
\begin{corollary*}
Let $\operatorname{V}_\mu(\vec{v}_{21},\vec x,
z):= \, z^{\Delta_{\vec u_2}-\Delta_{\vec u_1}} {\bar \operatorname{V}}_\mu(\vec{v}_{21}, \vec x, z)\ \exp\big(\log z\ \mathfrak{c}-\gamma_{21} \big)\, \exp\big(\gamma_{21}\,\big)$. Then we have
\begin{multline}
{\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}
= {\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\mathsf{q})^{\frac{1}{k}}\ \mathsf{q}^{\frac{1}{24}\,(1-\frac{1}{k})}\,\eta(\mathsf{q})^{\frac{1}{k}-1} \\
\times\ \operatorname{Tr}_{{\mathcal V}(\,\widehat{\omega}_{j_0}\,)} \,
\mathsf{q}^{L_0^{\widehat{\mathfrak{sl}}_k}}\ \vec\xi{}^{\ C^{-1}\vec h}\ \prod_{\upsilon=0}^r\
\sum_{(\vec{u}_\upsilon\in\mathfrak{U}^{\rm conf}_{j_\upsilon})} \,
\operatorname{V}_{\mu_\upsilon}(\vec{v}_{\upsilon,\upsilon+1},\vec x_\upsilon, z_\upsilon) \, \big\vert
_{{\mathbb{W}}_{\vec{u}_\upsilon,j_\upsilon}} \ ,
\end{multline}
where $\eta(\mathsf{q})$ is the Dedekind function, ${\mathcal V}(\,\widehat{\omega}_{j_0}\,)$ is the $j_0$-th dominant representation of $\widehat{\mathfrak{sl}}_k$ and $\mathfrak{U}^{\rm conf}_{j_\upsilon}$ is the subset of $\mathfrak{U}_{j_\upsilon}$ defined in Equation \eqref{eq:confchargeset}.
\end{corollary*}
Let ${\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}$ be the instanton partition function for the
${\mathcal N}=2$ superconformal $U(1)^{r+1}$ quiver gauge theory of type ${A}_{r}$ with holonomy at infinity associated with $\boldsymbol{j}:=(j_0,j_1, \ldots, j_r)$. We also prove the following AGT relation.
\begin{theorem*}[AGT relation for ${\mathcal N}=2$ $U(1)^{r+1}$ quiver gauge theory of type ${A}_{r}$]
The partition function of the $A_r$-theory on $X_k$ is given
by
\begin{multline}
{\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} \\
= \Big\langle |0\rangle_{\rm conf} \, , \, V_{\mu_0}(\vec x_0,z_0) \, \Big(\, \prod_{\upsilon=1}^{r}\,
\operatorname{V}_{\mu_{\upsilon}}^{j_{\upsilon}-1, j_{\upsilon}}(\vec x_\upsilon,z_\upsilon)\ \delta^{\rm
conf}_{{\upsilon-1},{\upsilon}} \, \Big)\, \operatorname{V}_{\mu_{r+1}}(\vec x_{r+1},z_{r+1}) |0\rangle_{\rm conf} \Big\rangle_{\bigoplus_{j=0}^{k-1}\,{\mathbb{W}}_j} \ ,
\end{multline}
where
$z_\upsilon:=
z_0\, \mathsf{q}_0\, \mathsf{q}_1\cdots \mathsf{q}_\upsilon$ and $(\vec x_\upsilon)_i:= (\vec
x_0)_i\, (\vec\xi_0)_i\, (\vec\xi_1)_i$ $\cdots (\vec\xi_{\upsilon-1})_i$ for
$\upsilon=1,\dots,r+1$, $i=1,\dots,k-1$, and $|0\rangle_{\rm conf}:= \prod_{\upsilon=0}^r\,\delta^{\rm
conf}_{0,\upsilon}\triangleright [\emptyset, \vec{0}\,]$ with
$[\emptyset, \vec{0}\,]$ the vacuum vector of the fixed point basis of
$\bigoplus_{j=0}^{k-1}\,{\mathbb{W}}_j$.
\end{theorem*}
Denote by ${\mathcal V}$ the direct sum of the $k$ level one dominant representations of $\widehat{\mathfrak{sl}}_k$. Similarly to before, we have the following characterization.
\begin{corollary*} We have
\begin{multline}
{\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}
= {\mathcal Z}_{{\mathbb{C}}^2}^{{A}_r}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})^{\frac1k} \\ \shoveleft{ \times \
\Big\langle |0\rangle_{\rm conf} \, , \, \Big(\, \sum_{j_0,j_0'=0}^{k-1} \
\sum_{\vec{u}_0\in\mathfrak{U}_{j_0},\vec{u}_0'\in\mathfrak{U}_{j_0'}} \
\operatorname{V}_{\mu_0}(\vec{v}_{0',0},\vec x_0, z_0)\, \big\vert
_{{\mathbb{W}}_{\vec{u}_0,j_0}} \, \Big) } \\ \times \
\prod_{\upsilon=1}^{r} \ \sum_{(\vec u_\upsilon\mathfrak{U}^{\rm conf}_{j_\upsilon})}\,
\operatorname{V}_{\mu_{\upsilon}}(\vec v_{\upsilon-1,\upsilon},\vec x_\upsilon,z_\upsilon)\, \big|_{{\mathbb{W}}_{\vec u_\upsilon,j_\upsilon}} \\ \times \ \Big(\,
\sum_{j_{r+1},j_{r+1}'=0}^{k-1} \
\sum_{\vec{u}_1\in\mathfrak{U}_{j_1},\vec{u}_1'\in\mathfrak{U}_{j_1'}} \
\operatorname{V}_{\mu_{r+1}}(\vec{v}_{1',1},\vec x_{r+1}, z_{r+1})\, \big\vert
_{{\mathbb{W}}_{\vec{u}_1,j_1}}\, \Big)
|0\rangle_{\rm conf} \Big\rangle_{{\mathcal V}} \ .
\end{multline}
\end{corollary*}
Another important aspect of the AGT correspondence that we address in
this paper is the relation of our construction with quantum integrable
systems. In particular, for any $j=0,1, \ldots, k-1$ we define an
infinite system of commuting operators which are diagonalized in the
fixed point basis of ${\mathbb{W}}_j$; geometrically these operators correspond
to multiplication by equivariant cohomology classes (see Section
\ref{sec:integrals}). The eigenvalues of these operators with respect
to this basis can be decomposed into a part associated with $k$
non-interacting Calogero-Sutherland models and a part which can be
interpreted as particular matrix elements of the vertex operators
$\operatorname{V}_\mu(\vec x,z)$ in highest weight vectors of $\widehat{\mathfrak{gl}}_k$. The
significance of this property is that this special orthogonal
basis manifests itself in the special integrable structure of the
two-dimensional conformal field theory and yields completely factorized matrix elements of composite vertex
operators explicitly in terms of simple rational functions of the
basic parameters, which from the gauge theory perspective represent the contributions of bifundamental matter fields.
The study of the AGT relation for pure ${\mathcal N}=2$ $U(1)$ gauge theories and the problem of constructing commuting operators associated with $\widehat{\mathfrak{gl}}_k$ is also addressed in \cite{art:belavinbershteintarnopolsky2013} from another point of view: there they consider the ``conformal" limit of the Ding-Iohara algebra, depending on parameters $q,t$, for $q,t $ approaching a primitive $k$-th root of unity and relate the representation theory of this limit to the AGT correspondence. However, their point of view is completely algebraic, so unfortunately it is not clear to us how to geometrically construct the action of the conformal limit on the equivariant cohomology groups.
\subsection{Outline}
This paper is structured as follows. In Section \ref{sec:combprelim} we briefly recall the relevant combinatorial notions that we use in this paper. In Section \ref{sec:infiniteliealg} we collect preliminary material on Heisenberg algebras and affine Lie algebras of type $\widehat{A}_{k-1}$, giving particular attention to the Frenkel-Kac construction of level one dominant representations of $\widehat{\mathfrak{sl}}_k$ and $\widehat{\mathfrak{gl}}_k$. In Section \ref{sec:AGTonR4} we review the AGT relations for ${\mathcal N}=2$ superconformal abelian quiver gauge theories on ${\mathbb{R}}^4$. In Section \ref{sec:sheaves} we briefly recall the construction of the orbifold compactification ${\mathscr{X}}_k$ and of moduli spaces of framed sheaves on ${\mathscr{X}}_k$ from \cite{art:bruzzopedrinisalaszabo2013}. Section \ref{sec:glkreps} addresses the construction of the action of $\widehat{\mathfrak{gl}}_k$ on ${\mathbb{W}}_j$ for $j=0,1, \ldots, k-1$: we perform a vertex algebra construction of the representation by using the Frenkel-Kac theorem. In Section \ref{sec:chiralvertex} we define the virtual bundle $\boldsymbol{E}_\mu$ and the vertex operator $\operatorname{V}_\mu(\vec x,z)$, and we characterize it in terms of vertex operators of an infinite-dimensional Heisenberg algebra $\mathfrak{h}$ and primary fields of $\widehat{\mathfrak{sl}}_k$ under the decomposition $\widehat{\mathfrak{gl}}_k=\mathfrak{h}\oplus \widehat{\mathfrak{sl}}_k$; moreover, we geometrically define an infinite system of commuting operators. In Section \ref{sec:quivergaugeXk} we prove our AGT relations, and furthermore provide expressions for our partition functions in terms of the corresponding partition functions on ${\mathbb{C}}^2$ and a part depending only on $\widehat{\mathfrak{sl}}_k$. The paper concludes with two Appendices containing some technical details of the constructions from the main text: in Appendix \ref{app:virasoroprimary} we give the proof that the vertex operator $\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)$ is a primary field, while in Appendix \ref{app:edgecontributions} we recall the expressions from \cite{art:bruzzopedrinisalaszabo2013} for the edge factors which appear in the definition of $\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)$ as well as in the eigenvalues of the integrals of motion.
\subsection{Acknowledgements}
We are grateful to M.\ Bershtein, A.\ Konechny, O.\ Schiffmann and E.\ Vasserot for helpful discussions. Also, we are indebted to the anonymous referee, whose remarks helped to improve the paper. This work was supported in part by PRIN ``Geometria delle variet\`a algebriche", by GNSAGA-INDAM, by the Grant RPG-404 from the Leverhulme Trust, and by the Consolidated Grant ST/J000310/1 from the UK Science and Technology Facilities Council. The bulk of this paper was written while the authors were staying at Heriot-Watt University in Edinburgh and at SISSA in Trieste. The last draft of the paper was written while the first and second authors were staying at IHP in Paris under the auspices of the RIP program. We thank these institutions for their hospitality and support.
\bigskip \section{Combinatorial preliminaries\label{sec:combprelim}}
\subsection{Partitions and Young tableaux}\label{sec:Young}
A \emph{partition} of a positive integer $n$ is a nonincreasing
sequence of positive numbers $\lambda=(\lambda_1\geq \lambda_2\geq
\cdots\geq \lambda_\ell> 0)$ such that
$\vert\lambda\vert:=\sum_{a=1}^\ell \, \lambda_a=n.$ We call
$\ell=\ell(\lambda)$ the \emph{length} of the partition $\lambda.$
Another description of a partition $\lambda$ of $n$ uses the notation
$\lambda=(1^{m_1}\, 2^{m_2}\, \cdots)$, where
$m_i=\#\{a\in{\mathbb{N}}\,\vert \, \lambda_a =i\}$ with $\sum_i \, i\,
m_i=n$ and $\sum_i\, m_i=\ell(\lambda)$. On the set of all partitions there is a natural
partial ordering called
\emph{dominance ordering}:
For two partitions $\mu$ and $\lambda$, we write $\mu\leq\lambda$ if
and only if $|\mu|=|\lambda|$ and
$\mu_1+\cdots+\mu_a\leq\lambda_1+\cdots+\lambda_a$ for all
$a\geq1$. We write $\mu<\lambda$ if and only if $\mu\leq\lambda$ and $\mu\neq\lambda.$
One can associate with a partition $\lambda$ its \emph{Young tableau},
which is the set $Y_\lambda=\{(a,b)\in{\mathbb{N}}^2\,\vert\, 1\leq a \leq
\ell(\lambda)\, ,\, 1\leq b \leq \lambda_a\}$. Then $\lambda_a$ is the
length of the $a$-th column of $Y_\lambda$; we write $\vert
Y_\lambda\vert=\vert\lambda\vert$ for the \emph{weight} of the Young
tableau $Y_\lambda$. We shall identify a partition $\lambda$ with its
Young tableau $Y_\lambda$. For a partition $\lambda$, the
\emph{transpose partition} $\lambda'$ is the partition whose Young
tableau is $Y_{\lambda'}:=\{(b,a)\in {\mathbb{N}}^2\,\vert\,(a,b)\in Y_\lambda\}$.
The elements of a Young tableau $Y$ are called the \emph{nodes} of
$Y$. For a node $s=(a,b)\in Y$, the \emph{arm length} of $s$ is
the quantity $A(s):=A_Y(s)=\lambda_a-b$ and the \emph{leg length} of $s$ the quantity $L(s):= L_Y(s)=\lambda_b'-a$. The \emph{arm colength} and \emph{leg colength} are respectively given by $A'(s):= A_Y'(s)=b-1$ and $L'(s):= L_Y'(s)=a-1$.
\subsection{Symmetric functions}\label{sec:symmetric}
Here we recall some preliminaries about the theory of symmetric functions in infinitely many variables which we shall use later on. Our main reference is \cite{book:macdonald1995}.
Let ${\mathbb{F}}$ be a field of characteristic zero. The \emph{algebra
of symmetric polynomials in} $N$ \emph{variables} is the subspace
$\Lambda_{{\mathbb{F}},N}$ of ${\mathbb{F}}[x_1,\ldots,x_N]$ which is invariant under the
action of the group of permutations $\sigma_N$ on $N$ letters. Then $\Lambda_{{\mathbb{F}},N}$ is a graded ring: $\Lambda_{{\mathbb{F}},N}=\bigoplus_{n\geq0}\, \Lambda_{{\mathbb{F}},N}^n$, where $\Lambda_{{\mathbb{F}},N}^n$ is the ring of homogeneous symmetric polynomials in $N$ variables of degree $n$ (together with the zero polynomial).
For any $M>N$ there are morphisms
$\rho_{MN}\colon\Lambda_{{\mathbb{F}},M}\to \Lambda_{{\mathbb{F}},N}$ that map the
variables $x_{N+1},\ldots,x_M$ to zero. They preserve the grading, and
hence we can define $\rho_{MN}^n\colon\Lambda^n_{{\mathbb{F}},M}\to
\Lambda^n_{{\mathbb{F}},N}$; this allows us to define the inverse limits
\begin{equation}
\Lambda^n_{\mathbb{F}}:= \lim_{\stackrel{\scriptstyle\longleftarrow}{\scriptstyle
N}} \, \Lambda^n_{{\mathbb{F}},N}\ ,
\end{equation}
and the \emph{algebra of symmetric functions in infinitely many variables} as $\Lambda_{\mathbb{F}}:=\bigoplus_{n\geq0}\, \Lambda^n_{\mathbb{F}}.$ In the following when no confusion is possible we will denote $\Lambda_{\mathbb{F}}$ (resp.\ $\Lambda_{\mathbb{F}}^n$) simply by $\Lambda$ (resp. $\Lambda^n$).
Now we introduce a basis for $\Lambda.$ For this, we start by defining a basis in $\Lambda_N.$ Let $\mu=(\mu_1,\ldots,\mu_t)$ be a partition with $t\leq N$, and define the polynomial
\begin{equation}
m_\mu(x_1,\ldots,x_N) =\sum_{\tau\in \sigma_N}\, x_1^{\mu_{\tau(1)}}\cdots x_N^{\mu_{\tau(N)}}\ ,
\end{equation}
where we set $\mu_j=0$ for $j=t+1,\ldots, N$. The polynomial $m_\mu$ is symmetric, and the set of $m_\mu$ for all partitions $\mu$ with $\vert \mu \vert\leq N$ is a basis of $\Lambda_N.$ Then the set of $m_\mu$, for all partitions $\mu$ with $\vert \mu \vert\leq N$ and $\sum_i\, \mu_i=n$, is a basis of $\Lambda^n_N$. Since for $M>N\geq t$ we have $\rho^n_{MN}(m_\mu(x_1,\ldots,x_M))=m_\mu(x_1,\ldots,x_N)$, by using the definition of inverse limit we can define the \emph{monomial symmetric functions} $m_\mu.$ By varying over the partitions $\mu$ of $n$, these functions form a basis for $\Lambda^n$.
Next we define the $n$\emph{-th power sum symmetric function} $p_n$ as
\begin{equation}
p_n := m_{(n)}= \sum_{i} \, x_i^n\ .
\end{equation}
The set consisting of symmetric functions $p_\mu:=p_{\mu_1}\dots p_{\mu_t}$, for all partitions $\mu=(\mu_1,\ldots,\mu_t)$, is another basis of $\Lambda$.
We now set ${\mathbb{F}}={\mathbb{C}}$ throughout and we fix a parameter $\beta\in {\mathbb{C}}$ (though
everything works for any field extension ${\mathbb{C}}\subseteq {\mathbb{F}}$ and $\beta\in
{\mathbb{F}}$). Define an inner product $\langle-,-\rangle_{\beta}$ on the vector space
$\Lambda\otimes{\mathbb{Q}}(\beta)$ with respect to which the basis of power sum symmetric functions $p_\lambda(x)$ are orthogonal with the normalization
\begin{equation}\label{eq:jackinnerprod}
\langle p_\lambda,p_\mu\rangle_{\beta} = \delta_{\lambda,\mu}\ z_\lambda\, \beta^{-\ell(\lambda)} \ ,
\end{equation}
where $\delta_{\lambda,\mu}:=\prod_{a}\, \delta_{\lambda_a,\mu_a}$ and
\begin{equation}
z_\lambda:= \prod_{j\geq1}\, j^{m_j}\, m_j! \ .
\end{equation}
This is called the \emph{Jack inner product}.
\begin{definition}\label{def:macdonald}
The monic forms of the {\em Jack functions} $J_\lambda( x;\beta^{-1} )\in\Lambda\otimes{\mathbb{Q}}(\beta)$ for $x=(x_1, x_2, \ldots)$ are uniquely defined by the following two conditions \cite{book:macdonald1995}:
\begin{itemize}
\item[(i)] Triangular expansion in the basis $m_\mu(x)$ of monomial symmetric functions:
\begin{equation}
J_\lambda(x;\beta^{-1}) = m_\lambda (x) + \sum_{\mu < \lambda} \,
\psi_{\lambda, \mu} (\beta) \, m_\mu (x) \qquad\mbox{with} \quad \psi_{\lambda, \mu} (\beta) \in {\mathbb{Q}}(\beta)\ .
\end{equation}
\item[(ii)] Orthogonality:
\begin{equation}\label{eq:orthogonalityjack}
\langle J_\lambda, J_\mu\rangle_{\beta} = \delta_{\lambda,\mu} \ \prod_{s\in Y_\lambda} \ \frac{\beta\,L(s) + A(s)+1}{\beta\,\big(L(s)+1 \big) + A(s)} \ .
\end{equation}
\end{itemize}
\end{definition}
\begin{lemma}\label{lem:p1}
For any integer $n\geq 1$ we have
\begin{equation}\label{eq:p1}
(p_1)^n=n!\, \sum_{\vert\lambda\vert=n} \ \prod_{s\in Y_\lambda} \ \frac{1}{\beta\,L(s) + A(s)+1}\ J_\lambda\ .
\end{equation}
\end{lemma}
\proof
The assertion follows straightforwardly from \cite[Proposition 2.3 and
Theorem 5.8]{art:stanley1989} after normalizing our Jack functions:
the Jack functions considered in \cite{art:stanley1989} are given by
\begin{equation}
\tilde J_\lambda=\beta^{-\vert\lambda\vert}\ \prod_{s\in Y_\lambda}\
\big[\beta\,\big(L(s)+1 \big) + A(s)\big] \ J_\lambda\ ,
\end{equation}
where the normalization factor is computed by using \cite[Theorem 5.6]{art:stanley1989}.
\endproof
\bigskip \section{Infinite-dimensional Lie algebras\label{sec:infiniteliealg}}
\subsection{Heisenberg algebras}
In this section we recall the representation theory of Heisenberg
algebras and the affine Lie algebras $\widehat{\mathfrak{sl}}_k$. Since the Lie algebra $\mathfrak{gl}_k$ coincides with ${\mathbb{F}}\,\mathrm{id}\oplus\mathfrak{sl}_k$, as a by-product we get the representation theory of $\widehat{\mathfrak{gl}}_k$.
Let ${\mathbb{C}}\subseteq {\mathbb{F}}$ be a field extension of ${\mathbb{C}}$. Let $\mathfrak{L}$ be a
{lattice}, i.e., a free abelian group of finite rank $d$ equipped with a symmetric nondegenerate bilinear form $\langle -,- \rangle_{\mathfrak{L}} \colon \mathfrak{L}\times \mathfrak{L} \to {\mathbb{Z}}$. Fix a basis $\gamma_1, \ldots, \gamma_d$ of $\mathfrak{L}$.
\begin{definition}
The \emph{lattice Heisenberg algebra} $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ associated with $\mathfrak{L}$ is the infinite-dimensional Lie algebra over ${\mathbb{F}}$ generated by $\mathfrak{q}^i_m $, for $m\in\mathbb{Z}\setminus\{0\}$ and $i\in\{1, \ldots, d\}$, and the {central element} $\mathfrak{c}$ satisfying the relations
\begin{equation}\label{eq:commutationrelations}
\left\{
\begin{array}{ll}
\left[\mathfrak{q}^i_m,\mathfrak{c}\right]=0 & \mbox{for} \quad m\in
{\mathbb{Z}}\setminus\{0\} \ ,\ i\in\{1, \ldots, d\}\ , \\[6pt]
\big[\mathfrak{q}^i_m,\mathfrak{q}^j_n\big]=m\, \delta_{m,-n} \,
\langle \gamma_i, \gamma_j \rangle_{\mathfrak{L}}\ \mathfrak{c} & \mbox{for}
\quad m,n\in {\mathbb{Z}}\setminus\{0\} \ ,\ i,j\in\{1, \ldots, d\}\ .
\end{array}
\right.
\end{equation}
\end{definition}
For any element $v\in \mathfrak{L}$ we define the element
$\mathfrak{q}^v_m\in\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ by linearity, with $\mathfrak{q}^i_m :=\mathfrak{q}_m^{\gamma_i}$. Set
\begin{equation}\label{eq:triangularheis}
\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^+:=\bigoplus_{m>0} \ \bigoplus_{i=1}^d \, {\mathbb{F}}\mathfrak{q}^i_m\qquad
\mbox{and} \qquad
\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^-:=\bigoplus_{m<0} \ \bigoplus_{i=1}^d \, {\mathbb{F}}\mathfrak{q}^i_m \ .
\end{equation}
Let us denote by $\mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}})$ (resp.\
$\mathcal{U}(\mathfrak{h}^\pm_{{\mathbb{F}}, \mathfrak{L}})$) the {universal enveloping
algebra} of $\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}}$ (resp.\ $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^\pm$),
i.e., the unital associative algebra over ${\mathbb{F}}$ generated by
$\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}}$ (resp.\ $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^\pm$).
We introduce some terminology similar to that used in \cite[Section 5.2.5]{book:frenkelben-zvi2004}.
\begin{definition}
For $v\in\mathfrak{L}$, define \emph{free bosonic fields} as the elements
\begin{equation}
\varphi_-^v(z):=\sum_{m=1}^\infty\,\frac{z^m}{m}\,
\mathfrak{q}^v_{-m}\qquad\mbox{and}\qquad\varphi_+^v(z):=
\sum_{m=1}^\infty\, \frac{z^{-m}}{m}\, \mathfrak{q}^v_{m}
\end{equation}
in $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^-[[z]]$ and $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^+[[z^{-1}]]$,
respectively. For $\alpha,\beta\in{\mathbb{F}}$, define the \emph{generalized
bosonic exponential}
\begin{equation}
\operatorname{V}_{\alpha,\beta}^v(z):=\exp\big(\alpha\,\varphi_-^v(z)\big)\, \exp\big(\beta\,\varphi_+^v(z)\big) =: \
:\, \exp\big(\alpha\,\varphi_-(z)+\beta\, \varphi_+(z)\big)\, :
\end{equation}
in $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}[[z,z^{-1}]]$, where the symbol $:-:$ denotes
normal ordering with respect to the decomposition $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}=
\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^-\oplus \mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^+$, i.e., all negative
generators $\mathfrak{q}_{-m}^v$ are moved to the left of all positive
generators $\mathfrak{q}_m^v$ for $m>0$. When $\beta=-\alpha$, we call
$\operatorname{V}_{\alpha,-\alpha}^v(z)$ a \emph{normal-ordered bosonic exponential}.
\label{def:bosonops}\end{definition}
\begin{remark}
The bosonic exponentials are \emph{vertex operators}, i.e., they are
uniquely characterized by their commutation relations in the
Heisenberg algebra $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$: For $v,v'\in\mathfrak{L}$ one has
\begin{equation}
\big[\mathfrak{q}_m^v,\operatorname{V}_{\alpha,\beta}^{v'}(z) \big] = \left\{
\begin{array}{ll}
\alpha\, \langle v,v'\,\rangle_\mathfrak{L}\, z^m\,\operatorname{V}_{\alpha,\beta}^{v'}(z)
& \mbox{for} \quad m>0 \ , \\[6pt]
-\beta\, \langle v,v'\,\rangle_\mathfrak{L}\, z^m\,\operatorname{V}_{\alpha,\beta}^{v'}(z)
& \mbox{for} \quad m<0 \ .
\end{array} \right.
\end{equation}
The compositions of vertex operators
$\operatorname{V}_{\alpha_1,\beta_1}^{v_1}(z_1)\cdots\operatorname{V}_{\alpha_n,\beta_n}^{v_n}(z_n)$
in $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}[[z_1^{\pm\,1},\dots,z_n^{\pm\,1}]]$ can be
easily calculated as
\begin{equation}\label{eq:VOnprod}
\prod_{i=1}^n\, \operatorname{V}_{\alpha_i,\beta_i}^{v_i}(z_i) = \bigg(\,
\prod_{1\leq j<i\leq n}\, \Big(\,1-\frac{z_i}{z_j}\,
\Big)^{-\alpha_i\, \beta_j\, \langle v_i,v_j\rangle_\mathfrak{L}}\, \bigg) \
:\, \prod_{i=1}^n\, \operatorname{V}_{\alpha_i,\beta_i}^{v_i}(z_i)\, : \ ,
\end{equation}
where the factors $(1-\frac{z_i}{z_j}\,
)^{-\alpha_i\, \beta_j\, \langle v_i,v_j\rangle_\mathfrak{L}}$ are
understood as formal power series in $\frac{z_i}{z_j}$.
\end{remark}
\begin{remark}
When $v=\gamma_i$ for $i=1, \ldots, d$, we simply denote
$\varphi_\pm^i(z):= \varphi_\pm^{\gamma_i}(z)$; if $d=1$, we further
simply write $\varphi_\pm(z)$. We use analogous notation for the generalized free boson exponentials.
\end{remark}
\begin{example}\label{ex:rankkheis}
Consider the lattice $\mathfrak{L}:={\mathbb{Z}}^k$ with the symmetric nondegenerate
bilinear form $\langle v, w \rangle_\mathfrak{L} = \sum_{i=1}^k \,
v_i\,w_i$. In this case $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ is called the {\em
Heisenberg algebra of rank $k$} over ${\mathbb{F}}$, and we denote it by $\mathfrak{h}_{\mathbb{F}}^k$. It is generated by elements $\mathfrak{p}^i_m$, $m\in{\mathbb{Z}}\setminus \{0\}$, $i=1,\ldots,k$, and the central element $\mathfrak{c}$ satisfying the relations \eqref{eq:commutationrelations} with $\langle \gamma_i, \gamma_j \rangle_{\mathfrak{L}}=\delta_{ij}$. When $k=1$, $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ is simply the infinite-dimensional \emph{Heisenberg algebra} $\mathfrak{h}_{\mathbb{F}}$ over the field ${\mathbb{F}}$.
\end{example}
\begin{example}\label{ex:latticeheisenebrg}
Fix an integer $k\geq 2$ and let $\mathfrak{Q}$ be the {root lattice} of
type $A_{k-1}$ endowed with the standard bilinear form $\langle
-,-\rangle_{\mathfrak{Q}}$ (see Remark \ref{rmk:rootlattice} below). Let
$\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$ be the lattice Heisenberg algebra over ${\mathbb{F}}$
associated to $\mathfrak{Q}$; we call $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$ the \emph{Heisenberg algebra of type} $A_{k-1}$ over ${\mathbb{F}}$. It can be realized as the Lie algebra over ${\mathbb{F}}$ generated by $\mathfrak{q}_m^i$ for $m\in{\mathbb{Z}}\setminus \{ 0 \}$, $i=1,\ldots, k-1$, and the central element $\mathfrak{c}$ satisfying the relations
\begin{equation}
\left\{
\begin{array}{ll}
\left[ \mathfrak{q}_m^i, \mathfrak{c} \right] = 0 \quad & \mbox{for}\quad
m\in{\mathbb{Z}}\setminus \{ 0 \} \ , \ i=1,\ldots, k-1\ ,\\[6pt]
\big[ \mathfrak{q}_m^i, \mathfrak{q}_n^j \big] = m\ \delta_{m,-n}\ C_{ij}\ \mathfrak{c}
\quad & \mbox{for}\quad m\in{\mathbb{Z}}\setminus \{ 0 \} \ , \ i,j=1,\ldots, k-1\ ,
\end{array}
\right.
\end{equation}
where $C=(C_{ij})$ is the Cartan matrix of type $A_{k-1}$.
\end{example}
\subsubsection{Virasoro generators}\label{sec:virasoro-heisenberg}
We construct the Viraroso algebra associated with the Heisenberg algebra $\mathfrak{h}_{{\mathbb{F}}}$. Define elements
\begin{align}
L^\mathfrak{h}_0=\sum_{m=1}^\infty\, \mathfrak{q}_{-m}\ \mathfrak{q}_{m} \qquad \mbox{and} \qquad
L^\mathfrak{h}_n = \frac{1}{2}\ \sum_{m\in{\mathbb{Z}}}\, \mathfrak{q}_{-m}\ \mathfrak{q}_{m+n}\quad
\mbox{for}\quad n\in{\mathbb{Z}}\setminus\{0\}
\end{align}
in the completion of the enveloping algebra $\mathcal{U}(\mathfrak{h}_{{\mathbb{F}}})$, where we set $\mathfrak{q}_0:=0$. They satisfy the relations
\begin{equation}
\big[L^\mathfrak{h}_n, L^\mathfrak{h}_m \big]=(n-m)\, L^\mathfrak{h}_{n+m}+\frac{n}{12}\,\big(n^2-1 \big)\, \delta_{m+n,0}\,\mathfrak{c}\ ,
\end{equation}
hence $\mathfrak{c}$ and $L^\mathfrak{h}_n$ with $n\in{\mathbb{Z}}$ generate a Virasoro
algebra $\mathfrak{Vir}_{\mathbb{F}}$ over ${\mathbb{F}}$.
\begin{remark}\label{rem:genbosprimary}
It is well-known (see Appendix~\ref{app:virasoroprimary}) that the
generalized bosonic exponential $\operatorname{V}_{\alpha,\beta}(z)$ is a primary
field of the Virasoro algebra $\mathfrak{Vir}_{\mathbb{F}}$ generated by $L_n^\mathfrak{h}$
with conformal dimension $\Delta(\alpha,\beta)=-\frac12\,
\alpha\,\beta$, i.e., it satisfies the commutation relations
\begin{equation}
\big[L_n^\mathfrak{h},\operatorname{V}_{\alpha,\beta}(z)\big] = z^n\,
\big(z\,\partial_z+\Delta(\alpha,\beta)\, (n+1) \big)\, \operatorname{V}_{\alpha,\beta}(z) \ .
\end{equation}
\end{remark}
\subsubsection{Fock space}
We are interested in a special type of representation of a given
lattice Heisenberg algebra $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ over ${\mathbb{F}}$.
\begin{definition}\label{def:fock}
Let $W$ be the trivial representation of $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^{+}$
(i.e., the one-dimensional ${\mathbb{F}}$-vector space with trivial
$\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^{+}$-action). The \emph{Fock space} representation
of the Heisenberg algebra $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}$ is the induced representation $\mathcal{F}_{{\mathbb{F}},\mathfrak{L}}:=\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}\otimes_{\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^+} W$.
\end{definition}
The Fock space is an irreducible \emph{highest weight representation} whereby any element $w_0\in W$ is a \emph{highest weight vector}, i.e., $\mathfrak{h}_{{\mathbb{F}},\mathfrak{L}}^{+}$ annihilates $w_0$ and the elements in $W$ of the form $\mathfrak{q}^v_{-m_1}\cdots \mathfrak{q}^v_{-m_l}\triangleright w_0$ generate $\mathcal{F}_{{\mathbb{F}},\mathfrak{L}}$ for $v\in \mathfrak{L}$, $l\geq 1$ and $m_i\geq 1$ for $i=1, \ldots, l$.
\begin{example}
For the Heisenberg algebra $\mathfrak{h}_{\mathbb{F}}$, the Fock space
$\mathcal{F}_{\mathbb{F}}$ is isomorphic to the polynomial algebra
$\Lambda_{\mathbb{F}}={\mathbb{F}}\left[p_1, p_2, \ldots \right]$ in the power sum
symmetric functions introduced in Section \ref{sec:symmetric}. In this realization, the actions of the generators are given for $m>0$ by
\begin{equation}\label{eq:fockheisenberg}
\mathfrak{p}_{-m}\triangleright f:=p_mf \ , \qquad \mathfrak{p}_m\triangleright
f:=m\, \frac{\partial f}{\partial p_m} \qquad \mbox{and} \qquad \mathfrak{c}\triangleright f:=f
\end{equation}
for any $f\in \Lambda_{{\mathbb{F}}}$.
\end{example}
\begin{example}
The Fock space ${\mathcal F}^k_{\mathbb{F}}$ of the rank $k$ Heisenberg algebra $\mathfrak{h}_{\mathbb{F}}^k$ can be realized as the tensor product of $k$ copies of the polynomial algebra $\Lambda_{\mathbb{F}}$:
\begin{equation}
{\mathcal F}^k_{\mathbb{F}} \simeq \Lambda_{\mathbb{F}}^{\otimes k}\ .
\end{equation}
In this realization, the action of the generators $\mathfrak{p}_m^i$ is obvious: each copy of the Heisenberg algebra generated by $\mathfrak{p}_m^i$ for $m\in{\mathbb{Z}}\setminus\{0\}$ acts on the $i$-th factor $\Lambda_{\mathbb{F}}$ as in Equation \eqref{eq:fockheisenberg}.
\end{example}
\subsubsection{Whittaker vectors}
We give the definition of Whittaker vector for Heisenberg algebras
following \cite[Section 3]{art:christodoulopoulou2008}; in conformal
field theory it has the meaning of a \emph{coherent state}.
\begin{definition}\label{def:whittakerheisenberg}
Let $\chi\colon \mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}}^+)\to {\mathbb{F}}$ be an algebra homomorphism such that $\chi\vert_{\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}}^+}\neq 0$, and let $V$ be a $\mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}})$-module. A nonzero vector $w\in V$ is called a \emph{Whittaker vector of type} $\chi$ if $\eta\triangleright w=\chi(\eta)\, w$ for all $\eta\in \mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{L}}^+)$.
\end{definition}
\begin{remark}\label{rmk:unicitawhittakerheis}
By \cite[Proposition~10]{art:christodoulopoulou2008}, if $w,w'$
are Whittaker vectors of the same type $\chi$, then $w'=\lambda \, w$ for some nonzero $\lambda\in {\mathbb{F}}$.
\end{remark}
\subsection{Affine algebra of type $\widehat{A}_{k-1}$}
Let $k\geq 2$ be an integer and let $\mathfrak{sl}_k:=\mathfrak{sl}(k,{\mathbb{F}})$ denote the finite-dimensional Lie algebra of rank
$k-1$ over ${\mathbb{F}}$
generated in the Chevalley basis by $E_i, F_i, H_i$ for $i=1, \ldots, k-1$ satisfying the relations
\begin{equation}
\begin{array}{ll}
\left[E_i, F_j\right] = \delta_{ij} \, H_j\ , & [H_i, H_j]=0\ ,\\[4pt]
\left[H_i, E_j\right] = C_{ij} \, E_j\ , & [H_i, F_j]=-C_{ij}\, F_j\ ,
\end{array}
\end{equation}
where $C=(C_{ij})$ is the Cartan matrix type $A_{k-1}$ (see Remark \ref{rmk:rootlattice} below).
An explicit realization of the generators of $\mathfrak{sl}_k$ in the algebra
$M(k,{\mathbb{F}})$ of $k\times k$ matrices over ${\mathbb{F}}$ is given in the following way. Let
$\boldsymbol{E}_{i,j}$ denote the $k\times k$ matrix unit with 1 in
the $(i,j)$ entry and 0 everywhere else for $i,j=1, \ldots, k$. Define
\begin{equation}
E_i:=\boldsymbol{E}_{i,i+1}\ ,\qquad F_i:= \boldsymbol{E}_{i+1,i}
\qquad \mbox{and} \qquad H_i:= \boldsymbol{E}_{i,i}- \boldsymbol{E}_{i+1,i+1}
\end{equation}
for $i=1, \ldots, k-1$. One sees immediately that $E_i, F_i, H_i$
satisfy the defining relations for $\mathfrak{sl}_k$.
Let us denote by $\mathfrak{t}$ the Lie subalgebra of $\mathfrak{sl}_k$
generated by $H_i$ for $i=1, \ldots, k-1$ and by $\mathfrak{n}_+$
(resp.\ $\mathfrak{n}_{-}$) the Lie subalgebra of $\mathfrak{sl}_k$
generated by $E_i$ (resp.\ $F_i$) for $i=1, \ldots, k-1$. Then there
is a {triangular decomposition}
\begin{equation}
\mathfrak{sl}_k=\mathfrak{n}_{-}\oplus\mathfrak{t}\oplus\mathfrak{n}_+
\end{equation}
as a direct sum of vector spaces.
\begin{remark}\label{rmk:rootlattice}
For $i=1, \ldots, k$, define $\boldsymbol{e}_i\in \mathfrak{t}^\ast$ by
\begin{equation}
\boldsymbol{e}_i\big(\mathrm{diag}(a_1, \ldots, a_k) \big)=a_i \ .
\end{equation}
The elements $\gamma_i:=\boldsymbol{e}_i-\boldsymbol{e}_{i+1}$ for
$i=1, \ldots, k-1$ form a basis of $\mathfrak{t}^\ast$. The
\emph{root lattice} $\mathfrak{Q}$ is the lattice
$\mathfrak{Q}:=\bigoplus_{i=1}^{k-1}\, {\mathbb{Z}}\gamma_i$. The elements of
$\mathfrak{Q}$ are called \emph{roots}, and in particular $\gamma_i$ are called the
\emph{simple roots}. The lattice of positive roots is
$\mathfrak{Q}_+:=\bigoplus_{i=1}^{k-1} \, {\mathbb{N}} \gamma_i$. Since $\boldsymbol{e}_i$ corresponds to the $i$-th
coordinate vector in ${\mathbb{Z}}^{k}$, there is a description of $\mathfrak{Q}$ and
$\mathfrak{Q}_+$ in ${\mathbb{Z}}^k$ given by
\begin{equation}
\mathfrak{Q}=\big\{\boldsymbol{e}_{i}-\boldsymbol{e}_{j}\, \big\vert\, i,j=1, \ldots,
k \big\}\qquad \mbox{and} \qquad
\mathfrak{Q}_+=\big\{\boldsymbol{e}_{i}-\boldsymbol{e}_{j}\, \big\vert\,
1\leq i<j\leq k \big\}\ .
\end{equation}
By setting $\langle \gamma_i, \gamma_j \rangle_\mathfrak{Q}:=\gamma_i(H_j)=C_{ij}$, we define a nondegenerate symmetric bilinear form $\langle-,- \rangle_\mathfrak{Q}$ on $\mathfrak{Q}$.
The \emph{fundamental weights $\omega_i$ of type $A_{k-1}$} are the elements of $\mathfrak{t}^\ast$ defined by $\omega_i(H_j)=\delta_{ij}$ for $i,j=1, \ldots, k-1$. In the standard basis of ${\mathbb{Z}}^k$, they are given explicitly by
\begin{equation}
\omega_i:=\sum_{l=1}^{i} \, \boldsymbol{e}_l-\frac{i}{k}\,
\sum_{l=1}^{k}\, \boldsymbol{e}_l
\end{equation}
for $i=1,\dots,k-1$. Let $\mathfrak{P}:=\bigoplus_{i=1}^{k-1}\, {\mathbb{Z}} \omega_i$
be the \emph{weight lattice}. Then $\mathfrak{Q}\subset \mathfrak{P}$, as
$\gamma_i=\sum_{j=1}^{k-1} \, C_{ij}\, \omega_j$. The set of
\emph{dominant weights} is $\mathfrak{P}_+:=\bigoplus_{i=1}^{k-1}\, {\mathbb{N}}
\omega_i$. There is a coset decomposition of $\mathfrak{P}$ given by
\begin{equation}
\mathfrak{P}=\bigcup_{j=0}^{k-1}\, (\mathfrak{Q}+\omega_j)\ ,
\label{eq:lateralweightlattice}\end{equation}
where we set $\omega_0:=\boldsymbol{0}$.
The \emph{coroot lattice} is the lattice $\mathfrak{Q}^\vee:=\bigoplus_{i=1}^{k-1}\, {\mathbb{Z}} H_i$.
\end{remark}
We now introduce the Kac-Moody algebra $\widehat{\mathfrak{sl}}_k$ of type $\widehat{A}_{k-1}$, first via its {canonical generators} and then as a central extension of the {loop algebra} of $\mathfrak{sl}_k$.
\begin{definition}
The \emph{Kac-Moody algebra} $\widehat{\mathfrak{sl}}_k$ \emph{of type} $\widehat{A}_{k-1}$ over ${\mathbb{F}}$ is the Lie algebra over ${\mathbb{F}}$ generated by $e_i, f_i, h_i$ for $i=0,1, \ldots, k-1$ satisfying the relations
\begin{equation}
\begin{array}{ll}
\left[e_i, f_j\right]= \delta_{ij} \, h_j\ ,& [h_i, h_j]=0\ ,\\[4pt]
\left[h_i, e_j\right] = \widehat{C}_{ij} \, e_j\ ,& [h_i,
f_j]=-\widehat{C}_{ij} \, f_j\ ,
\end{array}
\end{equation}
where $\widehat{C}=\big(\widehat{C}_{ij} \big)$ is the Cartan matrix of the extended Dynkin diagram of type $\widehat{A}_{k-1}$.
\end{definition}
The matrix $\widehat{C}$ is given for $k\geq 3$ by
\begin{equation}
\widehat{C}=\big(\widehat{C}_{ij} \big)=\begin{pmatrix}
2 & -1 & 0 & \dots & -1 \\
-1 & 2& -1& \dots & 0 \\
0 & -1& 2& \dots & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
-1 &0 &0 & \dots & 2
\end{pmatrix}
\end{equation}
and for $k=2$ by
\begin{equation}
\widehat{C}=\big(\widehat{C}_{ij} \big)=\begin{pmatrix}
2 & -2 \\
-2 & 2
\end{pmatrix}.
\end{equation}
Let us denote by $\widehat{\mathfrak{t}}$ the Lie subalgebra of $\widehat{\mathfrak{sl}}_k$
generated by $h_i$ for $i=0, 1, \ldots, k-1$ and by $\widehat{\mathfrak{n}}_+$
(resp.\ $\widehat{\mathfrak{n}}_{-}$) the Lie subalgebra of $\widehat{\mathfrak{sl}}_k$
generated by $e_i$ (resp.\ $f_i$) for $i=0, 1, \ldots, k-1$. Then
there is a {triangular decomposition}
\begin{equation}\label{eq:triangularslk}
\widehat{\mathfrak{sl}}_k=\widehat{\mathfrak{n}}_{-}\oplus\widehat{\mathfrak{t}}\oplus\widehat{\mathfrak{n}}_+
\end{equation}
as a direct sum of vector spaces.
Now we describe the relation between $\mathfrak{sl}_k$ and
$\widehat{\mathfrak{sl}}_k$. Define in $\mathfrak{sl}_k$ the elements
\begin{equation}
E_0:=\boldsymbol{E}_{k,1}\ ,\qquad F_0:= \boldsymbol{E}_{1,k}\qquad \mbox{and} \qquad
H_0:=\boldsymbol{E}_{k,k}- \boldsymbol{E}_{1,1} \ .
\end{equation}
Consider next the \emph{loop algebra} $\widetilde{\mathfrak{sl}}_k:=\mathfrak{sl}_k\otimes {\mathbb{F}}[t,t^{-1}]$. Set
\begin{equation}
\begin{array}{ll}
\tilde{e}_0:=E_0\otimes t\ , & \tilde{e}_i:=E_i\otimes 1\ ,\\[4pt]
\tilde{f}_0:=F_0\otimes t^{-1}\ , & \tilde{f}_i:=F_i\otimes 1\ , \\[4pt]
\tilde{h}_0:= H_0\otimes 1\ , & \tilde{h}_i:=H_i\otimes 1\ ,
\end{array}
\end{equation}
for $i=1, \ldots, k-1$. Let us denote by $\mathfrak{c}$ the {central element} of $\widehat{\mathfrak{sl}}_k$
given by $\mathfrak{c}=\sum_{i=0}^{k-1} \, h_i$. Then we can realize $\widehat{\mathfrak{sl}}_k$ as a one-dimensional central extension
\begin{equation}
0\ \longrightarrow \ {\mathbb{F}} \mathfrak{c}\ \longrightarrow \ \widehat{\mathfrak{sl}}_k \
\xrightarrow{ \ \pi \ } \ \widetilde{\mathfrak{sl}}_k \ \longrightarrow \ 0\ ,
\end{equation}
where the homomorphism $\pi$ is defined by
\begin{equation}
\pi\, :\, e_i\ \longmapsto \tilde{e}_i\ ,\qquad f_i \ \longmapsto\
\tilde{f}_i\ ,\qquad h_i\ \longmapsto \ \tilde{h}_i\ ,
\end{equation}
for $i=0, 1, \ldots, k-1$, and the Lie algebra structure of $\widehat{\mathfrak{sl}}_k$ is obtained through
\begin{equation}\label{eq:algebrastructslk}
[M\otimes t^m, N\otimes t^n]=[M, N]\otimes t^{m+n} +m \, \delta_{m,
-n}\,\mathrm{tr}(M\, N) \ \mathfrak{c}
\end{equation}
for every $M, N\in \mathfrak{sl}_k$ and $m,n\in\mathbb{Z}$. Thus the canonical generators of $\widehat{\mathfrak{sl}}_k$ are
\begin{equation}
\begin{array}{ll}
e_0:= E_0\otimes t\ ,& e_i:=E_i\otimes 1 \ ,\\[4pt]
f_0:= F_0\otimes t^{-1}\ , & f_i:= F_i\otimes 1\ ,\\[4pt]
h_0:= H_0\otimes 1 +\mathfrak{c}\ ,& h_i:=H_i\otimes 1 \ ,
\end{array}
\end{equation}
and we can realize $\widehat{\mathfrak{t}}$ as the one-dimensional extension
\begin{equation}
0\ \longrightarrow \ {\mathbb{F}} \mathfrak{c}\ \longrightarrow \ \widehat{\mathfrak{t}} \
\xrightarrow{ \ \pi \ } \ \mathfrak{t} \ \longrightarrow \ 0\ .
\end{equation}
\begin{remark}
Let $\gamma_0:=-\sum_{i=1}^{k-1} \gamma_i$. For $i=1,
\ldots, k-1$, let $\boldsymbol{e}_i$ be as in Remark
\ref{rmk:rootlattice}; then $\gamma_0=\boldsymbol{e}_k-\boldsymbol{e}_1$. We extend $\boldsymbol{e}_i$ from
$\mathfrak{t}^\ast$ to $\widehat{\mathfrak{t}}^\ast$ by setting
$\boldsymbol{e}_i(\mathfrak{c})=0$. Then $\gamma_i(\mathfrak{c})=0$ for $i=0,1, \ldots, k-1$. Thus the root lattice $\widehat{\mathfrak{Q}}$ of
$\widehat{\mathfrak{sl}}_k$ is the lattice $\widehat\mathfrak{Q}= \bigoplus_{i=0}^{k-1} \, {\mathbb{Z}}\gamma_i = {\mathbb{Z}}\gamma_0 \oplus \mathfrak{Q}$. In a similar way, one can define the lattice of positive roots and a nondegenerate symmetric bilinear form on $\widehat{\mathfrak{Q}}$.
Let $\widehat{\omega}_0$ be the element in $\widehat{\mathfrak{t}}^\ast$ defined by $\widehat{\omega}_0(\mathfrak{t}^\ast)=0$ and $\widehat{\omega}_0(\mathfrak{c})=1$. Define
\begin{equation}
\widehat{\omega}_i:=\omega_i+\widehat{\omega}_0\qquad\mbox{for} \quad i=1, \ldots, k-1\ .
\end{equation}
We call $\widehat{\omega}_0, \widehat{\omega}_1, \ldots,
\widehat{\omega}_{k-1}$ the \emph{fundamental weights of type
$\widehat{A}_{k-1}$}. Set $\widehat{\mathfrak{P}}:=\bigoplus_{i=0}^{k-1}\, {\mathbb{Z}}
\widehat{\omega}_i$. Any weight $\widehat{\lambda}=\sum_{i=0}^{k-1}\,
\lambda_i \, \widehat{\omega}_i\in \widehat{\mathfrak{P}}$ can be written as
$\widehat{\lambda}=\lambda+k_{\,\widehat{\lambda}}\,
\widehat{\omega}_0$, where $\lambda\in \mathfrak{P}$ and
$k_{\,\widehat{\lambda}}=\widehat{\lambda}(\mathfrak{c})=\sum_{i=0}^{k-1} \,
\lambda_i$ is the \emph{level} of $\widehat{\lambda}$.
\end{remark}
\subsubsection{Highest weight representations}
By declaring the degrees of generators $\deg e_i= -\deg f_i =1$ and $\deg h_i=0$ for $i=0,1, \ldots, k-1$, we endow $\widehat{\mathfrak{sl}}_k$ with the \emph{principal grading}
\begin{equation}
\widehat{\mathfrak{sl}}_k= \bigoplus_{n\in {\mathbb{Z}}} \, \big(\, \widehat{\mathfrak{sl}}_k \big)_n \ .
\end{equation}
The principal grading of
$\widehat{\mathfrak{sl}}_k$ induces a ${\mathbb{Z}}$-grading of its universal enveloping
algebra $\mathcal{U}\big(\, \widehat{\mathfrak{sl}}_k \big)$ over ${\mathbb{F}}$, which is written
as
\begin{equation}
\mathcal{U}\big(\, \widehat{\mathfrak{sl}}_k \big)=\bigoplus_{n\in {\mathbb{Z}}} \, \mathcal{U}_n\ .
\end{equation}
Set $\widehat{\mathfrak{b}}:=\widehat{\mathfrak{t}}\oplus\widehat{\mathfrak{n}}_+$. Let $\widehat{\lambda}$ be a linear
form on $\widehat{\mathfrak{t}}$. We define a one-dimensional $\widehat{\mathfrak{b}}$-module
${\mathbb{F}} v_{\, \widehat{\lambda}}$ by
\begin{equation}
\widehat{\mathfrak{n}}_+\triangleright v_{\, \widehat{\lambda}}=0 \qquad\mbox{and}\qquad
h_i\triangleright v_{\, \widehat{\lambda}}=\widehat{\lambda}(h_i)\, v_{\, \widehat{\lambda}} \qquad\mbox{for}\quad i=0,1, \ldots, k-1\ .
\end{equation}
Consider the induced $\widehat{\mathfrak{sl}}_k$-module
\begin{equation}
\tilde{{\mathcal V}}(\,\widehat{\lambda}\, ):=\mathcal{U}\big(\,\widehat{\mathfrak{sl}}_k \big)\otimes_{\mathcal{U}(\,\widehat{\mathfrak{b}}\,
)} {\mathbb{F}} v_{\,\widehat{\lambda}}\ .
\end{equation}
Setting $\tilde{{\mathcal V}}_n:=\mathcal{U}_n\triangleright v_{\,\widehat{\lambda}}$, we define
the principal grading $\tilde{{\mathcal V}}(\,\widehat{\lambda}\,
)=\bigoplus_{n\in {\mathbb{Z}}}\, \tilde{{\mathcal V}}_n$. The $\widehat{\mathfrak{sl}}_k$-module
$\tilde{{\mathcal V}}(\,\widehat{\lambda}\, )$ contains a unique maximal
proper graded $\widehat{\mathfrak{sl}}_k$-submodule $I(\,\widehat{\lambda}\, )$.
\begin{definition}\label{def:basicrepslk}
The quotient module
\begin{equation}
{\mathcal V}(\,\widehat{\lambda}\,):=\tilde{{\mathcal V}}(\,\widehat{\lambda}\,
)\,\big/\, I(\,\widehat{\lambda}\, )
\end{equation}
is called the \emph{highest weight representation of $\widehat{\mathfrak{sl}}_k$ at level $k_{\,\widehat{\lambda}}$}. The nonzero multiples of the image of $v_{\,\widehat{\lambda}}$ in
${\mathcal V}(\,\widehat{\lambda}\, )$ are called the \emph{highest weight vectors} of
${\mathcal V}(\,\widehat{\lambda}\, )$.
\end{definition}
The principal grading on $\tilde{{\mathcal V}}(\,\widehat{\lambda}\, )$ induces an ${\mathbb{N}}$-grading
\begin{equation}
{\mathcal V}(\, \widehat{\lambda}\, )=\bigoplus_{n\geq 0}\, {\mathcal V}_{-n}
\end{equation}
called the principal grading of ${\mathcal V}(\,\widehat{\lambda}\, )$.
\begin{definition}
The \emph{$i$-th dominant representation of $\widehat{\mathfrak{sl}}_k$ at level
one} is the highest weight representation
${\mathcal V}(\,\widehat{\omega}_i)$ of $\widehat{\mathfrak{sl}}_k$ for $i=0,1, \ldots,
k-1$. The module ${\mathcal V}(\,\widehat{\omega}_0)$ is also called the \emph{basic representation of $\widehat{\mathfrak{sl}}_k$}.\\
\end{definition}
\begin{remark}
One can define the Lie algebra $\widehat{\mathfrak{gl}}_k$ as the one-dimensional extension
\begin{equation}
0\ \longrightarrow \ {\mathbb{F}} \mathfrak{c}\ \longrightarrow \ \widehat{\mathfrak{gl}}_k \
\xrightarrow{ \ \pi \ } \ \mathfrak{gl}_k\otimes{\mathbb{F}}[t,t^{-1}] \ \longrightarrow \ 0\ .
\end{equation}
Since $\mathfrak{gl}_k={\mathbb{F}}\,\mathrm{id}\oplus\mathfrak{sl}_k$, the
representation theory of $\widehat{\mathfrak{gl}}_k$ is obtained by combining the
representation theory of the Heisenberg algebra $\mathfrak{h}_{\mathbb{F}}$ with that of
$\widehat{\mathfrak{sl}}_k$. For example, all highest weight representations of
$\widehat{\mathfrak{gl}}_k$ are of the form ${\mathcal F}_{{\mathbb{F}}}\otimes {\mathcal V}(\, \widehat
\lambda\, )$ for some weight $\widehat \lambda\in \widehat \mathfrak{P}$.
\end{remark}
\subsubsection{Whittaker vectors}
Let us denote by $\mathfrak{q}_m^i$ the element $H_i\otimes t^m$ for $i\in\{1, \ldots, k-1\}$ and $m\in{\mathbb{Z}}$. By Equation \eqref{eq:algebrastructslk}, these elements satisfy
\begin{equation}
\big[\mathfrak{q}_m^i, \mathfrak{q}_n^j\big]=m\ \delta_{m+n, 0}\ C_{ij}\ \mathfrak{c} \qquad\mbox{and}\qquad \big[\mathfrak{q}_m,\mathfrak{c}\big]=0\ ,
\end{equation}
for $i,j\in\{1, \ldots, k-1\}$ and $m,n\in{\mathbb{Z}}$. For a root $\gamma$, we
denote by $\mathfrak{q}_m^\gamma$ the element $H_\gamma\otimes t^m$ where
$H_\gamma\in\mathfrak{t}$ is defined by $\langle H,
H_\gamma\rangle_{\mathfrak{Q}^\vee\otimes_{\mathbb{Z}} {\mathbb{R}}} = \gamma(H)$ for any $H\in \mathfrak{t}$.
The subalgebra of $\widehat{\mathfrak{sl}}_k$ generated by $\mathfrak{q}_m^i$, for $i\in\{1, \ldots, k-1\}$ and $m\in{\mathbb{Z}}\setminus\{0\}$, and $\mathfrak{c}$ is isomorphic to the
Heisenberg algebra $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$. This motivates the following
definition of Whittaker vector for $\widehat{\mathfrak{sl}}_k$ (cf.\ \cite[Section 6]{art:christodoulopoulou2008}).
\begin{definition}\label{def:whittakersl}
Let $\chi\colon \mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{Q}}^+)\to {\mathbb{F}}$ be an algebra
homomorphism such that $\chi \vert_{\mathfrak{h}_{{\mathbb{F}}, \mathfrak{Q}}^+}\neq 0$, and let $V$ be a $\mathcal{U}\big(\, \widehat{\mathfrak{sl}}_k \big)$-module. A nonzero vector $w\in V$ is called a \emph{Whittaker vector of type} $\chi$ if $\eta\triangleright w=\chi(\eta)\, w$ for all $\eta\in \mathcal{U}(\mathfrak{h}_{{\mathbb{F}}, \mathfrak{Q}}^+)$.
\end{definition}
\subsection{Frenkel-Kac construction}\label{sec:frenkelkac}
Let ${\mathcal V}$ be a representation of $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$. We say that
it is a \emph{level one} representation if the central element $\mathfrak{c}$ acts by the
identity map. Henceforth we let ${\mathcal V}$ be a level one representation of $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$ such that for any $v\in {\mathcal V}$ there exists an integer $m(v)$ for which
\begin{equation}\label{eq:condition-fock}
\big(\mathfrak{q}^{l_1}_{m_1}\cdots \mathfrak{q}^{l_a}_{m_a}\big) \triangleright v=0
\end{equation}
if $m_i>0$ and $\sum_i\, m_i>m(v)$.
Fix an index $j\in\{0,1, \ldots, k-1\}$ and consider the coset $\mathfrak{Q}+\omega_j$. Denote by ${\mathbb{F}}[\mathfrak{Q}+\omega_j]$ the group algebra of $\mathfrak{Q}+\omega_j$
over ${\mathbb{F}}$. For a root $\gamma\in \mathfrak{Q}$ we define the generating function
$V(\gamma,z)\in \mathrm{End}({\mathcal V}\otimes {\mathbb{F}}[\mathfrak{Q}+\omega_j])[[z,z^{-1}]]$ of
operators on ${\mathcal V}\otimes {\mathbb{F}}[\mathfrak{Q}+\omega_j]$ by the
{bosonic vertex operator}
\begin{align}
V(\gamma,z)
&=\operatorname{V}_{1,-1}^\gamma(z) \, \exp(\log z\ \mathfrak{c}+\gamma)\\[4pt] & =\exp\Big(\sum_{m=1}^\infty\, \frac{z^m}{m}\,
\mathfrak{q}^\gamma_{-m}\, \Big)\, \exp\Big(-\sum_{m=1}^\infty\,
\frac{z^{-m}}{m}\, \mathfrak{q}^\gamma_{m}\Big)\, \exp(\log z\ \mathfrak{c}+\gamma)\ ,
\end{align}
where $\exp(\log z\ \mathfrak{c}+\gamma)$ is the operator defined by
\begin{equation}
\exp(\log z\ \mathfrak{c}+\gamma)\triangleright(v \otimes
[\beta+\omega_j]):=z^{\frac12\langle \gamma,\gamma\rangle_{\mathfrak{Q}} + \langle
\gamma, \beta+\omega_j\rangle_{\mathfrak{Q}}}\, (v \otimes [\beta+\gamma+\omega_j])
\end{equation}
for $v \otimes [\beta+\omega_j] \in {\mathcal V}\otimes
{\mathbb{F}}[\mathfrak{Q}+\omega_j]$.
\begin{remark}
Here for the operator $\exp(\log z\ \mathfrak{c}+\gamma)$ we follow the notation in \cite[Section 3.2.1]{art:nagao2009}. In the existing literature, this operator is denoted in various different ways.
\end{remark}
Let $V_m(\gamma)\in {\rm End}({\mathcal V}\otimes
{\mathbb{F}}[\mathfrak{Q}+\omega_j])$
denote the operator defined by the formal Laurent series expansion $V(\gamma,z)=\sum_{m\in{\mathbb{Z}}} \,
V_m(\gamma)\, z^m$.
We define a map $\epsilon\colon \mathfrak{Q}\times \mathfrak{Q}\to \{\pm 1\}$ by
\begin{equation}
\epsilon(\gamma_i,\gamma_j)=\left\{
\begin{array}{ll}
-1 \ , & j=i, i+1\ ,\\
1 \ , & \mbox{otherwise} \ ,
\end{array}\right.
\end{equation}
with the properties
$\epsilon(\gamma+\gamma',\beta)=\epsilon(\gamma,\beta)\, \epsilon(\gamma',\beta)$
and $\epsilon(\gamma,\beta+\beta'\,
)=\epsilon(\gamma,\beta)\, \epsilon(\gamma,\beta'\, )$.
\begin{theorem}[{\cite[Theorem 1]{art:frenkelkac1980}}]\label{thm:frenkelkac}
Let $j\in\{0,1, \ldots, k-1\}$ and let ${\mathcal V}$ be a level one representation of $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$ satisfying the condition \eqref{eq:condition-fock}. Then the vector space ${\mathcal V} \otimes {\mathbb{F}}[\mathfrak{Q}+\omega_j]$ carries a level one $\widehat{\mathfrak{sl}}_k$-module structure given by
\begin{align}
(H_i\otimes 1)\triangleright(v \otimes [\beta+\omega_j])&=\big(\langle
\gamma_i,\beta\rangle_\mathfrak{Q}+\delta_{ij}\big)\, (v \otimes [\beta+\omega_j])\ ,\\[4pt]
(H_i\otimes t^m)\triangleright(v \otimes [\beta+\omega_j])&= \big(\mathfrak{q}^i_m\triangleright v\big) \otimes [\beta+\omega_j]\ ,\\[4pt]
(E_i\otimes t^m)\triangleright(v \otimes
[\beta+\omega_j])&=\epsilon(\gamma_i,\beta)\, V_{m+\delta_{ij}}(\gamma_i)\triangleright(v \otimes [\beta+\omega_j])\ ,\\[4pt]
(F_i\otimes t^m)\triangleright(v \otimes [\beta+\omega_j])&=
\epsilon(\beta,\gamma_i)\, V_{-m-\delta_{ij}}(-\gamma_i)\triangleright(v \otimes [\beta+\omega_j])\ ,
\end{align}
for $i\in\{1, \ldots, k-1\}$ and $m\in{\mathbb{Z}}\setminus\{0\}$. If ${\mathcal V}$ is
the Fock space of $\mathfrak{h}_{{\mathbb{F}},\mathfrak{Q}}$, then ${\mathcal V} \otimes {\mathbb{F}}[\mathfrak{Q}+\omega_j]$ is the $j$-th dominant representation of $\widehat{\mathfrak{sl}}_k$.
\end{theorem}
\subsubsection{Virasoro operators}\label{sec:virasoro-affine}
Let $\{\eta_i\}_{i=1}^{ k-1}$ be an orthonormal basis of the vector
space $\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{R}}$. The Virasoro algebra associated with $\mathfrak{h}_{{\mathbb{F}}, \mathfrak{Q}}\subset\widehat{\mathfrak{sl}}_k$ has generators $\mathfrak{c}$ and $L^{\widehat{\mathfrak{sl}}_k}_n$ for $n\in{\mathbb{Z}}$ defined by \cite[Section 2.8]{art:frenkelkac1980}
\begin{align}
L^{\widehat{\mathfrak{sl}}_k}_0&=\sum_{i=1}^{k-1}\ \sum_{m=1}^\infty\, \mathfrak{q}_{-m}^{\eta_i}\
\mathfrak{q}_{m}^{\eta_i}+\frac{1}{2}\ \sum_{i=1}^{k-1}\,
\big(\mathfrak{q}_0^{\eta_i}\big)^2 \ ,\\[4pt]
L^{\widehat{\mathfrak{sl}}_k}_n &= \frac{1}{2}\ \sum_{i=1}^{k-1}\ \sum_{m\in{\mathbb{Z}}}\,
\mathfrak{q}_{-m}^{\eta_i}\ \mathfrak{q}_{m+n}^{\eta_i}\qquad \mbox{for}\quad n\in{\mathbb{Z}}\setminus\{0\}\ .
\end{align}
Note that distinct orthonormal bases of $\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{R}}$ give
rise to the
same Virasoro algebra $\mathfrak{Vir}_{\mathbb{F}}$.
\bigskip \section{AGT relations on ${\mathbb{R}}^4$\label{sec:AGTonR4}}
\subsection{Equivariant cohomology of $\hilb{n}{{\mathbb{C}}^2}$}\label{sec:C2}
In the following we shall give a brief survey of results concerning the
equivariant cohomology of the Hilbert schemes $\hilb{n}{{\mathbb{C}}^2}$ and representations of Heisenberg algebras thereon \cite{art:nakajima1996, art:grojnowski1996, book:nakajima1999, art:vasserot2001, art:liqinwang2004, art:qinwang2007, art:nakajima2014}.
Let us consider the action of the torus $T:=({\mathbb{C}}^*)^2$ on the
complex affine plane ${\mathbb{C}}^2$ given by $(t_1,t_2)\triangleright
(x,y)=(t_1\,x,t_2\,y)$, and the induced $T$-action on the
{Hilbert scheme of $n$ points} $\hilb{n}{{\mathbb{C}}^2}$ which is the
{fine} moduli space parameterizing zero-dimensional subschemes of
${\mathbb{C}}^2$ of length $n$; it is a smooth quasi-projective variety of
dimension $2n$. Following \cite{art:ellingsrudstromme1987}, the $T$-fixed points of $\hilb{n}{{\mathbb{C}}^2}$ are zero-dimensional subschemes of ${\mathbb{C}}^2$ of length $n$ supported at the origin $0\in{\mathbb{C}}^2$ which correspond to partitions $\lambda$ of $n$. We shall denote by $Z_\lambda$ the fixed point in $\hilb{n}{{\mathbb{C}}^2}^T$ corresponding to the partition $\lambda$ of $n$.
For $i=1,2$ denote by $t_i$ the $T$-modules corresponding to the
characters $\chi_i\colon (t_1,t_2)\in T \mapsto t_i \in{\mathbb{C}}^*$, and by
$\varepsilon_i$ the equivariant first Chern class of $t_i$. Then $H_T^*({\rm
pt}; {\mathbb{C}})={\mathbb{C}}[\varepsilon_1,\varepsilon_2]$ is the coefficient ring
for the $T$-equivariant cohomology. The equivariant Chern
character of the tangent space to $\hilb{n}{{\mathbb{C}}^2}$ at a fixed point
$Z_\lambda$ is given by
\begin{equation}
\mathrm{ch}_T\big(T_{Z_\lambda}\hilb{n}{{\mathbb{C}}^2} \big)=\sum_{s\in
Y_\lambda}\, \big({\,\rm e}\,^{(L(s)+1)\, \varepsilon_1-A(s)\,
\varepsilon_2}+{\,\rm e}\,^{-L(s)\, \varepsilon_1+(A(s)+1)\, \varepsilon_2}\big)\ .
\end{equation}
The equivariant Euler class is therefore given by
\begin{equation}
\operatorname{eu}_T\big(T_{Z_\lambda}\hilb{n}{{\mathbb{C}}^2} \big)=(-1)^n\,
\operatorname{eu}_+(\lambda)\, \operatorname{eu}_-(\lambda)\ ,
\end{equation}
where
\begin{equation}
\operatorname{eu}_+(\lambda)=\prod_{s\in Y_\lambda}\, \big((L(s)+1)\,
\varepsilon_1-A(s)\, \varepsilon_2\big)\qquad \mbox{and} \qquad
\operatorname{eu}_-(\lambda) =\prod_{s\in Y_\lambda}\, \big(L(s)\,
\varepsilon_1-(A(s)+1)\, \varepsilon_2\big)\ .
\end{equation}
\begin{remark}
By \cite[Corollary~3.20]{art:nakajima2014}, $\operatorname{eu}_+(\lambda)$ is the equivariant Euler class of the nonpositive part $T^{\leq 0}_{Z_\lambda}$ of the tangent space to $\hilb{n}{{\mathbb{C}}^2}$ at the fixed point $Z_\lambda$.
\end{remark}
Let $\imath_{\lambda}\colon \{Z_\lambda\}\hookrightarrow
\hilb{n}{{\mathbb{C}}^2}$ be the inclusion morphism and define the class
\begin{equation}\label{eq:fixed-class}
[\lambda]:={\imath_\lambda}_*(1) \ \in \ H^{4n}_T\big(\hilb{n}{{\mathbb{C}}^2}\big)\ .
\end{equation}
By the projection formula we get
\begin{equation}
[\lambda]\cup[\mu]=\delta_{\lambda,\mu}\,
\operatorname{eu}_T\big(T_{Z_\lambda}\hilb{n}{{\mathbb{C}}^2}
\big)\, [\lambda]=(-1)^n\, \delta_{\lambda,\mu}\,
\operatorname{eu}_+(\lambda)\, \operatorname{eu}_-(\lambda)\, [\lambda]\ .
\end{equation}
Denote
\begin{equation}
\imath_n:=\bigoplus_{Z_\lambda \in\hilb{n}{{\mathbb{C}}^2}^T}\, \imath_{\lambda}
\, \colon \, \hilb{n}{{\mathbb{C}}^2}^T\ \longrightarrow \ \hilb{n}{{\mathbb{C}}^2}\ .
\end{equation}
Let $\imath_n^!\colon H_T^\ast\left(\hilb{n}{{\mathbb{C}}^2}^T\right)_{\mathrm{loc}}\to
H_T^\ast\big(\hilb{n}{{\mathbb{C}}^2} \big)_{\mathrm{loc}}$ be the induced Gysin map, where
\begin{equation}
H_T^\ast(-)_{\mathrm{loc}}:=H_T^\ast(-)\otimes_{{\mathbb{C}}[\varepsilon_1,\varepsilon_2]}{\mathbb{C}}(\varepsilon_1,\varepsilon_2)
\end{equation}
is the localized equivariant cohomology. By the localization theorem, $\imath_n^!$ is an isomorphism and its inverse is given by
\begin{equation}
\big(\imath_n^! \big)^{-1}\, \colon \, A \ \longmapsto \ \Big(\,
\frac{\imath_{\lambda}^\ast(A)}{\operatorname{eu}_T\big(T_{Z_\lambda}\hilb{n}{{\mathbb{C}}^2}
\big)}\, \Big)_{Z_\lambda\in\hilb{n}{{\mathbb{C}}^2}^T}\ .
\end{equation}
Henceforth we denote $\mathbb{H}_{{\mathbb{C}}^2,n}:=H_T^{\ast}(\hilb{n}{{\mathbb{C}}^2})_{\mathrm{loc}}$. Define the bilinear form
\begin{align}\label{eq:bilinearform-C2}
\langle -,- \rangle_{\mathbb{H}_{{\mathbb{C}}^2,n}} \, \colon \,
\mathbb{H}_{{\mathbb{C}}^2,n}\times \mathbb{H}_{{\mathbb{C}}^2,n} & \ \longrightarrow \
{\mathbb{C}}(\varepsilon_1, \varepsilon_2)\ ,\\ \nonumber
(A, B)& \ \longmapsto \ (-1)^n \, p_n^! \, \big(\imath_n^!
\big)^{-1}(A \cup B) \ ,
\end{align}
where $p_n$ is the projection of $\hilb{n}{{\mathbb{C}}^2}^T$ to a point.
\begin{remark}
Our sign convention in defining the bilinear form is different from
the one used e.g. in \cite{book:nakajima1999,art:carlssonokounkov2012}. We choose this convention because,
under the isomorphism to be introduced later on in
\eqref{eq:isomorphism}, the form \eqref{eq:bilinearform-C2} becomes
exactly the Jack inner product \eqref{eq:jackinnerprod}. This convention produces various sign changes compared to
previous literature. Hence every time we state that a given result coincides with what is known in the literature, the reader should keep in mind ``up to the sign convention we choose".
\end{remark}
Following \cite[Section 2.2]{art:liqinwang2004} we define the distinguished classes
\begin{equation}
[\alpha_\lambda]:=\frac{1}{\operatorname{eu}_+(\lambda)}\, [\lambda] \
\in \ H^{2n}_T\big(\hilb{n}{{\mathbb{C}}^2} \big)_{\mathrm{loc}}\ .
\end{equation}
For $\lambda,\mu$ partitions of $n$ one has
\begin{align}\label{eq:product-C2}
\big\langle[\alpha_\lambda]\,,\,[\alpha_\mu]
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2,n}}&=\delta_{\lambda,\mu}\, \frac{\operatorname{eu}_-(\lambda)}{\operatorname{eu}_+(\lambda)}
\\[4pt] &=\delta_{\lambda,\mu}\, \prod_{s\in Y_\lambda} \,
\frac{L(s)\, \varepsilon_1-\big(A(s)+1 \big)\,
\varepsilon_2}{\big(L(s)+1 \big)\, \varepsilon_1-A(s)\varepsilon_2}
=\delta_{\lambda,\mu}\, \prod_{s\in Y_\lambda} \, \frac{L(s)\,
\beta+A(s)+1}{\big(L(s)+1 \big)\, \beta+A(s)}\ ,\nonumber
\end{align}
where
\begin{equation}\label{eq:beta}
\beta=-\frac{\varepsilon_1}{\varepsilon_2} \ .
\end{equation}
\begin{remark}
In \cite[Section 3(v)]{art:nakajima2014}, Nakajima gives a geometric interpretation of the class $[\alpha_\lambda]$.
\end{remark}
By the localization theorem and Equation \eqref{eq:product-C2}, the
classes $[\alpha_\lambda]$ form a
${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-basis for the infinite-dimensional vector space
$\mathbb{H}_{{\mathbb{C}}^2}:=\bigoplus_{n\geq 0} \, \mathbb{H}_{{\mathbb{C}}^2,n}$. Hence the symmetric bilinear form \eqref{eq:bilinearform-C2} is nondegenerate. The forms $\langle -,- \rangle_{\mathbb{H}_{{\mathbb{C}}^2,n}}$ define a symmetric bilinear form
\begin{equation}
\langle -,- \rangle_{\mathbb{H}_{{\mathbb{C}}^2}}\, \colon \,
\mathbb{H}_{{\mathbb{C}}^2}\times \mathbb{H}_{{\mathbb{C}}^2} \ \longrightarrow \ {\mathbb{C}}(\varepsilon_1, \varepsilon_2)
\end{equation}
by imposing that $\mathbb{H}_{{\mathbb{C}}^2,n_1}$ and $\mathbb{H}_{{\mathbb{C}}^2,n_2}$
are orthogonal for $n_1\neq n_2$. Then $\langle -,-
\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}$ is also nondegenerate.
The unique partition of $n=1$ is $\lambda=(1)$. Let us denote by $[\alpha]:=[\alpha_{(1)}]$ the corresponding class. Then
\begin{equation}
\big\langle[\alpha]\,,\,[\alpha] \big\rangle_{{\mathbb{H}}_{{\mathbb{C}}^2}}=\beta^{-1}\ .
\end{equation}
Let us denote by $D_x$ and $D_y$ respectively the $x$ and $y$ axes of
${\mathbb{C}}^2$. By localization, the corresponding equivariant cohomology
classes in $H_T^\ast({\mathbb{C}}^2)_{\mathrm{loc}}$ are given by
\begin{align}
\left[D_x\right]_T=\frac{[0]}{\varepsilon_1}=
\frac{[0]}{\operatorname{eu}_{+}(1)}=[\alpha]\qquad \mbox{and} \qquad
\left[D_y\right]_T=\frac{[0]}{\varepsilon_2}=-\beta\,
[\alpha]\ .
\end{align}
\subsection{Heisenberg algebra}
Following \cite{art:nakajima1996, book:nakajima1999}, for an integer $m>0$ define the
Hecke correspondences
\begin{equation}
D_x(n,m):= \big\{(Z,Z'\, )\in\hilb{n+m}{{\mathbb{C}}^2}\times\hilb{n}{{\mathbb{C}}^2}\
\big| \ Z'\subset Z \ , \
\mathrm{supp}(\mathcal{I}_{Z'}/\mathcal{I}_Z)=\{y\}\subset D_x \big\}\ ,
\end{equation}
where $\mathcal{I}_Z,\mathcal{I}_{Z'}$ are the ideal sheaves corresponding to $Z,Z'$ respectively. Let $q_1,q_2$ denote the projections of $\hilb{n+m}{{\mathbb{C}}^2}\times\hilb{n}{{\mathbb{C}}^2}$ to the two factors, respectively. Define linear operators $\mathfrak{p}_{-m}([D_x]_T)\in\mathrm{End}(\mathbb{H}_{{\mathbb{C}}^2})$ by
\begin{equation}
\mathfrak{p}_{-m}([D_x]_T)A :=q_1^!\big(q_2^\ast A\cup[D_x(n,m )]_T \big)
\end{equation}
for $A\in H_T^\ast(\hilb{n}{{\mathbb{C}}^2})_{\mathrm{loc}}$. We also define
$\mathfrak{p}_{m}([D_x]_T)\in\mathrm{End}(\mathbb{H}_{{\mathbb{C}}^2})$ to be the
adjoint operator of $\mathfrak{p}_{-m}([D_x]_T)$ with respect to the
inner product $\langle-,-\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}$ on
$\mathbb{H}_{{\mathbb{C}}^2}$. As the
class $[D_x]_T$ spans $H_T^\ast({\mathbb{C}}^2)_{\mathrm{loc}}$ over the field
${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$, we can define operators
$\mathfrak{p}_{m}(\eta)\in\mathrm{End}(\mathbb{H}_{{\mathbb{C}}^2})$ for every class
$\eta\in H_T^\ast({\mathbb{C}}^2)_{\mathrm{loc}}$.
\begin{theorem}[{see \cite{art:nakajima1996, art:nakajima2014}}]\label{thm:nakajimaoperators}
The linear operators $\mathfrak{p}_{m}(\eta)$, for $m\in{\mathbb{Z}}\setminus\{0\}$ and $\eta\in H_T^\ast({\mathbb{C}}^2)_{\mathrm{loc}}$, satisfy the Heisenberg commutation relations
\begin{equation}
\big[\mathfrak{p}_{m}(\eta_1)\,,\,\mathfrak{p}_{n}(\eta_2) \big]=m\,
\delta_{m,-n}\,
\langle\eta_1,\eta_2\rangle_{\mathbb{H}_{{\mathbb{C}}^2,1}}\,\mathrm{id}\qquad\mbox{and}\qquad
\big[\mathfrak{p}_{m}(\eta)\,,\,\mathrm{id} \big]=0\ .
\end{equation}
The vector space ${\mathbb{H}}_{{\mathbb{C}}^2}$ becomes the Fock space of the Heisenberg algebra $\mathfrak{h}_{{\mathbb{H}}_{{\mathbb{C}}^2,1}}$ modelled on ${\mathbb{H}}_{{\mathbb{C}}^2,1}=H_T^{\ast}({\mathbb{C}}^2)_{\mathrm{loc}}$ with the unit $\vert0\rangle$ in $H_T^0(\hilb{0}{{\mathbb{C}}^2})_{\mathrm{loc}}$ as highest weight vector.
\end{theorem}
\begin{remark}
Since $[D_x]_T=[\alpha]$, we have $\mathfrak{p}_{m}([\alpha])=\mathfrak{p}_{m}([D_x]_T)$.
\end{remark}
Henceforth we denote by $\mathfrak{h}_{{\mathbb{C}}^2}$ the Heisenberg algebra
$\mathfrak{h}_{{\mathbb{H}}_{{\mathbb{C}}^2,1}}$, and we define
\begin{equation}\label{eq:heisenberg-C2}
\mathfrak{p}_{m}:=\mathfrak{p}_{m}([D_x]_T) \qquad \mbox{for} \quad
m\in{\mathbb{Z}}\setminus \{0\} \ ,
\end{equation}
so that one has the nonzero commutation relations
\begin{equation}\label{eq:commutheisenberg-C2}
[\mathfrak{p}_{-m},\mathfrak{p}_m]=m \, \beta^{-1}\,\mathrm{id}\ .
\end{equation}
Since $[D_x]_T$ generates $H_T^\ast({\mathbb{C}}^2)_{\mathrm{loc}}$ over ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$, the operators $\mathfrak{p}_m$ generate $\mathfrak{h}_{{\mathbb{C}}^2}$.
Let $\lambda=(1^{m_1}\,2^{m_2}\,\cdots)$ be a partition. Define $\mathfrak{p}_\lambda:=\prod_i\, \mathfrak{p}_{-i}^{m_i}$. Then
\begin{equation}
\big\langle\mathfrak{p}_\lambda\vert0\rangle\,,\,\mathfrak{p}_\mu\vert0\rangle
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}=\delta_{\lambda,\mu}\, z_\lambda \, \beta^{-\ell(\lambda)}\ .
\end{equation}
Let us denote by $\Lambda_\beta$ the ring of symmetric functions in
infinitely many variables $\Lambda_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}$
over the field ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$, equiped with the
Jack inner product \eqref{eq:jackinnerprod}.
\begin{theorem}[{see \cite{art:nakajima1996, art:liqinwang2004, art:carlssonokounkov2012}}]\label{thm:actionC2}
There exists a ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear isomorphism
\begin{equation}\label{eq:isomorphism}
\phi\, \colon\, \mathbb{H}_{{\mathbb{C}}^2} \ \longrightarrow \ \Lambda_\beta
\end{equation}
preserving bilinear forms such that
\begin{equation}
\phi(\mathfrak{p}_\lambda\vert0\rangle)=p_\lambda(x) \qquad\mbox{and}\qquad \phi([\alpha_\lambda])=
J_\lambda(x;\beta^{-1})\ .
\end{equation}
Via the isomorphism $\phi$, the operators $\mathfrak{p}_m$ act on $\Lambda_\beta$
as multiplication by $p_{-m}$ for $m<0$ and as $m\, \beta^{-1}\,
\frac{\partial}{\partial p_m}$ for $m >0$.
\end{theorem}
\subsubsection{Whittaker vectors}
We characterize a particular class of Whittaker vectors
(cf. Definition \ref{def:whittakerheisenberg}) which will be useful in
our studies of gauge theories.
\begin{proposition}\label{prop:whittakerC2}
Let $\eta \in {\mathbb{C}} (\varepsilon_1,\varepsilon_2)$. In the completed Fock space $\prod_{n\geq 0}\,\mathbb{H}_{{\mathbb{C}}^2,n}$, every vector of the form
\begin{equation}
G(\eta):=\exp\left( \eta \ \mathfrak{p}_{-1}\right)\vert0\rangle
\end{equation}
is a Whittaker vector of type $\chi_\eta$, where the algebra
homomorphism $\chi_\eta\colon
\mathcal{U}(\mathfrak{h}^+_{{\mathbb{C}}^2} )\to{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ is defined by
\begin{equation}
\chi_\eta(\mathfrak{p}_1)=\eta \, \beta^{-1} \qquad
\mbox{and} \qquad \chi_\eta(\mathfrak{p}_n)=0 \qquad \mbox{for} \quad n>1 \ .
\end{equation}
\end{proposition}
\proof
The statement follows from the formal expansion
\begin{equation}\label{eq:formalexpansion}
G(\eta)= \sum_{n=0}^\infty\, \frac{\eta^n}{n!}\, (\mathfrak{p}_{-1})^n\vert 0\rangle
\end{equation}
with respect to the vector $\vert 0\rangle$, together with the relation $\mathfrak{p}_m\vert 0\rangle=0$ for $m>0$ and the identity
\begin{equation}
\mathfrak{p}_m\,(\mathfrak{p}_{-1})^n= n\, \beta^{-1}\, \delta_{m,1}\,(\mathfrak{p}_{-1})^{n-1}+ (\mathfrak{p}_{-1})^n\, \mathfrak{p}_m
\end{equation}
in $\mathcal{U}(\mathfrak{h}_{{\mathbb{C}}^2} )$ for $m\geq 1$.
\endproof
\subsection{Vertex operators}\label{sec:carlsson-okounkov}
Let $T_\mu={\mathbb{C}}^\ast$ and $H^\ast_{T_\mu}(\mathrm{pt};{\mathbb{C}})={\mathbb{C}}[\mu]$. Let us denote by ${\mathcal O}_{{\mathbb{C}}^2}(\mu)$ the trivial line bundle on ${\mathbb{C}}^2$
on which $T_\mu$ acts by scaling the fibers. In~\cite{art:carlssonokounkov2012}, Carlsson
and Okounkov define a vertex operator $\operatorname{V}({\mathcal L},z)$ for any
smooth quasi-projective surface $X$ and any line bundle ${\mathcal L}$ on
$X$. Here we shall describe only $\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)$;
see~\cite{art:carlssonokounkov2012} for a complete description of such
types of vertex operators.
Let $\boldsymbol{Z}_n\subset \hilb{n}{{\mathbb{C}}^2}\times {\mathbb{C}}^2$ be the universal
subscheme, whose fiber over a point $Z\in\hilb{n}{{\mathbb{C}}^2}$ is the
subscheme $Z\subset{\mathbb{C}}^2$ itself. Consider
\begin{equation}
{\boldsymbol{\mathcal Z}}_i:=p_{i3}^\ast({\mathcal O}_{\boldsymbol{Z}_{n_i}}) \ \in \ K\big(\hilb{n_1}{{\mathbb{C}}^2}\times
\hilb{n_2}{{\mathbb{C}}^2}\times {\mathbb{C}}^2\big) \qquad \mbox{for} \quad i=1, 2\ ,
\end{equation}
where $p_{ij}$ denotes the projection to the $i$-th and $j$-th factors. Define the virtual vector bundle
\begin{equation}
\boldsymbol{E}_\mu^{n_1, n_2}={p_{12}}_\ast\left(({\boldsymbol{\mathcal Z}}_1^\vee+{\boldsymbol{\mathcal Z}}_2-{\boldsymbol{\mathcal Z}}_1^\vee\cdot
{\boldsymbol{\mathcal Z}}_2)\cdot p_3^\ast({\mathcal O}_{{\mathbb{C}}^2}(\mu))\right) \ \in \
K\big(\hilb{n_1}{{\mathbb{C}}^2}\times \hilb{n_2}{{\mathbb{C}}^2} \big)\ ,
\end{equation}
where $p_3$ is the projection to ${\mathbb{C}}^2$. The fibre of $\boldsymbol{E}_\mu^{n_1, n_2}$ over $(Z_1, Z_2)\in
\hilb{n_1}{{\mathbb{C}}^2}^T\times \hilb{n_2}{{\mathbb{C}}^2}^T$ is given by
\begin{equation}
\boldsymbol{E}_\mu^{n_1, n_2}\big\vert_{(Z_1, Z_2)}=\chi\big({\mathcal O}_{{\mathbb{C}}^2}\,,\,
{\mathcal O}_{{\mathbb{C}}^2}(\mu)
\big)-\chi\big({\mathcal I}_{Z_1}\,,\,{\mathcal I}_{Z_2}\otimes{\mathcal O}_{{\mathbb{C}}^2}(\mu) \big)\ ,
\end{equation}
where $\chi(E,F):=\sum_{i=0}^2 \, (-1)^i\, \mathrm{Ext}^i(E,F)$ for any pair of
coherent sheaves $E,F$ on ${\mathbb{C}}^2$, while its rank is
\begin{equation}
\operatorname{rk}\big(\boldsymbol{E}_\mu^{n_1, n_2} \big)= n_1+n_2 \ .
\end{equation}
Define the operator
$\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)\in\mathrm{End}(\mathbb{H}_{{\mathbb{C}}^2})[[z,
z^{-1}]]$ by its matrix elements
\begin{multline}\label{eq:carlssonokounkov}
(-1)^{n_2} \, \big\langle \operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)A_1 \,,\, A_2
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}} \\ :=z^{n_2-n_1} \,
\int_{\hilb{n_1}{{\mathbb{C}}^2}\times \hilb{n_2}{{\mathbb{C}}^2}} \,
\operatorname{eu}_T\big(\boldsymbol{E}_\mu^{n_1, n_2}\big)\cup p_1^\ast(A_1) \cup p_2^\ast(A_2) \ ,
\end{multline}
where $A_i\in H_T^\ast(\hilb{n_i}{{\mathbb{C}}^2})_{\mathrm{loc}}$ and $p_i$ is the projection from
$\hilb{n_1}{{\mathbb{C}}^2}\times \hilb{n_2}{{\mathbb{C}}^2}$ to the $i$-th factor for
$i=1,2$. By \cite[Lemma
6]{art:carlssonokounkov2012}, the matrix elements \eqref{eq:carlssonokounkov} in the fixed point basis are given by
\begin{align}\label{eq:carlokfixedbasis}
\big\langle \operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)[\lambda_1] \,,\, [{\lambda_2} ]
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}& =(-1)^{|\lambda_2|}\,
z^{|\lambda_2|-|\lambda_1|} \
\operatorname{eu}_T\big(\boldsymbol{E}_\mu^{n_1, n_2}\big\vert_{(Z_{\lambda_1},Z_{\lambda_2})}
\big) \\[4pt] \nonumber
& = (-1)^{|\lambda_2|}\, z^{|\lambda_2|-|\lambda_1|} \ m_{Y_{\lambda_1}, Y_{\lambda_2}}(\varepsilon_1, \varepsilon_2,\mu)\ ,
\end{align}
where
\begin{align}\label{eq:m}
m_{Y_1, Y_2}(\varepsilon_1, \varepsilon_2,a):= &\prod_{s_1\in Y_1}\,
\big(a-L_{Y_2}(s_1)\,
\varepsilon_1+(A_{Y_1}(s_1)+1)\, \varepsilon_2\big) \\ \nonumber & \times \ \prod_{s_2 \in
Y_2}\, \big(a+(L_{Y_1}(s_2)+1)\,
\varepsilon_1-A_{Y_2}(s_2)\, \varepsilon_2 \big)
\end{align}
for a pair of Young tableaux $Y_1,Y_2$ and $a\in{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$. In gauge theory this factorized expression for the matrix elements represents the contribution of the
\emph{bifundamental hypermultiplet}.
We shall now describe the operator $\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)$ in terms
of the operators $\mathfrak{p}_{m}$ defined in Equation \eqref{eq:heisenberg-C2} for $m\in{\mathbb{Z}}\setminus\{0\}$. In our setting, \cite[Theorem 1]{art:carlssonokounkov2012} assumes the following form.
\begin{theorem}\label{thm:CO}
The operator $\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)$ is a vertex operator in
Heisenberg operators given by the generalized bosonic exponential
associated with the Heisenberg algebra $\mathfrak{h}_{{\mathbb{C}}^2}$ as
\begin{equation}\label{eq:carlssonokounkov-heisenberg}
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu),z)=\operatorname{V}_{-\frac{\mu}{\varepsilon_2}, \frac{\mu+ \varepsilon_1+\varepsilon_2}{\varepsilon_2}}(z)\ .
\end{equation}
\end{theorem}
\subsection{Integrals of motion}\label{sec:integralC2}
Let $\boldsymbol{V}^n$ be the pushforward of $\boldsymbol{E}_0^{n,n}$ with respect to the
projection of the product $\hilb{n}{{\mathbb{C}}^2}\times \hilb{n}{{\mathbb{C}}^2}$ to the
second factor. It is a $T$-equivariant vector bundle on
$\hilb{n}{{\mathbb{C}}^2}$ of rank $n$, which we shall call the natural bundle over $\hilb{n}{{\mathbb{C}}^2}$. The $T$-equivariant Chern character of $\boldsymbol{V}^n$ at the fixed point
$Z_\lambda$ is given by
\begin{equation}
\operatorname{ch}_T\big(\boldsymbol{V}^n\big\vert_{Z_\lambda}\big) = \sum_{s\in
Y_\lambda}\, {\,\rm e}\,^{-L'(s)\, \varepsilon_1- A'(s)\, \varepsilon_2} \ .
\end{equation}
\begin{remark}
The vector bundle $\boldsymbol{V}^n$ can equivalently be defined as the pushforward with
respect to the projection $\hilb{n}{{\mathbb{C}}^2}\times {\mathbb{C}}^2\to\hilb{n}{{\mathbb{C}}^2}$
of the structure sheaf of the universal subscheme $\boldsymbol{Z}_n$. In the
literature $\boldsymbol{V}^n$ is also called the tautological sheaf and denoted ${\mathcal O}^{[n]}$ (cf.\ \cite{art:lehn1999, book:lehn2004,art:okounkovpandharipande2010, art:nakajima2014}).
\end{remark}
Let us denote by $\boldsymbol{V}$ the natural bundle over $\coprod_{n\geq 0}\, \hilb{n}{{\mathbb{C}}^2}$. The operators
of multiplication by $\boldsymbol{I}_1:=\operatorname{rk}\big(\boldsymbol{V} \big)$ and $\boldsymbol{I}_p:=
(\operatorname{c}_{p-1} )_T\big(\boldsymbol{V} \big)$ for $p\geq 2$ on $\prod_{n\geq
0}\,\mathbb{H}_{{\mathbb{C}}^2,n}$ have even
degrees, are self-adjoint with respect to the inner product on $\prod_{n\geq
0}\,\mathbb{H}_{{\mathbb{C}}^2,n}$, and commute with each other; they can thus be simultaneously
diagonalized in the
fixed point basis $[\lambda]$ of $\prod_{n\geq
0}\,\mathbb{H}_{{\mathbb{C}}^2,n}$ (see \cite{art:okounkovpandharipande2010},
and also \cite[Section 4]{art:nakajima2014} where our operator
$\mathfrak{p}_m$ is denoted $P_m(\varepsilon_2)$ for $m\in{\mathbb{Z}}\setminus \{0\}$). For example, one has
\begin{align}\label{eq:integralofmotions-fixedpoints}
\boldsymbol{I}_1\triangleright [\lambda]= |\lambda|\, [\lambda] \qquad \mbox{and} \qquad
\boldsymbol{I}_2\triangleright [\lambda]= -\sum_{s\in Y_\lambda}\,
\big(L'(s)\, \varepsilon_1+ A'(s)\, \varepsilon_2\big) \
[\lambda] \ .
\end{align}
As a consequence, the operators of multiplication by $\boldsymbol{I}_p$ for $ p\geq 1$ can be written in terms of the Heisenberg
operators \eqref{eq:heisenberg-C2} as elements of a commutative subalgebra of $\mathcal{U}(\mathfrak{h}_{{\mathbb{C}}^2})$. For example, we have
\begin{align}\label{eq:rank-C2}
\boldsymbol{I}_1 =& \ \beta\,
\sum_{m=1}^\infty \, \mathfrak{p}_{-m}\, \mathfrak{p}_{m}\ ,\\[4pt]
\boldsymbol{I}_2 =
& \ \varepsilon_1\, \Big(\, \frac{\beta}{2}\, \sum_{m,n=1}^\infty \, \big(
\mathfrak{p}_{-m}\, \mathfrak{p}_{-n}\, \mathfrak{p}_{m+n}+\mathfrak{p}_{-m-n}\, \mathfrak{p}_{n}\,
\mathfrak{p}_{m} \big) - \frac{\beta-1}{2}\, \sum_{m=1}^\infty\, (m-1)\,
\mathfrak{p}_{-m}\, \mathfrak{p}_{m}\,\Big)\ . \label{eq:c1-C2}
\end{align}
Note that the \emph{energy operator} $\boldsymbol{I}_1$ coincides with the Virasoro generator
$L_0^{\mathfrak{h}}$ from Section \ref{sec:virasoro-heisenberg}, while the
operator $\boldsymbol{I}_2$ is equal to $\varepsilon_1\, \Box^{\beta^{-1}}$, where $\Box^{\beta^{-1}}$ is the bosonized Hamiltonian of the
quantum trigonometric Calogero-Sutherland model with infinitely many particles and coupling constant $\beta^{-1}$.
\subsection{${\mathcal N}=2$ gauge theory}\label{sec:R4pure}
The Nekrasov partition function for pure ${\mathcal N}=2$ $U(1)$ gauge
theory on $\mathbb{R}^4$ is given by the generating function \cite{art:nekrasov2003, art:bruzzofucitomoralestanzini2003}
\begin{align}
{\mathcal Z}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2; \mathsf{q}) :=& \
\sum_{n=0}^\infty\, \mathsf{q}^n \ \int_{\hilb{n}{{\mathbb{C}}^2}} \, \big[\hilb{n}{{\mathbb{C}}^2}\big]_T\\[4pt]
= & \ \sum_{n=0}^\infty \, (-\mathsf{q})^n\, \big\langle
[\hilb{n}{{\mathbb{C}}^2}]_T\,,\,[\hilb{n}{{\mathbb{C}}^2}]_T \big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}
\end{align}
where $\mathsf{q}\in{\mathbb{C}}^*$ with $|\mathsf{q}|<1$.
By the localization theorem we obtain
\begin{equation}
{\mathcal Z}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2; \mathsf{q}) = \sum_{\lambda} \,
\Big(-\frac{\mathsf{q}}{\varepsilon_2^2}\, \Big)^{\vert\lambda\vert} \
\prod_{s\in Y_\lambda}\, \frac{1}{\big((L(s)+1 )\,\beta+A(s)\big)\,
\big(L(s)\, \beta+(A(s)+1)\big)}
\end{equation}
as in \cite[Equation (3.16)]{art:bruzzofucitomoralestanzini2003},
where the sum runs over all partitions $\lambda$.
\begin{remark}\label{rmk:exponentialnekrasov}
By \cite[Equation (4.7)]{art:nakajimayoshioka2005-I}, the partition
function can be summed explicitly and written in the closed form
\begin{equation}\label{eq:nakajimayoshioka}
{\mathcal Z}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2; \mathsf{q}) = \exp\Big(\, \frac{\mathsf{q}}{\varepsilon_1\, \varepsilon_2}\, \Big) \ .
\end{equation}
\end{remark}
\subsubsection{Gaiotto state}\label{sec:gaiottoC2}
In \cite{art:gaiotto2009}, Gaiotto considers the inducing state of the
(completed) Verma module of the Virasoro algebra. It has the property
that it is a Whittaker vector for the Verma module, and the norm of
its $\mathsf{q}$-deformation coincides with the Nekrasov partition
function of pure ${\mathcal N}=2$ $SU(2)$ gauge theory on ${\mathbb{R}}^4$. Here we
consider the analogous vector for $U(1)$ gauge theory on ${\mathbb{R}}^4$.
Following \cite{art:schiffmannvasserot2013}, we define the \emph{Gaiotto state} to be the sum of all fundamental classes
\begin{equation}
G := \sum_{n\geq 0} \, \left[ \hilb{n}{{\mathbb{C}}^2} \right]_T
\end{equation}
in the completed Fock space $\prod_{n \geq 0} \, {\mathbb{H}}_{{\mathbb{C}}^2,n}$. We also introduce the \emph{weighted Gaiotto state} as the formal power series
\begin{equation}
G_\mathsf{q} := \sum_{n\geq 0} \, \mathsf{q}^n \, \left[ \hilb{n}{{\mathbb{C}}^2} \right]_T
\ \in \ \mbox{$\prod\limits_{n \geq 0}$}\, \mathsf{q}^n \, {\mathbb{H}}_{{\mathbb{C}}^2,n}\ .
\end{equation}
Consider the bilinear form
\begin{equation}
\langle-,-\rangle_{{\mathbb{H}}_{{\mathbb{C}}^2},\mathsf{q}} \,:\,
\mbox{$\prod\limits_{n \geq 0}$}\, \mathsf{q}^n \, {\mathbb{H}}_{{\mathbb{C}}^2,n} \times
\mbox{$\prod\limits_{n \geq 0}$}\, \mathsf{q}^n \, {\mathbb{H}}_{{\mathbb{C}}^2,n} \
\longrightarrow \ {\mathbb{C}}(\varepsilon_1,\varepsilon_2)[[\mathsf{q}]]
\end{equation}
defined by
\begin{equation}
\Big\langle \, \mbox{$\sum\limits_{n\geq 0}$} \, \mathsf{q}^n\, \eta_n \,,\,
\mbox{$\sum\limits_{n\geq 0}$} \, \mathsf{q}^n \, \nu_n \,
\Big\rangle_{{\mathbb{H}}_{{\mathbb{C}}^2},\mathsf{q}} := \sum_{n= 0}^\infty\, \mathsf{q}^n \
\int_{\hilb{n}{{\mathbb{C}}^2}}\, \eta_n \cup \nu_n = \sum_{n= 0}^\infty\,
(-\mathsf{q})^n \, \langle \eta_n, \nu_n \rangle_{{\mathbb{H}}_{{\mathbb{C}}^2}}\ .
\end{equation}
It follows immediately that the norm of the weighted Gaiotto state is the Nekrasov partition function for ${\mathcal N}=2$ $U(1)$ gauge theory on ${\mathbb{R}}^4$:
\begin{equation}\label{eq:gaiottonekrasovC2}
{\mathcal Z}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2;\mathsf{q}) = \langle G_\mathsf{q},
G_\mathsf{q} \rangle_{{\mathbb{H}}_{{\mathbb{C}}^2},\mathsf{q}} \ .
\end{equation}
By Proposition \ref{prop:whittakerC2} we have the following result.
\begin{proposition}\label{prop:gaiottowhittakerC2}
The Gaiotto state $G$ is a Whittaker vector of type $\chi$, where the
algebra homomorphism $\chi\colon
\mathcal{U}(\mathfrak{h}^+_{{\mathbb{C}}^2}) \to{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ is defined by
\begin{equation}
\chi(\mathfrak{p}_1)=-\frac{1}{\varepsilon_1} \qquad
\mbox{and} \qquad \chi(\mathfrak{p}_n)=0 \qquad \mbox{for} \quad n>1 \ .
\end{equation}
\end{proposition}
\proof
Let $\eta \in {\mathbb{C}} (\varepsilon_1,\varepsilon_2)$.
By using the formal expansion \eqref{eq:formalexpansion} and the
isomorphism $\phi$, we can write $\phi(G(\eta))$ in terms of powers
$p_1^n$. By Lemma \ref{lem:p1} and simple algebraic manipulations we
can then rewrite the vector $G(\eta)$ as
\begin{equation}\label{eq:whittaker}
G(\eta)=\sum_{n\geq 0} \, (\eta\,\varepsilon_2)^n\,
\big[\hilb{n}{{\mathbb{C}}^2} \big]_T
\end{equation}
and the result follows.
\endproof
\subsection{Quiver gauge theories\label{sec:R4quivers}}
We now add matter fields to the ${\mathcal N}=2$ gauge theory on ${\mathbb{R}}^4$. We consider the most general ${\mathcal N}=2$ superconformal quiver gauge
theory with gauge group $U(1)^{r+1}$ for $r\geq0$, following the
general ADE classification of \cite[Chapter
3]{art:nekrasovpestun2012}.
Let ${\tt Q}=({\tt Q}_0,{\tt Q}_1)$ be a quiver,
i.e., an oriented graph with a finite set of vertices ${\tt Q}_0$,
a finite set of edges
${\tt Q}_1\subset{\tt Q}_0\times{\tt Q}_0$, and two projection maps
${\tt s},{\tt t}:{\tt Q}_1\rightrightarrows {\tt Q}_0$ which assign to each oriented edge
its source and target vertex respectively. Representations of the
quiver encode the matter field content of the gauge theory. Fix a vector
$(n_\upsilon)_{\upsilon\in {\tt Q}_0}\in{\mathbb{N}}^{{\tt Q}_0}$ of integers labelled by the nodes of
the quiver ${\tt Q}$, and consider the product of Hilbert schemes
$\prod_{\upsilon\in{\tt Q}_0}\, \hilb{n_\upsilon}{{\mathbb{C}}^2}$. The vertices $\upsilon\in {\tt Q}_0$
label $U(1)$ gauge groups and $m_\upsilon\geq0$ (resp. $\bar m_\upsilon\geq 0$)
fundamental (resp. antifundamental) hypermultiplets of masses
$\mu^s_{\upsilon}$, $s=1,\dots,m_\upsilon$ (resp. $\bar\mu^{\bar s}_{\upsilon}$,
$\bar s=1,\dots,\bar m_\upsilon$) which correspond to the $T$-equivariant vector
bundles $\boldsymbol{V}_{\mu^s_{\upsilon}}^{n_\upsilon}$
(resp. $\bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{n_\upsilon}$) of rank $n_\upsilon$ on $\hilb{n_\upsilon}{{\mathbb{C}}^2}$ obtained by
pushforward of $\boldsymbol{E}_{\mu^s_\upsilon}^{n_\upsilon,n_\upsilon}$
(resp. $\boldsymbol{E}_{\bar\mu^{\bar s}_\upsilon}^{n_\upsilon,n_\upsilon}$) with respect to the
projection of $\hilb{n_\upsilon}{{\mathbb{C}}^2} \times\hilb{n_\upsilon}{{\mathbb{C}}^2}$ to the
second (resp. first) factor. The edges
$e\in {\tt Q}_1$ label $U(1)\times U(1)$ bifundamental hypermultiplets of masses $\mu_e$ which correspond to
the vector bundles $\boldsymbol{E}_{\mu_e}^{n_{{\tt s}(e)},n_{{\tt t}(e)}}$ of
rank $n_{{\tt s}(e)}+n_{{\tt t}(e)}$ on $\hilb{n_{{\tt s}(e)}}{{\mathbb{C}}^2}
\times\hilb{n_{{\tt t}(e)}}{{\mathbb{C}}^2}$; if the edge $e$ is a vertex loop, i.e., ${\tt s}(e)={\tt t}(e)$, then the restriction of
$\boldsymbol{E}_{\mu_e}^{n_{{\tt s}(e)},n_{{\tt s}(e)}}$ to the diagonal of $\hilb{n_{{\tt s}(e)}}{{\mathbb{C}}^2}
\times\hilb{n_{{\tt s}(e)}}{{\mathbb{C}}^2}$ describes an adjoint hypermultiplet
of mass $\mu_e$. The total matter field content of
the ${\mathcal N}=2$ quiver gauge theory associated to ${\tt Q}$
in the sector labelled by $(n_\upsilon)_{\upsilon\in {\tt Q}_0}\in{\mathbb{N}}^{{\tt Q}_0}$
is thus described by the bundle on $\prod_{\upsilon\in{\tt Q}_0}\,
\hilb{n_\upsilon}{{\mathbb{C}}^2}$ given by
\begin{equation}
\boldsymbol{M}^{(n_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar s}_\upsilon),
(\mu_e)}:= \bigoplus_{\upsilon\in{\tt Q}_0}\, p_\upsilon^* \Big(\, \bigoplus_{s=1}^{m_\upsilon}\,
\boldsymbol{V}_{\mu^s_{\upsilon}}^{n_\upsilon} \ \oplus \ \bigoplus_{\bar
s=1}^{\bar m_\upsilon}\, \bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{n_\upsilon}\, \Big) \ \oplus \ \bigoplus_{e\in{\tt Q}_1}\,
p_e^*\boldsymbol{E}_{\mu_e}^{n_{{\tt s}(e)},n_{{\tt t}(e)}} \ ,
\end{equation}
where $p_\upsilon$ is the projection of $\prod_{\upsilon\in{\tt Q}_0}\,
\hilb{n_\upsilon}{{\mathbb{C}}^2}$ to the $\upsilon$-th factor and $p_e$ the projection
to $\hilb{n_{{\tt s}(e)}}{{\mathbb{C}}^2}
\times\hilb{n_{{\tt t}(e)}}{{\mathbb{C}}^2}$.
For each vertex $\upsilon\in{\tt Q}_0$, the degree of the Euler class of the
pushforward of this
bundle to the $\upsilon$-th factor is the integer
\begin{align}
d_\upsilon := & \ \dim_{\mathbb{C}}\hilb{n_\upsilon}{{\mathbb{C}}^2}- \operatorname{rk}\Big(
\boldsymbol{M}^{(n_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar s}_\upsilon),
(\mu_e)}\Big\vert_{\hilb{n_\upsilon}{{\mathbb{C}}^2}}\,\Big) \\[4pt] = & \ n_\upsilon\,
\big(2-m_\upsilon-\bar m_\upsilon-\#\{e\in{\tt Q}_1\, | \, {\tt s}(e)=\upsilon\}-
\#\{e\in{\tt Q}_1\, | \, {\tt t}(e)=\upsilon\}\big) \ .
\end{align}
The ${\mathcal N}=2$ quiver gauge theory is said to be \emph{conformal} if
$d_\upsilon=0$ for all $\upsilon\in{\tt Q}_0$; it is
\emph{asymptotically free} if $d_\upsilon>0$. Note that with this
definition the pure ${\mathcal N}=2$ gauge theory of Section \ref{sec:R4pure}
is asymptotically free.
As explained in \cite[Chapter 3]{art:nekrasovpestun2012}, ${\mathcal N}=2$ asymptotically free quiver gauge theories can be recovered from conformal theories, so in the following we restrict our
attention to superconformal quiver gauge theories.
Introduce coupling constants $\mathsf{q}_\upsilon\in{\mathbb{C}}^*$ with
$|\mathsf{q}_\upsilon|<1$ at each vertex
$\upsilon\in{\tt Q}_0$, and let $T_{\boldsymbol{\mu}}$ be the maximal torus of the
total flavour symmetry group
\begin{equation}
G_{\rm f}=
\prod_{\upsilon\in{\tt Q}_0}\, GL(m_\upsilon,{\mathbb{C}})\times GL(\bar m_\upsilon,{\mathbb{C}}) \
\times \ \prod_{e\in{\tt Q}_1}\, {\mathbb{C}}^*
\end{equation}
with $H_{T_{\boldsymbol{\mu}}}^*({\rm
pt};{\mathbb{C}})={\mathbb{C}}[(\mu_e),(\mu_\upsilon^s),(\bar\mu_\upsilon^{\bar s})]$. Then the quiver
gauge theory partition function is defined by the generating function
\begin{multline}
{\mathcal Z}_{{\mathbb{C}}^2}^{{\tt Q}}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}):=
\sum_{(n_\upsilon)\in{\mathbb{N}}^{{\tt Q}_0}}\, \boldsymbol{\mathsf{q}}^{\boldsymbol n} \
\int_{\prod\limits_{\upsilon\in{\tt Q}_0}\, \hilb{n_\upsilon}{{\mathbb{C}}^2}}\, \operatorname{eu}_{T\times
T_{\boldsymbol{\mu}}}\big(\boldsymbol{M}^{(n_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)} \big) \\[4pt] = \sum_{(n_\upsilon)\in{\mathbb{N}}^{{\tt Q}_0}} \, \boldsymbol{\mathsf{q}}^{\boldsymbol n} \
\int_{\prod\limits_{\upsilon\in{\tt Q}_0}\, \hilb{n_\upsilon}{{\mathbb{C}}^2}}\
\prod_{\upsilon\in{\tt Q}_0}\, p_\upsilon^*\Big(\, \prod_{s=1}^{m_\upsilon}\,
\operatorname{eu}_T\big(\boldsymbol{V}_{\mu^s_{\upsilon}}^{n_\upsilon}\big) \
\prod_{\bar s=1}^{\bar m_\upsilon}\,
\operatorname{eu}_T\big(\bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{n_\upsilon} \big)\, \Big) \\
\times \ \prod_{e\in{\tt Q}_1}\,
p_e^*\, \operatorname{eu}_T\big(\boldsymbol{E}_{\mu_e}^{n_{{\tt s}(e)},n_{{\tt t}(e)}}
\big) \ ,
\end{multline}
where $\boldsymbol{\mathsf{q}}^{\boldsymbol{n}}:= \prod_{\upsilon\in{\tt Q}_0}\, \mathsf{q}_\upsilon^{n_\upsilon}$. By
the localization theorem, we obtain
\begin{multline}
{\mathcal Z}_{{\mathbb{C}}^2}^{{\tt Q}}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}) =
\sum_{(\lambda^\upsilon)}\, (-\boldsymbol{\mathsf{q}})^{\boldsymbol{\lambda}} \ \prod_{\upsilon\in{\tt Q}_0}\,
\frac{\prod\limits_{s=1}^{m_\upsilon}\,
m_{Y_{\lambda^\upsilon}}\big(\varepsilon_1,\varepsilon_2,\mu_\upsilon^s \big) \
\prod\limits_{\bar s=1}^{\bar m_\upsilon}\,
m_{Y_{\lambda^\upsilon}}\big(\varepsilon_1,\varepsilon_2,\bar
\mu_\upsilon^{\bar s} + \varepsilon_1 +
\varepsilon_2\big)}{m_{Y_{\lambda^\upsilon},
Y_{\lambda^\upsilon}}(\varepsilon_1,\varepsilon_2
,0)} \label{eq:ZcalC2quiver} \\ \times \
\prod_{e\in{\tt Q}_1}\,
m_{Y_{\lambda^{{\tt s}(e)}},Y_{\lambda^{{\tt t}(e)}}}(
\varepsilon_1,\varepsilon_2 ,\mu_e)
\end{multline}
where $\boldsymbol{\mathsf{q}}^{\boldsymbol{\lambda}}:= \prod_{\upsilon\in{\tt Q}_0}\,
\mathsf{q}_\upsilon^{|\lambda^\upsilon|}$ for a collection of partitions
$\lambda^\upsilon$ associated to the vertices of the quiver, and
\begin{equation}
m_Y(\varepsilon_1,\varepsilon_2,a):= \prod_{s\in Y}\, \big(a-L'(s)\,
\varepsilon_1-A'(s)\, \varepsilon_2\big)
\end{equation}
for a Young tableau $Y$ and $a\in{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$.
The conformal constraint
\begin{equation}
m_\upsilon+\bar m_\upsilon+ \#\big\{e\in{\tt Q}_1\, \big| \, {\tt s}(e)=\upsilon \big\} +
\#\big\{e\in{\tt Q}_1\, \big| \, {\tt t}(e)=\upsilon \big\} = 2
\label{eq:confconstr}\end{equation}
for each $\upsilon\in{\tt Q}_0$ severely restricts the possible quivers in the abelian gauge
theory. It is easy to check that the only admissible quivers in the ADE
classification of \cite[Chapter
3]{art:nekrasovpestun2012} are the linear (or chain)
quivers of the finite-dimensional $A_r$-type Dynkin diagram and the cyclic (or necklace)
quivers of the affine $\widehat{A}_{r}$-type extended Dynkin
diagram for some $r\geq0$\footnote{Here the $A_0$-type Dynkin diagram is the trivial quiver consisting of a single vertex with no arrows, and the $\widehat{A}_0$-type Dynkin diagram is the quiver consisting of a single vertex with a vertex edge loop.}. We consider in detail each case in turn.
\subsection{$\widehat{A}_{r}$ theories\label{sec:hatArtheories}}
For the cyclic quivers of type $\widehat{A}_{r}$
\begin{equation}
\begin{tikzpicture}[xscale=2,yscale=-1]
\node (A0_3) at (3,0) {$\circ$};
\node (A1_1) at (1,1.5) {$\circ$};
\node (A1_2) at (2,1.5) {$\circ$};
\node (A1_3) at (3,1.5) {$\ldots$};
\node (A1_4) at (4,1.5) {$\circ$};
\node (A1_5) at (5,1.5) {$\circ$};
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\path (A1_4) edge [->]node [auto] {$\scriptstyle{}$} (A1_5);
\path (A0_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A1_5) edge [->]node [auto] {$\scriptstyle{}$} (A0_3);
\end{tikzpicture}
\end{equation}
with $r+1$ vertices and arrows, one has $m_\upsilon=\bar m_\upsilon=0$ by
Equation \eqref{eq:confconstr}. We
label the vertices ${\tt Q}_0$ by $\upsilon=0,1,\dots, r$ with
counterclockwise orientation and read modulo $r+1$, and similarly for
the edges $e=(\upsilon,\upsilon+1) \in{\tt Q}_1$. The partition function for the ${\mathcal N}=2$
quiver gauge theory of type $\widehat{A}_{r}$ reads as
\begin{equation}\label{eq:ZC2hatAr}
{\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})
= \sum_{\boldsymbol{\lambda}}\, (-\boldsymbol{\mathsf{q}})^{\boldsymbol{\lambda}} \ \prod_{\upsilon=0}^r\
\frac{m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon+1}}}(\varepsilon_1,\varepsilon_2,\mu_\upsilon)}{m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon}}}(\varepsilon_1,\varepsilon_2,0)}
\ ,
\end{equation}
where the sum is over all $r+1$-vectors of partitions
$\boldsymbol{\lambda}=(\lambda^0,\lambda^1,\dots,\lambda^{r})$ with
$\lambda^{r+1}:= \lambda^0$ and
$\boldsymbol{\mathsf{q}}^{\boldsymbol{\lambda}}:=\prod_{\upsilon=0}^r \,
\mathsf{q}_\upsilon^{|\lambda^\upsilon|}$.
\subsubsection{Conformal blocks}
We will relate the partition function \eqref{eq:ZC2hatAr} to the trace of
vertex operators
$\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon)$. We shall also denote by $\mathfrak{h}$ the Heisenberg algebra $\mathfrak{h}_{{\mathbb{C}}^2}$ to simplify the presentation.
\begin{proposition}\label{prop:hatArtrace}
The partition function of the $\widehat{A}_r$-theory on ${\mathbb{R}}^4$ is
given by
\begin{equation}
{\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})
= \mathrm{Tr}_{\mathbb{H}_{{\mathbb{C}}^2}}\, \mathsf{q}^{L_0^\mathfrak{h}} \ \prod_{\upsilon=0}^r\,
\, \operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon)
\end{equation}
independently of $z_0\in{\mathbb{C}}^*$, where $\mathsf{q}:=\mathsf{q}_0\,
\mathsf{q}_1\cdots\mathsf{q}_r$ and $z_\upsilon:= z_0\, \mathsf{q}_1\cdots \mathsf{q}_\upsilon$
for $\upsilon=1,\dots,r$.
\end{proposition}
\proof
By
Equation \eqref{eq:integralofmotions-fixedpoints}, the Virasoro
operator $L_0^{\mathfrak{h}}$ acts in $\mathbb{H}_{{\mathbb{C}}^2}$ as
\begin{equation}
L_0^{\mathfrak{h}} \, \big\vert_{\mathbb{H}_{{\mathbb{C}}^2,n}}=n\
\mathrm{id}_{\mathbb{H}_{{\mathbb{C}}^2,n}} \ ,
\end{equation}
and so the trace of products of Ext vertex operators $\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon)$ is given by the sum of their matrix elements over the fixed point basis as
\begin{equation}
\mathrm{Tr}_{\mathbb{H}_{{\mathbb{C}}^2}}\, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \prod_{\upsilon=0}^r
\, \operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon) =\sum_{\boldsymbol{n}\in{\mathbb{N}}^{r+1}}
\, \mathsf{q}^{n_0} \
\sum_{\boldsymbol{\lambda}\, :\, \vert\lambda^\upsilon\vert=n_\upsilon}\
\prod_{\upsilon=0}^r\, \frac{\big\langle
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon)[\lambda^\upsilon]\,,\, [\lambda^{\upsilon+1}] \big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}}{\big\langle [\lambda^\upsilon]\,,\, [\lambda^\upsilon] \big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}}\ .
\end{equation}
By Equation \eqref{eq:carlokfixedbasis} we obtain
\begin{align}
\mathrm{Tr}_{\mathbb{H}_{{\mathbb{C}}^2}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \prod_{\upsilon=0}^r\,
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),1)& =\sum_{\boldsymbol{n}\in{\mathbb{N}}^{r+1}}
\, \mathsf{q}^{n_0} \ \prod_{\upsilon=0}^r\, (-1)^{n_\upsilon}\,
z_\upsilon^{n_\upsilon-n_{\upsilon+1}} \\ & \qquad \qquad \times \
\sum_{\boldsymbol{\lambda}\, :\, \vert\lambda^\upsilon\vert=n_\upsilon} \
\prod_{\upsilon=0}^r\,
\frac{\operatorname{eu}_T\big(
\boldsymbol{E}_{\mu_\upsilon}^{n_\upsilon,n_{\upsilon+1}} \big\vert_{
(Z_{\lambda^\upsilon} ,Z_{\lambda^{\upsilon+1}})}
\big)}{\operatorname{eu}_T\big(T_{Z_{\lambda^\upsilon}}\hilb{n_\upsilon}{{\mathbb{C}}^2} \big)}\\[4pt]
&=\sum_{\boldsymbol{n}\in{\mathbb{N}}^{r+1}}
\, (-\boldsymbol{\mathsf{q}})^{\boldsymbol{n}} \
\sum_{\boldsymbol{\lambda}\, :\, \vert\lambda^\upsilon\vert=n_\upsilon}\
\prod_{\upsilon=0}^r\
\frac{m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon+1}}}(\varepsilon_1,\varepsilon_2,\mu_\upsilon)}{m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon}}}(\varepsilon_1,\varepsilon_2,0)}
\ ,
\end{align}
and the result follows.
\endproof
\begin{remark}
Proposition \ref{prop:hatArtrace} shows that the partition function of
the $\widehat{A}_r$-theory coincides with the conformal block of the
Heisenberg algebra $\mathfrak{h}_{{\mathbb{C}}^2}$ on the
elliptic curve with nome $\mathsf{q}$ and $r+1$ punctures at
$z_0,z_1,\dots,z_r$; we can set $z_0=1$ without loss of
generality. The conformal dimension of the primary field inserted at
the $\upsilon$-th puncture is
\begin{equation}
\Delta(\mu_\upsilon;\varepsilon_1,\varepsilon_2)= \frac{\mu_\upsilon\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{2\varepsilon_1\,
\varepsilon_2} \ .
\end{equation}
This elliptic curve is the Seiberg-Witten curve of the ${\mathcal N}=2$
$U(1)^{r+1}$ quiver gauge theory on ${\mathbb{R}}^4$.
\end{remark}
By using the same arguments as in the proof of \cite[Corollary
1]{art:carlssonokounkov2012}, an explicit formula for the trace in
this case can be obtained using Equation
\eqref{eq:carlssonokounkov-heisenberg} and we arrive at the explicit evaluation of
the partition function. In the following $\eta(\mathsf{q}):=\mathsf{q}^{\frac{1}{24}}\,
\prod_{n=1}^\infty\, (1-\mathsf{q}^n)$ denotes the Dedekind function.
\begin{proposition}\label{prop:tracededekind}
\begin{equation}
{\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})
= \prod_{\upsilon=0}^r\,\big(\mathsf{q}_\upsilon^{-\frac{1}{24}}\, \eta(\mathsf{q}_\upsilon)\big)^{-\frac{\mu_\upsilon\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{\varepsilon_1\,
\varepsilon_2}}\ \mathsf{q}^{\frac{1}{24}}\,\eta(\mathsf{q})^{-1}\ .
\end{equation}
\end{proposition}
A similar formula for the $U(1)^{r+1}$ quiver gauge theory partition
function is conjectured in \cite[Appendix
C.2]{art:aldaygaiottotachikawa2010}.
\subsubsection{$\widehat{A}_{0}$ theory\label{sec:hatA0theory}}
The degenerate case $r=0$ of the $\widehat{A}_r$ quiver gauge theory
corresponds to the quiver consisting of a single
node with a vertex edge loop
\begin{equation}
\begin{tikzpicture}[auto]
\node (A0_0) at (0,0) {$\circ$};
\path[->] (A0_0) edge[loop above, in=130,out=50,looseness=10, shorten >=-2pt, shorten <=-2pt] node {$\scriptstyle{}$} (A0_0);
\end{tikzpicture}
\end{equation}
and is known as the ${\mathcal N}=2^*$ gauge theory; it describes a single adjoint matter hypermultiplet of mass $\mu$ in
the $U(1)$ ${\mathcal N}=2$ gauge theory on ${\mathbb{R}}^4$. Then the quiver gauge
theory partition function is the Nekrasov partition function
for ${\mathcal N}=2^\ast$ gauge theory \cite{art:nekrasov2003,art:bruzzofucitomoralestanzini2003} and is given
by
\begin{equation}
{\mathcal Z}_{{\mathbb{C}}^2}^{\widehat{A}_{0}}(\varepsilon_1,\varepsilon_2, \mu; \mathsf{q}) =
\sum_{\lambda} \, \mathsf{q}^{\vert\lambda\vert} \ \prod_{s\in Y_\lambda}
\frac{\big( (L(s)+1)\, \varepsilon_1-A(s)\, \varepsilon_2+\mu \big)\,
\big( L(s)\, \varepsilon_1-(A(s)+1)\, \varepsilon_2-\mu \big)}{\big(
(L(s)+1)\, \varepsilon_1-A(s)\, \varepsilon_2 \big) \, \big( L(s)\,
\varepsilon_1-(A(s)+1)\, \varepsilon_2 \big)}
\end{equation}
as in \cite[Equation (3.26)]{art:bruzzofucitomoralestanzini2003}.
By Proposition \ref{prop:tracededekind}, we have
\begin{equation}\label{eq:Ahat0dedekind}
{\mathcal Z}_{{\mathbb{C}}^2}^{\widehat{A}_{0}} (\varepsilon_1,\varepsilon_2, \mu;
\mathsf{q})=\big(\mathsf{q}^{-\frac{1}{24}}\, \eta(\mathsf{q})\big)^{-\frac{\mu\,
(\mu+\varepsilon_1+\varepsilon_2)}{\varepsilon_1\, \varepsilon_2}-1}\ .
\end{equation}
A similar formula is written in \cite[Equation (2.28)]{art:wyllard2009}.
In the case of an antidiagonal torus action, i.e.,
$\varepsilon_1=-\varepsilon_2$, this result coincides with the formula
derived in
\cite[Equation (6.12)]{art:nekrasovokounkov2006}.
We can then rewrite Proposition \ref{prop:hatArtrace} in the following way.
\begin{corollary}
\begin{equation}
{\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})
= \prod_{\upsilon=0}^r\, {\mathcal Z}_{{\mathbb{C}}^2}^{\widehat{A}_{0}} (\varepsilon_1,\varepsilon_2, \mu_\upsilon ;
\mathsf{q}_\upsilon)\ \frac{\prod\limits_{\upsilon=0}^r\,\eta(\mathsf{q}_\upsilon)}{\eta(\mathsf{q})}\ .
\end{equation}
\end{corollary}
\subsection{${A}_{r}$ theories\label{sec:Artheories}}
Consider now the linear quivers of type ${A}_{r}$
\begin{equation}
\begin{tikzpicture}[xscale=2.5,yscale=-.7]
\node (A1_0) at (0,0) {$\bullet$};
\node (A1_1) at (1,0) {$\circ$};
\node (A1_2) at (2,0) {$\ldots$};
\node (A1_3) at (3,0) {$\circ$};
\node (A1_4) at (4,0) {$\bullet$};
\path (A1_0) edge [->]node [auto] {$\scriptstyle{}$} (A1_1);
\path (A1_1) edge [->]node [auto] {$\scriptstyle{}$} (A1_2);
\path (A1_2) edge [->]node [auto] {$\scriptstyle{}$} (A1_3);
\path (A1_3) edge [->]node [auto] {$\scriptstyle{}$} (A1_4);
\end{tikzpicture}
\end{equation}
with $r+1$ vertices and $r$ arrows, where the solid nodes indicate the
insertion of a single fundamental or antifundamental
hypermultiplet. In this case we label vertices ${\tt Q}_0$ from left to
right with $\upsilon=0,1,\dots,r$ and edges ${\tt Q}_1$ with $e=(\upsilon,\upsilon+1)$;
for definiteness we take $\bar m_\upsilon=0$, so that $m_0=m_{r}=1$ by the
conformal constraints \eqref{eq:confconstr}. The partition function for the ${\mathcal N}=2$ quiver
gauge theory of type $A_r$ for $r\geq1$ reads as
\begin{equation}\label{eq:ZC2Ar}
{\mathcal Z}^{{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})
= \sum_{\boldsymbol{\lambda}}\, (-\boldsymbol{\mathsf{q}})^{\boldsymbol{\lambda}} \
\frac{m_{Y_{\lambda^0}}(\varepsilon_1,\varepsilon_2,\mu_0) \
\prod\limits_{\upsilon=0}^{r-1}\,
m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon+1}}}(\varepsilon_1,\varepsilon_2,\mu_{\upsilon+1})
\
m_{Y_{\lambda^r}}(\varepsilon_1,\varepsilon_2,\mu_{r+1})}{\prod\limits_{\upsilon=0}^r\,
m_{Y_{\lambda^\upsilon},Y_{\lambda^{\upsilon}}}(\varepsilon_1,\varepsilon_2,0)}
\ .
\end{equation}
\subsubsection{Conformal blocks}
We will express the partition function \eqref{eq:ZC2Ar} as a
particular matrix element of
Ext vertex operators.
\begin{proposition}\label{prop:fundamental-matter}
The partition function of the $A_r$-theory on ${\mathbb{R}}^4$ is given by
\begin{equation}
{\mathcal Z}^{{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}) =
\Big\langle \vert 0\rangle\,,\,
\prod_{\upsilon=0}^{r+1} \, \operatorname{V}\big({\mathcal O}_{{\mathbb{C}}^2}(\mu_{\upsilon}
),z_\upsilon \big)\vert 0\rangle\Big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}
\end{equation}
independently of $z_0\in{\mathbb{C}}^*$, where $z_\upsilon:=z_0\, \mathsf{q}_0\,
\mathsf{q}_1\cdots \mathsf{q}_{\upsilon-1}$ for $\upsilon=1,\dots,r+1$.
\end{proposition}
\proof
Arguing as in the proof of Proposition \ref{prop:hatArtrace}, we write
\begin{multline}
\Big\langle \vert 0\rangle\,,\,
\prod_{\upsilon=0}^{r+1} \, \operatorname{V}\big({\mathcal O}_{{\mathbb{C}}^2}(\mu_{\upsilon}
),z_\upsilon \big)\vert 0\rangle\Big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}} \\ =\sum_{\boldsymbol{n}\in{\mathbb{N}}^{r+1}}
\ \sum_{\boldsymbol{\lambda}\, :\, \vert\lambda^\upsilon\vert=n_\upsilon} \, \big\langle
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_0),z_0)\,[\lambda^{0}] ,\, |0\rangle
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}} \
\frac{\prod\limits_{\upsilon=0}^{r-1} \, \big\langle
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_{\upsilon+1}),z_{\upsilon+1} )[\lambda^{\upsilon+1}]\,,\,
[\lambda^{\upsilon}]
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}}{\prod\limits_{\upsilon=0}^r\,
\big\langle [\lambda^\upsilon]\,,\, [\lambda^\upsilon]
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}} \\ \times \ \big\langle
\operatorname{V}({\mathcal O}_{{\mathbb{C}}^2}(\mu_{r+1}),z_{r+1} )|0\rangle \,,\, [\lambda^{r}]
\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}} \ ,
\end{multline}
and by Equation \eqref{eq:carlokfixedbasis} and the orthogonality relation \eqref{eq:product-C2} the result then follows.
\endproof
\begin{remark}
Proposition \ref{prop:fundamental-matter} expresses the partition
function of the $A_r$-theory as a conformal block of the Heisenberg
algebra $\mathfrak{h}_{{\mathbb{C}}^2}$ on the Riemann
sphere with $r+4$ punctures at $\infty,z_0,z_1,\dots,z_{r+1},0$; again
we can set $z_0=1$ without loss of generality. The conformal dimension
of the primary field at the insertion point $z_\upsilon$ is
$\Delta(\mu_\upsilon;\varepsilon_1,\varepsilon_2)$, while at
$\infty,0$ they are given respectively by
$\Delta(\tilde\mu_{\infty,0}; \varepsilon_1,\varepsilon_2)$, where the
masses $\tilde\mu_{\infty,0}$ obey
\begin{equation}
\tilde\mu_\infty+\tilde\mu_0=\varepsilon_1+\varepsilon_2
+\sum_{\upsilon=0}^{r+1}\, \mu_\upsilon \ .
\end{equation}
The Seiberg-Witten curve of the ${\mathcal N}=2$ $U(1)^{r+1}$ quiver gauge
theory on ${\mathbb{R}}^4$ is a branched cover of this $(r+2)$-punctured Riemann
sphere, ramified over the points $\infty,0$.
\end{remark}
Using the vertex operator representation, we can again get a closed
formula for the combinatorial expansion \eqref{eq:ZC2Ar}.
\begin{proposition}\label{prop:ZC2Arexpl}
\begin{equation}
{\mathcal Z}^{{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}) =
\prod_{0\leq \upsilon<\upsilon'\leq r+1}\,
\big(1-\mathsf{q}_{\upsilon+1}\cdots
\mathsf{q}_{\upsilon'}\big)^{-\frac{\mu_{\upsilon'}\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{\varepsilon_1\,
\varepsilon_2}} \ .
\end{equation}
\end{proposition}
\proof
Using Equation \eqref{eq:VOnprod} to express the product of vertex
operators in Proposition \ref{prop:fundamental-matter} in normal
ordered form, we can write
\begin{multline}
\prod_{\upsilon=0}^{r+1}\,
\operatorname{V}\big({\mathcal O}_{{\mathbb{C}}^2}(\mu_\upsilon),z_\upsilon\big) \vert 0\rangle =
\prod_{0\leq \upsilon<\upsilon'\leq r+1}\, \Big(\,
1-\frac{z_{\upsilon'}}{z_\upsilon}\, \Big)^{-\frac{\mu_{\upsilon'}\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{\varepsilon_1\,
\varepsilon_2}} \\ \times \ \exp\Big(-\sum_{\upsilon=0}^{r+1}\,
\frac{\mu_\upsilon}{\varepsilon_2} \ \sum_{m=1}^\infty\,
\frac{z_\upsilon^m}{m}\, \mathfrak{p}_{-m}\Big)\vert 0\rangle
\end{multline}
since $\mathfrak{p}_m\vert 0\rangle=0$ for all $m>0$. Since $\mathfrak{p}_m$ is the
adjoint operator of $\mathfrak{p}_{-m}$ with respect to the scalar product
on $\mathbb{H}_{{\mathbb{C}}^2}$, we have
$\big\langle\vert 0\rangle,(\mathfrak{p}_{-m})^n\vert 0\rangle\big\rangle_{\mathbb{H}_{{\mathbb{C}}^2}}=0$
for all $m,n\geq1$ and the result follows.
\endproof
A similar formula for the $U(1)^{r+1}$ quiver gauge theory partition
function is conjectured in \cite[Appendix
C.1]{art:aldaygaiottotachikawa2010}.
\subsubsection{$A_0$ theory}
The degenerate limit $r=0$ of the $A_r$ quiver gauge theory is built
on the trivial quiver consisting of a single vertex with no arrows
\begin{equation}
\bullet
\end{equation}
and $m_0=2$ fundamental matter fields by Equation
\eqref{eq:confconstr}. Then the quiver gauge theory partition function
is the Nekrasov partition function for ${\mathcal N}=2$ gauge theory with two
fundamental matter hypermultiplets of masses $\mu_0,\mu_1$
\cite{art:nekrasov2003, art:bruzzofucitomoralestanzini2003} which is
given by
\begin{equation}
{\mathcal Z}_{{\mathbb{C}}^2}^{A_0}(\varepsilon_1,\varepsilon_2, \mu_0, \mu_1;
\mathsf{q}) = \sum_{\lambda} \, (-\mathsf{q})^{\vert\lambda\vert} \ \prod_{s\in Y_\lambda}\, \frac{\big(L'(s)\, \beta -A'(s)+\tilde \mu_0\big)\, \big(L'(s)\, \beta- A'(s)+\tilde \mu_1\big)}{\big((L(s)+1)\, \beta+A(s) \big)\, \big(L(s)\, \beta + A(s)+1\big)}
\end{equation}
as in \cite[Equation (3.22)]{art:bruzzofucitomoralestanzini2003},
where $\tilde \mu_0 =\mu_0/\varepsilon_2$ and $\tilde
\mu_1=\mu_1/\varepsilon_2$. By Proposition
\ref{prop:fundamental-matter} this partition function computes the
four-point conformal block for the Heisenberg algebra
$\mathfrak{h}_{{\mathbb{C}}^2}$ on the Riemann sphere with
primary field insertions at $\infty,1,\mathsf{q},0$, and by Proposition
\ref{prop:ZC2Arexpl} the combinatorial sum can be evaluated explicitly
with the result
\begin{equation}
{\mathcal Z}_{{\mathbb{C}}^2}^{A_0}(\varepsilon_1,\varepsilon_2, \mu_0, \mu_1;
\mathsf{q}) = (1-\mathsf{q})^{- \frac{\mu_1\, (\mu_0+\varepsilon_1+\varepsilon_2)}{\varepsilon_1\, \varepsilon_2}}\ .
\label{eq:fundamental}\end{equation}
A similar expression is written in \cite[Equation (2.27)]{art:wyllard2009}.
In the antidiagonal limit $\beta=1$, this formula coincides with the
partition function expression derived in \cite[Equation (49)]{art:marshakovmironovmorozov2010}.
\bigskip \section{Moduli spaces of framed sheaves\label{sec:sheaves}}
\subsection{Orbifold compactification of $X_k$}
In this subsection we recall the construction of the orbifold
compactification of the minimal resolution of ${\mathbb{C}}^2/{\mathbb{Z}}_k$ from
\cite[Section 3]{art:bruzzopedrinisalaszabo2013} and describe the main
results that we will use in this paper. For background to the theory of root and toric stacks used in the construction, see \cite[Section 2]{art:bruzzopedrinisalaszabo2013}, and to the theory of framed sheaves on (projective) Deligne-Mumford stacks, see \cite{art:bruzzosala2015}.
Fix an integer $k\geq 2$ and denote by $\mu_k$ the group of $k$-th
roots of unity in ${\mathbb{C}}$. A choice of a primitive $k$-th root of unity
$\omega$ defines an isomorphism of groups $\mu_k\simeq {\mathbb{Z}}_k$. We
define an action of $\mu_k\simeq {\mathbb{Z}}_k$ on ${\mathbb{C}}^2$ as $\omega\triangleright (x,y):= (\omega\, x, \omega^{-1}\, y)$.
The quotient ${\mathbb{C}}^2/{\mathbb{Z}}_k$ is a normal affine toric surface. The origin is the only singular point of ${\mathbb{C}}^2/{\mathbb{Z}}_k$, and is a particular case of a rational double point or du~Val singularity \cite[Definition~10.4.10]{book:coxlittleschenck2011}.
Let $\varphi_k\colon X_k\to{\mathbb{C}}^2/{\mathbb{Z}}_k$ be the minimal resolution of the
singularity of ${\mathbb{C}}^2/{\mathbb{Z}}_k$; it is a smooth toric surface with $k$
torus-fixed points $p_1, \ldots, p_k$ and $k+1$ torus-invariant
divisors $D_0, D_1, \ldots, D_k$ which are smooth projective curves of genus zero. For any $i=1, \ldots, k$ the divisors $D_{i-1}$ and $D_i$ intersect at the point $p_i$. Moreover, $D_1, \ldots, D_{k-1}$ are the irreducible components of the exceptional divisor $\varphi_k^{-1}(0)$.
By the McKay correspondence, there is a one-to-one
correspondence between the irreducible representations of $\mu_k$ and
the divisors $D_1, \ldots, D_{k-1}$ \cite[Corollary~10.3.11]{book:coxlittleschenck2011}. By \cite[Equation~(10.4.3)]{book:coxlittleschenck2011}, the intersection matrix $(D_i\cdot D_j)_{1\leq i,j\leq k-1}$ is given by minus the Cartan matrix $C$ of type $A_{k-1}$, i.e., one has
\begin{equation}
\left( D_i \cdot D_j \right)_{1\leq i,j\leq k-1} = -C=
\begin{pmatrix}
-2 & 1 & \cdots & 0 \\
1 & -2 & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots\\
0 & 0 & \cdots & -2
\end{pmatrix} \ .
\end{equation}
The surface $X_k$ is an ALE space of type $A_{k-1}$.
Let $U_i$ be the torus-invariant affine open subset of $X_k$ which is
a neighbourhood of the torus-fixed point $p_i$ for $i=1, \ldots,
k$. Its coordinate ring is given by ${\mathbb{C}}[U_i]:={\mathbb{C}}[T_1^{2-i}\, T_2^{1-i},T_1^{i-1}\, T_2^i]$ for $i=1, \ldots, k$. By imposing the change of variables $T_1=t_1^k$ and $T_2=t_2 \, t_1^{1-k}$, we have
\begin{equation}\label{eq:coordinatering-i}
{\mathbb{C}}[U_i]={\mathbb{C}}[t_1^{k-i+1}\,
t_2^{1-i},t_1^{i-k}\, t_2^{i}]\ .
\end{equation}
Define
\begin{equation}
\chi^i_1(t_1,t_2)=t_1^{k-i+1}\, t_2^{1-i}\qquad
\mbox{and}\qquad\chi_2^i(t_1,t_2)=t_1^{i-k}\, t_2^{i}\ .
\end{equation}
After identifying characters of $T$ with one-dimensional $T$-modules, let $\varepsilon_j^{(i)}$ denote the equivariant first Chern class of
$\chi^i_j$ for $i=1,\dots,k$ and $j=1,2$. Then
\begin{equation}
\varepsilon_1^{(i)}(\varepsilon_1,\varepsilon_2)=(k-i+1)\,
\varepsilon_1-(i-1)\,
\varepsilon_2\qquad\mbox{and}\qquad\varepsilon_2^{(i)}(\varepsilon_1,\varepsilon_2)=
-(k-i)\, \varepsilon_1+i\, \varepsilon_2\ .
\end{equation}
One can compactify the ALE space $X_k$ to a normal projective toric surface $\bar
X_k$ by adding a torus-invariant divisor $D_\infty\simeq {\mathbb{P}}^1$ such
that for $k=2$ the surface $\bar X_2$ coincides with the second
Hirzebruch surface ${\mathbb{F}}_2$. For $k\geq3$ the surface $\bar X_k$ is
singular, but one can associate with $\bar X_k$ its canonical toric
stack ${\mathscr{X}}_k^{\mathrm{can}}$ which is a two-dimensional projective
toric orbifold with Deligne-Mumford torus $T$ and coarse moduli space
$\pi_k^{\mathrm{can}}\colon {\mathscr{X}}_k^{\mathrm{can}}\to \bar{X}_k$. By
\emph{canonical} we mean that the locus over which
$\pi_k^{\mathrm{can}}$ is not an isomorphism has non-positive
dimension; for $k=2$ one has ${\mathscr{X}}_2^{\mathrm{can}}\simeq
{\mathbb{F}}_2$. Consider the one-dimensional, torus-invariant, integral closed
substack
$\tilde{{\mathscr{D}}}_\infty:=(\pi_k^{\mathrm{can}})^{-1}(D_\infty)_{\mathrm{red}}\subset{\mathscr{X}}_k^{\mathrm{can}}$.
By performing the $k$-th root construction on ${\mathscr{X}}_k^{\mathrm{can}}$
along $\tilde{{\mathscr{D}}}_\infty$ to extend the automorphism group of a
generic point of $\tilde{{\mathscr{D}}}_\infty$ by $\mu_k$, we obtain a two-dimensional projective toric orbifold ${\mathscr{X}}_k$ with Deligne-Mumford torus $T$ and coarse moduli space $\pi_k\colon{\mathscr{X}}_k\to\bar X_k$. The surface $X_k$ is isomorphic to the open subset ${\mathscr{X}}_k \setminus{\mathscr{D}}_\infty$ of ${\mathscr{X}}_k$, where ${\mathscr{D}}_\infty :=\pi_k^{-1}(D_\infty)_{\mathrm{red}}$. Let ${\mathscr{D}}_i:=\pi_k^{-1}(D_i)_{\mathrm{red}}$ be the divisors in ${\mathscr{X}}_k$ corresponding to $D_i$ for $i=1, \ldots, k-1$. The classes
\begin{equation}
- \sum_{j=1}^{k-1}\, \big(C^{-1}\big)^{ij} \, {\mathscr{D}}_j
\end{equation}
are integral for $i=1, \ldots, k-1$, where the inverse of the Cartan
matrix $C$ is given by
\begin{equation}
\big(C^{-1}\big)^{ij}=\frac{i\, (k-j)}k \qquad \mbox{for} \quad i\leq
j \ .
\end{equation}
Denote by ${\mathcal R}_i$ the associated line bundles on ${\mathscr{X}}_k$; the restrictions of ${\mathcal R}_i$ to $X_k$ are precisely the tautological line bundles of Kronheimer and Nakajima \cite{art:kronheimernakajima1990}.
\begin{proposition}[{\cite[Proposition 3.25]{art:bruzzopedrinisalaszabo2013}}]
The Picard group $\operatorname{Pic}({\mathscr{X}}_k)$ of ${\mathscr{X}}_k$ is freely generated over ${\mathbb{Z}}$ by ${\mathcal O}_{{\mathscr{X}}_k}({\mathscr{D}}_\infty)$ and ${\mathcal R}_i$ with $i=1, \ldots, k-1$.
\end{proposition}
The divisor ${\mathscr{D}}_\infty$ can be characterized as a toric Deligne-Mumford stack with Deligne-Mumford torus ${\mathbb{C}}^\ast\times {\mathscr{B}}\mu_k$ and coarse moduli space $r_k\colon {\mathscr{D}}_\infty\to D_\infty$.
\begin{proposition}[{\cite[Proposition 3.27]{art:bruzzopedrinisalaszabo2013}}]
The divisor ${\mathscr{D}}_\infty$ is isomorphic as a toric Deligne-Mumford stack to the global toric quotient stack
\begin{equation}
\left[\frac{{\mathbb{C}}^2\setminus\{0\}}{{\mathbb{C}}^\ast\times\mu_k}\right]\ ,
\end{equation}
where the group action is given in \cite[Equation (3.28)]{art:bruzzopedrinisalaszabo2013}.
\end{proposition}
\begin{corollary}[{\cite[Corollary 3.29]{art:bruzzopedrinisalaszabo2013}}]
The Picard group $\operatorname{Pic}({\mathscr{D}}_\infty)$ is isomorphic to ${\mathbb{Z}}\oplus
{\mathbb{Z}}_k$. It is generated by the line bundles ${\mathcal L}_1$ and ${\mathcal L}_2$
corresponding respectively to the characters
\begin{equation}
\chi_1\,\colon\, (t,\omega)\in{\mathbb{C}}^\ast\times\mu_k \ \longmapsto\
t\in{\mathbb{C}}^\ast\quad \mbox{and}\quad \chi_2\, \colon\,
(t,\omega)\in{\mathbb{C}}^\ast\times\mu_k \ \longmapsto\ \omega\in{\mathbb{C}}^\ast\ .
\end{equation}
\end{corollary}
For $j=0,1, \ldots, k-1$ define the line bundles
\begin{equation}
{\mathcal O}_{{\mathscr{D}}_\infty}(j)=\left\{
\begin{array}{ll}
{\mathcal L}_2^{\otimes j} & \mbox{for even $k$}\ ,\\[8pt]
{\mathcal L}_2^{\otimes j\, \frac{k+1}{2}} & \mbox{for odd $k$}\ .
\end{array}
\right.
\end{equation}
\begin{proposition}[{\cite[Corollary 3.34]{art:bruzzopedrinisalaszabo2013}}]
The restrictions of the tautological line bundles ${\mathcal R}_j$ to
${\mathscr{D}}_\infty$ are given by
\begin{equation}
{{\mathcal R}_j}\big\vert_{ {\mathscr{D}}_\infty}\simeq {\mathcal O}_{{\mathscr{D}}_\infty}(j)\ .
\end{equation}
\end{proposition}
\begin{remark}
In \cite{art:bruzzopedrinisalaszabo2013} the line bundles ${\mathcal O}_{{\mathscr{D}}_\infty}(j)$ are the line bundles ${\mathcal O}_{{\mathscr{D}}_\infty}(s,j)$ for $s=0$. Indeed, one can prove that the degree of ${\mathcal O}_{{\mathscr{D}}_\infty}(j)$ is zero. Moreover, ${\mathcal O}_{{\mathscr{D}}_\infty}(0,j)$ can be endowed with a unitary flat connection associated with the $j$-th
irreducible unitary representation $\rho_j$ of ${\mathbb{Z}}_k$ for $j=0,1, \ldots, k-1$ (cf.\ \cite[Remark 6.5]{art:eyssidieuxsala2013}).
\end{remark}
\subsection{Rank one framed sheaves}
\begin{definition}
Fix $j\in\{0,1,\dots,k-1\}$. A \emph{rank one $({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaf} on ${\mathscr{X}}_k$ is a pair $({\mathcal E}, \phi_{{\mathcal E}})$, where ${\mathcal E}$ is a
torsion-free sheaf on ${\mathscr{X}}_k$ of rank one which is locally free in a
neighbourhood of ${\mathscr{D}}_\infty$, and $\phi_{{\mathcal E}}\colon
{\mathcal E}\big\vert_{{\mathscr{D}}_\infty}\xrightarrow{\sim} {\mathcal O}_{{\mathscr{D}}_\infty}(j)$
is an isomorphism. We call $\phi_{{\mathcal E}}$ a \emph{framing} of
${\mathcal E}$. A \emph{morphism} between
$({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaves
$({\mathcal E},\phi_{\mathcal E})$ and $({\mathcal G},\phi_{\mathcal G})$ of rank one is a morphism
$f\colon {\mathcal E}\to{\mathcal G}$ such that $\phi_{\mathcal G}\circ
f\big\vert_{{\mathscr{D}}_\infty} = \phi_{\mathcal E}$.
\end{definition}
\begin{remark}
By \cite[Remark 4.3]{art:bruzzopedrinisalaszabo2013}, the Picard group
of ${\mathscr{X}}_k$ is isomorphic to the second singular cohomology group of
${\mathscr{X}}_k$ with integral coefficients via the first Chern class map $\operatorname{c}_1$. Thus fixing the determinant line bundle of a coherent sheaf ${\mathcal E}$ on ${\mathscr{X}}_k$ is equivalent to fixing its first Chern class.
\end{remark}
Given a vector $\vec{u}= (u_1,\ldots,u_{k-1}) \in{\mathbb{Z}}^{k-1}$, we denote by ${\mathcal R}^{\vec{u}}$ the line bundle $\bigotimes_{i=1}^{k-1}\, {\mathcal R}_i^{\otimes u_i}$ and by ${\mathcal R}_0$ the trivial line bundle ${\mathcal O}_{{\mathscr{X}}_k}$.
\begin{lemma}[{\cite[Lemma 4.4]{art:bruzzopedrinisalaszabo2013}}]
Let $({\mathcal E}, \phi_{{\mathcal E}})$ be a rank one $({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaf on ${\mathscr{X}}_k$. Then the determinant $\det({\mathcal E})$ of ${\mathcal E}$ is of the form ${\mathcal R}^{\vec{u}}$, where the vector $\vec{u}\in {\mathbb{Z}}^{k-1}$ satisfies the condition
\begin{equation}\label{eq:condition-determinant}
\sum_{i=1}^{k-1}\, i\, u_i= j \ \bmod{k} \ .
\end{equation}
\end{lemma}
\begin{remark}\label{rem:firstchern}
Set $\vec{v}:=C^{-1}\vec{u}$. Then Equation \eqref{eq:condition-determinant} implies the relations
\begin{equation}\label{eq:v-condition}
k\, v_l=-l\, j \ \bmod{k}
\end{equation}
for $l=1, \ldots, k-1$. Note that a component $v_l$ is integral if and
only if every component of $\vec v$ is integral. We subdivide the vectors $\vec{
u}\in{\mathbb{Z}}^{k-1}$ according to Equation \eqref{eq:condition-determinant} as
\begin{equation}
\mathfrak{U}_j:=\Big\{ \vec{u}\in{\mathbb{Z}}^{k-1} \ \Big\vert \
\mbox{$\sum\limits_{i=1}^{k-1}$}\, i\, u_i = j \ \bmod{k}
\Big\}\ .
\end{equation}
Define now a bijective map by identifying a vector
$\vec{u}\in{\mathbb{Z}}^{k-1}$ with $\sum_{i=1}^{k-1} \, u_i \operatorname{c}_1({\mathcal R}_i) =
\sum_{i=1}^{k-1}\, u_i \, \omega_i$ as
\begin{equation}
\psi\, \colon\, \vec{u}\in {\mathbb{Z}}^{k-1} \ \longmapsto \ \sum_{i=1}^{k-1}
\, u_i \, \omega_i\in \mathfrak{P}\ .
\end{equation}
It is natural to split this map according to the coset decomposition
\eqref{eq:lateralweightlattice} as
\begin{equation}
\psi^{-1}(\mathfrak{Q} + \omega_j) = \mathfrak{U}_j\ ,
\end{equation}
which means that $\psi(\vec{u}\, )$ for $\vec{u}\in \mathfrak{U}_j$ is naturally written as
a sum of the fundamental weight $\omega_j$ and an element $\gamma_{\vec{u}}$ of the root
lattice $\mathfrak{Q}$, which is given by
\begin{equation}
\gamma_{\vec{u}}:=\sum_{i=1}^{k-1}\, \Big(\, \sum_{l=1}^{k-1} \, \big(C^{-1}\big)^{
il}\, u_l-\big(C^{-1}\big)^{ij}\, \Big)\, \gamma_i =
\sum_{i=1}^{k-1}\, \Big( v_i -\big(C^{-1}\big)^{ij}\Big)\, \gamma_i \
\in \ \mathfrak{Q}\ .
\end{equation}
We write
\begin{equation}
\psi_j:=\psi\big\vert_{\mathfrak{U}_j} \, \colon \, \mathfrak{U}_j \ \longrightarrow \
\mathfrak{Q} + \omega_j \ .
\label{eq:psij}\end{equation}
\end{remark}
Following \cite[Section 4]{art:bruzzopedrinisalaszabo2013}, let
$\mathcal{M}(\vec{u},n,j)$ be the fine moduli space parameterizing
$({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaves of rank one on
${\mathscr{X}}_k$ with determinant line bundle ${\mathcal R}^{\vec{u}}$ and second
Chern class $n\in {\mathbb{Z}}$; the vector $\vec{u}$ belongs to $\mathfrak{U}_j$. Let $p_{{\mathscr{X}}_k}:{\mathscr{X}}_k\times \mathcal{M}(\vec{u},n,j) \to {\mathscr{X}}_k$ be the projection.
As explained in \cite[Remark 4.7]{art:bruzzopedrinisalaszabo2013}, by ``fine'' one means that there exists a \emph{universal framed sheaf} $(\boldsymbol{{\mathcal E}}, \boldsymbol{\phi}_{\boldsymbol{{\mathcal E}}})$, where
$\boldsymbol{{\mathcal E}}$ is a coherent sheaf on $\mathcal{M}(\vec{u},n,j)\times {\mathscr{X}}_k$ which is flat over $\mathcal{M}(\vec{u},n,j)$, and $\boldsymbol{\phi}_{\boldsymbol{{\mathcal E}}}\colon \boldsymbol{{\mathcal E}}\to
p_{{\mathscr{X}}_k}^\ast({\mathcal O}_{{\mathscr{D}}_\infty}(j))$ is a morphism such that
its restriction to $\mathcal{M}(\vec{u},n,j)\times {\mathscr{D}}_\infty$ is an isomorphism; the fibre over $[({\mathcal E},\phi_{\mathcal E})]\in \mathcal{M}(\vec{u},n,j)$ is itself the $({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaf $({\mathcal E},\phi_{\mathcal E})$ on ${\mathscr{X}}_k$. In the following we shall call $\boldsymbol{{\mathcal E}}$ the \emph{universal sheaf}.
\begin{theorem}[{\cite[Theorem 4.13]{art:bruzzopedrinisalaszabo2013}}]
The moduli space $\mathcal{M}(\vec{u},n,j)$ is a smooth quasi-projective variety of dimension $2n$. The Zariski tangent
space of
$\mathcal{M}(\vec{u},n,j)$ at a point $[({\mathcal E},\phi_{{\mathcal E}})]$ is $\mathrm{Ext}^1({\mathcal E},{\mathcal E}\otimes {\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty))$.
\end{theorem}
As explained in \cite[Section 4.3]{art:bruzzopedrinisalaszabo2013}, the Hilbert
scheme of $n$ points $\hilb{n}{X_k}$ of $X_k$ is isomorphic to
$\mathcal{M}(\vec u,n,j)$ for any $\vec{u}\in \mathfrak{U}_j$. For this, let $\imath:X_k\hookrightarrow {\mathscr{X}}_k$ be the
inclusion morphism. If $Z$ is a point of $\hilb{n}{X_k}$ and
$\vec{y}\in{\mathbb{Z}}^{k-1}$, then the coherent sheaf
${\mathcal E}:=\imath_\ast({\mathcal I}_Z)\otimes {\mathcal R}^{\boldsymbol e_j-C\vec{y}}$ is a
rank one torsion-free sheaf on $\mathscr{X}_k$ with a framing
$\phi_{{\mathcal E}}$ induced by the canonical isomorphism
${\mathcal R}^{\boldsymbol e_j-C\vec{y}}\, \vert_{{\mathscr{D}}_\infty} \xrightarrow{\sim}
{\mathcal O}_{{\mathscr{D}}_\infty}(j)$ such that $({\mathcal E}, \phi_{{\mathcal E}})$ is a rank
one $({\mathscr{D}}_\infty,{\mathcal O}_{{\mathscr{D}}_\infty}(j))$-framed sheaf with
determinant line bundle ${\mathcal R}^{\vec{u}}$, where $\vec{u}:=\boldsymbol
e_j-C\vec{y}$, and second Chern class $n$. Thus $Z$ induces a point $[(\mathcal{E},\phi_{\mathcal{E}})]$ in $\mathcal{M}(\vec{u},n,j)$. This defines an inclusion morphism
\begin{equation}
\tilde\imath_{\vec{u},n,j}\, \colon\, \hilb{n}{X_k} \ \hookrightarrow \ \mathcal{M}(\vec{u},n,j)
\end{equation}
which is an isomorphism of fine moduli spaces by \cite[Proposition 4.16]{art:bruzzopedrinisalaszabo2013}.
\begin{remark}\label{rem:quivervariety}
In \cite{art:kuznetsov2007} it is shown that the Hilbert scheme of points $\hilb{n}{X_k}$ is isomorphic to a Nakajima quiver variety of type $\widehat{A}_{k-1}$ with suitable dimension vectors. Thus $\mathcal{M}(\vec{u},n,j)$ is a quiver variety.
\end{remark}
\subsection{Equivariant cohomology}
We define a $T$-action on $\mathcal{M}(\vec{u},n,j)$ in the following way (cf.\ \cite[Section 4.6]{art:bruzzopedrinisalaszabo2013}). For $(t_1,t_2)\in T$ let $F_{(t_1,t_2)}$ be the automorphism of ${\mathscr{X}}_k$ induced by the torus action; then the $T$-action is given by
\begin{equation}
(t_1,t_2)\triangleright \big[({\mathcal E},\phi_{{\mathcal E}}) \big]:=\big[\big( (F_{(t_1,t_2)}^{-1})^\ast({\mathcal E})\,,\, \phi_{\mathcal E}'\big) \big] \ ,
\end{equation}
where $\phi_{\mathcal E}'$ is the composition of isomorphisms
\begin{equation}
\phi_{\mathcal E}' \, \colon\, \big(F_{(t_1,t_2)}^{-1} \big)^\ast{\mathcal E} \big\vert_{{\mathscr{D}}_\infty} \ \xrightarrow{(F_{(t_1,t_2)}^{-1})^\ast(\phi_{{\mathcal E}})} \
\big(F_{(t_1,t_2)}^{-1} \big)^\ast {\mathcal O}_{{\mathscr{D}}_\infty}(j)\ \longrightarrow \ {\mathcal O}_{{\mathscr{D}}_\infty}(j) \ ;
\end{equation}
here the last arrow is given by the $T$-equivariant structure induced
on ${\mathcal O}_{{\mathscr{D}}_\infty}(j)$ by restriction of the torus action of
${\mathscr{X}}_k$ to $\mathscr D_\infty$.
Note that the $T$-action on $X_k$ naturally lifts to $\hilb{n}{X_k}$ and the isomorphism $\tilde\imath_{\vec{u},n,j}$ is equivariant with respect to these torus actions.
\begin{proposition}[{\cite[Proposition 4.22]{art:bruzzopedrinisalaszabo2013}}]\label{prop:fixedpoint}
For a $T$-fixed point $[({\mathcal E},\phi_{\mathcal E})]\in \mathcal{M}(\vec{u},n,j)^T$
the underlying sheaf is of the form ${\mathcal E}=\imath_\ast ({\mathcal I}_Z)\otimes{\mathcal R}^{\vec{u}}$, where ${\mathcal I}_Z$ is the ideal sheaf of a $T$-fixed zero-dimensional subscheme $Z$ of $X_k$.
\end{proposition}
\begin{remark}\label{rem:fixedpoints}
Let $[({\mathcal E},\phi_{\mathcal E})]$ be a $T$-fixed point of
$\mathcal{M}(\vec{u},n,j)$, with
${\mathcal E}=\imath_\ast({\mathcal I}_Z)\otimes{\mathcal R}^{\vec{u}}$. The $T$-fixed
subscheme $Z$ of $X_k$ of length $n$ is a disjoint union of $T$-fixed
subschemes $Z_i$ for $i=1, \ldots, k$ supported at the $T$-fixed
points $p_i$ with $\sum_{i=1}^k\,
\mathrm{length}_{p_i}(Z_i)=n$. Put
$n_i=\mathrm{length}_{p_i}(Z_i)$. Since $p_i$ is the $T$-fixed point
of the smooth affine toric surface $U_i\simeq {\mathbb{C}}^2$, as explained in
Section \ref{sec:C2} the $T$-fixed subscheme $Z_i\in \hilb{n_i}{U_i}$
corresponds to a Young tableau $Y^i$ of weight $\vert Y^i\vert= n_i$
for $i=1, \ldots, k$. Thus the $T$-fixed point $Z$ corresponds to a
$k$-tuple of Young tableaux $\vec{Y}=(Y^1, \ldots, Y^k)$ with
$\vert\vec{Y}\vert:=\sum_{i=1}^k\, \vert Y^i\vert=n$. Hence we can
parametrize the point $[({\mathcal E}, \phi_{{\mathcal E}})]$ by the pair $(\vec{Y},
\vec{u}\, )$ which we call the \emph{combinatorial datum} of $[({\mathcal E}, \phi_{{\mathcal E}})]$.
\end{remark}
Consider the $T$-equivariant cohomology of the moduli spaces $\mathcal{M}(\vec{u},n,j)$ and set
\begin{equation}
{\mathbb{W}}_{\vec{u},n,j}:=H^\ast_T \big( \mathcal{M}(\vec{u},n,j) \big)_{\mathrm{loc}}\ .
\end{equation}
We endow ${\mathbb{W}}_{\vec{u},n,j}$ with the nondegenerate ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$-valued bilinear form
\begin{equation}
\langle A,B\rangle_{{\mathbb{W}}_{\vec{u},n,j}}:=(-1)^n \, p_{\vec{u}, n,j}^!
\, \big(\imath_{\vec{u}, n,j}^!\big)^{-1} (A \cup B)\ ,
\end{equation}
where $p_{\vec{u}, n,j}$ is the projection from $\mathcal{M}(\vec{u},n,j)$ to a point and $\imath_{\vec{u}, n,j}\colon \mathcal{M}(\vec{u},n,j)^T \hookrightarrow \mathcal{M}(\vec{u},n,j)$ is the inclusion of the fixed-point locus. Thus for $\vec{u}\in\mathfrak{U}_j$ we define
\begin{equation}\label{eq:L0eigenspace}
{\mathbb{W}}_{\vec{u},j}:= \bigoplus_{n \geq0 } \, {\mathbb{W}}_{\vec{u},n,j} \ ,
\end{equation}
and the total equivariant cohomology
\begin{equation}
{\mathbb{W}}_j:=\bigoplus_{\vec{u}\in \mathfrak{U}_j} \ \bigoplus_{n \geq0 }\,
{\mathbb{W}}_{\vec{u},n,j}
\end{equation}
which is an infinite-dimensional vector space over the field ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$
endowed with the nondegenerate ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$-valued bilinear form $\langle -,- \rangle_{{\mathbb{W}}_j}$
induced by the symmetric bilinear forms $\langle -,-\rangle_{{\mathbb{W}}_{\vec{u},n,j}}$.
Let us denote by $[\vec{Y}, \vec{u}\, ]$ the equivariant cohomology
class, defined similarly to \eqref{eq:fixed-class}, associated with the
$T$-fixed point $[({\mathcal E}, \phi_{{\mathcal E}})]$ with combinatorial datum
$(\vec{Y}, \vec{u}\, )$. By the localization theorem, the classes
$[\vec{Y}, \vec{u}\, ]$ with $\vec{u}\in \mathfrak{U}_j$ form a ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-basis of ${\mathbb{W}}_j$.
\bigskip \section{Representations of $\widehat{\mathfrak{gl}}_k$\label{sec:glkreps}}
\subsection{Overview}
The results collected so far imply the following result.
\begin{proposition}\label{cor:cohomologymodulihilbert}
There is an isomorphism
\begin{multline}\label{eq:isomWH}
\Psi_j \, \colon\, {\mathbb{W}}_j \ \xrightarrow{ \ \sim \ } \
\bigoplus_{\vec{u}\in\mathfrak{U}_j} \ \bigoplus_{ n\geq0 } \,
H^\ast_T\big(\hilb{n}{X_k} \big)_{\mathrm{loc}}\\[4pt]
\simeq \Big(\, \bigoplus_{n\geq0} \, H^\ast_T\big(\hilb{n}{X_k}
\big)_{\mathrm{loc}} \, \Big) \otimes {\mathbb{C}}(\varepsilon_1, \varepsilon_2)[\mathfrak{U}_j]\\[4pt]
\xrightarrow{ \ \sim \ } \ \Big(\, \bigoplus_{n\geq0 } \,
H^\ast_T\big(\hilb{n}{X_k} \big)_{\mathrm{loc}} \, \Big) \otimes {\mathbb{C}}(\varepsilon_1, \varepsilon_2)[\mathfrak{Q}+\omega_j]\ ,
\end{multline}
where the first arrow is induced by the morphisms
$\tilde\imath_{\vec{u},n,j}^{\,\ast}$ while the last arrow is induced
by the map $\psi_j$ introduced in \eqref{eq:psij}. There is also an isomorphism
\begin{equation}
\Psi:=\bigoplus_{j=0}^{k-1}\, \Psi_j\, \colon \, {\mathbb{W}} \ \xrightarrow{\
\sim \ } \ \bigoplus_{j=0}^{k-1}\ \bigoplus_{n\geq0 } \,
H^\ast_T\big(\hilb{n}{X_k} \big)_{\mathrm{loc}} \otimes {\mathbb{C}}(\varepsilon_1, \varepsilon_2)[\mathfrak{Q}+\omega_j]\ ,
\end{equation}
where ${\mathbb{W}}:=\bigoplus_{j=0}^{k-1}\, {\mathbb{W}}_j$.
\end{proposition}
In this section we first study the equivariant cohomology of
$\hilb{n}{X_k}$ and construct over it an action of the sum
(identifying central elements) $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}\oplus
\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2),\mathfrak{Q}}$. Then we use the
Frenkel-Kac construction (Theorem \ref{thm:frenkelkac}) to obtain an
action of $\widehat{\mathfrak{gl}}_k=\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}
\oplus\widehat{\mathfrak{sl}}_k$ on ${\mathbb{W}}_j$ for $j=0, 1, \ldots, k-1$.
\subsection{Equivariant cohomology of $\hilb{n}{X_k}$}
In this subsection we derive some results concerning the equivariant
cohomology of the Hilbert schemes $\hilb{n}{X_k}$ by generalizing similar results of \cite[Section 2]{art:qinwang2007} (see also \cite[Section 2]{art:maulikoblomov2009}).
As discussed in Remark \ref{rem:fixedpoints}, a $T$-fixed point
$Z\in\hilb{n}{X_k}$ corresponds to a $k$-tuple $(Z_1, \ldots, Z_k)$
where $Z_i$ is a $T$-fixed point of $\hilb{n_i}{U_i}$ for $i=1,\ldots,
k$ with $\sum_{i=1}^k\, n_i=n$, or equivalently to a $k$-tuple $\vec
Y= (Y^1,
\ldots, Y^k)$ of Young tableaux with $|\vec Y|:= \sum_{i=1}^k\, \vert Y^i\vert = n$. The following result is straightforward to prove.
\begin{lemma}\label{lem:decomposition}
Let $Z$ be a $T$-fixed point of $\hilb{n}{X_k}$. Then there is a $T$-equivariant isomorphism
\begin{equation}
T_Z\hilb{n}{X_k}\simeq \bigoplus_{i=1}^k \, T_{Z_i}\hilb{n_i}{U_i}\ ,
\end{equation}
where $Z=\bigsqcup_{i=1}^k\, Z_i$ and $n_i$ is the length of $Z_i$ at $p_i$ for $i=1, \ldots, k$.
\end{lemma}
By Lemma \ref{lem:decomposition} we get
\begin{equation}
\mathrm{ch}_T\big(T_Z\hilb{n}{X_k} \big)=\sum_{i=1}^k\, \mathrm{ch}_T
\big(T_{Z_i}\hilb{n_i}{U_i} \big)\ .
\end{equation}
By using the description \eqref{eq:coordinatering-i} of the coordinate
ring ${\mathbb{C}}[U_i]$ of $U_i$, one computes the equivariant Chern characters
\begin{equation}
\mathrm{ch}_T\big(T_{Z_i}\hilb{n_i}{U_i} \big)=\sum_{s\in Y^i}\,
\big({\,\rm e}\,^{(L(s)+1)\, \varepsilon_1^{(i)}-A(s)\,
\varepsilon_2^{(i)}}+{\,\rm e}\,^{-L(s)\, \varepsilon_1^{(i)}+(A(s)+1)\, \varepsilon_2^{(i)}}\big)\ .
\end{equation}
From now on we identify a torus-fixed point $Z$ of $\hilb{n}{X_k}$ with its $k$-tuple $\vec{Y}$ of Young tableaux.
Let $\vec{Y}=(Y^1, \ldots, Y^k)$ be a torus-fixed point. Define
\begin{align}
\operatorname{eu}_{+}(\vec{Y})&:=\prod_{i=1}^k \ \prod_{s\in Y^i} \,
\Big( \big(L(s)+1\big)\, \varepsilon_1^{(i)}-A(s)\, \varepsilon_2^{(i)}\Big)\ ,\\[4pt]
\operatorname{eu}_{-}(\vec{Y})&:=\prod_{i=1}^k \ \prod_{s\in Y^i}\,
\Big(L(s)\, \varepsilon_1^{(i)}-\big(A(s)+1\big)\, \varepsilon_2^{(i)}\Big)\ .
\end{align}
Then the equivariant Euler class of the tangent bundle at the fixed
point $\vec{Y}$ is given by
\begin{equation}
\operatorname{eu}_T\big(T_{\vec{Y}}\hilb{n}{X_k}\big)=(-1)^n \,
\operatorname{eu}_{+}(\vec{Y})\, \operatorname{eu}_{-}(\vec{Y})\ .
\end{equation}
\subsubsection{Equivariant basis I: \ Torus-fixed points}\label{sec:basisfixedpoints}
Let $\vec{Y}$ be a $k$-tuple of Young tableaux corresponding to a
fixed point in $\hilb{n}{X_k}$ and $[\vec{Y}]$ the equivariant
cohomology class defined similarly to \eqref{eq:fixed-class}. By the projection formula we get
\begin{equation}
[\vec{Y}]\cup[\vec{Y}'\,]=\delta_{\vec{Y},\vec{Y}'}\,
\operatorname{eu}_T\big(T_{\vec{Y}}\hilb{n}{X_k} \big)[Y]=(-1)^n\,
\delta_{\vec{Y},\vec{Y}'}\, \operatorname{eu}_+(\vec{Y})\, \operatorname{eu}_-(\vec{Y})[\vec{Y}]\ .
\end{equation}
Denote
\begin{equation}
\imath_n:=\bigoplus_{\vec{Y}\in\hilb{n}{X_k}^T}\, \imath_{\vec{Y}}\,
\colon \, \hilb{n}{X_k}^T\ \longrightarrow \ \hilb{n}{X_k}\ .
\end{equation}
In analogy to Equation \eqref{eq:bilinearform-C2}, define the bilinear form
\begin{align}\label{eq:bilinearform}
\langle -,- \rangle_{\mathbb{H}_n} \, \colon \, \mathbb{H}_n\times
\mathbb{H}_n &\ \longrightarrow \ {\mathbb{C}}(\varepsilon_1, \varepsilon_2) \
,\\ \nonumber
(A, B)&\ \longmapsto \ (-1)^n\, p_n^!\, \big(\imath_n^! \big)^{-1}(A \cup B)
\end{align}
where $\mathbb{H}_n:=H_T^{\ast}(\hilb{n}{X_k})_{\mathrm{loc}}$.
As in Section \ref{sec:C2}, for any class $[\vec{Y}]\in H_T^{4n}(\hilb{n}{X_k})$ we define a distinguished class
\begin{equation}
[\alpha_{\vec{Y}}]:=\frac{[\vec{Y}]}{\operatorname{eu}_+(\vec{Y})} \ \in
\ H_T^{2n}\big(\hilb{n}{X_k} \big)_{\mathrm{loc}}\ .
\end{equation}
Then by the same computation as in Equation \eqref{eq:product-C2} we get
\begin{equation}\label{eq:product}
\big\langle [\alpha_{\vec{Y}}]\,,\, [\alpha_{\vec{Y}'}]
\big\rangle_{\mathbb{H}_n} = \delta_{\vec{Y},\vec{Y}'} \,
\frac{\operatorname{eu}_-(\vec{Y})}{\operatorname{eu}_+(\vec{Y})} =
\delta_{\vec{Y},\vec{Y}'} \ \prod_{i=1}^k \ \prod_{s\in Y^i} \, \frac{
L(s)\, \beta_i+A(s)+1 }{ \big(L(s)+1 \big)\, \beta_i+A(s) } \ ,
\end{equation}
where analogously to \eqref{eq:beta} we defined
\begin{equation}\label{eq:betai}
\beta_i:=-\frac{\varepsilon_1^{(i)}}{\varepsilon_2^{(i)}}\ .
\end{equation}
Note that when $n=1$, $\vec{Y}$ is just a fixed point $p_i\in X_k^T$ with $i=1, \ldots, k$. Thus we have
\begin{equation}
\operatorname{eu}_{+}(p_i) =\varepsilon_1^{(i)}=(k-i+1)\,
\varepsilon_1-(i-1)\, \varepsilon_2 \qquad \mbox{and} \qquad
\operatorname{eu}_{-}(p_i)=-\varepsilon_2^{(i)}=(k-i)\,
\varepsilon_1-i\, \varepsilon_2\ ,
\end{equation}
and therefore
\begin{equation}
\beta_i=\frac{\operatorname{eu}_+(p_i)}{\operatorname{eu}_-(p_i)}\ .
\end{equation}
If for $i=1, \ldots, k$ we define $[\alpha_i]:=[\alpha_{p_i}]$, then we get
\begin{equation}
\big\langle [\alpha_i]\,,\, [\alpha_j] \big\rangle_{\mathbb{H}_1} =
\beta^{-1}_i \, \delta_{ij} \ \in \ {\mathbb{C}}(\varepsilon_1, \varepsilon_2)\ .
\end{equation}
By the localization theorem and Equation \eqref{eq:product}, the
classes $[\alpha_{\vec{Y}}]$ with $\vert \vec{Y}\vert=n$ form a
${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear basis of
$\mathbb{H}_n$. Hence the bilinear form \eqref{eq:bilinearform} is
nondegenerate; it extends to give a nondegenerate symmetric bilinear
form $\langle -,- \rangle_{\mathbb{H}}$ on the total equivariant cohomology $\mathbb{H}:=\bigoplus_{n\geq0}\, {\mathbb{H}}_n$ of the Hilbert schemes of points on $X_k$.
\begin{remark}\label{rem:description-Ui}
Let $i\in \{1, \ldots, k\}$. By the localization theorem, the
${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear subspace of $\mathbb{H}$
generated by all classes $[\vec{Y}]$ associated to fixed points
$\vec{Y}=(Y^1, \ldots, Y^k)$ such that $Y^j=\emptyset$ for every $j\in
\{1, \ldots, k\}$ with $j\neq i$ is isomorphic to
\begin{equation}\label{eq:cohomologyUi}
\bigoplus_{m\geq 0} \, H_T^\ast\big(\hilb{m}{U_i} \big)\otimes_{\mathbb{C}[\varepsilon_1^{(i)},\varepsilon_2^{(i)}]}\mathbb{C}\big(\varepsilon_1^{(i)},\varepsilon_2^{(i)}\big) \ .
\end{equation}
\end{remark}
Note that ${\mathbb{C}}\big[\varepsilon_1^{(i)},\varepsilon_2^{(i)}
\big]={\mathbb{C}}[\varepsilon_1,\varepsilon_2]$ and
${\mathbb{C}}\big(\varepsilon_1^{(i)},\varepsilon_2^{(i)}
\big)={\mathbb{C}}(\varepsilon_1,\varepsilon_2)$. Analogously to what we did for
${\mathbb{C}}^2$, we can thus define
\begin{equation}
{\mathbb{H}}_{U_i,m} := H_T^\ast\big(\hilb{m}{U_i}
\big)_{\mathrm{loc}}\qquad\mbox{and}\qquad{\mathbb{H}}_{U_i} := \bigoplus_{m\geq 0}
\, {\mathbb{H}}_{U_i,m}\ .
\end{equation}
By the localization theorem, there exists a ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear isomorphism
\begin{equation}\label{eq:omega}
\Omega\, \colon \, \mathbb{H} \ \xrightarrow{ \ \sim \ } \
\bigotimes_{i=1}^k\, {\mathbb{H}}_{U_i} \ .
\end{equation}
In particular, for a fixed point $\vec{Y}=(Y^1, \ldots, Y^k)$ we have
\begin{equation}
\Omega\, \colon \, [\alpha_{\vec{Y}}] \ \longmapsto \ [\alpha_{Y^1}]\otimes \cdots \otimes [\alpha_{Y^k}]\ .
\end{equation}
The isomorphism $\Omega$ interwines the bilinear forms $\langle -,-
\rangle_{\mathbb{H}}$ and $\prod_{i=1}^k\, \langle -,- \rangle_i$,
where $\langle -,- \rangle_i$ is the symmetric bilinear form on
${\mathbb{H}}_{U_i}$ defined analogously to \eqref{eq:bilinearform-C2}.
In a similar way, there is a ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear isomorphism
\begin{equation}\label{eq:omegak}
\Omega_k\, \colon\, {\mathbb{H}}_1 \ \xrightarrow{ \ \sim \ } \
\bigoplus_{i=1}^k \, {\mathbb{H}}_{U_i,1}\ .
\end{equation}
In this case $\displaystyle\Omega_k\colon [\alpha_i]\mapsto (0,
\ldots, [\alpha_i], \ldots, 0)$, where the class $[\alpha_i]$ on the
left-hand side belongs to ${\mathbb{H}}_1=H_T^\ast(X_k)_{\mathrm{loc}}$ while on the right-hand
side it belongs to ${\mathbb{H}}_{U_i,1}$ as defined in Section
\ref{sec:C2}. The isomorphism $\Omega_k$ also intertwines the symmetric bilinear forms.
\subsubsection{Equivariant basis II: \ Torus-invariant divisors}\label{sec:divisors}
Let $[D_i]_T$ be the class in ${\mathbb{H}}_1=H_T^\ast(X_k)_{\mathrm{loc}}$ given by the
$T$-invariant divisor $D_i$ for $i=0,1, \ldots, k$. For
$i=1,\dots,k-1$, we have
\begin{equation}\label{eq:divisor}
[D_i]_T=\frac{[p_i]}{\operatorname{eu}_T(T_{p_i}D_i)}+\frac{[p_{i+1}]}{\operatorname{eu}_T(T_{p_{i+1}}D_i)}=\frac{[p_i]}{\varepsilon_2^{(i)}}+\frac{[p_{i+1}]}{\varepsilon_1^{(i+1)}}=
-\beta_i\, [\alpha_i]+[\alpha_{i+1}]\ .
\end{equation}
Thus for $i,j =1, \ldots, k-1$ we obtain the pairings
\begin{equation}\label{eq:divisorcartan}
\big\langle[D_i]_T\,, \,[D_j]_T \big\rangle_{\mathbb{H}_1}=\left\{
\begin{array}{ll}
{2} \ ,& i=j\ ,\\
{-1} \ , & \vert i-j\vert =1\ ,\\
{0} \ , & \mbox{otherwise}\ .
\end{array}\right.
\end{equation}
By applying the localization theorem to $[D_0]_T$ and $[D_k]_T$ we
further obtain
\begin{equation}
\left[D_0\right]_T=\frac{[p_1]}{k\, \varepsilon_1}=\frac{[p_1]}{\varepsilon_1^{(1)}}=[\alpha_1]\qquad\mbox{and}\qquad
\left[D_k\right]_T=\frac{[p_k]}{k\,
\varepsilon_2}=\frac{[p_k]}{\varepsilon_2^{(k)}}=-\beta_k\, [\alpha_k]\ .
\end{equation}
By using these expressions, one can straightforwardly obtain the pairings
\begin{equation*}
\big\langle [D_0]_T\,,\,[D_i]_T \big\rangle_{\mathbb{H}_1}=\left\{
\begin{array}{ll}
{\beta_1^{-1}} \ , & i=0\ ,\\
{-1} \ , & i=1\ ,\\
{0} \ , & \mbox{otherwise}
\end{array}\right.
\qquad\mbox{and}\qquad \big\langle [D_k]_T\,,\,[D_i]_T \big\rangle_{\mathbb{H}_1}=\left\{
\begin{array}{ll}
{\beta_k} \ , & i=k\ ,\\
{-1} \ , & i=k-1\ ,\\
{0} \ , & \mbox{otherwise}\ .
\end{array}\right.
\end{equation*}
Now we can relate the classes $[\alpha_i]$ for $i=1, \ldots, k$ to the
classes $[D_j]_T$ for $j=0,1, \ldots, k$. By using Equation
\eqref{eq:divisor}, for $i=2, \ldots, k$ one obtains
\begin{equation}\label{eq:alpha}
[\alpha_i]=\sum_{j=0}^{i-2}\, \Big(\, \prod_{s=j+1}^{i-1}\,
\beta_s\, \Big)\, [D_j]_T+[D_{i-1}]_T\ .
\end{equation}
Since $\operatorname{eu}_+(p_l)=\operatorname{eu}_-(p_{l-1})$ for $l=2,
\ldots, k$, we get
\begin{equation}
\prod_{s=j+1}^{i-1}\,
\beta_s=\frac{\operatorname{eu}_+(p_{j+1})}{\operatorname{eu}_-(p_{i-1})} \
.
\end{equation}
By using the definition of $[\alpha_k]$ and Equation \eqref{eq:alpha} for $i=k$ we obtain
\begin{equation}
-\beta_k^{-1}\, [D_k]_T=[\alpha_k] =\sum_{j=0}^{k-1}\,
\frac{\operatorname{eu}_+(p_{j+1})}{\operatorname{eu}_-(p_{k-1})}\ [D_j]_T \
.
\end{equation}
If we formally put $\operatorname{eu}_+(p_{k+1}):=\operatorname{eu}_-(p_k)$,
we can reformulate this equation as
\begin{equation}\label{eq:sum}
\sum_{j=0}^{k}\, \operatorname{eu}_+(p_{j+1})\ [D_j]_T=0\ ,
\end{equation}
and in particular for all $i=0,1,\dots,k$ we have
\begin{equation}
[D_i]_T=-\sum_{\stackrel{\scriptstyle j=0} {\scriptstyle j\neq i}}^k\,
\frac{\operatorname{eu}_+(p_{j+1})}{\operatorname{eu}_+(p_{i+1})} \ [D_j]_T\ .
\end{equation}
\begin{remark}
If the action is antidiagonal, i.e., $t=t_1=t_2^{-1}$, Equation \eqref{eq:sum} implies that $\sum_{j=0}^k\, [D_j]_T=0$.
\end{remark}
As the classes $[\alpha_1], \ldots, [\alpha_k]$ form a
${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear basis of $\mathbb{H}_1$, by
Equations \eqref{eq:alpha} and \eqref{eq:sum} the classes
\begin{equation}\label{eq:bases}
\big\{[D_0]_T, [D_1]_T, \ldots, [D_{k-1}]_T
\big\}\qquad\mbox{and}\qquad \big\{[D_1]_T, [D_2]_T, \ldots, [D_{k}]_T
\big\}
\end{equation}
are also ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear bases for $\mathbb{H}_1$. Under the isomorphism $\Omega_k$ of Equation \eqref{eq:omegak}, we have
\begin{equation}
\Omega_k\, \colon\, [D_i]_T \ \longmapsto \ -\beta_i \, \big(0,
\ldots,0, [\alpha_i], 0, \ldots, 0 \big)+\big(0, \ldots,0,
[\alpha_{i+1}], 0, \ldots, 0 \big)
\end{equation}
for $i=1, \ldots, k-1$, together with a similar description for $[D_0]_T$ and $[D_k]_T$.
\subsection{Heisenberg algebras}\label{sec:heisenbergalbebra}
Let $m$ be a positive integer and $Y$ a torus-invariant closed curve
in $X_k$. Define the correspondences
\begin{equation}
Y_{n,m}:=\big\{(Z, Z'\,)\in \hilb{n+m}{X_k}\times \hilb{n}{X_k}\
\big\vert\ Z'\subset Z \ , \
\mathrm{supp}(\mathcal{I}_{Z'}/\mathcal{I}_Z)=\{y\}\subset Y \big\} \ .
\end{equation}
Let $q_1$ and $q_2$ be the projections of $\hilb{n+m}{X_k}\times
\hilb{n}{X_k}$ to the two factors respectively. We define the linear
operator $\mathfrak{p}_{-m}([Y]_T)\colon \mathbb{H} \to \mathbb{H}$ which
acts on $A\in \mathbb{H}_n$ as
$\mathfrak{p}_{-m}([Y]_T)(A):=q_1^!\big(q_2^\ast(A)\cup
[Y_{n,m}]_T \big)\in \mathbb{H}_{n+m}$. This definition is well-posed
because the restriction of $q_1$ to $Y_{n,m}$ is proper. Since the
bilinear form $\langle -,- \rangle_{\mathbb{H}}$ is nondegenerate on
$\mathbb{H}$, we can define $\mathfrak{p}_m([Y]_T)$ to be the adjoint
operator of $\mathfrak{p}_{-m}([Y]_T)$. By using one of the two bases in \eqref{eq:bases}, we extend by linearity in $\alpha$ to obtain the linear operator $\mathfrak{p}_m(\alpha)$ for every $\alpha\in\mathbb{H}_1=H_T^\ast(X_k)_{\mathrm{loc}}$.
\begin{theorem}[{see \cite{art:qinwang2007,art:maulikoblomov2009}}]\label{thm:heisenberg}
The linear operators $\mathfrak{p}_m(\alpha)$, where $m\in{\mathbb{Z}}\setminus\{0\}$ and $\alpha\in{H}_T^\ast({X_k})_{\mathrm{loc}}$, satisfy the Heisenberg commutation relations
\begin{equation}
\big[\mathfrak{p}_m(\alpha),\mathfrak{p}_n(\beta) \big]= m\, \delta_{m,-n}\,
\langle\alpha,\beta\rangle_{{\mathbb{H}}_1}\,\mathrm{id}\qquad \mbox{and}
\qquad\big[\mathfrak{p}_m(\alpha),\mathrm{id} \big]= 0\ .
\end{equation}
The vector space $\mathbb{H}$ is the Fock space of the Heisenberg algebra $\mathfrak{h}_{\mathbb{H}_1}$ modelled on $\mathbb{H}_1={H}_T^\ast({X_k})_{\mathrm{loc}}$ with highest weight vector the unit element $\vert 0\rangle$ in ${H}_T^0(\hilb{0}{X_k})_{\mathrm{loc}}$.
\end{theorem}
\subsubsection{Heisenberg algebra of rank $k$}
Let $i\in \{1, \ldots, k\}$. Consider the Heisenberg algebra $\mathfrak{h}_{i}$ over ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ generated by the operators
\begin{equation}
\mathfrak{p}_{-m}^i:=\mathfrak{p}_{-m}([\alpha_i])\qquad\mbox{and}\qquad\mathfrak{p}_{m}^i:=\mathfrak{p}_{m}([\alpha_i])
\end{equation}
for $m\in{\mathbb{Z}}_{>0}$. By Theorem \ref{thm:heisenberg}, the commutation relations are
\begin{equation}\label{eq:commutationalpha}
\big[\mathfrak{p}_m^i,\mathfrak{p}_n^j \big]=m\, \delta_{m,-n}\, \delta_{ij}\,
\big\langle[\alpha_i]\,,\,[\alpha_i]
\big\rangle_{\mathbb{H}}\,\mathrm{id}=m\, \delta_{m,-n}\,
\delta_{ij}\, \beta_i^{-1}\,\mathrm{id}\ .
\end{equation}
Since $\{[\alpha_1], \ldots, [\alpha_k]\}$ is a ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$-linear basis of ${\mathbb{H}}_1$, the Heisenberg algebra $\mathfrak{h}_{{\mathbb{H}}_1}$ is generated by $\mathfrak{p}_m^i$ for $i=1, \ldots, k$ and $m\in{\mathbb{Z}}\setminus\{0\}$.
Let ${\mathbb{H}}_{U_i}$ be the ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$-linear
subspace of $\mathbb{H}$ introduced in Section
\ref{sec:basisfixedpoints}. Then by Theorem
\ref{thm:nakajimaoperators} ${\mathbb{H}}_{U_i}$ is the Fock space for the
Heisenberg algebra $\mathfrak{h}_{i}$ for any $i\in \{1, \ldots, k\}$; therefore the
${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$-vector space ${\mathbb{H}}_{U_i}$ is
generated by the elements $\mathfrak{p}_\lambda^i\vert 0 \rangle$ where
$\mathfrak{p}_\lambda^i:=\prod_{l\geq 1}\, (\mathfrak{p}_{-l}^i)^{m_l}$ for a
partition $\lambda=(1^{m_1}\, 2^{m_2}\, \cdots)$. One can show that
\begin{equation}
\big\langle\mathfrak{p}_\lambda^i\vert 0\rangle\,,\,\mathfrak{p}_\mu^i\vert 0\rangle
\big\rangle_{{\mathbb{H}}_{U_i}} = \delta_{\lambda,\mu} \, z_\lambda \, \beta_i^{-\ell(\lambda)} \ .
\end{equation}
On the algebra $\Lambda_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2)}$ of
symmetric functions over the field ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$
we introduce the Jack inner product \eqref{eq:jackinnerprod} with
parameter $\beta_i$.
We shall denote with $\Lambda_{\beta_i}$ the algebra $\Lambda_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2)}$ endowed with the symmetric bilinear form $\langle -,- \rangle_{\beta_i}$. Thus by the isomorphism \eqref{eq:cohomologyUi} and Theorem \ref{thm:actionC2} there exists an isomorphism of ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$-vector spaces
\begin{equation}
\Phi_i\, \colon \, {\mathbb{H}}_{U_i} \ \xrightarrow{ \ \sim \ } \
\Lambda_{\beta_i}\ , \qquad
\mathfrak{p}_\lambda^i\vert 0 \rangle \ \longmapsto \ p_\lambda\ ,
\end{equation}
which intertwines the symmetric bilinear forms $\langle -,- \rangle_i$
and $\langle -,- \rangle_{\beta_i}$. For $m>0$ the operator
$\mathfrak{p}_{-m}^i$ acts as multiplication by $p_m$ on
$\Lambda_{\beta_i}$ while its adjoint $\mathfrak{p}_{m}^i$ with respect to
the symmetric bilinear form $\langle -,- \rangle_i $ acts as $m\,
\beta_i^{-1}\, \frac{\partial}{\partial p_m}$.
By Theorem \ref{thm:actionC2} we can also determine how $\Phi_i$ acts
on the ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$-linear basis
$\{[\alpha_{\vec{Y}}] \}$ of ${\mathbb{H}}_{U_i}$, where
$\vec{Y}=(Y^1, \ldots, Y^k)$ is a fixed point such that $Y^j=\emptyset$ for
every $j\in \{1, \ldots, k\}$ with $j\neq i$.
\begin{proposition}
Let $\vec{Y}=(Y^1, \ldots, Y^k)$ be a fixed point such that $Y^j=\emptyset$
for every $j\in \{1, \ldots, k\}$ with $j\neq i$. Then
\begin{equation}
\Phi_i([\alpha_{\vec{Y}}])=J_{\lambda_i}(x;\beta_i^{-1}) \ ,
\end{equation}
where $Y_{\lambda_i}:= Y^i$.
\end{proposition}
Define $\Lambda_{\vec{\beta}}=\bigotimes_{i=1}^k\,
\Lambda_{\beta_i}$ endowed with the symmetric bilinear form $\langle
p,q\rangle_{\vec{\beta}} :=\prod_{i=1}^k\, \langle
p_i,q_i\rangle_{\beta_i} $ for $p=p_1\otimes \cdots\otimes p_k$ and
$q=q_1\otimes \cdots\otimes q_k$ in $\Lambda_{\vec{\beta}}\, $.
For a $k$-tuple of Young tableaux $\vec{Y}$, define in
$\mathcal{U}(\mathfrak{h}_{{\mathbb{H}}_1}) $ the operators
$\mathfrak{p}_{\vec{Y}}=\prod_{i=1}^k \, \mathfrak{p}^i_{\lambda_{i}}$. We have thus
proven the following result.
\begin{theorem}\label{thm:cohomologyALE}
There exists a ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$-linear isomorphism
\begin{equation}
\Phi:= \bigotimes_{i=1}^k\, \Phi_i \, \colon \, {\mathbb{H}} \ \longrightarrow
\ \Lambda_{\vec{\beta}}
\end{equation}
preserving bilinear forms such that
\begin{equation}
\Phi \left( \mathfrak{p}_{\vec{Y}} \vert 0\rangle \right) = p_{\lambda_1} \otimes
\cdots \otimes p_{\lambda_k} \qquad \mbox{and} \qquad \Phi \left(
[\alpha_{\vec Y}] \right) = J_{\lambda_1}(x; \beta_1^{-1}) \otimes \cdots \otimes J_{\lambda_k}(x; \beta_k^{-1})\ .
\end{equation}
Via the isomorphism $\Phi$, the operators $\mathfrak{p}_m^i$ act on
$\Lambda_{\vec{\beta}}$ as multiplication by $p_{-m}$ on the $i$-th
factor for $m<0$ and as the derivation $m\, \beta_i^{-1}\, \frac{\partial}{\partial p_m}$ on the $i$-th factor for $m>0$.
\end{theorem}
\subsubsection{Lattice Heisenberg algebra of type $A_{k-1}$}
Let us now define
\begin{equation}
\mathfrak{q}^i_{-m}:=\mathfrak{p}_{-m}([D_i]_T)\qquad\mbox{and}\qquad\mathfrak{q}^i_m:=\mathfrak{p}_m([D_i]_T)
\end{equation}
for $m\in {\mathbb{Z}}_{>0}$ and $i=1, \ldots, k-1$. By Equation \eqref{eq:divisorcartan} the operators $\mathfrak{q}^i_m$ satisfy the commutation relations
\begin{equation}
[\mathfrak{q}^i_m, \mathfrak{q}^j_n] =m\, \delta_{m,-n} \, C_{ij}\,
\mathrm{id}\qquad \mbox{for} \quad i,j=1, \ldots, k-1 \ , \ m,n\in{\mathbb{Z}}\setminus\{0\}\ ,
\end{equation}
where $C=(C_{ij})$ is the Cartan matrix of the Dynkin diagram of type $A_{k-1}$.
Let $\mathfrak{L}\subset H_T^\ast(X_k)_{\mathrm{loc}}$ be the ${\mathbb{Z}}$-lattice generated by the classes $[D_1]_T, \ldots, [D_{k-1}]_T$ with the symmetric bilinear form given by the Cartan matrix $C$. Then the lattice Heisenberg algebra $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{L}}$ associated with $\mathfrak{L}$ over ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$, which has generators $\mathfrak{q}^i_m$ for $m\in {\mathbb{Z}}\setminus\{0\}$ and $i=1, \ldots, k-1$, is isomorphic to the Heisenberg algebra $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2), \mathfrak{Q}}$ of type $A_{k-1}$ over ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ (cf. Example \ref{ex:latticeheisenebrg}).
Let
\begin{equation}\label{eq:E}
E:=\sum_{i=0}^k \,a_i \, [D_i]_T
\end{equation}
where $a_i\in
{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ with $i=0,1, \ldots, k$ satisfy the relations
\begin{equation}\label{eq:relation1}
2a_j-a_{j-1}-a_{j+1}= 0\ , \quad j=1,\ldots, k-1 \qquad \mbox{and} \qquad
a_0 \, \varepsilon_2 + a_k \, \varepsilon_1 \neq 0\ .
\end{equation}
The first condition ensures that $\langle [D_i]_T, E\rangle_{{\mathbb{H}}_1} =0$ for
$i=1, \ldots, k-1$ while the second condition implies that $\{[D_1]_T,
\ldots, [D_{k-1}]_T, E\}$ is a ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$-linear basis of $H_T^\ast(X_k)_{\mathrm{loc}}$. By \eqref{eq:relation1} one has
\begin{equation}
\kappa:= \langle E, E\rangle_{\mathbb{H}_1} = a_0^2\,
\beta_1^{-1}-a_0\, a_1-a_k\, a_{k-1}+a_k^2\, \beta_k \ .
\end{equation}
From now on we set $a_i=i$ in Equation \eqref{eq:E} for $i=0, 1,
\ldots, k$, which is consistent with the conditions in Equation
\eqref{eq:relation1}. This implies that $\kappa=k\, \beta$. In the
following we normalize the equivariant cohomology class $E$ such that $\langle E,
E\rangle_{\mathbb{H}_1}=1$; we denote the normalized class with the
same symbol.
Define $\mathfrak{p}_{-m}:=\mathfrak{p}_{-m}(E)$ and $\mathfrak{p}_m:=\mathfrak{p}_m(E)$ for
$m\in {\mathbb{Z}}_{>0}$. Then the operators $\mathfrak{q}^i_m$ and $\mathfrak{p}_m$ satisfy the commutation relations
\begin{equation*}
\left\{
\begin{array}{ll}
\big[\mathfrak{q}^i_m, \mathfrak{q}^j_n\big]=m\, \delta_{m,-n} \,
C_{ij}\, \mathrm{id} & \mbox{for} \quad i,j=1, \ldots, k-1 \ , \ m,n\in{\mathbb{Z}}\setminus\{0\}\ ,\\[4pt]
\left[\mathfrak{q}^i_m, \mathfrak{p}_n\right]=0 & \mbox{for} \quad i=1,
\ldots, k-1 \ , \ m,n\in{\mathbb{Z}}\setminus\{0\}\ ,\\[4pt]
\big[\mathfrak{p}_m, \mathfrak{p}_n\big]=m\, \delta_{m,-n}\, \mathrm{id} &
\mbox{for} \quad m,n\in{\mathbb{Z}}\setminus\{0\}\ .
\end{array}\right.
\end{equation*}
Let $\mathfrak{L}'\subset H_T^\ast(X_k)_{\mathrm{loc}}$ be the ${\mathbb{Z}}$-lattice generated by
the classes $[D_1]_T, \ldots, [D_{k-1}]_T$ and $ E$. Then the operators
$\mathfrak{q}^i_m$ and $\mathfrak{p}_n$ for $m,n\in{\mathbb{Z}}\setminus\{0\}$ and
$1\leq i \leq k-1$ define the lattice Heisenberg algebra
$\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{L}'}$ associated with
$\mathfrak{L}'$ over ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$. In particular,
$\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{L}'}$ is the sum
(identifying central elements) of, respectively, the Heisenberg
algebra $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{Q}}$ of type
$A_{k-1}$ over ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ and the Heisenberg
algebra $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2)}$ over
${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ generated by $\mathfrak{p}_m$ for
$m\in{\mathbb{Z}}\setminus\{0\}$.
Since $\{[D_1]_T, \ldots, [D_{k-1}]_T\}\cup \{ E\}$ is a ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$-linear basis of $H_T^\ast(X_k)_{\mathrm{loc}}$, we have $\mathfrak{h}_{ {\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{L}'}\cong \mathfrak{h}_{\mathbb{H}_1}$. Hence $\mathbb{H}$ is the Fock space of $\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1, \varepsilon_2),\mathfrak{L}'}$.
In what follows we omit the symbol ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ from the notation for the Heisenberg algebras generated over the field ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$.
\begin{remark}\label{rem:divisors}
By Equation \eqref{eq:divisor}, for $l=1, \ldots, k-1$ one has
\begin{align}\label{eq:fixedpoint-divisor-1}
\mathfrak{q}^l_m &=-\beta_l\, \mathfrak{p}^l_m+\mathfrak{p}^{l+1}_m\ ,\\[4pt] \label{eq:fixedpoint-divisor-2}
\mathfrak{p}_m &= \sqrt{-k\, \varepsilon_1\, \varepsilon_2} \ \sum_{i=1}^k\,
\frac{1}{\varepsilon_2^{(i)}}\, \mathfrak{p}_m^i=\frac{1}{\sqrt{\big\langle
[X_k], [X_k] \big\rangle_{{\mathbb{H}}_1}}}\,\sum_{i=1}^k\,\frac{1}{\varepsilon_2^{(i)}}\, \mathfrak{p}_m^i\ .
\end{align}
\end{remark}
\subsection{Dominant representation of $\widehat{\mathfrak{gl}}_k$ on ${\mathbb{W}}_j$}
In the following we omit the dependence of the Lie algebras on the field ${\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ to simplify the presentation.
\begin{proposition}\label{prop:representation}
Let $j\in\{0,1, \ldots, k-1\}$. There is an action of $\widehat{\mathfrak{gl}}_k$ on ${\mathbb{W}}_j$ under which ${\mathbb{W}}_j$ is the $j$-th dominant representation of $\widehat{\mathfrak{gl}}_k$ at level one, i.e., the highest weight representation of $\widehat{\mathfrak{gl}}_k$ with fundamental weight $\widehat{\omega}_j$ of type $\widehat{A}_{k-1}$.
\end{proposition}
\proof
The vector space $\mathbb{H}$ is an irreducible highest weight
representation of the sum (identifying central elements)
$\mathfrak{h} \oplus
\mathfrak{h}_{\mathfrak{Q}}$ of the Heisenberg
algebra $\mathfrak{h}$ and the lattice
Heisenberg algebra $\mathfrak{h}_{\mathfrak{Q}}$
of type $A_{k-1}$ generated over the field ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)$, respectively, by the operators $\mathfrak{p}_m$ and $\mathfrak{q}^i_m$ for
$m\in{\mathbb{Z}}\setminus\{0\}$ and $i=1, \ldots, k-1$. We apply the Frenkel-Kac construction (Theorem \ref{thm:frenkelkac}) to the representation $\mathfrak{h}_{\mathfrak{Q}} \to \operatorname{End} ({\mathbb{H}} )$ to obtain a level one representation
\begin{equation}
\widehat{\mathfrak{sl}}_k \ \longrightarrow \ \operatorname{End}
\big({\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_j] \big) \ .
\end{equation}
We can extend the representation of $\mathfrak{h}$ from ${\mathbb{H}}$ to
${\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_j]$ by
letting it act as the identity on the group algebra of
$\mathfrak{Q}+\omega_j$. Thus we get a level one representation of
$\widehat{\mathfrak{gl}}_k$ with
\begin{equation}
\widehat{\mathfrak{gl}}_k \ \longrightarrow \ \operatorname{End}
\big({\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_j] \big) \ .
\end{equation}
Thanks to the isomorphism $\Psi_j$ introduced in \eqref{eq:isomWH},
this gives a level one representation of $\widehat{\mathfrak{gl}}_k$ with
\begin{equation}
\widehat{\mathfrak{gl}}_k \ \longrightarrow \ \operatorname{End}({\mathbb{W}}_j)\ .
\end{equation}
Since ${\mathbb{H}}$ is the Fock space of
$\mathfrak{h} \oplus\mathfrak{h}_{\mathfrak{Q}}$, the
module ${\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_j]$ is isomorphic to the $j$-th dominant representation $ {\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}\otimes {\mathcal V}(\,\widehat{\omega}_j\,)$ of $\widehat{\mathfrak{gl}}_k$ (cf.\ Theorem \ref{thm:frenkelkac}). Hence to complete the proof it is enough to note that the class $[\emptyset, \boldsymbol e_j]$ corresponding to the fixed point $({\mathcal R}_j, \phi_{{\mathcal R}_j})$ in $\mathcal{M}(\boldsymbol e_j, 0,j)$ is sent via $\Psi_j$ to $\vert 0\rangle\otimes [\omega_j]$, which is the highest weight vector of ${\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_j]$.
\endproof
\begin{remark}
Proposition \ref{prop:representation} is analogous to a previous result derived for Nakajima quiver varieties (see e.g. \cite[Section 10]{art:nakajima:1994-3} and \cite[Section 5.1]{art:smirnov2013}).
\end{remark}
Let us introduce the Virasoro operators of $\widehat{\mathfrak{gl}}_k$ by (cf.\ Sections \ref{sec:virasoro-heisenberg} and \ref{sec:virasoro-affine})
\begin{align}
L_0&:=L_0^{\mathfrak{h}}+L_0^{\widehat{\mathfrak{sl}}_{k}}=\sum_{m=
1}^\infty\, \mathfrak{p}_{-m}\,\mathfrak{p}_m +\sum_{i=1}^{k-1}\ \sum_{m=
1}^\infty\, \mathfrak{q}_{-m}^{\eta_i}\,\mathfrak{q}_{m}^{\eta_i}+\frac{1}{2}\,
\sum_{i=1}^{k-1} \, \big(\mathfrak{q}_0^{\eta_i}\big)^2\ ,\\[4pt]
L_n&:=L_n^{\mathfrak{h}}+L_n^{\widehat{\mathfrak{sl}}_{k}}=\frac12\,\sum_{m\in{\mathbb{Z}}}
\, \mathfrak{p}_{-m}\,\mathfrak{p}_{m+n} +\frac12\,\sum_{i=1}^{k-1}\
\sum_{m\in{\mathbb{Z}}}\, \mathfrak{q}_{-m}^{\eta_i}\,\mathfrak{q}_{m+n}^{\eta_i}\qquad
\mbox{for}\quad n\neq 0\ ,
\end{align}
where $\{\eta_i\}_{i=1}^{ k-1}$ is an orthonormal basis of the vector
space $\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{R}}$ and we set $\mathfrak{p}_0:=0$. Note that
$\{\eta_i\}_{i=1}^{k-1} \cup \{E\}$ is an orthonormal basis of the
vector space ${\mathbb{C}}(\varepsilon_1,\varepsilon_2)^k\simeq H_T^\ast(X_k)_{\mathrm{loc}}$, so after an orthonormal change of basis and a suitable normalization one can rewrite the operators $L_n$ in the form
\begin{align}
L_0&=\sum_{l=1}^k\, \beta_l \ \sum_{m= 1}^\infty \,
\mathfrak{p}^l_{-m}\,\mathfrak{p}^l_m +\frac{1}{2}\, \sum_{i=1}^{k-1} \, \big(\mathfrak{q}_0^{\eta_i}\big)^2\ ,\\[4pt]
L_n&=\frac12\,\sum_{l=1}^k\, \beta_l \ \sum_{\stackrel{\scriptstyle
m\in{\mathbb{Z}}}{\scriptstyle m\neq 0, -n}}\, \mathfrak{p}^l_{-m}\,\mathfrak{p}^l_{m+n}
+\sum_{i=1}^{k-1}\,
\mathfrak{q}_{0}^{\eta_i}\,\mathfrak{q}_{n}^{\eta_i} \qquad
\mbox{for}\quad n\neq 0\ .
\end{align}
\begin{proposition}\label{prop:L0eigen}
Let $j\in\{0,1, \ldots, k-1\}$, $\vec{u}\in\mathfrak{U}_j$ and $n\in{\mathbb{N}}$. Then
\begin{equation}
{L_0} \big\vert_{{\mathbb{W}}_{\vec{u}, n ,j}}=\big(n+\mbox{$\frac12$}\,
\vec{v}\cdot C\vec{v}\, \big)\,\operatorname{id}_{{\mathbb{W}}_{\vec{u}, n ,j}}\ ,
\end{equation}
where $\vec{v}:=C^{-1}\vec{u}$.
\end{proposition}
\proof
By Equation \eqref{eq:integralofmotions-fixedpoints}, we have
$\sum_{m=1}^\infty \, \mathfrak{p}^l_{-m}\,\mathfrak{p}^l_m\triangleright
[\vec{Y}]=\beta_l^{-1}\, \vert Y^l \vert \, [\vec{Y}]$ for $[\vec{Y}]=[(\emptyset,
\ldots, Y^l,\ldots, \emptyset)]$ in ${\mathbb{H}}_{U_l}$ and $l\in\{1,\ldots,k\}$. Then by using the isomorphism $\Omega$ introduced in \eqref{eq:omega} and the isomorphism $\Psi_j$ introduced in \eqref{eq:isomWH} we get
\begin{equation}
\Big(\, \sum_{l=1}^k\, \beta_l \ \sum_{m= 1}^\infty\,
\mathfrak{p}^l_{-m}\,\mathfrak{p}^l_m\, \Big)\Big\vert_{ {\mathbb{W}}_{\vec{u}, n ,j}}=n \,\operatorname{id}_{{\mathbb{W}}_{\vec{u}, n ,j}}\ .
\end{equation}
On the other hand, since $\{\eta_i\}_{i=1}^{ k-1}$ is an orthonormal basis of $\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{R}}$ we get
\begin{equation}
\Big(\, \sum_{i=1}^{k-1}\, \big(\mathfrak{q}_0^{\eta_i}\big)^2\,
\Big)\Big\vert_{ {\mathbb{W}}_{\vec{u}, n ,j}}=\Big(\, \sum_{i=1}^{k-1}\,
\Big\langle\eta_i\,,\, \sum_{s=1}^{k-1}\, v_s \, \gamma_s
\Big\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}}{\mathbb{R}}}^2 \, \Big) \,\operatorname{id}_{{\mathbb{W}}_{\vec{u}, n ,j}}
=\big(\,\vec{v}\cdot C\vec{v} \, \big) \, \operatorname{id}_{{\mathbb{W}}_{\vec{u}, n ,j}}\ ,
\end{equation}
where $\vec{v}:=C^{-1}\vec{u}$.
\endproof
Since
\begin{equation}
L_0\triangleright [\emptyset, \vec{u}\, ]=\mbox{$\frac12$}\,
{\vec{v}\cdot C\vec{v}} \, [\emptyset, \vec{u}\, ] \qquad\mbox{and}\qquad L_n\triangleright
[\emptyset,\vec{u}\, ] = 0 \quad \mbox{for}\quad n>0\ ,
\end{equation}
we have the following result.
\begin{corollary}\label{cor:hwrep}
${\mathbb{W}}_{\vec{u},j}$ is a highest weight representation
of the Virasoro algebra $\mathfrak{Vir}$ associated with $\widehat{\mathfrak{gl}}_k$, which is generated by the operators $L_n$ and $\mathfrak{c}$ with highest
weight vector $[\emptyset, \vec{u}\, ]$ and conformal dimension
\begin{equation}
\Delta_{\vec{u}}:=\mbox{$\frac12$} \, {\vec{v}\cdot C\vec{v}} \ .
\end{equation}
Moreover, the energy eigenspace decomposition of the
representation ${\mathbb{W}}_{\vec{u},j}$ is given by \eqref{eq:L0eigenspace}.
\end{corollary}
\begin{proposition}\label{prop:Wkweight}
Let $j\in\{0,1, \ldots, k-1\}$. The weight decomposition of ${\mathbb{W}}_j$ as
a $\widehat{\mathfrak{gl}}_k$-module is given by
\begin{equation}
{\mathbb{W}}_j=\bigoplus_{\vec{u}\in \mathfrak{U}_j} \, {\mathbb{W}}_{\vec{u},j}\ .
\end{equation}
\end{proposition}
\proof
For $j=0, 1, \ldots, k-1$ and for any element
$A\otimes [\gamma_{\vec{u}}+\omega_j]\in {\mathbb{H}}\otimes \mathbb{C}(\varepsilon_1, \varepsilon_2)[\mathfrak{Q}+\omega_j]$ with $\vec{u}\in \mathfrak{U}_j$, we have
\begin{align}
h_0\triangleright (A\otimes [\gamma_{\vec{u}}+\omega_j ])
&=\Big(1-\sum_{l=1}^{k-1}\, u_l\Big)\, (A\otimes
[\gamma_{\vec{u}}+\omega_j] ) \ , \\[4pt]
h_i\triangleright (A\otimes [\gamma_{\vec{u}}+\omega_j ])&=u_i \,
(A\otimes [\gamma_{\vec{u}}+\omega_j ] )\qquad
\mbox{for} \quad i=1, \ldots, k-1\ .
\end{align}
Under the $\widehat{\mathfrak{gl}}_k$-action, ${\mathbb{W}}_{\vec{u},j}$ decomposes as
\begin{equation}
{\mathbb{W}}_{\vec{u},j}\simeq {\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}\otimes{\mathcal V}(\widehat{\omega}_j)_{\gamma_{\vec{u}}}\ ,
\end{equation}
where
${\mathcal V}(\widehat{\omega}_j)_{\lambda}:=\{w \in
{\mathcal V}(\widehat{\omega}_j)\,\vert\, h_i\triangleright
w=(\widehat{\omega}_j+\lambda)(h_i)\,w \}$ for any weight $\lambda$.
The assertion now follows by showing that for a weight $\lambda$, the
vector space ${\mathcal V}(\widehat{\omega}_j)_\lambda$ is nonzero if and only if
$\lambda=\gamma_{\vec{u}}$ for some $\vec{u}\in\mathfrak{U}_j$. For this,
let $\vec{\xi}^{\
\vec{v}}:=\prod_{i=1}^{k-1} \, \xi_i^{v_i}$ for $\xi_i\in{\mathbb{C}}^*$ with
$|\xi_i|<1$ and
any vector $\vec{v}=(v_1,\dots,v_{k-1})$, and set $\vec{h}:=(h_1, \ldots, h_{k-1})$. Then
it is enough to note that the trace of the operator $\mathsf{q}^{L_0}\
\vec{\xi}^{\ C^{-1}\vec{h}}$ is given by
\begin{equation}
\operatorname{Tr}_{{\mathbb{W}}_j} \, \mathsf{q}^{L_0}\, \vec{\xi}^{\ C^{-1}\vec{h}} =
\sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{n=0}^\infty\, \mathsf{q}^{n+\frac12\,
\vec{v}\cdot C\vec{v}} \ \vec{\xi}^{\ \vec{v}}
\end{equation}
and the right-hand side is exactly the character of the $j$-th dominant
representation of $\widehat{\mathfrak{gl}}_k$ by \cite[Section 5.3]{art:bruzzopedrinisalaszabo2013}.
\endproof
\subsubsection{Whittaker vectors}
Consider now the completed total equivariant cohomology
\begin{equation}
\widehat{{\mathbb{W}}}_j:=\prod_{\vec{u}\in\mathfrak{U}_j} \ \prod_{n\geq0}\, {\mathbb{W}}_{\vec{u},n,j}\ .
\end{equation}
We can extend the isomorphism \eqref{eq:isomWH} to
\begin{equation}
\widehat{\Psi}_j\ \colon \ \widehat{{\mathbb{W}}}_j \
\xrightarrow{ \ \sim \ } \ \widehat{{\mathbb{H}}} \otimes \Big(\,
\prod_{\gamma_{\vec{u}}\in\mathfrak{Q}}\, {\mathbb{C}}(\varepsilon_1,\varepsilon_2)\,
(\gamma_{\vec{u}}+\omega_j) \, \Big) \ ,
\end{equation}
where $\widehat{{\mathbb{H}}}:=\prod_{n\geq0}\, H^\ast_T(\hilb{n}{X_k})_{\mathrm{loc}}$ is
the completed total equivariant cohomology of the Hilbert schemes of
points on $X_k$. In the following we drop the explicit symbols
$\widehat{\Psi}_j$ from the notation in order to simplify the
presentation, and we denote
\begin{equation}
|\omega_j\rangle:= \sum_{\vec{u}\in\mathfrak{U}_j}
\, [\emptyset, \vec{u}\, ] \ .
\end{equation}
\begin{proposition}\label{prop:whittakerALE}
Fix $\vec{\eta}\in{\mathbb{C}}(\varepsilon_1,\varepsilon_2)^k$. In the completed total equivariant cohomology $\widehat{{\mathbb{W}}}_j$ the vector
\begin{equation}
G_j(\vec{\eta}\,):=\exp\Big(\, \sum_{i=1}^k \, \eta_i \
\mathfrak{p}_{-1}^i\, \Big) |\omega_j\rangle
\end{equation}
is a Whittaker vector of type $\chi_{\vec\eta}\, $, where
$\chi_{\vec\eta}\colon
\mathcal{U}(\mathfrak{h}^+ \oplus
\mathfrak{h}_{\mathfrak{Q}}^+ )\to{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ is defined by
\begin{align}
\chi_{\vec\eta}(\mathfrak{q}_1^i)&= \eta_{i+1} \, \beta_{i+1}^{-1} -
\eta_i \qquad \mbox{and} \qquad \chi_{\vec\eta}(\mathfrak{q}_m^i) = 0
\quad \mbox{for} \quad m>1 \ , \ i=1,\ldots, k-1\ ,\\[4pt]
\chi_{\vec\eta}(\mathfrak{p}_1)&= \frac{1}{\sqrt{\big\langle [X_k], [X_k] \big\rangle_{{\mathbb{H}}_1}}}\,\sum_{i=1}^k\,
\frac{\eta_i}{\varepsilon_1^{(i)}} \qquad \mbox{and} \qquad
\chi_{\vec\eta}(\mathfrak{p}_m)= 0 \quad \mbox{for} \quad m>1 \ .
\end{align}
\end{proposition}
\proof
Let $\widehat{{\mathbb{H}}}_{U_i}:=\prod_{n\geq0}\, H_T^\ast(\hilb{n}{U_i})_{\mathrm{loc}}$
be the completed total equivariant cohomology of the Hilbert scheme $\hilb{n}{U_i}$ for
$i=1, \ldots, k$, and define $G(\eta_i):=\exp\left(\, \eta_i \
\mathfrak{p}_{-1}^i\, \right) \vert 0\rangle \in \widehat{{\mathbb{H}}}_{U_i}$. By using
Theorem \ref{thm:cohomologyALE} and the completed versions of the
isomorphisms \eqref{eq:omega} and \eqref{eq:isomWH}, we can rewrite
the vector $G_j(\vec{\eta}\, )$ as
\begin{equation}
G_j(\vec{\eta}\, ) = G(\eta_1)\otimes\cdots\otimes G(\eta_k) \ \otimes
\ \sum_{\vec{u}\in\mathfrak{U}_j}\, (\gamma_{\vec{u}}+\omega_j) \ .
\end{equation}
By Proposition \ref{prop:whittakerC2}, $G(\eta_i)$ for $i=1,\ldots, k$ is a Whittaker vector for the Heisenberg algebra $\mathfrak{h}_i$ of type $\chi_i$, where
\begin{equation}\label{eq:characterheisenberg}
\chi_i(\mathfrak{p}_1^i):= \eta_i \, \beta_i^{-1}\qquad\mbox{and}\qquad \chi_i(\mathfrak{p}_m^i):= 0\quad \mbox{for}\quad m>1\ .
\end{equation}
Again by Theorem \ref{thm:cohomologyALE}, each $\mathfrak{h}_i$ acts
trivially on ${\mathbb{H}}_{U_l}$ for $l\neq i$ and it is easy to see that
$G_j(\vec{\eta}\, )$ is a Whittaker vector for the Heisenberg algebra
$\mathfrak{h}_{{\mathbb{H}}_1}$ of type
$\tilde \chi_{\vec\eta}$,
where $\tilde \chi_{\vec\eta} \colon
\mathcal{U}\big(\mathfrak{h}_{{\mathbb{H}}_1}^+\big)\to{\mathbb{C}}(\varepsilon_1,\varepsilon_2)$
is defined by $\tilde \chi_{\vec\eta}(\mathfrak{p}_m^i):=\chi_i(\mathfrak{p}_m^i)$ for
$i=1, \ldots, k$ and $m\in{\mathbb{Z}}\setminus \{0\}$. Then by Remark
\ref{rem:divisors}, $G_j(\vec{\eta}\, )$ is a Whittaker vector for
$\widehat{\mathfrak{gl}}_k$ of type $\chi_{\vec\eta}$ with $\chi_{\vec\eta}\colon
\mathcal{U}(\mathfrak{h}^+ \oplus
\mathfrak{h}_{\mathfrak{Q}}^+)\to {\mathbb{C}}(\varepsilon_1,\varepsilon_2)$ defined for every $m>0$ by
\begin{align}
\chi_{\vec\eta}(\mathfrak{q}_m^i) & :=\tilde \chi_{\vec\eta}(\mathfrak{p}_m^{i+1})
- \beta_i \, \tilde \chi_{\vec\eta}(\mathfrak{p}_m^i) = \delta_{m,1} \,
\big(\eta_{i+1} \, \beta_{i+1}^{-1} - \eta_i \big)\ ,\\[4pt]
\chi_{\vec\eta}(\mathfrak{p}_m) &: = \sqrt{-k\, \varepsilon_1\, \varepsilon_2}\ \sum_{i=1}^k \,
\frac{1}{\varepsilon_2^{(i)}} \, \tilde \chi_{\vec\eta}(\mathfrak{p}_m^i) = \delta_{m,1} \,
\sqrt{-k\, \varepsilon_1\, \varepsilon_2} \ \sum_{i=1}^k\,
\frac{\eta_i}{\varepsilon_1^{(i)}}\ .
\end{align}
\endproof
\bigskip \section{Chiral vertex operators for $\widehat{\mathfrak{gl}}_k$\label{sec:chiralvertex}}
\subsection{Ext-bundles and bifundamental hypermultiplets\label{se:bifundXk}}
In this section we construct and study the natural generalizations of the
Ext vertex operators from Section
\ref{sec:carlsson-okounkov} for the moduli spaces
$\mathcal{M}(\vec{u},n,j)$.
For $n\in{\mathbb{N}}$, $j\in\{0,1,\ldots,k-1\}$ and $\vec u\in\mathfrak{U}_j$, let $\boldsymbol{{\mathcal E}}_{\vec u,n,j}$ denote the universal sheaf on
$\mathcal{M}(\vec{u},n,j)\times {\mathscr{X}}_k$. Define
\begin{equation}
\boldsymbol{{\mathcal E}}_i:=p_{i 3}^\ast\big(\boldsymbol{{\mathcal E}}_{\vec
u_i,n_i,j_i} \big) \ \in \
K\big(\mathcal{M}(\vec{u}_1,n_1;j_1)
\times\mathcal{M}(\vec{u}_2,n_2;j_2)\times {\mathscr{X}}_k \big) \qquad \mbox{for}
\quad i=1, 2 \ ,
\end{equation}
where $p_{ij}$ is the projection of $\mathcal{M}(\vec{u}_1,n_1;j_1)
\times\mathcal{M}(\vec{u}_2,n_2;j_2)\times {\mathscr{X}}_k$ onto the product of the $i$-th and $
j$-th factors. Denote by $p_3$ the
projection of the same product onto ${\mathscr{X}}_k$.
Let $T_\mu={\mathbb{C}}^\ast$ and $H^\ast_{T_\mu}(\mathrm{pt};{\mathbb{C}})={\mathbb{C}}[\mu]$. Denote
by ${\mathcal O}_{{\mathscr{X}}_k}(\mu)$ the trivial line bundle on ${\mathscr{X}}_k$ on which $T_\mu$ acts by scaling the fibers.
\begin{definition}[{\cite[Definition 4.17]{art:bruzzopedrinisalaszabo2013}}]
The \emph{Carlsson-Okounkov bundle} is the element
\begin{equation}
\boldsymbol{E}_\mu^{\vec{u}_1,n_1,j_1;\vec{u}_2,n_2,j_2}:={p_{12}}_\ast\big(-\boldsymbol{{\mathcal E}
}_1^\vee\cdot \boldsymbol{{\mathcal E}}_2\cdot p_3^\ast({\mathcal O}_{{\mathscr{X}}_k}(\mu)\otimes{\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty))
\big)
\end{equation}
in the K-theory $K\big(\mathcal{M}(\vec{u}_1,n_1;j_1)\times\mathcal{M}(\vec{u}_2,n_2;j_2)\big)$.
\end{definition}
By \cite[Section 4.5]{art:bruzzopedrinisalaszabo2013}, the fibre of
$\boldsymbol{E}_\mu^{\vec{u}_1,n_1,j_1;\vec{u}_2,n_2,j_2}$ over
$\big([({\mathcal E},\phi_{\mathcal E})]\,,\, [({\mathcal E}',\phi_{{\mathcal E}'})]\big)$ in
$\mathcal{M}(\vec{u}_1,n_1;j_1)\times\mathcal{M}(\vec{u}_2,n_2;j_2)$ is given by
\begin{equation}
\boldsymbol{E}_\mu ^{\vec{u}_1
,n_1,j_1;\vec{u}_2,n_2,j_2}\big|_{\left([({\mathcal E},\phi_{\mathcal E})]\,,\,[({\mathcal E}',\phi_{{\mathcal E}'})]\right)}=\operatorname{Ext}^1\big({\mathcal E},{\mathcal E}'\otimes
{\mathcal O}_{{\mathscr{X}}_k}(\mu)\otimes{\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty) \big)\ .
\end{equation}
One can compute the dimension of this vector space by a straightforward
generalization of the dimension computations of \cite[Appendix
A]{art:bruzzopedrinisalaszabo2013} to get the rank
\begin{equation}
\operatorname{rk}\big(\boldsymbol{E}_\mu^{\vec{u}_1,n_1,j_1;\vec{u}_2,n_2,j_2} \big) =
n_1+n_2+\mbox{$\frac12$}\, \vec v_{21}\cdot C\vec
v_{21}-\mbox{$\frac1{2k}$}\, j_{21}\, (k- j_{21} ) \ ,
\end{equation}
where $\vec v_{21}:=C^{-1}(\vec u_2-\vec u_1)$ and $j_{21}\in\{0, 1,\ldots, k-1\}$ is the equivalence class modulo $k$ of $j_2-j_1$.
Let ${\mathbb{W}}:=\bigoplus_{j=0}^{k-1}\, {\mathbb{W}}_j$ endowed with the nondegenerate ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)$-valued bilinear form $\langle -,- \rangle_{{\mathbb{W}}}$
induced by the symmetric bilinear forms $\langle
-,-\rangle_{{\mathbb{W}}_j}$. Define the operator $\operatorname{V}_\mu(\vec x,z)\in\mathrm{En
d}({\mathbb{W}}\, )[[
z^{\pm\, 1},x_1^{\pm\,1},$ $\dots,x_{k-1}^{\pm\,1}]]$ by its matrix elements
\begin{multline}\label{eq:vertexoperator}
(-1)^{n_2} \, \big\langle \operatorname{V}_\mu(\vec x,z)A_1\,,\, A_2
\big\rangle_{{\mathbb{W}}}
:=z^{n_2-n_1 +\Delta_{\vec u_2}-\Delta_{\vec u_1} } \, \vec x\,^{\vec
v_{21}}\\ \times \
\int_{\mathcal{M}(\vec{u}_1,n_1;j_1)\times\mathcal{M}(\vec{u}_2,n_2;j_2)} \,
\operatorname{eu}_T\big(\boldsymbol{E}_\mu^{\vec{u}_1,n_1,j_1;\vec{u}_2,n_2,j_2}
\big) \cup p_1^\ast(A_1) \cup p_2^\ast(A_2) \ ,
\end{multline}
where $A_i\in H_T^\ast(\mathcal{M}(\vec{u}_i,n_i;j_i))_{\mathrm{loc}}$ and $p_i$ is the projection from $\mathcal{M}(\vec{u}_1,n_1;j_1)\times\mathcal{M}(\vec{u}_2,n_2;j_2)$ to the $i$-th factor for
$i=1,2$. The extra isospin parameters $\vec x:=(x_1,\dots,x_{k-1})$
weigh the $\widehat{\mathfrak{sl}}_k$-action, and we abbreviated $\vec x\,^{\vec v}:= \prod_{i=1}^{k-1}\, x_i^{v_i}$ for a
vector $\vec v=(v_1,\dots,v_{k-1})$. By the computations of \cite[Section 4.7]{art:bruzzopedrinisalaszabo2013}, the matrix elements \eqref{eq:vertexoperator} in the fixed point basis are given by
\begin{multline}\label{eq:vertexoperator-fixedbasis}
\big\langle \operatorname{V}_\mu(\vec x,z)[\vec{Y}_1,\vec{u}_1\,] \,,\, [\vec{Y}_2,\vec{u}_2\, ]
\big\rangle_{{\mathbb{W}}}
\\ \shoveleft{=(-1)^{\vert \vec{Y}_2\vert}\,
z^{\vert \vec{Y}_2\vert-\vert \vec{Y}_1\vert+\Delta_{\vec u_2}-\Delta_{\vec u_1} } \, \vec x\,^{\vec v_{21}} \
\operatorname{eu}_T\big(\boldsymbol{E}_\mu^{\vec{u}_1,n_1,j_1;\vec{u}_2,n_2,j_2} \big\vert_{\left([({\mathcal E}_1,\phi_{{\mathcal E}_1})]\,,\,[({\mathcal E}_2,\phi_{{\mathcal E}_2})]\right)}
\big) } \\[4pt]
\shoveleft{ = (-1)^{\vert \vec{Y}_2\vert}\,
z^{\vert \vec{Y}_2\vert-\vert \vec{Y}_1\vert+\Delta_{\vec u_2}-\Delta_{\vec u_1} } \, \vec x\,^{\vec v_{21}} \ \prod_{i=1}^k \, m_{Y_1^{i},
{Y_2^{i}}} \big(\varepsilon_1^{(i)},\varepsilon_2^{(i)},
\mu-(\vec{v}_{21})_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\, \varepsilon_2^{(i)} \big) }
\\ \times \ \prod_{n=
1}^{k-1}\,
\ell^{(n)}_{\vec{v}_{21}}
\big(\varepsilon_1^{(n)},\varepsilon_2^{(n)},\mu \big)\ ,
\end{multline}
where $[({\mathcal E}_1,\phi_{{\mathcal E}_1})]$ and $[({\mathcal E}_2,\phi_{{\mathcal E}_2})]$
are the $T$-fixed points corresponding respectively to the
combinatorial data $(\vec{Y}_1,\vec{u}_1)$ and
$(\vec{Y}_2,\vec{u}_2)$, and we use the convention
$(\vec{v}_{21})_{0}=(\vec{v}_{21})_{k}=0$. Here $m_{Y_1, Y_2}$ is
defined in \eqref{eq:m}, while $\ell^{(n)}_{\vec{v}}$ is the \emph{edge
contribution} defined in Appendix \ref{app:edgecontributions}. This factorized expression for
the matrix elements represents the contribution of the $U(1)\times U(1)$ bifundamental
hypermultiplet for ${\mathcal N}=2$ quiver gauge theories on the ALE space $X_k$.
\subsection{Vertex operators and primary fields}
In this subsection we factorize the operators $\operatorname{V}_\mu(\vec x,z)$ defined in
\eqref{eq:vertexoperator} under the decomposition $\widehat{\mathfrak{gl}}_k=\mathfrak{h}\oplus\widehat{\mathfrak{sl}}_k$ as tensor products of generalized bosonic
exponentials associated to the Heisenberg algebra
$\mathfrak{h}$ from Definition \ref{def:bosonops} with primary
fields of the Virasoro algebra associated to the affine Lie algebra
$\widehat{\mathfrak{sl}}_k$ from Section \ref{sec:virasoro-affine}.
For $l=1,2$ fix $j_l\in\{0,1, \ldots, k-1\}$ and respectively $\vec{u}_l\in\mathfrak{U}_{j_l}$. Set
\begin{equation}
\gamma_{21}:=\sum_{i=1}^{k-1}\,(\vec{v}_{21})_{i}\,\gamma_i=\psi_{j_2}(\vec{u}_2
)-\psi_{j_1}(\vec{u}_1) \ \in \ \mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}\ .
\end{equation}
Note that $\gamma_{21}=\gamma_{\vec{u}_2}-\gamma_{\vec{u}_1}+\omega_{j_2}-\omega
_{j_1}$. We define the maps
\begin{align}
\exp\big(\gamma_{21}\,\big)\,\colon\, {\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)
[\mathfrak{Q}+\omega_{j_1}]\ &\longrightarrow \ {\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+
\omega_{j_2}]\ ,\\
v\otimes [\beta+\omega_{j_1}] \ &\longmapsto\ v\otimes [\beta+\gamma_{\vec{u}_2}-\gamma_{\vec{u}_1}+\omega_{j_2}]\ ,
\end{align}
and $\exp\big(\log z\ \mathfrak{c}+\gamma_{21}\big)\,\colon\, {\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)
[\mathfrak{Q}+\omega_{j_1}]\ \to \ {\mathbb{H}}\otimes{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+
\omega_{j_2}]$, given by
\begin{equation}
\exp\big(\log z\ \mathfrak{c}+\gamma_{21}\,\big)\triangleright(v \otimes
[\beta+\omega_{j_1}]):=z^{\frac12\, \langle
\gamma_{21},\gamma_{21}\rangle_{\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}} + \langle
\gamma_{21}, \beta+\omega_{j_1}\rangle_{\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}} }\, (v \otimes [\beta+\gamma_{\vec{u}_2}
-\gamma_{\vec{u}_1}+\omega_{j_2}])\ .
\end{equation}
Note that the operator $\exp\big(\log z\
\mathfrak{c}-\gamma_{21}\big)\,\exp\big(\gamma_{21}\big)\in\operatorname{End}({\mathbb{H}}\otimes
{\mathbb{C}}(\varepsilon_1,\varepsilon_2)[\mathfrak{Q}+\omega_{j_1}]) [[z,z^{-1}]]$ is
given by
\begin{multline}\label{eq:expopWj}
\exp\big(\log z\
\mathfrak{c}-\gamma_{21}\,\big)\,\exp\big(\gamma_{21}\,\big)\triangleright
(v\otimes [\beta +\omega_{j_1}]) \\ =z^{-\frac12\,\langle\gamma_{21},\gamma_{21}
\rangle_{\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}} -\langle\gamma_{21},
\beta+\omega_{j_1}\rangle_{\mathfrak{Q}\otimes_{{\mathbb{Z}}}{\mathbb{Q}}} }\, (v\otimes [\beta +\omega
_{j_1}]) \ .
\end{multline}
In the following we shall suppress the explicit isomorphism $\Psi$ from Theorem
\ref{cor:cohomologymodulihilbert} in our formulas in order to
simplify the presentation.
We will now rewrite the operator $\operatorname{V}_\mu(\vec x, z)$ in terms of chiral vertex operators in
$\operatorname{Hom}({\mathbb{W}}_{\vec{u}_1;j_1}, {\mathbb{W}}_{\vec{u}_2;j_2})$ $[[z^{\pm\,1},x_1^{\pm\,1},\dots,x_{k-1}^{\pm\,1}]]$ between two highest
weight representations of the Virasoro algebra associated with
$\widehat{\mathfrak{sl}}_k$. For this, let us define the vertex operator ${\bar \operatorname{V}}_\mu(\vec{v}_{21}, \vec x, z)$ of
$\operatorname{Hom}({\mathbb{W}}_{j_1}, {\mathbb{W}}_{j_2})[[z^{\pm\, 1},x_1^{\pm\,1},$ $\dots,x_{k-1}^{\pm\,1}]]$ by
\begin{equation}\label{eq:barVmudef}
{\bar \operatorname{V}}_\mu(\vec{v}_{21}, \vec x, z):=\vec x\,^{\vec v_{21}} \, \prod_{l=1}^{k-1}\,
\ell^{(l)}_{\vec{v}_{21}}
\big(\varepsilon_1^{(l)},\varepsilon_2^{(l)},\mu \big)\
\operatorname{V}_{1,-1}^{\gamma_{21}}(z) \ \exp\big(\log z\ \mathfrak{c}+\gamma_{21}\,\big)\ ,
\end{equation}
where $\operatorname{V}_{1,-1}^{\gamma_{21}}(z)$ is the normal-ordered bosonic exponential associated with the Heisenberg algebra $\mathfrak{h}_\mathfrak{Q}$.
\begin{theorem}\label{thm:virasoroprimary}
Under the decomposition $\widehat{\mathfrak{gl}}_k=\mathfrak{h}\oplus\widehat{\mathfrak{sl}}_k$, the operator $\operatorname{V}_\mu(\vec x, z)$ is given in terms of products of vertex
operators as
\begin{multline}
\operatorname{V}_\mu(\vec x, z)=\operatorname{V}_{-\frac{\mu}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}
}(z) \\ \otimes \ \sum_{j_1,j_2=0}^{k-1} \
\sum_{\vec{u}_1\in\mathfrak{U}_{j_1},\vec{u}_2\in\mathfrak{U}_{j_2}}\
\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)\, z^{\Delta_{\vec u_2}-\Delta_{\vec u_1}}\ \exp\big(\log z\ \mathfrak{c}-\gamma_{21}\,\big)\,\exp\big(\gamma_{21}\,\big)\big\vert_{ {\mathbb{W}}_{\vec{u}_1, j_1}}
\ ,
\end{multline}
where $\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)$ is a primary field of the Virasoro algebra
generated by $L_n^{\widehat{\mathfrak{sl}}_k}$ and $\mathfrak{c}$ with conformal
dimension $\Delta_{\vec u_2-\vec u_1}=\frac12\, \vec{v}_{21}\cdot C\vec{v}_{21}$, i.e., for any $n\in {\mathbb{Z}}$ we have
\begin{equation}\label{eq:primary}
\big[L^{\widehat{\mathfrak{sl}}_k}_n,\bar \operatorname{V}_\mu(\vec{v}_{21},\vec x, z)\big]=z^n \,
\big( z\, \partial_z + \mbox{$\frac12$}\, \vec{v}_{21}\cdot
C\vec{v}_{21}\, n \big)\bar\operatorname{V}_\mu(\vec{v}_{21},\vec x, z)\ .
\end{equation}
\end{theorem}
\proof
By using the isomorphism $\Psi$ and Equation \eqref{eq:carlokfixedbasis} we get
\begin{multline}\label{eq:factorization}
\operatorname{V}_\mu(\vec x, z)=\Psi^{-1} \circ \Big(\, \sum_{j_1,j_2=0}^{k-1} \
\sum_{\vec{u}_1\in\mathfrak{U}_{j_1},\vec{u}_2\in\mathfrak{U}_{j_2}} \,z^{\Delta_{\vec u_2}-\Delta_{\vec u_1}}\, \vec x\,^{\vec v_{21}} \ \prod_{n
=1}^{k-1}\,
\ell^{(n)}_{\vec{v}_{21}}
\big(\varepsilon_1^{(n)},\varepsilon_2^{(n)},\mu \big)\\
\times \ : \, \prod_{i=1}^k \ \operatorname{V}\big({\mathcal O}_{U_i}\big(\mu-(\vec{v}_{21})_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\,
\varepsilon_2^{(i)}\big)\,,\,z\big) \, : \
\otimes \ \big(\psi_{j_2}(\vec{u}_2)\otimes
\psi_{j_1}(\vec{u}_1)^\ast \big) \, \Big) \circ \Psi
\end{multline}
where ${\mathcal O}_{U_i}(\mu)$ is the trivial line bundle on $U_i\simeq{\mathbb{C}}^2$
with an action of $T_\mu$ which rescales the fibers, and $\psi_{j_1}(\vec{u}_1)^\ast$ denotes the dual vector to
$\psi_{j_1}(\vec u_1)$ in the dual vector space ${\mathbb{C}}(\varepsilon_1,
\varepsilon_2)[\mathfrak{Q}+\omega_{j_1}]^\ast$. By Theorem \ref{thm:CO} we
get an expression determined by the operators $\mathfrak{p}_m^i$ for
$m\in{\mathbb{Z}}\setminus\{0\}$ and $i=1,\dots,k$ as
\begin{multline}
\operatorname{V}\big({\mathcal O}_{U_i}\big(\mu-(\vec{v}_{21})_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\,
\varepsilon_2^{(i)}\big)\,,\,z\big) \\ = \exp
\Big(-\frac{\mu-(\vec{v}_{21})_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\, \varepsilon_2^{(i)}}{\varepsilon_2^{(
i)}}\,
\sum_{m=1}^\infty\, \frac{z^m}{m}\, \mathfrak{p}_{-m}^i\Big)\\
\times \ \exp\Big(\,
\frac{\mu+ \varepsilon_1+\varepsilon_2-(\vec{v}_{21})_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\, \varepsilon_2^{(i)}}{\varepsilon_2^{(
i)}}\,
\sum_{m=1}^\infty\, \frac{z^{-m}}{m}\, \mathfrak{p}_{m}^i\, \Big) \ .
\end{multline}
By using Equations \eqref{eq:fixedpoint-divisor-1} and
\eqref{eq:fixedpoint-divisor-2}, we can rewrite $: \, \prod_{i=1}^k\ \operatorname{V}\big({\mathcal O}_{U_i}\big(\mu-(\vec{v}_{21})
_{i}\,
\varepsilon_1^{(i)}-(\vec{v}_{21})_{i-1}\, \varepsilon_2^{(i)}\big)
\,,\, z\big) \, :$ in terms of Heisenberg operators $\mathfrak{q}_m^l$ and $\mathfrak{p}_m$, for $l=1,\ldots, k-1$
and $m\in{\mathbb{Z}}\setminus\{0\}$, and the first assertion now follows. The proof of Equation \eqref{eq:primary} is somewhat lengthy and can be found in Appendix \ref{app:virasoroprimary}.
\endproof
\begin{remark}
In the following we will denote by
$\operatorname{V}_\mu^{j_1,j_2}(\vec x, z)$ the restriction of the vertex operator
$\operatorname{V}_\mu(\vec x, z)$ to $\operatorname{Hom}({\mathbb{W}}_{j_1}, {\mathbb{W}}_{j_2})[[z^{\pm\,1},x_1^{\pm\,1},\dots,x_{k-1}^{\pm\,1}]]$.
\end{remark}
\subsection{Integrals of motion\label{sec:integrals}}
Let $\boldsymbol{V}^{\vec{u},n,j}$ be the pushforward of
$\boldsymbol{E}_0^{\vec{u},n,j;\vec{u},n,j}$ with respect to the projection of
the product $\mathcal{M}(\vec{u},n,j)\times\mathcal{M}(\vec{u},n,j)$ to the second
factor. It is a $T$-equivariant vector bundle on $\mathcal{M}(\vec{u},n,j)$
of rank $n+\frac{1}{2}\, (\vec{v}\cdot C\vec{v}-\frac{1}{k}\, j\, (k-j))$,
which we shall call the natural bundle over $\mathcal{M}(\vec{u},n,j)$.
The $T$-equivariant Chern character of $\boldsymbol{V}^{\vec{u},n,j}$ at a fixed point
$[({\mathcal E},\phi_{\mathcal E})]$ with combinatorial datum $(\vec Y,\vec u\,)$ is given by (cf.\ \cite[Section 4.7]{art:bruzzopedrinisalaszabo2013})
\begin{equation}
\operatorname{ch}_T\big(\boldsymbol{V}^{\vec{u},n,j}\big\vert_{[({\mathcal E},\phi_{\mathcal E})]}\big) = \sum_{i=1}^k \ \sum_{s\in
Y^i}\, {\,\rm e}\,^{-(v_i+L'_{Y^i}(s))\, \varepsilon_1^{(i)}- (v_{i-1}+
A_{Y^i}'(s))\, \varepsilon_2^{(i)}}+\sum_{n=1}^{k-1}\,
L_{\vec{v}}^{(n)} \big(\varepsilon_1^{(n)}, \varepsilon_2^{(n)}\big) \ ,
\end{equation}
where the edge contributions $L_{\vec{v}}^{(n)}
\big(\varepsilon_1^{(n)}, \varepsilon_2^{(n)}\big) $ are defined in
Appendix \ref{app:edgecontributions}.
Let us denote by $\boldsymbol{V}^j$ the natural bundle over
$\coprod_{\vec{u}\in\mathfrak{U}_j}\ \coprod_{n\geq 0}\,
\mathcal{M}(\vec{u},n,j)$, and consider the operators of multiplication by $\boldsymbol{I}_1:=\operatorname{rk}\big(\boldsymbol{V}^j \big)$ and $\boldsymbol{I}_p:=
(\operatorname{c}_{p-1} )_T\big(\boldsymbol{V}^j \big)$ for $p\geq 2$ on
$\prod_{\vec{u}\in\mathfrak{U}_j}\ \prod_{n\geq
0}\,{\mathbb{W}}_{\vec{u}, n ,j}$. For example, one has
\begin{align}
\boldsymbol{I}_1\triangleright [\vec{Y}, \vec{u}\, ]= & \ \big(\, \vert
\vec{Y}\vert+\mbox{$\frac{1}{2}$} \,
\vec{v}\cdot C\vec{v}\, \big)\, [\vec{Y}, \vec{u}\, ] -\mbox{$\frac{1}{2k}\, j\, (k-j)$}\, [\vec{Y},
\vec{u}\, ] \ , \\[4pt]
\boldsymbol{I}_2\triangleright [\vec{Y}, \vec{u}\, ]=& \
-\sum_{i=1}^k \ \sum_{s\in
Y^i}\, \Big( \big(v_i+L'_{Y^i}(s) \big)\, \varepsilon_1^{(i)}+
\big(v_{i-1}+ A_{Y^i}'(s) \big)\, \varepsilon_2^{(i)}\Big)\, [\vec{Y},
\vec{u}\, ]\\
& \ +\, \sum_{n=1}^{k-1}\,
\ell_{\vec{v}}^{(n)}\big(\varepsilon_1^{(n)},
\varepsilon_2^{(n)}\big)_{[1]} \, [\vec{Y}, \vec{u}\, ] \ ,
\end{align}
where the quantities $\ell_{\vec{v}}^{(n)}(\varepsilon_1^{(n)},
\varepsilon_2^{(n)})_{[1]} $ are defined in Appendix \ref{app:edgecontributions}. Note that, by Proposition \ref{prop:L0eigen}, the operator $\boldsymbol{I}_1$ coincides (up to a constant shift) with the Virasoro operator $L_0$ for $\widehat{\mathfrak{gl}}_k$.
By using the description in Section \ref{sec:integralC2}, these operators can be written partly in terms of the Heisenberg
operators $\mathfrak{p}_m^i$ of $\mathfrak{h}_{{\mathbb{H}}_1}$ and the $\widehat{\mathfrak{sl}}_k$ generators $\mathfrak{q}_0^i=h_i$; one has
\begin{align}
\boldsymbol{I}_1 =& \
\sum_{i=1}^k\,\beta_i \
\sum_{m=1}^\infty \, \mathfrak{p}_{-m}^i \, \mathfrak{p}_{m}^i +\frac{1}{2}\, \sum_{i=1}^{k-1} \, \big(\mathfrak{q}_0^{\eta_i}\big)^2
- \frac{1}{2k}\, j\, (k-j)\, \operatorname{id} \ , \\[4pt]
\boldsymbol{I}_2 =& \
\sum_{i=1}^k\,\varepsilon_1^{(i)}\, \Big(\, \frac{\beta_i}{2}\, \sum_{m,n=1}^\infty \, \big(
\mathfrak{p}_{-m}^i\, \mathfrak{p}_{-n}^i\, \mathfrak{p}_{m+n}^i+\mathfrak{p}_{-m-n}^i\, \mathfrak{p}_{n}^i\,
\mathfrak{p}_{m}^i \big) \\ & \qquad \qquad -\, \frac{\beta_i-1}{2}\, \sum_{m=1}^\infty\, (m-1)\,
\mathfrak{p}_{-m}^i\, \mathfrak{p}_{m}^i\, \Big)\\
& \ +\, \sum_{i=1}^k\, \varepsilon_1^{(i)} \ \sum_{j=1}^{k-1}\, \Big( \big(C^{-1}\big)^{ij}\, \beta_i-\big(C^{-1}\big)^{i-1,j}\Big) \ \sum_{m=1}^\infty\, \mathfrak{p}_{-m}^i \, \mathfrak{p}_{m}^i\, \mathfrak{q}_0^j + \boldsymbol{L}_1 \ ,
\end{align}
where $\boldsymbol{L}_1$ is the operator defined by $\boldsymbol{L}_1\triangleright [\vec{Y}, \vec{u}\, ]:= \sum_{n=1}^{k-1}\,
\ell_{\vec{v}}^{(n)}(\varepsilon_1^{(n)},\varepsilon_2^{(n)})_{[1]} \, [\vec{Y}, \vec{u}\, ]$ and we set $(C^{-1})^{0j}=0=(C^{-1})^{kj}$.
Following~\cite{art:belavinbershteinfeiginlitvinovtarnopolsky2011}, here we shall identify a quantum integrable system for each Heisenberg subalgebra $\mathfrak{h}$ of $\widehat{\mathfrak{gl}}_k$. Then each integral of motion associated to the Heisenberg algebra $\mathfrak{h}_{{\mathbb{H}}_1}$ is a sum of integrals of motion of $k$ non-interacting Calogero-Sutherland models from Section \ref{sec:integralC2} with prescribed couplings; in particular the Hamiltonian is given by $k$ copies of one-component bosonized Calogero-Sutherland Hamiltonians as $\sum_{i=1}^k\, \varepsilon_1^{(i)}\, \Box^{\beta_i^{-1}}$. This infinite system of commuting operators is diagonalized in the fixed-point basis $[\vec Y,\vec u\,]$. This simultaneous eigenbasis also factorizes the primary operators from Theorem \ref{thm:virasoroprimary}. The remaining $\vec v$-dependent parts of the eigenvalues are instead interpreted as particular matrix elements of our geometrically defined vertex operators $\operatorname{V}_\mu(\vec x,z)$ in highest weight vectors of $\widehat{\mathfrak{gl}}_k$, as we discuss in Appendix \ref{app:edgecontributions}.
\begin{remark}
By Remark \ref{rem:quivervariety}, the moduli spaces $\mathcal{M}(\vec{u},n,j)$ are Nakajima quiver varieties of type $\widehat{A}_{k-1}$. The descriptions of the integrable systems corresponding to quiver varieties are detailed in \cite{art:maulikokounkov2012,art:smirnov2013}.
\end{remark}
\bigskip \section{${\mathcal N}=2$ quiver gauge theories on $X_k$\label{sec:quivergaugeXk}}
\subsection{${\mathcal N}=2$ gauge theory\label{sec:N=2Xk}}
In this subsection we fix $j\in\{0,1, \ldots, k-1\}$ corresponding to a fixed holonomy at infinity.
The instanton partition function for the pure ${\mathcal N}=2$ $U(1)$ gauge
theory on the ALE space $X_k$ is the generating function (cf.\ \cite[Section~5.1]{art:bruzzopedrinisalaszabo2013})
\begin{align}
{\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j
:= & \ \sum_{\stackrel{\scriptstyle \vec{v}\in\frac{1}{k}\, {\mathbb{Z}}^{k-1}
}{\scriptstyle k\, v_{k-1}= j\bmod{k} }} \,
\vec{\xi}^{\ \vec{v}} \ \sum_{n=0}^\infty \, \mathsf{q}^{n+\frac{1}{2}\,
\vec{v}\cdot C\vec{v}} \ \int_{\mathcal{M}(\vec{u},n,j)} \,
\big[\mathcal{M}(\vec{u},n,j) \big]_T \\[4pt]
= & \ \sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{n=0}^\infty\,
\mathsf{q}^{\frac{1}{2}\, \vec{u}\cdot C^{-1}\vec{u}}\ \vec{\xi}^{\
C^{-1}\vec{u}}\, (-\mathsf{q})^{n} \, \big\langle
[\mathcal{M}(\vec{u},n,j)]_T\,,\,[\mathcal{M}(\vec{u},n,j)]_T \big\rangle_{\mathbb{W}_j}\ ,
\end{align}
where $\mathsf{q}\in{\mathbb{C}}^*$ with $|\mathsf{q}|<1$, and the fugacities
$\vec\xi:=(\xi_1,\dots,\xi_{k-1})\in({\mathbb{C}}^*)^{k-1}$ with $|\xi_i|<1$ can
be interpreted as coordinates on the maximal torus of the Lie group $SL(k,{\mathbb{C}})$.
In general, as described in \cite[Section
5.1]{art:bruzzopedrinisalaszabo2013}, the partition functions
factorize into products of the corresponding instanton partition
functions over the affine toric subsets $U_i\simeq {\mathbb{C}}^2$ of $X_k$ and
are weighted by the edge contributions $\ell^{(n)}_{\vec{v}}$ which
appear in the equivariant Euler classes of the Carlsson-Okounkov
bundle from Section \ref{se:bifundXk} (see Appendix \ref{app:edgecontributions}). The edge contributions for the rank one ${\mathcal N}=2$ gauge theory
on $X_k$ are roughly speaking the equivariant Euler classes of the
vector space $H^1({\mathscr{X}}_k, {\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty))$, which are
zero by the computation of the rank of the natural bundle in
\cite[Appendix A]{art:bruzzopedrinisalaszabo2013}, hence the edge
contribution is always equal to one. We thus obtain a factorization in
terms of Nekrasov partition functions for the pure ${\mathcal N}=2$ gauge
theory on ${\mathbb{R}}^4$ given by
\begin{equation}
{\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j =
\sum_{\vec{u}\in\mathfrak{U}_j}\, \mathsf{q}^{\frac{1}{2}\, \vec{u}\cdot
C^{-1}\vec{u}}\ \vec{\xi}^{\ C^{-1}\vec{u}} \ \prod_{i=1}^k\,
{\mathcal Z}_{{\mathbb{C}}^2}\big(\varepsilon_1^{(i)},\varepsilon_2^{(i)}; \mathsf{q} \big)\ .
\end{equation}
Let us denote by
\begin{equation}\label{eq:characterdef}
\chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big):=
\operatorname{Tr}_{{\mathcal V}(\,\widehat{\omega}_j\,)}\,
\mathsf{q}^{L_0^{\widehat{\mathfrak{sl}}_{k}}-\frac{k-1}{24} \, \operatorname{id}} \ \vec x^{\ \vec h}
\end{equation}
the character
of the $j$-th dominant highest weight representation of $\widehat{\mathfrak{sl}}_{k}$
at level one, with weight the $j$-th fundamental weight
$\widehat{\omega}_j$ of type $\widehat{A}_{k-1}$ for $j=0,1, \ldots,
k-1$; here $\vec\zeta:=\sum_{i=1}^{k-1}\, z_i\, H_i$ and $x_i:={\,\rm e}\,^{2\pi{\,{\rm i}\,}
z_i}$ for $i=1,\dots,k-1$, while $\vec h:=(h_1,\dots,h_{k-1})$. Setting $\xi_i={\,\rm e}\,^{2\pi \operatorname{i}\,
(2z_{i}-z_{i-1}-z_{i+1})}$ for $i=1,\dots,k-1$ with $z_0=z_k=0$, by explicit computation
of the character we get \cite[Section
5.3]{art:bruzzopedrinisalaszabo2013}
\begin{equation}\label{eq:character}
\chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big)=\frac{1}{\eta(\mathsf{q})^{k-1}} \,
\sum_{\vec{u}\in\mathfrak{U}_j}\, \mathsf{q}^{\frac{1}{2}\, \vec{u}\cdot
C^{-1}\vec{u}}\ \vec{\xi}^{\ C^{-1}\vec{u}}\ ,
\end{equation}
where
$\mathsf{q}^{\frac1{24}}\, \eta(\mathsf{q})^{-1} = \prod_{n=1}^\infty\, (1-q^n)^{-1} =\operatorname{Tr}_{{\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}}\,
\mathsf{q}^{L_0^\mathfrak{h}}$ is the character of the Fock space representation of
the Heisenberg algebra $\mathfrak{h}$.
By Equation \eqref{eq:nakajimayoshioka} and the identity
\begin{equation}\label{eq:identity}
\sum_{i=1}^k\, \frac{1}{\varepsilon_1^{(i)}\,
\varepsilon_2^{(i)}}=\frac{1}{k\, \varepsilon_1\, \varepsilon_2}
\end{equation}
we obtain explicitly
\begin{equation}
{\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j
= \eta(\mathsf{q})^{k-1} \ \chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \ \exp\Big(\, \frac \mathsf{q}{k\,\varepsilon_1\, \varepsilon_2}\, \Big) \ .
\end{equation}
\subsubsection{Gaiotto state}
Following Section \ref{sec:gaiottoC2}, we define the \emph{Gaiotto state} $G_j$ to be the sum, in the completed total equivariant cohomology $\widehat{{\mathbb{W}}}_j$, of all fundamental classes
\begin{equation}
G_j:= \sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{n\geq 0} \,
\big[\mathcal{M}(\vec{u},n,j) \big]_T\ .
\end{equation}
We also define the \emph{weighted Gaiotto state}
\begin{equation}
G_j\big(\mathsf{q},\vec{\xi} \ \big) := \sum_{\vec{u}\in\mathfrak{U}_j} \ \sum_{
n\geq 0} \, \mathsf{q}^{n+\frac{1}{2}\, \vec{u} \cdot C^{-1}\vec{u}}\,
\vec{\xi}^{\ C^{-1}\vec{u}}\ \big[\mathcal{M}(\vec{u},n,j) \big]_T
\end{equation}
in the completion
\begin{equation}
\widehat{{\mathbb{W}}}_j\big(\mathsf{q},\vec{\xi} \ \big):=\prod_{\vec{u}\in\mathfrak{U}_j}
\ \prod_{n\geq 0}\, \mathsf{q}^{n+\frac{1}{2}\, \vec{u} \cdot
C^{-1}\vec{u}}\ \vec{\xi}^{\ C^{-1}\vec{u}} \ {\mathbb{W}}_{\vec{u},n,j} \ .
\end{equation}
If we endow $\widehat{{\mathbb{W}}}_j\big(\mathsf{q},\vec{\xi} \ \big)$ with the scalar product
\begin{multline}
\Big\langle \, \mbox{$\sum\limits _{\vec{u}\in\mathfrak{U}_j}$} \ \mbox{$\sum\limits _{ n\geq0}$}\,
\mathsf{q}^{n+\frac{1}{2}\, \vec{u} \cdot C^{-1}\vec{u}}\ \vec{\xi}^{\
C^{-1}\vec{u}} \ \eta_{\vec{u},n} \,,\,
\mbox{$\sum\limits _{\vec{u}\in\mathfrak{U}_j} $}\ \mbox{$\sum\limits _{
n\geq0}$}\, \mathsf{q}^{n+\frac{1}{2}\, \vec{u} \cdot C^{-1}\vec{u}}\
\vec{\xi}^{\ C^{-1}\vec{u}} \ \nu_{\vec{u},n}
\Big\rangle_{\widehat{{\mathbb{W}}}_j(\mathsf{q},\vec{\xi}\ )}\\
:= \sum_{\vec{u}\in\mathfrak{U}_j}\ \sum_{ n=0}^\infty\, \mathsf{q}^{\frac{1}{2}\,
\vec{u} \cdot C^{-1}\vec{u}} \ (-\mathsf{q})^n \ \vec{\xi}^{\
C^{-1}\vec{u}} \ \big\langle \eta_{\vec{u},n}, \nu_{\vec{u},n} \big\rangle_{{\mathbb{W}}_{\vec{u},n,j}}\ ,
\end{multline}
\normalsize
then it is straightforward to see that the norm of the
weighted Gaiotto state is the instanton partition function for the ${\mathcal N}=2$ $U(1)$ gauge theory on $X_k$:
\begin{equation}
{\mathcal Z}_{X_k}\big(\varepsilon_1,\varepsilon_2; \mathsf{q}, \vec{\xi} \ \big)_j
= \big\langle G_j\big(\mathsf{q},\vec{\xi} \ \big)\,,\,
G_j\big(\mathsf{q},\vec{\xi} \
\big) \big\rangle_{\widehat{{\mathbb{W}}}_j(\mathsf{q},\vec{\xi} \ )}
\ .
\end{equation}
\begin{proposition}
The Gaiotto state is a Whittaker vector for $\widehat{\mathfrak{gl}}_k$ of type
$\chi$, where the algebra homomorphism
$\chi\colon \mathcal{U}(\mathfrak{h}^+ \oplus
\mathfrak{h}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2), \mathfrak{Q}}^+) \to {\mathbb{C}}(\varepsilon_1, \varepsilon_2)$ is defined by
\begin{align}
\chi(\mathfrak{q}^i_m) &= 0 \qquad \mbox{for} \quad m>0 \ , \ i = 1,\ldots, k-1\ ,\\[4pt]
\chi(\mathfrak{p}_m) & = \delta_{m,1}\
\sqrt{\big\langle [X_k], [X_k] \big\rangle_{{\mathbb{H}}_1}} \qquad \mbox{for}
\quad m>0\ .
\end{align}
\end{proposition}
\proof
We first note that under the isomorphism $\Psi_j$ defined in
\eqref{eq:isomWH} the fundamental class $[\mathcal{M}(\vec{u},n,j)]_T$ is
sent to $[\hilb{n}{X_k}]_T \otimes (\gamma_{\vec{u}}+\omega_j)$. Hence
under the isomorphism \eqref{eq:omega} the Gaiotto state becomes
\begin{equation}
G_j = \bigotimes_{i=1}^k \ \sum_{n\geq0}\, \big[\hilb{n}{U_i} \big]_T
\ \otimes \ \sum_{\vec{u}\in\mathfrak{U}_j} \, (\gamma_{\vec{u}}+\omega_j) \
\in \ \mbox{$\bigotimes\limits_{i=1}^k$} \, \widehat{{\mathbb{H}}}_{U_i} \ \otimes \
\mbox{$\prod\limits_{\vec{u}\in\mathfrak{U}_j} $} \, {\mathbb{C}}(\varepsilon_1,\varepsilon_2) (\gamma_{\vec{u}}+\omega_j)\ .
\end{equation}
By Proposition \ref{prop:gaiottowhittakerC2}, $\sum_{n\geq0}\,
[\hilb{n}{U_i}]_T$ is the Whittaker vector $G(\eta_i)$ for the
Heisenberg algebra $\mathfrak{h}_i$ with $\eta_i = \big(\varepsilon_2^{(i)}
\big)^{-1}$. It follows that $G_j$ is the Whittaker vector
$G_j(\vec{\eta}\, )$ for $\widehat{\mathfrak{gl}}_k$ as in Proposition
\ref{prop:whittakerALE} with $\vec{\eta}=(\eta_1, \ldots, \eta_k)$ of
type $\chi$ where
\begin{equation}
\chi(\mathfrak{q}_m^i) = \delta_{m,1} \, \big(\eta_{i+1}\,
\beta_{i+1}^{-1} - \eta_i \big)\qquad\mbox{and}\qquad
\chi(\mathfrak{p}_m) = \delta_{m,1} \ \sqrt{-k\, \varepsilon_1\,
\varepsilon_2} \ \sum_{i=1}^k\,
\frac{\eta_i}{\varepsilon_1^{(i)}}\ .
\end{equation}
By computing explicitly the quantities on the right-hand sides of these
equations, one gets the assertion.
\endproof
\subsection{Quiver gauge theories}
As we did in Section \ref{sec:R4quivers}, we will now add matter to
the ${\mathcal N}=2$ gauge theory and consider ${\mathcal N}=2$
superconformal quiver gauge theories on the ALE space $X_k$ with gauge group
$U(1)^{r+1}$ for $r\geq0$; we shall follow \cite{art:bruzzosalaszabo2015}, where superconformal quiver gauge theories on the ALE space $X_k$ are introduced.
For a quiver ${\tt Q}=({\tt Q}_0,{\tt Q}_1)$ we
fix vectors of integers
$(\vec
u_\upsilon,n_\upsilon,j_\upsilon)_{\upsilon\in{\tt Q}_0}$ representing
the topological numbers of the moduli spaces $\mathcal{M}(\vec u_\upsilon,
n_\upsilon, j_\upsilon)$ at the
vertices ${\tt Q}_0$ with $\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}$,
$n_\upsilon\in {\mathbb{N}}$, and $j_\upsilon\in\{0,1,\dots,k-1\}$. The fundamental (resp. antifundamental) hypermultiplets of masses
$\mu^s_{\upsilon}$, $s=1,\dots,m_\upsilon$ (resp. $\bar\mu^{\bar s}_{\upsilon}$,
$\bar s=1,\dots,\bar m_\upsilon$) at the nodes $\upsilon\in{\tt Q}_0$ correspond to the $T$-equivariant vector
bundles $\boldsymbol{V}_{\mu^s_{\upsilon}}^{\vec u_\upsilon, n_\upsilon,j_\upsilon}$
(resp. $\bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{\vec u_\upsilon,
n_\upsilon,j_\upsilon}$) on $\mathcal{M}(\vec u_\upsilon,n_\upsilon,j_\upsilon)$ obtained by
pushforward of $\boldsymbol{E}_{\mu^s_\upsilon}^{\vec
u_\upsilon,n_\upsilon,j_\upsilon; \vec u_\upsilon,n_\upsilon,j_\upsilon}$
(resp. $\boldsymbol{E}_{\bar\mu^{\bar s}_\upsilon}^{\vec
u_\upsilon,n_\upsilon,j_\upsilon; \vec u_\upsilon,n_\upsilon,j_\upsilon}$) with respect to the
projection of $\mathcal{M}(\vec u_\upsilon,n_\upsilon,j_\upsilon) \times \mathcal{M}(\vec u_\upsilon,n_\upsilon,j_\upsilon)$ to the
second (resp. first) factor. The bifundamental hypermultiplets of
masses $\mu_e$ at the edges
$e\in {\tt Q}_1$ correspond to
the vector bundles $\boldsymbol{E}_{\mu_e}^{\vec u_{{\tt s}(e)},
n_{{\tt s}(e)},j_{{\tt s}(e)};\vec
u_{{\tt t}(e)},n_{{\tt t}(e)},j_{{\tt t}(e)}}$ on $\mathcal{M}(\vec
u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)})\times \mathcal{M}(\vec
u_{{\tt t}(e)},n_{{\tt t}(e)}, j_{{\tt t}(e)})$; for vertex loops with ${\tt s}(e)={\tt t}(e)$ the restriction of
$\boldsymbol{E}_{\mu_e}^{\vec u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)};\vec u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)}}$ to the diagonal of $\mathcal{M}(\vec
u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)}) \times \mathcal{M}(\vec
u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)})$ describes an adjoint hypermultiplet
of mass $\mu_e$. The total matter field content of
the ${\mathcal N}=2$ quiver gauge theory on $X_k$ associated to ${\tt Q}$
in the sector labelled by $(\vec
u_\upsilon,n_\upsilon,j_\upsilon)_{\upsilon\in{\tt Q}_0}$
is thus described by the bundle on $\prod_{\upsilon\in{\tt Q}_0}\,\mathcal{M}(\vec u_\upsilon,n_\upsilon,j_\upsilon)$
given by
\begin{align}
\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)}:= & \ \bigoplus_{\upsilon\in{\tt Q}_0}\, p_\upsilon^* \Big(\, \bigoplus_{s=1}^{m_\upsilon}\,
\boldsymbol{V}_{\mu^s_{\upsilon}}^{\vec u_\upsilon, n_\upsilon,j_\upsilon} \
\oplus \ \bigoplus_{\bar s=1}^{\bar m_\upsilon}\,
\bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{\vec
u_\upsilon,n_\upsilon,j_\upsilon}\, \Big) \\ & \ \qquad \ \oplus \ \bigoplus_{e\in{\tt Q}_1}\,
p_e^*\boldsymbol{E}_{\mu_e}^{\vec
u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)};\vec u_{{\tt t}(e)},n_{{\tt t}(e)},j_{{\tt t}(e)}} \ ,
\end{align}
where $p_\upsilon$ is the projection of $\prod_{\upsilon\in{\tt Q}_0}\,
\mathcal{M}(\vec u_\upsilon,n_\upsilon,j_\upsilon)$ to the $\upsilon$-th
factor while $p_e$ is the projection
to the product $\mathcal{M}(\vec
u_{{\tt s}(e)},n_{{\tt s}(e)},j_{{\tt s}(e)})\times \mathcal{M}(\vec
u_{{\tt t}(e)},n_{{\tt t}(e)}, j_{{\tt t}(e)})$.
The degree of the Euler class of the hypermultiplet bundle $\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)}$ is given by
\begin{multline}
{\rm deg}\, \operatorname{eu}\big(\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)}\, \big):= \sum_{\upsilon\in {\tt Q}_0}\, \dim \mathcal{M}(\vec
u_\upsilon,n_\upsilon,j_\upsilon)- \operatorname{rk}\big(\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)}\,\big)\\[4pt] \shoveleft{=\sum_{\upsilon\in {\tt Q}_0}\,2\, n_\upsilon -\, \sum_{\upsilon\in {\tt Q}_0}\, \big(m_\upsilon+\bar m_\upsilon\big)\, \Big(n_\upsilon+\mbox{$\frac12$}\, \vec v_\upsilon\cdot C \vec
v_\upsilon-\mbox{$\frac1{2k}$}\,
j_\upsilon \, (k- j_\upsilon) \Big)} \\
-\sum_{e\in {\tt Q}_1}\, \Big(\, n_{{\tt t}(e)}+n_{{\tt s}(e)}
+\, \mbox{$\frac12$} \, \vec{v}_{{\tt s}(e)}\cdot C\vec{v}_{{\tt s}(e)}+\mbox{$\frac12$}\, \vec{v}_{{\tt t}(e)}\cdot C\vec{v}_{{\tt t}(e)} -\vec{v}_{{\tt s}(e)}\cdot C\vec{v}_{{\tt t}(e)}-\mbox{$\frac1{2k}$}\,
j_{{\tt t}(e){\tt s}(e)} \, (k- j_{{\tt t}(e){\tt s}(e)})\, \Big)\ ,
\end{multline}
where $\vec{v}_\upsilon:=C^{-1}\vec{u}_\upsilon$. Using \eqref{eq:confconstr} the degree becomes
\begin{equation}
{\rm deg}\, \operatorname{eu}\big(\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)}\,\big) =
\sum_{\upsilon\in {\tt Q}_0}\,d^{X_k}_\upsilon(\vec{\boldsymbol v}_\upsilon) \ ,
\end{equation}
where we defined
\begin{multline}
d^{X_k}_\upsilon(\vec{\boldsymbol v}_\upsilon) := \mbox{$\frac1{2k}$}\,
j_\upsilon \, (k- j_\upsilon)\,\big(2-\#\{e\in{\tt Q}_1\, | \, {\tt s}(e)=\upsilon\}-
\#\{e\in{\tt Q}_1\, | \, {\tt t}(e)=\upsilon\}\big) \\
+\mbox{$\frac1{4k}$}\,\Big( \sum_{\stackrel{\scriptstyle e\in{\tt Q}_1}{\scriptstyle {\tt s}(e)=\upsilon}} \,j_{{\tt t}(e)\upsilon} \, (k- j_{{\tt t}(e)\upsilon})
+\sum_{\stackrel{\scriptstyle e\in{\tt Q}_1}{\scriptstyle {\tt t}(e)=\upsilon}} \,j_{\upsilon{\tt s}(e)} \, (k- j_{\upsilon{\tt s}(e)}) \Big) \\
- \vec{v}_\upsilon\cdot C \vec{v}_\upsilon+\mbox{$\frac12$}\, \sum_{\stackrel{\scriptstyle e\in{\tt Q}_1}{\scriptstyle {\tt s}(e)=\upsilon}}\, \vec{v}_\upsilon\cdot C \vec{v}_{{\tt t}(e)}+\mbox{$\frac12$}\, \sum_{\stackrel{\scriptstyle e\in{\tt Q}_1}{\scriptstyle{\tt t}(e)=\upsilon}}\, \vec{v}_\upsilon\cdot C \vec{v}_{{\tt s}(e)}
\end{multline}
for each vertex $\upsilon\in{\tt Q}_0$; here $\vec{\boldsymbol v}_\upsilon:=\big(\vec v_\upsilon,(\vec v_{{\tt t}(e)})_{e\in{\tt Q}_1\,:\, {\tt s}(e)=\upsilon},(\vec v_{{\tt s}(e)})_{e\in{\tt Q}_1\,:\, {\tt t}(e)=\upsilon}\big)$. By analogy with the case of gauge theories on ${\mathbb{R}}^4$ (see Section \ref{sec:R4quivers}), we say that the ${\mathcal N}=2$ quiver gauge theory on $X_k$ is conformal if $d^{X_k}_\upsilon(\vec{\boldsymbol v}_\upsilon)=0$ for all $\upsilon\in{\tt Q}_0$; this is formally the requirement of vanishing beta-function for the running of the $\upsilon$-th gauge coupling constant. For any vertex $\upsilon\in {\tt Q}_0$, define the set of conformal fractional instanton charges by
\begin{equation}\label{eq:confchargeset}
\mathfrak{U}_{j_\upsilon}^{\rm conf}:=\big\{\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon} \
\big\vert \ d^{X_k}_\upsilon(\vec{\boldsymbol v}_\upsilon)=0 \big\}\ .
\end{equation}
The conformal constraint is always trivially satisfied by any $\vec u_\upsilon$ for the $\widehat{A}_0$-theory, while for the $A_0$-theory the set of conformal fractional instantons charges reduces to
\begin{equation}
\mathfrak{U}_{j}^{\rm conf}:= \big\{\vec u\in\mathfrak{U}_{j} \
\big\vert \ \vec u\cdot C^{-1}\vec
u= \mbox{$\frac1k$}\, j\,(k- j) \big\} \ .
\end{equation}
Note that in this case, this is a restriction on the
conformal dimension $\Delta_{\vec u}=\frac1{2} \,\langle \omega_{j},\omega_{j}\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{Q}}}$ of the highest weight representation ${\mathbb{W}}_{\vec u,j}$ of the Virasoro algebra.
Introduce topological couplings $\mathsf{q}_\upsilon\in{\mathbb{C}}^*$ with
$|\mathsf{q}_\upsilon|<1$ and $\vec\xi_v=\big((\xi_v)_1,\dots,
(\xi_v)_{k-1}\big) \in ({\mathbb{C}}^*)^{k-1}$ with $|(\xi_v)_i|<1$ at each vertex
$\upsilon\in{\tt Q}_0$. With notation as in Section \ref{sec:R4quivers},
the quiver gauge theory partition function on $X_k$ is then defined by
the generating function
\begin{multline}
{\mathcal Z}_{X_k}^{{\tt Q}}
\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi} \ \big)_{\boldsymbol{j}} :=
\sum_{(\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf})} \,
\vec\boldsymbol{\xi}{}^{\ C^{-1}\vec\boldsymbol{u}} \ \sum_{(n_\upsilon)\in{\mathbb{N}}^{{\tt Q}_0}}\,
\boldsymbol{\mathsf{q}}^{\boldsymbol{n} + \frac12\, \vec\boldsymbol{u}\cdot C^{-1}\vec\boldsymbol{u}} \\ \shoveright{
\times \
\int_{\prod\limits_{\upsilon\in{\tt Q}_0}\, \mathcal{M}(\vec
u_\upsilon,n_\upsilon,j_\upsilon)} \, \operatorname{eu}_{T\times
T_{\boldsymbol{\mu}}}\big(\boldsymbol{M}^{(\vec
u_\upsilon,n_\upsilon,j_\upsilon)}_{(\mu^s_{\upsilon}),(\bar\mu^{\bar
s}_\upsilon),
(\mu_e)} \big) } \\[4pt] \shoveleft{ = \sum_{(\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf})} \
\sum_{(n_\upsilon)\in{\mathbb{N}}^{{\tt Q}_0}}\,
\boldsymbol{\mathsf{q}}^{\boldsymbol{n} + \frac12\, \vec\boldsymbol{u}\cdot C^{-1}\vec\boldsymbol{u}} \
\vec\boldsymbol{\xi}{}^{\ C^{-1}\vec\boldsymbol{u}} } \\ \times \
\int_{\prod\limits_{\upsilon\in{\tt Q}_0}\, \mathcal{M}(\vec
u_\upsilon,n_\upsilon,j_\upsilon)} \
\prod_{\upsilon\in{\tt Q}_0}\, p_\upsilon^*\Big(\, \prod_{s=1}^{m_\upsilon}\,
\operatorname{eu}_T\big(\boldsymbol{V}_{\mu^s_{\upsilon}}^{\vec u_\upsilon, n_\upsilon,j_\upsilon}\big) \
\prod_{\bar s=1}^{\bar m_\upsilon}\,
\operatorname{eu}_T\big(\bar\boldsymbol{V}_{\bar\mu^{\bar s}_{\upsilon}}^{\vec u_\upsilon,
n_\upsilon,j_\upsilon}\big)\, \Big) \\
\times \ \prod_{e\in{\tt Q}_1}\,
p_e^*\, \operatorname{eu}_T\big(\boldsymbol{E}_{\mu_e}^{\vec u_{{\tt s}(e)},
n_{{\tt s}(e)},j_{{\tt s}(e)};\vec
u_{{\tt t}(e)},n_{{\tt t}(e)},j_{{\tt t}(e)}}
\big) \ ,
\end{multline}
where $\boldsymbol{\mathsf{q}}^{\boldsymbol{n} + \frac12\, \vec\boldsymbol{u}\cdot C^{-1}\vec\boldsymbol{u}}:=
\prod_{\upsilon\in{\tt Q}_0}\, \mathsf{q}_\upsilon^{n_\upsilon+\frac12\, \vec
u_\upsilon\cdot C^{-1}\vec u_\upsilon}$ and $\vec\boldsymbol{\xi}{}^{\
C^{-1}\vec\boldsymbol{u}} := \prod_{\upsilon\in{\tt Q}_0}\, \vec\xi_\upsilon^{\
C^{-1}\vec u_\upsilon}$. By applying
the localization theorem, and using Equations \eqref{eq:ZcalC2quiver}
and \eqref{eq:vertexoperator-fixedbasis}, we obtain a factorization in
terms of ${\mathcal N}=2$ quiver gauge theory partition functions on ${\mathbb{R}}^4$
weighted by edge contributions. For the fundamental and
antifundamental matter fields, the relevant edge contributions
$\ell_{\vec v_\upsilon}^{(n)}$ are the equivariant Euler classes of
$H^1({\mathscr{X}}_k,{\mathcal R}^{\vec{u}_\upsilon}\otimes
{\mathcal O}_{{\mathscr{X}}_k}(-{\mathscr{D}}_\infty))$ for $\upsilon\in{\tt Q}_0$; by
\cite[Section 5]{art:bruzzopedrinisalaszabo2013} this vector space is
zero if and only if $\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf}$ and the
corresponding edge contribution is equal to one. Thus only the arrows
of the quiver yield edge contributions and the partition function is
given by
\begin{multline}
{\mathcal Z}_{X_k}^{{\tt Q}}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} = \prod_{\upsilon\in{\tt Q}_0}\, \mathsf{q}_\upsilon^{\frac12\,
\langle \omega_{j_\upsilon},\omega_{j_\upsilon}\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{Q}}}} \ \sum_{(\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf})} \,
\vec\boldsymbol{\xi}{}^{\ C^{-1}\vec\boldsymbol{u}} \ \prod_{n=1}^{k-1} \ \prod_{e\in{\tt Q}_1}\, \ell^{(n)}_{\vec
v_e}\big(\varepsilon_1^{(n)},\varepsilon_2^{(n)},\mu_e \big) \\
\times \ \prod_{i=1}^k\,
{\mathcal Z}_{{\mathbb{C}}^2}^{{\tt Q}}\big(\varepsilon_1^{(i)},\varepsilon_2^{(i)},\boldsymbol{\mu}^{(i)}
;\boldsymbol{\mathsf{q}} \big) \ ,
\end{multline}
where $\vec v_e:= \vec v_{{\tt t}(e)}-\vec v_{{\tt s}(e)}$; the shifted
masses are
\begin{equation}
\big(\mu_v^s\big)^{(i)}:= \mu_v^s-(\vec v_v)_i\, \varepsilon_1^{(i)} -
(\vec v_v)_{i-1} \, \varepsilon_2^{(i)}
\end{equation}
for $v\in{\tt Q}_0$, $s=1,\dots,m_v$ and $i=1,\dots,k$, and similarly
for $\big(\bar\mu_v^{\bar s}\big)^{(i)}$, whereas
\begin{equation}
\mu_e^{(i)}:= \mu_e-(\vec v_e)_i\, \varepsilon_1^{(i)} -
(\vec v_e)_{i-1} \, \varepsilon_2^{(i)}
\end{equation}
for $e\in{\tt Q}_1$ and $i=1,\dots,k$.
In the remainder of this section we consider in detail each of the
admissible quivers in turn.
\subsection{$\widehat{A}_{r}$ theories}
With the conventions of Section
\ref{sec:hatArtheories}, the instanton partition function for the
${\mathcal N}=2$ $U(1)^{r+1}$ quiver gauge theory of type $\widehat{A}_{r}$ on the ALE
space $X_k$ reads
as
\begin{multline}
{\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} = \prod_{\upsilon=0}^r \, \mathsf{q}_\upsilon^{\frac12\,
\langle \omega_{j_\upsilon},\omega_{j_\upsilon}\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{Q}}}} \ \sum_{(\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf})} \,
\vec\boldsymbol{\xi}{}^{\
C^{-1}\vec\boldsymbol{u}} \ \prod_{n=1}^{k-1} \ \prod_{\upsilon=0}^r \, \ell^{(n)}_{\vec
v_{\upsilon,\upsilon+1}} \big(\varepsilon_1^{(n)},\varepsilon_2^{(n)},\mu_\upsilon \big) \\
\times \ \prod_{i=1}^k\,
{\mathcal Z}_{{\mathbb{C}}^2}^{\widehat{A}_r} \big(\varepsilon_1^{(i)},\varepsilon_2^{(i)},\boldsymbol{\mu}^{(i)}
;\boldsymbol{\mathsf{q}} \big) \ , \label{eq:ZXkhatAr}
\end{multline}
where $\vec\boldsymbol{\xi}{}^{\
C^{-1}\vec\boldsymbol{u}} := \prod_{\upsilon=0}^r\, \vec\xi_\upsilon^{\
C^{-1}\vec u_\upsilon}$ with $\vec u_{r+1}:= \vec u_0$, while
$\mu_\upsilon^{(i)}:= \mu_\upsilon-(\vec v_{\upsilon,\upsilon+1})_i\, \varepsilon_1^{(i)} -
(\vec v_{\upsilon,\upsilon+1})_{i-1} \, \varepsilon_2^{(i)}$ with
$\vec v_{\upsilon,\upsilon+1}:=\vec
v_{\upsilon+1}-\vec v_{\upsilon}$.
\subsubsection{Conformal blocks}
We will relate the partition function \eqref{eq:ZXkhatAr} to the trace of vertex
operators $\operatorname{V}_{\mu_\upsilon}^{j_\upsilon, j_{\upsilon+1}}(\vec x_\upsilon,z_\upsilon)$ from Section \ref{sec:chiralvertex},
analogously to what we did in
Section \ref{sec:hatArtheories}, and hence interpret it as a torus
$(r+1)$-point conformal block. For this, we fix vertices
$\upsilon,\upsilon'\in\{0,1,\dots, r\}$ and introduce the \emph{conformal
restriction operators} $\delta^{\rm
conf}_{\upsilon,{\upsilon'}}\colon{\mathbb{W}}\to {\mathbb{W}}$ which are defined by
their matrix elements in the fixed point basis of the vector space ${\mathbb{W}}$ by
\begin{equation}
\big\langle \delta^{\rm
conf}_{\upsilon,{\upsilon'}}\triangleright [\vec Y,\vec u\, ]\,,\, [\vec
Y',\vec u\,'\,] \big\rangle_{{\mathbb{W}}}:=
\left\{
\begin{array}{ll}
1 & \mbox{if} \quad \vec u\in\mathfrak{U}_{j_\upsilon}^{\rm conf} \ ,\ \vec u\,'\in
\mathfrak{U}_{j_{\upsilon'}}^{\rm conf} \ , \\[4pt]
0 & \mbox{otherwise} \ .
\end{array} \right. \label{eq:confrestrop}
\end{equation}
Suitable insertions of this operator restrict the first Chern classes $\vec
u_\upsilon\in\mathfrak{U}_{j_\upsilon}$ in the way required by the
superconformal constraints of the quiver gauge theory and the
constrained conformal dimensions of the associated Virasoro algebras
at the nodes of the $\widehat{A}_r$-type quivers. Using Proposition
\ref{prop:L0eigen} and Proposition \ref{prop:Wkweight}, by performing
analogous manipulations to those used in the proof of Proposition
\ref{prop:hatArtrace} we arrive at the following result.
\begin{proposition}\label{prop:ZcalXkArTr}
The partition function of the $\widehat{A}_r$-theory on $X_k$ is given
by
\begin{equation}
{\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} = \operatorname{Tr}_{{\mathbb{W}}_{j_0}} \, \mathsf{q}^{L_0}\ \vec\xi{}^{\ C^{-1}\vec h} \ \prod_{\upsilon=0}^r\,
\operatorname{V}_{\mu_\upsilon}^{j_\upsilon,j_{\upsilon+1}}(\vec x_\upsilon,z_\upsilon)\ \delta^{\rm
conf}_{\upsilon,{\upsilon+1}}
\end{equation}
independently of $z_0\in{\mathbb{C}}^*$ and $\vec x_0\in({\mathbb{C}}^*)^{k-1}$, where
$\mathsf{q}:= \mathsf{q}_0\, \mathsf{q}_1\cdots \mathsf{q}_r$, $( \, \vec\xi\ )_i:=
(\vec\xi_0)_i\, (\vec\xi_1)_i\, \cdots (\vec\xi_r)_i$, $z_\upsilon:=
z_0\, \mathsf{q}_1\cdots \mathsf{q}_\upsilon$, and $(\vec x_\upsilon)_i:= (\vec
x_0)_i\, (\vec\xi_1)_i\cdots (\vec\xi_\upsilon)_i$ for
$\upsilon=1,\dots,r$ and $i=1,\dots,k-1$.
\end{proposition}
By combining Theorem \ref{thm:virasoroprimary} and Proposition \ref{prop:ZcalXkArTr}, it follows that the quiver gauge theory partition function completely factorizes under the isomorphism of Proposition \ref{prop:representation} into partition functions associated to the affine algebras $\mathfrak{h}$ and $\widehat{\mathfrak{sl}}_k$.
\begin{corollary}\label{cor:trace}
Let $\operatorname{V}_\mu(\vec{v}_{21},\vec x,
z)$ be the vertex operator in
$\operatorname{Hom}({\mathbb{W}}_{j_1},
{\mathbb{W}}_{j_2})[[
z^{\pm\,1},x_1^{\pm\,1},\dots,x_{k-1}^{\pm\,1}]]$ given by
\begin{equation}\label{eq:Vmudef}
\operatorname{V}_\mu(\vec{v}_{21},\vec x,
z):= \, z^{\Delta_{\vec u_2}-\Delta_{\vec u_1}} {\bar \operatorname{V}}_\mu(\vec{v}_{21}, \vec x, z)\ \exp\big(\log z\ \mathfrak{c}-\gamma_{21} \big)\, \exp\big(\gamma_{21}\,\big)\ .
\end{equation}
Then
\begin{multline}
{\mathcal Z}_{X_k}^{\widehat{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}
= {\mathcal Z}^{\widehat{A}_{r}}_{{\mathbb{C}}^2}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\mathsf{q})^{\frac{1}{k}}\ \mathsf{q}^{\frac{1}{24}\,(1-\frac{1}{k})}\,\eta(\mathsf{q})^{\frac{1}{k}-1} \\
\times\ \operatorname{Tr}_{{\mathcal V}(\,\widehat{\omega}_{j_0}\,)} \,
\mathsf{q}^{L_0^{\widehat{\mathfrak{sl}}_k}}\ \vec\xi{}^{\ C^{-1}\vec h}\ \prod_{\upsilon=0}^r\
\sum_{(\vec{u}_\upsilon\in\mathfrak{U}^{\rm conf}_{j_\upsilon})} \,
\operatorname{V}_{\mu_\upsilon}(\vec{v}_{\upsilon,\upsilon+1},\vec x_\upsilon, z_\upsilon) \, \big\vert
_{{\mathbb{W}}_{\vec{u}_\upsilon,j_\upsilon}} \ .
\end{multline}
\end{corollary}
\proof
By Theorem \ref{thm:virasoroprimary} and Proposition \ref{prop:representation} we get
\begin{multline}
\operatorname{Tr}_{{\mathbb{W}}_{j_0}} \, \mathsf{q}^{L_0}\ \vec\xi{}^{\ C^{-1}\vec h} \ \prod_{\upsilon=0}^r\,
\operatorname{V}_{\mu_\upsilon}^{j_\upsilon,j_{\upsilon+1}}(\vec x_\upsilon,z_\upsilon)\ \delta^{\rm
conf}_{\upsilon,{\upsilon+1}}
=\operatorname{Tr}_{{\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \prod_{\upsilon=0}^r\,
\operatorname{V}_{-\frac{\mu_\upsilon}{\sqrt{-k\,\varepsilon_1\, \varepsilon_2}}, \frac{\mu_\upsilon+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\, \varepsilon_2}}}(z_\upsilon)\\
\times\ \operatorname{Tr}_{{\mathcal V}(\,\widehat{\omega}_{j_0}\,)} \, \mathsf{q}^{L_0^{\widehat{\mathfrak{sl}}_k}}\ \vec\xi{}^{\ C^{-1}\vec h} \ \prod_{\upsilon=0}^r \
\sum_{\vec{u}_1\in\mathfrak{U}_{j_\upsilon},\vec{u}_2\in\mathfrak{U}_{j_{\upsilon+1}}} \,
\operatorname{V}_\mu(\vec{v}_{21},z_\upsilon)\, \big\vert
_{{\mathbb{W}}_{\vec{u}_1,j_\upsilon}}\ \delta^{\rm
conf}_{\upsilon,{\upsilon+1}} \ .
\end{multline}
Then by using the same arguments as in the proof of \cite[Corollary 1]{art:carlssonokounkov2012} one gets
\begin{equation}
\operatorname{Tr}_{{\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \prod_{\upsilon=0}^r\,
\operatorname{V}_{-\frac{\mu_\upsilon}{\sqrt{-k\, \varepsilon_1\, \varepsilon_2}}, \frac{\mu_\upsilon+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\, \varepsilon_2}}}(z_\upsilon)=\prod_{\upsilon=0}^r\,\big(\mathsf{q}_\upsilon^{-\frac{1}{24}}\, \eta(\mathsf{q}_\upsilon)\big)^{-\frac{\mu_\upsilon\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{k\, \varepsilon_1\, \varepsilon_2}} \ \mathsf{q}^{\frac{1}{24}}\,\eta(\mathsf{q})^{-1} \ .
\end{equation}
The result now follows from Proposition \ref{prop:tracededekind}.
\endproof
\subsubsection{$\widehat{A}_{0}$ theory}
For the ${\mathcal N}=2^*$ gauge theory on $X_k$, similar arguments to those
of Section \ref{sec:N=2Xk} show that the edge contributions are also
equal to
one in this case. In this instance the gauge theory is automatically
conformal without further restriction of the first Chern classes $\vec
u\in\mathfrak{U}_j$. Then the instanton partition function for $U(1)$
gauge theory on the ALE space $X_k$ with a single adjoint
hypermultiplet of mass $\mu$ can be written in a factorized form in
terms of the Nekrasov partition function for ${\mathcal N}=2^\ast$ gauge
theory on ${\mathbb{R}}^4$ given by
\begin{equation}
{\mathcal Z}_{X_k}^{\widehat{A}_0} \big(\varepsilon_1,\varepsilon_2,\mu;
\mathsf{q}, \vec{\xi} \ \big)_j =\eta(\mathsf{q})^{k-1} \ \chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \
\prod_{i=1}^k \,
{\mathcal Z}_{{\mathbb{C}}^2}^{\widehat{A}_0} \big(\varepsilon_1^{(i)},\varepsilon_2^{(i)}, \mu ;
\mathsf{q} \big)\ .
\end{equation}
In this case Proposition
\ref{prop:ZcalXkArTr} may be stated in a factorized form under the decomposition of Theorem \ref{thm:virasoroprimary} in terms of characters of
$\mathfrak{h} \subset
\widehat{\mathfrak{gl}}_k$ and $\widehat{\mathfrak{gl}}_k$ as
\begin{equation}
{\mathcal Z}_{X_k}^{\widehat{A}_0} \big(\varepsilon_1,\varepsilon_2,\mu;
\mathsf{q}, \vec{\xi} \ \big)_j = \eta(\mathsf{q})^{k-1} \ \chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \
\operatorname{Tr}_{{\mathbb{H}}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \operatorname{V}_{-\frac{\mu}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}}(1) \ .
\end{equation}
By using the
identities \eqref{eq:Ahat0dedekind} and \eqref{eq:identity}, we obtain explicitly
\begin{equation}
{\mathcal Z}_{X_k}^{\widehat{A}_0} \big(\varepsilon_1,\varepsilon_2, \mu;
\mathsf{q}, \vec{\xi} \ \big)_j \ = \mathsf{q}^{\frac{k}{24}}\ \eta(\mathsf{q})^{-1} \ \chi^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \
\big(\mathsf{q}^{-\frac{1}{24}}\,
\eta(\mathsf{q})\big)^{-\frac{\mu\, (\mu+\varepsilon_1+\varepsilon_2)}{k\,
\varepsilon_1\, \varepsilon_2}}\ .
\end{equation}
\begin{remark}
Note that ${\mathbb{H}}$ is not the Fock space of $\mathfrak{h}$, as we have
\begin{equation}
\operatorname{Tr}_{{\mathbb{H}}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \operatorname{V}_{-\frac{\mu}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}}(1) =\big(\mathsf{q}^{\frac{1}{24}}\,
\eta(\mathsf{q})^{-1} \big)^{k-1}\ \operatorname{Tr}_{{\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}} \, \mathsf{q}^{L_0^{\mathfrak{h}}} \ \operatorname{V}_{-\frac{\mu}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}}(1)\ .
\end{equation}
\end{remark}
\subsection{$A_r$ theories}
With the conventions of Section
\ref{sec:Artheories}, the instanton partition function for the
${\mathcal N}=2$ $U(1)^{r+1}$ quiver gauge theory of type ${A}_{r}$ on the ALE
space $X_k$ reads
as
\begin{multline}
{\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} = \prod_{\upsilon=0}^r \, \mathsf{q}_\upsilon^{\frac12\,
\langle \omega_{j_\upsilon},\omega_{j_\upsilon}\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{Q}}}} \ \sum_{(\vec u_\upsilon\in\mathfrak{U}_{j_\upsilon}^{\rm conf})} \,
\vec\boldsymbol{\xi}{}^{\
C^{-1}\vec\boldsymbol{u}} \ \prod_{n=1}^{k-1} \ \prod_{\upsilon=0}^{r-1} \, \ell^{(n)}_{\vec
v_{\upsilon,\upsilon+1}} \big(\varepsilon_1^{(n)},\varepsilon_2^{(n)},\mu_{\upsilon+1} \big) \\
\times \ \prod_{i=1}^k\,
{\mathcal Z}_{{\mathbb{C}}^2}^{{A}_r} \big(\varepsilon_1^{(i)},\varepsilon_2^{(i)},\boldsymbol{\mu}^{(i)}
;\boldsymbol{\mathsf{q}} \big) \ . \label{eq:ZXkAr}
\end{multline}
\subsubsection{Conformal blocks}
By performing analogous manipulations to those used in the proof of
Proposition \ref{prop:fundamental-matter}, we can express the
partition function \eqref{eq:ZXkAr} as a particular matrix element of vertex operators and hence interpret it as an $(r+4)$-point conformal block on the sphere. For this, let
\begin{equation}
|0\rangle_{\rm conf}:= \prod_{\upsilon=0}^r\,\delta^{\rm
conf}_{0,\upsilon}\triangleright [\emptyset, \vec{0}\,] \ .
\end{equation}
\begin{proposition}\label{prop:ZcalXkArcorr}
The partition function of the $A_r$-theory on $X_k$ is given
by
\begin{multline}
{\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}} \\
= \Big\langle |0\rangle_{\rm conf} \, , \, V_{\mu_0}(\vec x_0,z_0) \, \Big(\, \prod_{\upsilon=1}^{r}\,
\operatorname{V}_{\mu_{\upsilon}}^{j_{\upsilon-1}, j_{\upsilon}}(\vec x_\upsilon,z_\upsilon)\ \delta^{\rm
conf}_{{\upsilon-1},{\upsilon}} \, \Big)\, \operatorname{V}_{\mu_{r+1}}(\vec x_{r+1},z_{r+1}) |0\rangle_{\rm conf} \Big\rangle_{{\mathbb{W}}}
\end{multline}
independently of $z_0\in{\mathbb{C}}^*$ and $\vec x_0\in({\mathbb{C}}^*)^{k-1}$, where
$z_\upsilon:=
z_0\, \mathsf{q}_0\, \mathsf{q}_1\cdots \mathsf{q}_\upsilon$ and $(\vec x_\upsilon)_i:= (\vec
x_0)_i\, (\vec\xi_0)_i\, (\vec\xi_1)_i$ $\cdots (\vec\xi_{\upsilon-1})_i$ for
$\upsilon=1,\dots,r+1$ and $i=1,\dots,k-1$.
\end{proposition}
Again, combining Theorem \ref{thm:virasoroprimary} and Proposition \ref{prop:ZcalXkArcorr} yields a completely factorized form for the quiver gauge theory partition function under the isomorphism of Proposition \ref{prop:representation}. In the following we denote ${\mathcal V}:=\bigoplus_{j=0}^{k-1}\, {\mathcal V}(\,\widehat{\omega}_j\,)$.
\begin{corollary}
Let $\operatorname{V}_\mu(\vec{v}_{21},\vec x,
z)$ be the vertex operator in
$\operatorname{Hom}({\mathbb{W}}_{j_1},
{\mathbb{W}}_{j_2})[[
z^{\pm\,1},x_1^{\pm\,1},\dots,x_{k-1}^{\pm\,1}]]$ given by \eqref{eq:Vmudef}. Then
\begin{multline}
{\mathcal Z}_{X_k}^{{A}_r}\big(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}}, \vec\boldsymbol{\xi}
\ \big)_{\boldsymbol{j}}
= {\mathcal Z}_{{\mathbb{C}}^2}^{{A}_r}(\varepsilon_1,\varepsilon_2,\boldsymbol{\mu};\boldsymbol{\mathsf{q}})^{\frac1k} \\ \shoveleft{ \times \
\Big\langle |0\rangle_{\rm conf} \, , \, \Big(\, \sum_{j_0,j_0'=0}^{k-1} \
\sum_{\vec{u}_0\in\mathfrak{U}_{j_0},\vec{u}_0'\in\mathfrak{U}_{j_0'}} \
\operatorname{V}_{\mu_0}(\vec{v}_{0',0},\vec x_0, z_0)\, \big\vert
_{{\mathbb{W}}_{\vec{u}_0,j_0}} \, \Big) } \\ \times \
\prod_{\upsilon=1}^{r} \ \sum_{(\vec u_\upsilon\in \mathfrak{U}^{\rm conf}_{j_\upsilon})}\,
\operatorname{V}_{\mu_{\upsilon}}(\vec v_{\upsilon-1,\upsilon},\vec x_\upsilon,z_\upsilon)\, \big|_{{\mathbb{W}}_{\vec u_\upsilon,j_\upsilon}} \\ \times \ \Big(\,
\sum_{j_{r+1},j_{r+1}'=0}^{k-1} \
\sum_{\vec{u}_{r+1}\in\mathfrak{U}_{j_{r+1}},\vec{u}_{r+1}'\in\mathfrak{U}_{j_{r+1}'}} \
\operatorname{V}_{\mu_{r+1}}(\vec{v}_{r+1',r+1},\vec x_{r+1}, z_{r+1})\, \big\vert
_{{\mathbb{W}}_{\vec{u}_{r+1},j_{r+1}}}\, \Big)
|0\rangle_{\rm conf} \Big\rangle_{{\mathcal V}} \ .
\end{multline}
\end{corollary}
\proof
The proof follows that of Corollary \ref{cor:trace}, and by repeating the proof of Proposition \ref{prop:ZC2Arexpl} to compute
\begin{equation}
\Big\langle |0\rangle\,,\, \prod_{\upsilon=0}^{r+1}\, \operatorname{V}_{-\frac{\mu_\upsilon}{\sqrt{-k\, \varepsilon_1\, \varepsilon_2}}, \frac{\mu_\upsilon+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\, \varepsilon_2}}}(z_\upsilon) |0\rangle\Big\rangle_{{\mathcal F}_{{\mathbb{C}}(\varepsilon_1,\varepsilon_2)}} = \prod_{0\leq \upsilon<\upsilon'\leq r+1}\,
\big(1-\mathsf{q}_{\upsilon+1}\cdots
\mathsf{q}_{\upsilon'}\big)^{-\frac{\mu_{\upsilon'}\,
(\mu_\upsilon+\varepsilon_1+\varepsilon_2)}{k\, \varepsilon_1\,
\varepsilon_2}}
\ .
\end{equation}
\endproof
\subsubsection{$A_0$ theory}
For the ${\mathcal N}=2$ superconformal gauge theory on $X_k$ with two
fundamental hypermultiplets of masses $\mu_0,\mu_1$, the set
$\mathfrak{U}_j^{\rm conf}$ coincides with the rank one limit of the more
general conformal charge sets obtained in \cite[Section
5.4]{art:bruzzopedrinisalaszabo2013}. Analogously to Equation
\eqref{eq:characterdef}, let us define the restricted
$\widehat{\mathfrak{sl}}_k$ characters
\begin{align}
\chi_{\mathrm{conf}}^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big):=& \ \operatorname{Tr}_{{\mathcal V}(\,\widehat{\omega}_j\,)}\,
\mathsf{q}^{L_0^{\widehat{\mathfrak{sl}}_{k}}-\frac{k-1}{24}\, \operatorname{id}}\ \vec x\,^{
\vec h} \ \delta^{\rm conf}_{j,j} \\[4pt] =& \ \frac1{\eta(\mathsf{q})^{k-1}} \
\sum_{\vec{u}\in\mathfrak{U}_j^{\mathrm{conf}}}\,
\mathsf{q}^{\frac{1}{2}\, \vec{u}\cdot C^{-1}\vec{u}} \ \vec{\xi}^{\
C^{-1}\vec{u}}= \frac{\mathsf{q}^{\frac{1}{2}\, \langle
\omega_j,\omega_j\rangle_{\mathfrak{Q}\otimes_{\mathbb{Z}} {\mathbb{Q}}}}}{\eta(\mathsf{q})^{k-1}} \
\sum_{\vec{u}\in\mathfrak{U}_j^{\mathrm{conf}}}\, \vec{\xi}^{\ C^{-1}\vec{u}}\ .
\end{align}
Then the instanton partition function is given by the factorization
\begin{equation}
{\mathcal Z}_{X_k}^{A_0}\big(\varepsilon_1,\varepsilon_2, \mu_0,\mu_1; \mathsf{q},
\vec{\xi} \ \big)_j
=\eta(\mathsf{q})^{k-1} \ \chi_{\mathrm{conf}}^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \
\prod_{i=1}^k \,
{\mathcal Z}_{{\mathbb{C}}^2}^{A_0}\big(\varepsilon_1^{(i)},\varepsilon_2^{(i)},
\mu_0,\mu_1 ; \mathsf{q} \big) \ .
\end{equation}
In this instance Proposition \ref{prop:ZcalXkArcorr} factorizes under the decomposition of Theorem \ref{thm:virasoroprimary} as
\begin{multline}
{\mathcal Z}_{X_k}^{A_0} \big(\varepsilon_1,\varepsilon_2, \mu_1,\mu_2; \mathsf{q},
\vec{\xi} \ \big)_j \\
=\eta(\mathsf{q})^{k-1}\ \chi_{\mathrm{conf}}^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \ \Big\langle \vert 0\rangle\,,\, \operatorname{V}_{-\frac{\mu_0}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu_0+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}}(1) \ \operatorname{V}_{-\frac{\mu_1}{\sqrt{-k\, \varepsilon_1\, \varepsilon
_2}},\frac{\mu_1+\varepsilon_1+\varepsilon_2}{\sqrt{-k\, \varepsilon_1\,
\varepsilon_2}}}(\mathsf{q}) \vert 0\rangle\Big\rangle_{\mathbb{H}}\ .
\end{multline}
By Equation \eqref{eq:fundamental} we then obtain explicitly
\begin{equation}
{\mathcal Z}_{X_k}^{A_0}\big(\varepsilon_1,\varepsilon_2, \mu_0,\mu_1; \mathsf{q},
\vec{\xi} \ \big)_j
=\eta(\mathsf{q})^{k-1} \ \chi_{\mathrm{conf}}^{\widehat{\omega}_j}\big(\mathsf{q},\vec\zeta \ \big) \
(1-\mathsf{q})^{-\frac{\mu_1\,(\mu_0+\varepsilon_1+\varepsilon_2)}{k\, \varepsilon_1\, \varepsilon_2}} \ .
\end{equation}
|
\section{INTRODUCTION}
The Standard Model (SM) of particle physics has given us enough reasons to look beyond its framework for dealing
with issues like tiny neutrino masses, matter-antimatter asymmetry of the present universe, Dark matter and Dark
energy, coupling unification of three fundamental interactions. Among all these, the observed baryon asymmetry
of the universe has motivated the scientific community to work upon it since a long time. The WMAP satellite data
\cite{Dunkley:2008,Komatsu:2011}, when combined with large scale structures (LSS) data, gives the baryon asymmetry
of the universe to be $\eta^{\mbox{\small CMB}} \simeq (6.3\pm 0.3) \times 10^{-10} $ while an independent measurement
of baryon asymmetry carried out by BBN \cite{Yao:2006} yields $\eta^{\mbox{\small BBN}} \simeq (3.4 -6.9) \times 10^{-10}$.
Two compelling mechanisms namely Leptogenesis \cite{Fukugita:1986} and Weak scale baryogenesis \cite{EWbary} have been
prime tools for explaining baryon asymmetry of the universe. In leptogenesis the desired lepton asymmetry is created
by the lepton number violating as well as out of equilibrium decays of heavy particles which is subsequently converted
into baryon asymmetry by the non-perturbative ($B+L$)-violating sphaleron interactions \cite{Buchmuller:2005eh,Davidson:2008bu}.
An inadequate knowledge about the nature of new physics beyond the standard model
leaves us with no choice but to explore all possibilities which may explain the origin of matter-antimatter asymmetry.
Recently a new idea behind baryon asymmetry has been explored named "Post-Sphaleron baryogenesis (PSB)" which
occurs via the decay of a scalar boson singlet under standard model having mass around few hundreds of GeV and
a high dimensional baryon number violating coupling \cite{Babu:2006xc, Babu:2012vc,Babu:2013yca}, where the
Yukawa coupling(s) of the scalar(s) act as the
source of CP-asymmetry. Apparently,this high dimensional baryon number violating coupling is generated via new physics
operative beyond standard model electroweak theory. The mechanism of PSB is based on the idea that
the required amount of baryon asymmetry of the universe can be generated below the scale of electroweak phase transition
where the sphaleron has decoupled from the Hubble expansion rate. Although the proposal seems interesting it has
not yet been incorporated in a realistic grand unified theory. Hence we attempt here to embed the proposal of PSB
in a non-SUSY SO(10) GUT with Pati-Salam (PS) symmetry and Left-Right (LR) symmetry as intermediate symmetry breaking steps.
A detail study of the literatures \cite{Mohapatra:1974gc,Pati:1974yy,Senjanovic:1975rk,
Mohapatra:1980yp,Mohapatra:1979ia,Deshpande:1990ip,Lazarides:1980nt,Dev:2013oxa}
gives an idea about many intriguing features of the $SO(10)$
grand unified theory (including both non-SUSY and SUSY). One of these features is that when left-right gauge symmetry
appears as an intermediate symmetry breaking step in a novel symmetry breaking chain, then seesaw mechanism can be naturally
incorporated into it. In conventional seesaw models associated with thermal leptogenesis the mass scale for heavy
RH Majorana neutrino is at $10^{10}$ GeV which makes it unsuitable for direct detectability at current
accelerator experiments like LHC. Therefore, it is necessary to construct a theory having
$SU(2)_L \times SU(2)_{R} \times U(1)_{B-L} \times SU(3)_C$ and $SU(2)_L \times SU(2)_{R} \times SU(4)_C$
gauge groups as intermediate symmetry breaking steps which results in low mass right-handed Majorana neutrinos along with
$W_R$, $Z^\prime$ gauge bosons at TeV scale. At the same time it should be capable of explaining post-sphaleron
baryogenesis elegantly along with other derivable predictions like proton decay and neutron-antineutron oscillation.
We intend to discuss TeV scale post-sphaleron baryogenesis, neutron-antineutron oscillation having mixing time close
to the experimental limit with the Pati-Salam symmetry or $SO(10)$ GUT as mentioned in a recent work \cite{Awasthi:2013ff}
slightly modifying the Higgs content where non-zero light neutrino masses can be accommodated via gauged extended
inverse seesaw mechanism along with TeV scale $W_R$, $Z^\prime$ gauge bosons. As discussed in the work \cite{Awasthi:2013ff}
the Dirac neutrino mass matrix is similar to the up-quark mass matrix even with low scale right-handed symmetry breaking.
Though the details has been already discussed in the above mentioned work we breifly clarify the point as follows.
In non-SUSY $SO(10)$, the type I seesaw \cite{typeI} contribution to neutrino mass is given by
$$m^{I}_\nu = - M_D M^{-1}_R M^T_D\,, $$
where $M_D$ is the Dirac neutrino mass matrix, $M_R$ is the Majorana neutrino mass matrix for right-handed
neutrinos and is related to the right-handed symmetry breaking scale. The Dirac neutrino mass matrix
and up-quark mass matrices are similar in a generic $SO(10)$ model that has high scale Pati-Salam symmetry
as an intermediate breaking step relating quarks and leptons with each other. Hence, $M_D \simeq M_u$, which further
implies that the $\tau-$ neutrino Dirac Yukawa coupling should be equal to top-quark Yukawa coupling.
With $M_D\simeq M_u \simeq 100$ GeV, the sub-eV scale of light neutrino consistent with oscillation data
requires the right-handed scale (seesaw scale) to be greater than $10^{13}$ GeV. Such high seesaw scale
makes this idea difficult to be probed at any foreseeable laboratory experiments.
Hence, as an alternative way, emphasizing on its verifiability at LHC, inverse seesaw
mechanism \cite{inv,Bdev-non} has been proposed, with an extra $SO(10)$ fermion singlet $S$ (in addition to the
existing fermion content of $SO(10)$), with light neutrino mass formula
$$m_\nu =\left(\frac{M_D}{M}\right) \mu \left(\frac{M_D}{M}\right)^T\, ,$$
where $M$ is the $N-S$ mixing matrix and $\mu$ is the small lepton number violating mass term for
sterile neutrino $S$. The above relation can be recasted as
$$\left( \frac{m_\nu}{\mbox{0.1\, eV}}\right) = \left(\frac{M_D}{\mbox{100\, GeV}} \right)^2
\left(\frac{\mu}{\mbox{keV}}\right) \left(\frac{M}{10^4\, \mbox{GeV}} \right)^{-2}\,.$$
Hence, sub-eV mass for light neutrinos are consistent with $M_D \simeq M_u$ (or, $Y_D \simeq Y_t$) which
is a generic predictions of high scale Pati-Salam symmetry and compatible with low right-handed
symmetry breaking scale ($M_R$) since inverse seesaw formula is independent of $M_R$.
We have utilised this particular property of low scale right-handed symmetry breaking in studying
Post-sphaleron baryogenesis and neutron-antineutron oscillation even though a complete discussion
on the origin of neutrino masses and mixing via low sacle extended inverse seesaw has been omitted.
Here we sketch out the complete work of our paper. In Sec.II, we briefly discuss
non-SUSY $SO(10)$ GUT with a novel symmetry breaking chain, having $\mathcal{G}_{2213}$ and $\mathcal{G}_{224}$
as intermediate symmetry breaking steps. In Sec.III we show how gauge coupling unification is achieved in our model.
In Sec.IV we discuss the TeV scale post-sphaleron baryogenesis and embed it within the novel chain of non-SUSY $SO(10)$
model with the self-consistent model parameters. In Sec.V, we estimate the mixing
time for neutron-antineutron oscillation. In Sec.VI, we present an idea how low mass scales for RH Majorana neutrino
as well as right-handed gauge bosons $W_R$, $Z^\prime$ are allowed in the model, while explaining light neutrino masses
via gauged extended seesaw mechanism. In Sec.VII we conclude our work with results and summary including a note on
viability of the model at LHC.
\section{THE MODEL}
In this section we shall discuss the one-loop gauge coupling unification and estimate the proton life time including
short distance enhancement factor to the $d=6$ proton decay operator by reviving the symmetry breaking chain \cite{Awasthi:2013ff}
{\small
\begin{eqnarray}
SO(10)
& &\stackrel{M_U}{\longrightarrow}SU(2)_L \times SU(2)_{R} \times SU(4)_C \times D \quad
\left[\mathcal{G}_{224D}, \, \, (g_{2L} = g_{2R}) \right]\nonumber \\
& &\hspace*{-1.5cm} \mathop{\longrightarrow}^{M_P}_{} SU(2)_L \times SU(2)_{R} \times SU(4)_C
\quad \left[\mathcal{G}_{224}, \, \, (g_{2L} \neq g_{2R}) \right]\nonumber \\
& &\hspace*{-1.5cm} \mathop{\longrightarrow}^{M_C}_{} SU(2)_L \times SU(2)_{R} \times U(1)_{B-L} \times SU(3)_C
\quad \left[\mathcal{G}_{2213} \, \right]\nonumber \\
& &\hspace*{-1.5cm} \mathop{\longrightarrow}^{M_\Omega}_{} SU(2)_L \times U(1)_{R} \times U(1)_{B-L} \times SU(3)_C
\quad \left[\mathcal{G}_{2113} \, \right]\nonumber \\
& &\hspace*{-1.5cm} \mathop{\longrightarrow}^{M_{B-L}}_{} SU(2)_L \times U(1)_{Y} \times SU(3)_C
\quad \left[\mathcal{G}_{\rm SM} \equiv \mathcal{G}_{\rm 213}\right] \nonumber \\
& &\hspace*{-1.5cm} \mathop{\longrightarrow}^{M_Z}_{}~U(1)_{\rm em}\times SU(3)_C \quad \quad \left[\mathcal{G}_{\rm 13}\right]\, .
\label{chain}
\end{eqnarray}
}
The chain breaks in a sequence, where $SO(10)$ first breaks down to $\mathcal{G}_{224D}, \, \, (g_{2L} = g_{2R})$ after
the Higgs representation $\langle (1,1,1) \rangle \subset \{54\}_H$ is given a VEV, then the spontaneous breakdown of
D-parity occurs in $\mathcal{G}_{224D}, \, \, (g_{2L} = g_{2R}) \to \mathcal{G}_{224}, \, (g_{2L} \neq g_{2R})$ with the
assignment of VEV to D-parity odd component $\langle(1,1,1) \rangle$ contained in the Higgs representation
$\{210\}_H$. The decomposition of $\{210\}_H$ under $\mathcal{G}_{224}$ is
\begin{eqnarray}
\{210\}_H&=&(1,1,1) \oplus (2,2,20) \oplus (3,1,15) \oplus (1,3,15) \nonumber \\
& &\quad \oplus (2,2,6) \oplus (1,1,15)\, .
\end{eqnarray}
Spontaneous D-parity mechanism is aptly utilized here, since the theory allows low mass scale for right-handed
Higgs fields around $\mathcal{O}$(TeV) while keeping all its left-handed components at D-parity breaking scale. Now assigning
a VEV to the neutral component $\langle(1,1,15)\rangle \subset \{210\}_H$, the Pati-Salam symmetry ($\mathcal{G}_{224}$)
breaks down to left-right symmetry ($\mathcal{G}_{2213}$). The next step of symmetry breaking $\mathcal{G}_{2213} \to
\mathcal{G}_{2113}$ occurs via the VEV $\langle(1,3,0,1)\rangle \subset \{210\}_H$. The right-handed gauge
boson $W_R$ acquires a mass in the range of few TeV and contributes sub-dominantly to neutrinoless
double beta decay.
The most desirable symmetry breaking step $\mathcal{G}_{2113} \to \mathcal{G}_{213}$ is achieved by the $\{126\}_H$ of $SO(10)$
though we have added another Higgs representation $\{16\}_H$ for realization of gauged inverse seesaw mechanism operative
at TeV scale. The decomposition of the Higgs $\{126\}_H$ under $\mathcal{G}_{224}$ is
\begin{eqnarray}
\{126\}_H&=&(3,1,10) \oplus (1,3,\overline{10}) \oplus (2,2,15) \oplus (1,1,6) \nonumber \\
\end{eqnarray}
As we have pointed earlier, due to D-parity mechanism, the right-handed triplet Higgs field
$\Delta_R (1,3,-2,1)$ contained in $(3,1,10)$ gets its mass at TeV scale while its left-handed partner
$\Delta_L (3,1,-2,1)$ has its mass at D-parity breaking scale $M_P$. As a result of this symmetry breaking,
the neutral component of right-handed gauge boson $Z^\prime$ gets its mass around $\mathcal{O}$(TeV) with
the experimental bound $M_{W_R} \geq 2.5$~TeV \cite{CMS:2012zv,ATLAS:2012ak}. The final stage of symmetry breaking $\mathcal{G}_{2113} \to \mathcal{G}_{213}$
is carried out by giving VEV to the neutral component of SM Higgs doublet $\langle \phi^0(2,1/2,1) \rangle$
contained in the bidoublet $\Phi \subset \{10\}_H$.
We shall now check whether $SO(10)$ having TeV scale post-sphaleron baryogenesis, neutron-antineutron oscillation
and gauged inverse seesaw mechanism is consistent with gauge coupling unification. It is found that the coupling constants
unify at ($10^{17}-10^{18.5}$) GeV with the Higgs fields {\bf $\{10\}_H$+$\{10\}_H^\prime$ + $\{16\}_H$ + $\{126\}_H$
+$\{210\}_H$}. Some good reasons behind taking these Higgs fields are; firstly, the TeV scale post-sphaleron baryogenesis
and neutron-antineutron oscillation can be well explained with these parameters while predicting $W_R$ gauge boson
in TeV range; secondly, it allows $B-L$ breaking ($M_{B-L}$) at TeV scale resulting
$Z^\prime$ mass $\geq 1.6$ TeV, moreover it explains tiny masses for light neutrinos consistent
with neutrino oscillation data via TeV scale gauged inverse seesaw mechanism and LFV decays with branching ratios
accessible to ongoing search experiments.
\section{GAUGE COUPLING UNIFICATION {\small AND} PROTON DECAY}
\label{sec:rge}
\subsection{One-loop renormalization group equations (RGEs) for gauge coupling evolution}
\label{subsec1:rge}
For simplicity, we consider only the one-loop renormalization group equations(RGEs) for gauge coupling evolution
which can be written as
\begin{eqnarray}
\mu\,\frac{d\,g_{i}}{d\,\mu}=\frac{a_i}{16 \pi^2} g^{3}_{i}\, \quad {\large \Longrightarrow}\, \quad
\frac{d\, \alpha^{-1}_{i}}{d\,t}=\frac{\pmb{a_i}}{2 \pi}
\end{eqnarray}
where, $t=\ln(\mu)$, $\alpha_{i}=g^2_{i}/(4 \pi)$ is the fine structure constant, and $\pmb{a_i}$ is the one-loop beta coefficients derived for the the corresponding
$i^{\rm th}$ gauge group for which coupling evolution has to be determined. Using the input parameters,
electroweak mixing angle $\sin^2\theta_W(M_Z)=0.2312$, electromagnetic coupling constant $\alpha(M_Z) = 127.9$
and strong coupling constant $\alpha_{S}(M_Z) = 0.1187$ taken from PDG \cite{Yao:2006,pdg}
the values of three coupling constants at electroweak scale $M_Z=91.187$ GeV can be calculated precisely to be,
\begin{equation}
\left(\begin{array}{cc}
\ \alpha_{2L}(M_Z) \\
\ \alpha_{1Y}(M_Z) \\
\ \alpha_{3C} (M_Z)
\end{array}\right)
= \left(\begin{array}{cc}
\ 0.033493^{+0.000042}_{-0.000038} \\
\ 0.016829 \pm 0.000017 \\
\ 0.118 \pm 0.003
\end{array}\right)\, ,
\label{pdg:alphas}
\end{equation}
where $\{\alpha_{2L}(M_Z), \alpha_{1Y}(M_Z), \alpha_{3C}(M_Z) \}$ denote the fine structure constants for the
SM gauge group $\mathcal{G}_{213}=SU(2)_L \times U(1)_Y \times SU(3)_C$.
\subsection{Higgs content for the model and corresponding one-loop beta coefficients $\pmb{a_i}$}
\label{subsec2:rge}
The Higgs contents for the model used in different ranges of mass scales under respective gauge symmetries ($\mathcal{G}_I$)
with a particular symmetry breaking chain as considered in a recent work \cite{Awasthi:2013ff} where the prime interest was to keep
the $W_R$, $Z_R$ gauge bosons at TeV scale are as follows,
\begin{eqnarray}
& &\hspace*{-1.0cm} {\bf \mbox{(i)}\, \mu=M_Z - M_{B-L} }: G={\rm SM} = G_{213},
\hspace*{0.2cm} \mbox{Higgs:\,}\Phi (2,1/2,1)\, ; \nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(ii)}\, \mu=M_{B-L} - M_\Omega}: G= G_{2113}, \nonumber \\
& & \hspace*{0.0cm} \mbox{Higgs:\,} \Phi_1 (2,1/2,0,1) \oplus \Phi_2 (2,-1/2,0,1) \oplus \chi_R (1,1/2,-1,1) \oplus \Delta_R (1,1,-2,1) \, ; \nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(iii)}\, \mu=M_\Omega - M_C}: G= G_{2213},
\nonumber \\
& & \hspace*{0.0cm} \mbox{Higgs:\,}\Phi_1 (2,2,0,1) \oplus \Phi_2 (2,2,0,1) \oplus
\chi_R (1,2,-1,1)\oplus \Delta_R (1,3,-2,1)\oplus \Omega_R(1,3,0,1) \nonumber \\
\label{higgs-a}
\end{eqnarray}
\begin{eqnarray}
& &\hspace*{-1.0cm} {\bf \mbox{(iv)}\, \mu=M_C - M_{\xi}}: G= G_{224}, \nonumber \\
& &\hspace*{0.2cm} \mbox{Higgs:\,}\Phi_1(2,2,1)_{10}\oplus \Phi_2(2,2,1)_{10^\prime} \oplus \Delta_R(1,3,\overline{10})_{126}
\oplus \chi_R(1,2,\overline{4})_{16} \nonumber \\
& & \hspace*{1.2cm}\oplus \Omega_R(1,3,15)_{210} \oplus \Sigma(1,1,15)_{210} \nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(v)}\, \mu=M_\xi - M_{P}}: G= G^\prime_{224}, \nonumber \\
& &\hspace*{0.2cm} \mbox{Higgs:\,}\Phi_1(2,2,1)_{10}\oplus \Phi_2(2,2,1)_{10^\prime} \oplus \Delta_R(1,3,\overline{10})_{126}
\oplus \chi_R(1,2,\overline{4})_{16} \nonumber \\
& & \hspace*{1.2cm}\oplus \Omega_R(1,3,15)_{210} \oplus \Sigma(1,1,15)_{210} \oplus \xi(2,2,15)_{126} \nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(vi)}\, \mu=M_P - M_U}: G= G_{224D}, \nonumber \\
& &\hspace*{0.2cm} \mbox{Higgs:\,}
\Phi_1(2,2,1)_{10}\oplus \Phi_2(2,2,1)_{10^\prime} \oplus \Delta_L(3,1,10)_{126} \oplus \Delta_R(1,3,\overline{10})_{126} \nonumber \\
& &\hspace*{1.2cm} \oplus \chi_L(2,1,4)_{16} \oplus \chi_R(1,2,\overline{4})_{16} \oplus \Omega_L(3,1,15)_{210} \oplus \Omega_R(1,3,15)_{210}\nonumber \\
& &\hspace*{1.2cm}\oplus \Sigma(1,1,15)_{210} \oplus \xi(2,2,15)_{126} \oplus \sigma(1,1,1)_{210}\, .\nonumber \\
\label{higgs-b}
\end{eqnarray}
Here we find two categories of Higgs spectrum; {\bf Model-I\,} having Higgs spectrum as given in eqn.(\,\ref{higgs-a})
and eqn.(\,\ref{higgs-b}) excluding the bitriplet Higgs scalar which estimates a proton life time that is far from the reach of
search experiments and {\bf Model-II\,} having the same Higgs spectrum, including the bitriplet Higgs scalar
$(3,3,1) \subset \mathcal{G}_{224}$ from mass scale $M_C$ onwards which estimates a proton life time very close to the
experimental limit. Thus {\bf Model-II\,} serves our purpose.
The one-loop beta coefficients are found to be the same for both the models at mass scale ranges $M_Z - M_{B-L}$,
$M_{B-L}-M_\Omega$, and $M_\Omega - M_{C}$ i.e.,
\begin{eqnarray}
& &\hspace*{-1.0cm} {\bf \mbox{(i)}\, \mu=M_Z - M_{B-L} }: G={\rm SM} = G_{2_L 1_Y 3_C}, \quad \pmb{a_i}=\left(-19/6,\, 41/10,\, -7\right)
\nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(ii)}\, \mu=M_{B-L} - M_\Omega}: G= G_{2_L 1_{R} 1_{(B-L)} 3_C},
\quad \pmb{a_i}=\left(-3,\, 19/4,\, 37/8,\, -7\right)
\nonumber \\
& &\hspace*{-1.0cm} {\bf \mbox{(iii)}\, \mu=M_\Omega - M_C}: G= G_{2_L 2_{R} 1_{(B-L)} 3_C},
\quad \pmb{a_i}=\left(-8/3,\, -2/3,\, 23/4,\, -7\right)\, ,
\end{eqnarray}
whereas, they differ at Pati-Salam scale $M_C$ to the Unification scale $M_U$ as shown in Table.\ref{tab:beta_coeff}.
\begin{table*}[htb]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
&&& \\[-4mm]
$G_{I}$ & {\bf Mass ranges} & $\pmb{a_i}$ for {\bf Model-I} & $\pmb{a_i}$ for {\bf Model-II} \\[4mm]
\hline \hline
&&&\\[-4mm]
${\small G_{2_L2_R4_C}}$
& {\bf $M_C-M_{\xi}$}
&
$\begin{pmatrix} -8/3 \\
29/3\\
-14/3 \end{pmatrix}$
&
$\begin{pmatrix} -2/3 \\
35/2\\
-14/3 \end{pmatrix}$\\[7mm]
\hline
&&&\\[-4mm]
${\small G_{2_L2_R4_C}}$
&{\bf $M_{\xi}-M_P$}
&
$\begin{pmatrix} 7/3 \\
44/3\\
2/3 \end{pmatrix}$
&
$\begin{pmatrix} -12/3 \\
35/3\\
-14/3 \end{pmatrix}$\\[7mm]
\hline
&&&\\[-4mm]
${\small G_{2_L2_R4_C D}}$
&{\bf $M_P-M_{U}$}
&
$\begin{pmatrix} 44/3 \\
44/3 \\
6 \end{pmatrix}$ &
$\begin{pmatrix} 35/3 \\
35/3 \\
2/3 \end{pmatrix}$\\[7mm]
\hline
\end{tabular}
\caption{One-loop beta coefficients for different gauge coupling evolutions, without Bitriplet Higgs scalar in {\bf Model-I} and with
a Bitriplet Higgs scalar (3,3,1) under the Pati-Salam group $SU(2)_L \times SU(2)_R \times SU(4)_C$ in {\bf Model-II}.}
\label{tab:beta_coeff}
\end{table*}
\begin{figure*}[htb!]
\begin{minipage}[t]{0.49\textwidth}
\hspace{-0.4cm}
\begin{center}
\includegraphics[scale=0.85,angle=0]{coup_unif_wobitriplet.eps}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[t]{0.49\textwidth}
\hspace{-0.4cm}
\begin{center}
\includegraphics[scale=0.85,angle=0]{coup_unif_bitriplet.eps}
\end{center}
\end{minipage}
\caption{Gauge coupling evolution plot having TeV scale $W_R$, $Z_R$ bosons where $M_U=2.65\times 10^{15.8}$ GeV}
\label{lightneutrino_general}
\end{figure*}
The gauge coupling unification for this work is shown in Fig.\,\ref{lightneutrino_general}
with the allowed mass scales desirable for our model predictions,
\begin{eqnarray}
& &M_{B-L}= 4-7~{\rm TeV},\, M_{\Omega} = 10~{\rm TeV},\, M_C = 10^{5}-10^{6}\, \mbox{GeV}\, ,\nonumber \\
& &M_P\simeq 10^{15.65}~{\rm GeV\,\, and\,\,} M_{\rm G}\simeq 10^{18.65}~{\rm GeV} \, .
\end{eqnarray}
\subsection{Estimation of Proton life time for $p \to \pi^0\, e^+$}
\label{subsec3:proton}
The decay rate for the gauge boson mediated proton decay in the channel $p \to \pi^0\, e^+$ including strong and electroweak
renormalization effects on the ${\rm d}=6$ operator starting from the GUT scale to the proton mass (i.e, 1 GeV) \cite{Babu:1992ia, Bertolini:2013vta}
comes out to be
\begin{eqnarray}
\Gamma\left( p\rightarrow \pi^0 e^+ \right) &=&\frac{\pi}{4}\, A^2_L\, \frac{|\overline{\alpha}_H|^2}{f^2_\pi}
\frac{m_p\, \alpha^2_U}{M^4_U} \left(1 + \mathcal{F} + \mathcal{D} \right)^2 \mathcal{R}\, .
\label{decay-width-proton}
\end{eqnarray}
In the eq.\,(\ref{decay-width-proton}), $A_L=1.25$ is renormalization factor from the electroweak scale to the proton mass, $\mathcal{D}=0.81$,
$\mathcal{F}=0.44$, $\overline{\alpha}_H=-0.011\, \, \mbox{GeV}^3$, and $f_\pi=139\, \, \mbox{MeV}$ which have
been extracted as phenomenological parameters by the chiral perturbation theory and lattice gauge theory.
Also $m_p=938.3\, \, \mbox{MeV}$ is the proton mass, and $\alpha_U \equiv \alpha_G$ is the gauge fine structure constant derived
at the GUT scale. It is worth to note here that the renormalization factor $\mathcal{R}=\left[\left(A_{SR}^2+A_{SL}^2\right)
\left(1+ |{V_{ud}}|^2\right)^2\right]$ for $SO(10)$, $V_{ud}=0.974=$ with $A_{SL}(A_{SR})$ being the short-distance
renormalization factor in the left (right) sectors, and $V_{ud}$ is the $(1,1)$ element of $V_{CKM}$ for quark mixings.
After re-expressing $\alpha_H = \overline{\alpha}_H \left(1 + \mathcal{F} + \mathcal{D} \right) = 0.012\, \, \mbox{GeV}^3$, and
$\mathcal{A}_R\simeq \mathcal{A}_{L} \mathcal{A}_{SL} \simeq \mathcal{A}_{L} \mathcal{A}_{SR}$, the proton life time can
be expressed as
\begin{eqnarray}
\tau_p = \Gamma^{-1}\left( p\rightarrow \pi^0 e^+ \right) &=&\frac{4}{\pi}\,\frac{f^2_\pi}{m_p}\frac{M^4_U}{\alpha^2_U}
\frac{1}{\alpha^2_H \mathcal{A}^2_{R}} \frac{1}{\mathcal{F}_q} \, ,
\label{lifetime-proton}
\end{eqnarray}
where $\mathcal{F}_q \simeq 7.6$.
\noindent
{\bf Short distance enhancement factor $\mathcal{A}_{SL}$ extrapolated from GUT scale to $1$ GeV:}
For estimating proton decay rate in the channel $p \to e^+\pi^0$ having dimension-6 operator, one needs to extrapolate
the operator from the GUT scale physics to the low energy physics at the scale of $m_p=1\, \mbox{GeV}$ \cite{Ibanez:1984ni, Buras:1977yy,
BhupalDev:2010he}. With the particular symmetry breaking chain allowed in the non-SUSY $SO(10)$ model (following the ref.\, \cite{BhupalDev:2010he}),
the whole energy range can be separated into following parts
\begin{enumerate}
\item[${\bf \tiny I.}$] from non-SUSY $S0(10)$ GUT scale,$M_U$, to the Pati-Salam symmetry with D-parity ($\mathcal{G}_{224D}$,
$g_{2L}=g_{2R}$) invariance scale, $M_P$\, ,
\item[${\bf \tiny II.}$] from $M_P$ to the Pati-Salam symmetry without D-parity ($\mathcal{G}_{224}$, $g_{2L} \neq g_{2R}$)
scale $M_C$\, ,
\item[${\bf \tiny III.}$] from $M_C$ to $SU(4)_C$ breaking scale, $M_\Omega$, where we have left-right symmetric model
({\bf LRSM}) $\mathcal{G}_{2213}$\, ,
\item[${\bf \tiny IV.}$] from left-right symmetry breaking scale ($M_\Omega$) to $\mathcal{G}_{2113}$ scale ($M_{B-L}$)\,
\item[${\bf \tiny V.}$] from $\mathcal{G}_{2113}$ scale ($M_{B-L}$) to standard model $\mathcal{G}_{213}$\, ,
\item[${\bf \tiny V.}$] from standard model to $1\, \mbox{GeV}$\,.
\end{enumerate}
As discussed in refs.\, \cite{Ibanez:1984ni, Buras:1977yy, BhupalDev:2010he}, the enhancement factor below SM
for the $LLLL$ operator is
$$\mathcal{A}^{\prime}_{L} = \bigg[\frac{\alpha_s (\mbox{1 GeV})}{\alpha_s (m_t)}\bigg]
^{-\frac{4}{2 \cdot \left(-11+\frac{2}{3} \, n_f \right)}}\, ,$$
where, $n_f$ denotes the number of quark flavors at the particular energy scale of our interest.
Neglecting the effect due to $\alpha_{2L}$ and $\alpha_{Y}$ since their contributions are suppressed
as compared to the strong coupling effect $\alpha_s$, this enhancement factor can be expressed
in a more explicit manner as
\begin{eqnarray}
\mathcal{A}^{\prime}_{L}=\bigg[\frac{\alpha_s (\mbox{1 GeV})}{\alpha_s (m_c)}\bigg]^{2/9}
\bigg[\frac{\alpha_s (m_c)}{\alpha_s (m_b)}\bigg]^{6/25}
\bigg[\frac{\alpha_s (m_b)}{\alpha_s (m_t)}\bigg]^{6/23}\, .
\end{eqnarray}
Since the model considered here is non-supersymmetric version of $SO(10)$ GUT, all other enhancement factors
can be written in the same way as
\begin{eqnarray}
& &\mathcal{A}^{\rm SM}_{SL}=\bigg[\frac{\alpha_i (m_t)}{\alpha_i (M^0_R)}\bigg]
^{\frac{- \gamma_i}{2\, {\Large \pmb a_i}}}\, ,
\end{eqnarray}
with $\gamma_i$ (${\bf \large \pmb a_i}$) as the anomalous dimension (one-loop beta coefficients)
for the corresponding gauge group $i=SU(2)_{L},\, U(1)_{Y},\, SU(3)_{C}$. Similarly, one can write
the enhancement factor valid for $\mathcal{G}_{2113}$, $\mathcal{G}_{2213}$, $\mathcal{G}_{224}$,
and $\mathcal{G}_{224D}$ as
\begin{eqnarray}
& &\mathcal{A}^{2113}_{SL}=\bigg[\frac{\alpha_i (M^0_R)}{\alpha_i (M^+_R)}\bigg]
^{\frac{- \gamma_i}{2\, {\Large \pmb a_i}}}\, , \mbox{with}\, i=SU(2)_{L},\,U(1)_R,\, U(1)_{B-L},\, SU(3)_{C}\, ,
\nonumber \\
& &\mathcal{A}^{2213}_{SL}=\bigg[\frac{\alpha_i (M^+_R)}{\alpha_i (M_C)}\bigg]
^{\frac{- \gamma_i}{2\, {\Large \pmb a_i}}}\, , \mbox{with}\, i=SU(2)_{L},\,SU(2)_R,\, U(1)_{B-L},\, SU(3)_{C}\, ,
\nonumber \\
& &\mathcal{A}^{224}_{SL}=\bigg[\frac{\alpha_i (M_C)}{\alpha_i (M_P)}\bigg]
^{\frac{- \gamma_i}{2\, {\Large \pmb a_i}}}\, , \mbox{with}\, i=SU(2)_{L},\,SU(2)_R,\, SU(4)_{C}\, ,
\nonumber \\
& &\mathcal{A}^{224D}_{SL}=\bigg[\frac{\alpha_i (M_P)}{\alpha_i (M_U)}\bigg]
^{\frac{- \gamma_i}{2\, {\Large \pmb a_i}}}\, , \mbox{with}\, i=SU(2)_{L},\,SU(2)_R,\, SU(4)_{C}\, \mbox{with D-parity}\, .
\nonumber
\end{eqnarray}
Hence, the complete short distance enhancement renormalization factor for this $d=6$ proton decay operator is found to be
\begin{equation}
\mathcal{A}_{SL}=\mathcal{A}^{\rm SM}_{SL} \cdot \mathcal{A}^{2113}_{SL} \cdot \mathcal{A}^{2213}_{SL}
\cdot \mathcal{A}^{224}_{SL} \cdot \mathcal{A}^{224D}_{SL} \,.
\end{equation}
We have earnestly followed the prescription given in ref.\cite{Ibanez:1984ni, Buras:1977yy} for the derivation of anomalous dimension for the
effective $d=6 (LLLL)$ proton decay operator. With a choice of TeV scale particle spectrum used in our model, the unification
scale is found to be $M_U=2.65\times 10^{18.5}$ GeV for {\bf Model-I} and $M_U=10^{15.8}$ GeV for {\bf Model-II}. We have estimated
the factor $\mathcal{A}_{R}=\mathcal{A}_L \cdot \mathcal{A}_{SL}$, approximately, to be $4.36$ with the value of long distance
renormalization factor $A_L=1.25$ which is the same for both the models.
With these input parameters, the model under consideration predicts the proton life time to be $$\tau (p \to e^+ \pi^0) = 2.6
\times 10^{34}\, \mbox{yrs} $$ that is closer to the latest Super-Kamiokande experimental bound \cite{Nishino:2012ipa,babuetal}
\begin{eqnarray}\tau (p \to e^+ \pi^0) \big|_{SK, 2011} > 8.2 \times 10^{33}\,
\mbox{yrs}\, ,
\end{eqnarray}
and ably supports planned experiments that can reach a bound \cite{hyperk}
\begin{eqnarray}
& &\tau (p \to e^+ \pi^0) \big|_{HK, 2025} > 9.0 \times 10^{34}\, \mbox{yrs} \nonumber \\
& &\tau (p \to e^+ \pi^0) \big|_{HK, 2040} > 2.0 \times 10^{35}\, \mbox{yrs} \nonumber \\
\end{eqnarray}
\section{TEV SCALE POST-SPHALERON BARYOGENESIS}
\label{sec:psb}
\subsection{Basic interaction terms}
\label{subsec1:psb}
As already discussed in Sec.\,\ref{sec:rge}, Pati-Salam symmetry survives till few $100$ TeV scale playing
an important role in the explanation of baryogenesis mechanism and neutron-antineutron oscillation. We need to know
all the basic interactions using quarks and di-quarks under high scale Pati-Salam symmetry as well as under
low scale SM like interactions around TeV scale in order to explain the above said phenomena successfully. For that,
we take a look at the decomposition of the Pati-Salam Higgs representation
$\Delta_R(1,3, \overline{10})$ under left-right symmetry group $SU(2)_L \times SU(2)_R \times U(1)_{B-L} \times
SU(3)_C$ and the SM gauge group $SU(2)_L \times U(1)_{Y} \times SU(3)_C$
\begin{eqnarray}
\Delta(1, 3, \overline{10}) &=& \{ \Delta_{\ell \ell} (1, 3, -2, 1) \oplus \Delta_{q \ell} (1, 3, -2/3, 3^*)
\oplus \Delta_{qq} (1, 3, 2/3, 6^*) \nonumber \\ & &
\hspace*{6cm}\mbox{under}\, \,\mathcal{G}_{2_L 2_R 1_{B-L} 3_C}\, , \\
&\supset& \Delta_{\nu \nu} (1,0,1) \oplus \Delta_{\nu e} (1,1,1) \oplus \Delta_{ee} (1,2,1)
\nonumber \\ &\oplus&
\Delta_{u \nu} (1,-2/3,3^*) \oplus \Delta_{d e} (1,1/3,3^*) \oplus \Delta_{u e} (1,1/3,3^*)
\oplus \Delta_{d \nu} (1,1/3,3^*) \nonumber \\
&\oplus& \Delta_{u u} (1, -4/3,6^*) \oplus \Delta_{u d} (1,-1/3,6^*) \oplus
\Delta_{d d} (1,2/3,6^*) \, \nonumber \\ & &
\hspace*{6cm}
\mbox{under}\quad \mathcal{G}_{2_L 1_Y 3_C}\, ,
\end{eqnarray}
where the electric charge is expressed in terms of the generators of the SM group and left-right symmetric group as,
\begin{equation}
Q=T_{3L}+T_{3R}+{\frac{B-L}{2}}=T_{3L}+Y\,.
\label{1}
\end{equation}
Since the fields $\Delta_{\nu \nu} \mbox{(S)}$, $\Delta_{uu}$, $\Delta_{ud}$, $\Delta_{ud}$ and quark fields are mainly
responsible for non-zero baryon asymmetry and neutron-antineutron oscillation,we need to know the exact
interactions among them. The desirable interaction Lagrangian for diquark Higgs scalars with the SM quarks at
TeV scale which will yield observable neutron-antineutron oscillation and post-sphaleron baryogenesis is
\begin{eqnarray}
\hspace*{-0.5cm}\mathcal{L} &\supset& \frac{f_{ij}}{2}\, \Delta_{d d} d_i d_j + \frac{h_{ij}}{2}\, \Delta_{u u} u_i u_j
+ \frac{g_{ij}}{2\sqrt{2}} \Delta_{ud} \left(u_i d_j + d_i u_j \right) \nonumber \\
&+& \frac{\lambda}{2} \Delta_{\nu \nu} \Delta_{dd} \Delta_{ud} \Delta_{ud}
+ \frac{\lambda^\prime}{2} \Delta_{\nu \nu} \Delta_{uu} \Delta_{dd} \Delta_{dd} + \mbox{h.c.} \nonumber \\
&\subset& F\, \left(\psi^T_{R a}\, C^{-1}\, \tau_2\, \vec{\tau} \cdot \Delta^\dagger_{ab}\, \psi_{R b}
+ \mbox{L}\leftrightarrow \mbox{R} \right) +\mbox{h.c.}\,\, \mbox{under}\quad \mathcal{G}_{224}\, ,
\label{eqn:BVint}
\end{eqnarray}
where $F$, $f,h,g$ are the Majorana couplings and $\tau$ is the generator for $SU(2)$ group.
Within the $SO(10)$ framework, the Yukawa couplings obey the boundary condition, $f_{ij} = h_{ij} = g_{ij}$ in the $SU(2)_L
\times SU(2)_R \times SU(4)_C \times D$ limit and the same holds true for quartic Higgs couplings $\lambda=\lambda^\prime$ as well.
All fermions are right-handed (when chiral projection on the operator is suppressed) and a fermion field under the
high scale Pati-Salam symmetry $\mathcal{G}_{224}$ transforms as,
\begin{eqnarray}
\psi_{L,R}=\left(
\begin{array}{cccc}
u_1 & u_2 & u_3 & \nu \\
d_1 & d_2 & d_3 & e
\end{array}
\right)_{L,R}
\end{eqnarray}
The diquark Higgs scalars transforming under the SM gauge group $SU(2)_L \times U(1)_Y \times SU(3)_C$
are denoted with quantum numbers as,
\begin{equation}
\Delta_{\nu \nu} (1, 0, 1),\, \Delta_{u^c u^c} (1, -4/3, 6^*),\, \Delta_{d^c d^c} (1, 2/3, 6^*),\,\, \mbox{and}\,\,
\Delta_{u^c d^c} (1, -1/3, 6^*)\, .
\end{equation}
It is clear from eqn\,(\ref{eqn:BVint}) that the Higgs field $\Delta_{\nu \nu} (1, 0, 1) \subset \Delta_R(1, 3, -2, 1)
\subset (1, 3, \overline{10}) $ is a neutral complex field. The breaking of $\mathcal{G}_{2113} \to \mathcal{G}_{213}$
is achieved by assigning a VEV to its neutral component $\Delta_{\nu \nu} \subset \Delta_{R} (1,0,-2,1)$. Its real component
acquires a VEV in the ground state which can be represented as $\Delta_{\nu \nu} =
v_{B-L} + \frac{1}{\sqrt{2}} \left(S_r + i \rho \right)$ while the field $\rho$ gets absorbed by the gauge boson corresponding
to the gauge group $U(1)_{B-L}$. Therefore, the remaining real scalar field $S_r$ is indeed the physical Higgs particle
which serves our purpose of explaining post-sphaleron baryogenesis and neutron-antineutron oscillation.
\subsection{General expression for CP-asymmetry}
Without loss of generality, if we consider the particle and antiparticle decay modes of $S_r$ ( $S_r$ being its own antiparticle)
i.e, $S_r \to u^c d^c u^c d^c d^c d^c$ which gives a change of baryon number $\Delta B_{(S_r \rightarrow 6 q^c)} = +2$, and
$S_r \to \overline{u^c} \overline{d^c} \overline{u^c}
\overline{d}^c \overline{d}^c \overline{d}^c$ which gives $\Delta B_{(S_r \rightarrow 6 \overline{q^c})} = -2$,
then the $CP$-asymmetry in baryon number produced by these decays can be quantified as,
\begin{align}
\varepsilon_{CP} &= \frac{\Delta B_{(S_r \rightarrow 6 q^c)}\, \Gamma(S_r \rightarrow 6 q^c)}{\Gamma_\text{tot}}
+\frac{\Delta B_{(S_r \rightarrow 6 \overline{q^c})}\, \Gamma(S_r \rightarrow 6 \overline{q^c})}{\Gamma_\text{tot}}
\;,\nonumber\\
&= \frac{(+2)\, \Gamma(S_r \rightarrow 6 q^c)
+(-2)\, \Gamma(S_r \rightarrow 6 \overline{q^c})}{\Gamma_\text{tot}} = 2\, \frac{\Gamma -\bar{\Gamma}}{\Gamma_\text{tot}}
\;,
\label{eq1:CP_asym_general}
\end{align}
where $\Gamma_\text{tot} = \Gamma +\bar{\Gamma}$ is the total decay rate with $\Gamma \equiv \Gamma(S_r \rightarrow 6 q^c)$
and $\bar{\Gamma}\equiv \Gamma(S_r \rightarrow 6 \overline{q^c})$. It is evident from eqn\, (\ref{eq1:CP_asym_general})
that we need divergent partial decay rates for particle and antiparticle decays in order to produce correct amount of baryon
asymmetry and hence we should derive the general conditions under which $\Gamma$ and $\bar{\Gamma}$ can be different. It
is worth to mention here that the other decay modes of $S_r$ have been ignored for simplicity by adjusting the
corresponding couplings involved in the respective decay modes.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.85,angle=0]{Sdecay.eps}
\caption{Feynman diagram representing the decay of $S\to 6 q$ at tree level in order to explain post-Sphaleron
baryogenesis operative at TeV scale. Since $S$ is a real scalar field, the decay mode $S \to 6 \overline{q}$
is possible by reversing the arrow direction of the quark field.}
\label{fig:post-sphaleron:tree}
\end{figure}
In generic situations where the theory is CPT-conserving, there can never be a difference between $\Gamma$ and $\bar{\Gamma}$
if one considers only the tree-level process depicted in Fig.~\ref{fig:post-sphaleron:tree} since $\Gamma =\bar{\Gamma}$ at
tree level. It is found that the nonzero contribution to $\varepsilon_{CP}$ comes from the interference between the tree-level graph
(shown in Fig.~\ref{fig:post-sphaleron:tree}) and the one-loop corrections (shown in Fig.~\ref{fig:post-sphaleron:loop}).
\subsection{Constraints on post-Sphaleron baryogenesis}
Here we illustrate how post-Sphaleron baryogenesis is slightly different from any other standard baryogenesis
process. For post-Sphaleron baryogenesis to be successful in explaining the required matter-antimatter
asymmetry of our Universe, few extra conditions must be satisfied by the model parameters along with the
\emph{Sakharov conditions} that says, particle interaction must {\bf (i)} violate baryon number, $B$, {\bf (ii)} violate $C$ and $CP$,
and {\bf (iii)} be out of thermal equilibrium. Firstly, the $S_r$ Higgs scalar should be lighter than other members
contained in the Pati-Salam multiplet $(1,3, \overline{10})$ i.e, the diquark
Higgs scalars $\Delta_{qq}$ so that the baryon number conserving decays involving on-shell $\Delta_{qq}$
are kinematically forbidden. Secondly, the out of equilibrium baryon number violating decays should
occur after the electroweak phase transition so that it will not be affected by the Sphaleron processes
which is proactive at $> $TeV scale. We make it a point here that ref.\,\cite{Babu:2013yca} neatly elaborates
the mechanism of post-sphaleron baryogenesis.
\subsection{Out of equilibrium condition}
For effectively creating the baryon asymmetry of the universe via post-Sphaleron baryogenesis, the decays
of $\Gamma(S_r \rightarrow 6 q^c)$ should satisfy the out of equilibrium condition, which is described
by $\Gamma_{S_r} \lesssim H(T)$ where $\Gamma= \Gamma(S_r \rightarrow 6 q^c)= \frac{36}{\left(2 \pi \right)^9}
\frac{\left(\textbf{Tr}[f^\dagger f] \right)^3\, \lambda^2 M^{13}_S}{6 M^{12}_\Delta}$ is the total decay width
and $ H \simeq 1.66 \sqrt{g^*_s}\, \frac{T^2}{M_\text{Pl}} \;,$ is the Hubble parameter with the reduced Planck
mass $M_{\textrm{Pl}} \simeq 1.2 \times 10^{18}\,\textrm{GeV}$ and $g^*_s$ is the number of relativistic degrees
of freedom. In order to satisfy the out of equilibrium condition, we should have
\begin{eqnarray}
& &\Gamma_{S_r} \simeq H \left|_{(T=T_d)}\right. \nonumber\\
&\Rightarrow& \mbox{T}_d =\bigg[ \frac{36\,\lambda^2\,\left(\textbf{Tr}[f^\dagger f] \right)^3\,M_{\textrm{Pl}} M^{13}_S}{
\left(2 \pi\right)^9\, 1.66\, g^{1/2}_{\ast} \left(6 M_\Delta \right)^{12} } \bigg]^{1/2} \simeq 6.1 \times
\left(\frac{M^{13}_S}{M^{12}_{\Delta}}\right)^{1/2} \,\mbox{GeV}^{1/2}
\label{eq2:Sdecay-tree}
\end{eqnarray}
To illustrate the mechanism of post-sphaleron baryogenesis, we require extra fields $\Delta_{uu}$, $\Delta_{ud}$ and
$\Delta_{dd}$ as color sextets and $SU(2)_L$ singlet scalar bosons that couple to the right-handed quarks contained
in the Pati-Salam multiplet $(1,3,\overline{10})$. For set of model parameters $M_S=500$ GeV, $M_\Delta \simeq 1000$ GeV,
the decoupling temperature is found to be 2 GeV which is well below the EW scale where the Sphaleron has been decoupled.
Hence, it is inferred from the above equation that the decay of $S$ goes out of equilibrium around $T\simeq M_S$. Below this
temperature ($T < M_S$), the decay rate falls very rapidly as the temperature cools down.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.95,angle=0]{Sdecay-loop.eps}
\caption{Feynman graphs of the one-loop vertex correction for $\Gamma(S_r \rightarrow 6 q^c)$.}
\label{fig:post-sphaleron:loop}
\end{figure}
\subsection{Estimation of net baryon asymmetry}
Now we concentrate on estimating the CP-asymmetry coming from the interference term between the tree level and the one-loop level
diagrams for the decay of $S_r$ which is shown in Fig.\ref{fig:post-sphaleron:tree} and Fig.\ref{fig:post-sphaleron:loop}
respectively. For discussion on baryon number violation in the loop diagram and necessary derivation of the interference diagram,
interested readers may go through reference \cite{Babu:2013yca}. In the present work, we only check whether or not the
representative set of model parameters provide the correct number for the required baryon asymmetry of the universe.
Hence, without going deep into the derivation, we simply note here down, the calculated CP-asymmetry
for post-sphaleron baryogenesis via decay of $S_r$ with baryon number violating interactions.
\begin{align}
\varepsilon_{\rm wave} &\simeq \frac{g^2}{64 \pi \mbox{Tr}(f^\dagger f)} f_{j \alpha} V^*_{j \beta} f_{i \alpha} \delta_{i3}
\frac{m_t m_j}{m^2_t-m^2_j} \sqrt{\left(1-\frac{m^2_W}{m^2_t} +\frac{m^2_\beta}{m^2_t}\right)^2- 4 \frac{m^2_\beta}{m^2_t}}
\nonumber \\
& \times \bigg[2 \left(1-\frac{m^2_W}{m^2_t} +\frac{m^2_\beta}{m^2_t}\right)
+ \left(1+\frac{m^2_\beta}{m^2_t}\right)\left(\frac{m^2_t}{m^2_W} +\frac{m^2_\beta}{m^2_t}-1\right)
- 4 \frac{m^2_\beta}{m^2_W}\bigg] \;,
\label{eq3:wave-cpasy}\\
\varepsilon_{\rm vertex} &\simeq \frac{g^2}{32 \pi \mbox{Tr}(f^\dagger f)} f_{i \beta} V^*_{i \beta} f_{i \alpha} \delta_{i3}
\frac{m_j m_\beta}{m^2_W} \bigg[1+\frac{9 m^2_W}{M^2_S}\mbox{ln}\left(1+\frac{M^2_S}{3 m^2_W} \right) \bigg] \;,
\label{eq3:vertex-cpasy} \\
\varepsilon_{\rm CP} &=\varepsilon_{\rm wave} + \varepsilon_{\rm vertex} \, .
\label{eq:CP_asym_expressn}
\end{align}
Here the expression in eq.(\ref{eq3:wave-cpasy}) represents the CP-asymmetry coming from interference between the tree and
one-loop self energy diagram while the expression in eq.(\ref{eq3:vertex-cpasy}) represents the CP-asymmetry due to interference
of the tree and one-loop vertex diagram (see ref.\cite{Babu:2013yca} for details).
In the above expression, $V$ is the well known CKM matrix in the quark sector, $i,j$ correspond to the up-quark indices
$u,c,t$ while $\alpha,\beta$ represent to down-quark indices $d,s,b$. Sum over repeated indices (Einstein convention)
is implicitly assumed here. The $\delta_{i3}$ is due to the fact that the CP asymmetry is non-zero only when we have
a top quark in the final state (since only the CKM elements involving third generation have a large imaginary part).
As mentioned earlier,the mechanism of post-sphaleron baryogenesis provides a natural explanation for the observed baryon
asymmetry of our universe i.e, $\eta_B\simeq 10^{-10}$. Using $m_c=1.27$ GeV, $m_b=4.25$ GeV, $m_t=172$ GeV, CKM mixing
elements $V_{CKM}$ and Yukawa couplings relevant for color scalar particles in their allowed range, the CP-asymmetry
via the decay of $S_r$ through loop diagrams with the exchanges of $W^{\pm}$ bosons is estimated to be $10^{-8}$. A
further dilution of the baryon asymmetry arises from the fact that $T_d \ll M_S$, since the decay of $S_r$ releases entropy
into the universe. As a result the final baryon asymmetry, taking into account the dilution factor, becomes
\begin{eqnarray}
\eta_B = \varepsilon_{\rm CP} \times \left(\frac{T_d}{M_S} \right)\, ,
\end{eqnarray}
where $T_d$ is the decoupling temperature of the color scalar and $M_S$ is the mass of the scalar. The condition $T_d/M_S \geq 10^{-2}$,
otherwise leads to suppressed baryon asymmetry, which finally results a baryon asymmetry in the range of $10^{-10}$.
The scatter plot between the final baryon asymmetry including dilution factor ($\eta_B$)
with this phase ($\delta_{i3}$) is shown in Fig.\,\ref{plot:CP-asymmetry}.
\begin{figure}[t]
\centering
\includegraphics[scale=1.2,angle=0]{psb_scatterplot_phase.eps}
\caption{Estimation of final baryon asymmetry in terms of CP-asymmetry with overall phase $\delta$ contained
in the CKM mixing matrix.}
\label{plot:CP-asymmetry}
\end{figure}
\section{OBSERVABLE NEUTRON-ANTINEUTRON OSCILLATION WITH TEV SCALE DIQUARK HIGGS SCALARS:}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.55,angle=0]{Sdecay-nnbar.eps}
\caption{Loop contributions to neutron-antineutron oscillation in the post-Sphaleron baryogenesis operative
at TeV scale.}
\label{fig:nnbar-loop}
\end{figure}
\subsection{Feynman amplitudes for neutron-antineutron oscillation}
We consider the contributions arising only from the RH diquark Higgs fields having masses at TeV scale while ignoring the
contributions from LH diquark Higgs fields since they have masses at around $\mbox{eV}$
range. The Feynman diagrams contributing to the neutron-antineutron oscillation are shown in Fig.\,\ref{fig:nnbar-loop} (loop-diagram),
Fig.\,6({\bf a}) and Fig.\,6({\bf b}). Our prime goal is to estimate the mixing time
for this loop diagram, clarifying why we have suppressed other contributions within our model parameters.
There are two types of contributions to $n-\overline{n}$ oscillation in the right-handed
sector at loop level (i) one involving one $u^c u^c$-type and two $d^c d^c$-type, (ii)
other one involving one $d^c d^c$-type and two $u^c d^c$-type $\Delta$-bosons. The Feynman
amplitude for the second type of contribution where one needs to change the two $b^c$ quarks
to two $d^c$ quarks from the already generated effective operator $u^c d^c b^c u^c d^c b^c$
via a second order weak interactions (given in Fig.\,\ref{fig:nnbar-loop}) can be written as,
\begin{eqnarray}
\mathcal{A}^{\rm 1-loop}_{n-\overline{n}} \simeq
\frac{\left(f_{ud}\right)_{11} \left(f_{ud}\right)_{13} \left(f_{dd}\right)_{13}\, \lambda v_{B-L}}
{M^4_{u^c d^c} M^2_{d^c d^c}}\, \frac{g^4\, V^2_{td}\, m^2_b\,m^2_t}{\left(16 \pi^2 \right)^2 M^4_{W_L}}\,
\mbox{\Large log}\left(\frac{m^2_b}{M^2_{W_L}}\right)
\label{eq:ampl-loop}
\end{eqnarray}
And, the Feynman amplitude for tree level processes shown in Fig.\,6({\bf a}) and Fig.\,6({\bf b})
(which are suppressed with the choice of our model parameters), can be written as,
\begin{eqnarray}
\mathcal{A}^{\rm tree}_{n-\overline{n}} &=& \mathcal{A}^{\rm (a)}_{n-\overline{n}}+ \mathcal{A}^{\rm (b)}_{n-\overline{n}} \nonumber \\
&\simeq&\frac{\left(f_{dd}\right)_{11} \left(f_{ud}\right)^2_{11}\, \lambda\, v_{B-L}}
{M^4_{u^c d^c} M^2_{d^c d^c}} + \frac{\left(f_{uu}\right)_{11} \left(f_{dd}\right)^2_{11}\, \lambda\, v_{B-L}}
{M^4_{d^c d^c} M^2_{u^c u^c}}
\label{eq:ampl-tree}
\end{eqnarray}
\begin{figure*}[htb]
\begin{minipage}[t]{0.49\textwidth}
\hspace{-0.4cm}
\begin{center}
\includegraphics[scale=0.43,angle=0]{nnbar1.eps}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[t]{0.49\textwidth}
\hspace{-0.4cm}
\begin{center}
\includegraphics[scale=0.43,angle=0]{nnbar2.eps}
\end{center}
\end{minipage}
\caption{Feynman diagrams contributing to neutron-antineutron oscillation. The figure in {\it left-panel}
involves two $\Delta_{u^c d^c}$ and one $\Delta_{d^c d^c}$ bosons whereas the figure in {\it right-panel} involves
two $\Delta_{d^c d^c}$ and one $\Delta_{u^c u^c}$ bosons. The structure of the theory is such that these tree-level
contributions are suppressed in the present work.}
\label{fig:supp-nnbar}
\end{figure*}
\subsection{Prediction for neutron-antineutron mixing time $\tau_{n-\overline{n}}$}
Before estimating the $n-\overline{n}$ oscillation mixing time one should carefully fix the input parameters
in order to satisfy flavor changing neutral current (FCNC) constraints and to give correct amount of baryon
asymmetry of the universe. For example, using diquark sextet Higgs scalar mass around TeV scale, the corresponding
Yukawa coupling $(f_{dd})_{11}\simeq 0.001-0.1$ along with other allowed range of model parameters contradicts
the FCNC constraints and hampers post-sphaleron baryogenesis even though it predicts neutron-antineutron oscillation
time (as shown in Fig.\,6) within the experimental search limits. So this means that one has to choose the Majorana Yukawa coupling
$f$ accordingly. Now we briefly discuss how this choice of $f$ can be achieved within the framework of SO(10) (elaborated in
ref \cite{Awasthi:2013ff}).
It is found in ref \cite{Awasthi:2013ff} that all charged fermion masses and CKM mixing can
be fitted well at GUT scale within the framework of SO(10) with two kinds of structures; I) with single Higgs representation
$126_H$, II) with two Higgs representations $126_H$, $126_H^\prime$. As it has been derived, structure-I with Yukawa coupling
$f_{126_H}$ = diag(0.0236, -0.38, 1.5) estimates $n-\bar{n}$ oscillation mixing time to be $10^9$ secs which doesn't
serve our purpose. Rather we consider structure-II where the dominant
contribution to $n-\overline{n}$ oscillation comes from the loop diagram while suppressing the tree level contribution.
This choice of having two Higgs $126_H$, $126_H^\prime$ not only fits fermion masses at GUT scale, but also allows RH neutrino
Majorana mass and hence corresponding Yukawa coupling $f_{126_H^\prime}$ as per our requirement.
Due to the second Higgs representation ${126}^{\prime}$ with its Yukawa coupling $f^{\prime}$ to fermions
we get $v_{\xi^{\prime}}= 1-100 $ MeV following the same procedure, provided all other components are at the
GUT scale except $\xi^{\prime}(2,2,15)$ which is at the intermediate scale $M_{\xi^{\prime}}=10^{13}-10^{14}$ GeV.
By treating the mass of $\xi(2,2,15) \subset 126$ to remain at its natural GUT-scale value, its induced VEV is
negligible and precision unification with large GUT scale value is unaffected except for phenomenologically
inconsequential additional threshold effects. Then defining $F= f^{\prime}v_{\xi^{\prime}}$ gives exactly
the same fit to the GUT scale fermion masses and mixings but now with the diagonal structure $f_i^{\prime}
=(0.0236, -0.38, 1.5)$. But since $<\Delta_R^{\prime}>=0$ and only $\Delta_R \subset {126}_H$ with VEV $v_R$
is used to break $G_{2113}$, the coupling $f$ and hence $M_N$ are allowed to have any $3 \times 3$ form
without any restriction. In order to suppress the tree level contributions to $n-\overline{n}$ oscillation
as shown in Fig.\,6 which otherwise causes problem in baryon asymmetry, we particularly choose
the Majorana coupling $f_{dd}$ as per our requirement, i.e, ${f_{dd}}_{11}\leq 10^{-5}$.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$f_{13}$ & $g_{11}$ & $g_{13}$ & $\lambda$ & ${M_\Delta}_{ud}$ (GeV) & ${M_\Delta}_{dd}$ (GeV) & $\tau_{n-\bar{n}}$ (sec)\\
\hline
0.001 & 0.01 & 0.01 & 0.1 & $10^3$ & $10^4$ & $3.96\times 10^8$\\
\hline
0.001 & 0.01 & 0.01 & 0.1 & $10^3$ & $10^5$ & $8.72\times 10^{10}$\\
\hline
0.001 & 0.01 & 0.01 & 1 & $10^3$ & $10^5$ & $3.29\times 10^9$\\
\hline
0.001 & 0.001 & 0.001 & 0.1 & $10^3$ & $10^4$ & $4.42\times 10^{10}$\\
\hline
\end{tabular}
\caption{Numerical estimation of neutrino-antineutrino oscillation time}
\label{tab:nnbar-mixing}
\end{center}
\end{table}
Using this particular choice of Yukawa couplings i.e, ${f_{dd}}_{11},\, {f_{dd}}_{22},\leq 10^{-5}$ and others in
the range of $0.001-1.0$, one can calculate the mixing time for neutron-antineutron oscillation as a function of
Mass of color Higgs scalar ($B-L$ breaking scale) as shown in Fig.\,\ref{plot:nnbar-Mud} (Fig.\,\ref{plot:nnbar-vBL}).
\begin{figure}[htb!]
\centering
\includegraphics[scale=1.1,angle=0]{nnbar_scatterplot_Mud.eps}
\caption{Estimation of $\tau_{n-\bar{n}}$ as a function of di-quarks mass $M_{\Delta_{ud}}$.}
\label{plot:nnbar-Mud}
\end{figure}
The $n-\bar{n}$ amplitude can be translated into the $n-\overline{n}$ oscillation time as,
\begin{equation}
\tau_{n-\bar{n}}^{-1} = {\delta m}_{n-\bar{n}} = C_{{\rm \small QCD}} (\mu_\Delta, \mbox{1\,GeV}) |A_{n-\bar{n}}^{\rm 1-loop}|
\end{equation}
with $C_{{\rm \small QCD}} (\mu_\Delta, \mbox{1\,GeV})$ = $0.1 \mbox{GeV}^6$ as used in ref.\cite{Babu:2013yca}. The estimated
$n-\bar{n}$ oscillation time for various choice of model parameters i.e, ${f_{ud}}_{11} \leq {10}^{-5}$, $M_S = (100-5000)GeV$,
B-L breaking scale from (3-5)TeV and the masses of ${M_\Delta}_{ud/dd}$ between $M_S$ and $V_{B-L}$, $\lambda\simeq 0.01-1.0$
is presented in Table.\ref{tab:nnbar-mixing}.
\begin{figure}[htb!]
\centering
\includegraphics[scale=1.1,angle=0]{nnbar_scatterplot_vBL.eps}
\caption{Estimation of $\tau_{n-\bar{n}}$ as a function of $B-L$ breaking scale $v_{B-L}$
while keeping other model parameters within their allowed range consistent with mechanism
of post-sphaleron baryogenesis.}
\label{plot:nnbar-vBL}
\end{figure}
\subsection{Coupling Unification including diquarks at TeV scale}
It is prominent that the post-sphaleron baryogenesis and neutron-antineutron oscillation phenomena require existence of
color Higgs scalars, having masses around TeV scale. In this subsection, we intend to examine whether unification of gauge couplings
is still possible after the addition of extra color scalars $\Delta_{ud}$, $\Delta_{dd}$, $\Delta_{uu}$ to the existing particle
content as noted in Sec.\ref{sec:rge}, by studying their respective renormalization group equations.
The one-loop beta coefficients derived for the present model along with their gauge
symmetry groups, range of mass scales and spectrum of Higgs scalars necessary for gauge
coupling unification to explain TeV scale post-sphaleron baryogenesis and
neutron-anti-neutron oscillation are given below
\begin{eqnarray}
& &{\bf \mbox{(i)}\, \mu=M_Z (91.817\, \mbox{GeV}) - M_T (1\, \mbox{TeV}) }: \quad \mathcal{G} = \mathcal{G}_{2_L 1_Y 3_C}\equiv \mbox{SM},
\nonumber \\
& &\mbox{Higgs:}\, \Phi(2, 1/2, 1)_{10}:\quad \quad
\pmb { a_i}=\left(-19/6,\, 41/10,\, -7\right)\,; \\
& &{\bf \mbox{(ii)}\, \, \mu=M_T(1\, \mbox{TeV}) - M_{B-L}(3\, \mbox{TeV}):} \quad \mathcal{G} = \mathcal{G}_{2_L 1_Y 3_C}
\nonumber \\
& &\mbox{Higgs:}\, \Phi(2,1/2,1)_{10} \oplus S(1, 0, 1)_{126}\subset \Delta_R \oplus \Delta_{u^c d^c} (1, -1/3, 6^*)_{126}
\oplus \Delta_{d^c d^c} (1, 2/3, 6^*)_{126}\, \nonumber \\
& &\hspace*{1.2cm} \oplus \Delta_{u^c u^c} (1, -4/3, 6^*)_{126}: \nonumber \\
& &\pmb{ a_i} = \left(-19/6,\, 207/30,\, -27/6\right)
\end{eqnarray}
\begin{eqnarray}
& &{\bf \mbox{(iii)}\, \, \mu=M_{B-L}(3\, \mbox{TeV}) - M_\Omega(10\, \mbox{TeV}):}
\quad \mathcal{G} = \mathcal{G}_{2_L 1_R 1_{B-L} 3_C} \nonumber \\
& &\mbox{Higgs:}\,\Phi_1(2, 1/2, 0, 1)_{10} \oplus \Phi_2(2, -1/2, 0, 1)_{10^\prime} \oplus \Delta_R(1, 1, -1, 1)_{126}
\oplus \chi_R(1, 1/2, -1/2, 1)_{16},\, \nonumber \\
& &\hspace*{1.2cm} \oplus \Delta_{u^c d^c}(1, 1, -2/3, 6^*)_{126} \oplus \Delta_{d^c d^c} (1, 0, -2/3, 6^*)_{126}
\oplus \Delta_{u^c u^c} (1, 0, -2/3, 6^*)_{126}: \nonumber \\
& &\pmb{a_i} = \left(-3,\, 35/4,\, 45/8, -27/6\right)
\end{eqnarray}
\begin{eqnarray}
& &\hspace*{-0.4cm} {\bf \mbox{(iv)}\, \, \mu=M_\Omega(10^{4}\, \mbox{GeV}) - M_C(10^{5}-10^{6}\, \mbox{GeV}):}
\quad \mathcal{G} = \mathcal{G}_{2_L 2_R 1_{B-L} 3_C} \nonumber \\
& &\mbox{Higgs:}\,\Phi_1(2, 2, 0, 1)_{10} \oplus \Phi_2(2, 2, 0, 1)_{10^\prime} \oplus \Delta_R(1, 3, -1, 1)_{126}
\oplus \chi_R(1, 2, -1/2, 1)_{16},\, \nonumber \\
& &\hspace*{1.2cm} \oplus \Delta_{u^c d^c} (1, 3, -2/3, 6^*)_{126} \oplus \Delta_{d^c d^c} (1, 3, -2/3, 6^*)_{126}
\oplus \Delta_{u^c u^c} (1, 3, -2/3, 6^*)_{126} \nonumber \\
& &\hspace*{1.2cm} \oplus \Omega_R(1, 3, 0, 1)_{210} \nonumber \\
& &\pmb { a_i} = \left(-8/3,\, 4/3,\, 55/4, -2\right)
\end{eqnarray}
In analogy to the above discussion, we have two scenarios; one without bitriplet and another with bitriplet Higgs scalar
(3,3,1) under the Pati-Salam group $SU(2)_L \times SU(2)_R \times SU(4)_C$ while its effect has been included from $M_C$
onwards to the unification scale $M_U$. Accordingly, we have estimated the one-loop beta coefficients for these two
scenarios as
\begin{eqnarray}
& &\hspace*{-0.4cm} {\bf \mbox{(v)}\, \, \mu=M_C - M_\xi:} \quad \mathcal{G} = \mathcal{G}_{2_L 2_R 4_C} \nonumber \\
& &\mbox{Higgs:}\,\Phi_1(2, 2, 1)_{10} \oplus \Phi_2(2, 2, 1)_{10^\prime} \oplus \Delta_R(1, 3, \overline{10})_{126}
\oplus \chi_R(1, 2, \overline{4})_{16} \oplus \Omega_R(1, 3, 15)_{210} \nonumber \\
& &\pmb{a_i} = \left(-8/3,\, 29/3,\, -14/3\right)
\end{eqnarray}
\begin{eqnarray}
& &\hspace*{-0.4cm} {\bf \mbox{(vi)}\, \, \mu=M_\xi - M_P:} \quad \mathcal{G} = \mathcal{G}_{2_L 2_R 4_C} \nonumber \\
& &\mbox{Higgs:}\,\Phi_1(2, 2, 1)_{10},\, \Phi_2(2, 2, 1)_{10^\prime},\,\Delta_R(1, 3, \overline{10})_{126},\,
\chi_R(1, 2, \overline{4})_{16}, \Omega_R(1, 3, 15)_{210} + \xi(2, 2, 15)_{126^\prime} \nonumber \\
& &\pmb { a_i} = \left(7/3,\, 44/3,\, 2/3\right)
\end{eqnarray}
\begin{eqnarray}
& &\hspace*{-0.4cm} {\bf \mbox{(vii)}\, \, \mu=M_P - M_U:} \quad \mathcal{G} = \mathcal{G}_{2_L 2_R 4_C} \nonumber \\
& &\mbox{Higgs:}\,\Phi_1(2, 2, 1)_{10},\, \Phi_2(2, 2, 1)_{10^\prime},\, \Delta_R(1, 3, \overline{10})_{126},\,
\Delta_L(3, 1, 10)_{126},\, \chi_R(1, 2, \overline{4})_{16},\, \chi_L(1, 2, 4)_{16},\, \nonumber \\
& &\quad \quad \quad \Omega_R(1, 3, 15)_{210},\, \Omega_L(3, 1, 15)_{210},\,
\xi(2, 2, 15)_{126^\prime},\, \Sigma^\prime(1,1,15)_{210},\, \nonumber \\
& &\pmb { a_i} = \left(44/3,\, 44/3,\, 6 \right) \nonumber \\
\end{eqnarray}
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.98,angle=0]{coup_unif.eps}
\caption{Coupling unification for the present model where $\Delta_{u^c d^c}$, $\Delta_{d^c d^c}$, and $\Delta_{u^c u^c}$
have been included at TeV scale keeping in mind that these particle mediate neutron-antineutron oscillation
and baryon asymmetry and including $\xi(2,2,15)$ around $10^{12}$ GeV in order to fit the fermions
masses at GUT scale.}
\label{fig:unif}
\end{figure}
The gauge coupling unification after the addition of extra color sextet scalars particles is shown in Fig.\,\ref{fig:unif}
with the allowed mass scales desirable for our model predictions,
\begin{eqnarray}
& &M_{B-L}= 4-7~{\rm TeV},\, M_\Omega= 10~{\rm TeV},\, M_C = 10^{5}-10^{6}\, \mbox{GeV}\, ,\nonumber \\
& &M_P\simeq 10^{14.65}~{\rm GeV\,\, and\,\,} M_{\rm U}\simeq 10^{16.25}~{\rm GeV} \, .
\end{eqnarray}
\section{VIABILITY OF THE MODEL}
As already known, the lepton flavor and lepton number violating dilepton signals can be probed
from the production of heavy RH Majorana neutrino via $p+p \to W^\pm_R \to \ell^\pm_\alpha +N_R$,
from which $N_R$ can be further decayed into $N_R \to W^*_R \to \ell^{\mp}_\beta = 2j$. This process,
being the main channel for $N_R$ production via on-shell $Z_R$ production and $W_R$ fusion, needs to be
verified at LHC and our model suits the purpose, since we have $W_R$, $Z_R$ gauge bosons and scalar
diquarks at TeV scale. A more pleasant situation is that the model, though non-supersymmetric, predicts
similar branching ratios as in supersymmetric models for LFV processes like $\mu \to e\gamma$, $\tau \to
\mu\gamma$, and $\tau \to e\gamma$. And the predicted branching ratios for these LFV decays, being closer
to the current experimental search limits can be used to verify the left-right framework in this model.
Moreover the estimated neutron-antineutron oscillation mixing time, gauge coupling unification and
proton life time in the model stay in the range of ongoing search experiments.
Besides all these points, the model can also predict a number of verifiable new physical quantities like
(i) new non-standard contribution to $0\nu 2\beta$ rate in the $W_L-W_L$ channel, (ii) contributions to branching
ratios of lepton flavor violating (LFV) decays, (iii) leptonic CP-violation due to non-unitarity effects, (iv)
experimentally verifiable proton decay modes such as $p\to e^+\pi^0$, provided the gauged inverse seesaw mechanism
is found to be operative. We find it appropriate to mention here that these physical quantities were also discussed
in a recent work \cite{Awasthi:2013ff}, but in that model the asymmetric left-right gauge symmetry was incorporated
at $\simeq 10$ TeV.
\section{CONCLUSION}
We have closely studied the mechanism of post-sphaleron baryogenesis, that can potentially explain matter-antimatter
asymmetry of the present universe, by analyzing the basic interactions using quarks and diquark Higgs scalars under
high scale Pati-Salam symmetry and low scale SM like interactions at TeV scale. The study estimates the total baryon
asymmetry to be $\eta_B \simeq {\cal O}(10^{-10})$ and neutron-antineutron oscillation with mixing time to be
$\tau_{n-\bar{n}}\simeq {\cal O}(10^{-10} -10^{-8})$ secs which can be accessible at ongoing search experiments.
We have made an humble attempt to embed the framework of PSB in a non-SUSY $SO(10)$ model with Pati-Salam symmetry as a
low scale intermediate breaking step
where we have shown a strong interlink between post-sphaleron baryogenesis
and neutron-antineutron oscillation operative at TeV scale and laid out a novel mechanism of inducing required CP-asymmetry via the SM
$W_L^{\pm}$ loops.
More essentially, we have embedded TeV scale LR model
within the framework of $SO(10)$ model where the predicted mass for light neutrinos matches with the neutrino oscillation
data.
Our calculations indicate that TeV scale masses of
$W^\pm_R$ and heavy RH neutrinos can also give dominant non-standard contributions to neutrinoless double beta decay which
may sound crucial to the experimentalists. Some more good features of the model are explanation of non-zero light neutrino masses
via extended/inverse seesaw mechanism, new non-standard contribution to neutrinoless double beta decay, leptonic CP-violation from
non-unitary effects.
\noindent
{\bf ACKNOWLEDGEMENT}
\noindent Sudhanwa Patra would like to thank the organizers of workshop entitled \textquotedblleft
Majorana to LHC: Origin of neutrino Mass\textquotedblright at ICTP, Trieste, Italy during 2-5 October, 2013
where the idea for this work was conceived. Both the authors sincerely acknowledge P.S. Bhupal Dev for his
useful clarification while preparing the manuscript. Prativa Pritimita is grateful to the Department
of Science and Technology, Govt. of India for INSPIRE Fellowship (IF140299). The work of Sudhanwa Patra
is supported by the Department of Science and Technology, Govt. of India under the financial grant
SERB/F/482/2014-15.
|
\section{Introduction}
Dark Matter (DM), which accounts for about 80 \% of all mass in the
universe, is one of the strong indications for physics beyond the
Standard Model (SM) of particle physics. The dominant paradigm is that
dark matter is made of new, neutral and stable (or very long-lived
particles). The most studied possibility is the neutralino, which is
the archetype of a weakly interacting massive particle (WIMP). A WIMP
is a particularly attractive DM candidate. If the WIMP was in thermal
equilibrium in the early universe, its relic abundance is elegantly
fixed by its annihilation cross section, the matching with
cosmological observations requiring $\langle \sigma
v\rangle~\sim~10^{-26} \text{cm}^3\cdot \text{s}^{-1}$. Also the WIMP hypothesis
may be tested at colliders, using low background detectors (direct
detection) or through the annihilation of DM into SM particles
(indirect detection). One of the issues with indirect searches is that
a potential DM signal may be (and, unfortunately, is expected to be)
obscured by an overwhelming astrophysical background. Hence the
importance of possible so-called smoking gun signatures, {\em i.e.}
signals that have no (or little) astrophysical counterparts, like a
strong gamma-ray line or, more generally, one or many peaks in the
gamma ray spectral energy density (which are called spectral features)
\cite{Bergstrom:1988fp,Rudaz:1989ij,Bergstrom:1989jr} (see
also~\cite{Bringmann:2012ez} for a recent review). Gamma ray features
are actively being searched by the Fermi satellite and the HESS
telescope in the GeV to multi-TeV range. Remarkably, the current constraints
on the annihilation cross section of DM into gamma ray lines are
rather strong, ranging from $\langle \sigma v\rangle~\lesssim~10^{-28}
\text{cm}^3\cdot \text{s}^{-1}$ for $M_{DM} \sim 10$ GeV, \cite{Ackermann:2013uma}
to $\langle \sigma v\rangle~\sim~10^{-26} \text{cm}^3\cdot \text{s}^{-1}$ for
$M_{DM} \sim 10$ TeV~\cite{Abramowski:2013ax}.
A WIMP is neutral and thus its annihilation in gamma rays lines
proceeds through radiative corrections. In general the annihilation
cross section is suppressed by (powers of) the fine structure constant
$\alpha$ compared to the leading, say, $2 \rightarrow 2$ or 2-body
tree level processes. A notable exception occurs if the 2-body
processes, while being relevant in the early universe, are themselves
suppressed in astrophysical environments, like at the center of our
galaxy. This is for instance possible if the annihilation cross
section is velocity dependent. A familiar example is the annihilation
of a pair of Majorana particles into SM model fermion pairs, in which
case the cross section is mass suppressed and is p-wave in the chiral
limit $\sigma v \propto v^2$~\cite{Goldberg:1983nd}. Another example,
which has been put forward very recently, is annihilation of a real
scalar, again into light fermions, which may be d-wave in the chiral
limit, $\sigma v \propto
v^4$~\cite{Toma:2013bka,Giacchino:2013bta}. In both cases, a simple
consequence is that Bremsstrahlung emission is relatively enhanced,
possibly leading to observable features in the gamma ray spectrum as
well as non negligible contribution at the time of freeze-out
\cite{Bergstrom:1988fp,Baltz:2002we,Bergstrom:2004cy,Toma:2013bka,Giacchino:2013bta}.
Bremsstrahlung of gamma rays and $W$ and $Z$ electroweak gauge bosons
have been extensively studied in the literature, both for their own
sake and with phenomenological applications in mind, see for
instance~\cite{Beacom:2004pe,Boehm:2006df,Bringmann:2007nk,Ciafaloni:2011sa,Bell:2011if,Barger:2011jg,Garny:2011ii,Weiler:2013hh,DeSimone:2013gj,Kopp:2014tsa}. Of
particular interest for the present contribution, Barger {\em et al}
have compared the Bremsstrahlung spectral energy density in three
simple DM scenarios in \cite{Barger:2011jg}. All three scenarios
involve a new charged particle (the mediator) that is chirally coupled
to the DM particle and to SM fermions (which may be leptons or a
quarks). The DM is assumed to be its own antiparticle. For instability
it is also assumed to be odd under some $Z_2$ symmetry, and so is the
charged mediator. The latter must be clearly heavier than the DM
particle. Concentrating on spin $0$ and $1/2$ DM candidates, there are
then three possible scenarios. In scenario $1$, a Majorana DM is
coupled to the SM fermions through a charged scalar. This is similar
to the neutralino, in which case the charged scalar is a slepton or a
squark. In scenario $2$, the DM is also a Majorana particle, but now
it couples to a charged gauged boson. This is possible in some variant
on the Left-Right model \cite{Ma:2009tc}, in which case the DM is some
sort of heavy Majorana neutrino. Finally, in scenario $3$, the DM
is a real scalar coupled to SM fermions through heavy, vector-like
charged fermions. This scenario, which has been developed for other
phenomenological purposes, has been dubbed the Vector-Like Portal in
\cite{Perez:2013nra} (see also \cite{Frandsen:2013bfa} for an
alternative appellation).
In the present article we complement the work of Barger {\em et al}
\cite{Barger:2011jg} and the work we have initiated in
\cite{Giacchino:2013bta}. Concretely, Barger {\em et al} have shown
that, in all three scenarios sketched above, the Bremsstrahlung
spectral signature is precisely the same, up to a normalization that
is scenario dependent. There are good reasons for this, which we
briefly discuss in the next section. In \cite{Giacchino:2013bta} (see
also \cite{Toma:2013bka} in which precisely the same conclusions have
been reached), we have shown that the 2-body annihilation of the
scalar DM candidate is d-wave suppressed in the chiral limit, and
furthermore, that the Bremsstrahlung signal is parametrically
larger. These two effects combined imply that that scenario $3$ may
lead to more significant gamma ray features than a Majorana particle
(specifically scenario $1$). In the same work we had also tentatively
incorporated the contributions of gamma ray lines to the spectral
signatures. In the present work, we compare all 3 scenarios, and in
particular provide analytical expressions for the annihilation of the
DM candidates into 2 gamma rays. In scenario $1$, the result is
well-known and has been derived many times in the literature
\cite{Rudaz:1989ij,Bergstrom:1989jr,Giudice:1989kc,Bergstrom:1997fh},
with which, having redone the calculation, we agree. In scenario $2$,
an analytical expression for annihilation cross section may be
extracted from the results of \cite{Bergstrom:1997fh} in the MSSM. For
lack of time, we do not provide a fully independent check of this
expression. It may be of interest to do so, but having reached the
same result as \cite{Bergstrom:1997fh} in scenario $1$, we have no
reason to doubt their results. In scenario $3$, the full expression is
not available in the literature, so we give it in the present work.
The amplitude is given in~\cite{Bertone:2009cb} in the chiral limit,
and has been used as such e.g. in \cite{Tulin:2012uq} for
phenomenological purposes, but we believe that the result reported
there is incorrect\footnote{We believe that the error, which
propagated in the literature, is actually just due to a misprint in
equation (13) of \cite{Bertone:2009cb}, see Sec.~\ref{sec:Sgg}. It
is however virtually impossible to spot it without knowledge of the
correct answer. }. An expression for large mediator mass limit is
also available in~\cite{Boehm:2006gu}.
The plan is as follows. In the next section we begin with a
presentation of the basic features of the three scenarios of
\cite{Barger:2011jg}, including the tree level 2-body annihilation
cross sections and the expressions of the Bremsstrahlung (for emission
of a gamma). Next we give some details on the one loop calculations
of the DM annihilation into two gamma. In the final section, we
compare the spectral signatures of the three scenarios, and then draw
some conclusions.
\section{3 simple scenarios}
\label{sec:3scenarios}
The 3 scenarios that we consider, following \cite{Barger:2011jg}, are
very simple. They have in common the fact that DM annihilates into SM
fermions through a charged mediator in the t and u channels (we
consider the case of self-conjugate DM candidates). For simplicity we
assume that those channels are the only ones that are relevant, {\em
i.e.} that other interactions that a given DM candidate may have can
be neglected in some appropriate range of parameters. Hence the
results we discuss may correspond to a corner of all the possible
outcomes of more sophisticated models (for instance scenario $1$
is contained in the MSSM).
\subsection{Scenario $1$: Majorana DM candidate $\chi$ and charged scalar $\tilde E$}
The couplings with SM fermions take the form
\begin{equation}
{\cal L} \supset y_\chi \tilde E^\dagger \bar \chi P_R \psi_l + h.c.\, .
\label{eq:majDM}
\end{equation}
with $P_R = (1+\gamma_5)/2$.
Although the notation suggests that $\chi$ is coupled only to
right-handed SM leptons $\psi_l$, and so that $\tilde E$ carries a fermionic
charge, the results apply to couplings with quarks, or to $SU(2)$
doublets (modulo more degrees of freedom). Which to choose depends on
the underlying model. Clearly the collider constraints on the mass of
heavy charged fermions (scalars or others) are weaker than those on
particles that carry colour but on the other hand interactions like
that of (\ref{eq:majDM}) are constrained by non-observation of lepton
flavour violating processes, so one may have to compromise (see for
instance \cite{Kopp:2014tsa}). As usually, stability may be simply
insured by imposing a discrete symmetry,
$$
\chi \rightarrow - \chi \quad \mbox{and} \quad \tilde E \rightarrow - \tilde E
$$
In the chiral limit, $m_l \rightarrow 0$, and in the non-relativistic
limit $v_\chi \rightarrow 0$, the 2-body annihilation cross section,
$\chi \chi \rightarrow \bar l l$ is given by
\begin{equation}
\sigma v (\chi \chi \rightarrow l\bar l) = {y_\chi^4 \over 48 \pi}\, {v^2\over M_\chi^2}\, {1+ r_\chi^4\over (1+r_\chi^2)^4}
\label{eq:Maj2bdy}
\end{equation}
where
$$
r_\chi = {M_{\tilde E}\over M_\chi} \geq 1.
$$ ($v$ is as usual the M\o ller velocity, $v = 2 v_\chi$ in the
center of mass frame \cite{Gondolo:1990dk}).
That the annihilation cross section is p-wave in the chiral limit is
well known \cite{Goldberg:1983nd} and may be stated as follows. A pair
of non-relativistic Majorana DM particles in a s-wave corresponds the
state $^1S_0(O^{-+})$ in the $^{2 S+1}L_J (J^{CP})$ spectroscopic
notation, which, in terms of bi-linear operators, is represented by
$\bar \chi\gamma_5 \chi$. Correspondingly, in a CP conserving theory,
the final state fermion pair is represented by the operator $\bar
\psi_l \gamma_5 \psi_l$, which involves a chirality flip, and is thus
mass suppressed. In a p-wave, the state is $^3 P_1 (1^{++})$, or $\bar
\chi \gamma_k \gamma_5 \chi$, which is coupled to the fermion pair
current, $\bar \psi_l \gamma^kP_R \psi_l$. Hence in the chiral limit,
the annihilation cross section is p-wave.
\subsection{Scenario $2$: Majorana DM candidate $N$ and charged gauge boson $W^\prime$}
In this case the mediator is a charged gauge boson, which we call
$W^\prime$. This scenario is akin to the models proposed by Ma {\em et
al} in \cite{Khalil:2009nb,Ma:2009tc} based on a Left-Right model,
in which the charged gauge boson that couples to right handed (RH)
current carries a generalized fermion number. In that model, unlike
the conventional LR models, the RH neutrino, which we write $N$, is
not the mass partner of the $\nu_L$, but is a viable Majorana DM
candidate. For our purpose we write the coupling of $N$ to
$W^{\prime}$ as
\begin{equation}
{\cal L} \supset {g_N\over \sqrt{2}} W_\mu^{\prime +} \bar N \gamma^\mu P_R \psi_l + h.c.
\label{eq:LN}
\end{equation}
Notice that we have included a factor of $1/\sqrt{2}$, like in the SM,
so our convention for the gauge coupling is different from that of
\cite{Barger:2011jg}.
While the tree level processes may be calculated in a unitary gauge,
the one-loop annihilation cross section that we will rely on has been
calculated in a 't Hooft-Feynman version ($\xi = 0$) of a non-linear
$R_\xi$ gauge (for some details on such gauges, see
\cite{Bergstrom:1994mg}). For this, we also need the coupling of the
$N$ to the nonphysical Goldstone charged scalars, $G^\prime$, which,
one may check, must be given by
\begin{equation}
{\cal L} \supset {g_N \over \sqrt{2} M_{W^\prime}} G^{\prime +} \bar N\left( M_N P_R - m_l P_L\right)\psi_l\,,
\end{equation}
with $P_L=(1-\gamma_5)$
.
Although it for sure exists somewhere, we have not found the 2-body
cross section $N N\rightarrow ~\bar l l$ in the literature, so we give
it here, again in the chiral limit $m_f \rightarrow 0$,
\begin{equation}
\sigma v (N N \rightarrow \bar l l) = {g_N^4 \over 192 \pi}\, {v^2\over M_N^2}\, {(1+ 4 r_N^2 + 13 r_N^4 + 12 r_N^6 + 4 r_N^8)\over r_N^4(1+r_N^2)^4}
\label{eq:Maj2bdy}
\end{equation}
where now
\begin{equation}
r_N = {M_{W^\prime}\over M_N} \geq 1.
\end{equation}
The dependence on $r_N$ is a bit complicated, but notice that for
large $r_N$ we simply have
\begin{equation}
\label{eq:NN}
\langle\sigma v\rangle (N N \rightarrow \bar l l) \approx \langle v^2\rangle \, {g_N^4 \over 48 \pi}\, { M_N^2\over M_{W^\prime}^4}
\end{equation}
the same as for $\chi$
\begin{equation}
\label{eq:chichi}
\langle \sigma v\rangle (\chi \chi \rightarrow \bar l l) \approx \langle v^2\rangle \, {y_\chi^4 \over 48 \pi}\, { M_\chi^2\over M_{\tilde E}^4}
\end{equation}
Of course the 2-body cross section is p-wave for precisely the same
reason as in scenario~$1$.
\subsection{Scenario $3$: Real scalar DM candidate $S$ and charged vector-like fermion $E$}
In this last scenario, DM is a real scalar particle, $S$ with Yukawa
couplings to a charged vector-like $E$ fermion and the SM fermions
(again we consider couplings to SM singlets for simplicity)
\begin{equation}
\label{eq:yuk1}
{\cal L} \supset y_S\; S\; \bar E P_R \psi_l + h.c.\,.
\end{equation}
with as above
$$
S \longrightarrow - S$$
and
$$ E\longrightarrow - E$$ under some discrete $Z_2$ symmetry.
Following~\cite{Perez:2013nra,Giacchino:2013bta} we call this scenario
the Vector-Like Portal. Being a scalar singlet, $S$ has also a
renormalizable coupling to the SM
scalar~\cite{Silveira:1985rk,Veltman:1989vw,McDonald:1993ex,Burgess:2000yq}.
\begin{equation}
{\cal L} \supset {\lambda_S\over 2} S^2 \vert H \vert^2\,.
\label{eq:ls}
\end{equation}
We assume that this coupling, if present, is sub-dominant.
An interesting point about this scenario is that the annihilation
cross section in SM fermions is d-wave in the chiral
limit~\cite{Toma:2013bka,Giacchino:2013bta},
\begin{equation}
\sigma v(S S \rightarrow \bar l l) = { y_S^4 \over 60 \pi}{v^4\over M_S^2}{1\over (1+r^2)^4}
\label{eq:sv2S}
\end{equation}
The suppression by a factor $ v^4$ is a bit unusual but is easy to
understand. A pair of non-relativistic real scalar DM particles in a
s-wave have quantum numbers $^1 S_0(O^{++})$, corresponding to the
bi-linear operator $S^2$, which may be coupled to SM fermions through
$\bar\psi_l \psi_l$. Hence the amplitude for s-wave annihilation is
mass suppressed, $\propto m_l$. For a S pair, the p-wave state is
$^1 P_1(1^{-+})$ to which corresponds no fermion bi-linear (in a CP
conserving setup)\footnote{If CP is not conserved, or if the S is
taken to be complex, the state $^1 P_1(1^{--})$ is possible,
$S^\dagger \partial_k S$, which may be coupled to $\bar \psi_l
\gamma^k \psi_l$.}. The next possibility is then a d-wave, with $^1
D_2(2^{++})$. This $J=2$ state may be coupled to to SM fermions
through their stress-energy tensor $\Theta^{ij}_l = {i\over 2} \bar
\psi_l (\gamma^i \partial^j - \gamma^j \partial^i) \psi_l$. Hence the
amplitude is d-wave in the chiral limit.
The $v^4$ behaviour has interesting phenomenological implications. In
the early universe one
has~\cite{Gondolo:1990dk,Toma:2013bka,Giacchino:2013bta},
\begin{equation}
\langle v^2 \rangle = {6 \over x_f}\approx 0.24 \quad \mbox{and} \quad \langle v^4 \rangle = {60 \over x_f^2}\approx 0.1
\label{eq:vfo}
\end{equation}
for $x_f= 25$ where $x_f=M_{DM}/T_f$ and $T_f$ is the temperature at
freeze-out. The averaged velocities in Eq.~(\ref{eq:vfo}) represent a
mild suppression but which is enforced by the distinct $r =
M_{med}/M_{DM}$ dependence of the 2-body cross sections,
\begin{equation}
\label{eq:SS}
\langle \sigma v\rangle (S S \rightarrow l\bar l) \approx {\langle v^4\rangle}\,
{ y_S^4 \over 60 \pi}\, {M_S^6\over M_E^8}\nonumber
\end{equation}
(compare with Eqs.~(\ref{eq:NN}) and (\ref{eq:chichi})). Hence, for fixed
$r$, DM mass and thermal velocity, it is clear that the coupling $y_S$
must be larger than $y_\chi$ or $g_N$ to match the observed relic
abundance. This, as shown in \cite{Toma:2013bka,Giacchino:2013bta},
has interesting implications for the strength of radiative processes.
\begin{figure}
\includegraphics[width=9cm]{plots/sv2all.png}
\caption{Annihilation cross sections into a SM fermion pair as a function of
$r =M_{med}/M_{DM}$ (thermal averages and in the chiral limit). The cross sections are given
for unit couplings and for $M_{DM}=100$ GeV.}
\label{fig:sv2all}
\end{figure}
\section{Spectral energy density: internal Bremsstrahlung}
\label{sec:vib}
In this section we discuss the contribution of so-called internal
Bremsstrahlung to the spectral energy density of gamma rays. This has
been discussed extensively in the literature, so we just recap the
salient features. Bremsstrahlung is of interest for two reasons.
First, the annihilation cross section in a s-wave through
Bremsstrahlung is no longer mass suppressed. For one thing, there is
no obstruction from conservation of angular momentum, but there is
more to it. Although the argument is somewhat gauge-dependent, this
result may be traced to emission of a gamma ray from the virtual
massive charged particle in the t- and u-channels, or so-called virtual
internal Bremsstrahlung (see {\em e.g.} \cite{Bringmann:2007nk}).
This implies that the Bremsstrahlung, a 3-body final state process,
may be more important than the 2-body tree level process, despite the
suppression of the former by a factor ${\cal O}(\alpha/\pi)$. This is
typically the case for annihilation in light fermions or equivalently
heavy dark matter $M_{DM}\gg m_f$, and when the velocity is
non-relativistic, like at the galactic center ($v \sim 10^{-3}$).
Second, emission from the virtual mediator, depending on the ratio $r
= M_{med}/M_{DM}$, may have a sharp spectral feature, possibly
mimicking a monochromatic gamma ray line.
For reference, we give here the expressions of the 3-body annihilation
cross section for the 3 scenarios. Defining
\begin{eqnarray}
vd\sigma_{2\rightarrow 3}&=&\frac{|{\cal M}|^2}{128\pi^3} dxdy
\label{eq:dsv3}
\end{eqnarray}
where $v={\sqrt{{k_1 \cdot k_2}-m_1m_2}}/{E_1E_2}$ refers to the
relative velocity of the $S$ particles and $x, y$ are the reduced
energy parameters $x=2E_\gamma/\sqrt{s}$ and $y=2E_f/\sqrt{s}$, with $s$
the Mandelstam variable corresponding to the center-of-mass energy
squared, we have:
\subsubsection{Scenario $1$: $\chi$ DM candidate \cite{Bergstrom:1989jr}}
\begin{equation}
\label{eq:amp3bodyChi}
{1\over 4} \sum_{\rm spin}|{\cal M}_\chi|^2=\frac{4 \, \pi \,\alpha
\,y_\chi^4}{M_\chi^2 } \, \frac{ 4 (1 - y) (2 + 2 x^2 + 2 x( y-2) - 2 y
+ y^2) }{(1 -r_\chi^2 - 2 x)^2 (3 +r_\chi^2 - 2 x - 2 y)^2}
\end{equation}
\subsubsection{Scenario $2$: $N$ DM candidate \cite{Barger:2011jg}}
\begin{equation}
\label{eq:amp3bodyN}
{1\over 4} \sum_{\rm spin}|{\cal M}_N|^2=\frac{ \, \pi \,\alpha
\,g_N^4}{M_N^2 }\left(2 + {1\over r_N^2}\right)^2 \, \frac{ 4 (1 - y) (2 + 2 x^2 + 2 x( y-2) - 2 y
+ y^2) }{(1 -r_N^2 - 2 x)^2 (3 +r_N^2 - 2 x - 2 y)^2}
\end{equation}
\subsubsection{Scenario $3$: $S$ DM candidate \cite{Barger:2011jg,Toma:2013bka,Giacchino:2013bta}}
\begin{equation}
\label{eq:amp3bodyS}
|{\cal M}_S|^2=\frac{32 \pi\, \alpha \,y_S^4 }{ M_S^2}\,\frac{ 4 (1 - y) (2 + 2 x^2 + 2 x ( y-2) - 2 y + y^2) }{
(1 - r_S^2 - 2 x)^2 (3 + r_S^2- 2 x - 2 y)^2}
\end{equation}
\bigskip
The acute reader will have noticed that, for fixed $r_{\chi,N,S}$, the
dependence of the Bremsstrahlung cross sections into gamma rays is
precisely the same in
Eqs.~(\ref{eq:amp3bodyChi},\ref{eq:amp3bodyN},\ref{eq:amp3bodyS}),
which implies that all three scenarios have the same spectral
signature \cite{Barger:2011jg}. Notice that we have already checked
this dependence for both scenarios $1$ and $3$
in~\cite{Giacchino:2013bta}. The Bremsstrahlung cross section in the
scalar case was also re-derived in~\cite{Toma:2013bka} and previously
obtained in the large mediator mass regime in~\cite{Boehm:2006df}.
\bigskip
An argument to explain this conclusion has been advanced in
\cite{Barger:2011jg} based on effective operators. First there should
be no distinction between scenarios $1$ and $2$ since they have
the same initial and final states. As above, the bi-linear operator
corresponding to the initial states is $\bar \chi \gamma_5 \chi$. Then
they have shown that Bremsstrahlung corresponds to the effective
operator
\begin{equation}
{\cal O}_\chi \sim \bar\chi \gamma_5 \chi\, \left(\partial_\mu \bar \psi_R \gamma_\nu \psi_R + \bar \psi_R \gamma_\nu \partial_\mu \psi_R\right) \tilde F^{\mu\nu}\, ,
\label{eq:effOggC}
\end{equation}
where $\tilde F^{\mu\nu}$ is the dual of $F^{\mu\nu}$. In scenario
$3$, the initial state corresponds to $S^2$ and the effective
coupling is given by
\begin{equation}
{\cal O}_S \sim S^2\, \left(\partial_\mu \bar \psi_R \gamma_\nu
\psi_R + \bar \psi_R \gamma_\nu \partial_\mu \psi_R\right) F^{\mu\nu}
\label{eq:effOggS}
\end{equation}
The only difference amounts to exchanging the role of the $\vec E$ and
the $\vec B$ of the photon and so the spectra are the same (up to
normalization). This argument is essentially based on a rephrasing of
the exact result in terms of effective operators, and thus is not {\em
per se} an explanation, but it may complemented as follows. Notice
that the effective couplings correspond to respectively a dimension 9
and 8 operator, but there are other, {\em a priori} independent,
operators. Incidentally a classification of all dimension 8 operators
contributing to photon Bremsstrahlung for both Majorana and scalar DM
has been given in \cite{DeSimone:2013gj}: for Majorana DM there are 5
operators, out of which 3 are CP even, while in the scalar case, there
are 7 operators, with 4 being CP even. Remarkably, while these
operators lead to different spectra for the emission of $W$ and $Z$,
the spectra for emission of gamma rays are precisely the same (see
Eqs.~(3.8)-(3.10) and Eqs.~(3.24)-(3.27) in
\cite{DeSimone:2013gj}). Moreover, it can be easily checked that they
have the same $x$ and $y$ dependence as the numerator of
Eqs.~(\ref{eq:amp3bodyChi}),~(\ref{eq:amp3bodyN}) and
(\ref{eq:amp3bodyS}) (after some obvious change of variables). This, in
passing, means that the dimension 9 operator for Majorana DM of
Eq.~(\ref{eq:effOggC}) must be equivalent to (a combination of) the
dimension 8 operators studied in \cite{DeSimone:2013gj}. While the
$r$ dependence of the denominator in
Eqs.~(\ref{eq:amp3bodyChi})-(\ref{eq:amp3bodyS}) can not be obtained from
an effective approach, it may be easily inferred from the propagators
of the intermediate particles. Hence the effective operator argument that Bremsstrahlung
from scalar and Majorana DM should have the same spectra is actually robust.
\bigskip
\begin{figure}
\includegraphics[width=9cm]{plots/sv3all.png}
\caption{Total 3-body cross sections (gamma ray Bremsstrahlung emision). Same conventions as in Fig.~\ref{fig:sv2all}.}
\label{fig:sv3all}
\end{figure}
Now it remains that the normalizations of the spectra are distinct, or
at least they are for the scalar, for scenarios $1$ and $2$ are quite
similar. For equal $M_{DM}$, couplings and $r$, the amplitude in the
gauge case is larger by a factor of
$$
{1\over 2}\left(2+ {M_N^2\over M_{W^\prime}^2}\right) \sim 1
$$
where the first term is from the 2 transverse polarization modes
and the second term from the longitudinal one. As emphasized in
\cite{Toma:2013bka,Giacchino:2013bta}, all things being taken to be
the same ({\em i.e.} DM and mediator masses and the couplings), the 3-body cross section is larger by a factor of 8 in
scenario $3$ compared to scenario $1$. This, together with the
relative suppression of the 2-body cross section leads to an enhanced
gamma ray feature in the scalar case compare to the Majorana cases.
\bigskip
The intermediate conclusion is that scenarios $1$ and $2$ are
essentially identical. They share the same spectra, with quasi the
same parametric dependence in the couplings and mass of DM and of the
mediator of the normalization of the 2 and 3-body processes. The
scalar case is distinct, in the sense that if the relic abundance is
thermal and is fixed by the 2-body annihilation process, then the
signal is stronger for the scalar case (by a factor which may be as
large as 2 orders of magnitude)
\cite{Toma:2013bka,Giacchino:2013bta}. If we relax the latter
constraint (for instance if the abundance is fixed by another
process), then the 3 scenarios become indistinguishable. It is thus of
interest to check whether other spectral features may help to lift
this degeneracy. The spectrum of gamma rays from say, $\pi_0$, being
featureless we focus on gamma ray lines. In \cite{Giacchino:2013bta}
we have tentatively included the features from annihilation into two
monochromatic gamma rays. However the one-loop cross sections, in
particular that relevant for scenario $3$, reported in the
literature has some peculiarities. We have thus felt compelled to
reanalyze this problem. Our results are presented in the next section.
\begin{figure}
\includegraphics[width=9cm]{plots/sv2gllallM.png}
\caption{Ratios of the 3-body (gamma ray Bremsstrahlung emision) to 2-body annihilation cross sections. The 2-body cross section is thermally averaged over velocities at the time of freeze-out, relevant to determine the relic abundance of DM. }
\label{fig:sv2gllallM}
\end{figure}
\section{SPECTRAL ENERGY DENSITY: GAMMA RAY LINES}
\label{sec:annih-into-gamma}
We consider the annihilation of non-relativistic DM into two on-shell gamma rays,
$$
DM(p_1) + DM(p_2) \rightarrow \gamma(\epsilon_1,k_1) + \gamma(\epsilon_2,k_2)
$$ with $p_{1, 2}$ and $k_{1,2}$ the momenta, $\epsilon_{1,2}$ the
polarization vectors, $p_1 + p_2 = k_1 + k_2$, and $\epsilon_1 \cdot
k_1 = \epsilon_2\cdot k_2 = 0$.
In the 3 scenarios we consider the amplitude for this process is
represented by a sum of box Feynman diagrams. Although the individual
diagrams may be infinite, the total amplitudes are finite and are
moreover non-vanishing in the s-wave. These features allow
substantial simplifications in the calculation of the Feynman
diagrams. In particular we can calculate the amplitude in the limit
$p_1 = p_2 = (M_{DM},\vec 0)$. In this limit, Gram determinants built on
the external momenta or their relevant linear combinations may be
vanishing. For instance
$$
\left\vert\begin{array}{ccc}
p_1^2 & p_1 \cdot p_2 & p_1 \cdot k_1\\
p_1\cdot p_2 & p_2^2 & p_2 \cdot k_1\\
p_1\cdot k_1 & p_2 \cdot k_1 & k_1 \cdot k_1\\
\end{array}
\right\vert = 0
$$ The standard Passarino-Veltman reduction of tensorial loop
integrals breaks down when Gram determinant are zero
\cite{Passarino:1978jh}. This is in particular an issue\footnote{To be fair we should mentioned that we have managed to obtain
numerical results from FormCalc that are in very good agreement with our
analytical expressions. } for automated
tools like FormCalc and is also a source of numerical
instabilities in LoopTools
\cite{Hahn:1998yk}. Fortunately, by the very same token, one may use
the degeneracy between the momenta to express 4-point integrals in
terms of 3-point integrals, and also some 3-points integrals in terms
of 2-points integrals \cite{Stuart:1987tt} (see also
\cite{Bergstrom:1997fh,Bertone:2009cb}). In particular this implies
that no 4-points loop integrals appear in the final expression of the
amplitude, which may be expressed in terms of finite, 3-points loop
integrals that are much easier to handle, and in particular may be
given in terms of rather simple analytic functions. As in the previous
section, we discuss the 3 scenarios separately.
\subsubsection{Scenario $1$: $\chi \chi \rightarrow \gamma \gamma$}
This process has been calculated several times, starting from the
seminal works of \cite{Bergstrom:1988fp,Rudaz:1989ij,Bergstrom:1989jr}
in the context of supersymmetry. As is well known, in the limit
$M_{\tilde E} \gg M_\chi$, the amplitude for $\chi \chi \rightarrow
\gamma \gamma$ may be related to the chiral anomaly. A remarkable
consequence is that the amplitude does not vanish in the chiral limit,
$m_f \rightarrow 0$.
For our own sake and to check our procedures, we have
redone the calculation of this amplitude in the non-relativistic
limit, $v_\chi \rightarrow 0$, albeit with the help of FeynCalc
\cite{Mertig:1990an}. To do so we have used the trick of
\cite{Bergstrom:1997fh} which consists of projecting the initial
$\chi$ pair into a s-wave state, using
$$
{\cal O} = -{M_\chi \over \sqrt{2}}\gamma_5 \left(1 - \gamma^0\right)\,.
$$
In the chiral limit the amplitude involves a single 3-points scalar
loop integral. For reference, we give its expression following the
standard nomenclature of 3-point loop integrals (the $C_0$ functions of
Passarino and Veltman) and then explicitly in terms of analytic
functions. We have found
\begin{eqnarray}
\langle\sigma v\rangle_{\gamma\gamma}^\chi=\frac{y_\chi^4\alpha^2M_\chi^2}{64\pi^3} \left|C_0\left(-M_\chi^2,M_\chi^2 , 0, r_\chi^2 M_\chi^2,0, r_\chi^2M_\chi^2\right)\right|^2\,.
\end{eqnarray}
Using
\begin{eqnarray}
\label{eq:C0_1}
C_0(- M^2,M^2,0,r^2 M^2,0,r^2 M^2) &=& {-1\over 2 M^2} \int_0^1 {dx\over x} \log\left(\left\vert{-x^2 + (1-r^2)x + r^2\over x^2 - (1+r^2) x+ r^2}\right\vert \right)\\
&=& {-1\over 2 M^2}\left(Li_2\left({1\over r^2}\right) - Li_2\left({-1\over r^2}\right) \right)
\end{eqnarray}
we are in complete agreement with previous results and in particular
with \cite{Bergstrom:1997fh,Bern:1997ng} (we have also calculated the amplitude
for the case $m_f \neq 0$ - the result may be read from
\cite{Bergstrom:1997fh}). Of interest for us will be the following
limits,
\begin{eqnarray}
\langle\sigma v\rangle_{\gamma\gamma}^\chi&=&\frac{y_\chi^4\alpha^2\pi}{64^2 M_\chi^2} \quad \mbox{for } r_\chi=1 \quad\mbox{and }\quad
\langle\sigma v\rangle_{\gamma\gamma}^\chi\approx\frac{y_\chi^4\alpha^2}{64 \pi^3}{M_\chi^2\over M_{\tilde E}^4} \quad \mbox{for } r_\chi\gg 1
\end{eqnarray}
\subsubsection{Scenario $2$: $ N N \rightarrow \gamma \gamma$}
For this process we refer to the work of \cite{Bergstrom:1997fh} in
which a related amplitude has been calculated. Specifically, the
process considered there is the annihilation of two neutralinos into
two photons through a chargino and SM $W^\pm$ gauge boson
loops. Making simple adjustments in the couplings and the particle
content, and taking into account the coupling to nonphysical Goldstone
modes (as alluded to above, the calculation of \cite{Bergstrom:1997fh}
has been done in a 't~Hooft-Feynman non-linear $R_\xi$ gauge), we get
\begin{eqnarray}
\label{eq:NNgg}
\langle\sigma v\rangle_{\gamma\gamma}^{N}& =&\frac{g_N^4\alpha^2 M_N^2}{64 \pi^3}\left|4 C_0(4 M_N^2,0,0,r_N^2 M_N^2,r_N^2 M_N^2,r_N^2 M_N^2)\right.\nonumber \\
&& - \left.\left(2+\frac{1}{r_N^2}\right) C_0(-M_N^2,M_N^2,0, r_N^2 M_N^2, 0, r_N^2 M_N^2)\right|^2
\end{eqnarray}
with
\begin{eqnarray}
\label{eq:C0_2}
C_0(4 M^2,0,0,r^2 M^2,r^2 M^2,r^2 M^2) &=&{1\over 4 M^2} \int_0^1\frac{dx}{x}\log\left(\left|4\frac{x^2}{r^2}-4\frac{x}{r^2}-1\right|\right)\cr
&=& {-1\over 2 M^2} \left( \arctan{1\over \sqrt{r^2-1}}\right)^2
\end{eqnarray}
where the last equality holds for $r\geq 1$ and the other $C_0$ are as
defined in Eq.~(\ref{eq:C0_1}). The limiting behaviours are now given
by
\begin{eqnarray}
\langle\sigma v\rangle_{\gamma\gamma}^N&=&\frac{g_N^4\alpha^2\pi}{64^2 M_N^2}\frac{25}{4} \quad \mbox{for } r_N=1 \quad\mbox{and }\quad
\langle\sigma v\rangle_{\gamma\gamma}^N\approx\frac{g_N^4\alpha^2}{64 \pi^3}{M_N^2\over M_{W^\prime}^4} \quad \mbox{for } r_N\gg 1\,.
\label{eq:limN}
\end{eqnarray}
As an independent check of this result, we have verified that we can
recover the cross-sections derived many years ago by Crewther {\em et
al} \cite{Crewther:1981wh}, for the annihilation of SM neutrinos into
two photons, in a regime corresponding to our limit $r_N\gg 1$ in
Eq.~(\ref{eq:limN}) (see Appendix~\ref{sec:appendixA}).
\subsubsection{Scenario $3$: $ S S \rightarrow \gamma \gamma$}
\label{sec:Sgg}
For this case we have redone the full one-loop calculation, including
finite SM fermion mass $m_f \neq 0$ contributions. The full expression
is given in the Appendix~\ref{sec:appendix-B}. Here we just give the expression in the
chiral limit, as for the other two scenarios.
To derive the amplitude, we have essentially followed the path of
\cite{Bertone:2009cb}. We made use of FeynCalc and calculated
amplitude and cross section for S particles at rest ($v_S = 0$).
This amounts to evaluate a combination of one-loop 4-points tensor
integrals. Since the Gram determinant is zero for DM particles at
rest, a straightforward application of Passarino-Veltman reduction
does not work, so instead we used the approach of \cite{Stuart:1987tt}
to express directly the 4-points loop integrals as a linear
combination of 3-points scalar integrals.
Writing
\begin{equation}
\langle\sigma v\rangle_{\gamma\gamma} = \frac{2\,y_S^4\alpha^2}{64\pi^3M_S^2}|{\cal A}|^2
\label{eq:sv}
\end{equation}
at an intermediate step we got the following expression (for $m_f =0$)
\begin{eqnarray}
{\cal A}&=& 2 +\frac{2}{(1 - r_S^2)}B_0\left(M_S^2, 0, r_S^2M_S^2\right) - B_0\left(4 M_S^2, 0, 0\right)
- \frac{1 + r_S^2}{1 - r_S^2}B_0\left(4 M_S^2, r^2M_S^2, r_S^2M_S^2\right)\nonumber\\
&& + M_S^2\left(
- (1 + r_S^2)\left(C_0\left(M_S^2, M_S^2, 4 M_S^2, 0, r_S^2M_S^2, 0\right)
+C_0\left(M_S^2, M_S^2, 4 M_S^2, r^2M_S^2, 0, r_S^2 M_S^2\right)\right) \right.\nonumber\\
&&\left.
- 2C_0\left(-M_S^2, M_S^2, 0, r_S^2 M_S^2, 0, r_S^2 M_S^2\right)
+ 4r_S^2C_0\left(4M_S^2, 0, 0, r_S^2 M_S^2, r_S^2 M_S^2, r_S^2 M_S^2\right)\right)\,.\nonumber\\
\label{eq:Sgg-us}
\end{eqnarray}
Despite the presence of divergent 2-point integrals, this expression
is finite. Actually it may be simplified using the fact that some
3-point scalars integral, whose momentum arguments have a vanishing
Gram determinant, may be reduced further
\begin{eqnarray}
C_0\left(M_S^2, M_S^2, 4M_S^2, M_E^2, 0, M_E^2\right) &=& \frac{B_0\left(M_S^2, 0, M_E^2\right)- B_0\left(4M_S^2, M_E^2, M_E^2\right)}{ - M_E^2 + M_S^2}\\
C_0\left(M_S^2, M_S^2, 4M_S^2, 0, M_E^2, 0\right) &=& \frac{B_0\left(M_S^2, 0, M_E^2\right) - B_0\left(4M_S^2, 0, 0\right)}{ M_E^2 + M_S^2}\label{eq:subst}
\end{eqnarray}
Using this substitution, we get the following simple expression
\begin{equation}
{\cal A}= 2 - 2 M_S^2 C_0(-M_S^2, M_S^2, 0, r_S^2 M_S^2, 0, r_S^2 M_S^2) + 4 r_S^2 M_S^2C_0(4 Ms^2, 0, 0,r_S^2 M_S^2, r_S^2M_S^2, r_S^2M_S^2)\,.
\label{eq:Sgg-uss}
\end{equation}
where the two 3-points loop integrals are those given in
Eqs.~(\ref{eq:C0_1},\ref{eq:C0_2}).\footnote{Notice that our result differs from that reported in Eq.~(13) of \cite{Bertone:2009cb}. We believe that the discrepancy is just a mere misprint in the last line of their Eq.~(13), in which $C_0\left(M_S^2, 0, M_S^2, 0, r_S^2 M_S^2, r_S^2 M_S^2\right)$ should read $C_0\left(-M_S^2, 0, M_S^2, 0,r_S^2M_S^2, r_S^2M_S^2\right)$. Although virtually impossible to spot, this becomes pretty clear when one realizes that no combination of external momenta of the box Feynman diagrams may lead to the arguments of $C_0\left(M_S^2, 0, M_S^2, 0, r_S^2 M_S^2, 0, r_S^2 M_S^2\right)$. The erroneous expression is divergent at $r=1$, and thus potentially leads to a large signal for annihilation into gamma rays ~\cite{Tulin:2012uq}. On the contrary we found that the correct expression is regular at $r=1$.\label{fn:typo}}
Finally for $r_S =1$ and $r_S \gg 1$, we have respectively
\begin{eqnarray}
\langle\sigma v\rangle_{\gamma\gamma}^S&=& \frac{y_S^4\alpha^2}{8 \pi^3 M_S^2}\left( 1-\frac{\pi^2}{8}\right)^2\quad \quad\mbox{and }\quad
\langle\sigma v\rangle_{\gamma\gamma}^S\approx\frac{y_S^4\alpha^2}{18\pi^3 M_S^2 r^4} \hspace*{1cm}
\label{eq:SSgglimits}
\end{eqnarray}
\begin{figure}
\includegraphics[width=9cm]{plots/svggall.png}
\caption{Annihilation into 2 photons. Same conventions as in Fig.\ref{fig:sv2all}.}
\label{fig:svggall}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{plots/sv2ggallM.png}
\caption{Ratios of the cross sections into 2 photons to the 2-body annihilation into a fermion pair.}
\label{fig:sv2ggallM}
\end{figure}
\section{Discussion of results}
In this section we compare the salient features of the 3 scenarios
considered in the previous section. The thermally averaged 2-body
annihilations cross sections into two fermions at the time of
freeze-out (i.e. for averaged velocities taken as in
Eq.~(\ref{eq:vfo})) in the chiral limit are shown in
Fig.~\ref{fig:sv2all}. For convenience we have normalized all the
couplings to 1 and we took $M_{DM}=$ 100 GeV. This figure shows that the
cross section in the scalar DM scenario (scenario $3$) is
parametrically smaller than that of scenarios with Majorana DM
(scenarios 1 and 2): while scenarios 1 and 2 (for $\chi$ and $N$ DM) share
the same asymptotic behaviour for large $r$, the thermally averaged
cross section in scenario 3 (for $S$ DM) for, say $r=2$, is suppressed by
almost 2 orders of magnitude. This result is due to a combination of
the d-wave dependence of the cross section and of a distinct
dependence on $r$, see Eqs.~(\ref{eq:chichi}), (\ref{eq:NN})
and~(\ref{eq:SS}). If the relic abundance is fixed by the 2-body
process, a larger coupling is thus required for the $S$ than for the
$\chi$ or $N$ scenarios \cite{Toma:2013bka,Giacchino:2013bta}.
In Fig.~\ref{fig:sv3all}, we give the 3-body annihilation cross
section (for photon Bremsstrahlung emission) again for unit couplings
and $M_{DM} = 100$ GeV. The $r$-dependences are very similar, but the
signal from $S$ is comparatively larger compared to the $\chi$ and $N$
scenarios, respectively by a factor $8$ and $8/(1+1/2 r^2)^2$. This,
combined with the previous feature, implies that the Bremsstrahlung is
potentially much stronger in the $S$ scenario, although the spectra
are the same \cite{Toma:2013bka,Giacchino:2013bta}. This is
illustrated in Fig.~\ref{fig:sv2gllallM} which shows the ratio of the
3-body to 2-body cross sections (note that this ratio is independent
of the couplings and of the mass of the DM candidate). If the 2-body
annihilation into two fermions is the one driving the relic abundance
in the early universe, the Bremsstrahlung spectral feature is expected
to be much stronger in the scalar DM scenario ($\langle\sigma v\rangle_{\gamma ll}$ is almost 3
orders of magnitude larger than for Majorana DM for $r\sim 2$).
\begin{figure}
\includegraphics[width=9cm]{plots/sv3ggall.png}
\caption{Ratios of the Bremsstrahlung to 2 photons cross sections.}
\label{fig:sv3ggall}
\end{figure}
Turning to the annihilation into monochromatic gamma rays, we compare
the three cross sections in Fig.~\ref{fig:svggall}. The dependence in
the $\chi$ scenario is well-known, and has been reported many times in
the literature. In particular, it may be related to the chiral
anomaly, which implies that the cross section is non-vanishing even in
the chiral limit \cite{Rudaz:1989ij,Bergstrom:1989jr}. Since the
initial and final states are the same in scenarios $1$ and $2$, one
may expect a similar behaviour, which is confirmed by the calculations
and is illustrated in Fig.~\ref{fig:svggall}. Although this result is
implicit in the literature (we derive the $N$ cross section from the
results on the neutralino discussed in \cite{Bergstrom:1997fh}), we
are not aware of an explicit discussion in the framework of simpler
models, like that of \cite{Khalil:2009nb,Ma:2009tc}. At any rate, the
cross section is that given in Eq.~(\ref{eq:NNgg}). It is slightly
larger than in the $\chi$ scenario for $r$ close to 1 (by a factor of
$25/4$ at $r=1$), but asymptots to the same result at large $r$. The
large $r$ behaviour is also in agreement with the result of Crewther
{\em et al} for the annihilation of Dirac neutrinos into two photons
\cite{Crewther:1981wh}. Incidentally, the latter result has been
derived making use of the chiral anomaly, as in the early derivation
of the cross section in the $\chi$ scenario.
The behaviour of the cross section in scenario $3$ is more
puzzling. Having done the calculation using different approaches, and
also having obtained the same results as those reported in
\cite{Ibarraetal}, we are confident that the expression is correct. In
particular it is finite at $r=1$, albeit with a strange value compared
to the Majorana cases. The large $r$ behaviour is also completely
mundane, having the same dependence in $r$ as in scenarios $1$ and
$2$. It shows however peculiar feature, since it has a maximum around
$r=1.15$ and then a dip, with a zero near $r=0$. The origin of
this destructive interference is unclear, at least to us, although it
may be traced to amplitudes that correspond to Feynman diagrams with a
distinct number of heavy fermion propagators. It may be of academic
interest to investigate this phenomenon further, which may perhaps be
related to the distinct property of the tree level annihilation cross
section into fermion pairs in the $S$ scenario compared to the
Majorana cases. Another, perhaps not unrelated question is whether the
annihilation cross section of $S$ into two photons may be derived from
the trace anomaly.
This being said, we see that the $S$ cross section is larger by a
factor of $32/9 \approx 3.5$ than in the Majorana DM scenarios, which
have the same asymptotic behaviour. The annihilation of a scalar may
thus also lead to a relatively stronger signal into monochromatic
photons. This is shown in Fig.~\ref{fig:sv2ggallM}, which displays
the ratio of the cross section into 2 gamma rays to that into a
fermion pair. The rise as a function of $r$ of the signal in the $S$
case is due to the fact that the 2-body cross section into fermion
pairs scales like $r^{-6}$, see Eq.~(\ref{eq:SS}), while the
annihilation into 2 photons is $\propto r^{-4}$, see
Eq.~(\ref{eq:SSgglimits}). On the contrary the ratios asymptote to a
small constant, ${\cal O}(10^{-5})$, for the Majorana scenarios.
\begin{figure}
\includegraphics[width=9cm]{plots/LSpectra-m100-r20-SN.png}
\caption{Comparison of normalized photon spectra for scenarios 2 and 3 for $M_{DM} = 100$ GeV.}
\label{fig:LSpectra-m100-r20-SC}
\end{figure}
For completeness we also give in Fig.~\ref{fig:sv3ggall} the ratio of
3-body annihilation cross section (photon Bremsstrahlung emission) to
the one into two photons. The Bremsstrahlung signal becomes
relatively less prominent for large $r$, as the cross sections drop
like $r^{-6}$, but we see it is more dominant in the scalar than in
the Majorana scenarios. Hence both signals are stronger for the
scalar. The relative importance of the $\gamma \gamma$ signal compared
to the $\gamma ee$ one is however comparatively smaller in the scalar
case than in the Majorana DM cases. This may be seen directly in the
photon spectra of Fig.~\ref{fig:LSpectra-m100-r20-SC} where we compare
scenario 2 and 3 ($N$ and $S$) for $r=2$; scenario 1 ($\chi$) would
have essentially the same signature as scenario 2. The quantity $x dN/dx=
E_\gamma/\sigma v_{\gamma} d\sigma v_{\gamma i}/dE_\gamma$, with
$i=\gamma, ee$ and $\sigma v_{\gamma} = \sum_i\sigma v_{\gamma i}$,
denotes the normalized photon spectrum multiplied by the photon
energy. We see that for $r=2$ the dominant Bremsstrahlung feature gets
an extra contribution from the $\gamma \gamma$ and $\gamma Z$
lines. Notice that we have assumed an energy resolution of $\Delta
E/E=0.1$, and that, although we did not explicitly derived $\langle\sigma v\rangle_{\gamma Z}$
here, we follow the same procedure as in~\cite{Giacchino:2013bta} to
estimate the $\gamma Z$ contribution. We refer to
\cite{Toma:2013bka,Giacchino:2013bta} for some discussion of scenarios
$1$ and $3$. Clearly scenario $2$ should give a phenomenology similar
to that of the $\chi$ Majorana. Here we just provide a few benchmark
values (see Table \ref{tab:m100}).
\begin{table}[t]
\begin{tabular}{c|cc|cc|cc}
&\multicolumn{2}{c|}{Scenario $1$ ($\chi$)} &\multicolumn{2}{c|}{Scenario 2 ($N$)}&\multicolumn{2}{c}{Scenario 3 ($S$)} \\
\hline
&$r=$1.2&$r=2$&$r=1.2$&$r=2$&$r=1.2$&$r=2$\\
\hline
$\langle\sigma v\rangle_{\gamma ll}$&$6.1\, 10^{-28}$&$1.3\, 10^{-29}$&$1.1\, 10^{-27}$&$1.7\, 10^{-29}$&$4.9\, 10^{-27}$&$1.1\, 10^{-28}$ \\
$\langle\sigma v\rangle_{\gamma\gamma}$&$2.9\, 10^{-29}$&$3.1\,10^{-30}$& $4.3\, 10^{-29}$&$ 3.4\, 10^{-30}$&$4.6\, 10^{-29}$&$9.4\,10^{-30}$\\
\hline
\end{tabular}
\caption{Cross sections in units of cm$^3$/s for $M_{DM}=100$ GeV and
unit couplings}
\label{tab:m100}
\end{table}
\section{Conclusions}
In this work we have complemented the work of Barger {\em et al}, in
which 3 simple scenarios of DM with Bremsstrahlung of photons are
discussed. It also complements the phenomenological studies initiated
in \cite{Toma:2013bka} and \cite{Giacchino:2013bta} in which it has
been shown that a real scalar DM candidate $S$ interacting with light
SM fermions could give a strong Bremsstrahlung signal. Specifically we
have considered the radiative annihilation of three DM candidates into
two photons. This lead us to re-calculate the amplitude for $SS
\rightarrow \gamma \gamma$. Our result differs from expressions that
may be found in the literature, but the discrepancy is minor, being
likely due to a misprint, which however is difficult to spot without
actually doing the full calculation. Hence we believe that it was
useful to provide an independent check of the expression for $SS
\rightarrow \gamma \gamma$. We have compared the result to those
expected in the case of a Majorana DM, interacting with SM fermions
either through a heavy charged scalar particle (scenario 1) or through
a charged gauge boson (scenario 2). The main outcome, which complement
the conclusions drawn in \cite{Toma:2013bka} and
\cite{Giacchino:2013bta}, is that radiative processes are
significantly more relevant for the scenario with scalar DM than for
the Majorana cases. These results from a combination of factors. First
the fact that the annihilation cross section into fermion pairs is
d-wave suppressed in the chiral limit in the scalar DM case, and
second, the fact that the radiative cross section are parametrically
larger. These results may be directly inferred from the analytical
expressions given in the body of the paper, and from a glance at the
figures.
|
\section{Introduction}
The significant progress recently made in the development
of new superconducting materials stirred a growing interest
in determining their key parameters. One of the most
important characteristics that allow analyzing the superconductor
behavior for the synthesis of new superconducting
compounds is the electron-phonon interaction (EPI) function.
One of the major techniques employed to determine the
EPI function is Yanson's point-contact (PC) spectroscopy.
This method has worked well for measuring nonlinearities of
current-voltage characteristics (CVC) in the point contacts
consisting of metals and compounds in the normal state\cite{Naidyuk1, Khotkevich1}.
On the other hand, for many superconductors it is very difficult
to realize the ballistic regime of current flow during their
transition to normal state, which is necessary for providing
the spectral mode of operation of a point-contact in
Yanson's point-contact spectroscopy. This problem is most
pronounced when the compounds belonging to the new
classes of superconducting materials are studied. The pointcontact
studies of such materials often have to deal with a
much distorted surface layer, which limits the possibility of
determining the EPI parameters. An effective solution to this
problem is to use the point-contact characteristics measured
in the superconducting state \cite{Bobrov}.
\section{Basic theoretical conceptions}
\subsection{Inelastic contribution}
The theory of inelastic spectroscopy of the EPI in a
superconductor considers ballistic point contacts with the
dimensions $d $ smaller than all characteristic lengths\cite{Khlus1, Khlus2, Khlus3}:
$d\ll \xi (0)$, $l_i$, $v_F$/$\omega_D$, where $\xi (0)$ is the
superconducting coherence length, $l_i$ is the scattering length on impurities,
$l_{\varepsilon} \sim {v_{F} } \mathord{\left/ {\vphantom {{v_{F} } {\omega_{D} }}}
\right. \kern-\nulldelimiterspace} {\omega_{D} }$ is the energy mean free path at
the Debye energy.
We should emphasize an important point: despite the
fact that most of the nonequilibrium phonons are generated
in the banks of the contact and any scattering process of the
Andreev electrons on nonequilibrium phonons is effective,
the existing theories consider only scattering in the region of
maximum concentration of nonequilibrium phonons since
the probability of their reabsorption by electrons depends on
the concentration of phonons. This region corresponds to the
highest current density and has a size, as also in the case of
Yanson's point-contact spectroscopy, of the order of the contact
diameter \cite{Kulik}.
The first publication\cite{Bobrov} on the reconstruction of EPI functions
from the spectra of superconducting point contacts has
addressed the cases that, to a certain extent, go beyond the
predictions of the theory of inelastic spectroscopy of EPI in
superconductors\cite{Khlus1, Khlus2, Khlus3}. In these cases, scattering in the banks
played an important role in the formation of nonlinearities in
such point contacts. However, the contacts which satisfy the
theoretical model to the fullest extent have not been considered.
In this paper, we will fill this gap and also consider the
point contacts in which the elastic contribution to electronphonon
scattering should be taken into account.
At the heart of the inelastic point contact spectroscopy
of superconductors lies the study of nonlinear current-voltage
characteristics of the contacts arising due to the
inelastic scattering of nonequilibrium phonons on electrons
undergoing Andreev reflection.
In Yanson's point contact spectroscopy the EPI function is
\vspace{5pt}
$G_{pc} \left( {eV} \right)=-\frac{3R_{0} \hbar v_{F} }{32ed}\cdot \frac{d^{2}I}{dV^{2}}$,
\vspace{5pt}
i.e., it is proportional to the second derivative of the currentvoltage
characteristic.\cite{Kulik} At the same time in the inelastic PC
spectroscopy of superconductors,\cite{Khlus1, Khlus2, Khlus3} the EPI function is
proportional to the first derivative of the excess current (the difference
between the current-voltage characteristics in the
normal and superconducting states at the same voltage). For
$S-c-S$ contacts, the following expression has been obtained\cite{Khlus1}:
\begin{equation} \label{eq__1}
\begin{array}{l} {\frac{dI_{exc} }{dV} =-\frac{64}{3R} \left(\frac{\Delta L}{\hbar \bar{v}} \right)\left[G^{N} (\omega )+\frac{1}{4} G^{S} (\omega )\right]} \\ {\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \omega ={eV\mathord{\left/ {\vphantom {eV \hbar }} \right. \kern-\nulldelimiterspace} \hbar } } \end{array}
\end{equation}
$G^{N} (\omega )$ is the PC EPI function identical to that of the point contact in the normal
state, $G^{S} (\omega )$ is the superconducting PC
EPI function different from $G^{N} (\omega )$ by a form factor. In contrast
to the normal form factor, which determines the contribution
to the current due to electron-phonon collisions
accompanied by a change in the $z$-component of the electron
velocity, in the case of the superconducting form factor
which is included in $G^{S} (\omega )$, it is the electron-phonon
collisions associated with Andreev reflection processes in
the contact region, i.e., conversion of quasi-electron into
quasi-hole excitations, that contribute to the current. The relative
magnitude of the phonon contribution to the excess
current is of the order of
${d\cdot \omega _{D} \mathord{\left/ {\vphantom {d\cdot \omega _{D} v_{F} }} \right. \kern-\nulldelimiterspace} v_{F} } $,
for $eV\sim \omega _{D} $, i.e., it is small
if the condition
$d\ll {v_{F} \mathord{\left/ {\vphantom {v_{F} \omega _{D} }} \right. \kern-\nulldelimiterspace} \omega _{D} } $
is fulfilled.
An analogous expression for $S-c-N$ contacts is \cite{Khlus2}
\begin{equation}\begin{split}
\label{eq__2}
\left. {\frac{1}{R(V)}- {\frac{1}{R(V)}} } \right|_{\Delta =0}
= \\ = -\frac{32d\Delta }{3R\hbar }\left[ {\frac{1}{v_{F}^{(1)} }G_{1} \left(
\omega \right)+ \frac{1}{v_{F}^{(2)} }G_{2} \left( \omega \right)} \right]
\end{split}\end{equation}
For the second derivative of the CVC in $S-c-N$ point contacts,
the following expression has been obtained:
\begin{equation}
\label{eq__3}
\frac{1}{R} \frac{dR}{dV}=\frac{16ed}{3\hbar }
\sum\limits_{a=1,2} {\frac{1}{v_{F}^{(a)} } \int\limits_0^\infty
{\frac{d\omega }{\Delta } S\left( {\frac{\omega -eV}{\Delta }}
\right)G_{a} (\omega )} }
\end{equation}
$G_{a} (\omega )$ are the EPI functions for the normal and superconducting
metals forming the heterojunction, $S(x)$ is the smearing factor,
\begin{equation}
\label{eq__4}
S(x)=\theta (x-1)\frac{2\left( {x-\sqrt {x^{2}-1} } \right)^{2}}{\sqrt
{x^{2}-1} },
\end{equation}
where $\theta (x)$ is the Heaviside theta-function. Thus, for $T\to 0$,
the resolution is determined by the value of $\Delta $.
From expression \eqref{eq__3}, given the relation between the derivative
of CVC and the PC EPI function, it can be written as
\begin{equation}
\label{eq__5}
\tilde{{g}}_{pc}^{S} =\int\limits_0^\infty {\frac{d\omega }{\Delta }S\left(
{\frac{\omega -eV}{\Delta }} \right)g_{pc}^{N} (\omega )}
\end{equation}
As a model $g_{pc}^{N} (\omega )$, we will take the EPI function of Cu-Sn
heterojunction reconstructed from its spectrum in the normal
state.
The calculation results obtained using \eqref{eq__5} in Fig.~1.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig1.eps}
\caption[]{${\rm g}_{{\rm PC}}^{{\rm N}} $ is the EPI function of Sn-Cu point
contact reconstructed from
the spectrum shown in Fig.\ref{Fig2} $\tilde{{\rm g}}_{{\rm PC}}^{{\rm S}} $
is the theoretically predicted point-contact EPI function upon transition into the
superconducting state (Eq.\eqref{eq__5}, see text for more detail) (a);
difference curve (b); integral of the difference curve, point-contact EPI function
obtained from the first derivative of the excess current
${\rm g}_{{\rm PC}}^{{\rm S}} $ in comparison with ${\rm g}_{{\rm PC}}^{{\rm N}} $.
For convenience of comparison, the maximum values of the curves are set the same.}
\label{Fig1}
\end{figure}
In comparison to the original $N$-curve, the $S$-curve
exhibits a shift of the EPI maxima towards lower energies by
the magnitude of the gap $\Delta $. Moreover, its amplitude is
somewhat smaller than the amplitude of the initial curve due
to an additional broadening by the smearing factor S, Eq.\eqref{eq__4}.
As already mentioned, in the superconducting state, the
EPI spectrum should appear in the first derivative of the
excess current. Indeed, if we subtract the initial N-curve
from the S-curve, we obtain the S-N curve: $\tilde{g}_{pc}^{S} -g_{pc}^{N} $.
As follows from Eq.\eqref{eq__2}, the EPI function can be reconstructed
from the first derivative of the excess current:
\begin{equation}
\label{eq__6}
g_{pc}^{S} (eV)=\frac{1}{\Delta }\int\limits_0^{eV} {\left[
{\tilde{{g}}_{pc}^{S} (\omega )-g_{pc}^{N} (\omega )} \right]d\omega }
\end{equation}
It should be emphasized that $\tilde{g}_{pc}^{S} $ and $g_{pc}^{S} $
are different functions. The former one, given by Eq.\eqref{eq__5}, is proportional
to the second derivative of the CVC and reflects the transformation
of the spectrum (broadening and the shift of the phonon
peaks) upon transition of the heterojunction into the
superconducting state. The latter one, given by Eq.\eqref{eq__6}, see
also Eq.\eqref{eq__2}, is proportional to the first derivative of the
excess current and does not contain any additional broadening.
The position of the phonon maxima in the $g_{pc}^{S} (eV)$ is intermediate
between $\tilde{g}_{pc}^{S} $ and $g_{pc}^{N} $.
Note that for $S-c-S$ contacts,
the position of maxima in the EPI function reconstructed
from the first derivative of the excess current match that for
the normal condition.
\subsection{Elastic contribution}
The CVC of a point contact in which one or both electrodes
contain a superconductor with strong EPI comprises,
along with the above nonlinearities, an additional \textbf{elastic}
component of the current related to the frequency dependence
of the superconducting energy gap. This additional nonlinearity
arises due to the electron-phonon renormalization
of the energy spectrum of the superconductor and is manifested
as \textit{differential conductance maxima} in the region of
characteristic phonon energies in the first derivative of the
excess current, which are shifted to higher energies by the
magnitude of the superconducting energy gap \cite{Omel'yanchuk}.
Equation \eqref{eq__7}, which describes the first derivative of the
CVC in a point contact with direct conductivity, differs from
the corresponding expression \eqref{eq__8} for a tunnel junction,
\cite{Omel'yanchuk}
\begin{equation}
\label{eq__7}
\left( {\frac{dI}{dV}} \right)_{S-c-N} =\frac{1}{R_{0} }\left\{ {1+\left|
{\frac{\Delta \left( \varepsilon \right)}{\varepsilon +\sqrt {\varepsilon
^{2}-\Delta^{2}\left( \varepsilon \right)} }} \right|_{\varepsilon =eV}^{2}
} \right\}
\end{equation}
\begin{equation}
\label{eq__8}
\left( {\frac{dI}{dV}} \right)_{S-I-N} =\frac{1}{R_{0} }Re\left\{
{\frac{\varepsilon }{\sqrt {\varepsilon^{2}-\Delta^{2}\left( \varepsilon
\right)} }} \right\}_{\varepsilon =eV}
\end{equation}
This difference is due to Andreev reflection processes
leading to an excessive current in the region $eV~\gg~\Delta_0$.
Note that, unfortunately, the above equations do not
cover the most frequently encountered experimental situation-
point contacts with arbitrary transparency of the tunnel
barrier between the electrodes. In this respect, the
situation is similar to the attempts to determine the superconducting
energy gap prior to the BTK theory (Blonder,
Tinkham, and Klapwijk)\cite{Blonder}, which has provided a method for
determining the gap that takes into account an arbitrary barrier
transparency. It should be noted that for inelastic superconducting
spectroscopy this gap has been filled by Ref. \cite{Khlus3}.
Obviously, for point contacts with low barrier transparency,
it is the elastic contribution that is predominant due to
the suppression of the excess current. It has been noted in
Ref.\cite{Omel'yanchuk} that for ballistic contacts, CVC nonlinearities of elastic
origin may be comparable with the inelastic contributions
in point contacts. For the point contacts with direct conductivity
or high barrier transparency, the ratio between the elastic
and inelastic contributions is determined by the
parameters of the superconductor. As follows from Ref. \cite{Wolf},
the expected elastic contribution to the spectrum is proportional
to {$\sim{}$($T_C$/$\theta_D)^2$}, where $\theta_D$ is the Debye temperature.
Table~1 shows the elastic contributions for a number of
superconductors normalized by that of lead $\delta_{rel}$, which have
been studied in the previous publication\cite{Bobrov} and in the present
paper. The data from Ref.\cite{Wolf} were taken as a basis. We normalized
the data by the elastic contribution of lead since it
has the highest elastic contribution among the considered
superconductors. Table~1 also shows the energy gap and
superconducting transition temperature.
\begin{table}[]
\caption[]{Estimated elastic spectral contribution normalized by that of Pb, $\delta_{rel}$,
superconducting gap and transition temperature for several superconductors (SC).}
\small \begin{tabular}{|p{28pt}|l|l|l|l|l|l|l|}
\hline
SC & Pb & In & Sn & Ta & Al & NbSe$_2$ & MgB$_2$ \\
\hline
$\delta_{rel}$ & 1 & 0.21 & 0.078 & 0.063 & 0.00168 & 0.023 & 0.24 \\
\hline
$\Delta_0$,mV&
1.365&
0.525&
0.575&
0.7&
0.17&
1.07\textdiv2.48&
1.8\textdiv7.4 \\
\hline
T$_C$, K&
7.2&
3.415&
3.722&
4.47&
1.181&
7.2&
39 \\
\hline
\end{tabular}
\label{Table1}
\end{table}
Recall that for $S-c-N$ point contacts, the \textbf{inelastic} superconducting
contribution to the spectrum manifests itself as
\textit{differential \textbf{resistance} maxima} in the first derivative of the
excess current, which are shifted to lower energies by the
distance of the order of the gap. On the other hand, there is
no such shift for $S-c-S$ point contacts. Therefore, \textit{these contributions
oppose each other} and, if their magnitude is similar,
might attenuate each other. Since the inelastic
contribution is proportional to the magnitude of the excess
current, i.e., $\Delta $, and the elastic contribution is proportional to
($\Delta /E)^2$ (Ref.\cite{Wolf}) (see also Eq.\eqref{eq__7}),
starting from a certain value of $\Delta $, the elastic contribution dominates.
It can be expected that the positions of the maxima in
the EPI functions reconstructed from $S-c-S$ and $S-c-N$ point
contacts, as well as that for weakly coupled superconductors,
will be different.
Both for tunneling and point contacts with direct conductivity,
the elastic contribution to the spectrum does not
explicitly contain the EPI function g($\omega $). However, it can
be reconstructed by inverting the Eliashberg equations
(similar to the case of Rowell-McMillan's elastic tunneling
spectroscopy \cite{Rowell1}).
\section{Reconstruction of the EPI functions}
\subsection{Sn-based point contacts}
Fig. \ref{Fig2} (a) shows the spectra of Sn-Cu point contacts in
the normal and superconducting states.\cite{Yanson} Markedly lower
level of background in the superconducting spectrum and the
presence of the gap peak in the region of low energies
requires, similar to the previous work,\cite{Bobrov} that the background
$B$ is subtracted from the difference curve $S-N$. The difference
curve with the background subtracted, $S-N-B$, is very
close to the theoretically calculated curve $\tilde{g}_{pc}^{S} -g_{pc}^{N} $
in Fig. \ref{Fig1}. Finally, the lower part of the figure shows a comparison of
the PC EPI function reconstructed from the spectrum in the
normal state (curve $g_{pc}^{N} $) and the EPI function reconstructed
from the superconducting contribution to the spectrum
(curve~$g_{pc}^{S}$). For convenience of comparison, the curves are
plotted with equalized amplitude. There is excellent agreement
with the theoretically predicted behavior of the superconducting
EPI function--the shift of the maxima to lower
energies by the distance of the order of the gap. A slight mismatch
in the shape of the curve reconstructed from the experimental
data as compared to the calculated EPI function
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig2.eps}
\caption[] {EPI spectra of Sn-Cu point contact in the normal and superconducting
states. Superconductivity is suppressed by a magnetic field (a); the difference
between the spectra in the superconducting and normal states and
the estimated background curve (b); difference curve (after subtracting the
background B) (c); point-contact EPI function reconstructed by integrating
the curve in panel (c) versus the EPI function of the normal state (d).}
\label{Fig2}
\end{figure}
in the region of large displacements can be related to certain
arbitrariness in defining the background or an increasing
contribution from the peripheral regions of the point contact.
The latter can occur due to the increasing concentration of
nonequilibrium phonons in these regions caused by decreasing
the electron energy relaxation length in the vicinity of
the Debye energy.
As already noted, in the case of $S-c-S$ contact, the EPI
function reconstructed from the first derivative of the excess
current exhibits the same position of the maxima as the EPI
function of the normal state. Although an expression similar
to Eq.\eqref{eq__3} describing the transformation of the second derivative
of CVC upon the transition of electrodes into the $S$-state
has not been given in Ref.\cite{Khlus1}, from the similarity of the
expressions \eqref{eq__1} and \eqref{eq__2}, we can assume that the algorithm
used for $S-c-N$ point contacts can be employed here as well.
Fig.\ref{Fig3}(a) shows a set of the second derivatives of CVC
obtained in the normal and superconducting states,\cite{Kamarchuk1} and
Fig.\ref{Fig3}(b) displays the spectral contribution associated with
superconductivity as well as the estimated background
curves.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig3.eps}
\caption[] {Point-contact spectra of Sn in the normal and superconducting state,
adopted from Ref.\cite{Kamarchuk1}. $H = 0$ (a).
Superconducting contribution to the spectrum at different
temperatures and the estimated background curves (b).
$T = 1.8 K$: $T/T_C = 0.48$, $\Delta = 0.96\Delta_0$;
$T = 2.35 K$: $T/T_C = 0.63$, $\Delta = 0.89\Delta_0$,
$T = 2.8 K$: $T/T_C = 0.75$, $\Delta = 0.78\Delta_0$,
$T = 3.5 K$: $T/T_C = 0.94$, $\Delta = 0.41\Delta_0$.}
\label{Fig3}
\end{figure}
Fig.\ref{Fig4} shows the difference curves after background
subtraction. Despite the fact that, unlike the previous case
for the Sn-Cu point contact, the temperatures of the normal
and superconducting states are not the same, the reconstruction
of the EPI functions from the superconducting spectral
contribution (Fig.\ref{Fig5}) is quite satisfactory.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig4.eps}
\caption[] {Superconducting contribution to the point-contact spectra of Sn at
different temperatures after subtracting the background curves (Fig.\ref{Fig3}).}
\label{Fig4}
\end{figure}
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig5.eps}
\caption[] {Point-contact EPI functions of Sn, which were reconstructed from
the difference curves shown in Fig.\ref{Fig4}.}
\label{Fig5}
\end{figure}
Minor variations in
the shape of the curves can be easily explained by certain
arbitrariness in defining the background. Even for temperatures
near $T_C$, the reconstructed curve matches the normal
state spectrum quite satisfactory and the agreement can be
further improved by a better choice of the background curve.
\subsection{Al-based point contacts}
Aluminum has a relatively low superconducting transition
temperature and a small value of the superconducting
energy gap (Table~1). This means a small superconducting
contribution to the spectrum. Together with the inevitable
inaccuracies arising when scanned experimental curves are
digitized, this leads to a relatively low accuracy of the difference
curve obtained. Nevertheless, the curves shown in
Fig.\ref{Fig6} (similar to the curves published in Ref.\cite{Chubov})
demonstrate that the normal and superconducting EPI functions
match each other sufficiently well.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig6.eps}
\caption[] {EPI spectra of an Al-Al point contact in the normal and
superconducting states. $T/T_C = 0.68$, $\Delta = 0.85\Delta_0$.
Superconductivity is suppressed by a magnetic field (a).
The difference between the superconducting and normal
spectra and the estimated background curve (b). Difference curve (after
background subtraction) (c). Point-contact EPI function reconstructed by
integrating the difference curve in panel (c) versus the EPI function of the
normal condition (d).}
\label{Fig6}
\end{figure}
\subsection{Pb-based point contacts}
Lead has highest elastic contribution of known superconductors
(Table~1). In Ref.\cite{Khotkevitch2}, the second derivatives of CVC
for Pb-Ru heterojunctions have been measured in both the
superconducting $S$ and normal $N$ states (Fig.\ref{Fig7}).
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig7.eps}
\caption[] {EPI spectra of a Pb-Ru point contact in the normal and superconducting
states. Superconductivity is suppressed by a magnetic field (a). The
difference between the superconducting and normal spectra and the estimated
background curve. The dashed line shows the theoretically calculated
elastic contribution to the spectrum (see text) (b). Difference curve (after
background subtraction) (c). Point-contact EPI function reconstructed by
integrating the difference curve in panel (c) versus the EPI function of the
normal condition (d).}
\label{Fig7}
\end{figure}
The EPI spectrum
of ruthenium does not overlap in energy with the spectrum
of lead and therefore was not taken into account. The
intensity of the reduced spectrum of lead is close to the maximum
for a symmetric heterojunction (0.4 of the maximum intensity
for a homojunction). Full contribution to the spectrum
associated with superconductivity (the difference curve $S-N$),
shown in Fig.\ref{Fig7}(b), is quite different from similar contributions
to the spectrum in metals with weak electron-phonon interaction
and does not allow to restore the EPI function using the
methods previously employed, in particular, subtraction of a
smooth background. The figure also shows the second derivative
$d^2V/dI^2(eV)$ of the elastic superconducting contribution to
the spectrum ("theory" curve), which was obtained from the
$dI/dV$ dependence found from Eq.\eqref{eq__7} by numerical differentiation.
When calculating $dI/dV$, the tables of real and imaginary
parts of $\Delta $($\varepsilon $) \cite{Rowell2} obtained from
tunneling experiments were used.
Although the calculated second derivative of the elastic
contribution ("theory" curve) is similar to the difference curve
$S-N$, there are notable differences, especially at high energies.
As already mentioned, the elastic superconducting contribution
manifests itself as maxima of the differential
conductance in the region of characteristic phonon energies in
the first derivative of the excess current. However, the difference
curve $S-N$ contains not only the elastic contribution, but
also inelastic one, and, apparently, in the same way as in the
superconductors with weak coupling, also additional nonlinearity
which is not accounted by the theory and is what we
call a superconducting background. Therefore, to obtain the
EPI spectrum by integration, as was done previously, let us
try to subtract the background from the difference curve $S-N$
using the same rules as in the case of weakly coupled superconductors.
\textit{After background subtraction, the areas under the
curve above and below the abscissa should be the same; for
energies above the Debye energy, the curve obtained after
background subtraction must be zero.}
The obtained background B is shown in Fig.\ref{Fig7}(b) as a
dashed curve, and the resulting curve after background subtraction,
$S-N-B$ is displayed in Fig.\ref{Fig7}(c). Fig.\ref{Fig7}(d) shows the
EPI $N$ function reconstructed from the spectrum in the normal
state and the EPI $S$ function obtained by integrating the curve
$S-N-B$. Here it is necessary to emphasize the following
points. First, as follows from the theoretical predictions, the
maxima of the function correspond to the maxima of the differential
conductance and not to the maxima of resistance as
in the case of the superconductors with weak EPI. After integration
the curve is inverted. Secondly, the positions of the
phonon peaks in the both curves coincide, and there is no shift
of the phonon peaks in the restored EPI function for S-c-N
point contacts. And finally, the background curve is not
smooth and monotonic but is similar in shape to the theoretically
calculated elastic contribution marked as "theory". Note
that for $S-c-S$ contacts, as discussed below, there is a shift of
the phonon maxima in the reconstructed EPI function to
higher energies by the distance of the order of $\Delta $.
Fig.\ref{Fig8} shows the spectra of a Pb-Pb point contact in the
normal and superconducting states \cite{Kamarchuk2}.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig8.eps}
\caption[] {EPI spectra of a Pb-Pb point contact in the normal and superconducting
states. Superconductivity is suppressed by a magnetic field (a). The
difference between the superconducting and normal spectra and the estimated
background curve (b). Difference curve (after background subtraction)
(c). Point-contact EPI function recovered by integrating the curve in
panel (c) versus the EPI function of the normal state (g).}
\label{Fig8}
\end{figure}
The curves were treated similar to the previous case. As anticipated above,
the position of the phonon peaks in the EPI function reconstructed
from a $S-c-S$ contact differs from that of a $S-c-N$
contact and is shifted toward higher energies by a distance of
the order of $\Delta $.
Finally, the data for a Pb-Sn point contact at temperatures
above $T_C$ for Sn are shown in Fig.\ref{Fig9}.\cite{Kamarchuk2}
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig9.eps}
\caption[] {EPI spectra of a $Pb-Sn$ point contact in the normal and superconducting
states. Measurement temperature is above $T_C$ of Sn
($T/T_C = 0.58$, $\Delta = 0.92\Delta_0$).
Superconductivity is suppressed by a magnetic field (a). The difference
between the superconducting and the normal spectra and the estimated
background curve (b). Difference curve (after background
subtraction) (c). Point-contact EPI function reconstructed through integrating
the difference curve in panel (c) versus the EPI function of the normal
state (d).}
\label{Fig9}
\end{figure}
In this case, the tin contribution is considerably higher than that of ruthenium,
and it overlaps in energy with the spectrum of lead.
Thus it should be necessarily taken into account in data processing.
Since this is an $S-c-N$ point contact, there is no shift
of the phonon peaks to higher energies in the EPI function
reconstructed from the superconducting contribution.
\subsection{In-based point contact}
In the case of lead, the superconducting contribution to
the spectrum associated with EPI is very large and, as can be
seen in Fig.\ref{Fig7}, its amplitude exceeds non-linearity in the
normal state. Indium is intermediate in EPI strength and, as
follows from the table, exhibits a five-fold smaller elastic
contribution to the spectrum compared to lead, but 2.7-fold
higher than tin. At the same time, the superconducting transition
temperature and the gap are only slightly ($\sim8\%$) less
than those of tin, so the inelastic contribution to the spectrum
must be very close for these metals. Since, as noted above,
the elastic and inelastic contributions counteract each other,
in the case of indium they should, to a large extent, weaken
each other. Fig.\ref{Fig10} shows the spectra of indium\cite{Kamarchuk1}.
\begin{figure}[t]
\includegraphics[width=8cm,angle=0]{Fig10.eps}
\caption[] {EPI spectra of an In-In point contact in the normal and superconducting
states at different temperatures (a). The differences between the
superconducting and normal spectra, as well as the estimated background curve.
$T = 2.1 K$: $T/T_C = 0.62$, $\Delta = 0.89\Delta_0$;
$T = 2.9 K$: $T/T_C = 0.85$, $\Delta = 0.63\Delta_0$;
$T = 3.2 K$: $T/T_C = 0.94$, $\Delta = 0.41\Delta_0$ (b).
Difference curve (after background subtraction) (c). Point-contact EPI function
reconstructed by integrating the difference
curve (c) versus the EPI function of the normal state (d).}
\label{Fig10}
\end{figure}
As can be seen in the figure, the contribution to the spectrum associated
with superconductivity in indium is very small and of elastic
nature. Unlike tin, where it was possible to restore the spectrum
from the superconducting contribution in a wide temperature
range (Figs.\ref{Fig3} and \ref{Fig4}), for indium, this was possible
only at the lowest temperature.
\section{Discussion}
In determining the function of electron-phonon interaction,
the traditional tunneling spectroscopy is limited to the
superconductors with strong coupling. At the same time
Yanson's point-contact spectroscopy focusses on metals in
the normal state. Inelastic superconducting point-contact
spectroscopy thus fills the gap and can be used to study
superconductors with weak coupling. Moreover, as follows
from the discussion of indium, the superconductors in which
the elastic and inelastic contributions are close in value are
the most complex objects since the elastic and inelastic
contributions counteract each other, which leads to a weakening
of the resulting contribution to the spectrum. When
evaluating the sign and magnitude of the expected effect in
such superconductors, first of all the transparency of the barrier
between the electrodes should be taken into account. A
situation is very likely to appear in which contacts with the
different transparency of barriers exhibit different positions
of the spectral maxima due to the predominance of elastic or
inelastic spectral contributions. Note that it is not so much
the magnitude of the elastic contribution to the spectrum as
the ratio between the elastic and inelastic contributions that
is important. For instance, as follows from the data in Table~1,
the elastic contribution to the spectrum in MgB$_2$ is slightly
larger than that in In. However, the resulting spectrum in
MgB$_2$ is inelastic\cite{Bobrov}. Here we should give attention to the
magnitude of the superconducting gaps in In and MgB$_2$. The
inelastic contribution is proportional to the energy gap,
which in MgB$_2$ is an order of magnitude larger than in In.
Even for superconductors with weak coupling there may
be deviations from the theoretical predictions. For instance
(see Table~1), the elastic contribution to the spectrum in Sn
is greater than that in Ta by $\sim12\%$. However, due to a
shorter electron energy relaxation length, the presence of
phonons with low group velocity results in a significant
influence of the contact region on the formation of the superconducting
contribution. As a result, sharpening of the
phonon peaks and other deviations from the theoretical predictions
are observed. Another typical example is 2H-NbSe$_2$,
a superconductor with covalent bonds between the atoms
within the layer and van der Waals forces between the
layers. Therefore, both the current spreading and dispersion
of phonons in the point contacts based on NbSe$_2$ are anisotropic.
This leads to a slower decrease of the concentration
of nonequilibrium phonons with increasing the distance
from the constriction, and thus, an increase of the superconducting
contribution to the spectrum. As can be seen in Fig.~3
in Ref.\cite{Bobrov}, this contribution is sufficiently large and only an
order of magnitude smaller than the gap peculiarity in the
spectrum.
To summarize, let us note that all of the observed deviations
from the theoretical predictions are, in varying degrees,
related to the influence of the contact region that requires further
theoretical and experimental studies. Moreover, point contacts
with a non-uniform distribution of impurities exhibiting
diffusion transport through the junction and ballistic banks
also require further studies. It should be noted that this situation
is most easily achieved for superconductors with covalent
bonds between atoms which have a rigid crystal lattice. For ordinary
metals, due to their ductility, the lattice distortions upon
forming point contacts could extend to the banks as well.
\section{Conclusions}
\begin{enumerate}
\item The EPI functions for Sn and Al were reconstructed from
the superconductivity-related contributions to the spectra of
point contacts based on these metals. The procedure of
reconstruction of the EPI functions produces similar results
across a wide range of temperatures. As follows from the
theory of superconductors with weak coupling, the superconducting
inelastic contribution to the spectrum manifests
itself as \textit{differential resistance maxima} in the first derivative
of the excess current in the range of characteristic phonon
energies. The position of these peaks coincides with the
phonon peaks observed in the normal state of $S-c-S$ point
contacts and, for $S-c-N$ point contacts, is shifted to \textit{lower}
energies by the value of the superconducting energy gap.
\item The EPI functions for Pb and In were reconstructed from
the superconductivity-related contributions to the spectra
of point contacts based on these metals. For superconductors
with strong coupling, the superconducting elastic
contribution to the spectrum manifests itself as \textit{differential
conductivity maxima} in the first derivative of the excess
current in the range of characteristic phonon energies.
The position of these peaks \textit{coincides} with the phonon
peaks observed in the normal state of $S-c-N$ point contacts
and, for $S-c-S$ point contacts, is \textit{shifted to higher energies}
by the value of the superconducting energy gap.
\end{enumerate}
|
\section{Introduction}
Predicting election winners is always an exciting activity: Who will
be the new president? Will the company merge with another one? Will
taxes be higher or lower? The goal of this paper is to establish the
computational complexity of a family of problems modeling a certain
type of winner-prediction problems.
Naturally, predicting winners is a hard task, full of
uncertainties. For example, we typically are not sure which voters
will eventually cast their votes or, sometimes, even which candidates
will participate in the election (consider, e.g., a candidate
withdrawing due to personal reasons). Further, typically we do not
have complete knowledge regarding each voters' preferences.
Nonetheless, elections are in everyday use both among humans
(consider, e.g., political elections, elections among companies'
shareholders, various polls on the Internet and social
media, or even sport events, where judges ``vote'' on who is the best
competitor) and among software agents (see, e.g., election
applications for planning in multiagent
systems~\cite{eph-ros:j:multiagent-planning}, for recommendation
systems~\cite{gho-mun-her-sen:c:voting-for-movies,bou-lu:c:chamberlin-courant,sko-fal-sli:c:multiwinner},
for web-search~\cite{dwo-kum-nao-siv:c:rank-aggregation}, or natural
language processing~\cite{kut-kit:j:nlp}) and,
thus, the problem of predicting election winners is far too important
to be abandoned simply because it is difficult.
In this paper, we focus on a variant of the winner-prediction problem
where we have complete knowledge regarding all possible candidates and
all eligible voters (including knowledge of their
preferences\footnote{Note that while full-knowledge assumption
regarding voters' preferences might seem very unrealistic, it is
standard within computational social choice literature, and in our
case can often be justified (e.g., election polls can provide a good
approximation of such knowledge).}), but we are uncertain as to
which candidates and which voters turn up for the actual election (see
Section~\ref{sec:related} for other approaches to the
problem). However, modelling uncertainty regarding both the candidate
set and the voter collection on one hand almost immediately leads to
computationally hard problems for typical election systems, and on the
other hand does not seem to be as well motivated as focusing on each
of these sets separately. Thus, we consider the following two
settings:
\begin{enumerate}
\addtolength{\itemsep}{-4pt}
\item The set of candidates is fixed, but for each possible subset
of voters we are given a probability that exactly these voters
show up for the vote.
\item The set of voters is fixed, but for each possible subset of
candidates we are given a probability that exactly these
candidates register for the election.
\end{enumerate}
The former setting, in particular, corresponds to political elections
(e.g., to presidential elections), where the candidate set is
typically fixed well in advance due to election rules, the set of all
possible voters (i.e., the set of all citizens eligible to vote) is
known, but it is not clear as to which citizens choose to cast their
votes. The latter setting may occur, for example, if one considers
software agents voting on a joint plan~\cite{eph-ros:j:multiagent-planning}. The set of agents participating in the
election is typically fixed, but various variants of the plan can be
put forward or dismissed dynamically. In either case, our goal is to
compute each candidate's probability of victory.
However, our task would very quickly become computationally
prohibitive (or, difficult to represent on a computer) if we allowed
arbitrary probability distributions. Thus, we have to choose some
restriction on the distributions we consider. Let us consider the
following example.
Let the set $C$ of candidates participating in the election be fixed
(e.g., because the election rules force all candidates to register
well in advance). We know that some set $V$ of voters will certainly
vote (e.g., because they have already voted and this information is
public\footnote{Naturally, in typical political elections such
information would not be public and we would have to rely on
polls. On the other hand, in multiagent systems there can be cases
where votes are public.}). The set of voters who have not decided to
vote yet is $W$. From some source (e.g., from prior experience) we
have a probability distribution $P$ on the number of voters from $W$
that will participate in the election (we assume that each equal-sized
subset from $W$ is equally likely to joint the election; we have no
prior knowledge as to which eligible voters are more likely to vote).
That is, for each $i$, $0 \leq i \leq \|W\|$, by $P(i)$ we denote the
probability that exactly $i$ voters from $W$ join the election (we
assume that $P(i)$ is easily computable). By $Q(i)$ we denote the
probability that a certain designated candidate $p$ wins provided that
exactly $i$, randomly chosen, voters from $W$ participate in the
election. The probability that $p$ wins is given by:
\[
P(0)Q(0) + P(1)Q(1) + \cdots P(\|W\|)Q(\|W\|).
\]
We use this formula to compute the probability of each candidate's
victory, which gives some idea as to who is the likely winner of the
election.
To compute $Q(i)$, we have to compute for how many subsets $W'$ of $W$
of size exactly $i$ candidate $p$ wins after adding the voters from
$W'$ to the election, and divide it by $\|W\| \choose i$. That is,
computing $Q(i)$ boils down to solving a counting variant of control
by adding voters problem. In the decision variant of this control
problem, introduced by Bartholdi, Tovey, and
Trick~\cite{bar-tov-tri:j:control}, we are given an election where
some voters have already registered to vote, some are
not-yet-registered, and we ask if it is possible to add some number of
these not-yet-registered voters (but no more than a given limit) to
ensure that a designated candidate wins. In the counting variant,
studied first in this paper, we ask how many ways are there to add
such a group of voters.
One can do reasoning analogous to the one presented above for the case
of control by deleting voters and for the case of control by
adding/deleting candidates. That is, our winner prediction problems,
in essence, reduce to solving counting variants of election control
problems. Our goal in this paper is, thus, to study the computational
complexity of these counting control problems.
Computational study of election control was initiated by Bartholdi,
Tovey, and Trick~\shortcite{bar-tov-tri:j:control}, and in recent
years was continued by a number of researchers (we discuss related
work in Section~\ref{sec:related}; we also point the reader to the
survey of Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:j:cacm-survey} for a more detailed
discussion of the complexity of election control). However, our paper
is the first one to study counting variants of control (though, as we
discuss in Section~\ref{sec:related}, the papers of Walsh and
Xia~\cite{wal-xia:c:lot-based} and of Bachrach, Betzler, and
Faliszewski~\cite{bac-bet-fal:c:counting-pos-win} are very close in
spirit).
It is well-known that election control problems tend to have fairly
high worst-case complexity. Indeed, we are not aware of a single
practical voting rule for which the decision variants of the four most
typical election control problems (the problems of adding/deleting
candidates/voters) are all polynomial-time solvable. Naturally, when a
decision variant of a counting problem is $\NPclass$-hard, then we cannot
hope that its counting variant would be polynomial-time solvable. In
effect, from the technical perspective, our research can be seen as
answering the following question: For which voting rules and for which
control types is the counting variant of the control problem no harder
than its decision variant? From the practical perspective, it is,
nonetheless, more important to simply have effective algorithms. To
this end, we follow Faliszewski et
al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences} and in addition
to the general setting where each voter can cast any possible vote, we
study the single-peaked case, where the voting is restricted to a
certain class of votes viewed as ``reasonable'' (intuitively, in the
single-peaked case we assume that the candidates are located on a
one-dimensional spectrum of possible opinions, each voter has some
favorite opinion on this spectrum, and the voters form their
preferences based on candidates' distances from their favorite
position). Single-peaked preferences, introduced by
Black~\cite{bla:b:polsci:committees-elections} over 50 years ago, are
often viewed as a natural (if somewhat simplified due to its
one-dimensional nature) model of how realistic votes might look like
(however, whenever one makes a statement of this form, one should keep
in mind the results of Sui, Francois-Nienaber, and
Boutilier~\cite{sui-fra-bou:c:single-peaked}). It turns out that for
the sinlge-peaked case, we obtain polynomial-time algorithms for
counting variants of all our control problems and all the rules that
we consider (except for the Approval rule, for which we have no
results in this case).
We mention that our model of winner prediction, based on counting
variants of control, is similar to, though more general than, the
model where we assume that each voter casts his or her vote with some
probability $p$ (universal for all the voters). We can simulate this
scenario in our model by providing an appropriate function $P$.
Specifically, if there are $n$ voters and the probability that each
particular voter $v$ casts his or her vote is $p$, then the
probability that exactly $i$ voters vote is given by the binomial
distribution:
\[
P(i) = {n \choose i}p^i(1-p)^{n-i}.
\]
On the other hand, our model does not capture the situation where each
voter $v$ has a possibly different probability $p_v$ of casting a
vote. We believe that this latter model deserves study as well, but we
do not focus on it in this paper.
The paper is organized as follows. First, in Section~\ref{sec:prelim},
we formally define elections and provide brief background on
complexity theory (focusing on counting problems). Then, in
Section~\ref{sec:problems}, we define counting variants of election
control problems and prove some of their general properties. Our main
results are in Section~\ref{sec:results}. There we study the
complexity of Plurality, $k$-Approval, Appoval, Condorcet rule, and
Maximin rule. For each of the rules, we consider the unrestricted case
and, if we obtain a hardness result, we consider the single-peaked
case (except for Approval, for which we were not able to obtain
results for the single-peaked case). Finally, we discuss related work
in Section~\ref{sec:related} and bring together our conclusions and
discuss future work in Section~\ref{sec:conclusions}.
\section{Preliminaries}
\label{sec:prelim}
\noindent\textbf{Elections.}\quad
An \emph{election} $E$ is a pair $(C,V)$ such that $C$ is a finite set
of candidates and $V$ is a finite collection of voters. We typically
use $m$ to denote the number of candidates and $n$ to denote the
number of voters. Each voter has a preference order in which he or
she ranks candidates, from the most desirable one to the most despised
one. For example, if $C = \{a,b,c\}$ and a voter likes $b$ most and
$a$ least, then this voter would have preference order $b > c >
a$. (However, under Approval voting, instead of ranking the candidates
each voter simply indicates which candidates he or she approves of.)
We sometimes use the following notation. Let $A$ be some subset of
the candidate set. Putting $A$ in a preference order means listing
members of $A$ in lexicographic order and putting $\overleftarrow{A}$
in a preference order means listing members of $A$ in the reverse of
the lexicographic order. For example, if $C = \{a,b,c,d\}$ then $a > C
- \{a,b\} > d$ means $a > c > d > b$ and $a > \overleftarrow{C -
\{a,b\}} > d$ means $a > d > c > b$.
Given an election $E = (C,V)$, we write $\N{E}(c,c')$ to denote the
number of voters in $V$ that prefer $c$ to $c'$. The function $N_E$ is
sometimes referred to as the \emph{weighted majority graph} of
election $E$.
In general, given a candidate set $C$, the voters are free to report
any preference order over $C$ (we refer to this setting as the
\emph{unrestricted case}). However, it is often more realistic to
assume some restriction on the domain of possible votes (for example,
in real-life political elections we do not expect to see many voters
who rank the extreme left-wing candidate first, then the extreme
right-wing candidate, and so on, in an interleaving fashion). Thus,
in addition to the unrestricted case, we also study the case of
\emph{single-peaeked} preferences of
Black~\cite{bla:b:polsci:committees-elections} (this domain
restriction is quite popular in the computational social choice
literature, and was already studied, e.g., by
Conitzer~\cite{con:j:eliciting-singlepeaked},
Walsh~\cite{wal:c:uncertainty-in-preference-elicitation-aggregation},
Faliszewski et al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences},
and others; see Section~\ref{sec:related}).
\begin{definition}\label{def:sp}
Let $C$ be a set of candidates and let $L$ be a linear order over
$C$ (we refer to $L$ as the societal axis). We say that a preference
order $>$ (over $C$) is single-peaked with respect to $L$ if for
each three candidate $a, b, c \in C$, it holds that
\[
(a \mathrel{L} b \mathrel{L} c \lor c \mathrel{L} b \mathrel{L} a) \implies (a > b \implies b > c)
\]
An election $E = (C,V)$ is single-peaked with respect to $L$ if each
vote in $V$ is single-peaked with respect to $L$. An election $E =
(C,V)$ is single-peaked if there is a societal axis $L$ such that
$E$ is single-peaked with respect to $L$.
\end{definition}
There are polynomial-time algorithms that given an election $E$ verify
if it is single-peaked and, if so, provide the appropriate societal
axis witnessing this
fact~\cite{bar-tri:j:stable-matching-from-psychological-model,esc-lan-ozt:c:single-peaked-consistency,bal-har:j:characterization-single-peaked}. Thus,
following
Walsh~\cite{wal:c:uncertainty-in-preference-elicitation-aggregation},
whenever we consider problems about single-peaked elections, we assume
that we are given appropriate societal axis as part of the input (if
it were not provided, we could always compute it.)\bigskip
\noindent\textbf{Voting Systems.}\quad
A \emph{voting system} (a \emph{voting rule}) is a rule which
specifies how election winners are determined. We allow an election to
have more than one winner, or even to not have winners at all. In
practice, tie-breaking rules are used, but here we disregard this
issue by simply using the unique winner model (see
Definition~\ref{def:control}). However, we point the reader
to~\cite{obr-elk-haz:c:ties-matter,obr-elk:c:random-ties-matter,obr-zic-elk:c:multiwinner-tie-breaking}
for a discussion regarding the influence of tie-breaking for the case
of election manipulation problems
(see~\cite{fal-hem-hem:j:cacm-survey,fal-pro:j:manipulation,bra-con-end:b:comsoc} for
overviews of the manipulation problem specifically and computational aspects
of voting generally).
We consider the following voting rules (in the description below we
assume that $E = (C,V)$ is an election with $m$ candidates and $n$
voters):
\begin{description}
\item[Plurality.] Under Plurality, each candidate receives a point for
each vote where this candidate is ranked first. The candidates with
most points win.
\item[$\bm{k}$-Approval.] For each candidate $c \in C$, we define
$c$'s $k$-Approval score, $\score{E}^k(c)$, to be the number of
voters in $V$ that rank $c$ among the top $k$ candidates; the
candidates with highest scores win. Note that Plurality is, in
effect, a nickname for $1$-Approval. (We mention that $k$-Veto means
$(m-k)$-Approval, though we do not study $k$-Veto in this paper).
\item[Approval.] Under Approval (without the qualifying ``$k$-''), the
score of a candidate $c \in C$, $\score{E}^a(c)$, is the number of
voters that approve of $c$ (recall that under Approval the voters do
not cast preference orders but 0/1 approval vectors, where for each
candidate they indicate if they approve of this candidate or
not). Again, the candidates with highest scores win. (Note that the
notion of single-peaked elections that we have provided as
Definition~\ref{def:sp} does not apply to preferences specified as
approval vectors; Faliszewski et
al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences} have provided
a natural variant of single-peakedness for approval vectors, but
since we did not obtain results for this case, we omitted this
definition.)
\item[Condorcet.] A candidate $c$ is a \emph{Condorcet winner} exactly
if $\N{E}(c,c') > \N{E}(c',c)$ for each $c' \in C - \{c\}$. A
candidate $c$ is a \emph{weak Condorcet winner} exactly if
$\N{E}(c,c') \geq \N{E}(c',c)$ for each $c' \in C - \{c\}$. Note
that if an election has a Condorcet winner, then this winner is
unique (though, an election may have several weak Condorcet
winners).\footnote{We mention that sometimes Condorcet's rule is not
considered a voting rule (but, for example, a consensus
notion~\cite{mes-nur:b:distance-realizability}) because there are
elections for which there are no Condorcet winners. However,
Condorcet's rule has been traditionally studied in the context of
election control, and--thus---we study it to (a) enable comparison
with other papers, and (b) because for single-peaked elections the
results for Condorcet rule translate to all Condorcet-consistent
rules.} We recall the classic result that if the voters are
single-peaked then the election always has (weak) Condorcet
winner(s) (if the number of voters is odd, then the election has a
unique Condorcet winner).
\item[Maximin.] Maximin rule is defined as follows. The score of a
candidate $c$ is defined as $\score{E}^m(c) = \min_{d \in C
-\{c\}}N_E(c,d)$. The candidates with highest Maximin score are
Maximin winners.
\end{description}
Maximin is an example of a so-called Condorcet-consistent rule. A
rule $R$ is Condorcet-consistent if the following holds: If $E$ is an
election with (weak) Condorcet winner(s), then the $R$-winners of $E$
are exactly these (weak) Condorcet winner(s). We also considered
studying the Copeland rule (see, e.g., the work of Faliszewski et
al.~\cite{fal-hem-hem-rot:j:llull} regarding the complexity of control
for the Copeland rule), but for this rule all relevant types of
control are $\NPclass$-complete and, so, at best we would obtain
$\sharpPclass$-completeness results based on simple generalizations of
proofs from the literature. Nonetheless, since Copeland's rule is
Condorcet-consistent, results for the single-peaked case are valid for
it.
\bigskip
\noindent\textbf{Notation for Graphs.}\quad We assume familiarity with basic
concepts of graph theory. Given an undirected graph $G$, by $V(G)$ we
mean its set of vertices, and by $E(G)$ we mean its set of
edges.\footnote{We use this slightly nonstandard notation to be able
to also use symbols $V$ and $E$ to denote voter collections and
elections.} Whenever we discuss a bipartite graph $G$, we assume
that $V(G)$ is partitioned into two subsets, $X$ and $Y$, such that
each edge connects some vertex from $X$ with some vertex from $Y$. We
write $X(G)$ to denote $X$, and $Y(G)$ to denote $Y$.\bigskip
\noindent\textbf{Computational Complexity.}\quad
We assume that the reader is familiar with standard notions of
computational complexity theory, as presented, for example, in the
textbook of Papadimitriou~\cite{pap:b:complexity}, but we
briefly review notions regarding the complexity theory of counting
problems. Let $A$ be some computational problem where for each
instance $I$ we ask if there exists some mathematical object
satisfying a given condition. In the counting variant of $A$, denoted
$\#A$, for each instance $I$ we ask for the number---denoted
$\#A(I)$---of the objects that satisfy the condition.
For example, consider the following problem.
\begin{definition}
An instance of X3C is a pair $(B,{{\mathcal{S}}})$, where $B = \{b_1, \ldots,
b_{3k}\}$ and ${{\mathcal{S}}} = \{S_1, \ldots, S_n\}$ is a family of
$3$-element subsets of $B$. In X3C we ask if it is possible to find
exactly $k$ sets in ${{\mathcal{S}}}$ whose union is exactly $B$. In \#X3C we
ask how many $k$-element subsets of ${{\mathcal{S}}}$ have $B$ as their union.
\end{definition}
The class of counting variants of $\NPclass$-problems is called $\sharpPclass$
and the class of functions computable in polynomial time is called
$\FPclass$. To reduce counting problems to each other, we will use one of
the following reducibility notions.
\begin{definition}
Let $\#A$ and $\#B$ be two counting problems.
\begin{enumerate}
\addtolength{\itemsep}{-4pt}
\item We say that $\#A$ Turing reduces to $\#B$ if there exists an
algorithm that solves $\#A$ in polynomial time given oracle access
to $\#B$ (i.e., given the ability to solve $\#B$ instances in unit
time).
\item We say that $\#A$ metric reduces to $\#B$ if there exist two
polynomial-time computable functions, $f$ and $g$, such that for
each instance $I$ of $\#A$ it holds that (1) $f(I)$ is an instance
of $\#B$, and (2) $\#A(I) = g(I, \#B(f(I)) )$.
\item We say that $\#A$ parsimoniously reduces to $\#B$ if $\#A$
metric reduces to $\#B$ via functions $f$ and $g$ such that for
each instance $I$ and each integer $k$, $g(I,k) = k$ (i.e., there
is a polynomial-time computable function $f$ such that for each
instance $I$ of $\#A$, we have $\#A(I) = \#B(f(I))$).
\end{enumerate}
\end{definition}
It is well-known that these reducibility notions are transitive. For
a given reducibility notion $R$, we say that a problem is
$\sharpPclass$-$R$-complete if it belongs to $\sharpPclass$ and every
$\sharpPclass$-problem $R$ reduces to it. For example, \#X3C is
$\sharpPclass$-parsimonious-complete~\cite{hun-mar-rad-ste-:j:planar-counting-problems}.
Throughout this paper we will write $\sharpPclass$-complete to mean
$\sharpPclass$-parsimonious-complete. Turing reductions were used, e.g., by
Valiant~\shortcite{val:j:permanent} to show $\sharpPclass$-hardness of
computing a permanent of a 0/1 matrix. As a result, he also showed
$\sharpPclass$-Turing-hardness of the following problem.
\begin{definition}
In \#PerfectMatching we are given a bipartite graph $G =
(G(X),G(Y),G(E))$ with $\|G(X)\| = \|G(Y)\|$ and we ask how many
perfect matchings does $G$ have.
\end{definition}
Zank{\'{o}}~\cite{zan:j:sharp-p} improved upon this result by showing
that \#PerfectMatching is \#P-many-one-complete (``many-one'' is yet
another reducibility type, more general than parsimonious reductions
but less general than metric reductions). From our perspective, it
suffices that, in effect, \#PerfectMatching is \#P-metric-complete.
Metric reductions were introduced by
Krentel~\shortcite{kre:j:optimization}, and parsimonious reductions
were defined by Simon~\shortcite{sim:thesis:complexity}.
\section{Counting Variants of Control Problems}
\label{sec:problems}
In this section we formally define counting variants of election
control problems and show some interconnections between some of
them. We are interested in control by adding candidates (AC), control
by deleting candidates (DC), control by adding voters (AV), and
control by deleting voters (DV). For each of these problems, we
consider its constructive variant (CC) and its destructive variant
(DC).
\begin{definition}\label{def:control}
Let $R$ be a voting system. In each of the counting variants of
constructive control problems we are given a candidate set $C$, a
voter collection $V$, a nonnegative integer $k$, and a designated
candidate $p \in C$. In constructive control by adding voters we are
additionally given a collection $W$ of unregistered voters, and in
constructive control by adding candidates we are additionally given
a set $A$ of unregistered candidates. In these problems we ask for
the following quantities:
\begin{enumerate}
\addtolength{\itemsep}{-4pt}
\item In control by adding voters ($R$-\#CCAV),
we ask how many sets $W'$, $W' \subseteq W$, are there such that
$p$ is the unique winner of $R$-election $(C,V \cup W')$, where
$\|W'\| \leq k$.
\item In control by deleting voters ($R$-\#CCDV), we ask how many
sets $V'$, $V' \subseteq V$ are there such that $p$ is the unique
winner of $R$-election $(C,V - V')$, where $\|V'\| \leq k$.
\item In control by adding candidates ($R$-\#CCAC), we ask how many
sets $A'$, $A' \subseteq A$, are there such that $p$ is the unique
winner of $R$-election $(C \cup A',V)$, where $\|A'\| \leq k$.
\item In control by deleting candidates ($R$-\#CCDC), we ask how
many sets $C'$, $C' \subseteq C$, are there such that $p$ is the
unique winner of $R$-election $(C - C',V)$, where $\|C'\|
\leq k$ and $p \not\in C'$.
\end{enumerate}
Destructive variants are defined identically, except that we ask for
the number of settings where the designated candidate is not the
unique winner.
\end{definition}
\noindent
(To obtain decision variants of the above problems, simply change the
question from asking for a particular quantity to asking if that
quantity is greater than zero. We mention that in the literature
researchers often study both the unique-winner model---as we do
here---and the nonunique-winner model, where it suffices to be one of
the winners. For the rules that we study, results for both models
are the same.)\medskip
The above problems are interesting for several reasons. First, as
described in the introduction, we believe that they are useful as
models for various scenarios pertaining to predicting election
winners. Second, they are counting variants of the well-studied
control
problems~\cite{bar-tov-tri:j:control,hem-hem-rot:j:destructive-control}.
Third, they expose intuitive connections between various types of
control. In particular, we have the following results linking the
complexity of destructive variants with that of the constructive ones,
and linking the complexity of the ``deleting'' variants with that of
the ``adding'' variants.
\begin{theorem}\label{thm:c-to-d}
Let $R$ be a voting system, let $\#{{\mathcal{C}}}$ be one of $R$-\#CCAC,
$R$-\#CCDC, $R$-\#CCAV, $R$-\#CCDV, and let $\#{{\mathcal{D}}}$ be the
destructive variant of $\#{{\mathcal{C}}}$. Then, $\#{{\mathcal{C}}}$ metric reduces to
$\#{{\mathcal{D}}}$ and $\#{{\mathcal{D}}}$ metric reduces to $\#{{\mathcal{C}}}$.
\end{theorem}
\begin{proof}
We give a metric reduction from $\#{{\mathcal{C}}}$ to $\#{{\mathcal{D}}}$. Let $I$ be
an instance of $\#{{\mathcal{C}}}$, where the goal is to make some candidate
$p$ the unique winner. We define $f(I)$ to be an instance of
$\#{{\mathcal{D}}}$ that is identical to $I$, except that the goal is to
ensure that $p$ is not the unique winner. Let $s_I$ be the total
number of control actions allowed in $I$\footnote{For example, if
$\#{{\mathcal{C}}}$ was $\#CCDV$ and $I = (C,V,p,k)$, then $s_I$ would be
the number of up-to-size-$k$ subsets of $V$ (i.e., the number of
subsets of voters that can be deleted from $V$).} (and, naturally,
also the total number of control actions allowed in $f(I)$). It is
easy to see that $s_I$ is polynomial-time computable and that the
number of solutions for $I$ is exactly $s_I - \#{{\mathcal{D}}}(f(I))$. Thus,
we define $g(I,\#{{\mathcal{D}}}(f(I))) = s_I - \#{{\mathcal{D}}}(f(I))$. We see that
the reduction is polynomial-time and correct.
The same argument shows that
$\#{{\mathcal{D}}}$ metric reduces to $\#{{\mathcal{C}}}$.
\end{proof}
\begin{theorem}\label{thm:d-to-a}
Let $R$ be a voting system, then $R$-\#CCDV Turing reduces to
$R$-\#CCAV and $R$-\#CCDC Turing reduces to $R$-\#CCAC.
\end{theorem}
\begin{proof}
Let us fix a voting system $R$. The proofs that $R$-\#CCDV Turing
reduces to $R$-\#CCAV and that $R$-\#CCDC Turing-reduces to
$R$-\#CCAV are very similar and thus we discuss them jointly.
%
%
%
Let $I = (C,V,p,k)$ be an input instance of $R$-\#CCDV (of
$R$-CCDC), where $C$ is the candidate set, $V$ is the collection of
voters, $p$ is the designated candidate, and $k$ is the upper bound
on the number of voters that can be removed. We define the following
transformations of $I$:
\begin{enumerate}\addtolength{\itemsep}{-4pt}
\item For the case of $R$-\#CCDV, for each nonnegative integer $g$,
$g \le \|V\|$, let $J_g$ be an instance of $R$-\#CCAV, where the
candidate set is $C$, the set of registered voters is empty, the
collection of unregistered voters is $V$, the designated candidate
is $p$, and the bound on the number of voters that can be added is
$g$, That is, $J_g = (C,\emptyset,V,p,g)$.
\item For the case of $R$-\#CCDC, for each nonnegative integer $g$,
$g \le \|C\|-1$, let $J_g$ be an instance of $R$-\#CCAC, where the
candidate set is $\{p\}$, the set of registered voters is $V$, the
set of unregistered candidates is $C-\{p\}$, the designated
candidate is $p$, and the bound on the number of candidates that
can be added is $g$. That is, $J_g = (\{p\},C-\{p\},V,p,g)$.
\end{enumerate}
To complete the reduction, it suffices to note that for
the case of $R$-\#CCDV it holds that:
\[
\text{$R$-\#CCDV}(I) = \left\{
\begin{array}{ll}
\text{$R$-\#CCAV}(J_{\|V\|}) & \text{, if $k \ge \|V\|$}\\
\text{$R$-\#CCAV}(J_{\|V\|}) - \text{$R$-\#CCAV}(J_{\|V\|-k-1}) & \text{, if $0 \leq k < \|V\|$}\\
\end{array}
\right.
\]
and for the case of $R$-\#CCDC it holds that:
\[
\text{$R$-\#CCDC}(I) = \left\{
\begin{array}{ll}
\text{$R$-\#CCAC}(J_{\|C\|-1}) & \text{, if $k \ge \|C\|-1$}\\
\text{$R$-\#CCAC}(J_{\|C\|-1}) - \text{$R$-\#CCAC}(J_{\|C\|-k-2}) & \text{, if $0 \leq k < \|C\|-1$}\\
\end{array}
\right.
\]
These expressions define a natural algorithm for solving $R$-\#CCDV
($R$-\#CCDC) given at most two calls to an $R$-\#CCAV (an
$R$-\#CCAC) oracle. It is clear that this Turing reduction runs in
polynomial time.
\end{proof}
Theorems~\ref{thm:c-to-d} and~\ref{thm:d-to-a} are very useful and, in
particular, we obtain all of our destructive-case hardness results via
Theorem~\ref{thm:c-to-d}, and all our ``deleting''-case easiness
results using Theorem~\ref{thm:d-to-a}.
Obtaining results similar to Theorems~\ref{thm:c-to-d}
and~\ref{thm:d-to-a} for the decision variants of control problems is
more difficult. Nonetheless, recently several researchers have made
some progress on this front. In particular, Hemaspaandra,
Hemaspaandra, and Menton~\cite{hem-hem-men:c:search-decision} have
shown that control by partition of candidates and control by runoff
partition of candidates (two types of control not discussed in this
paper) are equivalent in terms of their computational complexity.
Focusing on particular classes of voting rules, for the case of
$k$-Approval and $k$-Veto, Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:c:weighted-control} have shown that
control by deleting voters reduces to control by adding voters, and
Mi\k{a}sko~\cite{mia:t:priced-control} achieved the same for the case
of voting rules based on the weighted majority graphs (e.g., for
Borda, Condorcet, Maximin, and Copeland).\footnote{Mi\k{a}sko's
master's thesis gives this result for the case of Borda and
destructive control only, but it is clear that the proof technique
adapts to constructive control and that it applies to all voting
rules based on weighted majority graphs.}
\section{Results}
\label{sec:results}
We now present our complexity results regarding counting variants of
election control problems. We consider two cases: the unrestricted
case, where the voters can have any arbitrary preference orders, and
the single-peaked case, where voters' preference orders are
single-peaked with respect to a given societal axis. We summarize our
results in Table~\ref{tab:results}. For comparison, in
Table~\ref{tab:decision} we quote results for the decision variants of
the respective control problems.
Let us consider the unrestricted case first. Not surprisingly, for
each of our control problems whose decision variant is $\NPclass$-complete,
the counting variant is $\sharpPclass$-complete for some reducibility type.
However, interestingly, we have also found examples of election
systems and control types where the decision variant is easy, but the
counting variant is hard. For example, for $2$-Approval we have that
all decision variants of voter control problems are in
$\Pclass$~\cite{lin:thesis:elections,fal-hem-hem:c:weighted-control}, yet
all their counting variants are $\sharpPclass$-Turing-complete. Similarly,
for Maximin all decision variants of candidate control (except
constructive control by adding candidates) are in
$\Pclass$~\cite{fal-hem-hem:j:multimode}, yet their counting variants are
$\sharpPclass$-Turing-complete (or are complete for $\sharpPclass$ through even
less demanding reducibility types).
The situation for the single-peaked case is quite interesting as well.
While we found polynomial-time algorithms for all of our control
problems for Plurality, $k$-Approval, and all Condorcet-consistent
rules, we did not obtain results for voter control under Approval
(candidate control for Approval is easy even in the unrestricted
case). Decision variants of voter control are $\NPclass$-complete for
Approval in the unrestricted case, but---very interestingly---are in
$\Pclass$ for the single-peaked case. However, the algorithm for the
single-peaked case is a clever greedy approach that seems to be very
difficult to adapt to the counting case. Further, as opposed to the
$k$-Approval case, for voter control under Approval, there does not
seem to be a natural dynamic-programming-based approach. Naturally,
we also tried to find a hardness proof, but we failed at that as well
because the problem seems to have quite a rich structure. (However,
see the work of Mi\k{a}sko~\cite{mia:t:priced-control} for the case of
priced control under Approval with single-peaked preferences.)
\newcommand{com.}{com.}
\tabcolsep=0.1cm
\begin{table*}[t]
\small
\begin{center}
(a) The Unrestricted Case
\hfill{}
\begin{tabular}{c|cccccc}
\hline
Problem\rule{0cm}{0.5cm} & Plurality & $k$-Approval, $k \geq 2$ & Approval & Condorcet & Maximin\\[0.15cm]
\hline
\#CCAC\rule{0cm}{0.5cm}& $\sharpPclass$-com. & $\sharpPclass$-com. &-- & -- & $\sharpPclass$-com.\\
\#DCAC & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. & FP & FP & $\sharpPclass$-metric-com. \\
\#CCDC & $\sharpPclass$-com. & $\sharpPclass$-com. &FP & FP & $\sharpPclass$-Turing-com. \\
\#DCDC & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. & -- & -- & $\sharpPclass$-Turing-com. \\[0.4em]
\#CCAV & $\FPclass$ & $\sharpPclass$-Turing-com. & $\sharpPclass$-com. & $\sharpPclass$-com. & $\sharpPclass$-com. \\
\#DCAV & $\FPclass$ & $\sharpPclass$-Turing-com. & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. \\
\#CCDV & $\FPclass$ & $\sharpPclass$-metric-com. & $\sharpPclass$-com. & $\sharpPclass$-com. & $\sharpPclass$-com. \\
\#DCDV & $\FPclass$ & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. & $\sharpPclass$-metric-com. \\
\end{tabular}
\vspace{0.5cm}
(b) The Single-Peaked Case
\hfill{}
\begin{tabular}{c|cccc}
\hline
Problem\rule{0cm}{0.5cm} & Plurality & $k$-Approval, $k \geq 2$ & Approval & Condorcet-Consistent \\
& & & & (e.g., Maximin)\\[0.15cm]
\hline
\#CCAC\rule{0cm}{0.5cm}& $\FPclass$ & $\FPclass$ &-- & -- \\
\#DCAC & $\FPclass$ & $\FPclass$ & FP & FP \\
\#CCDC & $\FPclass$ & $\FPclass$ &FP & FP \\
\#DCDC & $\FPclass$ & $\FPclass$ & -- & -- \\[0.4em]
\#CCAV & $\FPclass$ & $\FPclass$ & ? & $\FPclass$ \\
\#DCAV & $\FPclass$ & $\FPclass$ & ? & $\FPclass$ \\
\#CCDV & $\FPclass$ & $\FPclass$ & ? & $\FPclass$ \\
\#DCDV & $\FPclass$ & $\FPclass$ & ? & $\FPclass$ \\
\end{tabular}
\hfill{}
\caption{\label{tab:results}The complexity of counting
variants of control problems for (a) the unrestricted case,
and for (b) the single-peaked case. A dash in an entry means
that the given system is \emph{immune} to the type of
control in question (i.e., it is impossible to achieve the
desired effect by the action this control problem allows;
technically this means the answer to the counting question
is always $0$). Immunity results were established by
Bartholdi, Tovey, and
Trick~(1989) %
for the constructive cases, and by Hemaspaandra,
Hemaspaandra, and
Rothe~(2007) %
for the destructive cases. }
\end{center}
\end{table*}
\tabcolsep=0.1cm
\begin{table*}[t]
\small
\begin{center}
(a) The Unrestricted Case
\hfill{}
\begin{tabular}{c|ccccccc}
\hline
Problem\rule{0cm}{0.5cm} & Plurality & $2$-Approval & $3$-Approval &$k$-Approval, $k \geq 4$ & Approval & Condorcet & Maximin\\[0.15cm]
\hline
CCAC\rule{0cm}{0.5cm}& $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. &-- & -- & $\NPclass$-com.\\
DCAC & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com.& $\Pclass$ & $\Pclass$ & $\Pclass$ \\
CCDC & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com.& $\Pclass$ & $\Pclass$ & $\Pclass$ \\
DCDC & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com.& -- & -- & $\Pclass$ \\[0.4em]
CCAV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. \\
DCAV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\NPclass$-com. & $\Pclass$ & $\Pclass$ & $\NPclass$-com. \\
CCDV & $\Pclass$ & $\Pclass$ & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. & $\NPclass$-com. \\
DCDV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\NPclass$-com. & $\Pclass$ & $\Pclass$ & $\NPclass$-com. \\
\end{tabular}
\vspace{0.5cm}
(b) The Single-Peaked Case
\hfill{}
\begin{tabular}{c|cccc}
\hline
Problem\rule{0cm}{0.5cm} & Plurality & $k$-Approval, $k \geq 2$ & Approval & Condorcet-Consistent \\
& & & & (e.g., Maximin)\\[0.15cm]
\hline
CCAC\rule{0cm}{0.5cm}& $\Pclass$ & $\Pclass$ &-- & -- \\
DCAC & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
CCDC & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
DCDC & $\Pclass$ & $\Pclass$ & -- & -- \\[0.4em]
CCAV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
DCAV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
CCDV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
DCDV & $\Pclass$ & $\Pclass$ & $\Pclass$ & $\Pclass$ \\
\end{tabular}
\hfill{}
\caption{\label{tab:decision}The complexity of decision
variants of control problems for (a) the unrestricted case,
and for (b) the single-peaked case. A dash in an entry means
that the given system is \emph{immune} to the type of
control in question. For the unrestricted case, we have the
following: The results for Plurality are due to Bartholdi,
Tovey, and Trick~\cite{bar-tov-tri:j:control} and
Hemaspaandra, Hemaspaandra, and
Rothe~\cite{hem-hem-rot:j:destructive-control}, the results
regarding $k$-Approval are due to
Lin~\cite{lin:thesis:elections} (see
also~\cite{elk-fal-sli:j:cloning,fal-hem-hem:c:weighted-control}),
the results regarding Approval are due to Hemaspaandra,
Hemaspaandra, and
Rothe~\cite{hem-hem-rot:j:destructive-control}, and the
results regarding Maximin are due to Faliszewski,
Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:j:multimode}. For the
single-peaked case, we inherit polynomial-time algorithms
from the unrestricted case. The remaining results (except
those for Condorcet-consistent rules) are due to Faliszewski
et al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences},
and the remaining results regarding Condorcet-consistent
rules are due to~\cite{bra-bri-hem-hem:c:sp2}.}
\end{center}
\end{table*}
In the following sections we give proofs for our results and provide
some more detailed discussion regarding particular voting rules.
\subsection{Plurality Voting}
Under plurality voting, counting variants of both control by adding
voters and control by deleting voters are in $\FPclass$ even for the
unrestricted case. Our algorithms are based on dynamic programming and
applications of Theorems~\ref{thm:c-to-d} and~\ref{thm:d-to-a}.
\begin{theorem} \label{th:plavc}
Plurality-\#CCAV, Plurality-\#DCAV, Plurality-\#CCDV, and
Plurality-\#DCDV are in $\FPclass$, even in the unrestricted case.
\end{theorem}
\begin{proof}
We give the proof for Plurality-\#CCAV only. The result for
Plurality-\#CCDV follows by applying Theorem~\ref{thm:d-to-a}, and
the destructive cases follows by applying Theorem~\ref{thm:c-to-d}.
Let $I = (C,V,W,p,k)$ be an input instance of Plurality-\#CCAV,
where $C = \{p,c_1,\ldots, c_{m-1}\}$ is the candidate set, $V$ is
the set of registered voters, $W$ is the set of unregistered voters,
$p$ is the designated candidate, and $k$ is the upper bound on the
number of voters that can be added. We now describe a polynomial-time
algorithm that computes the number of solutions for $I$.
Let $A_p$ be the set of voters from $W$ that rank $p$ first.
Similarly, for each $c_i \in C$, let $A_{c_i}$ be the set of voters
from $W$ that rank $c_i$ first. We also define
$\id{count}(C,V,W,p,k,j)$ to be the number of sets $W' \subseteq W - A_p$
such that
(1) $\|W'\| \leq k-j$, and
(2) in election $(C,V \cup W')$ each candidate $c_i \in C$, $1
\leq i \leq m-1$, has score at most $\score{(C,V)}^p(p)+j-1$.
Our algorithm works as follows. First, we compute $k_0$, the
minimum number of voters from $A_p$ that need to be added to $V$ to
ensure that $p$ has plurality score higher than any other candidate
(provided no other voters are added). Clearly, if $p$ already is the
unique winner of $(C,V)$ then $k_0$ is $0$, and otherwise $k_0$ is
$\max_{c_i \in C}(\score{(C,V)}^p(c_i)-\score{(C,V)}^p(p)+1)$. Then,
for each $j$, $k_0 \leq j \leq \min(k,\|A_p\|)$, we compute the number
of sets $W'$, $W' \subseteq W$, such that $W'$ contains exactly $j$
voters from $A_p$, at most $k-j$ voters from $W - A_p$, and $p$ is
the unique winner of $(C, V \cup W')$. It is easy to verify that for
a given $j$, there is exactly $h(j) =
\binom{\|A_p\|}{j}\cdot\id{count}(C,V,W\!,p,k,j)$ such sets. Our
algorithm returns $\sum_{j=k_0}^{\min(k,\|A_p\|)}h(j)$. The reader can
easily verify that this indeed is the correct answer. To complete
the proof it suffices to show a polynomial-time algorithm for
computing $\id{count}(C,V,W,p,k,j)$.
Let us fix $j$, $k_0 \leq j \leq \min(k,\|A_p\|)$. We now show how to
compute $\id{count}(C,V,W,p,k,j)$. Our goal is to count the number
of ways in which we can add at most $k-j$ voters from $W-A_p$ so
that no candidate $c_i \in C$ has score higher than
$\score{(C,V)}^p(p)+j-1$. For each candidate $c_i \in C$, we can
add at most
$
l_i =
\min\bigl(\|A_{c_i}\|,j+\score{(C,V)}^p(p)-\score{(C,V)}^p(c_i)-1\bigr)
$
voters from $A_{c_i}$; otherwise $c_i$'s score would exceed
$\score{(C,V)}^p(p)+j-1$.
For each $i$, $1 \leq i \leq m-1$, and each $t$, $0 \leq t \leq
k-j$, let $a_{t,i}$ be the number of sets $W' \subseteq A_{c_1}\cup
A_{c_2}\cup\cdots\cup A_{c_i}$ that contain exactly $t$ voters and
such that each candidate $c_1, c_2, \ldots, c_i$ has score at most
$\score{(C,V)}^p(p)+j-1$ in the election $(C,V \cup W')$. Naturally,
$\id{count}(C,V,W,p,k,j) = \sum_{t=0}^{k-j}a_{t,m-1}$. It is easy
to check that $a_{t,i}$ satisfies the following recursion:
\[
a_{t,i} =
\begin{cases}
\sum_{s=0}^{\min(l_i,t)}\binom{\|A_{c_i}\|}{s}a_{t-s,i-1}, & \text{if $t>0$, $i>1$}, \\[2mm]
1, & \text{if $t=0$, $i>1$}, \\[1mm]
\binom{\|A_1\|}{t}, & \text{if $t\le\|A_{c_1}\|$, $i=1$}, \\[1mm]
0, & \text{if $t>\|A_{c_1}\|$, $i=1$}.
\end{cases}
\]
Thus, for each $t,i$ we can compute $a_{t,i}$ using standard dynamic
programming techniques in polynomial time. Thus,
$\id{count}(C,V,W,p,k,j)$ also is polynomial-time computable. This
completes the proof that Plurality-\#CCAV is in $\FPclass$.
\end{proof}
On the other hand, for Plurality voting \#CCAC and \#CCDC are
$\sharpPclass$-complete and this follows from proofs already given in the
literature~\cite{fal-hem-hem:j:nearly-sp}.
\begin{theorem}\label{thm:placdcc}
In the unrestricted case, Plurality-\#CCAC and Plurality-\#CCDC are $\sharpPclass$-complete.
\end{theorem}
\begin{proof}
Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite[Theorem~6.4]{fal-hem-hem:j:nearly-sp} show that
decision variants of control by adding candidates and control by
deleting candidates are $\NPclass$-complete.\footnote{Naturally, these
results are originally due to Bartholid, Tovey, and
Trick~\cite{bar-tov-tri:j:control}. However, the proofs of Faliszewski, Hemaspaandra,
and Hemaspaandra are more useful in our case, because they work
for $k$-Approval for each fixed $k$ (which will be useful later),
and it is convenient to verify that they give parsimonious
reductions from a $\sharpPclass$-complete problem.} Their proofs work
by reducing X3C to appropriate control problems in a parsimonious
way. This means that the same reductions reduce \#X3C to the
counting variants of respective control problems.
\end{proof}
\noindent
Now, Corollary~\ref{cor:placdcdc} follows by combining
Theorems~\ref{thm:placdcc} and~\ref{thm:c-to-d}.
\begin{corollary}\label{cor:placdcdc}
In the unrestricted case, Plurality-\#DCAC and Plurality-\#DCDC are $\sharpPclass$-metric-complete.
\end{corollary}
However, for the single-peaked case, we get easiness results.
\begin{theorem}
For the case of single-peaked profiles, Plurality-\#CCAC,
Plurality-\#CCDC, Plurality-\#DCAC, and Plurality-\#DCDC are in
$\FPclass$.
\end{theorem}
\begin{proof}
This result follows from Theorem~\ref{thm:kapp-ccac-sp} (see the next section) for
the case of $1$-Approval.
\end{proof}
\subsection{$\boldsymbol{k}$-Approval Voting}
While $k$-Approval is in many respects a simple generalization of the
Plurality rule, it turns out that for $k \geq 2$, for the
unrestricted case, all counting variants of control
problems are intractable for $k$-Approval. This is quite expected for
candidate control because decision variants of these problems are
$\NPclass$-complete (see the work of
Lin~\cite{lin:thesis:elections,elk-fal-sli:j:cloning} and additionally
of Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:c:weighted-control}), but is more
intriguing for voter control (as shown by Lin, for $2$-Approval all
voter control decision problems are in $\Pclass$, and, as one can verify,
for $k$-Approval all destructive voter control decision problems are
in $\Pclass$).
On the other hand, for the single-peaked case all the counting
variants of control are polynomial-time solvable for $k$-Approval
(however, note that we rely on $k$ being a constant; our results do
not generalize to the case where $k$ is part of the input). \bigskip
\noindent
\textbf{Candidate Control.}
\quad
We start be quickly dealing with candidate control in the unrestricted
case. For the unrestricted case, as in the case of Plurality, we can
use the result of Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:j:nearly-sp}.
\begin{theorem}
In the unrestricted case, for each $k$, $k \geq 2$,
$k$-Approval-\#CCAC and $k$-Approval-\#CCDC are $\sharpPclass$-complete,
and $k$-Approval-\#DCAC and $k$-Approval-\#DCDC are
$\sharpPclass$-metric-complete.
\end{theorem}
\begin{proof}
For the constructive variants, the same approach as in the proof of
Theorem~\ref{thm:placdcc} works: The hardness proofs given by
Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite[Theorem~6.4]{fal-hem-hem:j:nearly-sp} apply to
the case of $k$-Approval as well. For the destructive variants, we
invoke Theorem~\ref{thm:c-to-d}.
\end{proof}
For the single-peaked case, we use a dynamic-programming approach. Our
algorithm is a very extensive generalization of the algorithm for the
single-peaked variant of Plurality-CCAC given by Faliszewski et
al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences}
\begin{theorem}\label{thm:kapp-ccac-sp}
For the case of single-peaked profiles, for each fixed positive $k$,
$k$-Approval-\#CCAC, $k$-Approval-\#CCDC, $k$-Approval-\#DCAC, and
$k$-Approval-\#DCDC are in $\FPclass$.
\end{theorem}
\begin{proof}
We give the proof for $k$-Approval-\#CCAC only. The result for
$k$-Approval-\#CCDC follows by applying Theorem~\ref{thm:d-to-a},
and the destructive cases follow by applying
Theorem~\ref{thm:c-to-d}.
We are given a candidate set $C$, a voter collection $V$, a
designated candidate $p \in C$, a collection $A$ of unregistered
candidates, a positive integer $\ell$, and a societal axis
$\mathrel{L}$ over $C \cup A$. We show a polynomial time algorithm
for determining the number of sets $A'$, $A' \subseteq A$, where
$\|A'\| \le \ell$, such that $p$ is the unique $k$-Approval winner
in election $(C \cup A', V)$. We assume that $\|C \cup A\| > 4k$
(otherwise we solve our problem by enumerating all possible solutions).
Intuitively, the idea of our algorithm is as follows: First, we fix
two groups of $k$ candidates, one ``to the left'' of $p$ (with
respect to $L$) and one ``to the right'' of $p$. We will argue that
each choice of such ``neighborhood'' of $p$ fixes $p$'s score.
Second, with $p$'s score fixed, we count how many ways are there to
add candidates so that every candidate has fewer points than $p$.
We sum these values over all choices of $p$'s ``neighborhood.''
\newcommand{{{\mathrm{before}}}}{{{\mathrm{before}}}}
\newcommand{{{\mathrm{after}}}}{{{\mathrm{after}}}}
Let us introduce some useful notation. For each set $X \subseteq C
\cup A$ of candidates we define ${{\mathrm{before}}}(X)$ to be the subset of
those candidates in $C \cup A$ that precede each member of $X$ with
respect to $L$ (i.e., ${{\mathrm{before}}}(X) = \{ d \in C \cup A \mid (\forall
x \in X)[d \mathrel{L} x]\}$). Analogously, we define ${{\mathrm{after}}}(X)$ to be the
subset of those candidates in $C \cup A$ that succeed all members of
$X$ with respect to $L$ (i.e., ${{\mathrm{after}}}(X) = \{ d \in C \cup A \mid
(\forall x \in X)[x \mathrel{L} d]\}$). The next lemma describes how we can
fix candidates' scores by fixing their ``neighborhoods.''
\begin{lemma}
\label{k-barrier-lemma}
For each set $H \subseteq C \cup A$, $\|H\| = k$, each set $X
\subseteq {{\mathrm{before}}}(H)$ and each candidate $c \in X$, if $z$ is the score
of candidate $c$ in k-Approval election $(X \cup H, V)$, then for
each $Y \subseteq {{\mathrm{after}}}(H)$ in k-Approval election $(X \cup H \cup Y,
V)$ the score of candidate $c$ is $z$ as well.
\end{lemma}
\begin{proof}
For a given subset $Y \subseteq {{\mathrm{after}}}(H)$ and candidate $c \in
X$, suppose that the score of candidate $c$ in $k$-Approval election
$(X \cup H \cup Y, V)$ is $z' \ne z$. Adding a candidate can
never increase the score of a candidate already present, so we
conclude that $z' < z$. It means that there is a voter $v \in V$
from which $c$ gains a point in election $E_1 = (X \cup H,V)$ but
not in election $E_2 = (X \cup H \cup Y, V)$. Thus there is a
candidate $c' \in Y$ that receives a point from $v$. Since $c$ is
among $v$'s top $k$ most preferred candidates across $X \cup H$
and $\|H\| = k$, there must exist a candidate $c'' \in H$ that is
less preferred than $c$ by $v$. Since $v$ prefers $c'$ over $c$,
it must be that $v$ prefers $c'$ over $c''$. However, we know that
$c \mathrel{L} c'' \mathrel{L} c'$, because $c \in X$, $c'' \in H$
and $c' \in Y$. Since $v$ is single-peaked with respect to
$L$, $v$ cannot prefer both $c$ and $c'$ over $c''$; a contradiction.
\end{proof}
Let $B_L$ be the set of the first $2k$ candidates from $C \cup A$
(with respect to $L$) and let $B_R$ be the set of the last $2k$
candidates from $C \cup A$ (with respect to $L$). Without loss of
generality, we assume that $B_L$ and $B_R$ contain candidates from
$C$ only, and that each voter has a preference order of the form $C
- (B_L \cup B_R) > B_L > B_R$ (if this were not the case then we
could add to $C$ two groups of $2k$ dummy candidates that all the
voters rank last, respecting the above requirement, and that are at
the extreme ends of the societal axis $L$). Note that by our
assumptions $p$ does not belong to $B_L \cup B_R$ and for each $A''
\subseteq A$ it holds that in election $(C \cup A'', V)$ the
candidates from $B_R$ receive $0$ points each.
We say that a set $Y \subseteq C \cup A$ is well-formed if for each
candidate $d \in C$ it holds that $d \in {{\mathrm{before}}}(Y) \cup Y \cup
{{\mathrm{after}}}(Y)$ (in other words, $Y$ is well-formed if it is an interval
with respect to the societal axis $L$ restricted to $C \cup Y$). We
define $K_L$ to be a collection of subsets of ${{\mathrm{before}}}(\{p\})$ such
that if $Y \in K_L$ then $Y\cup\{p\}$ is well-formed and $\|Y\| = k$.
%
%
%
(Intuitively, $K_L$ is the family of sets of candidates that can
form the ``left part'' of $p$'s neighborhood.) We define $K_R$
analogously, but ``to the right of $p$.'' That is, $K_R$ is a
collection of subsets of ${{\mathrm{after}}}(\{p\})$ such that if $Y \in K_R$
then $\{p\} \cup Y$ is well-formed and $\|Y\| = k$.
Note that there are at most $O(\|C \cup A\|^k)$ sets in each of
$K_L$ and $K_R$.
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
Our algorithm proceeds as follows. In a loop, we try each $Y_L \in
K_L$ and each $Y_R \in K_R$. For each choice of $Y_L$ and $Y_R$, we
set $Y = Y_L \cup \{p\} \cup Y_R$ to be the neighborhood of $p$, and
we set $\rho$ to be the $k$-Approval score of $p$ in election
$(Y,V)$. (If $\|Y \cap A\| > \ell$ then we drop this choice of $Y_L$
and $Y_R$ because picking this neighborhood requires adding more
candidates than we are allowed to.) We create two new sets, $C' = C
\cup Y$, $A' = A \cap ({{\mathrm{before}}}(Y) \cup {{\mathrm{after}}}(Y))$, and an integer
$\ell' = \ell - \|Y \cap A\|$. Finally, we compute the number $s(Y)$
of size-at-most-$\ell'$ subsets $A''$ of $A'$ such that in
election $(C' \cup A'',V)$ each candidate has less than $\rho$
points (except for $p$ who, by Lemma~\ref{k-barrier-lemma}, has
exactly $\rho$ points). Computing $s(Y)$ in polynomial time is a
crucial technical part of the algorithm and we describe it below. We
sum up all the values $s(Y)$ and return them as our output. By
Lemma~\ref{k-barrier-lemma} it is easy to see that this strategy is
correct.
We now describe how to compute $s(Y)$ in polynomial time. Let us
consider a situation where we have picked $p$'s neighborhood $Y$ and
computed $\rho$, $C'$, $A'$, and $\ell'$. If $\rho = 0$ then $p$
cannot be the unique winner so we assume that $\rho > 0$.
From now on, we assume that $C'$ and $A'$ take the roles of $C$ and
$A$ in the definition of a well-formed set.
\newcommand{{{\mathrm{prev}}}}{{{\mathrm{prev}}}}
\newcommand{\Prev}{{{\mathrm{Prev}}}}
%
%
%
%
For each well-formed set $Z = \{z_1, \ldots, z_{2k}\} \subseteq C'
\cup A'$, such that $z_1 \mathrel{L} \cdots \mathrel{L} z_{2k}$, and for each
nonnegative integer $s$ we define:
\begin{enumerate}
\item[] $f(Z,s) = $ the number of sets $A'' \subseteq (A' \cap
{{\mathrm{before}}}(Z))$ such that $\|A''\| \leq s$ and in election $(Z \cup
A'' \cup (C' \cap {{\mathrm{before}}}(Z)))$ each candidate in $\{z_1, \ldots,
z_k\} \cup A'' \cup (C' \cap {{\mathrm{before}}}(Z))$ other than $p$ (if $p$
is included in this set) has fewer than $\rho$
points.\footnote{Yes, we really mean the condition on scores to
apply to candidates $z_1, \ldots, z_k$ but not to candidates
$z_{k+1}, \ldots, z_{2k}$; we will implicitly ensure that the
scores of $z_{k+1}, \ldots, z_{2k}$ do satisfy the condition as
well.}
\end{enumerate}
It is easy to see that $s(Y)$ is simply $f(B_R,\ell')$ (recall that
$B_R$ is the set of ``right-hand side'' dummy candidates, who have score
$0$ in every election). For each $Z$ and $s$, we can compute
$f(Z,s)$ using dynamic programming. To provide appropriate recursive
expression for $f$, we need some additional notation. We define
${{\mathrm{prev}}}(Z)$ to be the last candidate from $C'$ with respect to $L$
that precedes the candidates from $Z$. That is ${{\mathrm{prev}}}(Z)$ is the
maximal (``rightmost'') element of $C' \cap {{\mathrm{before}}}(Z)$ with respect
to $L$. We define $\Prev(Z)$ to be the set that contains ${{\mathrm{prev}}}(Z)$
and all the candidates from $C' \cup A'$ that are between ${{\mathrm{prev}}}(Z)$
and $Z$, with respect to $L$. For a well-formed set $Z = \{z_1,
\ldots, z_{2k}\}$ and an element $z \in \Prev(Z)$ (provided that
$\Prev(Z)$ is defined) we let $\delta(Z,z)$ be $1$ if in election
$(\{z,z_1,\ldots, z_{2k}\},V)$ candidate $z_k$ is either $p$ or
obtains fewer than $\rho$ points. Otherwise we set $\delta(Z,z) =
0$. Now, for each nonnegative integer $s$ and each well-formed set
$Z = \{z_1, \ldots, z_{2k}\} \subseteq C' \cup A'$ such that $z_1
\mathrel{L} \cdots \mathrel{L} z_{2k}$ the following relation holds ($\chi_{A'}$
is the characteristic function of set $A'$, i.e., $\chi_{A'}(z)$ is
$1$ if $z \in A'$ and is $0$ otherwise):
\[ f(Z,s) = \left\{ \def1.5{1.5}
\begin{array}{l}
\displaystyle\sum_{z \in \Prev(Z)} \delta(Z,z)f(\{z,z_1,\ldots,z_{2k-1}\}, s-\chi_{A'}(z)) \text{, if $\Prev(Z)$ is defined,} \\
\quad 1 \text{, otherwise.}
\end{array} \right.
\]
Note that if $\Prev(Z)$ is not defined then $Z = B_L$ and by
single-peakedness of the election, $z_1, \ldots, z_k$ have zero
points each.
The correctness of this recursive expression follows by our
assumptions and by Lemma~\ref{k-barrier-lemma}. It is easy to see
that using dynamic programming and this recursive expression we can
compute $f(B_R,\ell')$ in polynomial time. This concludes the proof.
\end{proof}
\noindent
\textbf{Voter Control.} \quad We now turn to the case of voter
control and we start with the unrestricted case. This time, we need
quite a few new ideas: Under $k$-Approval (for fixed $k$, $k \geq 2$)
all types of counting voter control are hard, while the complexity of
decision variants is quite varied (see the works of
Lin~\cite{lin:thesis:elections,elk-fal-sli:j:cloning} and of Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:c:weighted-control}).
We start by considering $2$-Approval-\#CCAV and $2$-Approval-\#CCDV
separately, then we extend these results to $k \geq 3$, and finally we
invoke Theorem~\ref{thm:c-to-d} to obtain the results for the
destructive cases.
\begin{theorem}\label{thm:ccavdv-2}
In the unrestricted case,
$2$-Approval-\#CCAV is $\sharpPclass$-Turing-complete and
$2$-Approval-\#CCDV is $\sharpPclass$-metric-complete.
\end{theorem}
\begin{proof}
We first consider $2$-Approval-\#CCAV.
We give a Turing reduction from \#PerfectMatching to
$2$-Approval-\#CCAV. Let $G = (G(X),G(Y),G(E))$ be our input
bipartite graph, where $G(X) = \{x_1, \ldots, x_n\}$ and $G(Y) =
\{y_1, \ldots, y_n\}$ are sets of vertices, and $G(E) = \{e_1,
\ldots, e_m\}$ is the set of edges. We form an election $E = (C,V)$
and a collection $W$ of unregistered voters as follows. We set $C =
\{p,b_1,b_2\} \cup G(X) \cup G(Y)$ and we let $V = (v_1,v_2)$, where
$v_{1}$ has preference order $p > b_1 > C - \{p,b_1\}$ and $v_{2}$
has preference order $p > b_2 > C - \{p,b_2\}$. We let $W = (w_1,
\ldots, w_m)$, where for each $\ell$, $1 \leq \ell \leq m$, if
$e_\ell = \{x_i,y_j\}$ then $w_\ell$ has preference order $x_i > y_j
> C - \{x_i,y_j\}$.
Thus, in election $E$ candidate $p$ has score $2$, candidates $b_1$
and $b_2$ have score $1$, and candidates in $G(X) \cup G(Y)$ have score
$0$. We form an instance $I$ of $2$-Approval-\#CCAV with election $E
= (C,V)$, collection $W$ of unregistered voters, designated
candidate $p$, and the number of voters that can be added set to
$n$. We form instance $I'$ to be identical, except we allow to add
at most $n-1$ voters.
It is easy to verify that the number of $2$-Approval-\#CCAV
solutions for $I$ (for $I'$) is the number of matchings in $G$ of
cardinality at most $n$ (the number of matchings in $G$ of
cardinality at most $n-1$). (Each unregistered voter corresponds to an
edge in $G$ and one cannot add two edges that share a vertex as then
$p$ would no longer be the unique winner.) The number of
perfect matchings in $G$ is exactly the number of solutions for $I$ minus
the number of solutions for $I'$.\bigskip
Let us now consider the case of $2$-Approval-\#CCDV. We give a
metric reduction from \#PerfectMatching. As before, let $G =
(G(X),G(Y),G(E))$ be our input bipartite graph, where $X = \{x_1,
\ldots, x_n\}$ and $Y = \{y_1, \ldots, y_n\}$ are sets of vertices,
and $E = \{e_1, \ldots, e_m\}$ is the set of edges. For each vertex
$v$ of $G$, we write $d(v)$ to denote $v$'s degree. We set $D =
\max\{d(v) \mid v \in X \cup Y\}$ and we set $T = \sum_{v \in X \cup
Y}(D-d(v))$. W.l.o.g. we assume that $D \geq 2$. We form an
election $E = (C,V)$ as follows. We set $C = \{p\} \cup X \cup Y
\cup B$, where $B = \{b_1, \ldots, b_{T+D}\}$. We form the
collection $V = V_E + A_{X,Y} + A_p$ of voters as follows:
\begin{enumerate}\addtolength{\itemsep}{-4pt}
\item We set $V_E = (v_1, \ldots, v_m)$ and for each $e_\ell =
\{x_i, y_j\} \in E$, we set the preference order of $v_\ell$ to be
$x_i > y_j > C - \{x_i,y_j\}$.
\item We set $A_{X,Y} = (a_1, \ldots, a_T)$ and we set their
preference orders so that for each $v \in X \cup Y$, $A_{X,Y}$
contains exactly $D - d(v)$ voters with preference orders of the
form $v > b_t > C - \{v,b_t\}$ (see Item \ref{chooseb}. below regarding how
candidates $b_t$ are chosen).
\item We set $A_p = \{a_{T+1}, \ldots, a_{T+D}\}$, where each voter
$a_{T+i}$, $1 \leq i \leq D$, has preference order of the form $p
> b_t > C - \{p,b_t\}$ (see Item \ref{chooseb}. below regarding how
candidates $b_t$ are chosen).
\item\label{chooseb} We arrange the preference orders of voters in $A_{X,Y} + A_p$
so that each candidate $b_t$, $1 \leq t \leq T+D$, receives
exactly $1$ point.
\end{enumerate}
Thus, before deleting any of the voters, each candidate in $\{p\}
\cup X \cup Y$ has score $D \geq 2$ and each candidate in $B$ has
score $1$. We form instance $I$ of $2$-Approval-\#CCDV consisting of
$E = (C,V)$, designated candidate $p$, and with $n$ as the limit on
the number of voters we are allowed to delete.
We claim that the number of solutions for $I$ is equal to the number
of perfect matchings in $G$. Let $M \subseteq E$ be a perfect
matching in $G$. We form collection $V' = \{v_i \mid e_i \in
M\}$. Clearly, $\|V'\| \leq n$ and in election $E' = (C,V-V')$
candidate $p$ is the unique winner ($p$ has $D$ points, candidates
in $X \cup Y$ have $D-1$ points, and candidates in $B$ have $1$
point).
On the other hand, let $V'$ be a subcollection of $V$ such that
$\|V'\| \leq n$ and $p$ is the unique winner of election $E' =
(C,V-V')$. Since $p$ is the unique winner of $E'$, it must hold that
each of the $2n$ candidates in $X \cup Y$ has at most $D-1$ points
in $E'$. Thus, since $|V'| \leq n$, it must be the case that in
fact $\|V'\| = n$ and $V'$ is a subcollection of $V_E$ such that
each candidate from $X$ appears in the first position of exactly one
vote in $V'$ and each candidate from $Y$ appears in the second
position of exactly one vote from $V'$. As a result, $E' = \{e_i
\mid v_i \in V'\}$ is a perfect matching for $G$.
\end{proof}
\begin{theorem}
In the unrestricted case, for each $k \geq 3$, $k$-Approval-\#CCAV
and $k$-Approval-\#CCDV are $\sharpPclass$-metric-complete.
\end{theorem}
\begin{proof}
The proof for the case of \#CCDV follows by applying a natural
padding argument in the reduction from the proof of
Theorem~\ref{thm:ccavdv-2}. Thus we focus on the case of \#CCAV.
We first give a proof for the case $k=3$.
We give a Turing reduction from \#PerfectMatching to
$3$-Approval-\#CCAV. Let $G = (G(X),G(Y),G(E))$ be our input
bipartite graph, where $G(X) = \{x_1, \ldots, x_n\}$ and $G(Y) =
\{y_1, \ldots, y_n\}$ are sets of vertices, and $G(E) = \{e_1,
\ldots, e_m\}$ is the set of edges. We form an election $E = (C,V)$
and a collection $W$ of unregistered voters as follows. We set $C =
\{p,d\} \cup B \cup G(X) \cup G(Y)$ (where $B$ is a collection of dummy
candidates to be specified later) and we set $V = (v_1,\ldots, v_t)$,
where the preference orders of the voters in $V$ are such that
the score of $p$ is $0$, the score of $d$ is $n-1$, the score
of each candidate in $G(X) \cup G(Y)$ is $n-2$, and the score of
each dummy candidate in $B$ is either $1$.
We set $W = (w_1, \ldots, w_m)$, where for each $\ell$, $1 \leq \ell \leq m$, if
$e_\ell = \{x_i,y_j\}$ then $w_\ell$ has preference order $p > x_i > y_j
> C - \{p,x_i,y_j\}$.
We form an instance $I$ of $3$-Approval-\#CCAV with election $E
= (C,V)$, collection $W$ of unregistered voters, designated
candidate $p$, and the number of voters that can be added set to
$n$.
It is easy to verify that the number of $3$-Approval-\#CCAV
solutions for $I$ is the number of matchings in $G$ of cardinality
exactly $n$. (One has to add exactly $n$ voters for $p$ to defeat
$d$; each unregistered voter corresponds to an edge in $G$ and one
cannot add two edges that share a vertex as then the candidate
corresponding to that vertex would have score $n$, and $p$ would not
be able to become the unique winner.) Thus, the number of perfect
matchings in $G$ is exactly the number of solutions for $I$.
Further, clearly it is possible to implement our reduction in
polynomial time.
The case for $k > 3$ follows by natural padding arguments and
extending the dummy-candidates set $B$.
\end{proof}
Finally, we obtain the results for the destructive cases through
Theorem~\ref{thm:c-to-d} (we remark that it applies to
$2$-Approval-\#CCAV as well because Turing reductions are
generalizations of metric reductions; for the same reason in
Table~\ref{tab:results} we report the complexity for
$k$-Approval-\#CCAV and $k$-Approval-\#CCDV, $k \geq 2$, as
``$\sharpPclass$-Turing-complete,'' even though our results for $k \geq 3$
are slightly more precise).
\begin{corollary}\label{cor:dc-approval}
In the unrestricted case, $2$-Approval-\#DCAV is
$\sharpPclass$-Turing-complete and for each $k$, $k \geq 2$,
$(k+1)$-Approval-\#DCAV and $k$-Approval-\#DCDV are
$\sharpPclass$-metric-complete.
\end{corollary}
Let us now move on to the single-peaked case. Here we show
polynomial-time algorithms for all counting voter control problems
under $k$-Approval (for fixed $k$). Our algorithm is based on dynamic
programming. We use the following notation in the proof below: For a
given $j$-element integer vector ${\mathit{scores}}$ and integer $i$, $1 \le i \le j$,
we write ${\mathit{scores}}_i$ to denote the $i$-{th} element of vector
${\mathit{scores}}$ (i.e., ${\mathit{scores}} = ({\mathit{scores}}_1,...,{\mathit{scores}}_j)$).
\begin{theorem}
For the case of single-peaked profiles, for each fixed positive
integer $k$, $k$-Approval-\#CCAV, $k$-Approval-\#CCDV,
$k$-Approval-\#DCAV and $k$-Approval-\#DCDV are in $\FPclass$.
\end{theorem}
\begin{proof}
We give the proof for $k$-Approval-\#CCAV. The result for
$k$-Approval-\#CCDV follows by Theorem~\ref{thm:d-to-a} and the
results for the destructive cases follow by
Theorem~\ref{thm:c-to-d}.
We are given a candidate set $C$, a voter collection $V$, a
designated candidate $p \in C$, a collection $W$ of unregistered
voters, a positive integer $\ell$, and a societal axis
$\mathrel{L}$. We show a polynomial time algorithm for determining
the number of sets $W'$, $W' \subseteq W$, where $\|W'\| \le \ell$,
such that $p$ is a k-Approval winner in election $(C, V \cup W')$.
For each voter $v \in V \cup W$, we define ${\mathrm{top}}_k(v)$ to be the
set of $k$ most preferred candidates according to $v$. Let
$\mathrel{\widetilde{L'}}$ be a weak linear order over voter set $W
\cup V$, such that $v_1 \mathrel{\widetilde{L'}} v_2$, $v_1,v_2 \in
V \cup W$, if and only if there exists a candidate $c \in
{\mathrm{top}}_k(v_2)$ such that for each $c' \in {\mathrm{top}}_k(v_1)$ we have $c'
\mathrel{L} c$ or $c' = c$. Let $\mathrel{L'}$ be a strict order
obtained from $\mathrel{\widetilde{L'}}$ by breaking the ties in an
arbitrary fashion. For each voter $v \in V \cup W$ let $X^v = \{w
\mid w \in V \cup W \land w \mathrel{L'} v\}$ be the subset of
voters from $V \cup W$ that precede voter $v$ under $\mathrel{L'}$.
We let $v_l$ to be the last voter from $V$ under $\mathrel{L'}$.
For a given voter $v \in V \cup W$, integer $z \in \mathbb{Z}$, and
integer $s \in \mathbb{Z}$, we define $g(v,z,s)$ to be the number of
sets $\widetilde{W} \subseteq X^v \cap W$, where $\|\widetilde{W}\|
= s$, such that $p$ is a k-Approval winner in election $(C, (X^v
\cap V) \cup \widetilde{W} \cup \{v\})$ and in addition the score of
candidate $p$ in this election is exactly $z$. Equation
\eqref{k-app-ccav-main} below gives the result that we should output
on the given input instance of k-Approval-\#CCAV problem:
\begin{equation}\label{k-app-ccav-main}
\sum_{0 \le s \le \ell} \left( \sum_{0 \le z \le \|W \cup V\|} \left( g(v_l,z,s) + \sum_{\substack{w \in W \\ v_l \mathrel{L'} w}} g(w,z,s) \right) \right)
\end{equation}
Thus, it suffices to show how to compute $g(w,z,s)$ for $w \in V
\cup W$ and $z,s \in \mathbb{Z}$ in polynomial time. Before we give
an appropriate algorithm, we need to introduce some additional
notation.
For each candidate $c \in C$, let ${\mathrm{rank}}(c)$ be $c$'s rank under
$\mathrel{L}$ over all candidates from $C$ (so, for example, if $C =
\{a,b,c\}$ and $a \mathrel{L} b \mathrel{L} c$ then ${\mathrm{rank}}(a) = 1$,
${\mathrm{rank}}(b) = 2$ and ${\mathrm{rank}}(c) = 3$). For each voter $v$, let ${\mathrm{rank}}(v) =
\max \{ {\mathrm{rank}}(c) \mid c \in {\mathrm{top}}_k(v)\}$.
%
For each $v \in V \cup W$, let $\mathcal{P}^v= \{w \mid w \in (V
\cup W) \land w \mathrel{L'} v \land (\nexists t \in V)[ w
\mathrel{L'} t \mathrel{L'} v]\}$. In other words, $\mathcal{P}^v$
consists of the closest voter $u \in V$ that precedes $v$ under
$\mathrel{L'}$, and of all the voters that are between $u$ and $v$ under
$\mathrel{L'}$. When $v$ is the first element from $V$ under
$\mathrel{L'}$, then $\mathcal{P}^v$ contains all the voters from
$W$ preceeding $v$ under $\mathrel{L'}$.
%
For a $j$-element integer vector ${\mathit{scores}}$ and a nonnegative
integer $r$, let $\cutoff{{\mathit{scores}}}{r}$ denote $j$-element vector
$({\mathit{scores}}_1, \ldots, {\mathit{scores}}_r, 0,\ldots, 0)$ (that is, we replace
the last $j-r$ entries of vector ${\mathit{scores}}$ with zeros).
For each voter $w \in V \cup W$, we define ${\mathrm{approval}}(w)$ to be the
$\|C\|$-dimensional $0/1$ vector that for each candidate $c \in C$
has $1$ at position ${\mathrm{rank}}(c)$ if and only if $c \in {\mathrm{top}}_k(w)$.
(In other words, ${\mathrm{approval}}(w)$ is the $0/1$ vector that indicates
which candidates receive points from voter $w$.) Note that, due to
single-peakedness of the election, for each voter $w$,
${\mathrm{approval}}(w)$ contains exactly a single consecutive block of $k$
ones.
We are now ready to proceed with our algorithm for computing
function $g$. For a given voter $v \in V \cup W$, given integers
$z,s \in \mathbb{Z}$, and a $\|C\|$-dimensional integer vector
${\mathit{scores}}$, we define $f(v,z,s,{\mathit{scores}})$ to be the number of sets
$\widetilde{W} \subseteq X^v \cap W$ such that $\|\widetilde{W}\| =
s$ and in election $(C,(X^v \cap V) \cup \widetilde{W} \cup \{v\})$
the following holds: (1) $p$ scores exactly $z$ points and (2) each
candidate $c \in C -\{p\}$ scores no more than ${\mathit{scores}}_{{\mathrm{rank}}(c)}$ points.
For a given integer $r \in \mathbb{Z}$, let $\Gamma^{r}$ be the
vector $(r,...,r)$ of dimension $\|C\|$. Clearly, for a given voter
$v \in V \cup W$ and given integers $z,s \in \mathbb{Z}$ we have:
\begin{equation}\label{k-app-ccav-convert}
g(v,z,s) = f(v,z,s,\Gamma^{z-1})
\end{equation}
We now show a recursive formula for $f$. For a given voter $v \in V
\cup W$, a given vector ${\mathit{scores}}$ of (nonnegative) integers of
dimension $\|C\|$, and two integers $z,s \in \mathbb{Z}$, we have:
\begin{equation}\label{k-app-ccav-transform}
f(v,z,s,{\mathit{scores}}) = f(v,z,s,\cutoff{{\mathit{scores}}}{{\mathrm{rank}}(v)})
\end{equation}
This follows from the fact that in election $(C,X^v \cup \{v\})$
only candidates with ranks $j \le {\mathrm{rank}}(v)$ can score a point. It is
easy to see that $f(v,z,s,{\mathit{scores}}) = 0$ when $z < 0$ or ${\mathit{scores}}$
contains at least one negative entry, because the score is always a
nonnegative integer. When $s = 0$ then $f(v,z,s,{\mathit{scores}})$ is $1$ if
and only if in election $(C,(X^v \cap V) \cup \{v\})$ candidate $p$
scores $z$ points and each candidate $c \in C$ scores no more than
${\mathit{scores}}_{{\mathrm{rank}}(c)}$ points; otherwise $f(v,z,s,{\mathit{scores}})$ is $0$.
When $s > 0$, we note that each election consistent with the
condition for $f(v,z,s,{\mathit{scores}})$ contains at least one voter $w$
from $\mathcal{P}^v$. Eq.~\eqref{k-app-ccav-sum} below gives
formula for $f$ in case when $s > 0$; for each voter $w \in
\mathcal{P}^v$ we count all the sets $\widetilde{W} \subseteq X^v
\cap W$ such that $w$ directly precedes $v$ in $(X^v \cap V) \cup
\widetilde{W} \cup \{v\}$ under $L'$:
\begin{equation}\label{k-app-ccav-sum}
f(v,z,s,{\mathit{scores}}) = \sum_{w \in \mathcal{P}_v} f(w,z-z',s-s',{\mathit{scores}} - {\mathrm{approval}}(w)),
\end{equation}
where (1) $z' = 1$ if $p \in {\mathrm{top}}_k(v)$ and $z'=0$ otherwise, and
(2) $s' = 1$ if $v \in W$ and $s' = 0$ otherwise. By combining
Eq.~\eqref{k-app-ccav-transform} and \eqref{k-app-ccav-sum} we get:
\begin{equation}\label{k-app-ccav-sum-final}
f(v,z,s,{\mathit{scores}}) = \sum_{w \in \mathcal{P}_v} f(w,z-z',s-s',\cutoff{{\mathit{scores}} - {\mathrm{approval}}(w)}{{\mathrm{rank}}(w)}).
\end{equation}
We claim that function $g$ can be computed through
Eq.~\eqref{k-app-ccav-convert} in polynomial time, using standard
dynamic programming techniques to compute function $f$. The reason
is that to compute $f$ for the arguments as in
Eq.~\eqref{k-app-ccav-convert} using recursive formula
\eqref{k-app-ccav-sum-final}, we need to obtain $f$'s values for at
most $\|C\| (\|V\| + \|W\|)^{k+3}$ different arguments. This is
because starting from ${\mathit{scores}} = \Gamma^{z-1}$, the only allowed
transformations of ${\mathit{scores}}$ are given in
Eq.~\eqref{k-app-ccav-sum-final} and ensure that whenever we need to
compute $f$, the ${\mathit{scores}}$ vector is of the form
$(z-1,z-1,...,z-1,e_1,e_2,...,e_k,0,0,...,0)$, where $e =
(e_1,e_2,...,e_k)$ is some $k$-element vector of integers and for
each $i$, $1 \le i \le k$, we have $0 \le e_i \le z - 1$. Clearly,
there are no more than $\|C\|z^k$ vectors of this form. Thus, we
can compute $g$ in polynomial time and, through
Eq.~\eqref{k-app-ccav-main}, we can solve $k$-Approval-\#CCAV in
polynomial time.
\end{proof}
\subsection{Approval Voting and Condorcet Voting}
Let us now consider Approval voting and Condorcet voting. While these
two systems may seem very different in spirit, their behavior with
respect to election control is similar. Specifically, in the
unrestricted case, for both systems \#CCAV and \#CCDV are
$\sharpPclass$-complete, for both systems it is impossible to make some
candidate a winner by adding candidates, and for both systems it is
impossible to prevent someone from winning by deleting
candidates. Yet, for both systems \#DCAC and \#CCDC are in $\FPclass$ via
almost identical algorithms. There is, however, also one difference.
For the single-peaked case, we were able to find polynomial-time
algorithms for voter control under Condorcet, while the results for
Approval remain elusive.
\begin{theorem}\label{thm:approval-condorcet-avdv}
In the unrestricted case, each of Approval-\#CCAV, Approval-\#CCDV,
Condorcet-\#CCAV, and Condorcet-\#CCDV is $\sharpPclass$-complete. Their
destructive variants are $\sharpPclass$-metric-complete.
\end{theorem}
\begin{proof}
Note that the results for the destructive variants will follow by
Theorem~\ref{thm:c-to-d} and so we focus on constructive variants
only.
For the case of Approval, $\sharpPclass$-completeness of \#CCAV and
\#CCDV follows from the $\NPclass$-completeness proofs for their decision
variants given by Hemaspaandra, Hemaspaandra, and
Rothe~\shortcite{hem-hem-rot:j:destructive-control}; these proofs,
in effect, give parsimonious reductions from \#X3C to respective
control problems and, thus, establish $\sharpPclass$-completeness.
For the case of Condorcet, $\sharpPclass$-completeness of \#CCDV follows
from the proofs of Theorems 5.1 and 4.19 of Faliszewski et
al.~\cite{fal-hem-hem-rot:j:llull}, who effectively give a
parsimonious reduction from \#X3C to Condorcet-\#CCDV. The case of
Condorcet-\#CCAV appears to not have been considered in the
literature and thus we give a direct proof (naturally, we could
obtain $\sharpPclass$-Turing-completeness by noting that
Theorem~\ref{thm:d-to-a} gives a Turing reduction from
Condorcet-\#CCDV to Condorcet-\#CCAV, but $\sharpPclass$-completeness is
a stronger result). For the reminder of the proof we focus on
Condorcet-\#CCAV. The problem is clearly in $\sharpPclass$, so it
suffices to show that it is $\sharpPclass$-hard. We give a parsimonious
reduction from \#X3C.
Let $(B,\mathcal{S})$ be an instance of \#X3C problem, where
$B=\{b_1,\dots,b_{3k}\}$ and $\mathcal{S}=\{S_1,\dots,S_n\}$. We
create an election $E=(C,V)$, where $C=B\cup\{p\}$. Let $V$ consist
of $k-3$ voters with preferences $b_1\succ b_2\succ\cdots\succ
b_{3k}\succ p$. Thus, $b_1$ is the Condorcet winner of $E$, and
every candidate $b_i\in B$ beats $p$ in $k-3$ votes.
For each set $S_j\in\mathcal{S}$, $S_j =
\{b_{j_1},b_{j_2},b_{j_3}\}$, let $W$ contain a voter with
preference order $b_{j_1}\succ b_{j_2}\succ b_{j_3}\succ
p\succ\cdots$ (after $p$ the remaining candidates are ranked in
arbitrary order). We claim that every subset $W'$, $W'\subseteq W$,
such that $\|W'\|\le k$ and that $p$ is a Condorcet winner of election
$E'=(C,V\cup W')$ corresponds one-to-one to a $k$-element subfamily
${{\mathcal{S}}}'$ of ${{\mathcal{S}}}$ whose elements union up to $B$ (i.e., ${{\mathcal{S}}}'$
is an exact set cover of $B$).
First assume that there is a subfamily
$\mathcal{S'}\subseteq\mathcal{S}$ which is an exact set cover of
$B$. For each $S_j\in\mathcal{S'}$, we include the corresponding
voter from $W$ in $W'$. Let us consider election $E'=(C,V \cup W')$.
For each $b_i\in B$ we have $\N{E'}(b_i,p)=\N{E}(b_i,p)+1=k-2$ and
$\N{E'}(p,b_i)=\N{E}(p,b_i)+k-1=k-1$. Thus $p$ becomes the
Condorcet winner of $E'$.
Now assume that $p$ is the Condorcet winner in election $E'=(C,V
\cup W')$, where $W' \subseteq W$ and $\|W'\| \leq k$. For each $b_i
\in B$, there can be at most one voter in $W'$ that prefers $b_i$ to
$p$. This is so, because otherwise we would have
$\N{E'}(b_i,p)\ge\N{E}(b_i,p)+2=k-1$, and
$\N{E'}(p,b_i)\le\N{E}(p,b_i)+k-2=k-2$, and so $p$ would lose to
$b_i$. Thus, each $b_i$ is preferred to $p$ by either zero or one
voter from $W'$. If $b_i$ is preferred by one voter from $W'$, then
for $p$ to win, $p$ must be preferred to $b_i$ by $k-1$ voters from
$W'$, and since some voter must be added, it must hold that
$\|W'\|=k$.
If there are no voters in $W'$ who prefer $b_i$ to $p$, then since
each vote in $W'$ has some three candidates from $B$ ranked ahead of
$p$, there must be some other $b_{i'}$ that is ranked above $p$ by
more than one voter. This contradicts the requirement that for each
$b_j \in B$, at most one voter in $W'$ prefers $b_j$ to $p$. Hence,
each $b_i$ is preferred to $c$ by exactly one of the $k$ voters in
$W'$. Thus, the voters from $W'$ correspond to an exact set cover
of $B$.
\end{proof}
For the case of single-peaked preferences, we can use dynamic
programming to solve voter control problems under Condorcet.
\begin{theorem}
For the single-peaked case, Condorcet-\#CCAV, Condorcet-\#CCDV,
Condorcet-\#DCAV, and Condorcet-\#DCDV are in \FPclass.
\end{theorem}
\begin{proof}
We focus on Condorcet-\#CCAV. The remaining cases follow by applying
Theorems~\ref{thm:c-to-d} and~\ref{thm:d-to-a}.
We are given an election $E = (C,V)$ where $C$ is a set of
candidates and $V$ is a set of voters, a designated candidate $p \in
C$, a collection $W$ of unregistered voters, a nonnegative integer
$k$, and an order $\mathrel{L}$, the societal axis, such that $V$
and $W$ are single-peaked with respect to $\mathrel{L}$. We show a
polynomial-time algorithm for determining the number of sets $W'$,
$W' \subseteq W$, where $\|W'\| \le k$, such that $p$ is a Condorcet
winner in election $(C, V \cup W')$. Our algorithm is based on the
famous median-voter theorem.
We split $C$ into three sets, $C_a = \{c \mid c \in C \land c
\mathrel{L} p\}$, $C_b = \{c \mid c \in C \land p \mathrel{L} c\}$
and $C_m = \{p\}$; $C_a$ contains the candidates that are before $p$ on
the societal axis and $C_b$ contains the candidates that are after
$p$. For each voter $v$, by $c_v$ we mean the candidate that $v$
ranks first. For each collection $U$ of voters from $V \cup W$, we
define $U_a = \{v \mid v \in U \land c_v \in C_a\}$, $U_b = \{v \mid
v \in U \land c_v \in C_b\}$ and $U_m = \{v \mid v \in U \land c_v =
p\}$. In other words, $U_a$ and $U_b$ consist of those voters from $U$
for which the most preferred candidate is, respectively, in $C_a$ or
$C_b$, and $U_m$ contains those voters from $U$ that rank $p$ first.
For each $W' \subseteq W$, we define $\delta(W')$ to be $1$ exactly
if the following conditions hold:
\begin{enumerate}
\item $\|V_a\|+\|W'_a\| < \frac{1}{2}(\|V\|+\|W'\|)$, and
\item $\|V_b\|+\|W'_b\| < \frac{1}{2}(\|V\|+\|W'\|)$.
\end{enumerate}
Otherwise, we define $\delta(W') = 0$. The following lemma is an
expression of the well-known median voter theorem (we provide the
short proof for the sake of completeness).
\begin{lemma}
For each $W' \subseteq W$, $p$ is the Condorcet winner of election
$(C,V \cup W')$ if and only if $\delta(W') = 1$.
\end{lemma}
\begin{proof}
Assume that $\delta(W') = 1$ and consider an arbitrary candidate
$c$ other than $p$. For the sake of concreteness let us assume
that $c \in C_a$, but a symmetric argument holds for $c \in C_b$.
Due to single-peakedness of $V \cup W'$, no voter outside of $V_a$
and $W'_a$ prefers $c$ to $p$. However, since $\delta(W')=1$, we
have that $\|V_a\|+\|W'_a\| < \frac{1}{2}(\|V\|+\|W'\|)$, and so a
majority of the voters prefers $p$ to $c$. Since $c$ was chosen
arbitrarily, it holds that $p$ is a Condorcet winner.
For the other direction, assume that $\delta(W') = 0$. For the
sake of concreteness, let us assume that this is because
$\|V_a\|+\|W'_a\| \geq \frac{1}{2}(\|V\|+\|W\|)$. By this
assumption, there must be a candidate $c \in C$ that directly
precedes $p$ with respect to $L$ (that is, $c \mathrel{L} p$ and there is
no candidate $d$ such that $c \mathrel{L} d \mathrel{L} p$). Due to
single-peakedness of voters' preference orders, we have that every
voter in $V_a \cup W'_a$ prefers $c$ to $p$, and so $p$ is not a
Condorcet winner. A symmetric argument holds if $\|V_b\|+\|W'_b\|
\geq \frac{1}{2}(\|V\|+\|W\|)$.
\end{proof}
Thus our algorithm should output the number of sets $W'$, $W'
\subseteq W$, of cardinality at most $k$, such that $\delta(W')=1$
holds. However, to evaluate $\delta(W')$ it suffices to know the
values $\|W'_a\|$, $\|W'_b\|$, and $\|W'_m\|$. For each three
integers $\ell_a$, $\ell_b$, and $\ell_m$ we define
$\gamma(\ell_a,\ell_b,\ell_m)$ to be $1$ exactly if the following
two conditions hold (these conditions are analogous to those for
$\delta$):
\begin{enumerate}
\item $\|V_a\|+\ell_a < \frac{1}{2}(\|V\|+\ell_a+\ell_b+\ell_m)$, and
\item $\|V_b\|+\ell_b < \frac{1}{2}(\|V\|+\ell_a+\ell_b+\ell_m)$.
\end{enumerate}
Otherwise, $\gamma(\ell_a,\ell_b,\ell_m)=0$. It is easy to see that
for each $W' \subseteq W$ we have $\delta(W') =
\gamma(\|W'_a\|,\|W'_b\|,\|W'_m\|)$. Now it follows that the number
of subsets $W' \subseteq W$ of cardinality at most $k$ such that
$p$ is a Condorcet winner of election $(C,V\cup W')$ is exactly:
\[
\sum_{\ell=0,...,k}
\sum_{\substack{
\ell_a, \ell_b, \ell_m \in \mathbb{N} \\
\ell_a + \ell_b + \ell_m = \ell
}}
\gamma(\ell_a,\ell_b,\ell_m)
{\|W_a\| \choose \ell_a}
{\|W_b\| \choose \ell_b}
{\|W_m\| \choose \ell_m}.
\]
We can evaluate this sum in polynomial-time. This completes the
proof.
\end{proof}
For the case of candidate control, we get polynomial-time algorithms
even for the unrestricted case.
\begin{theorem} \label{th:apdcc} Approval-\#CCDC, and
Condorcet-\#CCDC are in $\FPclass$, even in the unrestricted case.
\end{theorem}
\begin{proof}
Let us handle the case of Approval first. Let $I = (C,V,p,k)$ be an
instance of approval-\#CCDC. The only way to ensure that $p \in C$
is the unique winner is to remove all candidates $c \in C-\{p\}$
such that $\score{(C,V)}^a(c)\geq\score{(C,V)}^a(p)$. Such
candidates can be found immediately. Let's assume that there are
$k_0$ such candidates. After removing all of them, we can also
remove $k-k_0$ or less of any remaining candidates other than $p$.
Based on this observation we provide the following simple algorithm.
\begin{codebox}
\Procname{$\proc{approval-\#CCDC}(C,V,p,k)$} \li Let $k_0$ be the
number of candidates $c \in C-\{p\}$,
\zi \>s.t.\ $\score{(C,V)}^a(c)\geq\score{(C,V)}(p)$. \\[-4mm]
\li \Return $\sum_{i=0}^{k-k_0}\binom{\|C\|-k_0-1}{i}$
\end{codebox}
Clearly, the algorithm is correct and runs in polynomial-time.
For the case of Condorcet voting, it suffices to note that if $p$ is
to be a winner, we have to delete all candidates $c \in C - \{p\}$
such that $\N{(C,V)}(p,c) \leq \N{(C,V)}(c,p)$. Thus, provided that
we let $k_0$ be the number of candidates $c \in C - \{p\}$ such that
$\N{(C,V)}(p,c) \leq \N{(C,V)}(c,p)$, the same algorithm as for the
case of approval voting works for Condorcet voting.
\end{proof}
\begin{theorem} \label{th:apacd}
Both Approval-\#DCAC and Condorcet-\#DCAC are in $\FPclass$.
\end{theorem}
\begin{proof}
We first consider the case of approval voting. Let $I =
(C,A,V,p,k)$ be an instance of Approval-\#DCAC, where $C = \{p,c_1,
\ldots, c_{m-1}\}$ is the set of registered candidates, $A = \{a_1,
\ldots, a_{m'}\}$ is the set of additional candidates, $V$ is the
set of voters, $p$ is the designated candidate, and $k$ is the upper
bound on the number of candidates that we can add. We will give a
polynomial-time algorithm that counts the number of
up-to-$k$-element subsets $A'$ of $A$ such that $p$ is not the
unique winner of election $(C \cup A',V)$.
Let $A_0$ be the set of candidates in $A$ that are approved by at
least as many voters as $p$ is. To ensure that $p$ is not the unique
winner of the election (assuming $p$ is the unique winner prior to
adding any candidates), it suffices to include at least one
candidate from $A_0$. Thus, we have the following algorithm.
\begin{codebox}
\Procname{$\proc{Approval-\#DCAC}(C,A,V,p,k)$}
\li \If $p$ is not the unique winner of $(C,V)$ \\[-4mm] \label{apacd:checking_c}
\li \Then \Return $\sum_{i=0}^k\binom{\|A\|}{i}$ \\[-4mm] \label{apacd:returning_result1}
\End
\li Let $A_0$ be the set of candidates $a_i \in A$\!,
\zi \>s.t.\ $\score{(C\cup A,V)}^a(a_i)\ge\score{(C\cup A,V)}^a(p)$. \label{apacd:defining_k0}
\li $\id{result}:=0$ \label{apacd:result_init}
\li \For $j:=1$ \To $k$ \label{apacd:counting_loop}
\li \Do $\id{result}:=\id{result}+\sum_{i=1}^{\min(\|A_0\|,j)}\binom{\|A_0\|}{i}\binom{\|A-A_0\|}{j-i}$ \label{apacd:counting}
\End
\li \Return $\id{result}$ \label{apacd:returning_result2}
\end{codebox}
The loop from line~\ref{apacd:counting_loop}, for every $j$, counts
the number of ways in which we can choose exactly $j$ candidates from
$A$; it can be done by first picking $i$ of the candidates in $A_0$
(who beat $p$), and then $j-i$ of the candidates in $A - A_0$. It is
clear that the algorithm is correct and runs in polynomial time.
Let us now move on to the case of Condorcet voting. It is easy to see
that the same algorithm works correctly, provided that we make two
changes: (a) in the first two lines, instead of testing if $p$ is an
approval winner we need to test if $p$ is a Condorcet winner, and (b)
we redefine the set $A_0$ to be the set of candidates $a_i \in A$ such
that $\N{(C \cup A,V)}(p,a_i) \leq \N{(C \cup A,V)}(a_i,p)$. To see that these two
changes suffice, it is enough to note that to ensure that $p$ is not a
Condorcet winner of the election we have to have that either $p$
already is not a Condorcet winner (and then we can freely add any
number of candidates), or we have to add at least one candidate from
$A_0$.
\end{proof}
We conclude our discussion of Condorcet voting by noting that for the
single-peaked case all our results for Condorcet directly translate to
all Condorcet-consistent rules. The reason for this is that if voters'
preferences are single peaked, then there always exist weak Condorcet
winners. Whenever weak Condorcet winners exist, they are the sole
winners under Condorcet-consistent rules by definition. Since we focus
on the unique-winner model, if $R$ is a Condorcet-consistent rule, $E$
is a single-peaked, and $c$ is a candidate, then $c$ is the unique
$R$-winner of $E$ if and only if $c$ is the unique Condorcet winner of
$E$. In effect, we have the following corollary.
\begin{corollary}\label{cor:condorcet-consistent}
Let $R$ be a Condrocet-consistent rule. For the single-peaked case,
$R$-\#CCDC, $R$-\#DCAC, $R$-\#CCAV, $R$-\#CCDV, $R$-\#DCAV, and
$R$-\#DCDV are in $\FPclass$.
\end{corollary}
\subsection{Maximin Voting}
For the case of Maximin, we consider the unrestricted case
only. Maximin is Condorcet-consistent and, thus, for the single-peaked
case we can use Corollary~\ref{cor:condorcet-consistent}.
The complexity of decision variants of control for Maximin was studied
by Faliszewski, Hemaspaandra, and
Hemaspaandra~\shortcite{fal-hem-hem:j:multimode}. In particular, they
showed that under Maximin all voter control problems are
$\NPclass$-complete and an easy adaptation of their proofs gives the
following theorem.
\begin{theorem}
Maximin-\#CCAV and Maximin-\#CCDV are $\sharpPclass$-complete, and
Maximin-\#DCAV and Maximin-\#DCDV are $\sharpPclass$-metric-complete.
\end{theorem}
On the other hand, among the candidate control problems for Maximin,
only Maximin-CCAC is $\NPclass$-complete (DCAC, CCDC, and DCDC are in
$\Pclass$). Still, this hardness of control by adding candidates translates
into the hardness of all the counting variants of candidate control.
\begin{theorem}
Maximin-\#CCAC is $\sharpPclass$-complete and Maximin-\#DCAC is
$\sharpPclass$-metric-complete.
\end{theorem}
\begin{proof}
Faliszewski et al.~\cite{fal-hem-hem:j:multimode}'s proof that
Maximin-CCAC is $\NPclass$-complete can be used without change; their
reduction from X3C to Maximin-CCAC is also correct as a parsimonious
reduction from \#X3C to Maximin-\#CCAC. The result for
Maximin-\#DCAC follow through Theorem~\ref{thm:c-to-d}.
\end{proof}
The cases of Maximin-\#CCDC and Maximin-\#DCDC are more complicated
and require new ideas because decision variants of these problems are
in $\Pclass$.
\begin{theorem}
Both Maximin-\#CCDC and Maximin-\#DCDC are $\sharpPclass$-Turing-complete.
\end{theorem}
\begin{proof}
We consider the \#CCDC case first. Clearly, the
problem belongs to $\sharpPclass$ and it remains to show hardness. We
will do so by giving a Turing reduction from \#PerfectMatching.
Let $G = (G(X),G(Y),G(E))$ be our input graph, where $G(X) = \{x_1,
\ldots, x_n\}$ and $G(Y) = \{y_1, \ldots, y_n\}$ are sets of
vertices, and $E = \{e_1, \ldots, e_m\}$ is the set of edges. For
each nonnegative integer $k$, define $g(k)$ to be the number of
matchings in $G$ that contain exactly $k$ edges (e.g., $g(n)$ is the
number of perfect matchings in $G$).
We define the following election $E = (C,V)$. We set $C = G(E) \cup
S \cup B \cup \{p\}$, where $S = \{s_0, \ldots, s_{n}\}$ and $B =
\{b_{i,j}^\ell \mid \text{$0 \leq \ell \leq n$, $i < j$, and $e_i$
and $e_j$ share a vertex}\}$. To build voter collection $V$, for
each two candidates $a, b \in C$, we define $v(a,b)$ to be a pair of
voters with preference orders $a > b > C - \{a,b\}$ and
$\revnot{C-\{a,b\}} > a > b$. We construct $V$ as follows:
\begin{enumerate}
\addtolength{\itemsep}{-4pt}
\item For each $s_i \in S$, we add pair $v(s_i,p)$.
\item For each $s_i \in S$, we add two pairs $v(s_i,s_{i+1})$,
where $i+1$ is taken modulo $n+1$.
\item For each $s_i \in S$ and each $e_t \in E$, we add two pairs
$v(s_i,e_t)$.
\item For each $e_i, e_j \in E$, $i < j$, where $e_i$ and $e_j$
share a vertex, and for each $\ell$, $0 \leq \ell \leq n$, we add
two pairs $v(e_i,b_{i,j}^\ell)$ and two pairs
$v(e_j,b_{i,j}^\ell)$.
\end{enumerate}
Let $T$ be the total number of pairs $v(a,b)$, $a,b \in C$, included
in $V$. By our construction, the following properties hold:
\begin{enumerate}
\addtolength{\itemsep}{-4pt}
\item $\score{E}^m(p) = T-1$ and it is impossible to change the
score of $p$ be deleting $n$ candidates or fewer (this is because
there are $n+1$ candidates $s_i \in S$ such that $N_E(p,s_i) =
T-1$.
\item For each $s_i \in S$, $\score{E}^m(s_i) = T-2$, but deleting
$s_{i-1}$ (where we take $i-1$ modulo $n+1$) increases the score
of $s_i$ to $T+1$.
\item For each $e_t \in E$, $\score{E}^m(e_t)$ is $T-2$.
\item For each $b_{i,j}^\ell \in B$, $\score{E}^m(b_{i,j}^\ell) = T-2$ and
it remains $T-2$ if we delete either $e_i$ or $e_j$, but it
becomes $T$ if we delete both $e_i$ and $e_j$.
\end{enumerate}
Note that $p$ is the unique winner of $E$. For each $k$, $0 \leq k
\leq n$, we form instance $I(k) = (C,V,p,k)$ of Maximin-\#CCDC. We
define $f(k) = \#I(k) - \#I(k-1)$. That is, $f(k)$ is the number of
solutions for $I(k)$ where we delete exactly $k$ candidates. We
claim that for each $k$, $1 \leq k \leq n$, it holds that $ f(k) =
\sum_{j=0}^k {\|B\| \choose j} g(k-j). $ Why is this so? First,
note that by the listed-above properties of $E$, deleting any subset
$C'$ of candidates from $C$ that contains some member of $S$
prevents $p$ from being a winner. Thus, we can only delete subsets
$C'$ of $C$ that contains candidates in $G(E) \cup B$. Let us fix a
nonnegative integer $r$, $0 \leq r \leq n$. Let $C' \subseteq G(E)
\cup B$ be such that $p$ is the unique Maximin winner of $E' =
(C-C',V)$ and $\|C'\| = r$. Let $r_B = \|C' \cap B\|$ and $r_{G(E)}
= \|C' \cap G(E)\|$. It must be the case that for each $e_i, e_j
\in G(E)$, $i < j$, where $e_i$ and $e_j$ share a vertex, $C'$
contains at most one of them. Otherwise, $E'$ would contain at least
one of the candidates $b_{i,j}^\ell$, $0 \leq \ell \leq n$, and this
candidate would have score higher than $p$. Thus, the candidates in
$C' \cap G(E)$ correspond to a matching in $G$ of cardinality
$r_{G(E)}$. On the other hand, since $r_B \leq n$, $C \cap B$
contains an arbitrary subset of $B$. Thus, there are exactly ${\|B\|
\choose r_B} g(r_{G(E)})$ such sets $C'$. Our formula for
$f(k)$ is correct.
Now, using standard algebra (a process similar to Gauss
elimination), it is easy to verify that given values $f(1), f(2),
\ldots, f(n)$, it is possible to compute (in this order) $g(0),
g(1), \ldots, g(n)$. Together with the fact that constructing each
$I(k)$, $0 \leq k \leq n$, requires polynomial time with respect to
the size of $G$, this proves that given oracle access to
Maximin-\#CCDC, we can solve \#PerfectMatching. Thus, Maximin-\#CCDC
is $\sharpPclass$-Turing-complete and, by Theorem~\ref{thm:c-to-d}, so is
Maximin-\#DCDC.
\end{proof}
\section{Related Work}
\label{sec:related}
The focus of this paper is on the complexity of predicting election
winners for the case, where we are uncertain about the structure of
the election (the exact identities of candidates/voters that
participate), yet we have perfect knowledge of voters' preference
orders. However, our model is just one of many approaches to winner
prediction, which in various forms and shapes has been studied in the
literature for some years already. For example, to model imperfect
knowledge regarding voters' preferences, Konczak and
Lang~\shortcite{kon-lan:c:incomplete-prefs} introduced the possible
winner problem, further studied by many other researchers (see, e.g.,
\cite{con-xia:j:possible-necessary-winners,bet-dor:j:possible-winner-dichotomy,bac-bet-fal:c:counting-pos-win,che-lan-mau-mon:c:possible-winners-adding,lan-mon-xia:c:new-alternatives-new-results}). In
the possible winner problem, each voter is represented via a partial
preference order and we ask if there is an extension of these partial
orders to total orders that ensures a given candidate's victory.
Bachrach, Betzler, and
Faliszewski~\shortcite{bac-bet-fal:c:counting-pos-win} extended the
model by considering counting variants of possible winner problems.
Namely, they asked for how many extensions of the votes a given
candidate wins, in effect obtaining the probability of the candidate's
victory. This is very similar to our approach, but there are also
important differences. In the work of Bachrach et al., we have full
knowledge regarding the identities of candidates and voters
participating in the election, but we are uncertain about voters'
preference orders. In our setting, we have full knowledge about
voters' preference orders, but we are uncertain about the identities
of candidates/voters participating in the election.
Another model of predicting election outcomes is that of Hazon et
al.~\cite{haz-aum-kra-woo:j:uncertain-election-outcomes}. They
consider a situation where each voter is undecided regarding several
possible votes. That is, for each voter we are given several possible
preference orders and a probability distribution over these votes.
The question is, what is the probability that a designated candidate
wins.
From a technical standpoint, our research continues the line of work
on the complexity of control. This line of work was initiated by
Bartholdi, Tovey, and Trick~\cite{bar-tov-tri:j:control}, and then
continued by Hemaspaandra, Hemaspaandra, and
Rothe~\shortcite{hem-hem-rot:j:destructive-control} (who introduced
the destructive cases), by Meir et
al.~\shortcite{mei-pro-ros-zoh:j:multiwinner} (who considered
multiwinner rules and who generalized the idea of the constructive and
destructive cases), by Faliszewski, Hemaspaandra, and
Hemaspaandra~\shortcite{fal-hem-hem:j:multimode} (who introduced
multimode model of control), by Faliszewski, Hemaspaandra, and
Hemaspaandra~\cite{fal-hem-hem:c:weighted-control} (who were first to
consider control for weighted elections), by Rothe and
Schend~\cite{rot-sch:c:experimental-control} (who initiated the
empirical study of the complexity of control problems), and by many
other researchers, who provided results for specific voting rules and
who introduced various other novel means of studying control problems
(see, e.g., the following
papers~\cite{bet-uhl:j:parameterized-complecity-candidate-control,erd-now-rot:j:sp-av,erd-rot:c:fallback-voting,fen-liu-lua-zhu:j:parameterized-control,liu-zhu:j:maximin,men-sin:c:schultze-control,par-xia:strategic-schultze-ranked-pairs};
we also point the readers to the
survey~\cite{fal-hem-hem:j:cacm-survey}).
Single-peaked elections were studied for a long time in social choice
literature, but they gained popularity in the computational social
choice world fairly recently, mostly due to the papers of
Walsh~\cite{wal:c:uncertainty-in-preference-elicitation-aggregation}
and Conitzer~\cite{con:j:eliciting-singlepeaked}. Then, Faliszewski et
al.~\cite{fal-hem-hem-rot:j:single-peaked-preferences} and, later,
Brandt et al.~\cite{bra-bri-hem-hem:c:sp2} studied the complexity of
control problems for single-peaked elections. Recently, Faliszewski,
Hemaspaandra, and Hemaspaandra~\cite{fal-hem-hem:j:nearly-sp}
complemented this line of work by studying nearly single-peaked
profiles.
Going in a different direction, our work is very closely related to
the paper of Walsh and Xia~\shortcite{wal-xia:c:lot-based} on
lot-based elections. Walsh and Xia study a model of Venetian
elections, where a group of voters of a given size is randomly
selected from a group of eligible voters, and the votes are collected
from these selected voters only. In this setting, the problem of
computing a candidate's chances of victory, in essence, boils down to
the counting variant of control by adding voters problem. Thus, our
paper and that of Walsh and Xia are quite similar on the technical
front. The papers, however, have no overlap in terms of results.
\section{Conclusions and Future Work}\label{sec:conclusions}
We have considered a model of predicting election winners in
settings where there is uncertainty regarding the structure of the
election (that is, regarding the exact set of candidates and the exact
collection of voters participating in the election). We have shown
that our model corresponds to the counting variants of election
control problems (specifically, we have focused on election control by
adding/deleting candidates and voters). We have considered Plurality,
Approval, Condorcet, $k$-Approval, and Maximin (see
Table~\ref{tab:results} for our results). For the former three, the
complexity of counting variants of control is analogous to the
complexity of decision variants of respective problems, but for the
latter two, some of the counting control problems are more
computationally demanding than their decision counterparts.
Many of our results indicate computational hardness of winner
prediction problems. To alleviate this issue to some extent, we also
considered single-peaked preferences that are more likely to appear in
practice. In this case, we got polynomial-time results only (except
the case of Approval, where have no results for the single-peaked
case). Still, sometimes in practice one might have to seek heuristic
algorithms or approximate solutions (e.g., sampling-based algorithms
similar to the one of Bachrach, Betzler, and
Faliszewski~\cite[Theorem~6]{bac-bet-fal:c:counting-pos-win}).
There are many ways to extend our work. For example, in the
intruduction we mentioned the model where for each voter $v$ (or
candidate $c$) we have probability $p_v$ (probability $p_c$) that this
voter (candidate) participates in the election. We believe that
studying this problem in more detail would be very interesting. As we
have argued, the model where all values $p_v$ ($p_c$) are identical,
reduces to our setting, but the cases where the probabilities can
differ remain open.
\medskip
\noindent\textbf{Acknowledgements.}
An early version of this paper appeared in IJCAI Workshop on Social
Choice and Artificial Intelligence (2011) and an extended abstract was
presented at the Twenty-Sixth AAAI Conference on Artificial
Intelligence (AAAI-12). We thank both the AAAI and WSCAI reviewers
for very helpful feedback. Piotr Faliszewski was in part supported by
AGH University of Technology Grant 11.11.230.015 (statutory project),
by Foundation for Polish Science's program Homing/Powroty, and by
Poland's National Science Center's grant 2012/06/M/ST1/00358.
\label{sec:summary}
\bibliographystyle{plain}
|
\section{Introduction}
The primary goal of the heavy ion collisions at the highest energy of the Relativistic
Heavy Ion Collider (RHIC) and the energy of the Large Hadron Collider (LHC) is to explore
the properties of the extreme hot and dense matter, the quark-gluon plasma (QGP), existed
in the early stages of the collisions. Recent investigations \cite{ISF1214,Sch12} indicate
that the initial systems created in the ultrarelativistic heavy ion collisions at the RHIC
and LHC energies are not uniform in space, and there are event-by-event fluctuations of the
system initial quantities, due to the fluctuations of nucleon distributions in the nuclei,
the fluctuations in the color charge distributions inside a nucleon, and combined with
highly Lorentz contraction. The studies of the system evolution with the fluctuating
initial conditions (FIC) and the influence of the FIC on final particle observables
are recently very interesting issues in high energy heavy ion collisions \cite{ISF1214}.
They are important to improve our understanding of the experimental results at the RICH
and the LHC.
Elliptic flow and two-particle Hanbury-Brown-Twiss (HBT) correlation functions are important
observables in high energy heavy ion collisions. They reflect the transverse (perpendicular
to the beam direction) anisotropic pressure property and the space-time structure of the
particle-emitting sources, respectively. In ultrarelativistic heavy ion collisions, the
spectators depart from the reaction region quickly after collision, and a very hot and
dense fireball is created in the mid-rapidity region. For uniform systems of the fireballs
the odd-order azimuthal flow harmonics are expected to be zero. However, recent studies
indicate that the fluctuating inhomogeneous density distributions of the initial systems
may lead to nonzero triangular flow, and thus inspires the investigations of azimuthal
triangular flow and even higher-order flow harmonics \cite{{ISF1214,Sch12,AlvRol10,Alv10a,
StaShu11,Sch11,vn-exp,Gal13,LXHan11,LMa14,Bra14}}. In Ref. \cite{Gal13}, the authors
reproduced well the experimental flow results of $v_2$, $v_3$, $v_3$, and $v_5$ in the
ultrarelativistic heavy ion collisions at the RHIC and the LHC, by using the IP-Glasma
FIC \cite{Sch12} with the viscous hydrodynamic model of MUSIC \cite{Sch11}. The
investigations imply that the fluctuations in the initial geometry state are important
and the created medium behaves as a nearly perfect liquid of nuclear matter because it
has an extraordinarily low ratio of shear viscosity to entropy density.
On event-by-event basis, the particle-emitting sources with the FIC are bumpy and
inhomogeneous \cite{Gyu97,Osa02,RenZha08,Wer10,Sch11,HuZha13}. This inhomogeneous
structure may lead to the fluctuations of final observables event-by-event. Although
the influence of the IFC can be analysed by some observables, for instance, nonzero
$v_3$, our motivation here is to detect the fluctuations of final observables directly
and try to look for the relationship between the fluctuations of final observables and
the granular inhomogeneity of the initial sources.
In this paper, we use the Heavy Ion Jet Interaction Generator (HIJING) \cite{Wan05} to
generate the FIC of the particle-emitting sources for the heavy ion collisions of the
Au-Au at $\sqrt{s_{NN}}=200$ GeV at the RHIC and the Pb-Pb at $\sqrt{s_{NN}}=2.76$ TeV
at the LHC. The system evolution is described by the relativistic ideal hydrodynamics
in (2+1) dimensions with the Bjorken longitudinal boost invariance \cite{Bjo83}, and
with the equation of state (EOS) of s95p-PCE \cite{She10}. We use the relativistic
Harten-Lax-Leer-Einfeldt (RHLLE) algorithm \cite{Ris98,HLLE,Ris9596,ZhaEfa,Zha04,Yu08Yin12}
in our hydrodynamic calculations. Although this algorithm is also valid for the viscous
hydrodynamic source with rest and smoothed initial conditions \cite{Efa12CPC}, it is hard
to obtain the stable numerical solutions for the viscous hydrodynamics with the FIC and
nonzero initial velocities of fluid-cells in the source with the RHLLE algorithm.
We calculate the pion elliptic flow, triangular flow, and HBT correlation functions for
the ideal hydrodynamic sources with the FIC and the nonzero initial fluid velocity.
Motivated by the works on the fluctuations of single-event HBT correlation functions for
granular sources \cite{WonZha04,Zha05,RenZha08,HuZha13}, we investigate the fluctuations
of the flow harmonics and HBT correlation functions of event subcollections for the FIC
sources. We introduce a granularity length to describe the granular inhomogeneity of
the initial sources and investigate its relationships with the fluctuations of the flow
harmonics and HBT correlation functions. Our investigations indicate that the FIC lead
to event-by-event fluctuations of the elliptic flow, triangular flow, and HBT correlation
functions. These FIC-caused fluctuations can be detected by the fluctuation distributions
of the observables of event subcollections. The fluctuations of the triangular flow of
event subcollections are sensitive to the granularity length of the initial source.
This dependence provide a way to investigate the granular inhomogeneity of the initial
sources through analysing the fluctuations of triangular flow in ultrarelativistic
heavy ion collisions.
The rest of the paper is organized as follows. In Sec. II, we will give a description
on solving the hydrodynamic equations in (2+1) dimensions and present the space-time
evolutions of the sources with the FIC. In Sec. III, we will investigate the fluctuations
of the pion elliptic flow, triangular flow, and HBT correlation functions in the event
subcollections for the heavy ion collisions of the Au-Au collisions at $\sqrt{s_{NN}}=200$
GeV at the RHIC and the Pb-Pb at $\sqrt{s_{NN}}=2.76$ TeV at the LHC. The relationship
between the fluctuations of the observables of event subcollections and the granular
inhomogeneity of initial source is investigated also in Sec. III. Finally, we will give
the summary and discussions in Sec. IV.
\section{Hydrodynamic evolution of the sources with HIJING FIC}
The description of ideal hydrodynamics for the system with zero net-baryon density is defined
by the local conservations of energy and momentum\cite{Ris98,KolHei03},
\begin{equation}\label{hy1}
\partial_{\mu}T^{\mu\nu}=0,
\end{equation}
where $T^{\mu\nu}\!=\!(\epsilon\!+\!{\cal P})u^{\mu}u^{\nu}\!-\!{\cal P} g^{\mu\nu}$ is the
density tensor of energy-momentum of ideal fluid, $\epsilon$ and ${\cal P}$ are the energy
density and pressure in the local rest frame of the fluid element which moving with velocity
$\textbf{\emph{v}}$, $u^{\mu}=\gamma(1,\textbf{\emph{v}})$ is the four-velocity, $\gamma\!=\!
(1\!-\!\textbf{\emph{v}}^{2})^{-1/2}$, and $g^{\mu\nu}\!=\mathrm{diag}(+,-,-,-)$ is Minkowski
metric tensor. Under the assumption of Bjorken longitudinal boost invariance \cite{Bjo83},
the hydrodynamics in (3+1) dimensions reduces to (2+1) dimensions. In this case we need only
to solve the transverse equations of motion in $z=0$ plane, and the hydrodynamic solutions
at $z\ne0~(v^z=z/t)$ can be obtained by the longitudinal boost invariance hypothesis
\cite{Bay83,Gyu97}.
From Eq. (\ref{hy1}) we have the transverse equations in $z=0$ plane,
\begin{eqnarray}\label{hyeq2}
&&\hspace*{-8mm}\partial_t {\cal E}+\partial_x [({\cal E}\!+\!{\cal P})v^x]+\partial_y
[({\cal E}\!+\!{\cal P}) v^y]=-{\cal F}({\cal E},{\cal P},t),
\nonumber\\
&&\hspace*{-8mm}\partial_t {\cal M}^x+\partial_x({\cal M}^x v^x\!+\!{\cal P})+\partial_y
({\cal M}^x v^y)=-{\cal G}({\cal M}^x,t),\\
&&\hspace*{-8mm}\partial_t {\cal M}^y+\partial_x({\cal M}^y v^x)+\partial_y({\cal M}^y
v^y\!+\!{\cal P})=-{\cal G}({\cal M}^y,t),
\nonumber
\end{eqnarray}
where ${\cal E}=T^{00}$, ${\cal M}^i=T^{0i},\,(i=x,y)$, ${\cal F}({\cal E},{\cal P},\,t)=
({\cal E}+{\cal P})/t$, and ${\cal G}({\cal M}^i,t)={\cal M}^i\!/t$. In equation set
(\ref{hyeq2}) there are $\epsilon$, ${\cal P}$, $v^x$, and $v^y$ four variables. So an EOS,
${\cal P}(\epsilon)$, is needed to enclose the equation set. In the calculations, we use
the EOS of s95p-PCE, which combines the hadron resonance gas at low temperature and the
lattice QCD results at high temperature \cite{She10}.
Assuming the local equilibrium of system is reached at time $\tau_{0}$, we construct
the initial energy density of the hydrodynamic source at $z=0$, by using the AMPT code
\cite{Lin05} in which the HIJING is used for generating the initial conditions, as
\cite{Gyu97,Pan12}
\begin{eqnarray}
\label{E-ic}
\epsilon(\tau_0,x,y;z=0)=K\sum_{\alpha}\frac{p_{\perp\alpha}}{\tau_0}\frac{1}{2\pi
\sigma_0^2}\,\exp \left\{-\frac{[x-x_{\alpha}(\tau_{0})]^{2}+[y-y_{\alpha}(\tau_{0}
)]^2}{2\pi\sigma_0^2}\right\}.
\end{eqnarray}
Here $p_{\perp\alpha}$ is the transverse momentum of parton $\alpha$ in the fluid element
at $(x,y)$, $x_{\alpha}(\tau_{0})$ and $y_{\alpha}(\tau_{0})$ are the transverse coordinates
of the parton at $\tau_{0}$, $\sigma_0$ is a transverse width parameter, and $K$ is a scale
factor which can be adjusted to fit the experimental data of produced hadrons \cite{Pan12}.
The initial velocity of the fluid element is then determined by the initial energy density
and the average transverse momentum of the partons in the element.
With the EOS and the initial values of energy density and velocities, we can solve equation
set (\ref{hyeq2}) using the relativistic HLLE scheme and Sod's operation splitting method
\cite{HLLE,Ris98,Ris9596,ZhaEfa,Zha04,Yu08Yin12,Sod77}: first getting the solutions for the
corresponding homogeneous equations in $x$ and $y$ directions; then obtaining the solutions
of (\ref{hyeq2}) with the corrections of ${\cal F}$ and ${\cal G}$ to the solutions of the
homogeneous equations. In our calculations the spatial grid sizes are taken to be $\Delta
x=\Delta y=0.1$ fm, and the time step is taken to be $\Delta t=0.99\Delta x$
\cite{Ris9596,ZhaEfa,Zha04,Yu08Yin12}. \\
\begin{figure*}[!htbp]
\begin{center}
\begin{minipage}{0.16\textwidth}
\hspace*{-1mm}\includegraphics[scale=0.5]{zfRHICb4_03}
\end{minipage}%
\begin{minipage}{0.16\textwidth}
\hspace*{-1mm}\includegraphics[scale=0.5]{zfRHICb4_06}
\end{minipage}%
\begin{minipage}{0.16\textwidth}
\hspace*{0mm}\includegraphics[scale=0.5]{zfRHICb4_09}
\end{minipage}%
\begin{minipage}{0.16\textwidth}
\hspace*{1mm}\vspace*{-1mm}\includegraphics[scale=0.5]{zfLHCb4_03}
\end{minipage}%
\begin{minipage}{0.16\textwidth}
\hspace*{1mm}\vspace*{-1mm}\includegraphics[scale=0.5]{zfLHCb4_06}
\end{minipage}%
\begin{minipage}{0.16\textwidth}
\hspace*{1mm}\vspace*{-1mm}\includegraphics[scale=0.5]{zfLHCb4_09}
\end{minipage}%
\caption{The transverse distributions of energy density at $z=0$ for the Au-Au collisions
at $\sqrt{s_{NN}}=200$ GeV at the RHIC and the Pb-Pb collisions at $\sqrt{s_{NN}}=2.76$
TeV at the LHC. The impact parameter $b$ for both the RHIC and LHC heavy ion collisions
is 4 fm. The $\sigma_0$ values are 0.3, 0.6, and 0.9 fm. The panels [(a1)-- (f1)],
[(a2)-- (f2)], [(a3)-- (f3)], [(a4)-- (f4)] are for the evolution time $t=$ 0, 3, 6, and
9 fm/$c$ after $\tau_0$, respectively. The unit of energy density is GeV/fm$^3$. }
\label{zf-hydro}
\vspace*{-5mm}
\end{center}
\end{figure*}
In Figs. \ref{zf-hydro}[(a1)-- (c1)] and \ref{zf-hydro}[(d1)-- (f1)], we show the transverse
distributions of the source initial energy density at $z=0$ for the Au-Au and Pb-Pb collisions
at the RHIC and LHC energies, respectively. Here the unit of energy density is GeV/fm$^3$.
The impact parameter $b$ for both the RHIC and LHC heavy ion collisions is 4 fm, and the
$\tau_0$ values for the RHIC and LHC collisions are taken to be 0.6 and 0.4 fm/$c$,
respectively. The energy densities at the evolution time 3, 6, and 9 fm/$c$ after $\tau_0$
are shown in the panels [(a2)-- (a4)], [(b2)-- (b4)], [(c2)-- (c4)], [(d2)-- (d4)], [(e2)--
(e4)], and [(f2)-- (f4)] for the RHIC and LHC sources with the different $\sigma_0$,
respectively. One can see from the panels [(a1)-- (c1)] and [(d1)-- (f1)] that the initial
energy density is fluctuated. There are hot spots and cold valleys in the systems.
We call this inhomogeneous structure the granular inhomogeneity of the initial source.
The maximum of the energy density of spot decreases when $\sigma_0$ increases. Also, the
spot number decreases with increasing $\sigma_0$. From the panels of time greater than
zero one can see that the sources are still inhomogeneous at the late stages of the evolution,
due to the initial fluctuations.
\vspace*{2mm}
\begin{figure}[!htb]
\includegraphics[scale=1.2]{zfdndptn.eps}
\caption{(Color online) The pion transverse momentum spectra for the Au-Au collisions at the
RHIC energy $\sqrt{s_{NN}}=200$ GeV [panel (a)] and the Pb-Pb collisions at the LHC energy
$\sqrt{s_{NN}}=2.76$ TeV [panel(b)]. The solid lines are the hydrodynamic results with the
FIC, and the dashed lines are the hydrodynamic results with the SIC which are obtained by
averaging the FIC over 100 events. The circle, triangle, and square symbols are the
experimental data at the RHIC \cite{PHEpt} and the LHC \cite{ALIpt}. }
\label{zf-dndpt}
\end{figure}
We show in Figs. \ref{zf-dndpt}(a) and \ref{zf-dndpt}(b) the pion transverse momentum spectra
calculated by the hydrodynamics with the FIC (solid lines) and the smoothed initial conditions
(SIC) which are obtained by averaging the FIC over 100 events, for the Au-Au and Pb-Pb collisions
at the RHIC and LHC energies, respectively. Here, the circle, triangle, and square symbols
are the experimental data at the RHIC \cite{PHEpt} and the LHC \cite{ALIpt}. In the hydrodynamic
calculations, we take the particle rapidity cuts the same as in the experimental analyses at the
RHIC \cite{PHEpt} and the LHC \cite{ALIpt}, respectively. The freeze out temperature is taken
to be 130 MeV, and the parameter $\sigma_0$ is 0.6 fm. For the centralities 0--5\%, 10--20\%,
and mini bias, the regions of impact parameter are taken to be 0--2.3, 4.2--5.9, and 0--10.2 fm,
respectively \cite{STAv2-05}. It can be seen that the hydrodynamic results with the FIC are
consistent with the experimental data. At large $p_T$, the spectrum of the hydrodynamic source
with the FIC is higher than the corresponding spectrum of the hydrodynamic source with the SIC.
\section{Fluctuations of pion flow harmonics and HBT correlation functions}
\subsection{Flow harmonics of event subcollections}
In high energy heavy ion collisions, the invariant momentum distribution of final particles can be
written in the form of a Fourier series \cite{SV-YZ96,AP-SV98},
\begin{eqnarray}
\label{pdis}
E\frac{d^3N}{d^3p}\!=\!\frac{1}{2\pi}\frac{d^2N}{p_Tdp_Tdy}\!\bigg[1+\sum_n{2v_n\cos(n\phi-n\Psi_R)}
\bigg],
\end{eqnarray}
where $\phi$ is the azimuthal angle of the particle and $\Psi_R$ is the azimuthal angle of event
reaction plane. The first term on the right side of Eq. (\ref{pdis}) is the transverse momentum
spectrum, and the coefficients in the summation, $v_n=\langle \cos[n(\phi-\Psi_R)]\rangle$, are
the azimuthal $n$th-order flow harmonics, where $\langle \dots \rangle$ denotes the average over
the particles and events. In experimental data analyses, the reaction plane is usually replaced
by the event plane which is determined with the measured particles in an event
\cite{PHEv2-03,Vol08}. An alternative technique in flow analyses is the measurement of the
two-particle cumulant of azimuthal correlations, $[v_n\{2\}]^2=\langle \cos[n(\phi_1-\phi_2)]
\rangle$ \cite{Bor01,ALIv2-10}, which avoids the uncertainty in estimating reaction plane.
In this work, we calculate the integrated flow harmonics $v'_n$ and $p_T$-differential flow
harmonics $v_n(p_T)$ with the two-particle cumulant method \cite{Bor01} as,
\begin{equation}
\label{intf}
v'_n\{2\}\!=\!\bigg[\frac{N_n\{2\}}{N_{\rm pair}}\bigg]^{\!\frac{1}{2}}\!\!\!
=\!\bigg\{\!\frac{1}{N_{\rm pair}} \sum_{\alpha=1}^{N_{\rm evt}}\sum_{i\ne j}^{M} \!
\cos n(\phi_i-\phi_j)\bigg\}^{\!\!\frac{1}{2}}\!\!,
\end{equation}
\begin{equation}
\label{diff}
v_n\{2\}(p_T)=\frac{N_n\{2\}(p_T)}{N_{\rm pair}(p_T)} \frac{1}{v'_n\{2\}}\,,
\end{equation}
where $N_{\rm pair}=N_{\rm evt}M(M-1)/2$ is the total number of the particle pairs in
$N_{\rm evt}$ events, $M$ is the particle multiplicity of event, $N_{\rm pair}(p_T)$ and
$N_n\{2\} (p_T)$ are the counts of the particle pairs in the $p_T$ bin with the weights
1 and $\cos[n(\phi_i-\phi_j)]$, respectively.
\begin{figure}[!htbp]
\vspace*{3mm}
\includegraphics[scale=0.70]{zfRLv2.eps}
\caption{(Color online) The pion elliptic flow of the hydrodynamic sources for the Au-Au
collisions at the RHIC energy $\sqrt{s_{NN}}=200$ GeV and the Pb-Pb collisions at the LHC
energy $\sqrt{s_{NN}} =2.76$ TeV with impact parameters $b=$ 4 and 8 fm. The star, circle,
and triangle symbols are for the parameter $\sigma_0=$ 0.3, 0.6 and 0.9 fm, respectively. }
\label{zfRLv2}
\end{figure}
In Fig. \ref{zfRLv2} we plot the pion $p_T$-differential elliptic flow of the hydrodynamic
sources for the Au-Au collisions at the RHIC energy $\sqrt{s_{NN}}=200$ GeV and the Pb-Pb
collisions at the LHC energy $\sqrt{s_{NN}} =2.76$ TeV with impact parameters $b=$ 4 and 8 fm.
In the calculations, the number of events $N_{\rm evt}$ is six thousand and the particle-pair
number for each event is taken to be 10$^6$. One can see that the values of elliptic flow
decrease with increasing $\sigma_0$ for the RHIC sources, and are almost independent of
$\sigma_0$ for the LHC sources.
\begin{figure}[!htbp]
\vspace*{3mm}
\includegraphics[scale=0.63]{zfRLv2_sb.eps}
\vspace*{0mm}
\caption{(Color online) The pion elliptic flow of the event subcollections (dashed lines) each
of them with 100 events, for the RHIC and the LHC sources with $b=$ 4 and 8 fm, $\sigma_0=$ 0.3,
0.6, and 0.9 fm. The number of particle pair for each event is $10^6$. The solid lines are
the average results over sixty the subcollections. }
\label{zfRLv2_sb}
\end{figure}
For the hydrodynamic sources with the FIC, the elliptic flow and triangular flow of single
event vary event to event greatly. These event-by-event fluctuations are associated with
the FIC as well as the sample statistics. To reduce the effect of sample statistics on the
fluctuations, we will examine the elliptic flow and triangular flow of the event subcollections,
each of them has enough events. In Fig. \ref{zfRLv2_sb} the dashed lines show the pion elliptic
flow of the event subcollections each of them with 100 events, for the RHIC and the LHC sources
with different impact parameter $b$ and $\sigma_0$ values. The number of particle pair for each
event is $10^6$. The solid lines are the results averaged over sixty the subcollections. The
error bars of the flow harmonics of the event subcollections are calculated based on the
definition (\ref{diff}) and the statistical counts, by
\begin{eqnarray}
\label{error}
&&\Delta[v_n\{2\}(p_T)]=\Delta\!\!\left[\frac{N_n\{2\}(p_T)}{N_{\rm pair}(p_T)}\right]
\frac{1}{v'_n\{2\}}+\frac{N_n\{2\}(p_T)}{N_{\rm pair}(p_T)}\, \Delta\!\!\left[\frac{1}{v'_n\{2\}}
\right]\nonumber\\
&&\hspace*{5mm}=v_n\{2\}(p_T) \Bigg\{\frac{\Delta \left[N_n\{2\}(p_T)\right]}{N_n\{2\}(p_T)}
+\frac{\Delta \left[N_{\rm pair}(p_T)\right]}{N_{\rm pair}(p_T)} +\frac{\Delta \left[ N_n\{2\}
\right]}{2N_n\{2\}} +\frac{\Delta N_{\rm pair}}{2N_{\rm pair}}\Bigg\}\nonumber\\
&&\hspace*{5mm}=v_n\{2\}(p_T) \Bigg[\frac{1}{\sqrt{N_n\{2\}(p_T)}} +\frac{1}{\sqrt{N_{\rm pair}
(p_T)}} + \frac{1}{2\sqrt{N_n\{2\}}} + \frac{1}{2\sqrt{N_{\rm pair}}}\Bigg]\nonumber\\
&&\hspace*{5mm} \approx v_n\{2\}(p_T) \Bigg[\frac{1}{\sqrt{N_n\{2\}(p_T)}}
+\frac{1}{\sqrt{N_{\rm pair}(p_T)}} \Bigg]\,.
\end{eqnarray}
One can see from Fig. \ref{zfRLv2_sb} that the elliptic flow calculated with the 100 events
are still with great fluctuations. They are larger than the statistical errors. So, the
fluctuations of the flow of event subcollections reflect the intrinsic properties of the
sources.
The azimuthal flow harmonics are related to the initial eccentricities of the source
\cite{AlvRol10,LXHan11},
\begin{equation}
\label{eccen}
\varepsilon_n=\frac{\sqrt{\{\rho^2\cos(n\phi')\}^2 +\{\rho^2\sin(n\phi')\}^2}}{\{\rho^2\}},
\end{equation}
where $\phi'$ is the azimuthal angle related to the reaction plane ($xz$ plane), ${\hbox{\boldmath $\rho$}}=(x,y)$
is the transverse coordinate of initial source point, and $\{\cdots\}$ denotes the average
over initial source. In order to give an approximate estimation of the elliptic flow for
the FIC sources, we use a simple distribution of the initial source with separated droplets
to calculate the eccentricity $\varepsilon_2$. The single-event transverse distribution of
the initial granular source is given by \cite{WonZha04,YanZha09}
\begin{equation}
\label{disdrop}
D({\hbox{\boldmath $\rho$}})=\frac{1}{N_d(2\pi a^2)}\sum_{i=1}^{N_d}\exp\bigg[-\frac{({\hbox{\boldmath $\rho$}}-\textbf{\emph{R}}_{\perp
i})^2}{2 a^2}\bigg] ,
\end{equation}
where $N_d$ is the number of droplet, $\textbf{\emph{R}}_{\perp i}=(X_i\,,Y_i)$ are the transverse
coordinates of droplet centers, and $a$ is the standard deviation of the Gaussian distribution.
So, for a single event we have
\begin{eqnarray}
\label{eccen2c1}
\{\rho^2\cos(2\phi')\}=\frac{1}{N_d}\sum_{i=1}^{N_d}(Y_i^2-X_i^2)\,,~~~~~~
\{\rho^2\sin(2\phi')\}=\frac{1}{N_d}\sum_{i=1}^{N_d}2X_iY_i\,,
\end{eqnarray}
\vspace*{-5mm}
\begin{eqnarray}
\label{eccen2c2}
\{\rho^2\}=\frac{1}{N_d}\sum_{i=1}^{N_d}(2a^2+X_i^2+Y_i^2)\,.
\end{eqnarray}
Further, assume that the central coordinates of the droplets obey the Gaussian distribution,
$P(X_i,Y_i)\sim \exp(-X_i^2/2{\cal R}_x^2 -Y_i^2/2{\cal R}_y^2)$. By substituting the summation
$\frac{1}{N_d}\sum_i$ in Eqs. (\ref{eccen2c1}) and (\ref{eccen2c2}) with $\int dX_idY_i P(X_i,
Y_i)$ and completing the integrations, we get
\begin{eqnarray}
\{\rho^2 \cos(2\phi')\}={\cal R}_y^2 -{\cal R}_x^2\,,~~~~\{\rho^2\sin(2\phi')\}=0\,,~~~~
\{\rho^2\}=2a^2+{\cal R}_x^2+{\cal R}_y^2\,,
\end{eqnarray}
and the $\varepsilon_2$ for a huge number of events is
\begin{equation}
\label{eccen2a}
\varepsilon_{2}=\frac{{\cal R}_y^2 -{\cal R}_x^2}{2a^2+{\cal R}_x^2+{\cal R}_y^2}\,.
\end{equation}
The initial eccentricity decreases with increasing $a$, and the effect of the droplet radius
$a$ on $\varepsilon_2$ becomes slight when ${\cal R}_x$ and ${\cal R}_y$ are large. These
conclusions are consistent with the results in Fig. \ref{zfRLv2}, where $\sigma_0\sim a$, the
values of elliptic flow decrease with increasing $\sigma_0$ for the RHIC sources, and the
values of elliptic flow for the LHC sources are almost independent of $\sigma_0$ because the
LHC sources have larger initial source sizes. In Eqs. (\ref{eccen2c1}) and (\ref{eccen2c2}),
$N_d$, $X_i$, and $Y_i$ vary chaotically event to event. It leads to the fluctuations of the
elliptic flow of single event and the event subcollections with finite number of events as
shown in Fig. \ref{zfRLv2_sb}.
\begin{figure}[!htbp]
\vspace*{3mm}
\includegraphics[scale=0.63]{zfRLv3_sb.eps}
\vspace*{0mm}
\caption{(Color online) The pion triangular flow of the event subcollections (dashed lines) as
in Fig. \ref{zfRLv2_sb}. The solid lines are the average results over sixty the subcollections. }
\label{zfRLv3_sb}
\end{figure}
In Fig. \ref{zfRLv3_sb} the dashed lines show the triangular flow for the event subcollections
as in Fig. \ref{zfRLv2_sb}. The solid lines are the results averaged over sixty the
subcollections. As compared to the results of elliptic flow in Fig. \ref{zfRLv2_sb}, the
fluctuations of triangular flow are larger. The values of triangular flow are related to the
initial eccentricity $\varepsilon_3$, which becomes nonzero due to the granular inhomogeneous
structure of the initial sources. Because involving elliptic integrations, we cannot obtain
an analytic expression of $\varepsilon_3$ even for the simple droplet model.
\begin{table*}
\begin{center}
\caption{The values of $\sigma_0$ (column 2), eccentricities (columns 3 and 6), average flow
harmonics columns 4 and 7), scaled average flow harmonics (columns 5 and 8), average transverse
diameter (columns 9), and granularity length (column 10) of the initial sources for the RHIC
and LHC collisions with impact parameters $b=$ 4 and 8 fm. }
\begin{tabular}{l|cccccccccc}
\hline\hline
~~&~~$\sigma_0\text{(fm)}$~~&~~$\varepsilon_2$&~\,$\langle v_2\{2\}(p_T)\rangle$\,&~
${\widetilde v}_2$~&~~~~\,$\varepsilon_3$~\,&~$\langle v_3\{2\}(p_T)\rangle$&~~${\widetilde
v}_3$~~&~~$2\{\rho\}$(fm)&~$L_{\xi}$(fm)~\\
\hline
~~RHIC~ ~ & 0.3 & ~0.189 & 0.082 & 0.434 & \,~~~~0.091 & 0.071 & 0.780 & ~~6.57 & 6.80\\
~$b=4$ fm~& 0.6 & ~0.183 & 0.072 & 0.393 & \,~~~~0.085 & 0.061 & 0.718 & ~~6.68 & 7.67\\
~ & 0.9 & ~0.172 & 0.065 & 0.378 & \,~~~~0.076 & 0.052 & 0.684 & ~~6.87 & 9.47\\
\hline
~~RHIC~~ & 0.3 & ~0.375 & 0.137 & 0.365 & \,~~~~0.143 & 0.085 & 0.594 & ~~5.20 & 5.49\\
~$b=8$ fm~& 0.6 & ~0.356 & 0.126 & 0.351 & \,~~~~0.129 & 0.062 & 0.480 & ~~5.37 & 6.71\\
~ & 0.9 & ~0.322 & 0.117 & 0.363 & \,~~~~0.110 & 0.045 & 0.409 & ~~5.62 & 9.53\\
\hline
~~LHC~~ & 0.3 & ~0.172 & 0.063 & 0.366 & \,~~~~0.076 & 0.071 & 0.934 & ~~7.30 & 7.50\\
~$b=4$ fm~& 0.6 & ~0.167 & 0.062 & 0.371 & \,~~~~0.075 & 0.071 & 0.947 & ~~7.40 & 8.27\\
~ & 0.9 & ~0.165 & 0.062 & 0.376 & \,~~~~0.065 & 0.068 & 1.046 & ~~7.58 & 9.79\\
\hline
~~LHC~~ & 0.3 & ~0.367 & 0.131 & 0.357 & \,~~~~0.112 & 0.075 & 0.670 & ~~5.81 & 6.08\\
~$b=8$ fm~& 0.6 & ~0.350 & 0.130 & 0.371 & \,~~~~0.105 & 0.071 & 0.676 & ~~5.97 & 7.11\\
~ & 0.9 & ~0.325 & 0.125 & 0.385 & \,~~~~0.092 & 0.061 & 0.663 & ~~6.19 & 9.35\\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
In the third and sixth columns of Table I, we list the values of $\varepsilon_2$ and
$\varepsilon_3$ calculated in Eq. (\ref{eccen}) by averaging with the initial distributions
of energy density (see the top panels in Fig. \ref{zf-hydro}) of the six thousand events for
the RHIC and LHC sources with the different $b$ and $\sigma_0$ values, respectively. The
values of $\varepsilon_2$ and $\varepsilon_3$ increase with increasing $b$ and decrease
with increasing $\sigma_0$ for fixed $b$. The columns 4 and 7 of Table I list the average
values of $v_2\{2\}(p_T)$ and $v_3\{2\}(p_T)$ over $p_T$ and sixty the subcollections
each of them with 100 events. For fixed $b$, the variations of $\langle v_2\{2\}(p_T)
\rangle$ and $\langle v_3\{2\}(p_T)\rangle$ with $\sigma_0$ are consistent with the
variations of $\varepsilon_2$ and $\varepsilon_3$ with $\sigma_0$, respectively.
However, the values of $\langle v_3\{2\}(p_T)\rangle$ and $\varepsilon_3$ for the same
$\sigma_0$ but different $b$ are not completely consistent. For instance, the value of
$\varepsilon_3$ for the RHIC source with $\sigma_0=0.9$ fm and $b=4$ fm is 0.076 and
smaller than the value 0.110 for the RHIC source with the same $\sigma_0$ and $b=8$ fm,
but the corresponding $\langle v_3\{2\}(p_T)\rangle$ value for the former source is 0.052
and larger than the $\langle v_3\{2\}(p_T)\rangle$ value 0.045 for the later source.
This contradiction of the values of $\langle v_3\{2\}(p_T)\rangle$ and $\varepsilon_3$
for different $b$ indicates that there exist other effects on the triangular flow values
for different impact parameter. In the fifth and eighth columns of Table I, we present
the ratios ${\widetilde v}_2 \equiv\langle v_2\{2\}(p_T)\rangle/\varepsilon_2$ and
${\widetilde v}_3\equiv\langle v_3\{2\}(p_T) \rangle/\varepsilon_3$, respectively.
One can see that the values of the scaled average elliptic flow and triangular flow,
${\widetilde v}_2$ and ${\widetilde v}_3$, are insensitive to $\sigma_0$.
\subsection{Fluctuations of the pion flow harmonics and HBT correlation functions of event
subcollections}
From Figs. \ref{zfRLv2_sb} and \ref{zfRLv3_sb} it can be seen that the elliptic flow and
triangular flow of the event subcollections (dashed lines) are fluctuated.
These fluctuations are smoothed out in the observables for all the event subcollections
(solid lines). To display the flow fluctuations of the event subcollections in the analysis
for all events (or all event subcollections), we introduce the distribution $dN/df$ of the
differences,
\begin{equation}
f_{vn}=\big|v_n^{(i)}\{2\}(p_T)-v_n^{(j)}\{2\}(p_T)\big|, ~~~~(j\ne i),
\end{equation}
accumulating for all the event subcollections and the $p_T$ bins in the considered region.
Here, the superscript denotes the event subcollection.
Unlike some cumulate quantities which smooth out the fluctuations in the analysis for all
events, $dN/df$ is a fluctuation distribution. It becomes wide for the variables $v_n^{(i)}
\{2\}$ with large fluctuations, and the similar analysis was used in the investigations of
the fluctuations of single-event HBT correlation functions \cite{Zha05,RenZha08,HuZha13}.
Because the event number of collision is very huge (in principle it may reach any large
number if prolonging experiment time), the number of the event subcollections will be large
enough to overcome the influence of statistic fluctuations on the distributions. In Fig.
\ref{zfRLdnfb48}, we plot the distributions $dN/df_{vn}$ for the RHIC and LHC sources
with different impact parameter $b$ and $\sigma_0$ values. The widths of the distributions
increase with increasing $\sigma_0$, and in most case the widths decrease with increasing
$b$ for fixed $\sigma_0$. The distributions of $dN/df_{v3}$ are wider as compared to the
corresponding $dN/df_{v2}$ distributions, and the width of the distribution of triangular
flow is sensitive to $\sigma_0$.
\begin{figure}[!htb]
\vspace*{0mm}
\includegraphics[scale=0.63]{zfRLdnfb48.eps}
\vspace*{0mm}
\caption{(Color online) The distributions $dN/df_{vn}$ for the RHIC and the LHC sources with
different impact parameter $b$ and $\sigma_0$. }
\label{zfRLdnfb48}
\end{figure}
The wider the distributions $dN/df_{vn}$, the larger the fluctuations are. The width of the
distribution reflects the fluctuation magnitude of the flow harmonics. The changes of the
widths of the distributions $dN/df_{v2}$ and $dN/df_{v3}$ with impact parameter $b$ and
$\sigma_0$ reflect the variations of the fluctuation magnitudes with the granular inhomogeneity
of the initial sources. In order to quantify the granular inhomogeneity of the initial source,
we introduce a granularity length of the initial source as the product of the initial transverse
radius ${\cal R}_{\perp}$ and the transverse granularity parameter $\xi_{\perp}$ \cite{YanZha09}
of the initial source,
\begin{eqnarray}
\label{Lxi}
L_{\xi}={\cal R}_{\perp}\xi_{\perp}={\cal R}_{\perp}\frac{({\cal R}_{\perp}/\sigma_0)^2}{N_d-2}
\approx\frac{2\{\rho\}^3}{\{\rho\}^2-4\sigma_0^2}, ~~~~~~(2\sigma_0<\{\rho\}).
\end{eqnarray}
Here, we replace ${\cal R}_{\perp}$ approximately with the average transverse radius of the
initial source $\{\rho\}$, and replace $N_d$ approximately with $\frac{1}{2}(\{\rho\}/
\sigma_0)^2$, considering approximately the same numbers of the hot spot and cold valley in
the initial source. Clearly, the concept of granular source requires $N_d\geq2$. So we have
$2\{\rho\} <L_{\xi}<\infty$. From Eq. (\ref{Lxi}), the granularity length increase when the
initial source radius $\{\rho\}$ (associated with collision impact parameter and energy)
increases. For fixed $\{\rho\}$, $L_{\xi}$ increases with increasing $\sigma_0$ because the
droplet number decreases with increasing $\sigma_0$.
In the right two columns of Table I, we presents the values of $2\{\rho\}$ (calculated with
the initial energy density of source) and $L_{\xi}$ for the RHIC and the LHC sources respectively.
The values of $\{\rho\}$ decrease with increasing impact parameter $b$ and increase slightly with
increasing $\sigma_0$ for fixed $b$. The values of $\{\rho\}$ for the LHC sources are larger
than the corresponding results for the RHIC sources. In most case, the value of $L_{\xi}$ for
smaller $b$ is larger than that for larger $b$ with the same $\sigma_0$, because the initial
source radius is larger for smaller $b$. However, for $\sigma_0=0.9$ fm, the $L_{\xi}$ value
for the RHIC source with $b=8$ fm is larger than that for $b=4$ fm. This is because that the
effect of droplet number on the granularity length becomes important when $(\{\rho\}/
\sigma_0)^2$ is small.
From Fig. \ref{zfRLdnfb48} and the $L_{\xi}$ values in Table I, we observe that the widths of
the distributions $dN/df_{v2}$ and $dN/df_{v3}$ become wider as $L_{\xi}$ increases, for the
RHIC sources or the LHC sources. This is obvious for the triangular flow distributions. We
conclude that the fluctuation magnitude of the triangular flow of event subcollections is
sensitively dependent on the granularity length of the initial source. For the sources of
the same collision energy, the fluctuation magnitude of the triangular flow increases with
$L_{\xi}$ monotonically. However, for different collision energies, we observe the variation
of the distribution width with $L_{\xi}$ is not completely in a monotonic manner. For instance,
the $L_{\xi}$ values for the LHC source with $b=8$ fm and $\sigma_0=0.6$ fm is 7.11 fm and
larger than the $L_{\xi}$ value 6.71 fm for the RHIC source with the same $b$ and $\sigma_0$,
but the distribution width for the LHC source is smaller than that for the RHIC source.
This indicates the limitation of $L_{\xi}$ for the collisions with greatly different energies.
\begin{figure}[!htb]
\vspace*{0mm}
\includegraphics[scale=0.63]{zfRLCqso6_sb.eps}
\vspace*{0mm}
\caption{(Color online) The HBT correlation functions of the event subcollections (dashed lines)
each of them with the 10 events for the RHIC and the LHC sources with different impact parameters
and $\sigma_0=0.6$ fm. For each event the particle pair number is $10^6$. The solid lines are
the average results over six hundred the subcollections. }
\label{zfRLCqso}
\end{figure}
For granular sources, the single- or several-event HBT correlation functions are fluctuated
\cite{WonZha04,Zha05,RenZha08,HuZha13}. We plot in Fig. \ref{zfRLCqso} the two-pion HBT
correlation functions of the event subcollections (dashed lines) each of them with 10 the
events for the RHIC and the LHC sources with different impact parameters and $\sigma_0=
0.6$ fm. Here, $q_{\rm side}$ and $q_{\rm out}$ are the relative transverse momenta of
the pion pair in the ``side" and ``out" directions \cite{Ber88,Pra90}. The solid lines
are the results averaged over six-hundred the subcollections. We observe the fluctuations
of the HBT correlations of the event subcollections, and the fluctuations are smoothed out
in the average results. As the analyses for the flow fluctuations, we investigate the
distribution, $dN/df_{Cq}$, for the HBT correlation functions $C(q_{\rm side})$ and
$C(q_{\rm out})$. In Fig. \ref{zfRLCdnf}, we plot the distributions of $dN/df_{Cq}$
the RHIC and the LHC sources with different $b$ and $\sigma_0$ values. We observe that
the distributions are insensitive to $\sigma_0$, although they become wider as compared
to the distributions for the source with smoothed initial conditions \cite{RenZha08,HuZha13}.
Unlike elliptic flow and triangular flow, HBT correlation functions reflect more about the
source freeze-out geometry and dynamics rather than the initial details of the sources.
\begin{figure}[!htb]
\includegraphics[scale=0.60]{zfRLCdnfb48.eps}
\caption{(Color online) The distributions $dN/df_{Cq}$ for the RHIC and the LHC sources
with different values of impact parameter $b$ and $\sigma_0$. }
\label{zfRLCdnf}
\end{figure}
\begin{figure}[!htb]
\includegraphics[scale=0.60]{zfRLrms.eps}
\caption{(Color online) The variations of the root-mean-square $f_{\rm rms}$ of the elliptic
flow, triangular flow, and HBT correlation functions with $L_{\xi}$, for the RHIC and LHC
sources with $b=$ 4 and 8 fm. }
\label{zfRLrms}
\end{figure}
The width of the distribution $dN/df$, thus the fluctuation magnitude, can be described
by the root-mean-square (RMS), $f_{\rm rms}$, quantitatively. Finally, we plot in Fig.
\ref{zfRLrms} the variations of the RMS of the elliptic flow, triangular flow, and HBT
correlation functions with the granularity length of the initial source. One can see
that the RMS values of the flow harmonics increase as $L_{\xi}$ increases. But the RMS
values of the HBT correlation functions are almost independent of $L_{\xi}$. The RMS
values of the triangular flow are very sensitive to $L_{\xi}$.
This dependence provide a way to study the granular inhomogeneity of the initial sources
through the analyses of the fluctuations of triangular flow in ultrarelativistic heavy
ion collisions.
\section{Summary and discussions}
We investigate the fluctuations of pion elliptic flow, triangular flow,
and HBT correlation functions for the hydrodynamic sources of the Au-Au collisions
at $\sqrt{s_{NN}}=200$ GeV and the Pb-Pb collisions at $\sqrt{s_{NN}}=2.76$ TeV.
The initial sources are generated by the HIJING, and then evolve described by ideal
relativistic hydrodynamics in (2+1) dimensions with Bjorken's longitudinal boost
invariance. The EOS of s95p-PCE, which combines the hadron resonance gas at low
temperature and the lattice QCD results at high temperature, is employed in the
hydrodynamic calculations. For the hydrodynamic sources with the FIC, the
elliptic flow, triangular flow, and HBT correlation functions are fluctuated
event-by-event. These FIC-caused fluctuations survive in the observables obtained
with event subcollections. To display the fluctuations we introduce the fluctuation
distribution, $dN/df$, of the observable of event subcollections, which becomes wide
for the observable with large fluctuations. We also introduce the granularity length
$L_{\xi}$ to describe the granular inhomogeneity of the initial sources.
The relationships between the granularity length and the fluctuations of the pion
elliptic flow, triangular flow, and HBT correlation functions of event subcollections
are investigated. Our investigations indicate that the FIC lead to event-by-event
fluctuations of the elliptic flow, triangular flow, and HBT correlation functions.
These FIC-caused fluctuations can be detected by the fluctuation distributions of the
observables of event subcollections. The fluctuations of the triangular flow of event
subcollections are sensitive to the granularity length of the initial source.
This dependence provide a way to investigate the granular inhomogeneity of the initial
sources through analysing the fluctuations of triangular flow in ultrarelativistic
heavy ion collisions.
For the initial source with fluctuating matter distribution, the initial velocities
of fluid-cells in the source is usually nonzero and also fluctuated. Considering the
very low ratio of shear viscosity to entropy density of the QGP matter and the
stability of the hydrodynamic evolution of the FIC sources with nonzero initial
velocities of fluid-cells in the source, we use an ideal hydrodynamics model in
this work. On the other hand, we also did not consider the initial fluctuation
in longitudinal direction in this work. Further investigations on the relationship
between the source initial granular inhomogeneity and the fluctuations of final
observables for different initial source models and based on viscous hydrodynamic
description of source evolution will be of great interest.
\begin{acknowledgments}
We thank Dr. Longgang Pang and Dr. Luan Cheng for helpful discussions. This work is supported
by the National Natural Science Foundation of China under Contract No. 11275037.
\end{acknowledgments}
|
\section{Introduction}
\label{intro}
One of the most popular ideas in quantum theory of electronic structure of molecular and condensed matter systems is the concept of {\it atoms in molecules} (AiM). Its use allows one to understand some chemical-physical properties of a whole system by analyzing the characteristic properties of constituent atoms. Though there is a well-known formulation of ``quantum theory of atoms in molecules'' developed by Bader \cite{Bader:94, Bader:98}, unambiguous and commonly accepted definitions of such terms as ``partial atomic charges'' or ``state of an atom in a molecule'' do not exist. Using different theoretical backgrounds and pursuing certain goals one derives different results, which can be useful for some applications and not so useful for the others (e.g., see the discussion about partial atomic charges on p.~309 of Ref.\cite{Cramer:04}).
Several basic concepts and quantitative tools are widely exploited in the literature on the subject.
{\it Basic concepts}
such as the oxidation state (number), valence, formal charges, etc.\ are commonly used to characterize the charge states of atoms in molecules to get a preliminary idea about the chemical structure of a compound of interest.
\comtav{Referee: For example, when it is said that quantities like oxidation state, valence are not correctly defined from quantum mechanical point of view, what does it mean? <<<}
However, there are no well-defined representations of these concepts by observable quantities which would be commonly accepted by the physical-chemical community.
This means that the following applies for ``intuitively useful'' quantities such as oxidation state, valence etc.\
{\bf (i)} They cannot be presented as expectation values of some
unambiguous quantum mechanical
operators; %
{\bf (ii)} they cannot be uniquely determined
from experiments (even nominally, see \Table{table:tabl2} for different charge states of Pb causing the same chemical shifts);%
{\bf (iii)} they can be well-defined theoretically and/or experimentally but not very helpful from practical point of view to be used for
analysis of vital electronic properties of a chemical system (see next paragraph and ``Class III charges'' in \cite{Cramer:04}).
\comtav{>>>End of change}
{\it Quantitative approaches}
to describe the effective states of atoms in compounds are mainly based on using
Hartree-Fock, natural or localized orbitals and one-electron density matrices; alternatively, they
are originated on analysis of the total electronic density $\rho(\vec{r})$ of a chemical system, utilizing spatial criteria or reproducing some experimental data within simple theoretical models (e.g., see Ref.~\cite{Cramer:04} and text below).
Each of the known definitions has not only advantages but drawbacks which can seriously weaken the former in specific applications. In particular, the methods based on the
Mulliken and L\"{o}wdin population analyses \cite{Lowdin:70} strongly depend on the basis set used and are not so useful for large basis sets. The more elaborate concept of natural atomic orbitals (NAOs) \cite{Reed:88, Weinhold:05} overcomes the problem but the valence NAOs of an atom can notably differ in various chemical compounds and they are generally not localized on an atom, thus complicating the comparison of effective states of the atom in different molecular environments in terms of occupation numbers, etc.
\comtav{??? corrected <<<}
The methods utilizing the electronic densities directly, including those by Hirshfeld \cite{Hirshfeld:77}, Bader (see Ref.~\cite{Bader:94} and next paragraph), Voronoi, etc.\ (see Refs.~\cite{Cramer:04, Ramos-Cordoba:13}), suggest partitioning of a chemical compound
by using some spatial criteria in a manner that
is not an intrinsic quantum mechanical
hallmark
even if it is compatible with quantum mechanics description of the compound.
\comtav{>>> End}
The auxiliary concepts introduced by such a way are logical when determining the charge states of a given atom for particular cases only.
The most sensible of the methods for determining the effective states of atoms (and not just their partial charges) in chemical compounds described in the literature are anyway based on evaluation of some kind of overlapping the electronic wave function of a chemically bonded atom with those of a free atom and its ions. In other words, a mapping of the effective state of an atom in a molecular environment on states of the free atom takes place.
One of the natural approaches of this group described in the literature is recently formulated in Ref. \cite{Ramos-Cordoba:13}. The ``charge states'' of atoms in a molecule are determined there
when assigning to an atom the parts of molecular orbitals truncated with using the Bader zero-flux surfaces.
\comtav{Referee: Similarly, when the authors talk about zero flux
condition of Bader, it would be good to write in a couple of lines
what it implies. <<<}
These surfaces are defined by the equation
\begin{equation}
\nabla\rho(\vec{r}) \cdot \vec{n}(\vec{r}) = 0,
\end{equation}
for every point $\vec{r}$ on the interatomic boundary surface, with the unit vector $\vec{n}(\vec{r})$ normal to the surface.
Thus, the atoms constituting the chemical system can be separated by the spatial criterion using total electronic density only.
\comtav{>>> End of change}
The overlap integrals of molecular orbitals with atomic functions are calculated not over the whole space, but only in the areas bounded by zero-flux surfaces. These surfaces define the boundaries of atoms in a molecule within the Bader analysis and the atomic expansion coefficients of the molecular orbitals thus obtained are, obviously, consistent with the atomic Bader charges.
However, as is discussed above, such a definition of partial atomic charges and spatial partitioning into atomic regions
is not completely consistent with the quantum-mechanical description of electronic structures. In particular, the relativistic atomic orbitals (spinors) with different $l$, $j$ components for valence electrons usually have notably different spatial localizations and the atomic regions cannot be separated by any surfaces unambiguously.
~
To calculate the properties of molecules described by quantum mechanical operators heavily concentrated in atomic cores but sensitive to variation of densities of valence electrons, combined (two-step) approaches have been developed \cite{Titov:96, Titov:99, Titov:05a, Skripnikov:13b} and applied to study hyperfine structure (HFS), space parity (P), and time-reversal invariance (T) nonconservation (PNC) effects.
\comtav{Table~I: <<<}
Earlier calculations of these properties, which are in the context of the subject of this paper are given in \Table{table:tabl1} (see also discussion in the next section).
\comtav{>>> End}
In Ref.~\cite{Lomachuk:13} a method of evaluating chemical shifts of x-ray emission spectra (XES) for compounds containing period four and heavier elements, that is consistent with the two step approach is also proposed.
On the basis of these developments
one can introduce a method for determining the effective state of a given atom
in a chemical compound (substance).
This method is originated from the relativistic pseudopotential theory \cite{Titov:99, Petrov:04b, Mosyagin:06amin} and one-center restoration approaches \cite{Titov:05a, Titov:06amin} to recover proper electronic structure in heavy-atom cores after the relativistic pseudopotential simulation of a chemical substance. The present research can also be considered as a generalization of our computational models utilized to study the HFS and PNC effects as well as XES chemical shifts in molecules and solids.
\section{Motivation}
\label{sec1}
The observable properties and effective Hamiltonian parameters of our interest here include those that can be measured and those that can only be calculated and, thus, usually need to be checked by some appropriate way. The group of measurable properties comprise magnetic dipole and electric quadrupole HFS constants, XES chemical shifts, isomeric (chemical) shifts of M\"{o}ssbauer spectra, the volume effect of isotopic shifts. The unmeasurable effective Hamiltonian parameters cover those required to study T- and P-nonconservation effects in nuclei, atoms, molecules and solids: effective electric field $E_{\rm eff}$\ on (unpaired) electrons required for the electron electric dipole moment ({{\it e}EDM}) search; electronic density gradients on nuclei for Schiff moment; electromagnetic field on nuclei for anapole moment etc. Some examples of these studies are discussed in the following three paragraphs.
In Refs.~\cite{Dmitriev:92, Mosyagin:98, Isaev:04, Petrov:05a, Isaev:05a} the calculated HFS constants in various compounds are compared to the corresponding experimental data \cite{Knight:81, Mawhorter:11, Hunter:02, Kawall:05, Chanda:95} to estimate the errors in the effective electric field calculations of HgF, PbF, YbF, HI$^+$, and PbO, see \Table{table:tabl1}.
The XES lines correspond to electronic transitions between the shells localized in atomic cores. Comparing the experimental and theoretical XES data one can analyze the electronic structure (effective state) of an atom in a compound (see below). The studies intended to extract the information about the electronic structure in atomic cores in different chemical compounds from the XES data are performed, in particular, in Refs.~\cite{Raj:98, Raj:02, Pawlowski:02}. The authors estimate the $3d$ shell occupancies in various metals by comparing the measured $K\beta$ to $K\alpha$ x-ray intensity ratios to the results of atomic multiconfiguration Dirac--Fock computations. In Refs.~\cite{Sumbaev:76, Sovestnov:09} the electronic structures of various compounds were studied by comparing the experimental data on the x-ray chemical shifts to the results of corresponding atomic calculations.
It can be shown \cite{Dickson:05} that the chemical shifts of M\"{o}ssbauer spectra (isomer shifts) in various compounds are proportional to the differences of the total electronic densities on a given atomic nucleus. The measured isomer shift values as functions of the oxidation states (ligands) of the Fe, K, Ir, As,
\comtav{Ph <<<}
and other
\comtav{>>> End}
atoms in different compounds are given in Ref.~\cite{Dickson:05} together with the calculated electronic densities on the atomic nuclei as functions of electronic configurations for the neutral and ionized Fe and Sn atoms.
The discused properties and parameters characterize the processes occurring in atomic cores or, for stationary states, they are mean values of the operators heavily concentrated on nuclei or in atomic cores. By other words, the properties and parameters of our interest
strongly depend on
the electronic configuration ({\it effective state}) of a given {\it atom in a compound} (AiC)%
\footnote{We will further use this terminology and acronym AiC to distinguish them from the widely used terms ``atoms in molecules'' and AiM.}
rather than on the {\it chemical bonds} between atoms.
We call such characteristics the {\it AiC characteristics} or {\it AiC properties and parameters} (together with the ``core characteristics'', ``core properties'', and ``core parameters'' as in our earlier papers) assuming that the processes and quantum mechanical operators considered are spatially localized near nuclei despite the fact that not core but valence electrons usually give a key contribution to the given properties and parameters.
The ultimate aims pursued in our consideration are:
(i)~to formulate a robust model for description of the effective states of atoms in compounds;\\
(ii)~to attain close and unambiguous connection between the quantities which can be measured (chemical shifts of x-ray emission lines etc.) and theoretical models for their evaluation;\\
(iii)~to give insight into the quality of {\it ab~initio} calculations or semiempirical estimates of the AiC characteristics which cannot be (or are not yet) measured;\\
(iv)~to provide a unified tool for indirect (or ``independent'') accuracy check of the evaluated AiC characteristics;\\
(v)~to give a theoretical background for development of advanced (combined) computational schemes, which would be optimal (in the ratio quality to price) for their study;\\
(vi)~to make calculations more feasible (easier, faster, and more reliable) for computationally difficult cases (e.g., for complicated molecules and condensed matter structures containing heavy $d$ and $f$ elements).
It is well known that not all the well observable properties can be used for testing the calculated AiC characteristics but only those that have comparable sensitivity to variation of the electronic densities (or, generally, density matrix) in the vicinity of a nucleus due to electronic structure reorganization from one compound to the other, perturbations or electronic excitations in the valence region, etc. On the other hand, some properties that can serve as a good check for a given AiC characteristic in one kind of compounds are not suitable for the other ones. As an example, for such molecules-radicals as BaF, YbF and HgF \cite{Titov:06amin} with $sp-$hybridized state of unpaired (valence) electron, a good semiempirical estimate for $E_{\rm eff}$\ can be written as $E_{\rm eff}$${\sim}\sqrt{A*A_d}$, where $A=(A_{\parallel}{+}2A_{\perp})/3, A_d=(A_{\parallel}-A_{\perp})/3$, $A_{\parallel}$ and $A_{\perp}$ are magnetic dipole hyperfine structure parameters
\cite{Kozlov:95}. However, this formula is not so useful for the systems with $d$ and $f$ unpaired electrons only \cite{Skripnikov:11a}. So, a systematic analysis of applicability of some (measurable) properties to test the other (unmeasurable) ones is required.
The results of our earlier calculated values and experimental data for
$A_{\parallel}$, $A_{\perp}$ and electric quadrupole hyperfine structure constant $eQq_0$
in HgF, PbF, YbF, HI$^+$, and PbO molecules are presented in \Table{table:tabl1}.%
\comtav{Referee: I also found that presenting Table-I did not really
serve any purpose. Modified: <<<}
As one can see, the calculated HFS values are within $10-15$\% agreement with the experimental data. We should note that the calculations given in \Table{table:tabl1} are performed with using the approximations that are compatible with the theory given in the next section, whereas accuracy of modern, more sophisticated calculations is notably better (e.g., see recent results on ThO in Ref.~\cite{Skripnikov:14b}).
\comtav{>>> End}
\input{tabl1}
\section{Theory}
\label{sec2}
According to the above-discussed motivation we are going to formulate the AiC model for applications that satisfy the following basic criteria:
(i) correct quantum mechanical description;
(ii) common features of the AiC characteristics should be taken into account;
(iii) a good {\it quantitative} agreement of the AiC-theory predictions with
experiment.
The importance of the first criterion is discussed above, first of all it is concerns the chemical shifts, for which the present status of theory cannot be considered as satisfactory.
Taking into account the common features of the AiC characteristics is discussed below; it assumes exclusion of those computational elements from the model which do not affect essentially the AiC properties and parameters, allows one to reduce the computational cost visibly and makes the calculation more transparent.
As to the latter criterion, for different AiC characteristics and various kinds of compounds the term ``a~good agreement'' can have different meanings. Consider as an example the XES chemical shifts for heavy elements discussed in the introduction. They can be three to six orders of magnitude smaller than the energies of $K,L$ lines and the difficulty of their evaluation is highly aggravated by the two-electron nature of energetic properties \cite{Lomachuk:13}. Therefore, calculation of such chemical shifts with a good accuracy is a very serious problem in practice. In general, however, we assume the disagreement of an AiC-model prediction with experiment should be at most on the level of ${\sim}30$\% to provide a satisfactory qualitative description of an experiment, though, a benchmark {\it ab~initio} \ calculation can provide, in principle, a better accuracy.
\comtav{Added: <<<}
Note here that typical errors, $10{\div}15$~\%, for the HFS constants (\Table{table:tabl1}) and error estimations for PNC effects \cite{Titov:06amin}, which were earlier evaluated within simple two-step models \cite{Titov:05a}, are also compatible with this limitation.
That is important since one of main goals of the AiC theory is to provide accuracy check for those {\it unmeasurable} molecular parameters which are required to study PNC effects.
\comtav{>>> End}
The principal common feature of the AiC characteristics given above is that the direct contributions from the spatially valence region ($r{>}R_c$, the choice of the core radius $R_c$ is discussed in the next paragraph) to their values are relatively small, though, these are the valence electrons which determine the effective state of a free or bounded atom and mainly control the AiC characteristics taking into account the inactivity of core shells. The valence states contribute directly to these properties by only their small parts with the electronic density share ${<}10$\% localized in the atomic core ($r{<}R_c$) and not by the valence and outer regions ($r{>}R_c$) with the share ${>}90$\%
\comtav{Referee: it is not clear how the inter-atomic region enters into the
calculations. For example, if the core wavefucntions (spinors) of
atoms are affected by the outside regions so that the density at the
nucleus change, the cause of this change is non-transparent.}
\footnote{In calculations of heavy-atom compounds the core shells of the heavy atom(s) are usually treated as frozen. However, one can partially account for a relaxation of the corresponding core shells of an atom in different compounds by performing calculations of the isolated atom after determining the state of the atom in the compound. We use such approximation for estimating XES chemical shifts values in various lead compounds below in the paper (see~\Table{table:tabl3} for details).}.
\comtav{<<<End of change}
In turn, the native atomic potential from the nucleus and core electrons is ``hard'' for the valence electrons in its own atomic core, i.e., it is much higher by the amplitude than the potentials of other atoms or external sources. It is also much more by amplitude than the energies of the one-electron {\it valence and low-lying virtual} (W) states. Note that the valence and low-lying virtual states may change places with each other in different environments of the atom, processes etc., so one should treat them on equal footing when studying the effective states of an atom in different compounds.
We take account of relativistic effects and, thus, use spherical spinors
($\Omega_{ljm}(\vec{r}/r)$ or $|ljm\rangle$,
where $l,j$~are the orbital and total angular momentum quantum numbers, $m$ is the projection of $j$) for the spin-angular part of W states.
Neglecting the outer potentials and energies of the W states in the atomic core,
the property of proportionality or homogeneous scaling of the W spinors takes place in the core ({\it W proportionality} below) as is shown in \Figs{fig1} and (\ref{fig2}) \cite{Petrashen:56, Flambaum:90, Titov:99, Titov:02Dis}.
\comtav{Referee: Furthermore, it is stated in the context of Fig. 2 that the valence
and virtual spinors are proportional to each other but it is not
explained what its implications are. <<<}
This property allows one to introduce the {\it W reduced density matrix} (given below)
that leads to a more intuitive
(basis-set independent and ``minimally-sufficient'')
formulation of the density matrix concept to evaluate the AiC characteristics (that is discussed in detail below).
\comtav{>>> End}
\begin{figure}[h!]
\caption{Large components of the $6p_{1/2}$, $7p_{1/2}$ spinors of Pb for the $5s^2\,5p^6\,5d^{10}\,6s^26p^2$ configuration. The large components of $6p_{1/2}$ and scaled $7p_{1/2}$ spinors in the core region are given in subfigure, where the scaling factor is chosen in such a way that the amplitudes of large components of these spinors are equal at $R_c=0.5$~a.u.}
\centering
\includegraphics[width=\columnwidth]{fig1}
\label{fig1}
\end{figure}
\begin{figure}[h!]
\caption{Large components of the $6p_{1/2}$ spinors of Pb for $[5s^2\,5p^6\,5d^{10}]6s^2\,6p^2$, $[\dots]\,6s^{1.11}\,6p^2$, and $[\dots]6s^{0.53}\,6p^{0.57}$ configurations. The first one corresponds to the ground state and the next two are roughly (according to Ref.~\cite{Kaupp:93}) equivalent to the states of Pb in the PbH$_4$ and PbF$_4$ molecules, correspondingly.}
\centering
\includegraphics[width=\columnwidth]{fig2}
\label{fig2}
\end{figure}
One can see from \Fig{fig1} that the valence and virtual spinors are proportional each other in the atomic core. \Fig{fig2} illustrates the proportionality of the valence spinor generated for different states of the atom.
The theoretical backgrounds of the AiC approach can be formulated as follows:
({\bf i}) the core radius $R_c$ is chosen for a given atom in a compound by such a way that the contributions to a characteristic of interest outside the sphere with this radius ($r{>}R_c$) can be neglected or small enough and valence spinors with same angular quantum numbers are proportional to each other
\comtav{Referee: Similarly, it is not apparent how $R_c$ is fixed, although the authors
do say in their conclusion that $R_c$ does not affect the results. <<<}%
\footnote{%
From a formal point of view, the smaller $R_c$, the more unambiguously $W$ proportionality can be defined. In practice, however, one needs to consider computational errors
at small $R_c$.
Besides, the larger radius of an AIC operator, the partial waves
with higher values of the angular quantum numbers $j$ and $l$
one should treat, in general, to attain appropriate accuracy for the characteristic of interest.
In applications such calculations are performed using (relativistic) pseudopotentials with the pseudoorbitals (pseudospinors) smoothed in some core region within a matching radius (detailed studies on the subject are given in \cite{Titov:99}). Within the hard-core (accurate) pseudopotential formulations the matching radii should be as small as possible. In practice, they are close to the last (by amplitude) maxima, $R^V_m$, of valence orbitals (thus the ``large-core'' pseudopotentials are generated) or, for better accuracy, to the last maxima, $R^{OC}_m$, of outermost core orbitals (for the ``small-core'' pseudopotentials). So,
for the all-electron four-component case
one can write
$R_c\le R^V_m$,
whereas, for the case of the pseudospinors smoothed in the atomic core, one should first restore proper four-component behavior of spinors in the core or use $R_c$ at least not less than the largest matching radius
\cite{Titov:05a}.};
\comtav{>>> End}
({\bf ii})
the W proportionality in the core is applied to generate unique relativistic
(four-component) reference spinors, $\eta_{ljm}(r)$ for $r{<}R_c$, which are not smoothed in the core. Being normalized within the sphere with radius $R_c$:
\begin{equation}
\int\limits_{r<R_c} r^2 dr |\eta_{ljm}(r)|^2 = 1\ ,
\label{eq:orthog}
\end{equation}
these functions become universal and almost independent from the state (states) of an atom and its ions for which they are constructed; here we assume the ground or low-lying states of the atom with excited or ionized valence electrons only. Consider further the set $\{\eta_{lj}(r)\}_{lj}$ such generated as the ``AiC basis'' for a given atom to describe its effective state in a chemical substance.
Designate the four-component W spinors on a given atom as
\begin{equation}
\varphi_{n_wljm}(r) =
\left(\begin{array}{l}
\varphi^{L}_{n_wljm}(r)\\
\varphi^{S}_{n_wljm}(r)
\end{array}\right) ,
\nonumber
\end{equation}
in accord to the conventional representation of atomic Dirac-Fock spinors, $\varphi_{n_xljm}(r)$, by means of large (L) and small (S) components.
The index $n_x$ can naturally be the principal quantum number not only for core ($n_c$) and valence shells but for the virtual ones when a finite and localized basis set or/and an external spherical wall-type potential on the atom is used to generate atomic spinors with $n_x$ that grows monotonically with the energies of $\varphi_{n_xljm}(r)$ states. Now the W states are those that match the appropriate choice of $n_w: n_c{<}n_w{<}n_r$, where the index
$n_r$ corresponds to the highly excited atomic states with energies comparable or more than the amplitude of the atomic potential within the sphere with radius $R_c$.
Thus, the W spinors in the core region with $r{<}R_c$ can be written as
\begin{equation}
\varphi^{<}_{n_wljm} (r) = k_{n_wljm} \eta^{\vphantom{W}}_{lj}(r)\, ,\ r{<}R_c\ ,
\label{eq:valorb}
\end{equation}
since the radial parts of W spinors with the same $l,j$ and different $m$ are also proportional each other in the core.
With this background the following reducing of a one-electron density matrix that describe the atomic, molecular and condensed matter structures can be performed:
(1) the valence and low-lying virtual molecular/crystal orbitals or spinors are reexpanded on a basis set of atomic (one-center) W spinors within the sphere with radius $R_c$:
\begin{equation}
\begin{split}
\psi^{W<}_i(\vec{r}) \approx \sum\limits_{n_wljm} c^i_{n_wljm}
\left(\begin{array}{l}
\varphi^{L}_{n_wljm}(r)\Omega_{l,jm}(\vec{r}/r)\\
\varphi^{S}_{n_wljm}(r)\Omega_{2j{-}l,jm}(\vec{r}/r)
\end{array}\right) , \\
|\vec{r}|{<}R_c\ ,
\end{split}
\nonumber
\end{equation}
where $\Omega_{l,jm}$ and $\Omega_{2j{-}l,jm}$ are the conventional spin-angular factors for large and small components of an atomic bispinor.
(2) the one-electron density matrix (DM) of these spinors,
${\bm\rho}^W[\psi^{W<}_i]\equiv(\rho^{<}_{n_wljm,n_w'l'j'm'})$,
is calculated and reexpanded on the chosen atom using a one-center basis set. The ${\bm\rho}^W$ matrix can be obtained from the one-electron density matrix,
${\bm\rho}$,
that describes all electrons of the system as follows:
\begin{equation}
\begin{split}
\bm{\rho}^{W}= P_\mathrm{W}\bm{\rho}P_\mathrm{W} \equiv\\
(1-P_\mathrm{C}-P_\mathrm{R})\bm{\rho}(1-P_\mathrm{C}-P_\mathrm{R})\ ,
\end{split}
\label{eq:rho_w_begin}
\end{equation}
where $P_\mathrm{W}$ is the projector on the W states only, $P_\mathrm{C}$ and $P_\mathrm{R}$ are the projectors on the core states of a selected atom and the states having negligible densities in the core of this atom, correspondingly.
Note that none of both diagonal and off-diagonal submatrices between the $P_C, P_W$ and $P_R$ projectors are zero in general
when correlation, relaxation and mixing of different harmonics take place for an atom in a chemical substance. However, the core shells (particularly, for the small-core cases) can usually be treated as frozen atomic spinors with a very high accuracy, the DM will be diagonal on the core states and the corresponding submatrix can be safely removed from consideration even for XES chemical shifts \cite{Lomachuk:13}.
Moreover, the space of R states describing mainly the core relaxation and correlation effects do not contribute to the DM in these cases, they can be neglected and the \Eq{eq:rho_w_begin} can be rewritten as
\begin{equation}
\begin{split}
\bm{\rho}^{W} \approx\ (1-P_\mathrm{C})\bm{\rho}(1-P_\mathrm{C})\ .
\end{split}
\label{eq:rho_w_mid}
\end{equation}
The only a question in practice is to partition the one-electron states on the C, W and R subspaces correctly to minimize computational efforts for an accuracy of the interest.
Taking account of \Eq{eq:valorb} the $\bm{\rho}^{W}$ matrix can be transformed to a new one:
\begin{equation}
\Delta_{ljm,l'j'm'} = \sum\limits_{n_w n_w'}k_{n_wljm}k_{n_w'l'j'm'}\rho_{n_wljm,n_w'l'j'm'}.
\label{eq:red_dens}
\end{equation}
The new matrix is already reduced (summed up) on the principal quantum numbers $n$ for only W spinors which are taken into account, whereas the core and high-energy virtual states are excluded, so, call this matrix as the {\it W reduced DM}.
Because of the W proportionality, a physical meaning for the AiC characteristics has only the W reduced DM since it is generally impossible to distinguish the distribution of electrons by W spinors with fixed $l,j,m$ numbers in atomic cores by any available data on AiC properties. Moreover, partitioning the W space on individual states has a meaning only for free or weakly bound atoms, it is almost meaningless for a chemically bound atoms and, particularly, for condensed matter structures.
Let us consider the diagonal terms $\Delta_{ljm,ljm}$ of the W reduced DM. Multiplying these terms by the charge of the corresponding reference states $\left\{\eta_{ljm}\right\}$ within the sphere of radius $R_c$, one obtains the $\left\{q^{W}_{lj}\right\}$ quantities for a given atom taking into account \Eq{eq:orthog}:
\begin{equation}
q^W_{lj} = \sum_m \Delta_{ljm,ljm} \int\limits_{r<R_c} r^2 dr |\eta_{ljm}(r)|^2 = \sum_m\Delta_{ljm,ljm}\ ,
\label{eq:partialwavecharges1}
\end{equation}
which we call below the core region {\it partial wave charges}.
Alternatively, as one can easily see, the partial wave charges of an atom ``$A$'',
$q^{WA}_{lj}$, may be defined as the expectation values of the projection operators
\begin{equation}
P^{<,A}_{lj}=\sum_m\left|ljm\right\rangle\theta(R_c-|\vec{r}-\vec{R_A}|)\left\langle ljm\right|
\label{eq:Abar_tailless}
\end{equation}
on the W density matrix $\bm\rho^W$:
\begin{equation}
q^{WA}_{lj} = \mathrm{Tr}\left[P^{<,A}_{lj}\bm\rho^W\right].
\label{eq:partialwavecharges1}
\end{equation}
The Heaviside step function $\theta(R_c-|\vec{r}-\vec{R_A}|)$ is equal to unity in the core of atom ``A'' and zero outside:
\[
\theta(R_c - |\vec{r}{-}\vec{R_A}|) = \left\{\begin{array}{l}
1,\ |\vec{r}{-}\vec{R_A}|<R_c\\
0,\ \mbox{otherwise}
\end{array}\right. ,
\]
and $\left| ljm\right\rangle$ is the spherical spinor discussed above.
The operator (\ref{eq:Abar_tailless}) can be interpreted also as a ``tailless'' semilocal model potential (pseudopotential) of
Abarenkov-Heine type \cite{Abarenkov:65}
that is independent on the radial parts features of atomic spinors.
Thus, it can be easily utilized in most available quantum chemical codes to evaluate the partial wave charges on an atom. The only a limitation is in its using together with the DFT wavefunctions which do not allow one to construct DM correctly
but the electronic density only, so the accuracy in DFT calculations of partial wave charges can be, in principle, quite low.
\input{tabl2}
It is shown in Ref.~\cite{Lomachuk:13} that the XES chemical shifts between two compounds can be calculated as the differences of mean values of an effective one-electron operator in these compounds. In \Table{table:tabl2} the evaluated chemical shifts of energies of the $2p_{1/2}{\to}1s_{1/2}$, $3p_{1/2}{\to}2s_{1/2}$, and $5d_{3/2}{\to}4p_{1/2}$ transitions with respect to Pb$^{2+}$ for various ionic and excited states of an unbound Pb atom are listed together with the values of partial wave charges $q^W_{p1/2}$ of $p_{1/2}$ spinors in the considered states. These states (electronic configurations) differ by only the occupation numbers of W shells, which are chosen in such a way that the partial wave charges
are the same in both ionic and excited states. Due to the W proportionality (\ref{eq:valorb}) the chemical shifts depend on the partial wave charges only. One can see from \Table{table:tabl2} that the chemical shift of the ionic and excited states with the same partial wave charges agree well to each other with the highest relative error of 10\% for the $5d_{3/2}{\to}4p_{1/2}$ transition. Note here that our formulation cannot be attributed to one of the four classes (charge models I--IV), discussed by Cramer in \cite{Cramer:04} since it is well defined theoretically and reproduce well the experimental data.
\input{tabl3}
In \Table{table:tabl3} the partial wave charges $q^{W}_{lj}$ and effective occupation numbers of the valence Pb (sub)shells in the PbH$_4$, PbF$_4$, and Pb$_2$ compounds and Pb$^{+2}$, Pb$^{+4}$ ions are listed together with the chemical shifts of the x-ray $K$-line in Pb with respect to the Pb$^{+2}$ ion. The partial wave charges and XES chemical shifts values are obtained with using the two-step restoration codes developed in Refs.\ \cite{Skripnikov:13b,Skripnikov:13c} after pseudopotential calculations carried out with the {\sc dirac} code \cite{dirac13}. Due to the stability of the computation procedure issues, calculations of the electronic structure of all listed compounds were performed with the core Pb shells frozen up to $5d$. The states belonging to these shells were taken from the Pb$^{+2}$ computation. The effective occupation numbers of the valence shells, $N^W_{n_vlj}$, were determined from the following equations:
\begin{equation}
N^W_{n_vlj}\int\limits_{r<R_c} (\left|\phi^L_{n_vlj}(r)\right|^2 + \left|\phi^S_{n_vlj}(r)\right|^2)r^2\,dr = q^W_{lj},
\label{confs_iter}
\end{equation}
where $\phi^L_{n_vlj}(r)$ and $\phi^S_{n_vlj}(r)$ are the radial functions of the large and small components of the corresponding valence state obtained from the relativistic average configuration computation of the isolated ion with frozen core states; the occupation numbers of the valence Pb shells in this computation are equal to the $N^W_{n_vlj}$ values. Thus \Eq{confs_iter} is a nonlinear self-consistent equation and must be solved iteratively. Corresponding calculations were carried out with the {\sc hfd} code \cite{hfd}.
It is possible to carry out all-electron calculations of an isolated ion with the given occupation numbers, $N^W_{n_vlj}$, and to take partially account of relaxation of the core states frozen earlier. The chemical shifts obtained from these calculations are also listed in \Table{table:tabl3}. The differences of these values and the chemical shifts obtained in the frozen core calculations are 10\%--20\% by the order of magnitude for the neutral Pb atom and molecules PbH$_4$, PbO, Pb$_2$, which are weaker bound compared to PbF$_4$. For the Pb$^{+4}$ ions and PbF$_4$ the relaxed chemical shifts are about two to three times lower than the corresponding values obtained in the frozen core calculations.
The experimental datum for the chemical shift of the XES $2p_{3/2}\to 1s_{1/2}$ line of lead in crystalline PbO with respect to metallic Pb is listed in Ref.~\cite{Egorov:92} and equals to $54\pm8$ meV. One can estimate the corresponding value from the performed calculations as difference between chemical shifts of PbO and Pb$_2$. For the results obtained in the Dirac--Fock approximation this estimate gives 9 meV without taking into the account the Pb core relaxation in the compounds and it is 15 meV for the relaxed core case.
It is possible to take account of the effects of electronic correlation within DFT. From the results of calculations with using the PBE0 functional \cite{pbe0} listed in \Table{table:tabl3} we conclude that the electronic correlation effects are important for the Pb$_2$ molecule, since the XES chemical shifts differ by the factor of two for Dirac--Fock and DFT calculations. The chemical shift of the lead XES $2p_{3/2}\to 1s_{1/2}$ line obtained at the DFT level is 54 meV when the core relaxation is not considered and is 32 meV when the relaxation is taken into account. The obtained values are much closer to the experimental datum than those obtained at the Dirac--Fock level.
\section{Conclusions}
Utilizing the property of proportionality of valence and low-lying virtual spinors within an atomic core region with radius $r{<}R_c$, the notions of ``W reduced density matrices'', ${\bm\rho}^W$, and ``partial wave charges'', $q^{W}_{lj}$, for valence electrons in the core region are introduced. Such properties as hyperfine structure constants
\comtav{for TableI+: <<<}
(\Table{table:tabl1}),
T- and P-violation effects (\cite{Titov:06amin}), XES chemical shifts (\Table{table:tabl3}),
\comtav{>>> End}
which are mainly sensitive to a variation of electronic densities in an atomic core region (or even on a nucleus), with a good accuracy depend on ${\bm\rho}^W$ only. For specific AiC properties or effective Hamiltonian parameters the more particular blocks of ${\bm\rho}^W$ like diagonal terms $q^{W}_{lj}$ (for XES chemical shifts), off-diagonal $s{-}p$, $p{-}d$, etc.
submatrices (for evaluation of $E_{\rm eff}$\ and some other P- and T,P-odd Hamiltonian parameters with the $sp$-, $pd$-, etc.\ hybridized unpaired electrons \cite{Kozlov:87, Dmitriev:92}) are sufficient to know. Thus, the $W$ reduced DM allows one to characterize the effective state of an atom in a chemical substance by an appropriate manner.
The features of the AiC approach are summarized as follows.\\
{\bf (i)} The $W$ reduced DM and, correspondingly, the AiC characteristics calculated on an atom are independent of the origin of one-electron basis set
used (whether it is a one-center, MO~LCAO, analytic, or numerical one) in the limit of its completeness in contrast to the cases of Mulliken and L\"owdin population analyses; the one-center AiC basis functions are independent on the valence structure of a chemical substance studied in contrast
to those in the NAO approach.\\
{\bf (ii)} it describes well (generally, in the range of accuracy of 10-30\%) the multitude of AiC characteristics.\\
{\bf (iii)} Due to the property of $W$ proportionality, the core radius $R_c$ is not a very critical parameter to be fixed as exactly as possible for an AiC characteristic but the accuracy of calculation is higher for those characteristics for which it can be chosen smaller since the $W$ reduced DM match better the original DM for smaller $R_c$%
\footnote{The $W$ proportionality can be interpreted as some kind of ``asymptotic {\it unfreedom}''
of valence and low-lying virtual spinors in an atomic core of a chemical substance
in contrast to a widely known property of ``asymptotic freedom'' in theory of strong interactions.};
the situation is similar to that in theory of ``transferable'' shape-consistent relativistic pseudopotentials with the $R_c$ treated as a matching radius (see Refs.~\cite{Goedecker:92a, Titov:99} and references).
\\
{\bf (iv)} The method allows one to give a correct quantum mechanical interpretation even for such difficult cases as XES chemical shifts and provide an unambiguous analysis of atomic (one-center) spinor contributions to AiC characteristics in complicated electronic structures (see discussion in \cite{Kudashov:13} concerning RaO);\\
{\bf (v)} it can provide a theoretical background for semiempirical models to evaluate (estimate) the AiC characteristics which are not known or cannot be calculated using available experimental data for corresponding properties \cite{Kozlov:95}.\\
Note as well that the approach can be easily implemented in codes when the relativistic pseudopotential theory \cite{Titov:99, Petrov:04b, Mosyagin:06amin} and one-center recovery (restoration) procedures \cite{Titov:05a, Titov:06amin} can be utilized for calculation of the AiC properties and parameters in heavy-element compounds.
However, within the AiC approach, the concept of a charge on an atom in a chemical substance becomes meaningless \cite{Cramer:04}.
\section*{Acknowledgement}
This work is supported by the Russian Science Foundation grant No.~14-31-00022.
We are grateful to A.Zaitsevskii for fruitful discussions.
|
\section{Introduction}
The considerable progress in thin film etching allows today the realization of almost any waveguide structure
requiring submicron resolution. For example, the fabrication of single mode optical waveguides, couplers, interferometers or ring resonators is well controlled.
Whatever their function, these structures are made of a superposition of several thin layers and they have to guide light. Therefore it is essential to accurately characterize the optical and geometrical properties of each layer of the stack. Several techniques exist to perform such a characterization in the case of a single
layer film, such as those based on the measurements of the reflection and transmission coefficients of the
sample~\cite{abeles1963,manifacier1976,martinezanton2000}, the ellipsometry~\cite{azzam1977}, and the m-lines
spectroscopy~\cite{tien1970,ulrich1970,ulrich1973}.
The case of multilayer waveguides has been much less investigated, although
the assumption that a layer in a stack shows the same properties as it has when measured individually, can become wrong. The characterization of the whole structure then becomes necessary. Some work was done on the case of two layers~\cite{tien1973,stutius1977,hewak1987,matyas1991,aarnio1995,kubica2002} (sometimes denoted as a "four layer film" since the stack is deposited on a substrate and the air is considered as a top layer), however, no literature is available for three layer structures.
These structures are of special interest since they consist of the minimum number of layers required in order to obtain a single mode waveguide which is thick enough to enable efficient coupling of the light. In this paper we will focus on three layer waveguides and show how the refractive indices and thicknesses of the individual layers can be retrieved simultaneously from m-lines spectroscopy measurements.
In classical m-lines devices, the sample is pressed against a face of a prism. The prism and the film are mounted on a rotating stage in order to allow the variation of the light incidence angle. A thin air gap between the sample and the prism face is maintained whose thickness should be approximately the fourth of the light probe wavelength.
The incoming light is refracted inside the prism and reaches the interface between the prism and the sample.
Since the refractive index of the prism is higher than that of the sample, the light is totally reflected at this interface, for a given range of incidence angles, and then emerges from the prism to be detected. For some angles, called
"synchronous angles", however, part of the light is coupled into the waveguide, hence substracted from the detected light. Therefore, a typical m-lines spectrum consists in several absorption-like peaks, centered around the synchronous angles. From the positions of the synchronous angles, it is possible to deduce the propagation constants of the guided modes of the sample under test and derive the optical and geometrical properties of the structure (refractive index and thickness) by solving the modal dispersion equations.
The first part of this paper is devoted to the determination of the explicit form of modal dispersion equations of the three
layer waveguide. We will then describe the numerical method used in order to solve these equations and present numerical tests that prove its accuracy. The validity of the method will be finally demonstrated experimentally by the analysis of m-lines spectra produced by three layer ZnO/PZT/ZnO waveguides.
\section{Three layer dispersion equations}
The studied structure is a stack of five transparent homogeneous layers shown on Fig.~1.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0.125cm,clip]{tschneider_fig1.eps}
}
\caption{Three layer waveguide.}
\label{figure1}
\end{figure}
The substrate (layer~\#0) and the superstrate (layer~\#5) are considered as semi-infinite since their thicknesses are several orders of magnitude greater than those of layers $1$, $2$ and $3$. The light is confined only in these three central layers and presents an evanescent decay in the substrate and the superstrate. We call this structure a "three layer waveguide", whereas in the nomenclature of other authors~\cite{hewak1987} it would be called a "5 layer waveguide". The central layer (layer~\#2) is the core of the structure and the layers~\#1 and \#3 are the claddings (true waveguide~\cite{zhang2002}). As a consequence, we assume for the refractive indices of the different layers that:
\begin{equation}
\label{eq1}
n_2>n_1,n_3>n_0,n_4
\end{equation}
In the following, we will restrict ourselves to the case of TE modes, the extension to the
case of TM modes is straightforward. The only component of the electrical field of the TE modes is along the (Oy) axis,
so the electric field in the layer j can be written as:
$E_{yj}(x,z)=A_j\exp[i(\gamma_j x+\beta_m z)]+B_j\exp[i(-\gamma_j x+\beta_m z)]$ and the tangential component
of the magnetic field is $H_{zj}(x,z)=i(\omega\mu_0)^{-1}\partial E_{yj}/\partial x$. In these expressions, $\omega$ is
the angular frequency and $\beta_m$ is the propagation
constant of the $m^{\rm th}$ guided mode. It is usually written as
$\beta_m =kN_m$, where $k$ is the wavevector modulus in vacuum and $N_m$
the effective index of the m$^{th}$ mode. Using the condensed notation $a_j=k\ |n_j^2-N^2|^{1/2}$, the x component of the wavevector, $\gamma_j$, which gives the nature of the wave in the layer $j$, becomes $\gamma_{j}=(\omega\mu_0)^{-1}a_j$ for a travelling wave and $\gamma_{j}=i (\omega\mu_0)^{-1}a_j$ for an evanescent wave. A transfer matrix $M_{j}$, which binds the electromagnetic fields at the backplane of the layer to the fields at its frontplane, can be associated to each layer~\cite{chilwell1984}:
\begin{equation}
\label{eq2}
M_j=
\left(\begin{array}{cc}
\cos (\omega\mu_0\gamma_j d_j) & \displaystyle \frac{i}{\gamma_j}\sin (\omega\mu_0\gamma_j d_j)\\
i\ \gamma_j \sin (\omega\mu_0\gamma_j d_j )& \cos (\omega\mu_0\gamma_j d_j)\\
\end{array}\right)
\end{equation}
The boundary conditions imply that the tangential components of the magnetic
and electrical fields must be continuous at the interface of the layers.
These conditions together with the condition for obtaining guiding lead to:
\begin{equation}
\label{eq3}
\left(\begin{array}{c} 1\\-\gamma_4 \end{array}\right) E_{4y}
=M_3M_2M_1 \left(\begin{array}{c} 1\\\gamma_0 \end{array}\right) E_{0y}
=M\left(\begin{array}{c} 1\\\gamma_0 \end{array}\right) E_{0y}
\end{equation}
which has solutions only for:
\begin{equation}
\label{eq4}
\gamma_4 m_{11}+\gamma_4\gamma_0 m_{12} + m_{21} + \gamma_0 m_{22}=0
\end{equation}
where $m_{ij}$ are the components of the matrix $M$.
With the definition~(\ref{eq2}) of the transfert matrix,
this equation can be written as~\cite{thesethomas}:
\begin{equation}
\label{eq5}
\begin{array}{ll}
{a_2}\,{d_{2}} & + \;{\rm arctan} \left[ \! {\displaystyle
\frac {i\gamma_{0}\,\gamma_1 + \gamma_1^2\,{\tan}({\omega\mu_0\gamma_1}\,{d_{1}})
}{\gamma_1\,\gamma_2 -i {\displaystyle \gamma_{0}\,\gamma_2\,
\tan({\omega\mu_0\gamma_1}\,{d_{1}})} }} \! \right] \\[8mm]
&
+ \;{\rm arctan} \left[ \! \displaystyle
\frac { i\gamma_3\,\gamma_{4} + \gamma_3^2\,\tan({\omega\mu_0\gamma_3}\,{d_{3}})
}{\gamma_2\,\gamma_3 -i \gamma_2\,\gamma_{4}\,
\tan(\omega\mu_0{\gamma_3}\,{d_{3}})} \! \right] - \;m
\,\pi =0
\end{array}
\end{equation}
which is the general modal dispersion equation for the true three layer waveguides.
As stated above, the $\gamma_j$ terms are either real or imaginary depending on the nature
of the waves in the $j^{\rm th}$ layer. As a consequence, an analysis in the complex plane is required in order to solve directly the equation~\ref{eq5}.
It is better
to take advantage of physical arguments to split the problem in several simpler ones.
Under the condition of Eq.~\ref{eq1}, three kinds of guided waves can exist:
\begin{itemize}
\item
The lowest order modes can only propagate in the layer \#2. Then $\gamma_2$ only is real
and Eq.~\ref{eq5} becomes:
\begin{equation}
\label{eq6}
\begin{array}{l}
{a_2}\,{d_{2}} - \;\mathrm{arctan} \left[ \! {\displaystyle
\frac {a_{0}\,a_1 + a_1^2\,\mathrm{tanh}({a_{1}}\,{d_{1}})
}{a_1\,a_2 + {\displaystyle a_{0}\,a_2\,
\mathrm{tanh}({a_{1}}\,{d_{1}})} }} \! \right] \\[4mm]
\quad
- \;\mathrm{arctan} \left[ \! \displaystyle
\frac { a_3\,a_{4} + a_3^2\,\mathrm{tanh}({a_{3}}\,{d_{3}})
}{a_2\,a_3 + a_2\,a_{4}\,
\mathrm{tanh}({a_{3}}\,{d_{3}})} \! \right] - \;m
\,\pi =0
\end{array}
\end{equation}
\item
As the order of the mode increases, the wavevector approaches the normal
of the interfaces. For a given mode $m_2$, the incidence angle on the interface between
the central layer and one cladding layer becomes smaller than the
limit angle for total internal reflection and the light propagates
inside these two layers. If $n_3>n_1$ then the light is guided in the layers~\#2 and \#3
and Eq.~\ref{eq5} becomes:
\begin{equation}
\label{eq7}
\begin{array}{l}
{a_2}\,{d_{2}}
- \;\mathrm{arctan} \left[ \! \displaystyle
\frac { a_0\,a_{1} + a_1^2\,\mathrm{tanh}({a_{1}}\,{d_{1}}) }
{a_1\,a_2 + a_0\,a_{2}\,\mathrm{tanh}({a_{1}}\,{d_{1}})} \! \right] \\[4mm]
\quad
- \;\mathrm{arctan} \left[ \displaystyle \! \frac {\displaystyle a_3\,a_4 -
a_3^2\,\mathrm{tan}({a_{3}}\,{d_{3}})}
{\displaystyle a_2\,a_3 + a_2\,a_{4}\,
\mathrm{tan}({a_{3}}\,{d_{3}}) } \! \right] - \;m
\,\pi =0
\end{array}
\end{equation}
Otherwise, the light propagates in the layers~\#1 and \#2. The corresponding modal dispersion equation
is obtained by interverting the functions $\tan$ and $\tanh$ in Eq.~\ref{eq7}.
\item
Finally, highest order modes can propagate in the three layers. The dispersion equation for these modes is:
\begin{equation}
\label{eq8}
\begin{array}{l}
{a_2}\,{d_{2}}
- \;\mathrm{arctan} \left[ \! \displaystyle
\frac { a_0\,a_{1} - a_1^2\,\mathrm{tan}({a_{1}}\,{d_{1}}) }
{a_1\,a_2 + a_0\,a_{2}\,\mathrm{tan}({a_{1}}\,{d_{1}})} \! \right] \\[4mm]
\quad
- \;\mathrm{arctan} \left[ \displaystyle \! \frac {\displaystyle a_3\,a_4 -
a_3^2\,\mathrm{tan}({a_{3}}\,{d_{3}})}
{\displaystyle a_2\,a_3 + a_2\,a_{4}\,
\mathrm{tan}({a_{3}}\,{d_{3}}) } \! \right] - \;m
\,\pi =0
\end{array}
\end{equation}
\end{itemize}
\section{Data analysis : algorithm and numerical tests}
At a first glance, the problem contains 6 unknown parameters which are the refractive indices and thicknesses of the layers \#1, 2 and 3.
\begin{figure}[htbp]
\centerline{\includegraphics[width=8cm,trim=0.5cm 0.4cm 0.5cm 0.85cm,clip]{tschneider_fig2.eps}}
\caption{Algorithm of the resolution of the system of 6 equations in 6 unknowns and $M$ measured modes.}
\label{figure2}
\end{figure}
However, depending on the guiding regime, one has to associate correctly the dispersion equation to each measured synchronous angle. Consequently, further parameters have to be determined:
\begin{itemize}
\item
$m_1$, the order of the first measured mode. It is often equal to 0, but the lower order modes are sometimes difficult to excite,
and hence may be not visible in the m-lines spectrum. There is no evidence on the value of $m_1$ in practice.
\item
$m_2$, the order of the first mode guided by two layers. Numerical simulations
with noisy data showed that $m_2$ is the value of $m$ such that $|N_m-N_{m-1}|>|N_{m+1}-N_m|$.
This criterion results in a correct value of $m_2$ or with a mismatch of +1~\cite{optmat2007}. In the following we will
call $m_2^{\rm th}$ the value given by this criterion.
\item
$m_3$, the order of the first mode guided by three layers.
\item
Finally, the modal dispersion equation in the case of two guiding layers
is not the same according to whether $n_1$ is smaller or greater than $n_3$.
Hence it is also necessary to make an initial hypothesis on the relative values of $n_1$ and $n_3$
and to verify this assumption during the resolution.
\end{itemize}
The analysis of a m-lines spectrum thus requires a somewhat complicated algorithm. For clarity
reasons, we splitted the presentation of this algorithm into two parts (Fig.~2 and~3).
\begin{figure}[htbp]
\centerline{\includegraphics[width=8cm,trim=0.125cm 0.125cm 0.125cm 0.125cm,clip]{tschneider_fig3.eps}}
\caption{Algorithm of the m-lines spectra analysis.}
\label{figure3}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0.05cm 0.05cm 0.05cm 0.05cm,clip]{tschneider_fig4.eps}
}
\caption{Distribution of errors on the parameter $n_2$.}
\label{figure4}
\end{figure}
The details of the "resolution" box of the full algorithm (shown on Fig.~3) are presented in Fig.~2,
aiming to solve
a system of 6 equations in 6 unknowns for $m_1$, $m_2$, $m_3$ given and for an assumption on
the relative values of $n_1$ and $n_3$. Obviously, this problem can be
solved only if the m-lines spectrum contains more than 6 modes. Let us call $M$ the number
of measured synchronous angles. There is $C_6^M$ combinations of 6 modes. For a given combination $j$,
we firstly use a Newton-Raphson method to get a solution $\{n_1,n_2,n_3,d_1,d_2,d_3\}_j$.
Then, we start from this solution and use a bissection algorithm to calculate the
set of $M$ corresponding synchronous angles : $\{\phi_{ij}^{\rm cal}\}$, $i\in\{1,2,\dots,M\}$.
The validity of the solution is evaluated by the standard deviation of the differences between
the values of the calculated synchronous angles and the measured ones:
\begin{equation}
\label{eq9}
S_j^2=\sum_{i=1}^M\displaystyle \frac{\left(\phi_{i}^{\rm meas}-\phi_{ij}^{\rm cal}\right)^2}{M}
\end{equation}
If $S_j$ is smaller than $0.1^\circ$ (upper limit of experimental uncertainty), the solution is considered as correct.
This procedure is used for the $C_6^M$ combinations but it does not find roots in all cases.
Let us call $\Gamma$ the number of combinations for which the procedure succeeds.
The solution we finally retain is the mean
value of these $\Gamma$ solutions and we associate a fitness $\sigma$ to this solution
defined by:
\begin{equation}
\label{eq10}
\sigma^2=\sum_{j=1}^\Gamma \frac{S_{j}^2}{\Gamma^2}
\end{equation}
$\Gamma^2$ was used rather than $\Gamma$, in order to give more weight to
the configurations which lead to a high number of acceptable solutions.
The parameters $m_1$, $m_2$, and $m_3$ are obtained with an iterative procedure schematized
on Fig.~3. We start with the $M$ measured synchronous angles
and under the hypothesis $H: n_1>n_3$. The value of $m_2$ is set to
$m_2^{\rm th}$, $m_1$ to 0 and $m_3$ to $m_2$ and the procedure of resolution described above is used in order to get a solution and its associated
fitness $\sigma_{H,m_1,m_2,m_3}$. This procedure is repeated with firstly an incrementation
of $m_3$ from $m_2$ to $M$, secondly an incrementation of $m_1$ from 0 to 5 (which in general is sufficient) and thirdly by setting $m_2$ to $m_2^{\rm th}-1$.
For the case where no solution was obtained at the end of these iterations, the whole procedure is repeated
with the opposite hypothesis ($n_1<n_3$).
The refractive indices and thicknesses finally selected are those which are associated to the lowest value
of $\sigma$ thus defining the values $m_1$, $m_2$, $m_3$ and $H$.
The validity of the method was verified by numerical simulations. Starting from a
theoretical waveguide defined by the parameters $\{n_1^{\rm th},n_2^{\rm th},n_3^{\rm th},d_1^{\rm th},d_2^{\rm th},d_3^{\rm th}\}$,
we calculated the corresponding set of synchronous angles $\{\phi_i^{\rm th}\},\ 0\leq i\leq 8$. In order to test
the sensitivity to the noise, this set was used for building one hundred
sets of noisy data $\{\phi_i^{\rm noisy}\}$ by adding to each theoretical synchronous angle,
$\phi_i^{\rm th}$, a random value, $\varepsilon_i$, standing
in the interval $[-\delta\phi,\delta\phi]$ with a uniform probability density.
The values of $\delta\phi$ used in the simulations were chosen in order to correspond to
the experimentally observed noise.
The hundred noisy data sets were analysed, giving rise to a number $\Lambda$ of sets of solutions $\{n_1^{\rm sol},n_2^{\rm sol},n_3^{\rm sol},d_1^{\rm sol},d_2^{\rm sol},d_3^{\rm sol}\}$. $\Lambda$ depends on the magnitude of the noise, it is very close to 100 for $\delta\phi=0.01^\circ$ and about 80 for $\delta\phi=0.1^\circ$. Each value $p^{\rm sol}_{j,\ 1\leq j\leq\Lambda}$ of the parameter
$p$ ($p=n_1,n_2,n_3,d_1,d_2$ or $d_3$) differed from the theoretical value $p_j^{\rm th}$. The distribution of errors
$\delta p_j=p_j^{\rm sol}-p_j^{\rm th}$ followed a normal law
for all parameters $p$ and all noises~\cite{thesethomas}, as showed on Fig.~4 for the example of $n_2$.
Finally, we defined the uncertainty $\Delta p$ on the parameter $p$ as three times the standard deviation
of the distribution, in order to be in the confidence interval of 99\%:
$\Delta p=3\left[\sum_{j=1}^{\Lambda}\Lambda^{-1}(\delta p_j-\langle \delta p\rangle)^2\right]^{1/2}$.
A total of eight thousands waveguides were simulated, whose characteristics are summarized
in Tab.~\ref{tab1}.
The thickness of the upper layer, $d_3$, was constant and smaller
than the penetration depth of the light, in order to make possible the evanescent
coupling between the prism and the central layer.
\begin{table}
\caption{Range of parameters explored with the statistical study.}
\begin{center}
\begin{tabular}{cccc}
\hline\hline
& Start & End & Step\\
\hline
$n_1$, $n_3$& 1.6 & $n_2$-0.1 & 0.1 \\
$n_2$ & 1.7 & 2.4 & 0.1 \\
$d_1$, $d_2$ & 1.0 $\mu$m & 2.0 $\mu$m & 0.1~$\mu$m \\
$d_3$ & \multicolumn{3}{c}{100 $nm$ } \\
\hline\hline
\end{tabular}
\end{center}
\label{tab1}
\end{table}
In order to get global estimators, we studied the distribution of
uncertainties $\Delta p$ of each parameter for $p$ for all the simulated guides and for each noise. It was always possible to fit
the distribution with a log-normal law (see Fig.~5):
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0.1cm 0.0cm 0.1cm 0.1cm,clip]{tschneider_fig5.eps}
}
\caption{Distribution of the uncertainties $\Delta n_2$ over all the simulated guides, fitted by a lognormal law.}
\label{figure5}
\end{figure}
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.0cm 0.0cm 0.0cm,clip]{tschneider_fig6.eps}
}
\caption{Evolution of the error on the refractive indices as a function of the noise on the synchronous angles.}
\label{figure8}
\end{figure}
\begin{equation}
\label{eq11}
f(\Delta p)=\frac{1}{ w \Delta p\sqrt{2\pi}}\ {\rm e}^{\displaystyle -\frac{(\ln \Delta p-\mu)^2}{2w^2}}
\end{equation}
where $w$ and $\mu$ are the free parameters of the model.
Hence, the mean error was $\langle \Delta p\rangle= \exp\left(\mu+w^2/2\right)$ and the standard deviation
of the distribution was: $\sigma_{\Delta p}^2=\left(\exp({w^2})-1\right)\exp(2\mu+w^2)$.
It followed that 85\% of the errors were in the range $\langle \Delta p\rangle \pm \sigma_{\Delta p}$.
We then used $\langle\Delta p \rangle$ as an estimate of the error on the parameter p with an uncertainty $\sigma_{\Delta p}$.
The results of the simulations are summarized on the graphics of Fig.~6 and Fig.~7.
For the central layer, the error on the refractive index is lower than $1 \times 10^{-3}$
and the error on the thickness remains lower than ten nanometers for a noise on the synchronous angles below 0.1$^\circ$.
For the other layers the errors on the refractive indices remain acceptable.
They are smaller than $1 \times 10^{-3}$ for the layer~\#1 and of the order of $1 \times 10^{-2}$
for the upper layer. Nethertheless, the errors on the thicknesses can reach large values, especially for $d_3$.
However, it should be emphasized that the number of synchronous angles
used for the statistical study was limited to 9 in order to limit
the computation time. A closer inspection of the guides leading to unaccurate results showed
that the accuracy can be considerably enhanced when increasing the number of modes
taken into account. As an example, Fig~8 and Fig~9
show the evolution of the errors on the different parameters as a function of the number
of modes for a noise of 0.01$^\circ$ and a guide defined by
$n_1=1.9$, $n_2=2.3$, $n_3=2.2$, $d_1=1.1\ \mu$m, $d_2=1.9\ \mu$m, and $d_3=100$~nm.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0cm,clip]{tschneider_fig7.eps}
}
\caption{Evolution of the error on the thicknesses as a function of the noise on the synchronous angles.}
\label{figure9}
\end{figure}
The errors on $d_1$ and $n_1$ can be reduced by two orders of magnitude
while the error on $d_3$ and $n_3$ is divided by 2 when the number of modes used in the calculation is increased from 9 to 15.
This means that one should always use the maximum number of modes in order to reach the best accuracy. However, the computation time also increases.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0.125cm,clip]{tschneider_fig8.eps}
}
\caption{Evolution of the error on the refractive indices as a function of the number of synchronous angles used in the resolution.}
\label{figure6}
\end{figure}
\section{Experimental results}
In order to experimentally validate the method,
several multilayer structures were elaborated. They were deposited on glass substrates and were composed of
lead zirconate titanate (PZT) for the core layer and Al doped zinc oxide (ZnO) for the claddings.
A SEM picture of one of these samples is shown on Fig.~10.
The zinc oxide layers were grown by RF magnetron sputtering at room temperature from
a 3'' in diameter ZnO/Al$_2$O$_3$ (98/2 wt.\%) ceramic target. Prior to the deposition,
a pressure lower than $5 \times 10^{-7}$ mbar was reached and pure argon was used as a sputter
gas at a partial pressure of $2 \times 10^{-3}$ mbar during the deposition process. An on-axis
growth rate of approximately 100 nm/min was achieved at a RF power of 200~W at a
target-substrate distance of 7.5~cm.
The films were annealed at 650$^\circ$C during 3~min and cooled down to room temperature during 3 hours.
Four substrates were placed side by side under the ZnO target, resulting in a non homogeneous thickness of the films.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0.125cm,clip]{tschneider_fig9.eps}
}
\caption{Evolution of the error on the thicknesses as a function of the number of synchronous angles used in the resolution.}
\label{figure7}
\end{figure}
In the following we will call "$A_\ell$" and "$A_r$" the samples directly on the left and on the right side of
the center of deposition, "$B_\ell$" the sample on the left of $A_\ell$ and "$B_r$" the sample on the right of
$A_r$. For symmetry reasons, we expect ZnO layers $A_\ell$ and $A_r$ as well as
$B_\ell$ and $B_r$, to be identical.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.25cm 0cm 0.125cm,clip]{tschneider_fig10.eps}
}
\caption{SEM photograph of a three layer waveguide.}
\label{figmeb}
\end{figure}
The PZT 36/64 layers were elaborated by Chemical Solution Deposition technique~\cite{jecs2005}.
A modified Sol-gel process was used for the elaboration of the PZT precursor solution,
which consisted of lead acetate dissolved in acetic acid, zirconium and titanium n-propoxide;
ethylene glycol was added in order to prevent from crack formation during the annealing process.
The final solution was spin-coated on the ZnO layer at 1000~rpm and a Rapid Thermal Annealing procedure
at 650$^\circ$C resulted in the formation of a polycrystalline perovskite without remaining pyrochlore phases.
A layer of PZT was spin-coated individually
on samples $A_\ell$, $A_r$, $B_\ell$ and $B_r$. After cristallisation, we expect a repeatability of $8 \times 10^{-3}$ on the refractive index
and of 20~nm on the thickness~\cite{jecs2005}.
The upper ZnO cladding layer was only deposited on samples
$A_r$ and $B_r$ with a thickness smaller than the penetration depth of the light. These two samples were not characterized, neither with m-lines
nor with other technique, until the third layer was deposited, in order to avoid pollution or any other
deterioration of the structure.
The two layers waveguides $A_\ell$ and $B_\ell$ were used as control samples.
Another three layer sample, called "C" in the following, was elaborated in the same way as the samples $A$ and $B$, but it
was characterized by m-lines after each deposition step.
We first consider the samples $A$ and $B$. An example of m-lines spectrum, obtained with sample
$B_r$, is shown on Fig.~11. The transition from the
single guiding layer to the two guiding layer regime appears clearly in the spectrum.
Indeed, the broad peaks correspond to the waves guided in the PZT layer only, while the narrow
peaks are associated to the waves also guided in the ZnO. This broadening is not
a peculiarity of the three layer structure, it can be also observed for PZT single layers and
may be due to light diffusion resulting in a loss along the direction of propagation~\cite{zhang2002}.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0.0cm 0.0cm 0.0cm 0.0cm,clip]{tschneider_figure11.eps}
}
\caption{m-lines spectrum obtained with sample $B_r$ measured at 29 $mm$ from the center of deposition.}
\label{figure10}
\end{figure}
The transition from the two guiding layer regime to the three guiding layer regime can not be infered from the spectrum and has to be determined by numerical computation.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0.125cm,clip]{tschneider_fig12.eps}
}
\caption{Evolution of the thicknesses of the different layers
of the three layer samples $A_r$, $B_r$ ($d_1:\square$,
$d_2:$ {\Large
$\circ $}, $d_3:\triangle$ ) and the two layer samples
$A_\ell$, $B_\ell$ ($d_1: \blacksquare$, $d_2:$ {\Large
$\bullet $}).}
\label{figure12}
\end{figure}
The spectrum of Fig.~11 was analyzed with the numerical method described in section 2, resulting in the
following characteristics:
\begin{itemize}
\item
layer \#1 : $n_1= 1.9701\pm 6 \times 10^{-4}$, $d_1= 1.588\pm 3 \times 10^{-3} \mu m$
\item
layer \#2 : $n_2= 2 2702\pm 6 \times 10^{-4}$, $d_2= 0.894\pm 6 \times 10^{-3} \mu m$
\item
layer \#3 : $n_3= 2.037\pm 9 \times 10^{-3}$, $d_3= 0.122\pm 3 \times 10^{-3} \mu m$
\end{itemize}
Due to the complexity of the system to solve, the uncertainties were estimated
with Monte Carlo simulations, in a similar way to what is described in section 3.
Starting from the solution $\{n_1,n_2,n_3,d_1,d_2,d_3\}$, we calculated
the associated synchronous angles $\{\phi_i\}$ and built one hundred sets
of noisy angles by adding a random value in the range of the experimental uncertainties.
We then computed the solutions corresponding to the differents noisy sets and considered
their distributions. The uncertainty on each parameter is defined
as three times the standard deviation of the distribution of the values of this parameter.
We performed m-lines measurements every 4~mm from the center of the ZnO deposition.
Fig.~12 and Fig.~13 show respectively the evolution
of the thickness and the refractive indices of the different layers of the samples $A$ and $B$. If we except the point located at 41~mm from the deposition center, where few modes were available due to the low thickness of the lower ZnO layer,
the error bars do not clearly appear since the uncertainties are small.
They are of the order of $5 \times 10^{-4}$ for $n_1$ and $n_2$, and
$1 \times 10^{-2}$ for $n_3$.
\begin{figure}[htbp]
\centerline{
\includegraphics[width=8cm,trim=0cm 0.125cm 0cm 0.125cm,clip]{tschneider_fig13.eps}
}
\caption{Evolution of the refractive indices of the different layers
of the three layer samples $A_r$,
$B_r$ ($n_1:\square$, $n_2:$ {\Large $\circ $}, $n_3:\triangle$ ) and the two layer samples
$A_\ell$, $B_\ell$ ($n_1: \blacksquare$, $n_2:$ {\Large $\bullet $} ).}
\label{figure11}
\end{figure}
The uncertainties on the thicknesses are of the order of 10~nm,
which corresponds to a relative accuracy of 0.5\% for $d_1$ and $d_2$,
and 10\% for $d_3$.
As expected, the thickness of the lower ZnO layers decreases
with the distance from the ZnO deposition center (Fig.~12). The thickness of the PZT films is rather constant except at the
border of the samples where a slight increase can be observed which is typical for the spin-coating process.
The refractive index of the PZT layer varies from 2.2227$\pm 5 \times 10^{-4}$ to 2.3610$\pm 5 \times 10^{-4}$. It is slightly higher
than the refractive index of the PZT deposited on glass under the same conditions~\cite{jecs2005}. This may be due
to a structural change of the PZT thin film induced by the ZnO buffer layer which acts as a diffusion barrier thus hindering diffusion of the lead
from the PZT into the glass substrate.
The index of the ZnO lower layer is rather constant, it oscillates between 1.9667$\pm 6 \times 10^{-4}$ and 1.991$\pm 5 \times 10^{-3}$. On the contrary, the index of the upper ZnO layer exhibits strong variations from 2.015$\pm 8 \times 10^{-3}$ to 2.123$\pm 8 \times 10^{-3}$. This may be explained by the existence of two different cristalline structures arising
when the thicknesses of the ZnO film is below 500~nm~\cite{lin2004}.
The differences between the thicknesses and refractive indices obtained for the
three layer waveguides ($A_r$, $B_r$) and the two layer waveguides ($A_\ell$, $B_\ell$)
essentially stay in the range of the repeatability of the elaboration procedure and the measurement techniques. So the method
of analysis of three layer guides m-lines spectra gives results in accordance
to those obtained with two layer guides. This is confirmed by the measurements
realized from sample $C$, where the refractive index and the thickness of the layers where measured before and after deposition of the upper cladding layer. The results are summarized in Tab.~\ref{tab2} showing that the differences between the values
obtained before and after the third deposition remain always smaller than the uncertainty and
the repeatability of the elaboration procedure and the measurement techniques. This good agreement proves the validity of our method.
\begin{table}
\caption{Comparison of the measured characteristics of a waveguide (sample $C$), before and after the deposition of the upper cladding layer.}
\begin{center}
\begin{tabular}{ccccccc}\hline\hline
&
$n_1$ &
$d_1$ ($\mu m$) &
$n_2$ &
$d_2$ ($\mu m$) &
$n_3$ &
$d_3$ ($\mu m$) \\
\hline
Two layers & 1.978 & 1.11 & 2.360 & 1.02 & & \\
Three layers & 1.979 & 1.09 & 2.368 & 0.96 & 1.987 & 0.13 \\ \hline\hline
\end{tabular}
\end{center}
\label{tab2}
\end{table}
\section{Conclusion}
In this paper, the general modal dispersion equation for a three layer
planar waveguide is shown, from which the modal dispersion equations that hold
for different guiding regimes are derived. A method to solve these equations is proposed,
where the input data are the synchronous angles measured by m-lines spectroscopy. Monte Carlo simulations show that the values of the refractives indices and the thicknesses of the central layer given by this method are as accurate
as those obtained for a single layer. The accuracy is not so good for the upper
layer, however, the error remains smaller than 2.10$^{-2}$ for the index and below 30~nm
for the thickness when the uncertainty on the measured angles remains smaller than 0.1$^\circ$.
The method was applied to the characterization of real three layer planar waveguide structures made of
one PZT layer embedded between two ZnO cladding layers deposited on glass substrate. The agreement between the results obtained
with three layer structures and those obtained with two layer structures ensures the validity of our method.
Moreover, the three layer analysis revealed changes in material properties, such as increasing
of the refractive index of PZT deposited on ZnO in comparison to deposition on glass
and increasing
of the refractive index of ZnO deposited on PZT in comparison to deposition on glass.
The proposed method allows to simultaneously characterize the optical and geometrical properties of each layer
of three layer waveguides. Consequently, it is a very interesting instrument
in order to verify whether the three layer structures are matching the parameters
defined during the design process of waveguide.
|
\section{INTRODUCTION}
\label{sec:introd}
Post-AGB (PAGB) stars, which represent the late stages of evolution of
low and intermediate mass stars (0.8 -- 8 M$_{\odot}$), are very important
diagnostic tools to understand the evolutionary processes including AGB
nucleosynthesis which is a major source of production of
elements such as C, N, F, Al, Na and s-process elements. These products are dredged up
to the surface due to mixing episodes thereby modifying the observed abundances
of these stars. The synthesized elements are ejected into the Interstellar Medium (ISM)
through strong mass-loss at the end of AGB evolution hence these stars
significantly affect the chemical evolution of galaxies.
In the super wind phase (at the end of AGB evolution), most of the outer stellar
envelope is lost; consequently PAGB stars
are surrounded by circumstellar envelopes and hence their spectral energy
distribution generally exhibits two peaks. However, exceptional
PAGBs without circumstellar envelopes have been detected.
The PAGB evolutionary model proposed by Iben and Renzini(1983)
had been improvised by Groenewegen \& de Jong (1993) and Boothroyd \& Sackmann (1999)
through better approximations for quantities like mass-loss and mixing length.
More comprehensive models including detailed calculations of thermal pulses
have been presented in Karakas, Lattanzio \& Pole (2002), Herwig (2004) etc.
A very comprehensive review of AGB evolutionary models for a full range
of masses and metallicities can be found in Herwig (2005).
The canonical definition of PAGB stars as objects showing
strong enhancement of carbon (C/O $>$ 1) and s-process elements is met by relatively
smaller fraction of stars. Evolutionary models show that objects with a
small range in masses (1.8 to 4.0 M$_{\odot}$) show carbon and s-process element
enrichment
while those with lower mass exhibit only the effect of CNO processing.
More massive stars (M$>$ 4.0 M$_{\odot}$) undergo Hot Bottom Burning (HBB) where
carbon produced by He burning is converted to N, hence C/O remains $<$ 1 but
N is enhanced and transient production of Li is predicted.
The observational and theoretical developments made in the last two decades
on PAGBs has been summarized in the excellent review by Van Winckel (2003)
where the diverse chemical compositions observed for the central stars are
compared with the predictions from evolutionary models.
More recent developments based on extended sample of
PAGBs, made using the extended wavelength coverage
have augmented our understanding of PAGB classes
including the morphology and chemistry of circumstellar shells as
described in reviews by Garc\'ia-Lario (2006) and Giridhar (2011).
While the surface compositions of C-rich and O-rich PAGBs can
be explained with AGB nucleosynthesis, mixing events and mass-loss;
it has been noticed that chemical peculiarities not attributable to
AGB nucleosynthesis are observed in
a large fraction of PAGB stars including RV Tauri stars which
show abundance patterns reflecting a systematic removal of condensable elements.
Well known examples of such objects are HR4049, HD~52961, BD +39$^{o}$4926, HD
44179; a good summary could be found in Van Winckel (2003).
The depletion of elements show strong correlation with condensation temperature
T$_{C}$ (defined as the temperature at which half of a
particular element in a gaseous environment condenses into grains) ;
those elements with higher T$_{C}$ (easily condensable) are
heavily depleted while those with lower T$_{C}$ are largely unaffected.
This removal of condensable elements onto the dust grain
is commonly referred as dust-gas winnowing or DG-effect.
The full mechanism of DG effect is not yet fully understood, but
it is generally believed to operate in a dusty circumbinary disks surrounding
these objects as initially suggested by Waters, Trams \& Waelkens (1992).
From the study of SED of a large sample of PAGB stars De Ruyter et al.
(2006) reported the presence of dust at or near sublimation temperature
very close to the star for most depleted objects irrespective
of their temperatures and ascribed it to the presence of gravitationally
bound disk. With the help of high spatial resolution interferometry in
mid IR, dusty disks around depleted objects Red Rectangle, HR 4049 and
HD 52961 have been resolved (Deroo et al. 2007).
Sumangala Rao, Giridhar \& Lambert (2012) have compiled the
stellar parameters and abundances for PAGB stars and
find that s-process enhanced group contains
a very small number of binaries. However the depleted group
contains larger fraction of binaries in accordance with the hypothesis
of dusty disks surrounding binary post-AGB stars.
Sumangala Rao \& Giridhar (2014) have presented a similar compilation for RV Tauri
stars.
These authors also revisit the boundary conditions for discernible DG effect
discussed in earlier papers by Giridhar et al. (2000, 2005).
The values of minimum temperature and intrinsic metallicities are very similar for
PAGB and RV Tauri stars.
The goal of this paper is to enlarge PAGB sample by a study
of four candidate PAGB stars chosen from their location in the
two color IRAS diagram. We have carried out
detailed atmospheric abundance analysis using high resolution spectra
for IRAS 13110\,-\,6629,
IRAS 17579\,-\,3121, IRAS 18321\,-\,1401, IRAS 18489\,-\,0629,
previously identified as
post-AGB/PN objects (Preite-Mart\'inez 1988; Garc\'ia-Lario et al. 1997; Pereira \&
Miranda 2007). We have also studied their spectral energy distribution to
understand the properties of circumstellar material surrounding them.
The structure of the paper is as follows: $\S$ 2 presents the sample selection.
We describe our observations and data analysis procedures in $\S$ 3. The estimation
of atmospheric parameters, abundances and the errors associated with these
parameters are presented, in $\S$ 4.
The derived abundances for individual objects are reported. In $\S$ 5 these
results are discussed. In $\S$ 6 and $\S$ 7 we present summary and conclusions.
\section {The sample}
\label{sec:sample}
Fig.~\ref{fig:figure1} shows the IRAS color-color diagram of van der Veen \&
Habing (1988)
with different zones signifying the emergence and evolution of the circumstellar shell
produced during AGB evolution. The objects in zone IV are undergoing super wind phase
or slightly beyond; objects with only cold dust are found in V while VI b contains
objects with warm and cold shells. Zone VIII and dashed area defined
by Garc\'ia-Lario et al. (1997) is a region where, most
post-AGB and planetary nebulae (PN) are found.
We have selected sample stars whose IR colors places them in zone VIII (IRAS 18321\,-\,1401)
and the remaining three in the region where according to Garc\'ia-Lario et al. (1997)
PAGB and PPN are found.
IRAS 13110\,-\,6629 (GLMP 342) was included in a sample of candidate PN objects
by Silva et al. (1993) attempting to detect OH (1612 MHz, 1665 MHz and 1667 MHz)
masers without success.
Its position in the IRAS color-color diagram $[12]-[25]$ =$+$4.08 and
$[25]-[60]$ = $+$0.30, indicates that this object falls into the region defined
by Garc\'ia-Lario et al. (1997) for PAGB stars and planetary nebulae. The presence
of circumstellar material was detected by Clarke et al. (2005). This object
has also been included in the Toru\'n catalog of Galactic PAGB (Szczerba et al.
2007; Szczerba et al. 2011) as a likely PAGB object.
The star IRAS~17579\,-\,3121 (GLMP 686) was classified as candidate PN by
Preite-Mart\'inez (1988) and Ratag et al. (1990) due to its infrared fluxes.
Silva et al. (1993) reported the detection of a maser emission of OH at 1612 MHz
(with double peaks) where they found velocities for the blue peak of 21 km s$^{-1}$
and red peak 0 km s$^{-1}$ respectively. Su\'arez et al. (2006) classified this star
as a PAGB star from its optical counterpart based on a low resolution spectrum.
A more recent interferometric study using VLA by G\'omez et al. (2008) did not confirm
the association of IRAS 17579\,-\,3121 with the position of the radio continuum emission
and conclude that this IRAS source is not likely to be a PN.
Given its position in the IRAS two color diagram IRAS 18321\,-\,1401 (PM 1-243)
was
classified as a possible PN by Preite-Mart\'inez et al. (1988). This classification
was questioned in a subsequent study by van de Steene \& Pottasch (1995) who employed
radio interferometric measurements to detect radio continuum emission at 6~cm
which is characteristics of PN; but no emission was detected for this object.
More recently, Pereira \& Miranda (2007) classified the star as a PAGB star by
comparing its spectrum with the spectrum of the known PAGB star GLMP 982.
These authors suggest a spectral type F for this object.
The star IRAS 18489\,-\,0629 (PM 1-261) has been classified as a probable
Planetary
Nebula by Preite-Mart\'inez (1988) following its location in the IRAS two color diagram.
The large value of [12]-[25] = $+$3.05 color \textbf{might indicate} the presence of
relatively warm dust resulting from a strong mass loss episode (Clarke et al. 2005).
Pereira \& Miranda (2007) from their study of 16 PAGB candidates
using low resolution flux calibrated spectra have classified IRAS 18489\,-\,0629 as PAGB
since the spectrum closely resembles that of GLMP 1058. By comparison
with the spectrum of the supergiant star HD 9973, taken from the library of Jacoby
et al. (1984), Pereira \& Miranda (2007) suggested that the observed
continuum energy distribution of IRAS 18489\,-\,0629 was similar to that of
a reddened F5Iab star.
Basic information of the sample stars such as celestial and galactic coordinates,
brightness, IRAS fluxes and its quality (in parenthesis) is given in
Table~\ref{tab:table1}.
The apparent magnitudes $V$ were obtained only for IRAS 18489\,-\,0629,
IRAS 17579\,-\,3121 and IRAS 13110\,-\,6629 taken from different sources (Hog et al. 2000,
SIMBAD\footnote{http://simbad.u-strasbg.fr/simbad/sim-fid} database
and Perryman 1997).
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figg1.eps}
\caption{The location of our sample stars in IRAS two color diagram is shown. The
dotted-line
represents the region limit by Garc\'ia-Lario et al. (1997). The regions represented
with
Roman numerals have been defined by van der Veen \& Habing (1988).}
\label{fig:figure1}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{reg5130.eps}
\caption{Representative spectra of the sample stars
IRAS 13110\,-\,6629 and IRAS 17579\,-\,3121.
IRAS 18321\,-\,1401, IRAS 18489\,-\,0629,
The location of lines
of certain important elements have been indicated by dashed lines.}
\label{fig:figure2}
\end{center}
\end{figure}
\begin{table*}
\centering
\caption{Basic data of the program stars. The flux quality is given in parenthesis
coded from low to high as from 1 to 3 respectively.}
\label{tab:table1}
\begin{tabular}{lccccccccc}
\hline
\hline
\multicolumn{1}{c}{No. IRAS}&
\multicolumn{1}{c}{$\alpha_{2000}$}&
\multicolumn{1}{c}{$\delta_{2000}$}&
\multicolumn{1}{c}{{\it V}}&
\multicolumn{1}{c}{{\it l}}&
\multicolumn{1}{c}{{\it b}}&
\multicolumn{1}{l}{F$_{12}$ $\mu$m}&
\multicolumn{1}{c}{F$_{25}$ $\mu$m}&
\multicolumn{1}{c}{F$_{60}$ $\mu$m}&
\multicolumn{1}{c}{F$_{100}$ $\mu$m}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(h m s)}&
\multicolumn{1}{c}{(h m s)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{($^{0}$)}&
\multicolumn{1}{c}{($^{0}$)}&
\multicolumn{1}{c}{(Jy)}&
\multicolumn{1}{c}{(Jy)}&
\multicolumn{1}{c}{(Jy)}&
\multicolumn{1}{c}{(Jy)}\\
\hline
13110\,-\,6629&13 14 27.4&$-$66 45 35.0&10.74&305.20&$-$3.98&0.55(3)&23.51(3)&30.91(3)&12.48(3) \\
17579\,-\,3121&18 01 13.3&$-$31 21 56.5&11.40&359.61&$-$4.13&2.01(3)&62.39(3)&58.21(3)&14.71(1) \\
18321\,-\,1401&18 34 57.2&$-$13 58 49.0&$\ldots$&18.62&$-$2.75&0.36(1)&1.10(3)&1.52(3)&115.90(1) \\
18489\,-\,0629&18 51 39.1&$-$06 26 07.1&11.70&27.22&$-$2.97&0.27(1)&4.48(3)&7.49(3)&78.49(1) \\
\hline
\end{tabular}
\end{table*}
\subsection {Photometric Variability}
\label {sec:variability}
Light variability in PAGB stars is of common occurrence. RV Tauris exhibit
large amplitude light curve of distinct shape and periodicity but other PAGBs are
known to exhibit small amplitude light variation as described by Kiss et
al. (2007).
PAGB stars of intermediate temperature are found near or above the upper
luminosity limit of the instability strip hence
semi-regular pulsations are commonly seen. Their periodicities are not always firmly
established and show irregular amplitude modulations and even cessation in their
variability. Classical examples of variable PAGB stars are 89 Her (HR 6685),
HD~161796 and LN Hya (HR 4912) with semi-periods of 63, 42-60 and 44-80 days (Arellano
Ferro 1985; Arellano Ferro 1981).
\begin{figure*}
\begin{center}
\includegraphics[width=15.0cm,height=10.0cm]{2BOXES.eps}
\caption{Light variations in the stars IRAS13110$-$6629 and IRAS18489$-$0629 as seen
from the V-band data from
the All Sky Automated Survey. Each independent data set is
depicted with a different Color.
Year numbers are placed at the beginning of the calendar year for a reference. See
$\S$ \ref{sec:variability}. for more details}
\label{fig:lightcurves}
\end{center}
\end{figure*}
Defining periodicities even using the radial velocity
variations for PAGB is complicated since their radial
velocities are often the combination of pulsations and differential
motions in the extended atmosphere (Kloshkova \& Panchuck 2012). On the other hand,
dense and continuous photometry of PAGB stars is rare. In recent times the All Sky
Automated Survey (ASAS) (Pojmanski 2002) provided valuable photometric data for a
large number of objects.
We have examined the V-band data obtained between 2001 and 2009 for all four stars in
our sample.
IRAS13110\,-\,6629 and IRAS18489\,-\,0629 display clear signs of variability of
semi-regular nature, with amplitude modulations within 0.3 mag
(Fig. \ref{fig:lightcurves}). A frequency analysis of these data reveal
characteristic times or semi-period between 52 and 58 d for IRAS13110\,-\,6629 and
between
51.8 and 60.0 d for IRAS18489\,-\,0629. No secondary frequencies were identified. IRAS
17579\,-\,3121 and IRAS 18321\,-\,1401 do not show coherent variations but irregular
fluctuations that range between 0.2 and 1.0 mag.
\section{Observations and Reductions}
\label{sec:data}
The high-resolution spectra of the stars analyzed in this work
were obtained with FEROS, (Fiber-fed Extended Range Optical
Spectrograph) at the 2.2\,m ESO telescope at La Silla (Chile) on the
nights of August 27, 2007 (IRAS 18489\,-\,0629), May 14, 2009 (IRAS
17579\,-\,3121), July 29, 2009 (IRAS 18321\,-\,1401) and July 31, 2009
(IRAS 13110\,-\,6629). For each star two exposure times of 3\,600 secs
were obtained. FEROS has a CCD with an array of 2048 $\times$ 4096
with each pixel size of 15$\mu$, manufactured by EEV. The wavelength
range obtained with FEROS covers the spectral region between 3900 and
9200\AA\, and is distributed in 39 orders with a resolving power of
$\lambda$/$\Delta$$\lambda$\,=\,48\,000 or 2.2 pixels per
$\Delta$$\lambda$, which gives, at 5000\AA\, approximately 0.05\AA/pixel
(Kaufer et al. 1999).
The spectra were reduced following the standard procedure
including bias subtraction, flat-fielding, order
extraction and wavelength calibration with the MIDAS pipeline reduction package.
The cosmic ray hits causing narrow spikes were manually removed and
all spectra were normalized to continuum. Two exposures were combined
to get higher S/N ratios.
The S/N ratio of the spectra was generally in the 120-150 range, however for
the spectrum of IRAS 13110\,-\,6629 it was about 70 because one of the exposures was
cut short due to cirrus. Since
most of the sample stars are of intermediate temperatures, line blending was
not severe. It was generally possible to measure line strengths with an
accuracy of 5 to 8 percent.
Fig.~\ref{fig:figure2} shows the representative spectra of the sample stars
arranged in descending temperature sequence.
\section{Atmospheric parameters}
\label{sec:param}
The spectral line strengths strongly depend on atmospheric parameters
such as effective temperature $T_{\rm eff}$, gravity log~$g$ and microturbulence
velocity $\xi_{t}$ in addition to the abundances; hence a good determination of
atmospheric parameters are mandatory to derive accurate abundances.
The estimation of the effective temperature and gravity can be made
from photometric and spectroscopic data.
\subsection{Photometric data}
\label{sec:photome}
Table~\ref{tab:table2} lists the magnitudes JHK for
IRAS 13110\,-\,6629 and IRAS 17579\,-\,3121
IRAS 18489\,-\,0629, taken from the 2MASS photometry (Cutri et al. 2003).
The color excess $E(B-V)$ was estimated from the reddening map of Schlegel et al.
(1998).
For $E(B-V) > 0.15$, we adopted the correction suggested by Bonifacio, Caffau \&
Molaro (2000). Then the intrinsic colors $J_{0}$,
$(J-H)_{0}$ and $(H-K)_{0}$, were derived using the ratios
$\frac{E(J-H)}{E(B-V)} = 0.322$ and $\frac{E(H-K)}{E(B-V)} = 0.183$
of Fiorucci \& Munari (2003).
\begin{table*}
\tiny{
\caption{2MASS photometry, intrinsic color and photometric atmospheric parameters
for the sample stars.}
\label{tab:table2}
\centering
\begin{tabular}{lccccccccccc}
\hline
\hline
\multicolumn{1}{l}{No. IRAS}&
\multicolumn{1}{c}{$J\pm \sigma_J$}&
\multicolumn{1}{c}{$H \pm \sigma_H$}&
\multicolumn{1}{c}{$K \pm \sigma_{K}$}&
\multicolumn{1}{c}{$E(B-V)_{IS}$}&
\multicolumn{1}{c}{$(J-H)_{0}$}&
\multicolumn{1}{c}{$(H-K)_{0}$}&
\multicolumn{1}{c}{$J_{0}$}&
\multicolumn{1}{c}{$M_{J}$}&
\multicolumn{1}{c}{$T_{\rm eff,Tok}$}&
\multicolumn{1}{c}{$T_{\rm eff,Mol}$}&
\multicolumn{1}{c}{log~$g$,$_{Mol}$}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{l}{(mag)}&
\multicolumn{1}{c}{(mag)}&
\multicolumn{1}{c}{(K)}&
\multicolumn{1}{c}{(K)}&
\multicolumn{1}{c}{}\\
\hline
13110\,-\,6629 & 8.06 & 7.60 & 7.40 & 0.50 & 0.30 & 0.15 & 7.62 &
$-$3.63$\pm$0.28 & 5864 & 6094$\pm$220 & 0.75$\pm$0.27 \\
17579\,-\,3121 & 8.396$\pm$0.017 & 7.890$\pm$0.046 & 7.585$\pm$0.021 & 0.60 & 0.312 & 0.195 & 7.863 &
$-$3.48$\pm$0.28 & 5722 & 6023$\pm$220 & 0.78$\pm$0.27 \\
18321\,-\,1401 & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ &
$\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ \\
18489\,-\,0629 & 9.006$\pm$0.027 & 8.63$\pm$0.05 & 8.398$\pm$0.029 & 0.51 & 0.213 & 0.139 & 8.560 &
$-$3.90$\pm$0.28 & 6733 & 6576$\pm$220 & 0.79$\pm$0.27 \\
\hline
\end{tabular}
}
\end{table*}
The effective temperatures for program stars were first
derived using the $(J-H)_o$
calibrations for supergiant stars by Tokunaga (2000).
These temperatures are given in Table~\ref{tab:table2} as $T_{\rm eff,Tok}$.
An independent estimates of temperatures, gravity and absolute magnitudes
were made from intrinsic colors
$(J-H)_o$ and $(H-K_{s})_o$ calibrated by Molina (2012) for post-AGB and RV Tauri
stars. They are presented as
$T_{\rm eff_{Mol}}$, log~$g_{Mol}$ and $M_{J}$ in Table~\ref{tab:table2}.
As shown in the figure 6 of Molina (2012), the calibration made using F-G
supergiants gives systematically lower temperatures.
These photometric estimates of atmospheric parameters serve as
starting values which are further refined by a detailed spectral analysis.
\subsection{Spectroscopic data}
\label{sec:spectro}
We have used new grids of ATLAS9 model atmospheres available at the database of
Kurucz (see Castelli and Kurucz 2003).
We have used 2010 version of MOOG developed by C. Sneden (1973) in both line
and spectrum synthesis mode. The assumptions made are local thermodynamic
equilibrium (LTE), plane parallel atmosphere and hydrostatic equilibrium.
The atmospheric parameters $T_{\rm eff}$, log~$g$ and the microturbulence velocity
$\xi_{t}$ are obtained from the lines of well represented elements such as Fe, Ti and
Cr.
The temperature was determined by requiring that there was no dependence of derived
abundances on lower excitation using Fe~I lines. Finally gravity was estimated from
the ionization equilibrium of Fe~I/Fe~II, i.e. $\log$ n(Fe II)$=$$\log$ n(Fe I).
Attempts have been made to get a unique solution for temperature
and gravity by studying the ionization equilibrium of several elements;
Mg~I \& Mg~II, Si~I \& Si~II, Ti~I \& Ti~II and Cr~I \& Cr~II.
The Balmer line profiles were affected by emission filling in hence
we have explored the loci of Paschen lines in the temperature-gravity plane as
it will be illustrated later in this paper (see Fig. 5).
The microturbulence $\xi_{t}$ was estimated by requiring that the derived abundances
are independent of the line strengths for a given specie. Generally, lines of Fe~II
are preferred as it is known that they are not seriously affected
by departure from LTE (Schiller \& Przybilla 2008). For objects with
very few measurable Fe~II lines, Fe~I lines were employed.
In addition we have also employed the method described in
Sahin \& Lambert (2009) for estimating the microturbulence velocity
based upon the measurement of
standard deviation as a function of microturbulence velocity $\xi_{t}$
for Fe I, Fe II, Ti~II, Cr~I and Cr~II lines in the 0 to 9 km~s$^{-1}$ range.
Fig. \ref{fig:turbulence} illustrates the method using
the star IRAS 18489\,-\,0629 as an example. The minimum standard deviation is reached for
$\xi_{t} \sim 3.9$\,km~s$^{-1}$.
The measurement error in the equivalent widths (about 5 to 8 \%),
corresponds to uncertainty in
microturbulence velocity of $\pm$ 0.2 km s$^{-1}$ and in
temperature and gravity of
$\pm$ 200\, K and $\pm$ 0.20 respectively. The adopted values of $T_{\rm eff}$,
log~$g$ and $\xi_{t}$ for our sample are listed in Table \ref{tab:adopted}.
\begin{table*}
\centering
\caption{Adopted atmospheric parameters for the program stars}
\label{tab:adopted}
\begin{tabular}{lcccrrr}
\hline
\hline
\multicolumn{1}{c}{No. IRAS}&
\multicolumn{1}{c}{$T_{\rm eff}$}&
\multicolumn{1}{c}{log~$g$}&
\multicolumn{1}{c}{$\xi_{t}$}&
\multicolumn{1}{c}{$V_{r}(hel)$}&
\multicolumn{1}{c}{$V_{r}(LSR)$}&
\multicolumn{1}{c}{Date}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(K)}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(km s$^{-1}$)}&
\multicolumn{1}{c}{(km s$^{-1}$)}&
\multicolumn{1}{c}{(km s$^{-1}$)}&
\multicolumn{1}{c}{}\\
\hline
13110$-$6629 & 6500 & 0.50 & 5.0 & $+$11.5$\pm$0.5 & $+$8.1 & July 31, 2009\\
17579$-$3121 & 6500 & 0.00 & 4.7 & $+$24.7$\pm$0.6 & $+$34.3 & May 14, 2009 \\
18321$-$1401 & 5500 & 0.20 & 4.0 & $+$32.7$\pm$0.6 & $+$46.9 & July 29, 2009 \\
18489$-$0629 & 6500 & 0.50 & 3.9 & $+$159.6$\pm$0.5 & $+$175.3 & August 27, 2007\\
\hline
\end{tabular}
\end{table*}
\begin{figure}
\begin{center}
\includegraphics[width=8.cm,height=8cm]{VTURB2.eps}
\caption{Determination of microturbulence velocity for the star IRAS
18489\,-\,0629 using the plot of standard deviation as a function of
microturbulence velocity $\xi_{t}$ for several species in the 0 to 9 kms$^{-1}$ range.}
\label{fig:turbulence}
\end{center}
\end{figure}
\subsection {Uncertainties in the elemental abundances}
\label{sec:uncert}
The derived abundances are affected by errors in equivalent width
measurements (which in turn depends upon the spectral resolution,
S/N ratio, continuum placement and also on the spectral type of the star), errors
in atomic data such as oscillator strengths (log {\rm gf}) and errors
in deriving the atmospheric parameters. The equivalent
width measurement errors are random errors
while errors caused by uncertainty in log {\rm gf} and those in
estimating atmospheric parameters are systematic errors.
The effect of random errors in equivalent widths for a single star is well
represented by $\sigma_{1}$, the standard
deviation from the mean abundance based on the whole set
of lines. These errors can be reduced by measuring a large number of lines.
The error in log gf values vary from element to element. For example,
experimental values
for Fe~I and Fe~II of high accuracy, better than 5 \%, are
available for a large fraction of lines. For other Fe-peak
elements, errors in their ${\rm gf}$ values may range between 10 to 25\%.
For neutron-capture elements the accuracy of recent estimates are in 10\% to 25\%
range. An extensive list of ${\rm gf}$ values for all important elements can be
found in Sumangala Rao, Giridhar \& Lambert (2012).
The effect of systematic errors in abundances (in dex)
caused by errors in estimating the atmospheric
parameters are illustrated in Table~\ref{tab:table4} for two representative stars
covering the temperature and gravity range of the sample stars.
We present change in abundances caused by varying atmospheric parameters
by 200\,K, 0.20 and 0.2 km s$^{-1}$ with respect to the chosen model for each star.
Following the standard procedure, the total systematic error $\sigma_{2}$ is
estimated by taking the square root of the sum of the squares
of each individual errors associated with uncertainties in temperature,
gravity and microturbulence.
The total error $\sigma_{tot}$ for each element is quadratic sum
of $\sigma_{1}$ and $\sigma_{2}$. The error bars in abundance plot corresponds
to this total error.
\begin{table*}
\caption{The sensitivity of abundances to the changes in the model atmospheric
parameters for two values of temperature of our sample stars.}
\label{tab:table4}
\begin{center}
\begin{tabular}{lrrrrcrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\noalign{\smallskip}
& \multicolumn{4}{c}{IRAS 18489\,-\,0629} & & \multicolumn{4}{c}{IRAS 18321\,-\,1401}\\
\cline{2-5}\cline{7-10}
& \multicolumn{4}{c}{(6500\,K)} & & \multicolumn{4}{c}{(5500\,K)}\\
\multicolumn{1}{l}{Species}&
\multicolumn{1}{c}{$\Delta$ T$_{\rm eff}$}&
\multicolumn{1}{c}{$\Delta$ $\log$~$g$}&
\multicolumn{1}{r}{$\Delta$ $\xi_{t}$}&
\multicolumn{1}{c}{$\sigma_{tot}$}&
\multicolumn{1}{l}{}&
\multicolumn{1}{c}{$\Delta$ T$_{\rm eff}$}&
\multicolumn{1}{c}{$\Delta$ $\log$~$g$}&
\multicolumn{1}{r}{$\Delta$ $\xi_{t}$}&
\multicolumn{1}{c}{$\sigma_{tot}$}\\
\multicolumn{1}{l}{}&
\multicolumn{1}{c}{$+$200~K}&
\multicolumn{1}{c}{$+$0.20}&
\multicolumn{1}{c}{$+$0.20}&
\multicolumn{1}{c}{}&
\multicolumn{1}{l}{}&
\multicolumn{1}{c}{$+$200~K}&
\multicolumn{1}{c}{$+$0.20}&
\multicolumn{1}{c}{$+$0.20}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
C I & $-0.07$ & $-0.01$ & $+0.02$ & $0.07$ & & $-0.06$ & $-0.12$ & $-0.08$ & $0.15$\\
N I & $-0.09$ & $-0.06$ & $+0.02$ & $0.11$ & & $+0.07$ & $-0.04$ & $ 0.00$ & $0.08$\\
O I & $+0.02$ & $-0.04$ & $+0.02$ & $0.05$ & & $-0.02$ & $-0.03$ & $+0.01$ & $0.03$\\
Na I & $-0.17$ & $+0.05$ & $+0.01$ & $0.18$ & & $-0.05$ & $+0.01$ & $+0.01$ & $0.05$\\
Mg I & $-0.22$ & $+0.06$ & $+0.01$ & $0.23$ & & $-0.06$ & $ 0.00$ & $+0.02$ & $0.06$\\
Al I & & & & & & $-0.05$ & $+0.01$ & $+0.01$ & $0.05$\\
Si I & $-0.17$ & $+0.05$ & $+0.01$ & $0.18$ & & $-0.06$ & $+0.01$ & $+0.01$ & $0.06$\\
Si II& $+0.08$ & $-0.05$ & $+0.02$ & $0.10$ & & & & & \\
S I & $-0.13$ & $+0.01$ & $+0.02$ & $0.13$ & & $+0.01$ & $-0.02$ & $+0.01$ & $0.02$\\
Ca I & $-0.24$ & $+0.05$ & $+0.02$ & $0.25$ & & $-0.08$ & $+0.01$ & $+0.03$ & $0.08$\\
Sc II& $-0.15$ & $-0.05$ & $+0.02$ & $0.16$ & & $-0.04$ & $-0.04$ & $+0.03$ & $0.06$\\
Ti I & $-0.27$ & $+0.04$ & $+0.01$ & $0.27$ & & $-0.12$ & $+0.01$ & $+0.01$ & $0.12$\\
Ti II& $-0.14$ & $-0.06$ & $+0.04$ & $0.16$ & & $-0.04$ & $-0.03$ & $+0.03$ & $0.05$\\
V I & & & & & & $-0.12$ & $+0.01$ & $+0.01$ & $0.12$\\
V II & $-0.13$ & $-0.06$ & $+0.04$ & $0.15$ & & $-0.03$ & $+0.03$ & $+0.03$ & $0.05$\\
Cr I & $-0.25$ & $+0.04$ & $+0.02$ & $0.25$ & & $-0.09$ & $+0.02$ & $+0.03$ & $0.09$\\
Cr II& $-0.07$ & $-0.07$ & $+0.05$ & $0.11$ & & $-0.01$ & $-0.04$ & $+0.02$ & $0.04$\\
Mn I& $-0.25$ & $+0.04$ & $+0.06$ & $0.26$ & & $-0.09$ & $-0.04$ & $+0.01$ & $0.09$\\
Fe I& $-0.21$ & $+0.05$ & $+0.03$ & $0.22$ & & $-0.09$ & $+0.01$ & $+0.01$ & $0.09$\\
Fe II& $-0.06$ & $-0.05$ & $+0.05$ & $0.09$ & & $-0.05$ & $-0.04$ & $+0.04$ & $0.07$\\
Co I& $-0.30$ & $+0.04$ & $ 0.00$ & $0.30$ & & $-0.10$ & $+0.01$ & $+0.01$ & $0.10$\\
Ni I & $-0.22$ & $+0.04$ & $+0.01$ & $0.22$ & & $-0.08$ & $+0.03$ & $+0.01$ & $0.08$\\
Ni II& $-0.07$ & $-0.06$ & $+0.01$ & $0.09$ & & & & & \\
Cu I & $-0.27$ & $+0.04$ & $+0.01$ & $0.27$ & & $-0.12$ & $+0.01$ & $+0.02$ & $0.12$\\
Zn I & $-0.22$ & $+0.05$ & $+0.02$ & $0.23$ & & $-0.08$ & $+0.01$ & $+0.05$ & $0.09$\\
Y II& $-0.16$ & $-0.05$ & $+0.02$ & $0.17$ & & $-0.04$ & $-0.03$ & $+0.02$ & $0.05$\\
Zr II& $-0.16$ & $-0.06$ & $+0.01$ & $0.17$ & & $-0.04$ & $-0.03$ & $+0.01$ & $0.05$\\
Ba II& $-0.30$ & $+0.01$ & $+0.03$ & $0.30$ & & $-0.05$ & $-0.02$ & $+0.01$ & $0.05$\\
La II& $-0.24$ & $-0.03$ & $-0.01$ & $0.24$ & & $-0.05$ & $-0.03$ & $ 0.00$ & $0.05$\\
Ce II& $-0.22$ & $-0.04$ & $ 0.00$ & $0.22$ & & $-0.06$ & $-0.03$ & $+0.02$ & $0.07$\\
Nd II& & & & & & $-0.07$ & $-0.03$ & $+0.01$ & $0.07$\\
Eu II& $-0.21$ & $-0.03$ & $+0.01$ & $0.21$ & & $-0.06$ & $-0.03$ & $ 0.00$ & $0.06$\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\end{tabular}
\end{center}
\end{table*}
\section { Results for Individual stars}
\label{sec:abund}
The observed elemental abundances of the stars are
governed by the composition of the natal ISM, the nucleosynthesis
and mixing process in the course of its evolution and in some
cases external effects such as mass exchange. For certain PAGBs
and RV Tauri stars the DG effect introduces abundance anomalies.
The atmospheric abundances of all represented elements are given in Table~
\ref{tab:abund} for the sample stars.
\begin{table*}
\scriptsize{
\caption{Elemental abundances for the sample stars.}
\label{tab:abund}
\centering
\begin{tabular}{lccrcccrccrccr}
\hline
& &\multicolumn{3}{c}{IRAS 13110\,-\,6629} &\multicolumn{3}{c}{IRAS
17579\,-\,3121}
&\multicolumn{3}{c}{IRAS 18321\,-\,1401}&\multicolumn{3}{c}{IRAS
18489\,-\,0629}\\
\cline{3-14}\\
\multicolumn{1}{l}{Species}&
\multicolumn{1}{c}{$\log \epsilon_{\odot}$}&
\multicolumn{1}{c}{[X/H]}&
\multicolumn{1}{c}{N}&
\multicolumn{1}{r}{[X/Fe]}&
\multicolumn{1}{c}{[X/H]}&
\multicolumn{1}{c}{N}&
\multicolumn{1}{r}{[X/Fe]}&
\multicolumn{1}{c}{[X/H]}&
\multicolumn{1}{c}{N}&
\multicolumn{1}{r}{[X/Fe]}&
\multicolumn{1}{c}{[X/H]}&
\multicolumn{1}{c}{N}&
\multicolumn{1}{r}{[X/Fe]}\\
\hline
C I &8.39&$-0.01\pm$0.13&18&$+0.27$ &$+0.07\pm$0.13&15&$+0.39$ &$-0.11\pm$0.10&
7&$+0.32$ &$-0.09\pm$0.10&18&$+0.27$\\
N I &7.78&$+0.47\pm$0.14& 5&$+0.75$ &$+0.60\pm$0.07& 5&$+0.92$ &$-0.19\pm$0.16&
3&$+0.24$ &$+0.01\pm$0.07& 5&$+0.37$\\
O I &8.66&$+0.04\pm$0.06& 3&$+0.32$ &$+0.29\pm$0.13& 3&$+0.61$ &$+0.10\pm$0.13&
3&$+0.53$ &$-0.18\pm$0.21& 4&$+0.18$\\
Na I &6.17&$-0.02\pm$0.05& 3&$+0.26$ &$+0.27\pm$0.08& 3&$+0.59$ &$-0.21\pm$0.09&
4&$+0.22$ &$-0.06\pm$0.10& 4&$+0.30$\\
Mg I &7.53&$+0.09\pm$0.15& 4&$+0.37$ &$+0.03\pm$0.09& 1&$+0.35$ &$+0.07\pm$0.10&
3&$+0.50$ &$-0.13\pm$0.14& 3&$+0.23$\\
Mg II&7.53&$+0.09\pm$0.20& 2&$+0.37$ &$+0.26\pm$0.09& 1&$+0.58$ & & &
& & & \\
Al I &6.37& & & & & & &$-0.66\pm$0.06&
2&$-0.23$ & & & \\
Si I &7.51& & & &$+0.02\pm$0.09& 8&$+0.34$
&$-0.06\pm$0.10&18&$+0.37$ &$-0.03\pm$0.09&11&$+0.33$\\
Si II&7.51&$-0.06\pm$0.11&11&$+0.22$ &$+0.09\pm$0.04& 1&$+0.41$ & & &
&$-0.15\pm$0.04& 1&$+0.21$\\
S I &7.14&$-0.04\pm$0.05&syn&$+0.24$&$+0.14\pm$0.10& 2&$+0.46$ &$-0.14\pm$0.13&
3&$+0.29$ &$+0.22\pm$0.12& 8&$+0.58$\\
Ca I &6.31&$-0.35\pm$0.13&16&$-0.07$ &$-0.52\pm$0.09& 9&$-0.20$ &$-0.62\pm$0.11&
8&$-0.19$ &$-0.52\pm$0.08&16&$-0.16$\\
Ca II&6.31&$-0.28\pm$0.04& 2&$ 0.00$ &$-0.44\pm$0.04& 1&$-0.12$ & & &
& & & \\
Sc II&3.05&$-0.69\pm$0.17& 6&$-0.41$ &$-0.87\pm$0.12& 5&$-0.55$ &$-0.70\pm$0.15&
7&$-0.27$ &$-1.06\pm$0.12& 5&$-0.70$\\
Ti I &4.90&$-0.60\pm$0.06& 1&$-0.32$ & & &
&$-0.67\pm$0.10&11&$-0.24$ &$-0.70\pm$0.16& 3&$-0.34$\\
Ti II&4.90&$-0.52\pm$0.10& 7&$-0.24$ &$-0.72\pm$0.10& 8&$-0.40$ &$-0.66\pm$0.13&
8&$-0.23$ &$-0.59\pm$0.13&12&$-0.23$\\
V I &4.00& & & & & & &$-0.35\pm$0.21&
2&$+0.08$ & & & \\
V II &4.00&$-0.69\pm$0.12& 2&$-0.41$ &$-0.87\pm$0.02& 1&$-0.55$ &$-0.38\pm$0.16&
2&$+0.05$ &$-0.40\pm$0.08& 3&$-0.04$\\
Cr I &5.64&$-0.37\pm$0.13& 8&$-0.09$ &$-0.54\pm$0.10& 4&$-0.22$
&$-0.54\pm$0.15&12&$-0.11$ &$-0.48\pm$0.17& 9&$-0.12$\\
Cr II&5.64&$-0.35\pm$0.16&12&$-0.07$ &$-0.38\pm$0.11& 6&$-0.06$
&$-0.58\pm$0.13&11&$-0.15$ &$-0.41\pm$0.18&12&$-0.05$\\
Mn I&5.39&$-0.40\pm$0.17& 5&$-0.12$ &$-0.40\pm$0.14& 4&$-0.08$ &$-0.74\pm$0.12&
7&$-0.31$ &$-0.55\pm$0.12& 5&$-0.19$\\
Fe I&7.45&$-0.28\pm$0.13&84& &$-0.34\pm$0.14&73& &$-0.45\pm$0.12&132&
&$-0.33\pm$0.12&55& \\
Fe II&7.45&$-0.28\pm$0.11&19& &$-0.29\pm$0.09&17& &$-0.41\pm$0.10&17&
&$-0.39\pm$0.12& 9& \\
Co I &4.92&$-0.45\pm$0.08& 1&$-0.17$ & & & &$-0.31\pm$0.14&
4&$+0.12$ &$-0.33\pm$0.08& 1&$+0.03$\\
Ni I &6.23&$-0.26\pm$0.14&13&$+0.02$ &$-0.37\pm$0.09& 4&$-0.05$
&$-0.47\pm$0.10&34&$-0.04$ &$-0.42\pm$0.10&21&$-0.06$\\
Ni II&6.23&$-0.20\pm$0.12& 2&$+0.08$ &$-0.39\pm$0.04& 1&$-0.07$ & & &
&$-0.44\pm$0.04& 2&$-0.08$\\
Cu I &4.21&$-0.16\pm$0.04& 1&$+0.12$ &$-0.22\pm$0.10& 2&$+0.10$ &$-0.53\pm$0.04&
2&$-0.10$ &$-0.53\pm$0.04& 1&$-0.17$\\
Zn I &4.60&$-0.37\pm$0.07& 3&$-0.09$ &$-0.36\pm$0.08& 2&$-0.04$ &$-0.25\pm$0.14&
4&$+0.18$ &$-0.35\pm$0.10& 3&$+0.01$\\
Y II &2.21&$-0.96\pm$0.05& 3&$-0.68$ &$-1.33\pm$0.06& 4&$-1.01$ &$-1.28\pm$0.10&
5&$-0.85$ &$-1.41\pm$0.16& 4&$-1.05$\\
Zr II&2.59&$-0.67\pm$0.07& 2&$-0.39$ &$-1.42\pm$0.04& 1&$-1.10$ &$-0.93\pm$0.04&
1&$-0.50$ &$-1.28\pm$0.04& 2&$-0.92$\\
Ba II&2.17&$-0.40\pm$0.07& 1&$-0.12$ &$-0.85\pm$0.07& 1&$-0.53$ &$-1.55\pm$0.07&
1&$-1.12$ &$-0.84\pm$0.07& 1&$-0.48$\\
La II&1.13&$-1.06\pm$0.05& 1&$-0.78$ & & & &$-1.05\pm$0.05&
1&$-0.62$ &$-0.53\pm$0.10& 3&$-0.17$\\
Ce II&1.58& & & &$-1.48\pm$0.09& 1&$-1.16$ &$-0.99\pm$0.16&
4&$-0.56$ &$-1.02\pm$0.15& 2&$-0.66$\\
Nd II&1.45& & & & & & &$-1.02\pm$0.12&
8&$-0.59$ & & & \\
Sm II&1.01& & & & & & &$-0.92\pm$0.09&
5&$-0.49$ & & & \\
Eu II&0.52&$-0.22\pm$0.06& 2&$+0.06$ &$-1.21\pm$0.06& 1&$-0.89$ & & &
&$-0.45\pm$0.06& 1&$-0.09$\\
\hline
\end{tabular}
}
\end{table*}
\subsection {IRAS 13110\,-\,6629}
\label{sec:iras13110}
For IRAS 13110\,-\,6629 (GLMP 342) the atmospheric parameters $\xi_{t}$,
T$_{\rm eff}$ were derived from
the study of a large number of Fe~I and Fe~II line; for the estimation of
log~$g$ the ionization equilibrium between Fe~I \& Fe~II as well as that of
Mg~I \& Mg~II, Ca~I \& Ca~II, Ti~I \& Ti~II, Cr~I \& Cr~II were used.
The spectroscopically derived values T$_{\rm eff}$ 6500K, log~$g$ 0.5
and $\xi_{t}$ of 5.0 kms$^{-1}$ as given in Table~\ref{tab:adopted}
are used for abundance analysis.
The temperatures derived from photometry are systematically lower; errors
in reddening correction for object surrounded by circumstellar material
could be one of the possible reasons.
The heliocentric radial velocity $+$11.5$\pm$ 0.5 km s$^{-1}$ on HJD 2455043.479
was derived using 167 lines.
This star also shows a marginal Fe deficiency with [Fe/H]
= $-$0.28. Similar to IRAS 18489\,-\,0629, it exhibits a large
number of C, N lines but [N/Fe] is a little larger, $+$0.75.
Replenishment of C (used in CN processing) from triple $\alpha$
reactions is indicated by [C/Fe] of $+$0.27.
The C/O ratio is $\sim$ 0.5.
[$\alpha$/Fe] as estimated from Mg, Si and S is $+$0.27 which together
with small radial velocity does not make a very compelling case for thick
disk population. Fig.~\ref{fig:depletion} indicates a mild DG effect.
\subsection {IRAS 17579\,-\,3121}
\label{sec:iras17579}
For this star IRAS~17579\-\,3121 (GLMP 686) we estimated T$_{\rm eff}$=5722K from
the Tokunaga (2000) calibrations and T$_{\rm eff}$=6023 $\pm$ 220K,
log~$g$=0.78 $\pm$ 0.27 and $M_J$=$-$3.48 $\pm$ 0.28 mag from the Molina (2012).
From excitation equilibrium of Fe I and Paschen line fit we estimated a temperature
of 6500K and the ionization equilibrium of Fe~I/Fe~II, Ti~I/Ti~II, Cr~I/Cr~II
led to the estimate of log~$g$=0.0.
The photometric estimates of temperature are systematically smaller and log~$g$
larger. Although light variations
are not yet reported for this object but variability in parameters observed
for PAGBs could in addition to the errors in
reddening estimate may possibly explain systematic differences in atmospheric
parameters. The adopted parameters for IRAS~17579\,-\,3121
are T$_{\rm eff}$=6500K, log~$g$=0.0 and
$\xi_{t}$=4.7 km s$^{-1}$ are used for deriving abundances.
\begin{figure*}
\begin{center}
\includegraphics[width=15cm,height=15cm]{SEDs.eps}
\caption{The SED for the program stars is generated
using the existing photometric data and IR colors as described in the text.}
\label{fig:SED}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{plotpcygni2.eps}
\caption{H$\alpha$ profiles in IRAS 18321\,-\,1401, IRAS 18489\,-\,0629, IRAS
17579\,-\,3121 and IRAS 13110\,-\,6629.}
\label{fig:figure5}
\end{center}
\end{figure}
The heliocentric radial velocity measured for the spectrum taken on HJD 2454966.844
using 86 lines is $+$24.7 $\pm$ 0.6 km s$^{-1}$.
The photospheric abundances of IRAS 17579\,-\,3121
shows a moderate metal-deficiency ([Fe/H]= $-$0.32). Among light
elements, N shows significant enrichment [N/Fe] = $+$0.92 indicating
strong signatures of CN cycle; while [C/Fe] of $+$0.39 points to the
products of He burning being brought to the surface.
We estimate [$\alpha$/Fe] of $+$0.4 similar to thick disk value.
We again find the signature of DG effect as illustrated in
Fig.~\ref{fig:depletion}.
\subsection {IRAS 18321\,-\,1401}
\label{sec:iras18321}
There is no $JHK$ photometry available for IRAS 18321\,-\,1401 to estimate the
initial temperature. H$\alpha$ and H$\beta$ being affected by emission were not used
as temperature indicators.
However the matching of P15 (8545\,\AA) and P18 (8438\,\AA) profiles of Paschen
lines with
those synthesized for a set of model atmospheres covering a good range in temperature
and gravities and the combination with the ionization equilibrium requirement for Fe,
Cr, Ti and V led to a temperature value of $\sim$5500 K and surface gravity of log~$g$=0.20.
Fig.~\ref{fig:loci18321-1401} shows the corresponding loci in the temperature-gravity
plane.
\begin{figure}
\begin{center}
\includegraphics[width=8.cm,height=8cm]{LogTeff18321.eps}
\caption{ The P15 (8545\,\AA) \& P18 (8438\,\AA) and the ionization equilibrium loci
for Fe, Cr, V, Ti are plotted in the temperature gravity plane for IRAS18321\,-\,1401.
The red dot indicates the adopted values of T$_{\rm eff}$ and log~$g$ for the calculation
of abundances.}
\label{fig:loci18321-1401}
\end{center}
\end{figure}
From our spectrum obtained on HJD 2455042.632, we measure a heliocentric radial
velocity of $+32.7 \pm 0.6$ km s$^{-1}$ using 141 clean lines.
The star is moderately metal-poor ([Fe/H]=$-$0.43). Like IRAS 13110\,-\,6629 and IRAS
18489\,-\,0629, IRAS 18321\,-\,1401 has a significant number of C~I lines. Very mild
enrichment of N and C is present. However this stars has smaller C/O
ratio of $\sim$0.3.
We find [$\alpha$/Fe] of $+$0.38 similar to thick disk objects.
This star also shows a moderate depletion
of condensable elements similar to that observed in IRAS 18489\,-\,0629 (see
Fig.~\ref{fig:depletion}). It should be noted that the scatter in the
plot is much larger indicating the influence of other parameter/process.
\subsection {IRAS 18489\,-\,0629}
\label{sec:iras18489}
For IRAS 18489\,-\,0629 (PM 1-261) the estimated values of the atmospheric
parameters based upon a large number of lines covering a large
range equivalent widths, LEP and both stages of
ionization used for the abundance analysis are listed in Table \ref{tab:adopted}.
The temperature estimated by spectroscopy is in good agreement with that indicated
by the spectral type mentioned above as well as the photometric estimate
given in Table 2.
From the spectrum taken on HJD 2454340.496 a heliocentric radial
velocity of $+159.6 \pm 0.5$ km s$^{-1}$ is measured using
223 clean unblended absorption lines that cover a wide spectral range.
The present study is the first detailed abundance analysis made
for this star using high resolution spectrum.
The spectrum contained a large number of C and N lines.
The star shows a mild iron deficiency ([Fe/H]= $-0.36\pm0.12$).
From Table~\ref{tab:abund} it can be seen that not only
N shows small but significant enhancement indicating CN processing;
C is replenished by the mixing of triple $\alpha$ products.
We derive a C/O ratio of $\sim$ 0.8.
IRAS 18489$-$0629 shows a mild Na-enrichment, [Na/Fe] = $+$0.30, probably
caused by proton capture on $^{22}$Ne.
Although our study covers five $\alpha$ elements Mg, Si, S, Ti and Ca,
we employ only Mg, Si and S to measure [$\alpha$/Fe] since
Ti and Ca are susceptible to non-LTE and are also affected
by DG effect. The derived [$\alpha$/Fe] of $+$0.38 together
with large radial velocity makes it a likely thick disk object not withstanding
a modest Fe deficiency.
The Fe-peak elements V, Cr, Mn, Co, Ni vary in lockstep with Fe
while s-process elements are significantly deficient.
The abundance pattern can be easily understood via Fig.~\ref{fig:depletion}
which clearly shows the dependence of [X/H] on T$_{C}$. IRAS 18489$-$0629
belongs to the group of PAGBs showing systematic depletion
of condensable elements.
\begin{figure*}
\begin{center}
\includegraphics[width=15cm,height=15cm]{plotTcXFe.eps}
\caption{Elemental abundances [X/H] as a function of condensation temperature
T$_{c}$ (Lodders 2003) for the sample stars. }
\label{fig:depletion}
\end{center}
\end{figure*}
\section {Discussion}
\label{sec:discussion}
In our study of unexplored PAGB candidates we find the signature of
CN processing in all four sample stars although [N/Fe] ranges from
$+$0.24 to $+$0.92. The measured [C/Fe] between +0.27 and +0.39
clearly shows the mixing of He burning products via the third dredge up (TDU).
The sample stars have SEDs displaying a double peak (see Fig.~\ref{fig:SED})
and the observed [$\alpha$/Fe] in excess of $+$0.2 dex commonly seen
in thick disk objects, which contain a large fraction of
s-process enhanced PAGBs (see Table 9 of Sumangala Rao et al. 2012).
However, none of the sample stars exhibit the expected enhancement of s-process
elements. Instead, the observed [X/H] shows a dependence on T$_{C}$
as demonstrated in (Fig.~\ref{fig:depletion}),
although the extent of depletion varies over the sample.
To quantify this effect, we could use the ratio of abundance of least
affected elements like S and Zn (with low T$_{C}$), and the most affected ones such
as Sc, Al and s-process elements.
The potential candidates for the representation of intrinsic
(unaffected) metallicity are S and Zn. But S being an $\alpha$ element,
shows relative enrichment for the thick disk and halo stars.
Zn on the other hand, remains unchanged over a large range in metallicities.
Hence Zn is a better indicator of initial metallicity and has been used
as reference to judge the relative enrichment of various elements.
Among the elements most affected by the DG effect, Al and Zr are not
well represented by many lines while Sc and Y measurements are based upon
4-5 lines. Hence we chose to define DG index as [Zn/H] - ([Sc/H]+[Y/H])/2.
In our sample,
IRAS 18489\,-\,0629 with DG of 0.88 leads the group followed by IRAS 17579\,-\,3121
(0.74), IRAS 18321\,-\,1401 (0.67) and IRAS 13110\,-\,6629 (0.45) being the
least depleted of the four. Another striking feature of the depletion pattern
shown in Fig.~\ref{fig:depletion} is the large scatter found for
IRAS 18321\,-\,1401 and IRAS 13110\,-\,6629. The intriguing aspect of the
sample is that three objects showing a range in
DG index have the same T$_{\rm eff}$ of 6500K and not withstanding its
lower temperature, IRAS 18321-1401 does not have lowest DG index.
All four sample stars have temperatures larger than lower
temperature limit required to sustain the depletion pattern against
the mixing caused in the extended convective envelopes of cooler stars, which
according to Sumangala Rao, Giridhar \& Lambert (2012) is 4800K.
\subsection {The observed SEDs and H${\alpha}$ profile}
\label{sec:profileH}
The SEDs for program stars were generated by comparing the archival
photometric data with the theoretical SED fluxes given in Kurucz (1991)
for models with atmospheric parameters given in Table 3
and adopting the
interstellar extinction model by Steeman \& The (1991).
The ratio of total to selective extinction R= A$_V$/E(B-V) was estimated
iteratively. We have used archival data from
USNO, 2MASS, IRAS, AKARI, WISE and MSX6C available at NASA/IPAC
Infrared Science Archive. We have omitted data with large errors
e.g. IRAS with quality flag =1, 2MASS quality flag = E/X/U/F, MSX6C quality flag =1.
The SEDs constructed for the program stars are shown in Fig.~\ref{fig:SED}.
A Planckian fit to the observed fluxes between 1.0 to 200$\mu$m
leads to dust temperatures of
130K for IRAS 13110\,-\,6629, 110 K for IRAS 17579\,-\,3121,
120K for IRAS18324\,-\,1401 and 110K for IRAS 18489\,-\,0629.
For IRAS 18321\,-\,1401 the scanty photometric data resulted in
poor fit with the theoretically constructed SED.
Fig.~\ref{fig:figure5} shows the profiles of H$\alpha$ at 6562.8 \AA~ of the
program stars. All profiles have a broad shallow component and a deep
narrow component with significant asymmetry.
All profiles show indication of filling in by a rather central
emission. Incipient P-Cygni structure is also seen in all cases.
The P-cygni profiles (or inverse P-Cygni profiles) are of common
occurrence in PAGB stars and are ascribed to the presence of a shock
propagation in the pulsating atmospheres of these stars.
\subsection {Overall abundance patterns}
\label{sec:Oabun}
We have observed a mild enrichment of [Na/Fe] $\sim$ $+$0.2 to
$+$0.3 in
IRAS 18489\,-\,0629, IRAS 13110\,-\,6629 and IRAS 18321\,-\,1401 and of $+$0.6 in IRAS
17579\,-\,3121. The enrichment of the latter object was measured from lines at
4979, 5682, 5688, 6160\,\AA, for which the non-LTE
corrections do not exceed $-$0.10 dex (Gehren et al. 2004; Lind et al. 2011).
This Na enhancement is probably due to a deep mixing that bring products of the
Ne-Na cycle to the surface.
The [$\alpha$/Fe] measured using S, Si, Mg is similar to those found in
thick disk objects although the observed metal deficiency is quite moderate.
The observed [C/Fe] in the range $+$0.2 to $+$0.39 points to the mixing of He burning
products
being brought to the surface but accompanying s-process enhancement either did not
take place or was obliterated by the depletion.
All four stars show depletion but a remarkable aspect of the abundance pattern is
that only elements with T$_{C}$ greater than 1300K
are significantly affected.
It is not clear if the observed depletion pattern is caused by insufficient
mixing of depleted material with photospheric gas or temperature and compositional
structure of circumstellar material favored condensation process only for
high T$_{C}$ elements.
\begin{figure*}
\begin{center}
\includegraphics[width=12.cm,height=10.cm]{plotAbunFIPSIP1.eps}
\caption{Abundances [X/H] versus FIP and SIP for IRAS18321\,-\,1401.}
\label{fig:18321-1401}
\end{center}
\end{figure*}
To understand
the large scatter present in the depletion diagram for IRAS18321\,-\,1401,
we investigated the possible dependence of abundances on
First Ionization Potentials (FIP) as suggested by Kameswara Rao \& Reddy (2005).
A plot of observed abundances as a function of FIP
left panel of Fig.~\ref{fig:18321-1401}
clearly shows a reduced scatter. It appears that elements with FIP
8eV or lower show systematic depletion.
We also explored the possible dependence of
observed abundance on the Second Ionization Potential (SIP)
right column of Fig.~\ref{fig:18321-1401}
and do notice a significant depletion of elements
with SIP lower than 13.6eV. It may be recalled that
Luck \& Bond (1989) had proposed that hydrogen
Lyman continuum produced from a shock in the atmosphere
over ionized elements with a SIP lower than 13.6eV, the
Lyman limit. The Ca, Sc, and s-process elements Y, Zr, Ba, La, Ce, Sm
Eu have lower SIP and hence could be over ionized relative to
LTE estimate hence the LTE abundance analysis would
yield systematically low abundance of these elements.
However, in the SIP plot the elements with similar SIP show
different depletion. On the other hand,
a smooth variation with FIP (with exception
of Mg and Na) implies the systematic removal of species mostly present
in first ionized state. The magnetically driven stellar wind
proposed by Garc\'ia-Segura et al. (2005) for PAGB stars has the potential to
remove the ionized species, but the presence of the stellar wind
need to be established for this object. Although P-Cygni structure
is mildly present in H$_{\alpha}$ the emission barely rises above continuum.
Na D indeed exhibits a complex structure. This star deserves a multi-wavelength
monitoring to understand the observed complex abundance pattern.
\section {Conclusions}
\label {sec:conclusion}
In this paper, we have performed a detailed atmospheric abundance analysis from high
resolution spectra aimed to verify the evolutionary stage of a sample of four PAGB
candidates IRAS13110\,-\,6629, IRAS 17579\,-\,3121, IRAS 18321\,-\,1401
and IRAS 18489\,-\,0629,
selected from their infrared characteristics.
None of them exhibit the s-process enhancement but clear indication of modest
DG effect are observed for IRAS 18489\,-\,0629 and IRAS 17579\,-\,3121.
For IRAS 13110\,-\,6629
and IRAS 18321\,-\,1401 the depletion diagram displays a large scatter.
The scatter for IRAS 18321\,-\,1401 is significantly reduced in [X/H] vs FIP
plot. Hence the possibility of outflowing magnetized column of gas
systematically removing the singly ionized specie need to be explored.
Another competing scenario is proposed by Luck \& Bond (1989) wherein the
Lyman $\alpha$ photons from a shock in the atmosphere over ionized elements
with SIP less than the Lyman limit, 13.6eV.
While SIP plot displays a larger scatter than FIP; the later requires
magnetic field driven wind which are not easy to explain.
In either case a systematic time series spectral monitoring to detect the events of
strong shocks and possible manifestation of stellar wind is required.
The long-term photometric monitoring from the All Sky Automated Survey (ASAS) shows
that IRAS13110\,-\,6629 and IRAS18489\,-\,0629 display periodic variations with characteristic
times between 52 and 60 days, with amplitudes in the 0.2-0.3 magnitude
range. The origin of these variations is most likely pulsational and are of a very
similar nature to those in other well known PAGB of intermediate temperature, such
as 89 Her, HD 161796 and LN Hya.
IRAS 17579\,-\,3121, IRAS 18321\,-\,1401 display fluctuations of a few hundredths of
magnitudes but no periodicities were found.
Multi-wavelength observations are required for better understanding of these objects.
\section*{Acknowledgments}
AAF acknowledges the support of DGAPA-UNAM grant through project IN104612. This work
has made extensive use of SIMBAD database, 2MASS (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),
IRAS, AKARI/IRC data (a JAXA project
with participation of ESA), WISE(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),
MSX6C( Midcourse Space Experiment funded by the Ballistic Missile
Defense Organization with additional support from NASA Office of
Space Science) and the ADS-NASA to which we are
thankful. We would also like to express our gratitude to anonymous referee for
his/her comments which have helped in improving the paper considerably.
|
\section{Introduction}
Accretion in Low-Mass X-ray Binaries (LMXBs) takes place in a region of the spacetime around compact objects where strong gravity effects play an important role. In many cases, the presence of quasiperiodic oscillations (QPOs) with frequencies in the Hz to kHz range has been detected. Various approaches have been used for explaining the QPOs: relativistic precession models, relativistic resonance models, beat frequency models, preferred radii models and others (for a review see e.g. \cite{Klis2006}). Most of these models are related in one way or another to the three characteristic frequencies of particles orbiting around compact objects, namely the orbital frequency and the radial and vertical epicyclic frequencies. On the other hand, a different approach has been taken in \cite{Rezzolla2003,Rezzolla2003a,Montero2004} where the global oscillations of thick accretion discs (accretion torii) were considered, but some of the oscillations of tori are modified (due to finite pressure effects)\ epicyclic frequencies.
The fact that QPOs originate in a strong gravity regime, naturally leads us to the idea of using them as a test of alternative theories of gravity, since almost all of the current observations constraining the alternative theories of gravity are in the weak field regime (see, e.g. \cite{Will2006}). One of the simplest and most natural generalizations of General Relativity (GR) are the scalar-tensor theories of gravity (STTs). Their essence is in one or several scalar fields which are mediators of the gravitational interaction, in addition to the spacetime metric. A very important property of STT is that in the physical Jordan frame there is no direct interaction between the matter and the scalar field, and consequently the weak equivalence principle is satisfied.
Different proposals of STTs have been examined in the literature and an interesting subclass are cases which are indistinguishable from GR in the weak field regime, but which can have appreciable deviations in strong fields. In such cases interesting phenomena, such as non-uniqueness and bifurcations of the solutions, could exist. An example is the so-called spontaneous scalarization of neutron stars \cite{Damour1993}. In a density regime near the maximum mass for neutron stars, new solutions that have a nontrivial scalar field appear, in addition to the trivial (GR) solution. It turns out that the scalarized solution is energetically more favorable \cite{Damour1993} than the GR solution and its stability was studied in \cite{Harada1998,Harada1997}. Different astrophysical implications of scalarized neutron stars were examined in \cite{Sotani04,Sotani2005,DeDeo2003,DeDeo2004,Barausse2013,Palenzuela2014,Shibata2014,Novak1998,Novak1998a,Harada1997a,Sotani2014}. Similar nonlinear phenomena in STTs are present also in the black hole case \cite{Stefanov2008,Doneva2010,Pani2011a,Cardoso2013a}.
Most studies of neutron stars in STT were for either nonrotating models or in the slow rotation approximation (of linear order in the angular velocity) \cite{Damour1996,Sotani2012}. The only exception is our recent work \cite{Doneva2013} where we obtained equilibrium models of scalarized, rapidly rotating neutron stars, up to the mass-shedding limit. In \cite{Doneva2013} we found that rapidly rotating scalarized neutron stars exist for a significantly wider range of central densities, compared to the nonrotating case. In addition, the mass and radius of rapidly rotating scalarized models differ significantly more from their GR counterparts, than in the nonrotating case. Thus, even for moderate values of the scalar-field coupling parameter, for which scalarized nonrotating solutions either do not exist or differ only marginally from the GR solutions, rapidly rotating scalarized neutron stars exist and can deviate considerably from neutron stars in pure Einstein's theory.
The effect of the scalar field on the epicyclic frequency (for a particular version of STT) was examined in \cite{DeDeo2004} for the case of nonrotating neutron stars. Accreting neutron stars in LMXBs can be rapidly rotating, with spin frequencies reaching approximately $700$Hz \cite{Hessels2006} and thus we present here the first study of epicyclic frequencies around rapidly rotating scalarized neutron stars. We find that although the effect of a nontrivial scalar field on the epicyclic frequencies is quite small for nonrotating models (for values of the coupling parameter that are in agreement with the observational constraints \cite{Freire2012,Antoniadis13}), it is more appreciable for rapidly rotating stars.
The paper is organized as follows:\ In Section II we give the theoretical background and the equations for the orbital and epicyclic frequencies. The results are presented in Section III in the case of neutron and strange stars. In the Appendix we derive analytical formulae for the orbital and epicyclic frequencies in the nonrotating case.
\section{MAIN equations} \label{Sec:BasicEquation}
The action in scalar-tensor theories of gravity, in the physical \textit{Jordan} frame, is given by \cite{Fujii2003,Damour1992}:
\begin{eqnarray} \label{JFA}
S = {1\over 16\pi G_{*}} \int d^4x \sqrt{-{\tilde
g}}\left[{F(\Phi)\tilde R} - Z(\Phi){\tilde
g}^{\mu\nu}\partial_{\mu}\Phi
\partial_{\nu}\Phi -2 U(\Phi) \right] +
S_{m}\left[\Psi_{m};{\tilde g}_{\mu\nu}\right] ,
\end{eqnarray}
where $G_{*}$ is the bare gravitational constant, $\Phi$ is the scalar field, ${\tilde R}$ is the Ricci scalar curvature with respect to the spacetime metric ${\tilde g}_{\mu\nu}$ , ${\tilde g}$ is the determinant of the metric and $F, Z, U$ are functions of the scalar field. The matter fields are collectively denoted by $\Psi_{m}$ and their action is $S_{m}$. As one can notice, the scalar field does not appear explicitly in the action of the matter, in order for the weak equivalence principle to be satisfied\footnote{Of course one can consider more general cases with a direct coupling between the matter and the scalar field, but this is out of the scope of the current paper.}. Instead, the scalar field influences the matter only through the spacetime metric. This means that the equations of motion of test particles in scalar-tensor theories are the same as the equations of motion in general relativity.
Having this in mind and using the procedure described in \cite{Ryan1995,Pappas2012b}, one can easily derive the orbital and epicyclic frequencies of a particle orbiting around a compact object by examining the geodesics and their perturbations. For rapidly rotating neutron stars we consider a stationary, axisymmetric spacetime (see \cite{Friedman2013} for the general form of the metric). It is easy to show that in the usual quasi-isotropic coordinates, the orbital frequency for a circular equatorial orbit with coordinate radius $r_c$ is given by:
\begin{eqnarray}\label{Eq:OmegaP}
\Omega_{p} = \left.\frac{-\partial_r {\tilde g}_{t\phi} \pm \sqrt{(\partial_r {\tilde g}_{t\phi})^2 - \partial_r {\tilde g}_{tt}\partial_r {\tilde g}_{\phi\phi}}}{\partial_r {\tilde g}_{\phi\phi}}\right|_{r=r_c, \theta=\frac{\pi}{2}},
\end{eqnarray}
where ${\tilde g}_{tt}$, ${\tilde g}_{\phi\phi}$ and ${\tilde g}_{t\phi}$ are the corresponding components of the metric.
If a particle on a stable circular orbit is perturbed, then it oscillates with some characteristic epicyclic frequencies $\omega_r$ and $\omega_\theta$ in the radial or vertical direction, respectively. These frequencies can be obtained by perturbing the equations of motion and after some calculations we arrive at \cite{Ryan1995,Pappas2012b}:
\begin{eqnarray}
\omega_r^2 &=& \frac{1}{2{\tilde g}_{rr}}\left[({\tilde g}_{tt}+{\tilde g}_{t\phi}\Omega_p)^2\;\partial_r^2 \left(\frac{{\tilde g}_{\phi\phi}}{Y}\right) -2 ({\tilde g}_{tt}+{\tilde g}_{t\phi}\Omega_p)({\tilde g}_{t\phi}+{\tilde g}_{\phi\phi}\Omega_p) \;\partial_r^2 \left(\frac{{\tilde g}_{t\phi}}{Y}\right) \right. + \notag \\ \notag \\
&&\left.+ ({\tilde g}_{t\phi}+{\tilde g}_{\phi\phi}\Omega_p)^2\;\partial_r^2 \left(\frac{{\tilde g}_{tt}}{Y}\right)\right]_{r=r_c, \theta=\frac{\pi}{2}},
\label{eq:epicrad}\\ \notag \\
\omega_\theta^2 &=& \frac{1}{2{\tilde g}_{\theta\theta}}\left[({\tilde g}_{tt}+{\tilde g}_{t\phi}\Omega_p)^2\;\partial_\theta^2 \left(\frac{{\tilde g}_{\phi\phi}}{Y}\right) -2 ({\tilde g}_{tt}+{\tilde g}_{t\phi}\Omega_p)({\tilde g}_{t\phi}+{\tilde g}_{\phi\phi}\Omega_p) \;\partial_\theta^2 \left(\frac{{\tilde g}_{t\phi}}{Y}\right) \right. + \notag \\ \notag \\
&&\left.+ ({\tilde g}_{t\phi}+{\tilde g}_{\phi\phi}\Omega_p)^2\;\partial_\theta^2 \left(\frac{{\tilde g}_{tt}}{Y}\right)\right]_{r=r_c, \theta=\frac{\pi}{2}},
\label{eq:epicvert}
\end{eqnarray}
where
\begin{eqnarray}
Y = {{\tilde g}_{tt}{\tilde g}_{\phi\phi}-\tilde g}_{t\phi}^2 .
\end{eqnarray}
Eqns. (\ref{eq:epicrad}) and (\ref{eq:epicvert}) for the epicyclic frequencies of particles orbiting a compact object are the same in STT as in GR.
For a nonrotating star the vertical epicyclic frequency is equal to the orbital frequency, i.e. $\omega_\theta = \Omega_p$ for $\Omega=0$, where $\Omega$ is the angular velocity of the star, while $\omega_r^2$ can vanish at the innermost stable circular orbit (ISCO) near the neutron star surface, signaling radial instability of the circular orbits at smaller radii (an effect absent in Newtonian gravity for spherical, ideal fluid sources). In simple accretion disk models, the ISCO defines the inner edge of the disc. For some neutron star models, $\omega_r^2$ does not become negative outside the stellar surface so that all circular orbits are stable (the accretion disc can extend down to the surface).
While the Jordan frame is the physical frame in which observable quantities are measured,
the field equations take a somewhat simpler form in the so-called \textit{Einstein} frame, where the metric $g_{\mu\nu}$ is related to the physical metric $\tilde g_{\mu\nu}$
by the conformal transformation
\begin{equation}\label {CONF1}
g_{\mu\nu} = F(\Phi){\tilde g}_{\mu\nu} .
\end{equation}
If one introduces a new scalar field $\varphi$ (the dilaton field) satisfying
\begin{equation}\label {CONF2}
\left(d\varphi \over d\Phi \right)^2 = {3\over
4}\left({d\ln(F(\Phi))\over d\Phi } \right)^2 + {Z(\Phi)\over 2
F(\Phi)},
\end{equation}
then the inverse transformation can be written as
\begin{equation}\label {CONF4}
{\tilde g}_{\mu\nu} = {\cal A}^2(\varphi) g_{\mu\nu}
\end{equation}
and the Einstein frame action has the following simpler form:
\begin{eqnarray}
S= {1\over 16\pi G_{*}}\int d^4x \sqrt{-g} \left(R -
2g^{\mu\nu}\partial_{\mu}\varphi \partial_{\nu}\varphi -
4V(\varphi)\right)+ S_{m}[\Psi_{m}; {\cal A}^{2}(\varphi)g_{\mu\nu}],
\end{eqnarray}
where
\begin{eqnarray}\label {CONF3}
{\cal A}(\varphi) &=& F^{-1/2}(\Phi), \\
V(\varphi) &=& U(\Phi)F^{-2}(\Phi)/2.
\end{eqnarray}
The complication that arises is that in this frame the scalar field appears explicitly in the action of the matter via the coupling function ${\cal A}(\varphi)$. Nevertheless, the field equations are simpler in the Einstein frame so that has become the preferred frame for solving the field equations to construct models of neutron stars or black holes in STT.
In order to obtain the solutions describing scalarized rapidly rotating neutron stars, we used a modification of the {\tt rns} code \cite{Stergioulas95} implemented recently in \cite{Doneva2013}. In \cite{Doneva2013} only the case of a polytropic equation of state (EOS) was considered -- here we use a sample of tabulated, microphysical EOSs.
More details on the field equations governing rapidly rotating neutron stars in STT and on their properties such as mass, radius, angular momentum, etc., can be found in \cite{Doneva2013}.
After obtaining the numerical solution describing a neutron star equilibrium model in the Einstein frame, the physical metric is obtained from Eq. \eqref{CONF4} and used in Eqs. \eqref{Eq:OmegaP}, (\ref{eq:epicrad}) and (\ref{eq:epicvert}) for evaluating the orbital and epicyclic frequencies.
We consider the case of a vanishing scalar field potential, $V(\varphi)=0$, and choose the following standard form of the coupling function:
\begin{equation}
{\cal A}(\varphi) = e^{\frac{1}{2} \beta \varphi^2},
\end{equation}
which leads to a scalar-tensor theory that is indistinguishable from GR in the weak field regime, but which can differ significantly when strong fields are considered. More precisely, for negative values of $\beta $, spontaneous scalarization of the neutron stars can be observed for a density range near the maximum mass model \cite{Damour1993,Damour1996,Doneva2013} -- the scalarized solution exists in addition to the (trivial) GR solution with vanishing scalar field, but is energetically favoured over the latter.
Rapid rotation both extends the range of central densities and values of $\beta$ for which scalarization occurs, and significantly enhances the differences in mass and radius with respect to GR.
\section{NUMERICAL\ Results}\label{Sec:Results}
For the EOS, we consider both hadronic and strange matter EOSs. In addition to the properties of orbital and epicyclic frequencies, we point out that we present the first rapidly rotating, scalarized equilibrium models for tabulated hadronic EOSs and for strange star EOSs\footnote{To our knowledge, scalarization of strange stars has not been studied before even in the nonrotating case.}.
\subsection{Hadronic EOSs}
We consider two representative hadronic EOSs that cover a large domain in the presently uncertain mass-radius relation of neutron stars. EOS APR \cite{AkmalPR} has an average stiffness, while EOS L \cite{Pandharipande1976} is one of the stiffest tabulated EOSs that have been proposed. Both have a maximum mass larger than the $2 M_\odot$ observational constraint \cite{Demorest10,Manousakis12,Antoniadis13,Lattimer12}. The mass vs. radius relation for these two EOSs is shown in Fig. \ref{Fig:M(R)}. Two sequences of models are shown -- nonrotating (solid lines) and models rotating up to the mass-shedding (Kepler) limit, where $\Omega=\Omega_K$ (dashed lines). The unstable (to collapse) branch is shown as a dotted line.\footnote{For the nonrotating models, the instability to collapse sets in at the maximum mass model. We use the same criterion for models at the mass-shedding limit to indicate \textit{approximately} the onset of the quasi-radial instability (the actual marginally stable model will be nearby). } In each case, we calculate neutron star solutions with $\beta=0$ (the GR case) and solutions with $\beta=-4.5$ and $\beta=-4.8$, which are scalarized in a certain range of central densities (shown in green and red colors) and coincide with the GR solution outside of this range. The current constraint on the coupling parameter from astrophysical observations is $\beta\geq-4.5$ \cite{Freire2012,Antoniadis13}. However, we also include the somewhat stronger case of $\beta=-4.8$ in order to demonstrate the sensitivity of our results on the value of the coupling parameter.
Similarly to the case of polytropes that we studied in \cite{Doneva2013}, we find that also for the hadronic EOSs rapid rotation both enlarges the range of central densities for which scalarized solutions exist and causes significantly larger deviations from the GR solution, compared to the nonrotating limit. Even though for the current bound of $\beta=-4.5$ the scalarization has only a marginal effect on the structure of nonrotating neutron stars, rapidly rotating models still show a significant effect. In addition to the nonrotating and mass-shedding sequences, Fig. \ref{Fig:M(R)} also shows (with dotted lines) the models for which the corotating ISCO\ touches the surface of the star. Models above the dotted lines possess a gap between the ISCO and the surface, while for models below the dotted lines there is no region of unstable circular orbits and an accretion disk can reach the surface.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.48\textwidth]{M_R_e_APR.eps}
\includegraphics[width=0.48\textwidth]{M_R_eosL.eps}
\caption{The mass as a function of the radius for EOS APR and EOS L, for two different values of $\beta$. The solid lines correspond to nonrotating solutions and the dashed lines to models rotating at the mass-shedding (Kepler) limit (in each case, the branches with models unstable to collapse are shown with thin dotted lines). Models that possess a gap between the ISCO\ and the surface are above the nearly horizontal dotted lines, while models where all circular orbits in the equatorial plane are stable are below those lines. }
\label{Fig:M(R)}
\end{figure}
Next, we investigate the deviations of orbital properties of STT\ solutions from GR solutions in two cases: for nonrotating models and for models rotating at the mass-shedding limit. All of the quantities are plotted as a function of mass along these two sequences. Fig. \ref{Fig:risco(M)} shows the radius of the ISCO. Notice that for low masses, when the ISCO touches the surface, the radius of the surface is shown instead and we indicate the model for which a gap between the ISCO and the surface starts to form with an asterisk. In the nonrotating limit, the ISCO is only marginally affected by scalarization. In contrast, for stars rotating at the mass-shedding limit, the effect is much more pronounced and the radius of the ISCO for the maximum mass model increases by 25\% for EOS APR and by 33\% for EOS L. Here and below, all quantitative comparisons with the GR solution will be reported for $\beta=-4.5$ (we remind that we show results for $\beta=-4.8$ only as a indication of the sensitivity of the results on the value of $\beta$).
Figure \ref{Fig:OmK1(M)} shows the orbital frequency $\nu_p=\Omega_p/2\pi$ at the ISCO for nonrotating and masss-shedding sequences, as a function of mass. For the maximum mass models of the mass-shedding sequence, the orbital frequency at the ISCO decreases by 17\% for EOS\ APR and by 22\% for EOS L for STT models, compared to the corresponding models in GR.
Figure \ref{Fig:OmTh(M)} shows the difference between the orbital frequency $\nu_p$ and the vertical epicyclic frequency $\nu_\theta=\omega_\theta/2\pi$ at the ISCO. For nonrotating models this difference vanishes, while for models at the mass-shedding limit
this difference is reduced by 47\% for the maximum mass model of EOS APR and by 48\% for EOS L. Finally, Fig. \ref{Fig:omr1(M)} shows the corresponding plots for the maximum value of the radial epicyclic frequency $\nu_r=\omega_r/2\pi$, which is reduced by 22\% for the maximum mass model at the mass-shedding limit of EOS APR and by 26\% for EOS L.
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{rISCO_M_APR.eps}
\includegraphics[width=0.48\textwidth]{rISCO_M_eosL.eps}
\caption{The radius of the corotating ISCO for nonrotating models (solid lines) and for models rotating at the mass-shedding limit (dashed lines), as a function of mass along these sequences. An asterisk separates low-mass models, for which there are no unstable circular orbits, from higher-mass models, for which the ISCO is outside the surface of the star. }
\label{Fig:risco(M)}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{OmK1_M_APR.eps}
\includegraphics[width=0.48\textwidth]{OmK_ISCO_Mass_eosL.eps}
\caption{The orbital (Keplerian) frequency of a particle at the ISCO (or at the surface for low-mass models) as a function of mass. The notation is the same as in Fig. \ref{Fig:risco(M)}. }
\label{Fig:OmK1(M)}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{Omtheta1_M_APR.eps}
\includegraphics[width=0.48\textwidth]{Omtheta1_M_eosL.eps}
\caption{The difference between the orbital frequency and the vertical epicyclic frequency at the ISCO as a function of mass. The notation is the same as in Fig. \ref{Fig:risco(M)}.}
\label{Fig:OmTh(M)}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{omr1_max_M_APR.eps}
\includegraphics[width=0.48\textwidth]{omr1_max_mass_eosL.eps}
\caption{Maximum value of the radial epicyclic frequency as a function of mass. The notation is the same as in Fig. \ref{Fig:risco(M)}.}
\label{Fig:omr1(M)}
\end{figure}
In summary, while for nonrotating models the particular STT we consider has an almost negligible effect on both the structure and the orbital and epicyclic frequencies, the opposite is true for models at the mass-shedding limit, where differences to GR can become significant.
For astrophysical observations, we consider two specific models of neutron stars rotating with (approximately) the maximum observed frequency for millisecond pulsars, $\nu=700Hz$ \cite{Hessels2006}. The two models are for EOS APR and EOS L, and they have masses $M=1.8 M_\odot$ and $M=2.3 M_\odot$ respectively\footnote{A larger mass is chosen for EOS L, because normally scalarization is stronger close to the maximum mass and the maximum mass for EOS L is quite high.}. Fig. \ref{Fig:omega(rc)_APR} shows the profiles of the frequencies $\nu_p$, $\nu_r$ and $\nu_p - \nu_\theta $ as a function of radius . The solid lines correspond to neutron stars in GR (with $\beta=0$) , while dashed lines are for the scalarized models with $\beta=-4.5$. For stable orbits (outside the ISCO) the differences between the GR and STT solutions are generally small. Only $\nu_p-\nu_\theta$ shows an appreciable difference at $R_{\rm ISCO}$ for the $M=2.3 M_\odot$\ model of EOS L.
In principle, this deviation could become important in models of quasi-periodic oscillations in low-mass x-ray binaries and could serve as a test of strong gravity (if other parameters are well constraint). Should more rapidly rotating neutron stars in LMXBs be discovered in the future, then stronger deviations are possible (up to the maximum deviations reported above for the maximum mass rotating models).
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{omega_rc_APR_M1.8.eps}
\includegraphics[width=0.48\textwidth]{omega_rc_L.eps}
\caption{Profiles of $\nu_p$, $\nu_r$ and $(\nu_p - \nu_\theta)\times 10$ outside the star, where $r$ is the circumferential radius of an orbit. The left panel considers a model with $M=1.8 M_\odot$, $\nu=700 Hz$ and EOS APR, and the right panel corresponds to $M=2.3 M_\odot$, $\nu=700 Hz$ and EOS L. Models with $\beta=-4.8$, $\beta=-4.5$ and $\beta=0$ are shown in each case. }
\label{Fig:omega(rc)_APR}
\end{figure}
\subsection{Strange stars}
We consider the standard strange matter EOS with parameters described in \cite{Gondek-Rosinska2008} (denoted by SQSB60 there), for which the maximum mass for nonrotating model is very close to $2M_\odot$ (other versions of the strange star EOS can somewhat surpass this value, but quantitatively the results would be similar to our chosen EOS). Fig. \ref{Fig:M(R_e)_SQS60} shows the mass-radius relation for nonrotating models and for models at the mass-shedding limit, for the same values of the coupling constant $\beta$ as for the hadronic EOSs in the previous section. To our knowledge, these are the first results for scalarized strange star models, even in the nonrotating case. The mass and radius of rapidly rotating models at the mass-shedding limit are dramatically larger than for the GR soution, much more than for hadronic EOSs. The maximum mass reaches $5M_\odot$, compared to $2.4 M_\odot$ in GR, with the corresponding radius reaching 25 km, compared to 17 km. Strange stars can reached a higher oblateness compared to hadronic EOSs when rotating at the mass-shedding limit. Already, the GR solution deviates at mass-shedding more from the nonrotating models than the case for hadronic EOSs. Because the effect of scalarization on the equilibrium structure becomes more dramatic with increasing rotation \cite{Doneva2013}, the end result is a significantly larger deviation of STT solutions from GR solutions than for hadronic EOSs.
Similar to the neutron star case, the ISCO for the less massive nonrotating strange star models in GR ($\beta=0$) coincides with the surface and only for the more massive models it is outside the stellar surface. But for all the models (above $1 M_\odot$ at least) rotating at the mass shedding limit, the ISCO is always located outside the star.
In Fig. \ref{Fig:M(R_e)_SQS60} the sequence of models for which the ISCO barely touches the surface is shown as a dotted line starting from the nonrotating sequence. The ISCO is above the stellar surface for all of the scalarized strange star models.
\textcolor[rgb]{0,0.537255,1}{}\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{M_R_e_SQS60.eps}
\caption{Mass vs. radius relation for the strange star EOS SQSB60 for two different values of $\beta$. The solid lines correspond to nonrotating solutions and the dashed ones to models rotating at the mass-shedding limit. The dotted line starting from the nonrotating GR solutions identifies the sequence of models where the corotating circular equatorial orbit at the stellar surface is marginally stable, i.e. models for which $\omega^{2}_r(r=r_e,\theta=\pi/2)=0$. The branches of equilibrium solutions above the maximum mass, which are unstable to collapse, are shown with a thin dotted line.}
\label{Fig:M(R_e)_SQS60}
\end{figure}
The left panel of Fig. \ref{Fig:rISCO_OmK_SQS60} shows the radius of the ISCO (or the radius of the star, when there are no unstable circular orbits), as a function of the mass for the nonrotating and mass-shedding sequences. For the maximum mass model on the mass-shedding sequence, the radius of the ISCO is 70\% larger than for the corresponding model in the GR case. The right panel of Fig. \ref{Fig:rISCO_OmK_SQS60} shows the orbital frequency at the ISCO along the nonrotating and mass-shedding sequences. In \cite{Stergioulas1999} it was found that for strange star models in GR, rotating at the mass-shedding limit, the orbital frequency at the ISCO deviates more from the nonrotating case than for hadronic EOSs, which can be seen in the right panel of Fig. \ref{Fig:rISCO_OmK_SQS60} for $\beta=0$. For the scalarized solutions, the orbital frequency decreases even more and for the maximum mass model at the mass-shedding limit it is 38\% smaller than in GR.
The left panel of Fig. \ref{Fig:omr_omth_SQS60} shows the maximum value of the radial epicyclic frequency, which decreases by 38\% for the scalarized model with maximum mass at the mass-shedding limit, compared to the corresponding model in GR.
The right panel of Fig. \ref{Fig:omr_omth_SQS60} shows that
$\nu_p - \nu_\theta$ for the maximum mass models on the mass-shedding limit has similar values in GR and STT, but for significantly different masses.
Finally, Fig. \ref{Fig:omega(rc)_SQS60} shows the profiles of $\nu_p$, $\nu_r$ and $\nu_p - \nu_\theta$ as a function of radius for $\beta=0, -4.5$ and $-4.8$. All three models are rotating with a spin frequency of 700Hz and have a mass of $1.8M_\odot$. In this case, the rotation is too slow to significantly alter the above frequencies in STT, compared to GR.
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{rISCO_M_SQS60.eps}
\includegraphics[width=0.48\textwidth]{OmK1_M_SQS60.eps}
\caption{{\it (Left panel)} The radius of the ISCO (or the stellar radius for low masses), as a function of mass. {\it (Right panel)} The Kepler frequency of a particle at the ISCO.}
\label{Fig:rISCO_OmK_SQS60}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{omr1_max_M_SQS60.eps}
\includegraphics[width=0.48\textwidth]{Omtheta1_M_SQS60.eps}
\caption{{\it (Left panel)} The maximum value of the radial epicyclic frequency as a function of mass for strange stars. {\it (Right panel)} The difference between the orbital frequency and the vertical epicyclic frequency as a function of mass for strange stars rotating at the mass-shedding limit. }
\label{Fig:omr_omth_SQS60}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[width=0.48\textwidth]{omega_rc_SQS60_M1.8.eps}
\caption{The profiles of the quantities $\nu_p$, $\nu_r$ and $(\nu_p - \nu_\theta)\times 10$ outside the star, where $r$ is the circumferential radius of an orbit, for scalarized strange stars ($\beta=-4.8$ and $\beta=-4.5$) and for a strange star with zero scalar field ($\beta=0$).
All models have the same mass of $M=1.8 M_\odot$ and rotational frequency of $\nu=700 Hz$.}
\label{Fig:omega(rc)_SQS60}
\end{figure}
\section{Conclusions}\label{Sec:Conclusions}
In the present paper we examined the orbital and epicyclic frequencies of particles orbiting rapidly rotating neutron stars within a particular class of scalar-tensor theories of gravity. Because geodesics formally only depend on the metric in scalar tensor theories, our formalism can be applied to any other class of STT. The interest in the orbital and epicyclic frequencies stems from the fact that these are involved in models that attempt to explain QPOs that are regularly observed in accretion disks.
We consider a class of STT that is indistinguishable from GR in the weak field regime, but it can differ significantly for strong fields. For this class of STT, scalarization of the solution in a certain range of central densities is observed, i.e. the neutron stars can develop a strong, nontrivial scalar field. The effect of scalarization is more pronounced for rapidly rotating neutron stars.
The orbital and epicyclic frequencies of scalarized neutron stars were examined in \cite{DeDeo2004}, but only for the case of nonrotating neutron stars. The extension to rapid rotation, considered here, is important for two reasons. First, neutron stars in LMXBs can reach high rotational frequencies, due to the accretion spin-up. Second, the developed nontrivial scalar field is much stronger for rapid rotation, compared to the static case. As a matter of fact, if we consider values of $\beta$ that are in agreement with the current observational constrains, only the rapidly rotating case could give us significant deviations from GR.
Using two representative hadronic EOSs and one strange star EOS, we studied in detail the effect of a nontrivial scalar field on the orbital and epicyclic frequencies, along sequences of models that are nonrotating, rotating at the mass-shedding limit or rotating with a spin frequency of 700 Hz. The effect of scalarization is marginal for nonrotating models, when one considers values of $\beta>-4.5$. However, as the rotation increases, these differences become more appreciable, and at the mass-shedding limit, particles orbiting the scalarized compact stars have significantly different orbital and epicyclic frequencies than in GR. The position of the ISCO also changes considerably. The deviations from GR for strange star models can be larger compared to hadronic EOSs.
We studied in more detail neutron star models with masses around $2M_\odot$, and the maximum observed rotational frequency of $\sim 700{\rm Hz}$. It turns out that the presence of nontrivial scalar field is evident in the profiles of the orbital and epicyclic frequencies (even though EOS uncertainties are comparable or larger than this effect, at present). Should the EOS become tightly constraint in the future, then the current bound $\beta>-4.5$ allows of strong-field tests of alternative theories of gravity, especially in the case of a very stiff EOS, such as EOS L and for a sufficiently high mass. If neutron stars in LMXBs spinning faster than 700 Hz are observed in the future, then it is more likely that the orbital properties of particles in accretion disks could serve in strong-field gravity tests.
Notice that although our present paper focuses on orbital and epicyclic frequencies, we actually present here the first results for equilibrium configurations of rapidly rotating scalarized compact stars with either realistic hadronic EOS or strange matter EOS. The detailed investigation of the equilibrium properties of these object will be considered elsewhere, but from our current results we can make one important observation: the scalarized rapidly rotating strange stars can deviate stronger from the GR case in their equilibrium parameters, such as mass and radius, compared to neutron stars. The main reason is that the oblateness of strange stars can reach larger values at the mass-shedding limit, with respect to the neutron star case, which amplifies the effect of the scalar field.
For nonrotating models, we also provide tables of model parameters that allow the analytic construction of the exterior spacetime around a scalarized neutron star, as well as analytic expressions for the orbital and radial epicyclic frequencies, which are useful for checking numerical codes.
\acknowledgments
DD would like to thank the Alexander von Humboldt Foundation for a stipend. KK and SY would like to thank the Research Group Linkage Programme of the Alexander
von Humboldt Foundation for the support. The support by the Bulgarian National Science Fund
under Grant DMU-03/6, by the Sofia University Research Fund under Grant 63/2014 and by the German Science Foundation (DFG) via
SFB/TR7 is gratefully acknowledged. Partial support comes from ``New-CompStar'', COST Action MP1304. NS is grateful for the hospitality of the T\"ubingen group during an extended visit.
|
\section{Introduction}
Voting has been recognized as a common approach for preference aggregation and collective decision making whenever there exists more than one
alternative for a community to choose from. It comes with a wide variety of applications which ranges from
multi-agent systems, political elections, recommendation systems, machine learning etc. \cite{DBLP:journals/cj/PittKSA06,DBLP:conf/hci/Popescu13b,DBLP:conf/atal/Xia13}. Unfortunately, due to the Gibbard-Satterthwaite theorem \cite{Gibbard73,Satterthwaite75}, any voting system which satisfies a set of desirable criteria is not strategy-proof, that is, there exists a voter who can make himself better off by misreporting his vote. To address this issue, Bartholdi et al. \cite{BARTHOLDI89} introduced the computational complexity to the study of strategic voting problems. The point is that if it is {\nph} to successfully perform a specific strategic behavior, the strategic individual(s) may give up performing such a strategic behavior.
Since then, exploring the complexity of strategic behaviors in voting systems has been one of the main focus of computational social choice community. We refer to \cite{DBLP:conf/birthday/BetzlerBCN12,DBLP:conf/sofsem/ChevaleyreELM07,Lindner08} for comprehensive surveys on this topic.
Recently, this purely worst-case analysis, which ignores real-world settings, was criticized by researchers. See \cite{DBLP:conf/aaai/ConitzerS06,DBLP:journals/aim/FaliszewskiP10,DBLP:conf/aldt/MatteiW13,DBLP:journals/jair/ProcacciaR07,DBLP:journals/amai/Walsh11} for detailed discussions. For their purpose, they proposed diverse measurements to evaluate the feasibility of strategic behaviors in practical elections. For example, Procaccia and Rosenschein \cite{DBLP:journals/jair/ProcacciaR07} introduced
the concept of junta distributions
(generally speaking, these are distributions over the elections that satisfy several constraints.)
and argued that
if an (heuristic) algorithm often solve the manipulation problem when the instances are
distributed according to a junta distribution, it would
also often solve the manipulation problem when the
instances are distributed according to many other plausible
distributions.
In this paper, we study the strategic behaviors in a variety of voting systems from the parameterized complexity perspective. The parameterized complexity was first systematically introduced by Downey and Fellows \cite{fellows99}. Differently from the classical complexity, the parameterized complexity deals with problems in two dimensions. More specifically, an instance of a {\it{parameterized problem}} consists of a main part
and a parameter ${\para}$ which is normally a positive integer.
The main task in the parameterized complexity is to explore how the parameters affect the complexity of the problems. It turned out that under the framework of the parameterized complexity, many {\nph} problems become tractable with respect to specific parameterizations. More precisely, many {\nph} problems turned out to be solvable in $f(\para)\cdot |{\ins}|^{O(1)}$ time. Here, $f$ is a computable function that depends only on the parameter $\para$. All the parameterized problems which fall into this category are called fixed-parameter tractable ({\fpt} for short). However, the parameters do not always behave in this way. There are parameterized problems which do not admit {\fpt}-algorithms unless the parameterized complexity hierarchy collapses at some level, which is commonly believed to be unlikely. This discussion is beyond our focus in this paper. For a comprehensive understanding of parameterized complexity, we refer to the text of Niedermeier \cite{rolf06}.
For recent developments of parameterized complexity applied to computational social choice, we refer to \cite{DBLP:conf/birthday/BetzlerBCN12,Lindner08}.
A natural parameter in the voting scenario is the number of candidates. This parameter is relatively small in some real-world settings. For example, a political election normally contains only a few candidates. A reference library of preference data assembled by Mattei and Walsh \cite{DBLP:conf/aldt/MatteiW13} also reveals such a situation. Out of their 14 sets of election data from the real-life settings, 5 data sets contain less than 10 candidates each (\today).
In this paper, we aim at deriving a general framework for achieving {\fpt} results with respect to the number of candidates.
To this end, we adopt the concept of the class of generalized scoring rules which was introduced by Xia and Conitzer \cite{DBLP:conf/sigecom/XiaC08a}. In particular, we prove that the manipulation, control (all the 22 standard types) and bribery problems are {\fpt} for most of the generalized scoring rules, with respect to the number of candidates. Since many common voting rules fall into the category of the generalized scoring rules, these tractability results hold for these voting rules, among which are all the positional scoring rules (e.g., Borda, $r$-Approval, Veto, Plurality), Copeland$^{\alpha}$, Maximin, Bucklin, Ranked pairs, Schulze, Nanson's and Baldwin's.
\bigskip
\noindent{\bf{Related Works.}} Hemaspaandra et al. \cite{DBLP:conf/atal/HemaspaandraLM13} recently studied the manipulation, control and bribery problems in Schulze's and Ranked pairs voting systems. They proved that all these strategic problems in Schulze and Ranked pairs voting systems are {\fpt} with respect to the number of candidates. Gaspers et al.~\cite{DBLP:conf/atal/GaspersKNW13} proved that the manipulation problem in Schulze voting system is indeed polynomial-time solvable for any number of manipulators. Faliszewski et al. \cite{DBLP:journals/jair/FaliszewskiHHR09} studied Copeland$^{\alpha}$ control problems and achieved {\fpt} results for most of the control problems in Copeland$^{\alpha}$ voting with respect to the number of candidates, for every $0\leq \alpha\leq 1$. Besides the manipulation, (22 standard forms of) control and bribery problems, many other strategic voting problems were also studied from the parameterized complexity perspective by researchers. Faliszewski et al. \cite{DBLP:journals/jair/FaliszewskiHH11} studied a multimode control problem (in this model, the strategy individuals are allowed to add votes, delete votes, add candidates, delete candidates, and change votes simultaneously) and proved that this problem is {\fpt} with respect to the number of candidates for voting rules which are integer-linear-program implementable.
Dorn and Schlotter \cite{DBLP:journals/algorithmica/DornS12} proved that the swap bribery problem is {\fpt} with respect to the number of candidates for any voting system which is described by linear inequalities.
Betzler et al. \cite{DBLP:conf/ijcai/BetzlerHN09} proved that the possible winner problem is {\fpt} with respect to the number of candidates for Maximin, Copeland and Ranked pairs voting rules.
Elkind et al. \cite{DBLP:conf/atal/ElkindFS10} devised a general framework for classifying the fixed-parameter tractability of the winner determination problems for voting rules which are ``distance-rationalizable''. For parameterized complexity of strategic voting problems with respect to other parameters than the number of candidates, we refer to \cite{DBLP:conf/birthday/BetzlerBCN12} for a survey.
\section{Preliminaries}\label{prelim}
{\bf{Common Rules}}. We follow the terminology of the work of Xia and Conitzer \cite{DBLP:conf/ijcai/XiaC09}. Let $\mathcal{C} = \{c_1, . . . , c_m\}$ be a set of candidates. A {\it{linear order}} on $\mathcal{C}$ is a transitive,
antisymmetric, and total relation on $\mathcal{C}$. The set of all linear
orders on $\mathcal{C}$ is denoted by $L(\mathcal{C})$. An $n$-{\it{voter profile}} $P$ on $\mathcal{C}$ consists of $n$ votes defined by linear orders on $\mathcal{C}$. That is, $P = (V_1, . . . , V_n)$, where for every $i\leq n, V_i\in L(\mathcal{C})$. Each vote represents the preferences of the respective voter over the candidates. In particular, a candidate $c$ is ranked higher than another candidate $c'$ in a vote, if the voter prefers $c$ to $c'$. For convenience, we also use $\succ_i$ to denote a vote. Throughout this paper, we use the words vote and voter interchangeably. The set of all profiles
on $\mathcal{C}$ is denoted by $P(\mathcal{C})$. In the remainder of the paper, $m$
denotes the number of candidates and $n$ denotes the number
of voters. A {\it{(voting) rule}} is a function that maps a voting profile to a single candidate, the winner.
\begin{itemize}
\item Positional scoring rules. Every candidate gets a specific score from each vote according to the position of the candidate in the vote. More specifically, a scoring voting rule is defined by a scoring vector $\vec{\lambda}=(\lambda_1,\lambda_2,...,\lambda_m)$ with $\lambda_1\geq \lambda_2\geq,...,\geq \lambda_m$.
The candidate ranked in the $i$-th position in a vote gets $\lambda_i$ points from this vote. The winner is the candidate with the highest score.\footnote{If more than one candidate has the highest score, we break the tie by a fixed deterministic tie-breaking rule. This applies to all the other voting rules discussed in this paper.} Following are some well-known positional scoring rules.
\begin{tabular}{l|l}
voting rules & scoring vectors \\ \hline
Borda & $(m-1,m-2,...,0)$ \\
$r$-Approval & $(1,...,1,0,...,0)$ with exactly $r$ many 1's. \\
Plurality & $(1,0,0,...,0)$ \\
Veto & $(1,1,...,1,0)$ \\
\end{tabular}
\item Maximin. For two candidates $c$ and $c'$, let $N(c,c')$ denote the number of votes which prefer $c$ to $c'$. We say $c$ beats $c'$ if $N(c,c')>N(c',c)$. The maximin score of a candidate $c$
is defined as $\min_{c'\in \mathcal{C}\setminus \{c\}}N(c,c')$. The winner is the candidate with the highest maximin score.
\item Copeland$^\alpha$.
Each candidate is compared with every other candidate. In each comparison, the one which beats its rival gets one point and its rival gets zero points. If they are tied, both get $\alpha$ points. The winner is the candidate with the highest score.
\item Instant-runoff (STV): If a candidate is ranked in the first position by more than half of the votes, the candidate wins.
Otherwise, the candidate which is ranked in the first position by the least number of votes is eliminated.
This is repeated until there is a candidate which is ranked in the first position by more than half of the votes.
\end{itemize}
\bigskip
{\bf{Generalized Scoring Rules.}} In the following, we give the definition of the class of the generalized scoring rules which was introduced by Xia and Conitzer \cite{DBLP:conf/ijcai/XiaC09}.
Let $K = \{1, . . . , k\}$. For any $\vec{a},\vec{b} \in \mathbb{R}^k$, we say that $\vec{a}$ and $\vec{b}$ are
{\it{equivalent}} with respect to $K$, denoted by $\vec{a} \sim _K \vec{b}$, if for
any $i, j \in K, \vec{a}[i] > \vec{a}[j] \Leftrightarrow \vec{b}[i] > \vec{b}[j]$ and $\vec{a}[i] < \vec{a}[j] \Leftrightarrow \vec{b}[i] < \vec{b}[j]$ (where $\vec{a}[i]$ denotes the $i$-th component of the vector $\vec{a}$, etc.).
A function $g : \mathbb{R}^k \rightarrow \mathcal{C}$ is {\it{compatible}} with $K$ if
for any $\vec{a},\vec{b} \in \mathbb{R}^k, \vec{a} \sim _K \vec{b} \Rightarrow g(\vec{a})
= g(\vec{b})$.
Let $k \in \mathbb{N}, f : L(\mathcal{C})\rightarrow \mathbb{R}^k$, and $g : \mathbb{R}^k \rightarrow \mathcal{C}$ where $g$ is compatible with
$K$. The functions $f$ and $g$ determine the
{\it{generalized scoring rule}} $GS(f, g)$ as follows. For any profile of votes
$V_1, . . . , V_n \in L(\mathcal{C}), GS(f, g)(V_1, . . . , V_n) =g(\sum_{i=1}^n f(V_i))$. That is, every vote results in a vector of scores according to
$f$, and $g$ decides the winner based on comparisons between the total scores. Here we call $f$ the {\it{generalized scoring function}} and $g$ the {\it{decision function}}. Moreover, we say that $GS(f, g)$ is of {\it{order}} $k$ and $\sum_{i=1}^n f(V_i)$ is the {\it{total score vector}} of the profile according to $GS(f,g)$. For convenience, we also use $f(P)$ to denote $\sum_{i=1}^n f(V_i)$, where $P$ is the profile with the votes $V_1,...,V_n$.
Many common voting rules fall into the category of the generalized scoring rules. For example, for the Borda voting rule, the corresponding generalized scoring rule is specified as follows.
$k_{Borda} = m$.
$f_{Borda}(V) = (s(V, c_1), . . . , s(V, c_m))$, where $s(V,c_i)$ is the score of $c_i$ from the vote $V$.
$g_{Borda}(f_{Borda}(P)) =argmax_i (f_{Borda}(P))$, that is, the winner is the one with highest Borda score.
We point out that the class of generalized scoring rules also encapsulates many runoff voting rules, that is, voting rules where the winners are determined via several rounds. A typical example is the STV voting rule. Moreover, the definition of generalized scoring rules can be
generalized to voting rules selecting more than one winner \cite{DBLP:conf/sigecom/XiaC08a}. The following lemma summarizes the common voting rules known to fall into the category of generalized scoring rules.
\begin{lemma}\label{commonrulesfall}
\cite{procaccia13jair,DBLP:conf/ijcai/XiaC09} The following voting rules are generalized scoring rules: all the positional scoring rules, Copeland$^{\alpha}$, Maximin, STV, Baldwin's, Nanson's, Ranked pairs, Bucklin.
\end{lemma}
A common voting rule which is precluded by the class of generalized scoring rules is the Young's voting rule \cite{DBLP:conf/ijcai/XiaC09}. Goldsmith~et~al.~\cite{AAAI14a} recently studied a new class of voting
rules (rank-dependent scoring rules, RDSRs for short) and showed by an example that there are voting rules in this class which do not fall into the category of of the generalized scoring rules.
\bigskip
{\bf{Strategic Behaviors.}} We make use of the standard definitions of strategic behaviors in computational social choice. In the following, we briefly introduce the problems discussed in this paper. We refer to \cite{DBLP:journals/mlq/ErdelyiNR09,DBLP:journals/jair/FaliszewskiHHR09} for all the detailed definitions, including the manipulation, bribery and all the 22 standard control problems. In all these problems, we have as input a set $\mathcal{C}\cup \{p\}$ of candidates where $p$ is a distinguished candidate, and a profile $P=\{V_1,...,V_n\}$ of votes. The question is whether the distinguished candidate $p$ can become a winner (in this case, $p$ is not the winner in advance) or become a loser (in this case, $p$ is the winner in advance) by imposing a specific strategic behavior on the voting. The former case of making $p$ a winner is called a {\it{constructive}} strategic behavior, and the latter case is called a {\it{destructive}} strategic behavior. Observe that if the problem of a specific constructive strategic behavior is {\fpt} with respect to the number of candidates, so is the corresponding destructive case. To check this, suppose that we have an {\fpt} algorithm $Algo$ for a specific constructive strategic behavior problem. Then, we can guess a candidate $p'\in \mathcal{C}$ and run the algorithm $Algo$ but with the distinguished candidate being $p'$. Since we have at most $m$ guesses, the destructive case is solved in {\fpt}-time. Due to this fact, we consider only the problems of constructive strategic behaviors.
\medskip
{\bf{Manipulation.}}
In addition to the aforementioned input, we have a set $\mathcal{V}'$ of voters who did not cast their votes yet. We call these voters {\it{manipulators}}. The question is whether the manipulators can cast their votes in a way so that $p$ becomes a winner.
\medskip
{\bf{Bribery.}}
The bribery problem asks whether we can change at most $\kappa$ votes (in any way but still linear orders over the candidates) so that $p$ becomes a winner, where $\kappa\in \mathbb{N}$ is also a part of the input.
\medskip
{\bf{Control.}} There are 11 standard constructive control behaviors in total. Among them 7 are imposed on the candidate set and 4 are imposed on the vote set. We first discuss the candidate control cases. In these scenarios, we either add some candidates (limited or unlimited), or delete some candidates, or partition the candidate set into two sets (runoff or non-runoff partitions with ties-promote or ties-eliminate models). Since the number of the candidates is bounded by the parameter $m$, we can enumerate all the possibilities of performing the control strategic behaviors in {\fpt}-time with respect to $m$. Thus, if the winner is computable in {\fpt} time (which holds for all the common voting rules studied in this paper) with respect to $m$, the candidate control problems are {\fpt}. In the following, we restrict our attention to the vote control problems.
\smallskip
{\bf{Deleting votes:}} The problem of control by deleting votes asks whether we can remove at most $\kappa$ votes from the given profile so that $p$ becomes the winner, where $\kappa\in \mathbb{N}$ is also a part of the input.
\smallskip
{\bf{Partition votes:}} In the control by partitioning of votes, we are asked the following question: is there a partition of $P$ into $P_1$ and $P_2$ such that $p$ is the winner of the two-stage election where the winners of election $(\mathcal{C}\cup \{p\},P_1)$ compete
against the winners of $(\mathcal{C}\cup \{p\},P_2)$? We distinguish the ties-promote model and the ties-eliminate model. In the ties-promote model, all the candidates which are tied as winners in the first-stage election
are promoted to the second stage election. In the ties-eliminate model, if there is more than one winner, then all these winners will not be moved to the second stage election.
{\bf{Adding votes:}} In addition to the aforementioned input, we have another list $P'$ of {\it{unregistered votes}}, and are asked whether we can add at most $\kappa$ votes in $P'$ to $P$ so that the distinguished candidate $p$ becomes the winner.
\section{The General Framework}
In this section, we investigate the parameterized complexity of strategic behaviors under the class of generalized scoring rules. Our main result is summarized in the following theorem.
\begin{theorem}\label{thmframework}
Let $\varphi=(f,g)$ be a generalized scoring rule of order $k$, where $f$ is the generalized scoring function and $g$ is the decision function. If $k$ is bounded by a function of the number of candidates, and $f$ and $g$ are computable in \fpt-time with respect to the number of candidates, then the manipulation, bribery and all the 22 standard control problems are \fpt under $\varphi$, with respect to the number of candidates.
\end{theorem}
\begin{proof}
According to the discussion in Sec. \ref{prelim}, we can restrict our attention to the constructive strategic behaviors of manipulation, bribery, control by adding/delting/partition votes. We derive \fpt-algorithms for these problems. Our algorithms rely on the theorem by Lenstra \cite{lenstra83}, which implies that the integer linear programming (ILP) is {\fpt} with respect to the number of variables. Specifically, we reduce the instances of the stated problems to instances of ILP with the number of variables bounded by some function in $m$.
Let $\psi$ be the function in $m$ with $k\leq \psi({m})$. Due to the definition of the generalized scoring rules, we need focus on at most $3^{k\choose 2}$ different types of total score vectors (for each pair of subindices $i,j\in \{1,2,...,k\}$, we have either $\vec{a}[i]>\vec{a}[j]$ or $\vec{a}[i]=\vec{a}[j]$, or $\vec{a}[i]<\vec{a}[j]$). Here we say two vectors $\vec{a}, \vec{b}\in \mathbb{R}^k$ have the same type if they are equivalent with respect to $K=\{1,2,...,k\}$. Since the decision function $g$ is computable in \fpt-time with respect to $m$, we can enumerate all the types of total score vectors in the final election (that is, elections after performing strategic behaviors) which result in $p$ being the winner. Each enumerated total score vector $\vec{a}$ is specified by, for each pair of subindices $i,j$, either $\vec{a}[i]>\vec{a}[j]$ or $\vec{a}[i]=\vec{a}[j]$, or $\vec{a}[i]<\vec{a}[j]$. Then, we reduce the subinstances to ILP instances. To this end, we assign variables to different types of votes and derive restrictions to ensure that the currently enumerated total score vector coincides with the final election. If the given instance is a true-instance, then at least one of the total score vector leads to a correct answer. We fix $\vec{a}$ as the currently enumerated total scoring vector. In the following, we show how to reduce these instances to ILP instances.
\medskip
{\bf{Manipulation.}} Let $P=(V_1,V_2,...,V_n)$ be the profile of non-manipulators, and let $\vec{b}=f(P)$ be the total score vector of $P$. Clearly, $\vec{b}$ can be calculated in {\fpt} time since the generalized scoring function is computable in {\fpt} time.
To reduce the manipulation problem to ILP, we assign variables to all the $m!$ possible linear orders over the candidates, one for each. Let $x_{\succ}$ denote the variable assigned to the linear order $\succ$. These variables indicate how many manipulators cast their votes which are defined as $\succ$. Now we introduce the restrictions.
(1) Let $t$ be the number of manipulators, we have
\[\displaystyle{\sum_{\succ}x_{\succ}}=t\]
Here, $\succ$ runs through all the linear orders in $L(\mathcal{C}\cup \{p\})$.
(2) For convenience, for each linear order $\succ$, we use $f_{\succ}[i]$ instead of $f(\succ)[i]$ to denote the $i$-th entry of the score vector of $\succ$ by the generalized scoring function $f$. For each pair $i,j\in \{1,2,...,k\}$ with $\vec{a}[i]-\vec{a}[j]\rhd 0$, where $\rhd\in \{>,=,<\}$ we have
\[\vec{b}[i]+\displaystyle{\sum_{\succ}(f_{\succ}[i]\cdot x_{\succ})}-\vec{b}[j]-\displaystyle{\sum_{\succ}(f_{\succ}[j]\cdot x_{\succ})}\rhd 0\]
{\bf{Bribery.}}
We divide the votes into $P_{\succ_1},P_{\succ_2},...,P_{\succ_t}$ with $P_{\succ_i}$ containing all the votes defined as the linear order ${\succ_i}$.
For every two distinguished linear orders $\succ$ and $\succ'$, we assign a variable denoted by $x_{\succ}^{\succ'}$, which specifies how many voters from $P_{\succ}$ are bribed to recast their votes as ${\succ'}$. Clearly, we have at most $m!^2$ variables.
For each $\succ$, let $N(\succ)$ be the number of the votes which are defined as $\succ$ after changing the votes according to the variables assigned to the instance. More precisely, $N(\succ)$ is given by
\[N(\succ)=|P_{\succ}|-\displaystyle\sum_{\succ'\neq \succ}x_{\succ}^{\succ'}+\sum_{\succ'\neq \succ}x_{\succ'}^{\succ}\]
Now we introduce the restrictions. First, we have the following restriction since we can bribe at most $\kappa$ votes in total.
\[\displaystyle\sum_{\succ\neq \succ'} x_{\succ}^{\succ'} \leq \kappa\]
Here, $\succ$ and $\succ'$ with $\succ\neq \succ'$ run through all the linear orders over the candidates.
In addition, for each $P_{\succ}$, at most $|P_{\succ}|$ can be bribed. Hence, for each $P_{\succ}$, we have
\[\displaystyle\sum_{\succ'\neq \succ} x_{\succ}^{\succ'} \leq |P_{\succ}|\]
Finally,
for every~$i,j$~\text{with}~$\vec{a}[i]- \vec{a}[j]\rhd 0$, where $\rhd \in \{>,<,=\}$, we have
\[\displaystyle\sum_{\succ} N(\succ)\cdot f_{\succ}[i]-\displaystyle\sum_{\succ} N(\succ)\cdot f_{\succ}[j]\triangleright 0\]
{\bf{Control by Adding/Deleting Votes.}}
We first consider the adding votes case. We divide the unregistered votes into parts each containing all the votes defined as the same linear order.
For each part containing the votes defined as linear order $\succ$, we assign a variable $x_{\succ}$, which specifies how many votes from this part are included in the solution. Now we introduce the restrictions.
For each linear order $\succ$, let $N_1(\succ)$ be the number of registered votes defined as $\succ$, and $N_2(\succ)$ be the number of unregistered votes defined as $\succ$.
Since we can add at most $\kappa$ votes, we have the following restriction.
\[
\displaystyle\sum_{\succ}x_{\succ} \leq \kappa
\]
In addition, for each $\succ$, we have
\[x_{\succ}\leq N_2(\succ)\]
Finally, for every pair $i,j$ with $\vec{a}[i] -\vec{a}[j]\rhd 0$, where $\rhd \in \{>,<,=\}$, we have,
\[\displaystyle\sum_{\succ} (x_{\succ}+N_1(\succ))\cdot f_{\succ}[i]-\displaystyle\sum_{\succ} (x_{\succ}+N_1(\succ))\cdot f_{\succ}[j]\triangleright 0\]
The algorithm for the deleting votes case is analogous.
{\bf{Partition of Votes with Ties-Eliminate.}}
This case is slightly different from the above cases. First observe that $p$ has chance to be the final winner if $p$ is a temporary winner in at least one of the two subelections.
Therefor, to solve the problem, we enumerate all possible candidates $p'$ which will compete with $p$ in the second-stage election. We immediately discard the enumerations for which $p$ is not the winner when competing with $p'$. The above procedure clearly takes polynomial time and leads to polynomially many subinstances, each asking whether we can partition the profile into two parts $P_1$ and $P_2$ so that $p$ is the winner in the voting with profile $P_1$, and $p'$ is the winner in the voting with profile $P_2$. Therefore, instead of enumerating all the possible total score vectors as discussed for the above controls, we enumerate all the possible vector pairs $\vec{a},\vec{b}\in \mathbb{R}^k$, where $\vec{a}$ is the potential total score vector for the voting profile $P_1$, and $\vec{b}$ is the potential total score vector for voting profile $P_2$. We discard all the enumerations for which $p$ (resp. $p'$) is not the winner with respect to $\vec{a}$ (resp. $\vec{b}$). Now, we adopt the similar method as discussed above to reduce each subinstance to an ILP instance. To this end, again we partition the votes into parts each containing all the votes defined as the same linear order over the candidates. We still use $x_{\succ}$ to denote the variable assigned to the part $P_{\succ}$ of votes defined as $\succ$. Here, $x_{\succ}$ indicates how many votes in $P_{\succ}$ go to $P_1$.
The restrictions are as follows.
For each $\succ$, we have
\[x_{\succ}\leq |P_{\succ}|\]
For every pair $i,j$ with $\vec{a}[i]- \vec{a}[j]\rhd 0$, we have,
\[\displaystyle\sum_{\succ}(x_{\succ}\cdot f_{\succ}[i])-\displaystyle\sum_{\succ}(x_{\succ}\cdot f_{\succ}[j])\rhd 0\]
This equality is to ensure the that $p$ is the winning candidate in the voting with profile $P_1$. The following equality is to ensure that $p'$ is the winning candidate in the voting with profile $P_2$.
For every pair $i,j$ with $\vec{b}[i]- \vec{b}[j]\rhd 0$, where $\rhd \in \{>,<,=\}$, we have,
\[\displaystyle\sum_{\succ}(|P_{\succ}|-x_{\succ})\cdot f_{\succ}[i])-\displaystyle\sum_{\succ}(|P_{\succ}|-x_{\succ})\cdot f_{\succ}[j])\rhd 0\]
The proof for the ties-promote model is similar. However, in this case, we should adopt the multiwinner variant of the generalized scoring rules (which is possible \cite{DBLP:conf/sigecom/XiaC08a}) as a tool. Besides, instead of enumerating a candidate $p'$, we need to enumerate all the pairwise disjoint subsets $C_1$ and $C_2$ of the candidates, with $p\in C_1$. If $p$ is not the winner in the election restricted to $C_1\cup C_2$, we discard the enumeration at the moment. Moreover, the restrictions are derived to making all the candidates in $C_1$ the co-winners in the voting profile $P_1$, and $C_2$ the co-winners in the voting profile $P_2$. \end{proof}
To use the framework described in Theorem \ref{thmframework}, we require that the order of the generalized scoring rule must be bounded by a function of the number of the candidates, and the scoring function and the decision function must be computable in {\fpt} time with respect to the number of candidates. In the following, we show that both the conditions are fulfilled for all the voting rules stated in Lemma \ref{commonrulesfall}.
\begin{lemma}\label{commonrulesfpt1}
For all the positional scoring rules, Copeland$^{\alpha}$, Maximin, STV, Baldwin's, Nanson's, Ranked pairs, Bucklin, the orders of the corresponding generalized scoring rules are bounded by functions of the number of candidates, and the scoring and decision functions of the corresponding generalized scoring rule are computable in {\fpt}-time.
\end{lemma}
\begin{proof}
We refer to \cite{DBLP:conf/sigecom/XiaC08a} for all the positional scoring rules, STV, Maximin, Ranked pairs and Copeland$^{\alpha}$ (\cite{DBLP:conf/sigecom/XiaC08a} described only for $\alpha=0.5$. However, with slight modification, the arguments work for all the $0\leq \alpha\leq 1$). In the following, we prove for the Baldwin's, Nanson's and Bucklin, by describing in detail the specifications of the respective generalized scoring rules. Hereby, let $\mathcal{C}$ denote the set of candidates.
\medskip
\medskip
{\bf{Bucklin.}} The Bucklin score of a candidate $c$ is the smallest
number $x$ such that more than half of the votes rank $c$ among
the top $x$ candidates. The winner is the candidate that has
the smallest Bucklin score. The respective generalized scoring rule is as follows.
$k_{Bucklin}=m^2$; the components are indexed by pairs $(c,i)$ where $c$ is a candidate and $i$ is a positive integer with $1\leq i\leq m$.
The score vector of a vote $\succ$ is calculated in the following way.
\[f_{Bucklin}(\succ)_{c,i}=
\begin{cases}
1 & ~\text{if}~ c~\text{is ranked among the top}~i~\text{positions in}~\succ\\
0 & ~otherwise~
\end{cases}\]
The decision function $g_{Bucklin}$ works as follows. First, sum up all the score vectors of the votes. Let $\vec{a}$ be the total score vector. It is clear that the Bucklin score of a candidate $c$ is the minimum value $i$ for which $\vec{a}[c,i]> n/2$, where $n$ is the number of votes. Then, the winner is the one with the minimum Bucklin score.
\medskip
{\bf{Nanson's and Baldwin's.}} These two voting rules are multiround runoff rules, meaning that the winner is selected via rounds in each some candidates are removed from the election. Specifically, in the Nanson's voting, all the candidates with Borda score no greater than the average Borda score are
eliminated in each round. In the next round, the Borda scores of the remaining candidates are recomputed, as
if the eliminated candidates were not in the voting. This is repeated until there is a
final candidate left. The Baldwin's is similar to the Nanson's with difference that in each round the eliminated candidate is the one with least Borda score.
Xia and Conitzer \cite{DBLP:conf/sigecom/XiaC08a} proved that for any voting rule with finitely many runoff rounds, if in each step the rule used to rule out the eliminated candidates is also a generalized scoring rule, then the multiround runoff rule is a generalized scoring rule. Moreover, their (constructive) proof implies that if in each step the generalized scoring rule has {\fpt}-time computable functions $f'$ and $g'$, with respect to the number of candidates, the respective generalized scoring rule of the multiruound voting rule (with polynomially many rounds) also has a {\fpt}-time computable functions $f$ and $g$. We refer to Appendix 1 in \cite{DBLP:conf/sigecom/XiaC08a} for checking further details. Due to this fact, it is sufficient to show that in each round the procedure of selecting the eliminated candidates is a generalized scoring rule with bounded order, {\fpt}-time computable decision function $g$ and {\fpt}-time computable scoring function $f$.
We consider first for the Nanson's voting. Let $(c_1,c_2,...,c_m)$ be a arbitrary fixed order of the candidates.
$k = m$.
$f(V) = (s(V, c_1), . . . , s(V, c_m))$, where $s(V,c_i)$ is the Borda score of $c_i$ from the vote $V$.
The decision function $g$ selects the winner(s) as follows. Let $avg=\frac{1}{m}\cdot \sum_{i=1}^m f(P)[i]$. The winners (the candidates which are eliminated) are the candidates $c_i$ with $f(P)[i]< avg$.
The Baldwin's voting is similar with the difference that the candidates $c_i$ with minimum ${f(P)[i]}$ are eliminated.
\end{proof}
Based on Theorem \ref{thmframework} and Lemma \ref{commonrulesfpt1}, we have the following corollary.
\begin{corollary}
All the manipulation, bribery and the 22 standard control problems for the following voting rules are {\fpt} with respect to the number of candidates: all the positional scoring rules, Copeland$^{\alpha}$, Maximin, STV, Baldwin's, Nanson's, Ranked pairs and Bucklin.
\end{corollary}
\section{Discussion}
The class of generalized scoring rules was first introduced by Xia and Conitzer \cite{DBLP:conf/sigecom/XiaC08a} to investigate the frequency of coalitional manipulability. In this scenario, the focus is how the probability of a random profile being manipulable changes as the number of manipulators increases from 1 to infinite. The class of generalized scoring rules was also used in investigating the margin of victory in voting systems~\cite{DBLP:conf/sigecom/Xia12}. Moreover, Xia and Conitzer \cite{DBLP:conf/ijcai/XiaC09} characterized the class of generalized scoring rules as the class of voting rules that are anonymous and finitely locally consistent. A highly related class of voting rules is the class of hyperplane rules introduced by Mossel et al. \cite{procaccia13jair}. Mathematically,
the generalized scoring rules are equivalent to the hyperplane rules \cite{procaccia13jair}.
In this paper, we extend the application of generalized scoring rules by exploring the parameterized complexity of strategic voting problems. In particular, we show that from the viewpoint of parameterized complexity, the manipulation, bribery and control problems which are {\nph} in many voting systems turned out to be fixed-parameter tractable ({\fpt}), with respect to the number of candidates. The key point of our {\fpt} algorithms is the compatibility of the decision function $g$ in the generalized scoring rules, which enables us to enumerate all the desirable total score vectors in {\fpt} time.
To date, many strategic problems have been proved {\nph}. A challenging task would be to explore the connections between these {\nph}ness via the notion of the generalized scoring rules. For this purpose, a deeper exploitation of the functions $f$ and $g$ is needed. Besides, there are also many other
problems in computational social choice that pertain to conducting strategic
voting, some of which were introduced quite recently \cite{DBLP:conf/atal/FaliszewskiHH13,DBLP:conf/atal/PerekFPR13}. Exploring the parameterized complexity of these newly proposed voting problems via the framework of generalized scoring rules is also an interesting topic.
{\bf{Note.}} Similar results of this paper were independently announced by Xia~\cite{lirongxiaaamas2014}. However, there are several differences. First, our results apply to all the 22 standard control problems, while the results in~\cite{lirongxiaaamas2014} does not include the control by partition votes. Second, Xia studied the winner determination problem which is not discussed in this paper.
|
\section{Introduction}
\label{s:intro}
\vspace{-2mm}
The booming growth in popularity of the cellular communications and the exponential rise of cellular data traffic pushed the technology manufactures to their limits in such a way that they could not keep pace with the current demand growth in mobile user's applications~\cite{Bhushan2014Commag}. This made the cellular network industry open to new proposals more than ever. Among various proposals to ameliorate the cellular capacity shortcoming, Device-to-Device (D2D) communication stood out because it detected the paradigm shift in cellular data flow~\cite{asadi2014survey}. The cellular communication ends used to be distant a decade ago, while the emergence of new mobile applications into people's life (e.g., social networking) created significant traffic among nearby users. The literature on D2D communication is abundant. In fact both academia and industry have been actively exploring use-cases and techniques of D2D communications~\cite{asadi2014survey}.
Academia proposes a wide range of use-cases for D2D communications such as relay~\cite{asadi2014dronee}, multicasting~\cite{zhou_intracluster_2013}, and cellular offloading~\cite{bao_dataspotting_2010}. Initial D2D proposals focused on D2D communication underlaying cellular network transmissions, i.e., using the same spectral resources used for cellular communications~\cite{Doppler2009Commag}. Later, other D2D techniques have been proposed, which either fall under either {\it inband} or {\it outband} D2D communication. Inband D2D communications allow D2D users to communicate over the cellular spectrum, while outband schemes demands the D2D users to access unlicensed bands for D2D transmissions~\cite{asadi2014survey}.
Each of these D2D operational {\it modes} poses its own merits and disadvantages in terms of interference management, implementation complexity, achievable spectral efficiency, and therefore in terms of performance guarantees. However, the available literature proposes solutions for efficiently implementing each mode in isolation, i.e., {\it mode selection} has not been addressed. Nevertheless, according to the definition provided by 3GPP standards, ``D2D communication is the communication between two users in proximity using a direct link between the devices in order to bypass the eNB(s)\footnote{eNB is the 3GPP term referring to cellular base stations.} or core network''~\cite{lin2013arxiv}. Therefore, any of these modes or perhaps all shall be used for D2D communications. Moreover, promising studies on D2D communication moved industry leaders such as Qualcomm to invest on future implementation of D2D communications, and 3GPP is considering to include generic D2D support in the next release of LTE-A standard as a public safety feature~\cite{lin2013arxiv}.
In such a framework, we believe that different D2D modes should not be treated as competitors but as complementary techniques. Co-existing D2D modes can immensely increase the system complexity because there should exist a mechanism to select the correct D2D mode according the overall system conditions.
\section{Acknowledgements}
\bibliographystyle{abbrv}
\renewcommand{\baselinestretch}{0.96}
\scriptsize
\section{Problem Formulation}
We solve the problem hierarchically at the beginning of each mode interval $ j $, i.e., each $T$ seconds. Let $ \setL ( j ) $ be the set of all existing arcs
during mode interval $ j $, i.e., such that $ Z_{ n , m } ( j ) = 1 $.
For an active arc $ ( n , m ) $ under an LTE mode $ i \in \{0, 1, 2 \} $ in mode interval $ j $ we define the energy consumption $ E^{ i }_{ n , m } ( j ) $
and the transferred data $ \theta^{ i }_{ n , m } ( j ) $ (both per mode interval $T$) as follows:
\vspace{-2mm}
\begin{eqnarray}
E^{ i }_{ n , m } ( j ) & =& \left( p^{ i, \text{TX} }_{ n } + p^{ i, \text{RX} }_{ m } \right) B^{ i }_{ n , m } ( j ), \label{eq:einband} \\
\theta^{ i }_{ n , m } ( j ) & =& B^{ i }_{ n , m } ( j ) R^{ i, \text{CQI} }_{ n , m } ( j ), \label{eq:tinband2}
\end{eqnarray}
where
we do not consider the baseline energy consumed by a user in LTE in one mode interval, since it cannot be changed unless the node is switched off,
$ p^{ i, \text{TX} }_{ n } $ and $ p^{ i, \text{RX} }_{ m } $ are the energy consumed by user $ m $ per transmitted and received RB, respectively, $ B^{ i }_{ n , m } ( j ) $ is the number of RBs allocated to arc $ ( n , m ) $, and $ R^{ i, \text{CQI} }_{ n , m } ( j ) $ is the
number of transmitted bits per RB of arc $ (n , m) $ under mode $ i $ during mode interval $ j $.
For an active arc $ ( n , m ) $ under mode $ 3 $ (i.e., WiFi) in mode interval $ j $ we define the energy consumption $ E^{ 3 }_{ n , m } ( j ) $
and the throughput $ \theta^{ 3 }_{ n , m } ( j ) $ (both per mode interval) as follows:
\vspace{-1mm}
\begin{eqnarray}
\label{eq:pWifi}
E^{ 3 }_{ n , m } ( j ) &=& 2 \beta^{ \texttt{WiFi} } + \left( p^{ 3, \text{TX} }_{ n } + p^{ 3, \text{RX} }_{ m } \right) \theta^{ 3 }_{ n , m } ( j ), \\
\theta^{ 3 }_{ n , m } ( j ) &=& T \cdot R^{ i, \text{CQI} }_{ n , m } ( j ),
\end{eqnarray}
where $ \beta^{ \texttt{WiFi} } $ is the baseline WiFi energy consumed by a user in one mode interval, and $R^{ i, \text{CQI} }_{ n , m }$ is the WiFi rate. Note that the energy consumption as defined here can incorporate both the consumption
due to transmission/reception and packet processing (see \cite{asadi2014dronee}).
The utility function for an active arc $ ( n , m ) $ under mode $ i $ in mode interval $ j $ is defined as follows:
\begin{equation}
\label{eq:u}
U_{ n , m }^i ( j ) = \theta^{ i }_{ n , m } ( j ) - \alpha E^{ i }_{ n , m } ( j ),
\end{equation}
where $ \alpha $ is a relative cost of energy.
We use a set of binary decision variables $\{Y^i_{n,m}(j)\}$,
to formulate the problem of mode selection for mode interval $j$, preceding the RB allocation procedure in the above described system (note that at mode selection time
it is not yet possible to predict the exact interference caused by/to D2D users, so we account for the worst-case interference). The problem is formulated as follows (we omit the dependency on $ j $ from utilities, interferences, and decision variables):
\vspace{-2mm}
\begin{align*}
\text{maximize} & \sum_{i=0}^{3} \sum_{(n, m)\in\setL( j )} U_{ n, m }^{ i } Y_{ n, m }^{ i }; \\
\text{s.t.:} \quad & \sum_{i=0}^{3} \sum_{n|(n, m)\in \setL(j)} Y_{ n , m }^{ i } \le 1 \quad \forall m \in \setN; \\
& \sum_{i=0}^{3} \sum_{m|(n, m)\in \setL(j)} Y_{ n , m }^{ i } \le 1 \quad \forall n \in \setN; \\
& \sum_{\substack {(n, m)\in \setL(j) }} Y_{ n , m }^{ 1 } I_{ n,x} \le \gamma \quad \forall x\in \setS_{N+1} \cup \{N+1\};\\
& \sum_{i \in \{0, 1\}} \sum_{(x, y) \in \setL(j) \setminus \{ (n, m) \} } Y_{ x , y }^{ i } Y_{n, m}^ {1} I_{ x, m} \le \gamma \quad \forall (n, m) \in \setL(j);\\
& \sum_{(x, y) \in \setL(j) \setminus \{ (n, m) \} } Y_{ x , y }^{ 2 } Y_{n, m}^ {2} I_{ x, m} \le \gamma \quad \forall (n, m) \in \setL(j);
\end{align*}
The formulated problem maximizes the sum of utilities over all possible combinations of users and modes.
The first and second constraints ensure that at most one active connection can be allowed for each user (but for the eNB, which is labeled as $N+1$).
The third constraint imposes that the interference caused by inband underlay D2D users to cellular users and to the eNB is below a threshold $\gamma$.
The fourth constraint ensures that the interference caused by cellular and inband underlay transmissions to other inband underlay users is below a threshold.
Finally, the fifth constraint ensures that the interference caused by inband overlay transmissions (mode $2$) is below the threshold $\gamma$.
The challenge to be tackled in future work consists in plugging the resource allocation scheme into the computation of $\theta^{ i }_{ n , m }$ and $E^{ i }_{ n , m }$, which, in turn, depend on mode selection and connection activation decisions through the resource allocation scheme.
\vspace{-4mm}
\section{System model}
Our system consists of $ N $ users labelled as $ n \in \setN := \{ 1, 2, \dots, N \} $ in a single-cell LTE network with $20$MHz bandwidth eNB. For notational
consistency, the eNB is labelled as $ N + 1 $. Downlink/uplink channels are open separated bands (i.e., using an FDD scheme).
Each LTE {\it subframe} ($1$ms) the eNB has 100 time-frequency Resource Blocks (RB)s for downlink and uplink transmission~\cite{johnson2010lte}.
Users may communicate with other users in the cell or with those outside the cell.
If a user wants to communicate with another user that is physically close to her, she can use D2D communication. We call such a pair of users a D2D pair.
We assume that each user wants to communicate only with (at most) one user at any given time.
\textbf{User states.}
The users are allowed to move, and therefore their availability for communication can change over time, so we will say that each user is in a particular
\textit{state} which can change over time. We will denote the state of user $ n \in \setN $ at time $ t $ by $ X_{ n } ( t ) \in \{ 0, 1, 2, \dots, N + 1 \} $,
where each state can be categorized in one of the following types:
\begin{itemize}
\vspace{-2mm}
\itemsep -0.7 mm
\item \textbf{Dormant user (state $ 0 $)}: this is a user who either $ (i) $ has no data to transceive, or $ (ii) $ has a
poor channel quality in which communication is not feasible.
\item \textbf{Cellular user (state $ N + 1 $)}: this is a user who wants to communicate, and can only communicate with the eNB, labelled as $ N + 1 $;
\item \textbf{D2D user (states $ 1 \le m \le N $)}: this is a user who wants and can communicate with her D2D pair labelled $ m $ directly (i.e., she is
in D2D reach of the user with whom she wants to communicate).
\vspace{-2mm}
\end{itemize}
Consequently, the number of users in each state will vary in time. However, we assume that state changes occur
(or are detected by the mode selection mechanism) at regular {\it mode intervals} of duration $T$ seconds.
We denote by $\setS_m (j)$ the set of users in state $m \in \{0, .. , N+1\}$ in mode interval $j$.
Each cellular user and D2D pair is associated with a flow and each pair can only have one active flow at any given time. Moreover, the D2D communication is
assumed to be symmetric, i.e., if user $ n \in \setN $ is in state $ m \in \setN $, then user $ m $ is in state $ n $.
\begin{figure} [!t]
\centering
\includegraphics[scale=0.40]{./Network.eps}
\vspace{-3mm}
\caption{An illustration of a cell with dormant, cellular, and D2D users}
\label{fig:network}
\vspace{-5mm}
\end{figure}
\textbf{Graph model.}
We can map the network with $ N $ users and one eNB to a graph with $ N + 1 $ nodes, where nodes $ 1 $ to $ N $ represent the users and node $ N + 1 $
represents the base station. The location of nodes in the graph does not necessarily correspond to a physical position of the users (which are moreover allowed
to move within a cell). The users' physical location and mobility affect the arcs of the graph rather than the nodes. An arc between two nodes represents
the communication feasibility between the two nodes. Thus, at every given time, there is an arc between two nodes if these two nodes want to communicate and
their physical channel allows a non-zero transmission rate. Thus, dormant users are isolated (without any arc), cellular users have an arc with eNB, and D2D
users have an arc with their pairs and with eNB. In particular, if a user $ n $ is in state $ m \in \setN $, then there is an arc between users $ n $ and $ m $
and another one between user $ n $ and eNB $ N + 1 $; if a user $ n $ is in state $ N + 1 $, then there is an arc between user $ n $ and eNB $ N + 1 $. Thus,
the state of the user indicates her neighbour(s). See \autoref{fig:network} for an illustration. Due to users mobility and communication needs, which affect
users' states, the arcs will change over time (which is fully captured by state changes each $T$ seconds).
Note that there are at most $ 3N/2 $ arcs in the graph, because each cellular user creates $ 1 $ arc and each D2D pair create $ 3 $ arcs. The arcs will be
denoted by their end-nodes, $ ( n, m ) $. We will further denote the existence of arc $ ( n, m ) $ at time $ t $ by $ Z_{ n, m } ( t ) \in \{ 0, 1 \} $.
\textbf{Cellular mode.} Users in state $N + 1$ use normal cellular communication. We define this as mode $ 0 $.
\textbf{D2D modes.}
Every D2D pair can communicate via any of the following modes (see \autoref{fig:in-outband}):
\begin{itemize}
\vspace{-2mm}
\itemsep -0.7 mm
\item \emph{Underlay inband} (mode $1$): D2D users reuse the RBs which are available to the cellular users (and therefore share resources with connections in mode $0$).
\item \emph{Overlay inband} (mode $2$): D2D communications occur over dedicated RBs, subtracted from cellular users.
\item \emph{Outband} (mode $3$): D2D users switch to WiFi.
\vspace{-2mm}
\end{itemize}
In both underlay and overlay modes, D2D pairs can use the same RBs used by other D2D pairs simultaneously as long as interference allows. \autoref{tb:adv-disadv} summarizes the merits and drawbacks of each method. Note that the major issue in inband is interference control, while outband D2D suffers from the power consumption of WiFi interface.
\begin{figure} [!t]
\centering
\includegraphics[scale=0.23]{./in-outband.eps}
\vspace{-3mm}
\caption{Schematic representation of overlay inband, underlay inband, and outband D2D.}
\label{fig:in-outband}
\vspace{-5mm}
\end{figure}
\begin{table}[t!]
\scriptsize
\centering
\caption{Cons and pros of each D2D mode}
\label{tb:adv-disadv}
\resizebox{1\columnwidth}{!}{
\begin{tabular}{|p{3.5cm}|cccc|}
\cline{2-5}
\cline{2-5}
\multicolumn{1}{c|}{} &Underlay &Overlay &Cellular &WiFi \\
\hline
\hline
Interference between D2D and cellular users &$\checkmark$ &$\times$ &$\times$ &$\times$ \\
\hline
Interference among D2D users &$\checkmark$ &$\checkmark$ &$\times$ &$\times$ \\
\hline
Requires dedicated resources for D2D users &$\times$ &$\checkmark$ &$\times$ &$\times$ \\
\hline
Controlled interference environment &$\checkmark$ &$\checkmark$ &$\checkmark$ &$\times$ \\
\hline
Simultaneous D2D and cellular transmission &$\times$ &$\times$ &$\times$ &$\checkmark$ \\
\hline
Increased spectral efficiency &$\checkmark$ &$\checkmark$ &$\times$ &$\checkmark$ \\
\hline
Energy cost &Eq.\eqref{eq:einband} &Eq.\eqref{eq:einband} &Eq.\eqref{eq:einband} &Eq.\eqref{eq:pWifi} \\
\hline
\end{tabular}
}
\vspace{-5mm}
\end{table}
\vspace{-2mm}
\subsection{Joint scheduling and mode selection}
At a given time, every existing arc represents a possible data transmission, and can be either active (allowed to transmit) or inactive (not allowed to
transmit). We have to design a mechanism that selects the arcs to be used in each mode interval, and assign RBs to the arcs.
There are three tiers of decision making in our system:
\begin{itemize}
\vspace{-2mm}
\itemsep -0.7 mm
\item \textbf{Mode selection}: we have to decide about the operating mode for D2D pairs (modes $1$ to $3$);
\item \textbf{Connection activation}: we have to decide which connections (arcs) are active given the interference constraints of the selected mode;
\item \textbf{Connection scheduling}: we have to decide which connections transmit at what transmission rate (i.e., how the RBs are allocated).
\vspace{-2mm}
\end{itemize}
The three tiers are intertwined, since the interference depends on mode selection but cannot be known before connection activation and scheduling. In turn, connection activation and scheduling depend on which connection is active and on which mode is used in each connection. For sake of tractability, we implement first mode selection, assuming a worst-case interference scenario for activating connections, and then we implement a conventional opportunistic cellular scheduler, Proportional Fair (PF) for connection scheduling at eNB. Specifically, scheduling priorities are computed on
instantaneous or expected instantaneous channel quality of the users, and RBs allocated to inband overlay are fixed (they are used by users in mode $2$, or released for modes $0$ and $1$ if no connection selects mode $2$).
In each {\it subframe} (lasting $1$ ms), only one user is scheduled for direct communication to the eNB, while the number of concurrent D2D transmissions is not limited {\it a priori}. Therefore, mode $0$ users do not interfere with each other, mode $1$ users interfere with users in modes $0$ and $1$, and mode $2$ only causes interference among users in
mode $2$.
Our system operates in discrete time units and there is a central \textit{controller}, who schedules all the transmissions.
For tractability, we build the model hierarchically. The controller observes the actual CQI (the LTE Channel Quality Indicator, which corresponds to a particular transmission rate)
of each connection and takes the fundamental scheduling decisions every \textit{frame} (consisting of 10 subframes, hence lasting $10$ ms). All scheduled transmissions in
each subframe occur simultaneously and use the maximal transmission rate permitted by the CQI observed by the eNB. A connection scheduled in a subframe will use all the RBs assigned to the specific mode selected.
The controller further estimates at the beginning of every \textit{mode interval} $T$ the future CQI of all possible connections (both
WiFi and cellular), and decides upon the mode for the duration of the mode interval, which may imply setting up new connections or closing existing ones,
i.e., changes the arcs of the graph. From the graphical point of view, there is hence a new random graph (on a fixed number of nodes $ N $) at the beginning of
every mode interval $T$. The controller also decides which of the arcs are active (allowed to transmit) over the mode interval. In practice, the random
graphs will be strongly correlated, because the mode interval length has to be short enough (say, $200$ frames) to prevent users to move and experience deep channel fading.
\vspace{-2mm}
\subsection{CQI and interference estimation}
As mentioned above, CQI information is needed for each connection. We assume that the eNB can estimate the CQI of each connection by using the reports produced by the users, containing the signal strength they receive from each and all their neighboring transmitters. By extending this legacy LTE scheme, the interference can be estimated as well. Thus, the eNB can build the interference table, whose elements $I_{n,m}(j)\ge0$ represent the interference caused by user $ n $ to user $m$ ($\forall n,m \in \setN \cup \{N+1\}$), in mode interval $j$. Hence, decisions upon connections (set up/close) can be made on the observations from previous mode intervals.
|
\section{Introduction}
Fragmenting collisions are important in a range of astrophysical systems. While the slope of the fragment size distribution and the size of the largest fragment are well characterized and can be used confidently in models, the smallest fragment size is less well understood and is usually assumed to be constant for all collisions. We provide a framework for self-consistently calculating the smallest fragment size as a function of material and collision parameters (Section~\ref{sec:theory}), and discuss its implications for modelling debris disks (Section~ \ref{sec:applications}).
Numerous experimental studies have looked at the fragment size distribution of destructive collisions, focussing on the slope of the power law(s), and on the size of the largest fragment \citep{davis1990,ryan1991,nakamura1991,ryan2000}. The smaller end of the size distribution has received considerably less attention; the smallest fragments are hard to count experimentally, and require a very high resolution to be captured in numerical simulations. Fragment distributions are therefore incomplete below sizes of $100$~$\mu$m, or masses below $10^{-3}$~gr \citep{fujiwara1977,takagi1984}. Molecular dynamics \citep[e.g.][]{dominiktielens1997} or smooth particle hydrodynamics \citep{geretshauser2010} simulations have limited resolution and tend to focus on the fragmentation threshold velocity rather than the smallest fragments.
\section{Minimum fragment size in a single collision}\label{sec:theory}
We consider collisions below the hypervelocity regime, i.e. the relative velocity of the colliders is much smaller than their internal sound speed, generally implying $v_{\rm rel}\lesssim1~$km/s. Based on experiments, we adopt the standard fragment size distribution
\begin{equation}
n(s) = C\cdot s^{-\alpha},
\end{equation}
with $3 < \alpha < 4$, and $C$ a coefficient we express below. While the mass is dominated by the largest particles, the surface area and thus the surface energy is dominated by the smallest fragments. As the creation of infinitely small fragments would require an infinite amount of energy, while the amount of kinetic energy available in a collision is finite, the power law must stop or flatten at some small fragment size. To the best of our knowledge, however, the regime of fragment sizes relevant for the analysis below has not yet been probed by available experimental data nor described theoretically in an astrophysical context.
Assuming spherical fragments with sizes between $s_{\rm min}$ and $s_{\rm max} (\gg s_{\rm min})$, the total fragment mass and surface area are
\begin{equation}\label{eq:M_frag}
M_{\rm frag} = \frac{4\pi \rho C}{3(4-\alpha)} s_{\rm max}^{4-\alpha}, ~~~~~~
A_{\rm frag} = \frac{4\pi C}{\alpha-3} s_{\rm min}^{3-\alpha}.
\end{equation}
For a collision between two bodies of size $s_0$ and mass $M_0=(4\pi /3)\rho s_0^3$, mass conservation implies $M_{\rm frag}=2M_0$, and thus
\begin{equation}\label{eq:C}
C = 2(4-\alpha)s_0^3 s_{\rm max}^{\alpha-4}.
\end{equation}
The pre-collision kinetic energy is simply $U_{\rm K} = (1/2) \mu v_{\rm rel}^2$, where $\mu=M_0/2$ denotes the reduced mass. The difference in surface energy before and after the collision equals $\Delta U_{\rm S}=\gamma(A_{\rm frag} - 8\pi s_0^2)$, where $\gamma$ equals the surface energy per unit surface of the material. Assuming that only a fraction $\eta$ of the kinetic energy is used for creating new surface, we can combine Eqs.~\ref{eq:M_frag} and \ref{eq:C} to obtain a lower limit for the smallest fragment size. For the specific case of $\alpha=3.5$, this reduces to
\begin{equation}\label{eq:s_min3.5}
s_{\rm min} = \left(\frac{24 \gamma s_0}{\eta \rho s_0 v_{\rm rel}^2 + 24\gamma}\right)^2 s_{\rm max}^{-1},
\end{equation}
and gives the size of the smallest fragments created in a collision at $v_{\rm rel}$, assuming $\alpha=3.5$ and $s_{\rm max} \gg s_{\rm min}$.
Instead of forming a fragment distribution, we imagine the limiting case in which the kinetic energy just suffices to split both colliders in half\footnote{One could imagine splitting only one of the colliders, or indeed chipping off only small parts of one of the collider bodies. Since less surface area is created, this would still be allowed at very low velocities. However, in that case the largest fragment is of the same size as $s_0$. We refrain from identifying this as fragmentation, and use the size derived in Eq.~\ref{eq:s_split} as the size below which fragmentation becomes inefficient.}, i.e. $\eta U_{\rm K} = 2\pi s_0^2 \gamma$. Solving for $s_0$, we obtain
\begin{equation}\label{eq:s_split}
s_0^{\rm split} = \frac{3 \gamma}{\eta \rho v_{\rm rel}^2},
\end{equation}
which is the smallest particle that can be split in half. The smallest fragment is slightly smaller, but does not have a rigorously defined radius because we assume spherical particles. Equation~\ref{eq:s_split} is similar to the result of \citet{biermann1980} if $\eta=1$.
The same limit can be explored using Eq.~\ref{eq:s_min3.5}, by forcing $s_{\rm min} \sim s_{\rm max} \sim 2^{-1/3}s_0$. This results in
\begin{equation}
\label{eq:biermann}
s_{\rm min} \simeq \frac{5 \gamma}{\eta \rho v_{\rm rel}^2},
\end{equation}
which is very similar to Eq.~\ref{eq:s_split}. To summarise, in an energetic collision in which many fragments are created, the size of the smallest fragment is given by Eq.~\ref{eq:s_min3.5}. When the relative velocity is decreased, the fragment distribution becomes more and more discrete, until we reach the limit described by Eq.~\ref{eq:biermann}, in which particles can only just be split into two.
Figure \ref{fig:fig1} shows the minimum size from Eq.~\ref{eq:s_min3.5} as a function of collider size, assuming $v_{\rm rel}=20$~m/s, $\eta=10^{-2}$, and a maximum fragment that carries half of the initial collider mass. Gravity is important for bodies larger than $100$~m (see below). Smaller bodies are weaker, and can produce fragments down to the $s_{\rm min}$ indicated by the solid curve. For example, SiO$_2$ fragments smaller than a micron can, at this velocity, \emph{only} be formed by collisions of bodies larger than a few centimetres. The shaded region, top left, is forbidden, as there $M_{\rm frag} > 2M_0$ and mass is not conserved. Close to $s_{\rm min} \sim s_{0}$, the solid curves are non-linear as the pre- and post-collision surface areas become comparable.
\begin{table}
\caption{Material properties for silicate and ice used in this work. The values for the typical aggregate are explained in Appendix \ref{sec:A}.}
\label{tab:materials}
\centering
\begin{tabular}{l c c}
\hline\hline
Material & $\rho$ (\ensuremath{\rm g\,cm^{-3}}) & $\gamma$ (\ensuremath{\rm J\,m^{-2}}) \\
\hline
Silicate & 2.6 & 0.05 \\
Ice & 1.0 & 0.74 \\
Aggregate & $\sim10^{-1}$ & $\sim10^{-4}-10^{-3}$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}[!ht]
\includegraphics[clip=,width=.95\linewidth]{fig1.pdf}
\caption{\emph{Top:} Minimum fragment size $s_{\rm min}$ in a 20~m/s collision (Eq.~\ref{eq:s_min3.5}) for three materials, as a function of collider size $s_0$. We have assumed $\eta=10^{-2}$ and $s_{\rm max}=2^{-1/3}s_0$. The grey shaded areas are excluded because of mass conservation (left), and self-gravity (right). The dotted lines indicate the limit of Eq.~\ref{eq:biermann}. \emph{Bottom:} Critical energies versus size for equal-sized collisions. The energy needed to split a particle (e.g. Eq.~\ref{eq:biermann}) is shown for $\eta=10^{-3}$ (dotted) and $\eta=1$ (dashed). The solid curves correspond to catastrophic fragmentation of aggregates \citep{beitz2011}, and ice and basalt \citep{benz1999}, showing both the strength-dominated (small sizes) and self-gravity dominated regimes (large sizes).}
\label{fig:fig1}
\end{figure}
It is interesting to compare Eq.~\ref{eq:biermann} with the traditional form of the catastrophic fragmentation threshold velocity in equal-sized collisions: $v_f^2 = 8Q^*$. The critical energy, $Q^*$, has units of $\rm{erg/g}$, and varies with particle mass. For small bodies, the strength is dominated by cohesion, and for large ones by gravity \citep{benz1999}. For solid bodies, this transition occurs around 100 metres in size. Values of $Q^*\sim10^7 \rm{~erg/g}$ are often taken as typical for asteroids, and experimentally obtained values for small grains (mm to cm sizes) can be several orders of magnitude smaller \citep{blummunch1993, beitz2011}. Figure \ref{fig:fig1} shows the critical energy for splitting predicted by Eq.~\ref{eq:biermann} as a function of size for the materials in Table \ref{tab:materials} assuming $\eta=10^{-3}$ and $\eta=1$. The solid lines indicate critical fragmentation energies for basalt and ice \citep{benz1999} and silicate aggregates \citep{beitz2011}. The critical fragmentation energies exceed the splitting energy, indicating that substantially more energy is required to destroy -- rather than split -- colliders. The values plotted for ice and basalt were obtained at collision velocities of 3~km/s, substantially higher than the velocities considered here, and $Q^*$ is known to depend on velocity \citep{Leinhardt2012}. While such a velocity dependence appears absent in the splitting energy, it might be implicitly included in $\eta$. In fact, $\eta$ is expected to vary with material and impact energy. We adopt a constant value of $\eta=10^{-2}$.
Appendix \ref{sec:B} investigates similar limits for colliders with different mass ratios, and shows that collisions with a mass ratio close to unity are the most effective at creating small fragments.
\section{Application to debris disks}\label{sec:applications}
Debris disks are leftovers of planet formation, and are usually described by a birth-ring of km-sized asteroid-like particles orbiting their parent star, together with a population of smaller bodies formed in a collisional cascade \citep[for a recent review, see][]{matthews2014}. A steady-state and scale-independent population of bodies will follow a size distribution given by a power law with $\alpha=3.5$ \citep{dohnanyi1969}. Some variation in $\alpha$ has been found in different simulations. \citet{PanSchlichting2012} find up to $\alpha=4$ for cohesion-dominated collisional particles, and up to $\alpha=3.26$ for gravity-dominated ones.
Models of debris disks most often assume a smallest fragment size equal to the blow-out size, $s_{\rm blow}$ \citep{WyattDent2002, Wyattetal2010}, or some constant, but arbitrary, $s_{\rm min} < s_{\rm blow}$ for all collisions \citep{Thebaultetal2003, Krivovetal2008}. The blow-out size corresponds to particles with $\beta=1/2$, where $\beta~=~1.15 Q_{\rm pr} (L_{\star}/L_{\odot}) (M_{\star}/M_{\odot})^{-1} (\rho/\mathrm{g~cm^{-3}})^{-1}(s/\mathrm{\mu m})^{-1}$ is the ratio of the radiation and gravitational force. Particles with $\beta>1/2$ are removed from the system by radiation pressure. Alternatively, \citet{Gasparetal2012a} calculate a collision-dependent $s_{\rm min}$ from mass conservation, but do not study the surface energy.
If, however, for any relevant collision Eq.~\ref{eq:s_min3.5} predicts $s_{\rm min}>s_{\rm blow}$, extrapolating the fragment size distribution down to these sizes is not justified. For example, starting from Eq.~2 of \citet{Krivovetal2008}, cm-sized bodies have $Q^*\simeq5\times10^6$ erg/g. In the Krivov et al. framework, a collision between two such bodies at 70~m/s will then result in fragmentation, as the kinetic energy ($\simeq 12$ J, assuming $\rho=2.35\rm{~g/cm^3}$) slightly exceeds the critical energy ($=2mQ^* \simeq 10$ J), and fragments will be created from a size comparable to the impactor \citep[Eq.~21 of][]{Krivovetal2006} down to the blow-out size. For this particular collision, Eq.~\ref{eq:s_min3.5} yields $s_{\rm min} \ll s_{\rm blow}$ for $\eta=0.1$ and $\gamma=0.1\rm{~J/m^2}$, but $s_{\rm min} \simeq 6\rm{~\mu m}$ for $\eta = 10^{-3}$. Thus, the difference between our results and the fragment sizes of Krivov et al. may be substantial, depending on the true value of $\eta$. We stress that our theory is valid for $3>\alpha>4$, and does not apply to models that use shallower power laws, for example Section 4.2 of \citet{Krivovetal2013}.
The importance of the limit given by Eq.~\ref{eq:s_min3.5} depends on the parameters, and can vary per individual collision, depending on the collision velocity and choice for $s_{\rm max}$. In the rest of this section, we explore in which cases this limit is most relevant.
In a debris disk, a particle of size $s_0$ is most likely formed in a collision between only slightly larger particles. In addition, we focus on collisions between equal-sized particles, as these are most efficient at forming small fragments (Appendix \ref{sec:B}). Therefore, we use Eq.~\ref{eq:biermann} as an indication for the lower limit of the particle size distribution. Quantitative comparisons require relative collision velocities, which for the largest bodies are often written in terms of the Keplerian orbital velocity at the corresponding distance from the central object. For bodies on orbits with identical semi-major axes, the relative velocity can then be written in terms of orbital eccentricity and inclination as $ f \equiv v_{\rm rel} / v_{\rm K} = (1.25 e^2 + i^2)^{1/2}$, with $v_{\rm K}=(GM_{\star} / a)^{1/2}$ the Keplerian orbital velocity \citep{WyattDent2002}. In a debris disk, a range of eccentricities and inclinations will be present. For a rough comparison, we use average quantities $\langle e \rangle$ and $\langle i \rangle$ to obtain typical collision velocities. In reality, $\langle e \rangle$ and $\langle i \rangle$ are poorly constrained. Estimates range from $\langle e\rangle \sim \langle i\rangle \sim10^{-3}-10^{-1}$, depending on the level of stirring \citep{matthews2014}.
\begin{figure}[!h]
\includegraphics[clip=,width=.95\linewidth]{robinhood.pdf}
\caption{Predicted $s_{\rm min}/s_{\rm blow}$ in debris disks, as a function of disk radius. Arrows indicate how the ratio changes with stellar luminosity, surface energy, $f$, and $\eta$. Coloured regions indicate observational constraints on $s_{\rm min}/s_{\rm blow}$ for various systems (see text), and the diagonal solid lines give our calculations for each system. We fix $\gamma=0.1\mathrm{~J~m}^{-2}$, $f=10^{-2}$, and $\eta=10^{-2}$ for this comparison.}
\label{fig:arrows}
\end{figure}
The ratio between the smallest grain size from Eq.~\ref{eq:biermann} and the blow-out size then becomes
\begin{equation}\label{eq:s_ratio}
\frac{s_{\rm min}}{s_{\rm blow}} = 2.4 \left( \frac{a}{5\mathrm{AU}} \right) \left( \frac{L_{\star}}{L_{\odot}} \right)^{-1} \left( \frac{f}{10^{-2}} \right)^{-2} \left( \frac{\eta}{10^{-2}} \right)^{-1} \left( \frac{\gamma}{0.1~\mathrm{J~m^{-2}}} \right),
\end{equation}
where both the stellar mass and the material density drop out, and we assumed $Q_{\rm pr}=1$. Figure~\ref{fig:arrows} compares this ratio with observations of debris disks at large radii, where we predict the most pronounced effect. We have used a fixed $\gamma=0.1\mathrm{~J~m}^{-2}$, $f=10^{-2}$, and $\eta=10^{-2}$, and the arrows indicate the dependence of $s_{\rm min}/s_{\rm blow}$ on various parameters. For the main dust belt around \object{Fomalhaut}, \citet{Minetal2010} found the scattering properties to be consistent with predominantly $\sim100 \mu$m silicate grains \citep[$s_{\rm blow}=13\mu$m][]{Ackeetal2012}. The relative velocities in Fomalhaut are typically taken a factor of 10 higher \citep{WyattDent2002}. For \object{HD105}, \citet{Donaldsonetal2012a} derived $s_{\textrm{min}}=8.9 \mu$m ($s_{\rm blow}=0.5\mu$m) at orbital distances above $\sim50$~AU. Notably, very large grain sizes of $\sim100\mu$m ($s_{\rm blow}\lesssim1\mu$m) are inferred for the recently discovered ``\emph{Herschel} cold debris disks'' \citep{Krivovetal2013}, which are seen around F, G, and K type stars. \citeauthor{Krivovetal2013} were not able to model these systems with a collisional cascade reaching down to $s_{\rm min}=3\mu$m, and proposed that the large grains in these systems are primordial, unstirred material. Our calculations suggest that they can also be explained as the outcome of a collisional cascade. However, the model is highly degenerate, as material properties ($\eta$ and $\gamma$) and belt properties ($f$) are usually poorly known, and all have a large impact on $s_{\rm min}$.
In Figure~\ref{fig:arrows}, we assume constant and equal relative velocities for all particles. In reality, radiation pressure will also increase the eccentricities of small particles with $\beta \lesssim 0.5$. The enhanced eccentricity can be written as $e_{\beta}=\beta / (1-\beta)$. The relative velocity of such a radiation-influenced particle scales with its size as $v_{\rm rel} \propto \beta \propto s^{-1}$, while Eq.~\ref{eq:biermann} predicts the fragmentation velocity scales as $v_{\rm rel}\propto s_0^{-1/2}$. Hence, the relative velocity between the smallest particles increases \emph{faster} than the velocity needed for fragmentation. As a result, particles can reach arbitrarily small sizes in this regime. Particles smaller than $s_{\rm blow}$ are then removed on a short timescale. For a more detailed estimation, $\beta$ should be evaluated for each particular case, considering the optical properties of the material and the shape of the stellar spectrum.
A dearth of small grains in weakly stirred disks is also predicted by \citet{Thebault2008}, but the cause is not a limit on $s_{\rm min}$. In their scenario, $s_{\rm min}$ is fixed and the production rate of the smallest grains decreases with weaker stirring, while the destruction rate is determined by radiation forces and is unaffected by stirring. While the smallest grains present are always of blow-out size, their abundance is set by the balance between their creation and destruction (their Figure~7).
While the theory developed in this work predicts that the smallest particles that can fragment further can be quite large and sizes just below this will be depleted, some smaller particles will still be created as a result of erosive collisions, collisions between larger bodies, and collisions that occur above the average collision velocity. Detailed debris disk models implementing the surface energy constraint are needed to determine the resulting size distribution.
\section{Discussion}
Of the fundamental parameters in our model, the largest uncertainty affects $\eta$, the fraction of kinetic energy used for the creation of new surface. While information may be available about the kinetic and surface energy of the largest fragments, it is hard to quantify whether the remainder of the available energy went into surface creation, heat generation, or kinetic energy of the smallest fragments. Experimentally, studying $\eta$ is challenging, since it requires sensitive and complete measurements down to very small sizes. Once the functional form of $\eta$ is quantified by laboratory and numerical experiments, observations of $s_{\textrm{min}}$ in a system of interest may constrain $f$ and thus the local relative velocities.
During the preparation of this manuscript, we discovered that a similarly defined $s_{\textrm{min}}$ to the one we present has been explored in a more abstract framework, and without elaborating on applications, by \citet{Bashkirov1996}. We note that the lack of data on the size distribution of small collision fragments, as well as the fraction of kinetic and internal energy in the fragments already noted by \citeauthor{Bashkirov1996}, still prevails and we encourage further experiments to quantify these important parameters.
Thus far, we have focussed on equal-sized collisions. While collisions with a larger mass ratio might not lead to catastrophic fragmentation, cratering and erosion may still be important, and might be able to form small particles (Appendix \ref{sec:B}). Assuming a fixed relative velocity, we focus on a particle of size $s_1$. We define a mass loss rate $\dot{m}(s)$ for the larger particle, dependent on collider size $s$. Assuming a collision with a particle of size $s<s_1$ erodes a mass $\propto s^3$, and noting the collision timescale is proportional to the particle density and collision cross-section, we obtain for the \emph{total} mass loss rate $\dot{M} = \dot{m}(s)\,\mathrm{d} s \propto s^{-3.5} s^3 (s+s_1)^2 \,\mathrm{d} s$. If the collisional cross-section is dominated by $s_1$, we find $\dot{M} \propto s_1^2 s^{1/2} $, and thus the mass loss is dominated by the \emph{larger} bodies. When $s \sim s_1$, we obtain $\dot{M}\propto s^{5/2}$.
We have adopted a constant value of 50\% of the collider mass for the largest fragment. However, experiments show $s_{\rm max}$ can be substantially smaller as a function of material and impact velocity \citep[e.g.][]{davis1990,ryan1991}. Such results can easily be implemented in Eq.~\ref{eq:s_min3.5} (and Eqs.~\ref{eq:s_min_uneq} and \ref{eq:s_min_eros}) as necessary. Since $s_{\rm min} \propto s_{\rm max}^{-1}$, smaller sizes for the largest remnant will make the production of small particles even more difficult.
Other collisional systems where the proposed fragment size limit operates include planetary rings. Our calculations are consistent with the observed dominant grain sizes in the rings of Saturn, Jupiter, and Uranus. Because of additional relevant physics, such as tidal and electromagnetic effects, consistency does not directly imply the dominant grain size in all these rings is fragmentation-dominated.
The full implications of an energy-limited $s_{\textrm{min}}$ on systems such as debris disks and planetary rings can only be assessed with models tracking the full particle population with all relevant processes included. For example, if small particles cannot be destroyed in collisions, Poynting-Robertson (PR) drag will influence their orbits, and cause particles to drift towards the star on timescales of Gyrs in the outer parts of disks \citep{wyatt2005, lieshout2014}. Such modelling is outside the scope of the present paper.
\section{Conclusions}
We investigated the energetic constraints on the lower size limit in a distribution of collision fragments. A quantification of the lower limit of such size distributions is relevant for the modelling of debris disks and other astrophysical systems where collisional fragmentation is important.
\begin{enumerate}
\item{Based on surface energy constraints, we derive a parameterised recipe for the smallest fragment size in individual grain-grain collisions.}
\item{The smallest size in a distribution of fragments from a two-particle collision, constrained by the collision energy, is given by Eq.~\ref{eq:s_min3.5}, and illustrated in Figure~\ref{fig:fig1}. For example, at 20 m/s, submicron silicate particles can only be effectively produced by centimeter-sized colliders.}
\item{In the limit where the colliding bodies are split in half, the fragmentation threshold velocity is given by Eq.~\ref{eq:biermann}.}
\item{While dedicated models are needed to reveal the full implications of the fragment size limit, Figure~\ref{fig:arrows} offers an indication of where the size distribution is expected to be influenced.}
\item{In systems where the collision velocities are low, our theory may offer a natural explanation for a paucity of small grains in debris at large orbital distances, such as observed in Fomalhaut and the \emph{Herschel} cold debris disks (Figure~\ref{fig:arrows}).}
\end{enumerate}
\begin{acknowledgements}
The authors wish to thank Carsten Dominik and Xander Tielens for comments and discussions. Dust studies at Leiden Observatory are supported through the Spinoza Premie of the Dutch science agency, NWO. Astrochemistry in Leiden is supported by NOVA, KNAW and EU A-ERC grant 291141 CHEMPLAN.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section*{References}
|
\section*{Acknowledgement}
\noindent {\bf Acknowledgements }
This work is supported by
Grant-in-Aid for Scientific research
from the Ministry of Education, Science, Sports, and Culture (MEXT), Japan, No. 25400248 (M.K.), No. 21111006 (M.K.);
the World Premier International Research Center Initiative (WPI Initiative),
MEXT, Japan (A.K., M.K., and M.Y.);
the Program for Leading Graduate Schools, MEXT, Japan (M.Y.);
and JSPS Research Fellowships for Young Scientists (M.Y.).
|
\section{Introduction}
Graphene, a new material with promising application possibilities and
important fundamental physics aspects, is a two-dimensional allotrope of
carbon which has become one of the most significant topics in solid state
physics (\cite{novo},\cite{intro1},\cite{intro2}, \cite{B}, \cite{BBBB}).
The carbon atoms form a honey-comb lattice made of two interpenetrating
triangular sublattices, $A$ and $B$. A special feature of the graphene band
structure is the linear dispersion at the Dirac points which are dictated by
the $\pi $ and $\pi ^{\prime }$ bands that form conical valleys touching at
the high symmetry points of the Brillouin zone \cite{A}. Electrons near
these symmetry points behave as massless relativistic Dirac fermions with an
effective Dirac-Weyl Hamiltonian \cite{B}.
The understanding of charge transport mechanism of pristine graphene/doped
graphene is crucial for future applications in nanoelectronics. As clean
graphene being a gapless semimetal, is useless for electronic development,
therefore it is necessary to turn graphene from semimetallic to a gap
semiconductor, which can be realized in several ways (with substrates \cit
{sh} by confinement \cite{han} and quantum dots \cite{pono}). In the other
side, a quasigap in the vicinity of the Dirac point can be obtained in two
dimensional system lattice sites with two different site energies and
different probabilities \cite{yu}. In this work we report a theoretical
model to describe the transport mechanism in the ballistic regime at the
interface of clean graphene and graphene with adatoms with arbitrary
energies (positive values as donor and negative values as acceptor) placed
on a site-like position. The impurities are randomized and averaged over its
possible positions which transform the diffusive system in a ballistic one.
In this sense, the system can be considered in thermodynamic equilibrium
with a fixed number of impurity concentration. The diffusive character of
systems with disorder can be study through Green function techniques (see
\cite{janis}, \cite{zie}, \cite{rammer}) and Kubo formalism, which allow to
obtain the quantum corrections to the conductivity and other effects as weak
antilocalization.\footnote
Although for ideal graphene, the dynamics of the electrons produce the same
shot noise as that found in classical diffusion (see \cite{mario}).} For the
purpose of this work, we will consider that the quantum mechanical coherence
length is longer than the sample size $L$, in this case, the disorder of
random impurities is transformed in a mass term in the Hamiltonian by the
averaging procedure. Then by applying Landauer formalism (\cite{landa}) is
enough to study the ballistic behavior of Bloch electrons through the sample
\footnote
Other theoretical methods can be applied to obtain the effects of impurities
in the electronic spectra of graphene (see \cite{feher}).} Graphene-based
devices have not been fully investigated due to the complex processes
required to achieve p- and n-doped semiconducting graphene. By chemical
doping, graphene-based p-n junctions can be obtained (\cite{will}, \cite{lo
), but graphene retain its semimetallic character. In this work, from the
impurity averaged tight-binding Hamiltonian in the long wavelength limit, is
possible to obtain a Dirac equation with mass for the Bloch electrons. In
this sense, a gap in the energy band near the Fermi energy is obtained. The
tunnel junction studied is based on pristine graphene/doped
graphene/pristine graphene (PG/DG/PG). The transport mechanism is highly
dependent on the impurity concentrations and several predictions can be
obtained through theoretical calculations to be applied to different
metal/semiconductor junctions (see \cite{bai}).
The paper is organized as follows. In Section II, we introduce the
tight-binding Hamiltonian with impurities and the averaging procedure. Then,
the long-wavelength approximation is applied to obtain Dirac equation with
mass. In Section III, Landauer formalism to the PG/DG/PG system is applied
for low temperatures. The conductance as a function of the applied voltage
is obtained. Minimum conductance is computed for zero voltage. Finally, an
equation relating the length of the sample as a function of the
concentration of impurities can be computed to obtain a transmission
coefficient which is indepenent on the impurity concentration.
In the final section, the conclusions are presented. In appendix A, the
coefficients of the Taylor expansion of the conductance as a function of the
impurities are computed.
\section{Tight binding model with impurities}
The tight-binding Hamiltonian of graphene for nearest neighbors read
\begin{equation}
H_{0}=-t\underset{\left\langle i,j\right\rangle ,\sigma }{\overset{}{\sum }
(a_{i,\sigma }^{\dag }b_{j,\sigma }+b_{i,\sigma }^{\dag }a_{j,\sigma })
\label{1}
\end{equation
where $a_{i,\sigma }^{\dag }$($a_{i,\sigma }$) creates (annihilates) an
electron on site $\mathbf{r_{i}}$ with spin $\sigma $, where $\sigma
=\uparrow ,\downarrow $ on sublattice $A$ and $b_{i,\sigma }^{\dag }$(
b_{i,\sigma }$) creates (annihilates)\ an electron on site $\mathbf{r_{i}}$
with spin $\sigma $, on sublattice $B$ and $t$ is the nearest neighbor
\left\langle i,j\right\rangle $ hopping energy. Impurities can be included
in the tight-binding description by the addition of a local energy ter
\begin{equation}
H_{imp}=\underset{i,\sigma }{\overset{N_{i}}{\sum }}(V_{i}a_{i,\sigma
}^{\dag }a_{i,\sigma }+W_{i}b_{i+\delta ,\sigma }^{\dag }b_{i+\delta ,\sigma
}) \label{2}
\end{equation
where $V_{i}$ is a random potential at site $\mathbf{r_{i}}$ and $W_{i}$ is
a random potential at site $\mathbf{r_{j}}$ and where $\mathbf{\delta
=a(1,0,0)$. By introducing the Fourier transform of the annihilation and
creation operators $a_{i,\sigma }$ and $b_{i,\sigma }$:
\begin{equation}
a_{i,\sigma }=\frac{1}{\sqrt{N}}\underset{\mathbf{k}}{\overset{}{\sum }}e^{
\mathbf{k\mathbf{r_{i}}}}a_{\mathbf{k},\sigma }\text{ \ \ \ \ \ \
b_{i,\sigma }=\frac{1}{\sqrt{N}}\underset{\mathbf{k}}{\overset{}{\sum }}e^{
\mathbf{k\mathbf{r_{i}}}}b_{\mathbf{k},\sigma } \label{3a}
\end{equation
The Hamiltonian read
\begin{equation}
H=\underset{\mathbf{k,}\sigma }{\overset{}{\sum }}\left[ \phi (\mathbf{k)}a_
\mathbf{k},\sigma }^{\dag }b_{\mathbf{k},\sigma }+\phi ^{\ast }(\mathbf{k)
b_{\mathbf{k},\sigma }^{\dag }a_{\mathbf{k},\sigma }\right] +\frac{1}{N
\underset{\mathbf{k,\mathbf{q,}}\sigma }{\overset{}{\sum }}\left( V(\mathbf{
})a_{\mathbf{k},\sigma }^{\dag }a_{\mathbf{k+\mathbf{q}},\sigma }+W(\mathbf{
})b_{\mathbf{k},\sigma }^{\dag }b_{\mathbf{k+\mathbf{q}},\sigma }\right)
\label{4a}
\end{equation
wher
\begin{equation}
V(\mathbf{q})=\underset{i}{\overset{N_{i}}{\sum }}V_{i}e^{i\mathbf{qr_{i}}
\text{ \ \ \ \ \ \ \ \ }W(\mathbf{q})=\underset{i}{\overset{N_{i}}{\sum }
W_{i}e^{i\mathbf{qr_{i}}} \label{5a}
\end{equation
A configurational averaging over the impurites can be applied over the
Hamiltonian of last equation, that is, the lattice points where the
impurities are located can be placed at random with different random
configurities.\footnote
In general, configurational averaging is applied over the Green function
(see \cite{rammer}). The average restore the translation symmetry of the
system, but transforms the original system of non-interacting electrons to a
correlated one. Configurational averaging applied directly over the
Hamiltonian restore traslation invariance and the impurities appears as an
effective mass term. In this sense, the average over the Hamiltonian
disables the disorder introduced in eq.(\ref{4a}).} We can sum over all the
configurations of possible positions of impurities in the lattice. If there
are $N_{i}$ impurities, then the configurational averaging can be computed
as (see \cite{rammer}
\begin{equation}
\left\langle F\right\rangle =\frac{1}{A^{N_{i}}}\underset{\mathbf{r_{1}}}
\overset{}{\sum }}...\underset{\mathbf{r_{N_{i}}}}{\overset{}{\sum }}F
\mathbf{r_{1},...,r_{N_{i}}}) \label{6a}
\end{equation
where $A$ is the area of graphene sheet. The configurational averaged
Hamiltonian read
\begin{equation}
\left\langle H\right\rangle =\underset{\mathbf{k}}{\overset{}{\sum }}\left[
\phi (\mathbf{k)}a_{\mathbf{k},\sigma }^{\dag }b_{\mathbf{k},\sigma }+\phi
^{\ast }(\mathbf{k)}b_{\mathbf{k},\sigma }^{\dag }a_{\mathbf{k},\sigma }+Va_
\mathbf{k},\sigma }^{\dag }a_{\mathbf{k},\sigma }+Wb_{\mathbf{k},\sigma
}^{\dag }b_{\mathbf{k},\sigma }\right] \label{6.1a}
\end{equation
wher
\begin{equation}
V=\frac{1}{N}\underset{i}{\overset{N_{i}}{\sum }}V_{i}\text{ \ \ \ \ \ }W
\frac{1}{N}\underset{i}{\overset{N_{i}}{\sum }}W_{i} \label{7a}
\end{equation
an
\begin{equation}
\left\vert \phi (\mathbf{k})\right\vert =-t\sqrt{1+4\cos ^{2}(\frac{\sqrt{3
}{2}k_{y}a)+4\cos (\frac{3}{2}k_{x}a)\cos (\frac{\sqrt{3}}{2}k_{y}a)}
\label{7.1a}
\end{equation
and where we have used eq.(\ref{6a}). Hamiltonian of eq.(\ref{6.1a}) can be
diagonalized and the spectrum read
\begin{equation}
E_{\lambda }(\mathbf{k})=\frac{V+W}{2}+\lambda \sqrt{(\frac{V-W}{2
)^{2}+\left\vert \phi (\mathbf{k})\right\vert ^{2}} \label{9a}
\end{equation
where $\lambda =\pm 1$, where the plus sign is for the conduction band and
the minus sign for the valence band. To study the behavior of electrons at
the Dirac point, we can expand the energy near the $K$ Dirac point$(\frac
2\pi }{3a},\frac{2\pi }{3\sqrt{3}a})
\begin{equation}
E_{\lambda }(\mathbf{p})=\lambda \sqrt{\frac{\gamma ^{2}}{4}+\frac{3}{4
\frac{t^{2}a^{2}}{\hbar ^{2}}p^{2}} \label{10a}
\end{equation
where $\gamma =V-W$ and $\mathbf{p=\hbar k}$ and where we the constant term
\frac{V+W}{2}$ has been absorbed into a redefinition of the energy. This
low-energy description is valid as long as the characteristic energy is
smaller than a cutoff $E_{C}\sim \frac{\hbar v_{F}}{a}\sim 2.6eV$ of the
order of the inverse lattice spacing (see \cite{peres-guinea}).
The last Hamiltonian is the Hamiltonian of a massive Dirac fermions, where
the mass and velocity read
\begin{equation}
m=\frac{2\gamma \hbar ^{2}}{3t^{2}a^{2}} \label{11a}
\end{equation
an
\begin{equation}
v_{F}=\frac{\sqrt{3}ta}{2\hbar } \label{12a}
\end{equation
which is identical to the Fermi velocity of clean graphene. In this sense,
graphene with impurities in the long wavelength approximation can be
considered as fermions satisfying the Dirac Hamiltonian with mass given by
eq.(\ref{11a}) and velocity given by eq.(\ref{12a}). The mass of electrons
is proportional to $\gamma $, which depends on the impurity potentials. In
the case that $V_{i}=V_{0}$ and $W_{i}=W_{0}$, $V=n_{i}V_{0}$ and $W=$
n_{i}W_{0}$, where $n_{i}=N_{i}/N$ is the impurity concentration. The mass
term due to the impurities can be interpreted as if graphene is altered by a
periodic potential which originates an effective mass for the propagation of
electrons. This mass term is not a diffusive term in the Hamiltonian, then
it cannot gives information about many-body effects for electrons in
graphene with impurities. Nevertheless, impurities are still there in the
Hamiltonian as a mass term and can give a detailed description for ballistic
transport phenomena, provided that the quantum mechanical coherence length
is longer than the sample size $L$. Even more, graphene with random
impurities is no longer a disorder system when the average is applied to the
Hamiltonian, but inertial effects appears in electrons.
\subsection{Dirac equation}
By apply the quantization procedure to the energy of eq.(\ref{9a}) we can
obtain the Dirac equatio
\begin{equation}
E\psi =(v_{F}\overrightarrow{\sigma }\cdot \overrightarrow{p}+\sigma
_{z}mv_{F}^{2})\psi \label{d1}
\end{equation
where $\overrightarrow{\sigma }$ are the Pauli matrices $\sigma _{x}$ and
\sigma _{y}$. The solution of the Dirac equation read
\begin{equation}
\psi _{\lambda }(\mathbf{r})=\frac{1}{\sqrt{2}}\left[
\begin{array}{c}
\sqrt{1+\frac{\lambda mv_{F}^{2}}{\sqrt{m^{2}v_{F}^{4}+v_{F}^{2}p^{2}}}} \\
\lambda e^{i\phi _{p}}\sqrt{1-\frac{\lambda mv_{F}^{2}}{\sqrt
m^{2}v_{F}^{4}+v_{F}^{2}p^{2}}}
\end{array
\right] e^{\frac{i}{\hbar }\mathbf{p\cdot r}} \label{d2}
\end{equation
where
\begin{equation}
\phi _{p}=arctg(\frac{p_{y}}{p_{x}}) \label{d3}
\end{equation
In the limit of no impurities,~$m=0$ the spinor of eq.(\ref{d2}) is
identical to the one of clean graphene as it is expected.\footnote
The limit $m=0$ can be obtained with $\gamma =0$ which implies that both
impurity potentials $V$ and $W$ can be identical.} In the limit of high
concentration of impurities $m\rightarrow \infty $ the spinor decouples the
pseudospi
\begin{equation}
\underset{m\rightarrow \infty }{\lim }\psi _{+}(r)=\left[
\begin{array}{c}
1 \\
\end{array
\right] e^{\frac{i}{\hbar }\mathbf{p\cdot r}} \label{tb20}
\end{equation
and
\begin{equation}
\underset{m\rightarrow \infty }{\lim }\psi _{-}(r)=\left[
\begin{array}{c}
0 \\
\end{array
\right] e^{\frac{i}{\hbar }\mathbf{p\cdot r}} \label{tb21}
\end{equation
which implies that the impurities break the symmetry of $A$ and $B$
sublattice.
As we said before, graphene with random impurities has been transformed to
graphene with a periodic potential that introduces an effective mass which
is proportional to the impurity concentration. In this sense, the wave
function of eq.(\ref{d2}) encodes the probability amplitude of finding a
Bloch electron in the whole sample of lentgh $L$. This wave function do not
contains information about a possible localization of the electron due to
the interference effects of the impurities. Nevertheless, if the sample of
length $L~$is placed near a sample of clean graphene, the wave functions
must match in the boundary, which introduces several restrictions to the
energies involved for the electron propagation. The fact that electrons in
graphene with a configurational averaging of impurities applied to the
Hamiltonian introduces an effective mass proportional to the impurity
concentration, implies that the restrictions to the energies involved will
be sensitive to the impurities through the mass term. From this result, it
is possible to obtain a detailed description of the transport phenomena in
the ballistic regime where the wave function is not localized.
\section{Tunneling transport through a potential barrier}
Consider the following two-dimensional model where doped graphene is placed
between two reservoirs with different chemical potentials $\mu _{L}$ and
\mu _{R}$ connected through pristine graphene leads (see Figure 1). By
matching the solution of Dirac equation for $x<0$ with $m=0$ with the
solution of Dirac equation for $0<x<L$ (see eq.(\ref{d2})) and with the
solution of Dirac equation for $x>L$ we obtain for the transmission as a
function of the incident energy\footnote
The $y$ direction has a length $D$ which introduces a quantization of the
p_{y}$ component of the momentum.
\begin{equation}
T_{n}(E)=\frac{16\left\vert \eta _{n}\right\vert ^{2}}{\left\vert e^{-i\xi
_{n}L}(\eta _{n}+1)^{2}-e^{i\xi _{n}L}(\eta _{n}-1)^{2}\right\vert ^{2}}
\label{tt1}
\end{equation
where
\begin{equation}
\eta _{n}(E)=\frac{\hbar v_{F}(\xi _{n}-i\alpha _{n})}{E-E_{0}} \label{tt2}
\end{equation
an
\begin{equation}
\xi _{n}(E)=\frac{1}{\hbar v_{F}}\sqrt{E^{2}-E_{0}^{2}-\hbar
^{2}v_{F}^{2}\alpha _{n}^{2}} \label{tt3}
\end{equation
where $E_{0}=mv_{F}^{2}$ and $\alpha _{n}=\frac{n\pi }{D}$, where $D$ is the
width of the sample ribbon. The allowed energy are in the range
-E_{C}<E<E_{C}$. For the $n=0$ mode, the transmission coefficient read
\begin{equation}
T_{0}(E)=\frac{16\eta _{0}^{2}}{16\eta _{0}^{2}+4(\eta _{0}^{2}-1)^{2}\sin
^{2}(\xi _{0}L)} \label{tt3.8}
\end{equation
\begin{figure}[tbp]
\centering
\includegraphics[width=110mm,height=50mm]{dibujo2.eps}
\caption{Schematic illustration of the tunnel junction composed of pristine
graphene in the left and right leads and graphene with impurities in the
scattering region.}
\label{schematic}
\end{figure}
where the maximum are located when $\xi _{0}L=m\pi $, for $m=\in
\mathbb{Z}
\begin{equation}
E_{\max }=\sqrt{\left( \frac{\hbar v_{F}m\pi }{L}\right) ^{2}+E_{0}^{2}}
\label{tt3.9}
\end{equation
The first mode $n=0$ is independent of the width of the ribbon $D$ due to
the the factor $\alpha _{n}=\frac{n\pi }{D}$, which implies that the
conductance and transmission are valid for any value of $D$.\footnote
This could be appropiate for some experimental purpose, since it is not
necessary to consider one of the sample dimension.} The Landauer formula for
the total current flowing from the left to right lead in the $n=0$ mode is
given b
\begin{equation}
I_{T}=I_{L\rightarrow R}-I_{R\rightarrow L}=g_{s}\frac{2\left\vert
e\right\vert }{\hbar }\int_{-E_{C}}^{E_{C}}dET_{0}(E)(f_{L}(E,\mu
_{L})-f_{R}(E,\mu _{R})) \label{tt4}
\end{equation
where $g_{s}=2$ is the spin degeneracy factor and where $f(E,\mu )$ is the
Fermi-Dirac distributio
\begin{equation}
f_{L(R)}(E,\mu _{L(R)})=\frac{1}{1+\exp [\beta (E-\mu _{L(R)})]} \label{tt5}
\end{equation
where $\beta =1/k_{B}T$. If we assume that the chemical potentials are
related as $\mu _{R}=\mu _{L}+\left\vert e\right\vert V_{SD}$ where $V_{SD}$
is the source-drain voltage difference between the leads and we shift the
energy as $E\rightarrow E+\frac{eV_{SD}}{2}$, then the difference of Fermi
distributions can be written as a gate functio
\begin{equation}
g(\beta )=f_{L}(E,\mu _{L})-f_{R}(E,\mu _{R})=-\frac{\sinh (\frac{\beta
eV_{SD}}{2})}{\cosh (\frac{\beta eV_{SD}}{2})+\cosh (\beta (E-\mu _{L}))}
\label{tt6}
\end{equation
\begin{figure}[tbp]
\centering
\includegraphics[width=95mm,height=55mm]{distribution.eps}
\caption{Gate function for different temperatures (red line $\protect\beta
=0.3$, blue line $\protect\beta =2$, green line $\protect\beta =10$ and
black line $\protect\beta =15$, $\left\vert e\right\vert V_{SD}=20$). }
\label{gate}
\end{figure}
which behaves at low temperatures as $f_{L}(E,\mu _{L})-f_{R}(E,\mu
_{R})\sim 1\,$between $\mu _{L}-\frac{eV_{SD}}{2}$ and $\mu _{L}+\frac
eV_{SD}}{2}$ and zero in the remaining energy values (see \ref{gate}). As
\beta $ goes down, the gate function goes to zero by relaxing the behavior
of rectangular function. For $\beta \rightarrow \infty $, the integral of
eq.(\ref{tt4}) read
\begin{equation}
\int_{-E_{C}}^{E_{C}}dET_{0}(E)(f_{L}(E,\mu _{L})-f_{R}(E,\mu
_{R}))=\int_{\mu _{L}}^{\mu _{L}+eV_{SD}}T_{0}(x)dx \label{au20}
\end{equation
\begin{figure}[tbp]
\centering
\includegraphics[width=95mm,height=55mm]{transmission.eps}
\caption{Transimission as a function of incident energy for different
concentration of impurities (red line $E_{0}=0.1$, green line $E_{0}=0.4$,
blue line $E_{0}=1.25$, black line $E_{0}=1.7$).}
\label{transmission}
\end{figure}
In last equation, the limits of the integral are located between $\mu _{L}$
and $\mu _{L}+\left\vert e\right\vert V_{SD}$ in the case that $E_{C}>\mu
_{L}+\left\vert e\right\vert V_{SD}$. In the other side $E_{C}<$ $\mu
_{L}+\left\vert e\right\vert V_{SD}$, the upper limit will be $E_{C}$. The
conductance $G=I_{T}/V_{SD}$ can be computed up to order $E_{0}^{6}$ due to
the complexity of the integrand of last equation. The limit $\mu
_{L}\rightarrow 0$ can be taken without loss of generality, the
\begin{equation}
G=\frac{4\left\vert e\right\vert }{\hbar V_{g}}\left(
f_{0}(V_{SD})+f_{2}(V_{SD})E_{0}^{2}+f_{4}(V_{SD})E_{0}^{4}+f_{6}(V_{SD})E_{0}^{6}+O(E_{0}^{8})\right)
\label{au20.1}
\end{equation
where the coefficients $f_{j}(V_{SD})$ are shown in Appendix A. In figure
\ref{conductance}, the contributions of the different orders in $E_{0}$ can
be obtained for the conductance $G$ as a function of a dimensionless
variable $y=\frac{L\left\vert e\right\vert V_{SD}}{\hbar v_{F}}$, which is
proportional to $V_{SD}$. The limit $V_{SD}\rightarrow 0$ can be taken in $G$
and read
\begin{equation}
\underset{V_{SD}\rightarrow 0}{\lim }G=\frac{4\left\vert e\right\vert ^{2}}
\hbar }T_{0}(0)=\frac{4\left\vert e\right\vert ^{2}}{\hbar }\frac{1}{1+\sinh
^{2}(x)} \label{au20.2}
\end{equation
where $x$ is a dimensionless variable which reads $x=\frac{LE_{0}}{\hbar
v_{F}}$. Last result is the correction to the minimum conductivity without
source-drain voltage. Several theoretical explanations can be found in the
literature related to the chiral nature of low energy excitations (see \cit
{kat}, \cite{tru}, \cite{zie}).
\begin{figure}[tbp]
\centering
\includegraphics[width=105mm,height=65mm]{conductance.eps}
\caption{Conductance as a function of the source-drain voltage for different
concentration of impurities (red line $E_{0}=0.2$, blue line $E_{0}=0.4$,
green line $E_{0}=0.8$, black line $E_{0}=0.84$). The unit of the $x$-axis
is dimensionless.}
\label{conductance}
\end{figure}
In figure \ref{transmission}, the electron transmission as a function of the
electron energy for different concentration of impurities is shown.\footnote
The energy unit of figures \ref{transmission} and \ref{length} are $eV$.} As
it is expected, the transmission probability is suppressed for energy values
less than $E_{0}$, which can be ascribed to the enhancement of the
reflection. For low impurity concentration ($E_{0}<0.2$, red line in figure
\ref{transmission})\ transmission probability do not show transmission gap
for incident energies below $E_{0}$, similar to the results found in \cit
{jian}.
\begin{figure}[tbp]
\centering
\includegraphics[width=115mm,height=60mm]{dconductance.eps}
\caption{$\frac{dG}{dy}$ as a function of the source-drain voltage (the
values of $E_{0}$ correspond to figure \protect\ref{conductance}).}
\label{schematic}
\end{figure}
In order to see what extent the transmission properties are reflected in
measurable quantities which involve averaging over the impurities in the
ballistic regime, we plot the conductance as a function of the source-drain
voltage for different concentration of impurities (see figure \re
{conductance}). As it is expected, the conductance decrease when the
source-drain voltage goes to zero.
\begin{figure}[tbp]
\centering
\includegraphics[width=95mm,height=55mm]{minimum.eps}
\caption{Minimum conductance as a function of concentration of impurities.}
\label{minimum}
\end{figure}
The value of the minimum of the conductance is lower for higher values of
impurity concentrations, in corcondance with experimental reports (see \cit
{yanbin}, fig. 2a, \cite{haijan}, fig.8, \cite{muatez}, fig. 4), where the
energy introduced by the impurities plays the rol of $kT$. In turn, the
variation of the conductance with respect the source-drain voltage can
reflect the effects found in \cite{raj} (fig. 4). The peculiar minimum value
for $V_{SD}=0$ is different than zero as it is expected for graphene. In eq.
\ref{au20.2}), the minimum of the conductance is computed (see figure \re
{minimum}). For no impurities $E_{0}=0$, the conductance is $G=\frac
4\left\vert e\right\vert ^{2}}{\hbar }$, although the density of states has
no charge carriers at the Fermi energy. For high values of impurities, the
minimum conductivity goes to zero, which correspond to the disentangling of
sublattices $A$ and $B$ (see eq.(\ref{tb20}) and eq.(\ref{tb21})) and no
Zitterbewegung effect.
Another relevant point is wether the effective mass $m$ of eq.(\ref{11a})
can be introduced in the Drude-Sommerfield model for diffusive conductivity.
For a simple inspection, the conductivity $\sigma $ is proportional to
n_{i}^{-1}$ which is a typical behavior of solids (see \cite{mahan}). In the
other side, using that the relaxation time $\tau =l/v_{F}$ and $l\sim 1
\sqrt{n_{i}}$ (see \cite{stauber}, \cite{jason}), the conductivity is
proportional to $\beta =n/n_{i}^{3/2}$ which is a dimensionless parameter
that separates the diffusive ($\beta <<1$) from the ballistic regime ($\beta
>>1$) (see \cite{fogler}). In this sense, the effective mass description for
the diffusive regime in graphene is highly sensitive in the relation between
charge carriers and impurity concentration.
With the purpose of obtaining a transmission coefficient with no no
dependence in the concentration of impurities, PG/DG/PG junctions can be
realized by traslating the $E_{0}$ dependence into $L$. For this, the
following renormalization equation can be obtaine
\begin{equation}
\frac{dT_{0}}{dE_{0}}=\frac{\partial T_{0}}{\partial E_{0}}+\frac{\partial
T_{0}}{\partial L}\frac{dL}{dE_{0}}=0 \label{r1}
\end{equation
which is a non-linear first order differential equatio
\begin{equation}
\frac{dL}{dE_{0}}=-\frac{\hbar v_{F}E^{2}}{E_{0}(E^{2}-E_{0}^{2})^{3/2}}\tan
(\frac{L\sqrt{E^{2}-E_{0}^{2}}}{\hbar v_{F}})+\frac{E_{0}L}{E^{2}-E_{0}^{2}}
\label{r2}
\end{equation
Transmission gap is below $E_{0}$, which implies that for practical
purposes, the r.h.s. of last equation can be expanded in Taylor series
around $E=E_{0}
\begin{equation}
\frac{dL}{dE_{0}}=-\frac{L}{E_{0}}-\frac{L^{3}E_{0}}{3\hbar ^{2}v_{F}^{2}}
\label{r3}
\end{equation
which is valid at order $O(E^{0})$. The solution of last differential
equation read
\begin{equation}
L(E_{0})=\pm \frac{L(1)\sqrt{3}\hbar v_{F}}{\sqrt{3\hbar
^{2}v_{F}^{2}E_{0}^{2}+2L^{2}(1)E_{0}^{2}\ln (E_{0})}} \label{r4}
\end{equation
where $L(1)$ is the length of the sample of doped graphene when $E_{0}=\frac
V-W}{2}=1$. Last equation is valid only for incident energy $E\sim E_{0}$.
For impurity concentrations below $E_{0}=e^{-\frac{3}{2}(\frac{\hbar v_{F}}
L(1)})^{2}}$, real solution for $L$ does not exist.
\begin{figure}[tbp]
\centering
\includegraphics[width=105mm,height=65mm]{length.eps}
\caption{Length of the doped graphene sample as a function of the impurity
concentration. Black line for $L(1)=10^{-9}m$ and blue line for
L(1)=10^{-6}m.$}
\label{length}
\end{figure}
In fig. \ref{length}, the reference value is $L(1)=10^{-9}m$ and
L(1)=10^{-6}m$. For values $E_{0}<1$, the length of the doped graphene
sample in both cases increase considerably. These interesting phenomena can
provide an important reference to the design of various electronic devices
based on graphene with energy gap, but where the impurities only introduces
a minimum threshold for the electron transmission coefficient. In turn, the
renormalization methods that relates the length of the sample with the
impurity concentration can give some insight of how the diffusive and
ballistic regimes are related by applying Shot noise measurements (see \cit
{mario}), where the conductance variations are measured as a function of the
sample length of doped and clean graphene.
Finally, is interesting to note that the averaging procedure introduces a
gap in the energy band as it is occur in the Haldane model (see \cit
{haldane} and for a more accurate version see \cite{sti}). The Haldane model
depends on an inversion symmetry breaking on-site energy $M$ for the
sublattice $A$ and $-M$ on sublattice $B$ and a complex hopping amplitude
between next-nearest neighbor due to the Peierls subtitution, which is
obtained by applying a staggered magnetic field, which is positive near the
center of each hexagon and negative near the edges which results in a zero
net flux in the hexagon. In our model, random impurities are introduced as
adsorbates that change the on-site energies where they are located. In the
model introduced in this work, the particle-hole symmetry is not destroyed
because we do not take into account next-nearest neighbor bonds. The crucial
point is that the random impurities breaks the translational symmetry and
the inversion symmetry. With the averaging procedure, the translational
symmetry is restored, but the inversion symmetry is not. In fact, the mass
term in our model depends on $V-W=\frac{1}{N}\underset{i=1}{\overset{N_{i}}
\sum }}(V_{i}-W_{i})$, where $V_{i}$ is the impurity located in sublattice
A $ and $W_{i}$ is the impurity located in sublattice $B$. If both values
are the same, $V_{i}=W_{i}$, then the mass term is zero, which implies that
Bloch electrons cannot distinguish between the upper side of the sample with
respect the other side, that is, the inversion symmetry is restored and no
gap is obtained.\footnote
The condition $V_{i}=W_{i}$ is not the unique condition for no gap in the
energy bands. Because we are taking an averaging over the possible locations
of the impurities, no matter if an on-site energy $V_{i}$ is not equal to
W_{i}$ provided that there is another on-site energy that is $W_{j}=-V_{i}$.}
In this sense, our model is a particular case of the Haldane model, where
t_{2}=0$ and $-W_{i}=V_{i}=M$.
\section{Conclusion}
In this paper we have studied the transport mechanism of
pristine/doped/pristine graphene junction in the ballistic regime. The
conductance as a function of the impurity concentration has been computed
using Landauer formalism with the application of the averaging procedure on
the impurity positions over the Hamiltonian. Minimum conductivity is
obtained exactly for low temperatures. A renormalization equation was
obtained for the sample length and the impurity concentration with the
purpose to obtain a transmission coefficient that do not depends on the
impurities. This result can be of importance for the manufacturing of
Schootky junctions with gapped doped graphene. Finally, the model introduced
in Section II is related to the Haldane model, finding that the band gap
obtained is due to breaking inversion symmetry introduced by the
position-averaged impurities.
\section{Appendix}
The coefficients of the Taylor expansion of the transmission of eq.(\ref{tt1
) as a function of $E_{0}$ up to order $O(E_{0}^{6})$ read
\begin{equation}
f_{0}(y)=\frac{\gamma y}{L} \label{a1}
\end{equation
\begin{equation}
f_{2}(y)=\frac{L}{\gamma }(\frac{1}{y}\sin ^{2}(y)-Si(2y)) \label{a2}
\end{equation
\begin{eqnarray}
f_{4}(y) &=&\frac{L^{3}}{24\gamma ^{3}y^{3}}(-12y^{2}\cos
(2y)+(8y^{2}-1)\cos (4y)+1 \label{a3} \\
&&-6y\sin (2y)+2y\sin (4y)-24y^{3}Si(2y)+32y^{3}Si(4y))) \notag
\end{eqnarray
an
\begin{gather}
f_{6}(y)=\frac{L^{5}}{960\gamma ^{5}y^{5}}(12+2(3+29y^{2}-58y^{4})\cos (2y)-
\label{a4} \\
4(3+16y^{2}-128y^{4})\cos (4y)-6(1-3y^{2}+54y^{4})\cos (6y) \notag \\
-y(33+58y^{2})\sin (2y)+16y(-3+8y^{2})\sin (4y)+(9y-54y^{3})\sin (6y) \notag
\\
-8y^{5}(29Si(2y)-256Si(4y)+243Si(6y))) \notag
\end{gather
where $\gamma =\hbar v_{F}$ and $y=\frac{L\left\vert e\right\vert V_{SD}}
\gamma }$, $Si(x)$ is the sine integral function.
|
\section{Introduction}
Equisingularity has been first studied for hypersurfaces by O. Zariski. Different concepts can be used to describe the idea of all the fibers in a family of varieties having similar singularities. Many of these concepts consist in measuring some numerical invariant and requiring its constancy along the subspace. However, apart from the case of families of reduced plane curves, where almost all the natural concepts of equisingularity coincide, the general situation is quite ambiguous and unclear.
Even in the case of one parameter family of non planar reduced curves, the classical concepts split into different levels of strength. However, in this case the situation is more or less clear and understood. We refer to \cite{MGranger-equisingularite} for a description of the case of families of reduced curves. We also recommend the expository paper by J. Lipman
\cite{JLipman-equisingularity-obergurgl}, for a general overview of equisingularity problems.
In this work we focus on the case of families of generically reduced curves. These are families where the special fiber may have an embedded component. This happens always when the surface described by the family of curves is not Cohen-Macaulay. When one works with families of curves that are not complete intersections, it is natural to deal with these kind of situations. The examples we give in this work show how natural it is to meet such non Cohen-Macaulay surfaces.
When the family of curves is no longer reduced, many of the ingredients of the classical proofs, and even some definitions, fail. The Milnor number, somehow the star invariant goes to infinity.
Our main purpose in this work consists in comparing some of those equisingularity criteria that can be expressed without requiring the fibers to be reduced. More precisely we consider the Whitney regularity, Zariski's discriminant criterion and topological triviality. We state criteria that are close to normalization in a family and equisaturation.
Our main results are the equivalence between Zariski's discriminant criterion and Whitney regularity together with the fact that topological triviality implies smoothness of the normalized surface. Both results were already known in the case of reduced curves. Specialists expected them to hold also in our case, but no written proof of it is known to the authors. As a tool in one of the proofs, we establish a natural lemma which can be viewed as a partial analogue of the Rolle theorem in the one dimensional complex case.
All along this work we give examples. Some of them illustrate the concepts which appear in the text and others are counterexamples to seemingly natural conjectures. The reader will find an example of a surface Whitney regular along its singular locus that is not Cohen Macaulay and examples that illustrate how some equisingularity criteria are stronger than others.
We also explore the conservation of Whitney regularity and Zariski's discriminant criterion after two types of modifications: the blow up of the singular locus and the Nash modification. We give examples showing that neither of these properties is stable under these modifications.
\section{Equisingularity on surfaces}
Let $(S,0)$ be a germ of a reduced and irreducible complex surface with one-dimensional smooth singular locus.
A generic projection $\pi : (S,0) \rightarrow (\mathbb{C},0)$ exhibits the surface $S$ as a one-parameter flat
deformation of the curve $X_0 : = \pi^{-1}(0)$.
For a generic projection $\pi$ the curve $X_0$ is reduced at its generic point. Hence, when $S$ is Cohen-Macaulay at the origin, such a projection makes $S$ into a flat deformation of a reduced curve.
In the paper \cite{MGranger-equisingularite}, the authors give a full description of different types of
equisingularity conditions and explain the relations between them, especially in the case of deformations of
reduced curves. Many of these criteria are given in terms of invariants such as Milnor number,
multiplicity, Milnor number of a generic planar projection, dimensions of various ``tangent cones'', and also
Whitney regularity conditions, equisaturation, Zariski discriminant criterion, and simultaneous resolution.
Since we do not want to restrict to the Cohen-Macaulay case, we will not have deformations of reduced curves, therefore we will not be able to use the Milnor number invariant. We will mainly focus on two equisingularity criteria:
Whitney regularity conditions and Zariski's discriminant criterion. However we will also briefly consider topological triviality, normalization in a family and equisaturation or strong equisingularity.
\begin{definition}\label{whitney}
Let $(S,0)$ be a germ of a complex surface with one dimensional smooth singular locus ${\mathcal C}$.
We will say that $(S,0)$ is \textbf{Whitney regular} if a small representative $(S\setminus {\mathcal C}, {\mathcal C})$ satisfies Whitney conditions $(a)$ and $(b)$ at 0, {\it that is}:
For any sequences of points $(x_n)\subset S\setminus {\mathcal C}$ and $(y_n)\subset {\mathcal C}\setminus
\{0\}$ both converging to $0$ and such that the sequence of lines $(x_ny_n)$ converges to a line $l$ and the sequence of directions of tangent spaces, $T_{x_n}S$, to $S$ at $x_n$,
converges to a linear space $T$ we have:
$a)$ the direction of the tangent space $T_0{\mathcal C}$ to ${\mathcal C}$ at the origin is such that
$T_0{\mathcal C}\subset T$
and
$b)$ $l\subset T$.
\end{definition}
It is not hard to prove that condition $(b)$ implies condition $(a)$, see for example \cite{DTrotman-ENS}.
\begin{definition}
Let $(S,0)$ be a germ of a complex surface and $\pi : S \rightarrow \mathbb{C}^2$ a finite map.
The critical locus of $\pi$ is the set of points in $S$ where the map $\pi$ does not
induce a local isomorphism. The discriminant locus of $\pi$ is the set-theoretic image of the critical locus of $\pi$.
\end{definition}
Notice that the singular locus of a surface $S$ is always a subset of the critical
locus of any finite map to $\mathbb{C}^2$.
Zariski's discriminant criterion in dimension two can be stated as follows:
\begin{definition}
Let $(S,0)$ be a germ of a complex surface with one-dimensional smooth singular locus ${\mathcal C}$.
We will say that $(S,0)$ satisfies \textbf{Zariski's discriminant criterion} if any finite map induced by a generic
linear projection from the ambient space to $\mathbb{C}^2$ has smooth discriminant locus.
\end{definition}
The discriminant of a map can be endowed with a scheme structure using Fitting ideals as
in \cite{BTeissier-hunting}. Zariski's discriminant criterion can then be stated in terms
of constancy of the multiplicity of the discriminant.
In \cite[Theorem III. 5]{MGranger-equisingularite} the authors prove that when $(S,0)$ as above is a complete
intersection then Zariski's discriminant criterion is equivalent to Whitney regularity.
Their proof is based on the fact that in this case, the discriminant space is
a hypersurface of $\mathbb{C}^2$ and they apply the L\^e-Greuel formula, relating the Milnor number of a curve
and the multiplicity of the discriminant.
When the surface is not a complete intersection, the discriminant space (with scheme structure)
may have an embedded component. However, if we define the divisorial discriminant to be the closure
of the discriminant space without the origin ({\it i.e.} we ignore the embedded component)
then we still can use L\^e-Greuel formula whenever we have a deformation of a reduced curve
as in \cite[Definition 3.2]{JSnoussi-CMH} and \cite[3.4]{RBondil-DTLe-Trends}.
Therefore, when $(S,0)$ is a Cohen-Macaulay surface, we can still prove the
equivalence between Whitney regularity and Zariski's discriminant criterion
using the same technique as in \cite[Theorem III. 5]{MGranger-equisingularite}.
Before we continue, we would like to give an example showing that a surface as above which is
Whitney regular need not be Cohen Macaulay.
This fact is known to specialists, however we do believe it is useful to make it explicit.
\begin{example}\label{whitneynocm}
\end{example}
Let $(S,0)$ be the surface parametrized by the map:
$$\rho: (a,t) \mapsto (a, t^3, t^4, at^5);$$
it is a family of space curves degenerating to a planar cusp.
The singular set of the surface $(S,0)$ is the curve $(\mathbb{C}\times \{(0,0,0)\}, 0)$.
Consider the linear projection
$$r: \mathbb{C}^4 \rightarrow \mathbb{C} \times \{(0,0,0)\}.$$
In order to prove the condition $(b)$, it is enough to prove the condition $(a)$ and that,
for every sequence of points $(p_n)$ in the non-singular locus of $S$ converging
to the origin, the sequence of lines generated by $p_n$ and $r(p_n)$ converges
to a line contained in the limit $T$ of tangent spaces $T_{p_n}S$.
The Jacobian matrix of $\rho$ is
$$\left(
\begin{array}{cccc}
1 & 0 & 0 & t^5 \\
0 & 3t^2 & 4t^3 & 5at^4\\
\end{array}
\right)
$$
At each non-singular point $p_n$ of $S$ the tangent plane is spanned by the rows of
the Jacobian matrix at $p_n$. At the limit all these planes contain the vector
$(1,0,0,0))$, which spans the tangent line to the singular locus at $0$.
Hence conditon $(a)$ is verified.
Let $l_n$ be the line containing the points $p_n= \rho(a_n,t_n)$ and $r(p_n)$.
The direction of $l_n$ is given by the vector
$(0, t_n^3, t_n^4, a_nt_n^5)$.
After dividing through by $t_n^3$, we see that all these lines, independently of
the choice of $a_n$ and $t_n$ converge to the line $l$ generated by the vector $(0,1,0,0)$.
On the other hand, the tangent space $T_{p_n}$ is spanned by the vectors $(1, 0, 0, t_n^5)$
and $(0, 3t_n^2, 4t_n^3, 5a_nt_n^4)$. Dividing the second vector by $t_n^2$,
we see that all these tangent spaces, independently of the choice of $a_n$ and $t_n$ converge
to the linear space spanned by the vectors $(1,0,0,0)$ and $(0,1,0,0)$ which contains the line $l$.
Thus the surface $S$ is Whitney regular in the sense of definition \ref{whitney}.
Let us now compute equations of the surface $S$.
If we use the local co-ordinates $(x,y,z,w)$ in the ambient space, the surface $S$
can be defined by the polynomials
$$y^4 - z^3, yw - xz^2, zw - xy^3, x^3y^5 - w^3, x^2y^2z - w^2.$$
It is a reduced surface. The hyperplane section $S \cap (x=0)$ has an embedded component at the origin.
Thus the surface $(S,0)$ is not Cohen-Macaulay at the origin.
\section{Whitney regularity is equivalent to Zariski's discriminant criterion in dimension two}
We will prove now that Whitney regularity conditions are equivalent to Zariski's discriminant
criterion for surfaces. This equivalence has been proved for complete intersections, still in dimension two,
by Brian\c{c}on, Galligo and Granger in \cite[Theorem III. 5]{MGranger-equisingularite}.
\begin{theorem}\label{equisingularidad}
Let $(S,0)$ be a germ of reduced and irreducible complex surface with a non-singular one-dimensional singular locus ${\mathcal C}$.
The surface is Whitney regular if and only if it satisfies
Zariski's discriminant criterion.
\end{theorem}
In order to prove the theorem we will need to use a criterion for Whitney regularity, established in a general setting by B. Teissier
in \cite[Theorem V.1.2]{BTeissier-rabida}. It relates Whitney regularity of a pair of strata to the constancy of
the multiplicity of the family of polar varieties along the small stratum.
Let us first recall some definitions.
Consider a germ of analytic surface $(S,0) \subset (\mathbb{C}^N,0)$ and a linear projection $p: \mathbb{C}^N \rightarrow \mathbb{C}^2$.
\begin{definition}
When the restriction of the projection $p$ to a representative $S$ of $(S,0)$ is finite, the closure in $S$ of the critical locus of the restriction of $p$ to the non-singular locus of $S$ is called the polar curve associated to $p$ on $S$ at $0$.
\end{definition}
It is well known that for a generic projection $p$, the associated polar curve is either empty or
one-dimensional. Furthermore, its multiplicity at the origin does not depend on the choice of the
(generic) projection; see for example \cite[Chapter IV]{BTeissier-rabida}.
B. Teissier proved in \cite[Theorem V.1.2]{BTeissier-rabida}, that Whitney regularity is equivalent to equimultiplicity of the polar varieties along the small stratum. In dimension two, there are only two polar varieties: the polar curve and the surface itself. Since a generic projection at a given point is still generic at a nearby point, equimultiplicity of the polar curves along the singular locus of a surface implies that the general polar curve is empty.
Therefore B. Teissier's result, restricted to surfaces, can be restated as follows:
\begin{lemma}\label{teissier}
Let $S$ be a representative of a complex surface germ $(S,0)$ whose singular locus is smooth and one-dimensional.
The surface is Whitney regular if and only if the general polar curve at the origin is empty and the surface is
equimultiple along its singular locus near the origin.
\end{lemma}
\begin{proof} (of Theorem \ref{equisingularidad})
Let $S$ be a small representative of a germ of a reduced and irreducible complex surface $(S,0)$, with one-dimensional smooth singular
locus ${\mathcal C}$.
Assume that the strata $(S\setminus {\mathcal C}, {\mathcal C})$ satisfy Whitney conditions.
Let $\pi: S \rightarrow \mathbb{C}^2$ be a finite map induced by a general linear projection. By lemma \ref{teissier},
the polar curve associated to $\pi$ is empty, so the critical locus of $\pi$ is the curve ${\mathcal C}$.
Since this curve is smooth and $\pi$ general, its image, which is the discriminant locus, is again smooth.
Thus $(S,0)$ satisfies Zariski's discriminant criterion.
Conversely, let $\pi: S \rightarrow \mathbb{C}^2$ be a finite map induced by a general linear projection
$p : \mathbb{C}^N \rightarrow \mathbb{C}^2$, for which the discriminant locus is non-singular.
We are going to prove that the polar curve associated to $\pi$ is empty and that the multiplicity of the surface along its singular locus is constant.
The proof will follow from the six steps below:
{\it Step 1:} The normalization of the germ $(S,0)$ is non-singular.
Indeed, the projection $\pi : S \rightarrow \mathbb{C}^2$ is finite and has a smooth discriminant locus. The composition map $\pi \circ n$ of
$\pi$ with the normalization $n$ is also finite and its discriminant locus will be either smooth or empty. A normal surface singularity with smooth or empty discriminant (in $\mathbb{C}^2$) is non-singular (see \cite[Theorem 5.2]{Barth}).
{\it Step 2:} Let $0 \in D\subset \mathbb{C}^2$ be a line and $t\in D_t$ a sufficiently close parallel line to it.
We are going to show that in a sufficiently small neighborhood of the origin, the hyperplane section $\pi^{-1}(D_t)$ is connected.
\begin{lemma}\label{connected}
There exists $\epsilon_0 >0$ such that for every $0 < \epsilon \leq \epsilon_0$ the curve $\pi^{-1}(D_t)\cap \mathbb{B}_{\epsilon}$ is connected, for every $t$ with sufficiently small absolute value; where $\mathbb{B}_{\epsilon}$ is the ball of $\mathbb{C}^N$ centered at the origin with radius $\epsilon$.
\end{lemma}
\begin{proof}
The germ of surface $(S,0)$ is assumed to be irreducible. So its normalization is still a germ and moreover, by {\it Step 1}, it is the germ of a smooth surface. So a sufficiently small neighborhood of $n^{-1}(0)$ is isomorphic to an open neighborhood $U$ of the origin in $\mathbb{C}^2$. The composition map:
$$U \hflp{\cong}{} n^{-1}(S) \hflp{n}{} S \hflp{\pi}{} \mathbb{C}^2$$
is a finite map with discriminant locus contained in a smooth curve of $\mathbb{C}^2$.
Following \cite[Corollary 5.3]{FArocaJSnoussi-quasi-ordinary}, there exist a system of coordinates $(x,y)$ in $U$ and a natural number $a$ such that, up to isomorphism in the basis, the composition map is of the form $(x,y) \mapsto (x^a, y)$. This is a consequence of the classification of normal quasi-ordinary singularities; see also \cite[III. 5]{Barth}.
So the inverse image of a line $D_t$ will be a curve in $U$ with an equation of the form $\alpha x^a + \beta y +c = 0$. Such a curve is connected. Its image by the normalization is still a connected curve in $S$. Hence, the inverse image by $\pi$ of $D_t$ is a connected curve in a small representative of $S$.
\end{proof}
{\it Step 3:} A Rolle--type lemma for complex curves.
\begin{lemma}\label{rolle}
Let ${\mathcal T} \subset \mathbb{C}^N$ be an analytic curve. Let ${\mathbb{B}}_{\tau , \epsilon}$ be the open ball of $\mathbb{C}^N$ centered at a point
$\tau \in {\mathcal T}$ with radius $\epsilon$.
Suppose the intersection ${\mathcal T} \cap {\mathbb{B}}_{\tau, \epsilon}$ is connected.
Consider a linear projection $ \mathbb{C}^N \rightarrow \mathbb{C}$ that induces a finite ramified covering $\rho : {\mathcal T}\cap {\mathbb{B}}_{\tau, \epsilon} \rightarrow \mathbb{C}$, with $\rho (\tau) =0$.
If $\rho^{-1}(0) \neq \{\tau\}$ then the map $\rho$ has a critical point different from the points in the fiber $\rho^{-1}(0)$.
\end{lemma}
\begin{proof}
This lemma will follow from a Hurwitz formula for possibly singular curves. Let us call:
$d$ the degree of $\rho$
$n$ the cardinality of $\rho^{-1}(0)$ ; by hypothesis $n\geq 2$
$\chi$ the Euler Characteristic of ${\mathcal T} \cap {\mathbb{B}}_{\tau, \epsilon}$
$\chi_0$ the Euler characteristic of $\mathbb{C}$
Let us suppose that there are no critical points for $\rho$ in ${\mathbb{B}}_{\tau, \epsilon}$ outside $\rho^{-1}(0)$.
We can then chose a triangulation of a ball of $\mathbb{C}$, in such a way that the origin is a vertex and no critical value lies in any edge nor a face. This triangulation is lifted by $\rho$ to a triangulation where all the points of $\rho^{-1}(0)$ are vertices.
So we obtain a Hurwitz formula for this situation
$$\chi = d \chi_0 + n - d.$$
Knowing that Euler characteristic of $\mathbb{C}$ is 1 we have
$$\chi = n.$$
On the other hand, Euler characteristic is the alternating sum of dimensions of homology spaces.
Since ${\mathcal T}\cap {\mathbb{B}}_{\tau , \epsilon}$ is connected of real dimension two and not compact, we have
$$\chi = 1 - h_1({\mathcal T}\cap{\mathbb{B}}_{\tau, \epsilon}) \leq 1,$$
which contradicts the fact that $n\ge2$.
Hence there is necessarily a critical point of $\rho$ outside $\rho^{-1}(0)$. Note that this point may be a singular point of ${\mathcal T}\cap {\mathbb{B}}_{\tau , \epsilon}$.
\end{proof}
{\it Step 4:} The inverse image by $\pi$ of the discriminant locus is the singular locus ${\mathcal C}$ of $S$.
In fact, consider a point $d$ in the discriminant locus of $\pi$, close to the origin. Consider a general line $D_d\in \mathbb{C}^2$ containing $d$. Since the discriminant locus of $\pi$ is non singular, the intersection of $D_d$ with the discriminant is precisely the point $d$.
The inverse image $\pi^{-1}(D_d)$ is a curve in $S$. By Lemma \ref{connected}, it is connected.
The restriction of $\pi$ to $\pi^{-1}(D_d)$ has all its critical values in the intersection of $D_d$ with the discriminant locus of $\pi$. It has then only one critical value, $d$. By Lemma \ref{rolle} the inverse image $\pi^{-1}(d)$ consists of one point that lies in the singular locus ${\mathcal C}$ of $S$.
{\it Step 5:} The general polar curve is empty. In fact, the critical locus of $\pi$ is in the inverse image of the discriminant locus. By {\it Step 4}, this inverse image is the singular locus of $S$.
{\it Step 6:} The multiplicity of $S$ along ${\mathcal C}$ is constant. In fact, let $c\in {\mathcal C}$ be close to the origin. Since
$\pi$ is generic, the multiplicity of $S$ at $0$ is the degree of $\pi$ at $0$; {\it i.e.} $$m(S,0) = {\rm deg}_0 \pi.$$
In the same way
$$m(S,c) = {\rm deg}_c \pi.$$
The conservation of the degree implies
$${\rm deg}_0 \pi = {\displaystyle \sum_{x\in \pi^{-1}(\pi(c))}}{\rm deg} _x \pi.$$
By {\it Step 4}, $$\pi^{-1}(\pi(c)) = c.$$
So
$$m(S,0) = {\rm deg} _0 \pi = {\rm deg}_c \pi = m(S,c).$$
This ends the proof of \ref{equisingularidad}
\end{proof}
\begin{remark}
1) D.T. L\^e and B. Teissier proved in \cite[Theorem 5.3.1]{DTLe-BTeissier-CESPCW2} that Whitney conditions
are equivalent to the constancy
of the Euler characteristic of plane sections of all possible dimensions along the small stratum. This appears implicitly in our proof, since the main tool was an Euler characteristic calculation.
2) In lemma \ref{connected} we used an argument on quasi-ordinary singularities to conclude on the connectedness of the fibers we consider. We could have used a much stronger result by H. Hamm and D.T. L\^e in \cite[Theorem II.1.4]{HHamm-DTLe-creil1988}.
\end{remark}
\section{Topological triviality and smoothness of the normalization}
Let us state two other equisingularity criteria valid for non Cohen-Macaulay surfaces: smoothness of the normalization and topological triviality. The first one is a weaker version of simultaneous normalization in a family. The second one states that the family of curves is homeomorphic to a product of the special fiber with the base. We will compare these criteria to the previous ones.
\vglue .3cm
{\bf Normalization in a family}
In the case of a flat family of reduced curves, one has normalization in a family when the normalization of the surface induces a normalization on each curve of the family. In particular, it implies that the normalized surface is non-singular.
When the surface is not Cohen-Macaulay, the special fiber has an embedded component at the origin.
So its total ring of quotients is equal to its ring of holomorphic functions and hence, there is no ``reasonable" notion of normalization.
Instead of normalization in a family we can use the weaker condition of smoothness of the normalized surface.
From {\it Step 1} in the proof of theorem \ref{equisingularidad} we obtain the following consequence:
\begin{corollary}\label{smoothnormalization}
Let $(S,0)$ be a germ of a reduced complex analytic surface satisfying Whitney's conditions along its smooth one-dimensional singular locus. The normalization of $(S,0)$ produces a non-singular surface.
\end{corollary}
It is well known that the converse is not true. See Example \ref{topotrivial} below.
\vglue .3cm
{\bf Topological triviality}
Let $(S,0)$ be a germ of a reduced surface with smooth singular locus ${\mathcal C}$. Let $r : (S,0) \rightarrow ({\mathcal C}, 0)$ be a retraction making $(S,0)$ into a family of curves.
\begin{definition}
The family of curves $r: S \rightarrow {\mathcal C}$ is said to be topologically trivial along ${\mathcal C}$ if there exists a homeomorphism $h: S \rightarrow r^{-1}(0) \times {\mathcal C}$ such that $r =\pi \circ h$, where $\pi: r^{-1}(0) \times {\mathcal C} \rightarrow {\mathcal C}$ is the natural projection.
\end{definition}
It is well known that Whitney regularity implies topological triviality (see \cite{JMather-notes}).
In the surface $y^2 - x^2(x+z)=0$, with projection to the $z$-axis, the special fiber is a cusp, and the general one has two branches. It shows that one can have smoothness of normalized surface without having topological triviality.
When we have a family of reduced curves $f: S \rightarrow D$ with a section $\sigma$ such that the fibers $f^{-1}(t)$ are all non singular outside $\sigma(t)$, R.-O. Buchweitz and G.-M. Greuel proved in \cite[Thm.5.2.2]{ROBuchweitz-GMGreuel}, that topological triviality is equivalent to the constancy of the Milnor number of the fibers and equivalent to weak simultaneous resolution. In particular, topological triviality implies smoothness of the normalization.
As a consequence we obtain:
\begin{lemma}\label{topotriv-normal}
A normal surface singularity is a topologically trivial family of curves if and only if the surface is non singular.
\end{lemma}
We are now going to prove that topological triviality implies smoothness of the normalization, even when the surface is not
Cohen-Macaulay.
\begin{theorem}\label{topotri-smoothness}
Let $(S,0)$ be a germ of a reduced surface with a smooth one dimensional singular locus ${\mathcal C}$.
Let $r: S \rightarrow {\mathcal C}$ be a topologically trivial family of curves. Then the normalization of the surface $(S,0)$ is
non-singular.
\end{theorem}
\begin{proof}
It is enough to prove the statement for each analytically irreducible component of the surface $(S,0)$. So we may assume the surface is reduced and irreducible.
Let $r: S \rightarrow {\mathcal C}$ be the retraction making $S$ into a topologically trivial family of curves. The triviality, together with the irreducibility of $S$ at the origin, imply that the fibers $r^{-1}(t)$ are irreducible and generically reduced for all $t$. The surface $S$ is then analytically irreducible at every point.
Call $n: {\bar S} \rightarrow S$ the normalization of $(S,0)$. The inverse image by $n$ of any point of ${\mathcal C}$ is a single point. Then the normalization $n$ induces a homeomorphism between ${\bar S}$ and $S$. Call $\rho = r\circ n$ the composition map. The projection $\rho : {\bar S} \rightarrow {\mathcal C}$ is a topologically trivial family of curves.
Since the surface ${\bar S}$ is normal, and hence Cohen-Macaulay, the family of curves $\rho : {\bar S} \rightarrow {\mathcal C}$ is a topologically trivial family of reduced curves. So we can apply lemma \ref{topotriv-normal}. The surface ${\bar S}$ is then non singular.
\end{proof}
\begin{remark}
1) Since Whitney conditions imply topological triviality, Theorem \ref{topotri-smoothness} is a stronger statement than Corollary \ref{smoothnormalization}. In fact, It is well known that Whitney regularity is not equivalent to topological triviality; \cite[Chap. V]{MGranger-equisingularite}.
2) In order to prove lemma \ref{topotriv-normal} we can use Mumford's criterion for smoothness of normal surfaces. Indeed, if $(S,0)$ is a normal surface with a topologically trivial family of reduced curves along a non-singular curve, then a suitable representative $S$ of $(S,0)$ is a product of two discs. Its link will be homeomorphic to a sphere.
\end{remark}
Let us now give an example of a topologically trivial family of curves for which the surface is not Whitney regular. We do believe it is worth to give such an example, where the special curve is not reduced, and therefore we do not use the constancy of Milnor number.
\begin{example}\label{topotrivial}
\end{example}
Consider the parametrized surface $S\subset \mathbb{C}^4$ given by
$$n: (a,t) \mapsto (a,t^3, t^5, at^2)$$
and consider local coordinates $x$, $y$, $z$ and $w$ in $\mathbb{C}^4$.
The singular locus of $S$ is the $x$-axis.
One can easily check that the map $n$ is bijective and is an isomorphism outside the $x$-axis. So it is the normalization of $S$.
The family of curves $(t^3, t^5, at^2)$ parametrized by $a$ is a topologically trivial family.
In fact the map
$$((0,t^3,t^5),a) \mapsto (a, t^3, t^5, at^5/t^3), {\rm whenever} \, t \neq 0$$
and $$((0,0,0),a) \mapsto (a,0,0,0)$$
is a homeomorphism.
However the surface $S$ does not satisfy Whitney conditions along its singular locus.
In fact, consider the sequence of points $p_k = n(1/k, 1/k)$ and $q_k = (1/k , 0, 0, 0)$.
The line $(p_kq_k)$ is spanned by the vector $(0, 1, 1/k^2, 1)$. so the sequence of lines $(p_kq_k)_k$ converges to the line spanned by $(0, 1, 0, 1)$.
On the other hand, the tangent space $T_{p_k}S$ is spanned by the vectors $(1, 0, 0, 1/k^2)$ and $(0, 3, 5/k^2, 2)$. The limit is the plane spanned by $(1, 0, 0, 0)$ and $(0, 3, 0, 2)$ which does not contain the vector $(0, 1, 0, 1)$.
\section{Equisaturation}
O. Zariski in \cite{OZariski-equising3}, and F. Pham together with B. Teissier in \cite{FPham-BTeissier-fractions}, introduced concepts of saturation of reduced local analytic algebras. In both cases these saturated algebras are intermediate rings between the ring of holomorphic functions and its integral closure.
Since in this paper we are interested in the case of curves which may not be reduced, we cannot use this definition of saturation.
Following \cite{MGranger-equisingularite, FPham-BTeissier-fractions, Stutz-equising}, one knows that a flat deformation of a reduced plane curve is equisaturated if and only if the resulting surface is Whitney regular and the generic plane projection of the fibers has a fixed topological type.
We can use the latter condition as an equisingularity criterion instead of equisaturation.
More precisely, let $(S,0)$ be a germ of a complex surface singularity with a one-dimensional smooth singular locus. Via a projection to
$\mathbb{C}$ we can view this surface as a one-parameter deformation of a curve. We say that $(S,0)$ is strongly equisingular along its singular locus near the origin if it is Whitney regular and the topological type of a generic planar projection of the curves is constant; see \cite{MGranger-equisingularite}.
\begin{example}\label{whitneynosaturado}
\end{example}
Consider the surface given by the parametrization:
$$(a,t) \mapsto (a, t^4, at^6, t^7)$$
It is a non-Cohen-Macaulay surface with one-dimensional smooth singular locus.
One can see that the limit of tangent spaces at the origin is unique; it is given by the linear space spanned by the vectors $(1,0,0,0)$ and $(0,1,0,0)$. For sequences of points $p_n =(a_n, 0, 0, 0)\in {\rm Sing}(S)$ and $q_n = (a_n, t_n^4, a_nt_n^6, t_n^7)\in S\setminus {\rm Sing}(S)$, the corresponding secants $l_n= (p_nq_n)$ converge to the line $l$ spanned by $(0,1, 0, 0)$. So the limit of tangent spaces contain the tangent line to the singular locus and the limit $l$. The surface is then Whitney regular.
However a generic plane projection will have the following parametrization
$$x = t^4, \ y= at^6 + t^7$$
The characteristic exponents change when $a$ takes the value $0$.
This surface is not strongly equisingular.
\section{Equisingularity criteria are not stable under modifications}
One of our early motivations in this work was to investigate equisingularity criteria possibly stable under Nash modification in dimension two.
We obtained two negative examples, showing that neither Whitney regularity nor strong equisingularity is stable under Nash modification or the blow-up of the singular locus; {\it i.e.} the surface obtained by the modification does not satisfy the equsingularity criterion satisfied by the original surface.
Let us first recall the definition of Nash modification and establish some properties in case of equisingular surfaces.
Let $(X,0) \subset (\mathbb{C}^N,0)$ be a germ of reduced equidimensional analytic space of dimension $d$.
Call ${\bf G}(d,N)$ the Grassmannian of $d$-dimensional linear subspaces of $\mathbb{C}^N$.
Let us denote by $\gamma : X \setminus {\rm Sing}(X) \rightarrow {\bf G}(d,N)$ the Gauss map that associates to any non-singular point
$x$ in a representative of $(X,0)$ the direction of the tangent space $T_xX$.
Call ${\tilde X}$ the closure of the graph of $\gamma$ in $X \times {\bf G}(d,N)$.
The induced map
$$\nu : {\tilde X} \rightarrow X$$
is, by definition, the Nash modification. For its properties, one can read \cite[Chap. II]{BTeissier-rabida}, \cite{Nobile-nash} and for its desingularization properties see \cite{MSpivakovsky-sandwich, GGonzalez-Sprinberg-doubles}, and also \cite{PGonzalez-BTeissier-nash-toric, DDuarte-nash-toric-preprint}.
Recall that if a subspace $Y\subset X$ is defined by an ideal $I = (f_1, \cdots, f_r)$, then the blow-up of $X$ along $Y$ is obtained as the closure of the graph of the map
$$\begin{array}{rcl}
X \setminus Y & \rightarrow & \mathbb{P}^{r-1}\\
x & \mapsto & (f_1(x) : \cdots : f_r(x))
\end{array}
$$
Assume that we have a germ of reduced and irreducible analytic surface $(S,0)$ with one-dimensional smooth singular locus. Furthermore, assume it is Whitney regular in the sense of definition \ref{whitney}.
\begin{proposition}
A surface $(S,0)$ as above admits a bijective parametrization $n: (\mathbb{C}^2, 0) \rightarrow (S,0)$ that factors through the Nash modification and the blow-up of the singular locus.
\end{proposition}
\begin{proof}
We have seen in \ref{smoothnormalization} that the normalization of such a surface is smooth.
Since the surface $(S,0)$ is irreducible, the inverse image of the origin by the normalization map is a single point.
So, up to an analytic isomorphism, the normalization induces a finite map $(\mathbb{C}^2, 0) \rightarrow (S,0)$, that is an isomorphism outside the singular locus.
Since the surface is topologically trivial along its singular locus, it is analytically irreducible at each point of the singular locus. So, the normalization induces a homeomorphism $(\mathbb{C}^2,0) \rightarrow (S,0)$.
In \cite[V. I.2]{BTeissier-rabida} it is shown that for a Whitney regular surface, both Nash modification and the blow-up of the singular locus are finite maps. So, by universal property of the normalization, the normalization map factors through Nash modification, and through the blow-up of the singular locus.
\end{proof}
\begin{example}\label{whitneyunstable}
\end{example}
We give an example of an irreducible Whitney regular surface whose Nash modified surface and blow-up along the singular locus are no longer Whitney regular.
Consider the parametrized surface of Example \ref{whitneynosaturado}
$$\begin{array}{rcl}
n: \mathbb{C}^2 & \rightarrow & S \\
(a, t) & \mapsto & (a, t^4, at^6, t^7)\\
\end{array}$$
We have already seen that it is a Whitney regular surface.
The blow-up of the singular locus is given by the parametrization
$$\tau: (a, t) \mapsto (a, t^4, at^2, t^3).$$
Consider the sequence of points $p_n = \tau(\frac{1}{n}, 0)$ and $q_n = \tau(\frac{1}{n}, \frac{1}{n})$, with $n$ a positive integer.
The sequence of corresponding lines $(p_nq_n)$, converges to the line spanned by $(0, 0, 1, 1)$.
The tangent spaces $T_{q_n}S$ are spanned by the vectors $(1, 0, \frac{1}{n^2}, 0)$ and $(0, \frac{4}{n^3}, \frac{2}{n^2},
\frac{3}{n^2})$. They converge to the linear space spanned by $(1,0, 0, 0)$ and $(0, 0, 2, 3)$; it does not contain the vector
$(0, 0, 1, 1)$.
The Nash modification with Plucker coordinates is given by the parametrization
$$\sigma: (a, t) \mapsto (a, t^4, \frac{3}{2} at^2, \frac{7}{4} t^3)$$
With the sequence of points $p_n = \sigma(\frac{1}{n}, 0)$ and $q_n = \sigma (\frac{1}{n}, \frac{1}{n})$, we obtain as limit of lines, the one spanned by $(0, 0, 6, 7)$; and as limit of directions of tangent spaces the plane spanned by $(1, 0, 0, 0, 0)$ and
$(0, 0, 12, 21)$. The limit of lines is not contained in the limits of tangent spaces.
So, neither Nash modification nor the blow-up of the singular locus is Whitney regular.
\begin{remark}
Example \ref{whitneyunstable} shows that one can obtain a finite map from a surface to $\mathbb{C}^2$ whose discriminant locus is smooth, but the surface does not satisfy Zariski discriminant criterion. In fact, the original surface is Whitney regular and hence satisfies Zariski's discriminant criterion. There exists then a finite generic projection $\pi : S \rightarrow \mathbb{C}^2$ whose discriminant locus is non-singular. Composing $\pi$ with any of the modifications considered is still a finite map with the same discriminant locus. But it is not Whitney regular. So a non-generic finite map can have a smooth discriminant locus while the generic one has a singular discriminant locus.
\end{remark}
\begin{example}\label{equisaturationunstable}
\end{example}
We give here a surface that is strongly equisingular, however, the surfaces obtained by Nash modification and the blow-up of the singular locus are no more strongly equisingular.
Consider the surface $S$ given by the parametrization:
$$n: (a, t) \mapsto (a, t^5, t^8, at^9).$$
It is not difficult to check that it is a Whitney regular surface. Furthermore, the plane curve parametrized by
$$x= t^5, \, y= t^8 + a t^9$$
has a topological type that does not depend on the parameter $a$. So the surface is strongly equisingular along its singular locus.
The parametrization/normalization $n$ factors through the blow-up of the singular locus and through the Nash modification.
The blow-up of the singular locus is given by
$$\tau : (a, t) \mapsto (a, t^3, at^4, t^5)$$
and the Nash modification is given by
$$\sigma : (a,t) \mapsto (a, t^5, \frac{8}{5} t^3, \frac{9}{5} at^4)$$
Now consider $a$ as a parameter. The generic projection of a curve to $\mathbb{C}^2$ has the following parametrization:
$$x = t^3, \, y= at^4 + t^5$$
whose topological type changes depending on whether $a=0$ or $a\neq 0$.
So the modified surfaces are not strongly equisingular.
\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}
Arrays are widely used to model parts of systems. In software model checking,
for example, the heap of a program can be modelled by an array that represents
the main memory. A software model checker using such a model can check for
illegal accesses to memory or even memory leaks. While checking for illegal
accesses can be done using only the axioms proposed by McCarthy, leak checking
typically is done using extensionality. In this setting, extensionality is
used to ensure that the memory after executing a program does not contain more
allocated memory cells than it contained at the beginning of the program.
The theory of arrays was initially proposed by
McCarthy~\cite{DBLP:conf/ifip/McCarthy62}. It specifies two operations: (1)
The store operation \store{a}{i}{v} creates a new array that stores at every
index different from $i$ the same value as array $a$ and the value $v$ at
index $i$. (2) The select operation \select{a}{i} retrieves the value of
array $a$ at position $i$. The theory is parametric in the index and element
theories.
The store operation only modifies an array at one index. The values stored at
other indices are not affected by this operation. Hence, the resulting
array and the array used in the store operation are \emph{weakly equal}
in the sense that
they differ only at finitely many indices.
Current decision procedures do not fully exploit such dependencies between
arrays. Instead, they use a series of instantiations of the axiom proposed by
McCarthy to derive weak equivalences.
In this paper we present a new algorithm to decide the quantifier-free
fragment of the theory of arrays. The decision procedure is based on the
notion of \emph{weak equivalence}, a property that combines equivalence
reasoning with array dependencies. The new algorithm only produces a few new
terms not present in the input formula during preprocessing. This is possible
since the decision procedure does not instantiate the axiom proposed by
McCarthy, but axioms derived from them.
\paragraph*{Related Work}
Since the proposal of the theory of arrays by
McCarthy~\cite{DBLP:conf/ifip/McCarthy62} several decision procedures have
been proposed. We can identify two basic branches: \emph{rewrite-based} and
\emph{instantiation-based} techniques.
Armando et al.~\cite{DBLP:journals/tocl/ArmandoBRS09} used
rewriting techniques to solve the theory of arrays. They showed how to
construct simplification orderings to achieve completeness. The benchmarks
used in this paper test specific properties of the array operators like
commutativity of stores if the indices differ. While these benchmarks require
a lot of instantiations of McCarthy's axioms, they are easy for the decision
procedure presented in this paper since the properties tested by these
benchmarks are properties satisfied by the weak equivalence relation presented
in this paper.
Bruttomesso et al.~\cite{DBLP:journals/corr/abs-1204-2386} present a rewrite
based decision procedure to reason about arrays. This approach exploits some
key properties of the store operation that are also captured by the weak
equivalence relation described in this paper. Contrary to our method, the
rewrite based approach is not designed for Nelson--Oppen style theory
combination and thus not easily integratable into an existing SMT solver.
They extended the solver into an interpolating solver for computing
quantifier-free interpolants. In contrast to our method their solver depends
on the partitioning of the interpolation problem. We create a SMT proof
without any knowledge of the partitioning and can use proof tree preserving
interpolation~\cite{DBLP:conf/tacas/ChristHN13}, which only requires a
procedure to interpolate the lemmas.
A decision procedure for the theory of arrays based on instantiating
McCarthy's axioms is given by de Moura et
al.~\cite{DBLP:conf/fmcad/MouraB09}. The decision procedure saturates several
rules that instantiate array axioms under certain conditions. Several filters
are proposed to minimise the number of instantiations.
Closest to our work is the decision procedure published by Brummayer et
al.~\cite{DBLP:journals/jsat/BrummayerB09}. Their decision procedure produces
lemmas that can be derived from the axioms for the theory of arrays proposed
by McCarthy. They consider the theory of arrays with bitvector indices and
prove soundness and completeness of their approach in this setting. In
contrast to our method, they do not allow free function symbols (i.~e., the
combination of the theory of arrays with the theory of uninterpreted function
symbols) since they only consider a limited form of extensionality where the
extensionality axiom is only instantiated for arrays $a$ and $b$ if the
formula contains the literal $a\neq b$. We do not have this limitation, but
add some requirements on the index theory that prevent the procedure presented
in this paper from using the theory of bitvectors as index theory.
\section{Notation}\label{sec:notation}
A first order theory consists of a signature $\Sigma$ and a set of models
$\mathbb{M}$. We assume the equality symbol $=$ with its usual interpretation
is part of any signature. Every model contains for every sort interpreted by
this model a non-empty domain and a mapping from constant or function symbol
into the corresponding domain. A theory \T is \emph{stably infinite} if and
only if every satisfiable quantifier-free formula is satisfied in a model of
\T with an infinite universe.
The theory of arrays \ta is parameterised by an index theory \ti and an
element theory \te. The signature of \ta consists of the two functions
\select{\cdot}{\cdot} and \store{\cdot}{\cdot}{\cdot}. Every model of the
theory of arrays satisfies the select-over-store-axioms proposed by
McCarthy~\cite{DBLP:conf/ifip/McCarthy62}:
\begin{align}
\forall a\,i\,v.\ \select{\store{a}{i}{v}}{i} & = v\tag{idx}\label{ax:idx}\\
\forall a\,i\,j\,v.\ i\neq j\implies\select{\store{a}{i}{v}}{j} & =
\select{a}{j}\tag{read-over-write}\label{ax:read-over-write}
\end{align}
Additionally we consider the extensional variant of the theory of arrays.
Then, every model has to satisfy the extensionality axiom:
\begin{align}
\forall a\,b.\ a = b \lor\exists i.\ \select{a}{i} \neq
\select{b}{i}\tag{ext}\label{ax:ext}
\end{align}
We use $a,b$ to denote array-valued variables, $i,j,k$ to denote index
variables, and $v,w$ to denote element variables. Additionally we use
subscripts to distinguish different variables. We use $P$ to denote a
path in a graph. A path in a graph is interpreted as a sequence of
edges.
In the remainder of this paper, we consider quantifier-free \ta-formulae.
Furthermore we fix the index \ti to a stably infinite theory and the element
theory \te to a theory that contains at least two different
values\footnote{Note that \ta is stably infinite under these conditions.}.
\section{Towards a Nelson--Oppen-based Array Solver}
Multiple theories are usually combined with a variant of the Nelson--Oppen
combination procedure~\cite{DBLP:journals/toplas/NelsonO79}. The procedure
requires the participating theories to be stably infinite and to only share
the equality symbol $=$.
The procedure first transforms the input such that every literal is
\emph{pure} with respect to the theories. Let $f(t)$ be a term in the
input. If $f$ is interpreted by theory $\mT_1$ and $t$ is interpreted
by theory $\mT_2$, then $f(t)$ is not pure. The first step of the
Nelson--Oppen procedure then generates a fresh variable $v$, rewrites
$f(t)$ into $f(v)$, and adds the definition $v=t$ as a new conjunct to
the formula. The fresh variable is shared between theories $\mT_1$
and $\mT_2$. This step is repeated until all terms are pure. By
abuse of notation, we name the shared variable after its
defining term $t$, e.\,g., we use $\select{a}{i}$ to denote the shared
variable that is defined as $\select{a}{i}$.
Let $V$ be the set of fresh variables introduced in the first step of the
combination procedure. The second step of the procedure tries to find an
\emph{arrangement} of $V$, i.~e., an equivalence relation between variables in
$V$ such that $\mT_1$ and $\mT_2$ produce partial models that agree with this
equivalence relation. Finding such an arrangement is typically done by
propagating equalities or providing case split lemmas. In the following, we
call this arrangement strong equivalence to distinguish it from weak
equivalence defined in the next section. We write $a \strongeq b$ to
denote that $a$ and $b$ are strongly equivalent, i.e., that in the current
arrangement the shared variables $a$ and $b$ are equal.
For the theory of arrays, we consider every term of the form
$\store{\cdot}{\cdot}{\cdot}$ or $\select{\cdot}{\cdot}$ as being
interpreted by the array theory.
We consider all array terms, store, and select terms to be shared and
thus they have to occur in the arrangement.
Furthermore, every index term that
appears in a store or select is considered shared between the array
theory and the index theory. Then the goal is to find a suitable
arrangement to these shared terms such that all theories agree on this
arrangement.
For an array solver to be used in Nelson--Oppen combination we have to
propagate equalities between shared array terms and shared select terms.
Furthermore, the other theories have to propagate equalities between terms
used as index in a select or store. In the remainder of this paper we will
first show how to propagate equalities between select terms and afterwards
deal with extensionality to propagate equalities between array-valued terms.
\section{Weak Equivalences over Arrays}
The theory of arrays has two constructors for arrays: array variables, and
store terms $\store{\cdot}{\cdot}{\cdot}$. Assuming quantifier-free input, we can only
constrain the values of a finite number of indices. These constraints can
either be explicity like $\select{a}{i}=v$, or implicit like $\store{a}{i}{v}$
where axiom (\ref{ax:idx}) produces the corresponding
$\select{\store{a}{i}{v}}{i}=v$. Hence, for quantifier-free input, arrays
that are connected via a sequence of $\store{\cdot}{\cdot}{\cdot}$ can only
differ in finitely many positions. We call such arrays \emph{weakly
equivalent}. In this section we formally define weak equality and show how
to exploit this to produce a decision procedure for the (extensional) theory of
arrays.
Let \ST be the set of all terms of the form \store{\cdot}{\cdot}{\cdot} in the
input formula and \A be the set of all array-valued terms that are not in \ST.
Since $\store{a}{i}{v}$ modifies $a$ only at index $i$, these two arrays are
guaranteed to be equal on all indices except on index $i$. We generalise this
observation to chains of the form $\store{\store{\ldots}{j}{w}}{i}{v}$ to
extract a set of indices for which two arrays might store different values.
\begin{definition}[weak equivalence]\label{def:weakeq}
A \emph{weak equivalence graph} $G^W$ contains vertices
$\ST\cup\A$ and undirected edges defined as follows:
\begin{enumerate}
\item $a\leftrightarrow b$ if $a\strongeq b$, and
\item $a\stackrel{i}{\leftrightarrow} b$ if $a$ has form
\store{b}{i}{\cdot}.
\end{enumerate}
We write $a\weakpath{P} b$ if there exists a path $P$ between nodes $a$ and
$b$ in $G^W$. In this case, we call $a$ and $b$ \emph{weakly equal}. The weak
equivalence class containing all elements that are weakly equal to $a$ is
defined as $\Weakeq{a} := \{b\ |\ \exists P.\ a\weakpath{P} b\}$.
\end{definition}
For a path $P$ we define $\Stores{P}$ as the set of all indices corresponding
to edges of the form $\stackrel{\cdot}{\leftrightarrow}$, i.~e.,
$\Stores{P}:=\{i\ |\ \exists a\,b.\ a\stackrel{i}{\leftrightarrow} b\in P\}$.
\begin{example}\label{ex:weakeq}
Consider the formula $a=\store{b}{j}{v}\land b=\store{c}{i}{w}\land d=e\land
\select{c}{i}=w$. The weak equivalence graph for this example is shown in
Figure~\ref{fig:weakeq}. Note that the last conjunct is not important for
the construction of the weak equivalence graph.
\begin{figure}[htpb]
\centering
\begin{tikzpicture}[<->,node distance=1.5cm]
\node (d) at (0,0) {$d$};
\node[right of=d] (e) {$e$};
\node[above of=d,node distance=.5cm] (b) {$b$};
\node[right of=b] (storec) {$\store{c}{i}{w}$};
\node[right of=storec] (c) {$c$};
\node[left of=b] (storeb) {$\store{b}{j}{v}$};
\node[left of=storeb] (a) {$a$};
\draw (d) -- (e);
\draw (a) -- (storeb);
\draw (storeb) --node[above]{$j$} (b);
\draw (b) -- (storec);
\draw (storec) --node[above]{$i$} (c);
\end{tikzpicture}
\caption{\label{fig:weakeq}Weak Equivalence Graph for
Example~\ref{ex:weakeq}}
\end{figure}
We get two different weak equivalence classes. The first one contains the
nodes $a$, $\store{b}{j}{v}$, $b$, $\store{c}{i}{w}$, and $c$. The second
contains $d$ and $e$. Note that $d$ and $e$ are actually strongly equivalent.
Thus, they store the same value at every position. Let $P$ denote the path
from $a$ to $c$ in the weak equivalence graph. Then, $\Stores{P}=\{i,j\}$.
Thus, arrays $a$ and $c$ can only differ in at most the values stored at the
indices $i$ and $j$.
\end{example}
If we want to know if $\select{a}{i}$ and $\select{b}{i}$ should be equal, we
check if $a\weakpath{P}b$ for a path $P$ such that $i\not\in\Stores{P}$. If
this is the case, $P$ witnesses the equivalence between the select terms.
\begin{definition}[weak equivalence modulo $i$]\label{def:weakeqi}
Two arrays $a$ and $b$ are \emph{weakly equivalent modulo~$i$} if and only if
they are weakly equivalent and connected by a path that does not contain an
edge of the form $\stackrel{j}{\leftrightarrow}$ where $j\strongeq i$. We
denote weak equivalence modulo~$i$ by $a\weakeqi b$ and define it as
$a\weakeqi b := \exists P.\ a\weakpath{P} b\land \forall
j\in\Stores{P}.\ j\not\strongeq i$.
\end{definition}
Using this definition we can propagate equalities between shared selects if
the arrays are weakly equivalent modulo the index of the select.
\begin{lemma}[read-over-weakeq]\label{lem:read-over-weakeq}
Let $\strongeq$ be an arrangement satisfying the array axioms.
Let $\select{a}{i}$ and $\select{b}{j}$ be two selects such that $i\strongeq
j$ and $a\weakeqi b$. Then, $\select{a}{i} \strongeq \select{b}{j}$ holds.
\end{lemma}
\begin{proof}
We induct over the length of the path $P$ witnessing $a\weakeqi b$.
\paragraph{Base case.}
In this case, $a$ and $b$ are the same term. Hence
$\select{a}{i}\strongeq\select{b}{j}$ holds by congruence.
\paragraph{Step case.}
Let the step from $c$ to $b$ be the last step of path $P$.
By induction hypothesis we know that
$\select{a}{i}\strongeq\select{c}{j}$ holds.
If the edge between $c$ and $b$ is due to a strong equivalence (i.~e.,
$c\strongeq b$), then $\select{c}{j}\strongeq\select{b}{j}$ follows from
congruence.
If the edge between $c$ and $b$ is of the form
$c\stackrel{k}{\leftrightarrow} b$, then either $c$ is $\store{b}{k}{\cdot}$
or $b$ is $\store{c}{k}{\cdot}$. In both cases, we get the lemma
$j=k\lor\select{c}{j}=\select{b}{j}$ from axiom (\ref{ax:read-over-write}).
Since $j\strongeq i$ and $i\not\strongeq k$, we get $j\not\strongeq k$ and
thus $\select{c}{j}\strongeq\select{b}{j}$. We conclude
$\select{a}{i}\strongeq\select{b}{j}$ by transitivity.
\end{proof}
This lemma allows us to propagate equalities between shared selects. Note
that it depends upon disequalities between index terms needed to ensure
$a\weakeqi b$.
If two arrays are weak equivalent modulo~$i$ they store the same value at the
index $i$. The reverse is not necessarily true. Therefore, we define a
weaker relation weak congruence modulo~$i$.
\begin{definition}[weak congruence modulo~$i$]\label{def:weakcongi}
Arrays $a$ and $b$ are \emph{weak congruent modulo~$i$} if and only if they
are guaranteed to store the same value at index $i$. We denote weak
congruence modulo~$i$ by $\weakcongi$ and define $a\weakcongi b := a\weakeqi
b\lor\exists a'\,b'\,j\,k.\ a\weakeqi
a'\land i\strongeq j\land
\select{a'}{j}\strongeq\select{b'}{k}\land k \strongeq i
\land b'\weakeqi b$.
\end{definition}
We use weak congruences to decide extensionality. Intuitively, if for all
indices $i$ the weak congruence modulo~$i$ $a\weakcongi b$ holds, then $a=b$
should be propagated. But this na\"{\i}ve approach requires checking every
index occurring in the formula. To minimise the number of indices we need to
consider, we exploit the weak equivalence graph.
\begin{lemma}[weakeq-ext]\label{lem:weakeq-ext}
Let $\strongeq$ be an arrangement satisfying the array axioms.
Let $a$ and $b$ be two arrays such that $a\weakpath{P}b$ holds. If for all
indices $i\in\Stores{P}$ we have $a\weakcongi b$, then $a\strongeq b$ holds.
\end{lemma}
\begin{proof}
Follows from Lemma~\ref{lem:read-over-weakeq},
Definition~\ref{def:weakcongi} and (\ref{ax:ext}).
\end{proof}
\section{A Decision Procedure Based on Weak Equivalences}
Our decision procedure is based on weak equivalences and the
Nelson--Oppen combination scheme. It propagates equalities between
terms shared by multiple theories. We limit the propagation to shared array
terms and array select terms.
The \ta-formulae are preprocessed as follows. For every
\store{a}{i}{v} contained in the input, we (1) instantiate
the axiom (\ref{ax:idx}) and (2) add \select{a}{i} to the set of terms
contained in the input\footnote{This can be achieved by adding the
equality $\select{a}{i} = \select{a}{i}$.}. Thus, the preprocessing
step adds at most two select operations for every store.
We propagate new equalities from weak equivalence relations and weak
congruence relations based on lemmas
\ref{lem:read-over-weakeq}~and~\ref{lem:weakeq-ext}. These relations
depend on the arrangement $\strongeq$, which represents logical
equality ($=$). We now define a function $\Cond\cdot$ that computes a
condition (a conjunction of equalities and inequalities) under which
a weak equivalence or weak congruence holds. To denote
the condition for a path that does not contain an edge labelled with
index $i$ we use $\Condi{\cdot}$. For an edge in the weak equivalence
graph that represents an equality, the condition reflects this equality.
For an edge that comes from a $\store{\cdot}{j}{\cdot}$, no condition
is needed. However, $\Condi{\cdot}$ should ensure that $i$ does not
occur on the path, so $i \neq j$ needs to hold.
\begin{align*}
\Cond{a\leftrightarrow b} &:= a=b &
\Condi{a\leftrightarrow b} &:= a=b\\
\Cond{a\stackrel{j}{\leftrightarrow} b} &:= \mathrm{true} &
\Condi{a\stackrel{j}{\leftrightarrow} b} &:= i\neq j
\end{align*}
We can extend these definitions to paths by conjoining the conditions for all
edges on that path. Then, we can compute $\Cond{a\weakeqi b}$ using the path
that witnesses $a\weakeqi b$.
\[
\Cond{a\weakeqi b} := \Condi{P}\text{ where }a\weakpath{P}b\land\forall
j\in\Stores{P}.\ i\not\strongeq j
\]
Finally, to define $\Cond{a\weakcongi b}$, we use the definition of
$\weakcongi$.
\[
\Cond{a\weakcongi b} :=\begin{cases}
\Cond{a\weakeqi b}&\text{if }a\weakeqi b\\
\begin{array}{l}
\Cond{a\weakeqi a'}\land i = j \land\select{a'}{j}=\select{b'}{k}\\
\quad{}\land k = i\land\Cond{b'\weakeqi b}
\end{array}
&\text{if }
\begin{array}{l}
a\weakeqi a'\land i\strongeq j \land \select{a'}{j}\strongeq\select{b'}{k}\\
\quad{}\land k \strongeq i \land b'\weakeqi b
\end{array}
\end{cases}
\]
\begin{example}\label{ex:condi}
Consider again the formula $a=\store{b}{j}{v}\land b=\store{c}{i}{w}\land
d=e\land \select{c}{i}=w$ from Example~\ref{ex:weakeq} whose weak
equivalence graph is shown in Figure~\ref{fig:weakeq}. Assume $i\not\strongeq j$.
Then we have $a\weakeqi\store{c}{i}{w}$ since no edge contains a label that
is equivalent to $i$. We get $\Cond{a\weakeqi\store{c}{i}{w}}\equiv
a=\store{b}{j}{v}\land i\neq j\land b=\store{c}{i}{w}$.
From Axiom~(\ref{ax:idx}) we get $\select{\store{c}{i}{w}}{i}=w$. With
$\select{c}{i}=w$ we conclude $a\weakcongi c$ since
$a\weakeqi\store{c}{i}{w}$ and $\select{\store{c}{i}{w}}{i} =
\select{c}{i}$. We have $\Cond{a\weakcongi c} \equiv
\Cond{a\weakeqi\store{c}{i}{w}}\land\select{\store{c}{i}{w}}{i} =
\select{c}{i}$.
\end{example}
To decide the theory of arrays we define two rules to generate
instances of array lemmas. We present the rules as inference rules.
The rule is applicable if the current arrangement $\strongeq$ on
the shared variables $V$
satisfies the conditions above the line. The rule then generates a
new (valid) lemma that can propagate an equality under the current
arrangement.
The first rule is based on Lemma~\ref{lem:read-over-weakeq}. Two select
terms are equivalent if the indices of the selects are congruent and the
arrays are weakly equivalent modulo that index. We only create this lemma
if the select terms existed in the formula. Note that we create for
select terms in the formula a shared variable with the same name in $V$.
\[
\inferrule*{a\weakeqi b\\i\strongeq j\\\select{a}{i},\select{b}{j}\in V}
{i\neq j\lor\lnot\Cond{a\weakeqi b}\lor
\select{a}{i}=\select{b}{j}}
\tag{read-over-weakeq}\label{rule:read-over-weakeq}
\]
The next rule is based on Lemma~\ref{lem:weakeq-ext} and used to
propagate an equality between two extensionally equal array terms. Two
arrays $a$ and $b$ have to be equal if there is a path $P$ such that
$a\weakpath{P}b$ and for all $i\in\Stores{P}$, $a\weakcongi b$ holds.
\[
\inferrule*{a\weakpath{P}b\\ \forall i\in\Stores{P}.\ a\weakcongi b\\a,b\in V}
{\lnot\Cond{P}\lor\bigvee_{i\in\Stores{P}}\lnot\Cond{a\weakcongi b}
\lor a=b}
\tag{weakeq-ext}\label{rule:weakeq-ext}
\]
The resulting decision procedure is sound and complete for the existential
theory of arrays assuming sound and complete decision procedures for the index
and element theories.
\begin{lemma}[soundness]\label{lem:soundness}
Rules (\ref{rule:read-over-weakeq}) and (\ref{rule:weakeq-ext}) are sound.
\end{lemma}
\begin{proof}
Soundness of the rules follows directly from the lemma with the
corresponding name.
\end{proof}
\begin{restatable}[completeness]{lemma}{completeness}\label{lem:completeness}
The rules (\ref{rule:read-over-weakeq}) and (\ref{rule:weakeq-ext}) are
complete.
\end{restatable}
\begin{longversion}
\begin{proof}
Assume all rules are saturated. Let \M be the model generated by
the theories different from the array theory. In this model, arrays
are considered uninterpreted and only subject to congruence. Note
that $v_1\strongeq v_2$ if and only if $\M(v_1)=\M(v_2)$ is
guaranteed in this model. We create a new model $\M_A$ that extends
\M by the interpretation of the array terms in the formula.
First, for every array type $\sigma\Rightarrow\tau$, we define its domain as
the set of all functions from $\sigma$ to $\tau$. The interpretation of
$\select{\cdot}{\cdot}$ is function application and the interpretation of
$\store{\cdot}{\cdot}{\cdot}$ is function update. This definition trivially
satisfies the array axioms (\ref{ax:idx}) and (\ref{ax:read-over-write}).
For every other function symbol $f$ and every constant $v$ that is not of array type
we define $\M_A(f) = \M(f)$ and $\M_A(v) = \M(v)$.
Next we define the interpretation for all constants of type array in $\M_A$,
such that $\M_A$ satisfies the input formula.
Let $\prec$ be a partial order on types such that for every type
$\sigma\Rightarrow\tau$ we have $\sigma\prec\sigma\Rightarrow\tau$ and
$\tau\prec\sigma\Rightarrow\tau$. We define the interpretation of constants
according to $\prec$. Thus, when defining the interpretation for a constant
of type $\sigma\Rightarrow\tau$ we assume all constants of type $\sigma$
resp.\ $\tau$ are already defined.
For each sort $\tau$ we assume two different values $\First_\tau$ and
$\Second_\tau$. Furthermore we assume every sort $\sigma$ that is used as
index sort for an array sort contains an infinite supply of fresh domain
elements denoted by $\fresh{\cdot}$. Then, for an array constant $a$ of sort
$\sigma\Rightarrow\tau$, we define
\[
\M_A(a)(\mbox{\j}):=\begin{cases}
\M(\select{b}{i})&\text{if }\select{b}{i}\text{ occurs in input, }\M(i)=\mbox{\j}\text{ and } a\weakeqi b\\
\Second_\tau&\text{if }\mbox{\j}=\fresh{\Weakeq{a}}\\
\First_\tau&\text{otherwise}
\end{cases}
\]
The first case is well defined. Given two
select terms $\select{b_1}{i_1}$ and $\select{b_2}{i_2}$ with
$a\weakeqi[i_1] b_1$, $a\weakeqi[i_2] b_2$ and $\M(i_1) = \mbox{\j} = \M(i_2)$ we
need to show $\M(\select{b_1}{i_1})=\M(\select{b_2}{i_2})$.
Since $\M(i_1)= \M(i_2)$, we have $i_1 \sim i_2$ and
$b_1\weakeqi[i_1] b_2$ (since $\weakeqi[i_1]$ and $\weakeqi[i_2]$ is
the same relation). The rule (\ref{rule:read-over-weakeq})
generated the lemma
$i_1\neq i_2\lor\lnot\Cond{b_1\weakeqi[i_1] b_2}\lor
\select{b_1}{i_1}=\select{b_2}{i_2}$.
\M guarantees that
$i_1=i_2$ and $\Cond{b_1\weakeqi[i_1] b_2}$ hold.
Thus, $\M(\select{b_1}{i_1})=\M(\select{b_2}{i_2})$ has to hold in
order to satisfy this lemma.
We have to show that the input formula is satisfied by $\M_A$ if it
is satisfied by $\M$.
We assume that the array operations in the input formula were
flattened by introducing fresh variables, i.\,e., that
$\select\cdot\cdot$ and $\store\cdot\cdot\cdot$ occur only in
definitions $v=\select{a}{i}$ and $b = \store{a}{i}{v}$. By
definition of $\M_A$ it already satisfies the parts of the input
formulae that do not involve arrays. It remains to show that
\begin{enumerate}
\item for all definition $v=\select ai$
$\M_A(v) = \M_A(\select ai)$ holds.
\item for all definition $b=\store a i v$
$\M_A(b) = \M_A(\store aiv)$ holds.
\item $\M_A(a) = \M_A(b)$ if and only if $\M(a) = \M(b)$ for array constants
$a$ and $b$, and
\end{enumerate}
\begin{enumerate}
\item For a definition $v=\select{a}{i}$ the model $\M$ already guarantees
$\M(v) = \M(\select{a}{i})$. The definition of $\M_A$ gives us
$\M_A(v) = \M(v) = \M(\select ai) = \M_A(a)(\M(i)) = \M_A(\select ai)$
as required.
\item For $b=\store{a}{i}{v}$ we need to show
$\M_A(b)(\mbox{\j}) = \M_A(a)(\mbox{\j})$ for $\mbox{\j}\neq \M(i)$ and $\M_A(b)(\M(i)) =
\M(v)$.
Our preprocessing step adds the equality $\select{\store{a}{i}{v}}{i}=v$
to the input. Thus, $\M_A(b)(\M(i)) = \M(\select{\store{a}{i}{v}}{i})
= \M(v)$. For $\mbox{\j}\neq \M(i)$, we can derive
$\M_A(b)(\mbox{\j}) = \M_A(a)(\mbox{\j})$ from $\Weakeq{b} = \Weakeq{a}$ and the fact
that $a\weakeqi[j] b$ for every $j\not\strongeq i$.
\item We show that our extended model satisfies equalities on arrays. If
$a\strongeq b$ holds for two arrays $a$ and $b$, then by construction
$\M_A(a)=\M_A(b)$. If $a\not\strongeq b$ holds for two arrays $a$ and $b$,
then we distinguish two cases. If $a$ and $b$ are not connected in the weak
equivalence graph, they differ at the indices $\fresh{\Weakeq{a}}$ and
$\fresh{\Weakeq{b}}$. Hence, $\M_A(a)\neq\M_A(b)$ holds. Otherwise, since
rule (\ref{rule:weakeq-ext}) is saturated and $a\weakpath{P} b$ for some
path $P$ but $a\neq b$, there exists an $i\in\Stores{P}$ such that
$a\not\weakcongi b$ holds.
Let $a'$ be the first array on the path $P$ that involves an edge
$\stackrel{i}\leftrightarrow$ and $b'$ the last such array.
In the preprocessing step we added $\select {a'}{i}$ and $\select {b'}{i}$.
By the choice of $a'$ and $b'$ we have $a\weakeqi a'$ and $b' \weakeqi b$.
Since $a\not\weakcongi b$ holds, we have $\select{a'}{i} \not
\strongeq\select{b'}{i}$. By definition $\M_A(a)(\M(i)) = \M(\select{a'}{i})$
and $\M_A(b)(\M(i)) = \M(\select{b'}{i})$, so $\M_A(a)$ and $\M_A(b)$ are
different arrays as desired.
\end{enumerate}
\end{proof}
\end{longversion}
\begin{shortversion}
\noindent The proof of this lemma can be found in the extended version of this
paper~\cite{longversion}.
\end{shortversion}
\section{Restricting Instantiations}
The preprocessor is the only component of our decision procedure that produces
new select terms and thus might trigger new lemmas. These lemmas only
generate new (dis-)equality literals between existing terms. Thus, reducing
the number of select terms might reduce the number of lemmas generated by our
decision procedure and speed up the procedure.
If the element theory is stably infinite we can omit the preprocessor step
that adds for every $\store{a}{i}{v}$ the select $\select{a}{i}$. Instead, we
simply assume $\select{a}{i}$ to be different than any other $\select{b}{i}$
unless $a\strongeq b$. This method preserves soundness and
completeness.
\begin{lemma}(soundness of modified procedure)
The modified procedure is sound.
\end{lemma}
\begin{proof}
Follows directly from Lemma~\ref{lem:soundness} since it does not rely on
the addition of $\select{a}{i}$ for every $\store{a}{i}{\cdot}$.
\end{proof}
For the completeness lemma we take into account the fact that the element
theory is stably infinite. Thus, if $\select{a}{i}$ is not present we use a
fresh element in the value domain.
\begin{restatable}[completeness of modified
procedure]{lemma}{modifiedcompleteness}\label{lem:modifiedcompleteness}
The modified procedure is complete.
\end{restatable}
\begin{longversion}
\begin{proof}
We redefine the generation of the model $\M_A$ in the proof of
Lemma~\ref{lem:completeness}.
Let $\Weakeqi{a} := \{b\ |\ b\weakeqi a\}$
denote the set of all array terms that are weakly equivalent modulo $i$ to
$a$. We construct $\M_A$ in the following way.
\[
\M_A(a)(\mbox{\j}):=\begin{cases}
\M(\select{b}{i})&\text{if } \select{b}{i}\text{ occurs in the input, } \M(i)=\mbox{\j}\text{ and }
a\weakeqi b\\
\fresh{\Weakeqi{a}}&\text{if }\M(i)=\mbox{\j}, i\in\Stores{\Weakeq{a}}\text{ and there are
no $b$ and $j$}\\
&\quad\text{such that }\select{b}{j}\text{ occurs in the input,
}\M(j)=\mbox{\j}\text{ and }a\weakeqi[j] b\\
\Second_\tau&\text{if }\mbox{\j}=\fresh{\Weakeq{a}}\\
\First_\tau&\text{otherwise}
\end{cases}
\]
Note that we only change the model for array $a$ and index $i$ if
$\store{a}{i}{\cdot}$ exists in the input, but $\select{a}{i}$ does not.
Hence, for definition of the form $v=\select{a}{i}$ and $b =
\store{a}{i}{v}$ the old proof can be reused. Thus, we only need to
show extensionality.
If $\M(a)=\M(b)$ for two array terms $a$ and $b$, then $a\strongeq
b$. Thus, for all indices $i$, $\M_A(a)(i)=\M_A(b)(i)$ holds by
construction. Otherwise we distinguish two cases. If $a$ and $b$
are not connected in the weak equivalence graph, $\M_A(a)$,
$\M_A(b)$ differ, e.\,g., on index $\fresh{\Weakeq{a}}$. If the
arrays are connected ($a\weakpath{P} b$), there is an
$i\in\Stores{P}$ such that $a \not\weakcongi b$. If there is no
$\select{a'}{j}$ with $a\weakeqi a'$ and $j \strongeq i$, then
$\M_A(a)(\M(i)) = \fresh{\Weakeqi{a}}$ differs from
$\M_A(b)(\M(i))$. Similarly for $b$, if there is no $\select{b'}{k}$ with
$b\weakeqi b'$ and $k\strongeq i$.
Otherwise $\M_A(a)(\M(i)) = \M(\select{a'}{j})$ and
$\M_A(b)(\M(i)) = \M(\select{b'}{k})$ and these values differ since
$a\not\weakcongi b$.
\end{proof}
\end{longversion}
\begin{shortversion}
\noindent The proof of this lemma can be found in the extended version of this
paper~\cite{longversion}.
\end{shortversion}
This optimisation enables us to limit the number of additional terms in the
input. Since we only need to generate (\ref{rule:read-over-weakeq}) lemmas if
the select terms in the conclusion are present after preprocessing, this
optimisation also reduces the number of lemmas. Furthermore, it is widely
applicable. In fact, the non-bitvector logics defined in the
SMTLIB~\cite{BarST-SMT-10} only allow array sorts where the element theory is
stably infinite. Thus, only the terms corresponding to instantiations of
Axiom~(\ref{ax:idx}) are required. In an actual implementation even these
terms could be omitted (see~\cite{DBLP:journals/jsat/BrummayerB09}).
\section{Implementation and Evaluation}\label{sec:implementation}
We implemented the decision procedure described in this paper in our SMT
solver SMTInterpol~\cite{DBLP:conf/spin/ChristHN12}.
Besides the aforementioned preprocessing step that applies (\ref{ax:idx}) to
every \store{\cdot}{\cdot}{\cdot} in the input, we also simplify \ta-formulae by
applying (\ref{ax:read-over-write}) if the index of the store and the index of
the select are syntactically equal. Furthermore, we contract terms of the
form $\store{\store{a}{i}{v_2}}{i}{v_1}$ to $\store{a}{i}{v_1}$. We only add
\select{a}{i} to the set of terms contained in the formula if we have
\store{a}{i}{v} in the input and the domain of $v$ is finite.
We represent the weak equivalence relation and the weak equivalence
modulo~$i$ relations in a forest structure, similarly to the
representation of equivalence graph in congruence
solvers~\cite{DBLP:conf/rta/NieuwenhuisO05}. Every node has an
outgoing edge, and these edges build a spanning tree for every
equivalence class. The edges point from a child node to the parent
node. The root node of every tree has no outgoing edge and is the
representative of its equivalence class.
We have to distinguish between strong equivalence, weak equivalence,
and weak equivalence modulo~$i$. The strong equivalence classes are
already handled by the equality solver. In our implementation of the
array solver we treat them as indivisible and create a single node for
every strong equivalence class. To represent the weak equivalence
relations the nodes have up to two outgoing edges, a primary $p$ and a
secondary $s$, see Figure~\ref{alg:data-structures}. The edges come
from a store operation and correspond to the edges
$\stackrel{i}{\leftrightarrow}$ in the weak equivalence graph. The
index of the primary edge is stored in the $pi$ field.
\begin{figure}[htbp]
\begin{minipage}{.5\textwidth}
\begin{tabbing}
\textbf{struct} $\type{node}$\\
\qquad\=
$p : \type{node}$\\
\> $pi : \type{index}$\\
\> $s : \type{node}$
\end{tabbing}
\begin{codebox}
\Procname{$\proc{get-rep}(n : \type{node})$}
\zi \If $n.p = \const{nil}$ \Then $n$
\zi \Else $\proc{get-rep}(n.p)$
\End
\end{codebox}
\begin{codebox}
\Procname{$\proc{make-rep}(n : \type{node})$}
\zi \If $n.p \neq \const{nil}$ \Then
\zi $\proc{make-rep}(n.p)$
\zi $\left.\begin{array}{@{}l@{}}
n.p.p := n \\
n.p.pi := n.pi\\
n.p := \const{nil}
\end{array}\right\}\mbox{\parbox{2.5cm}{\raggedright invert primary edge}}$
\zi $\proc{make-rep${}_i$}(n)$
\End
\end{codebox}
\end{minipage}%
\begin{minipage}{.5\textwidth}
\begin{codebox}
\Procname{$\proc{get-rep${}_i$}(n : \type{node}, i : \type{index})$}
\zi \If $n.p = \const{nil}$ \Then $n$
\zi \ElseIf $n.pi \neq i$ \Then $\proc{get-rep${}_i$}(n.p, i)$
\zi \ElseIf $n.s = \const{nil}$ \Then n
\zi \Else $\proc{get-rep${}_i$}(n.s, i)$
\End
\end{codebox}
\begin{codebox}
\Procname{$\proc{make-rep${}_i$}(n : \type{node})$}
\zi \If $n.s \neq \const{nil}$ \Then
\zi \If $n.s.pi \neq n.pi$ \Then
\zi $n.s := n.s.p$\quad\mbox{\parbox{2.2cm}{move towards representative}}
\zi $\proc{make-rep${}_i$}(n)$
\zi \Else
\zi $\proc{make-rep${}_i$}(n.s)$
\zi $\left.\begin{array}{@{}l@{}}
n.s.s := n.s\\
n.s := \const{nil}\\
\end{array}\right\}\mbox{\parbox{2.5cm}{\raggedright invert secondary edge}}$
\End
\End
\end{codebox}
\end{minipage}%
\caption{Data structure and functions to represent weak equivalence
relations. A \type{node} structure is created for every strong
equivalence class on arrays. It contains two outgoing edges $p,s$
pointing towards the representative of the weak equivalence classes.
The functions \proc{get-rep} and \proc{get-rep${}_i$} are used to
find the representative of the weak equivalence (resp. weak
equivalence modulo~$i$) class. The functions $\proc{make-rep}$ and
$\proc{make-rep${}_i$}$ invert the edges to make a node the
representative of its weak equivalence classes.\label{alg:data-structures}}
\end{figure}
The primary edge points towards the representative of the weak
equivalence class. Every primary edge $p$ connects the node
representing (the strong equivalence class of) a store $\store ajv$
with the node representing $a$ and the corresponding index in the $pi$
field is $j$. Note, however, that the direction of the edge can be
arbitrary, as we invert the edges during the execution of the
algorithm. If the primary edge is missing the node is the
representative of its weak equivalence class and of all its weak
equivalence modulo~$i$ classes.
While the primary edge is enough to represent the weak equivalence
relation we need another edge to represent weak equivalence modulo
$i$. The representative of weak equivalence modulo~$i$ is
also found by following the primary edges. However, if the store of
the primary edge is on the index $i$, the secondary edge is followed
instead. If the secondary edge is missing the node is the
representative of its weak equivalence modulo~$i$ class.
The equivalence classes are represented as follows. Two
arrays $a$ and $b$ are weakly equivalent iff
$\proc{get-rep}(a) = \proc{get-rep}(b)$ and
$a\weakeqi b$ iff $\proc{get-rep${}_i$}(a, i) =
\proc{get-rep${}_i$}(b,i)$.
\begin{figure}[htbp]
\begin{minipage}{.5\textwidth}
\begin{codebox}
\Procname{$\proc{add-secondary}(S : \type{index set}, a, b : \type{node})$}
\zi \If $a = b$ \Then
\zi \Return
\End
\zi \If $a.pi \notin S
\land \proc{get-rep${}_i$}(a, a.pi) \neq b$ \Then
\zi $\proc{make-rep${}_i$}(a)$
\zi $a.s := b$
\End
\zi $\proc{add-secondary}(S \cup \{a.pi\}, a.p, b)$
\end{codebox}
\end{minipage}%
\begin{minipage}{.5\textwidth}
\begin{codebox}
\Procname{$\proc{add-store}(a,b : \type{node}, i: \type{index})$}
\zi \proc{make-rep}(b)
\zi \If $\proc{get-rep}(a) = b$ \Then
\zi $\proc{add-secondary}(\{i\}, a, b)$
\zi \Else
\zi $b.p := a$
\zi $b.pi := i$
\End
\end{codebox}
\end{minipage}
\caption{The algorithm \proc{add-store} adds a new store edge to the
data structure updating the weak equivalence classes. In the else
case a new primary edge is added to merge two disjoint weak
equivalence classes. Otherwise, \proc{add-secondary} inserts new
secondary edges to merge the necessary weak equivalence modulo~$i$
classes. \label{alg:add-store}}
\end{figure}
The algorithm proceeds by inserting the store edges one by one,
similarly to the algorithm presented
in~\cite{DBLP:conf/rta/NieuwenhuisO05}. The algorithm that inserts a
store edge is given in Figure~\ref{alg:add-store}. The algorithm
first inverts the outgoing edges of one node to make it the
representative of its weak equivalence class. If the other side of
the store edge lies in a different weak equivalence classes, the store
can be inserted as a new primary edge.
If the nodes are already weakly equivalent the
procedure~\proc{add-secondary} is called. This procedure follows the
path from the other array $a$ to the array $b$ that was made the
representative. For every node on this path it checks if a secondary
edge needs to be added. If the primary edge of the node is labelled
with a store on~$i$, the algorithm first checks if the node is weakly
equivalent modulo~$i$ with $b$ due to the new store edge.
This is the case if no store on $i$ occurred on the path so far and the
new store is also on an index different from $i$. We use the set $S$
to collect these forbidden indices. Then if $b$ is not already the
representative of the weak equivalence modulo~$i$ class, the outgoing
secondary edges are reversed and a new secondary edge is added.
The complexity of the procedure \proc{add-store} is worst case
quadratic in the size of the weak equivalence class. This stems from
\proc{make-rep${}_i$} being linear in the size and being called a
linear number of times. The overall complexity is cubic in the number
of stores in the input formula. The space requirement, however, is
only linear. In our current implementation in SMTInterpol this
procedure was not a bottleneck so far. In SMTInterpol we also keep
the stores that created the primary and secondary edge in the
data-structure. This allows for computing the paths needed for lemma
generation in linear time.
\begin{example}\label{ex:addedge}
Figure~\ref{fig:addedge} shows an example of the data structure
where the primary edges are labelled by the index of the
corresponding store. This data structure represents only one weak
equivalence class with the representative node 0. The resulting
data structure after adding a store with index $k$ between nodes 0
and 4 is shown on the right. Since nodes 0 and 4 were
already in the same weak equivalence class, secondary edges were
added. \pgfdeclarelayer{back} \pgfsetlayers{back,main}
\begin{figure}[htbp]
\hfill
\begin{tikzpicture}[edge from parent/.style={draw,red,<-,thick},level
distance=0.75cm,sibling distance=1.4cm,index/.style={black,inner sep=0cm}]
\begin{scope}
\node[name=n0] {0}
child[thick] {
node[name=n1] {1}
child {
node[name=n2] {2}
child {
node[name=n3] {3}
child {
node[name=n4] {4}
edge from parent node[above left,index] {$k$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$i$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$j$}
}
child {
node[name=n5] {5}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above left,index] {$i$}
}
child[thick] {
node[name=n6] {6}
child[missing] {
node {}
}
child {
node[name=n7] {7}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above right,index] {$j$}
};
\end{scope}
\node[draw,shape=single arrow] at(3,-1.5) {$4 = \store{0}{k}{v}$};
\begin{scope}[xshift=7.3cm]
\node[name=n0] {0}
child[thick] {
node[name=n1] {1}
child {
node[name=n2] {2}
child {
node[name=n3] {3}
child {
node[name=n4] {4}
edge from parent node[above left,index] {$k$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$i$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$j$}
}
child {
node[name=n5] {5}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above left,index] {$i$}
}
child[thick] {
node[name=n6] {6}
child[missing] {
node {}
}
child {
node[name=n7] {7}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above right,index] {$j$}
};
\draw[green,thick,->] (n3) .. controls +(0,2.3) .. (n0);
\draw[green,thick,->] (n2) .. controls +(0,1.3) .. (n0);
\end{scope}
\end{tikzpicture}
\hfill\hbox{}
\caption{\label{fig:addedge} Weak equivalence classes represented by a
graph using primary and secondary edges. The short direct edges are
primary edges, the long bended edges are secondary edges. Each
primary edge represents a store edge between the connected nodes and is
labelled by the index of the store. The secondary edges in the right graph
were created by a store edge on index $k$ between node 0 and 4 as described
in Example~\ref{ex:addedge}.
}
\end{figure}
These secondary edges are needed to connect the weak equivalence
modulo $i$ and modulo $j$ classes. Figure~\ref{fig:mergei} shows how
the first secondary edge connects the two weak equivalence modulo $i$
classes rooted at nodes 0 resp.\ 3. This is necessary since there is
now a new path using the edge from 4 to 0. Note that no secondary
edge is added to node~1, since nodes 1, 2, and 5 are still not weakly
equivalent modulo $i$ to the other nodes.
Figure~\ref{fig:mergej} shows the connection between the two weak equivalence
modulo $j$ classes rooted at nodes 0 resp.\ 2. The weak equivalence modulo
$j$ class rooted at node 6 is not affected by a new edge between nodes 0 and 4
since these nodes are on a different path.
\begin{figure}[htbp]
\subfigure[Merging weak equivalence modulo $i$ classes.\label{fig:mergei}]{
\begin{tikzpicture}[edge from parent/.style={draw,red,<-,thick},level
distance=0.75cm,sibling distance=1.4cm,index/.style={black,inner sep=0cm}]
\node[name=n0] {0}
child[thick] {
node[name=n1] {1}
child {
node[name=n2] {2}
child {
node[name=n3] {3}
child {
node[name=n4] {4}
edge from parent node[above left,index] {$k$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$i$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$j$}
}
child {
node[name=n5] {5}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above left,index] {$i$}
}
child[thick] {
node[name=n6] {6}
child[missing] {
node {}
}
child {
node[name=n7] {7}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above right,index] {$j$}
};
\draw[green,thick,->] (n3) .. controls +(0,2.3) .. (n0);
\draw[black,dashed,thick] (n4) .. controls +(0,3.5) .. node[left]{$k$} (n0);
\begin{pgfonlayer}{back}
\draw[rotate=45,fill=white!80!black] (n6) ellipse [x radius=.5cm,y radius=1.7cm];
\draw[rotate=-45,fill=white!80!black] ($(n3)!.5!(n4)$) ellipse [x radius=.5cm,y radius=1cm];
\draw[fill=white!90!black] ($(n1)-(0,.6)$) circle [radius=.9cm];
\end{pgfonlayer}
\end{tikzpicture}
}
\hfill
\subfigure[Merging weak equivalence modulo $j$ classes.\label{fig:mergej}]{
\begin{tikzpicture}[edge from parent/.style={draw,red,<-,thick},level
distance=0.75cm,sibling distance=1.4cm,index/.style={black,inner sep=0cm}]
\node[name=n0] {0}
child[thick] {
node[name=n1] {1}
child {
node[name=n2] {2}
child {
node[name=n3] {3}
child {
node[name=n4] {4}
edge from parent node[above left,index] {$k$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$i$}
}
child[missing] {
node {}
}
edge from parent node[above left,index] {$j$}
}
child {
node[name=n5] {5}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above left,index] {$i$}
}
child[thick] {
node[name=n6] {6}
child[missing] {
node {}
}
child {
node[name=n7] {7}
edge from parent node[above right,index] {$k$}
}
edge from parent node[above right,index] {$j$}
};
\draw[green,thick,->] (n2) .. controls +(0,1.3) .. (n0);
\draw[black,dashed,thick] (n4) .. controls +(0,3.5) .. node[left]{$k$} (n0);
\begin{pgfonlayer}{back}
\draw[rotate=-45,fill=white!80!black] (n3) ellipse [x radius=.5cm,y radius=1.6cm];
\draw[rotate=45,fill=white!90!black] ($(n6)!.5!(n7)$) ellipse [x radius=.5cm,y radius=.9cm];
\draw[rounded corners,fill=white!80!black] (n0.north west) -- ($(n1.west)-(.2,0)$) --
($(n5.south east)-(0,.2)$) -- (n0.north east) -- cycle;
\end{pgfonlayer}
\end{tikzpicture}
}
\caption{\label{fig:mergeweakeqi}Secondary edges merge weak equivalence modulo
$i$ classes.}
\end{figure}
\end{example}
We implemented this decision procedure in our SMT solver
SMTInterpol~\cite{DBLP:conf/spin/ChristHN12} and tested it on the benchmarks
from the QF\_AX and QF\_AUFLIA devisions of the SMTEVAL 2013 benchmarks. We
solved all benchmarks in 1:32 resp.\ 10:45 minutes without running into a
timeout of 10 minutes. According to the data from the SMTEVAL, no other
solver was able to solve all benchmarks in these divisions. We defer an
up-to-date comparison to the SMTCOMP 2014.
\section{Conclusion and Future Work}
We presented a new decision procedure for the extensional theory of arrays.
This procedure exploits weak equalities to limit the number of axiom
instantiations. The instantiations produced by the decision procedure
presented in this paper can be restricted to terms already present in the
input formula. Furthermore we discussed an implementation based on a graph
structure similar to congruence closure graphs. This decision procedure is
implemented in our SMT solver \si~\cite{DBLP:conf/spin/ChristHN12}. We plan
to implement a variant of the quantifier-free interpolation for
arrays~\cite{DBLP:journals/corr/abs-1204-2386} based on the lemmas generated
by this decision procedure. Since these lemmas only generate mixed
equalities, proof tree preserving
interpolation~\cite{DBLP:conf/tacas/ChristHN13} can be used.
\bibliographystyle{splncs03}
|
\section*{Supplemental Material}
\subsection*{Nonperturbative corrections}
\begin{figure*}[t!]
\includegraphics[width=0.8\textwidth]{figs/yJ_dependence}
\caption{Boost by $-y_J$ along the beam direction and rotation by $-\phi_J$ around the beam direction used to show that $\Omega_\kappa$ is independent of $y_J$ and $\phi_J$.}
\label{fig:yJboost}
\end{figure*}
\begin{figure*}[t!]
\hfill%
\includegraphics[width=0.9\columnwidth]{plots/Omegahad_pT_AU2}%
\hfill\hfill%
\includegraphics[width=0.9\columnwidth]{plots/Omegahad_pT_Hw}%
\hspace*{\fill}
\vspace{-2ex}
\caption{$p_T^J$ dependence of $\Omega_\kappa^{\rm had}(R)$ for \textsc{Pythia}8\xspace (left panel) and \textsc{Herwig++}\xspace (right panel).}
\label{fig:Omegahad_pTJ}
\vspace{-1ex}
\end{figure*}
\begin{figure*}[t!]
\hfill%
\includegraphics[width=0.9\columnwidth]{plots/Omegahad_yJ_AU2}%
\hfill\hfill%
\includegraphics[width=0.9\columnwidth]{plots/Omegahad_yJ_Hw}%
\hspace*{\fill}
\vspace{-2ex}
\caption{Jet rapidity dependence of $\Omega_\kappa^{\rm had}(R)$ for \textsc{Pythia}8\xspace (left panel) and \textsc{Herwig++}\xspace (right panel).}
\label{fig:Omegahad_yJ}
\vspace{-1ex}
\end{figure*}
The leading hadronization effects in the first jet mass moment and in the tail of the jet mass spectrum are described by the parameter $\Omega_\kappa(R)$, which is defined by
\begin{align} \label{eq:Omega2}
\Omega_\kappa(R)
&= \int_0^1\! \mathrm{d} r \int_{-\infty}^{\infty}\! \mathrm{d} y \int_0^{2\pi}\! \mathrm{d} \phi\, f(r,y-y_J,\phi-\phi_J,R)
\nonumber \\ &\quad \times\!
\bigl\langle 0 \bigl\lvert \bar T[Y_J^\dagger Y_b^\dagger Y_a^\dagger]\, \hat {\mathcal E}_T(r,y,\phi) T[Y_a Y_b Y_J] \bigr\rvert 0 \bigr\rangle
.\end{align}
Here, $Y_a$ and $Y_b$ are incoming soft Wilson lines along the beam directions in the color representation of the incoming primary hard partons. $Y_J \equiv Y_J(y_J, \phi_J)$ is an outgoing soft Wilson line along the jet direction in the color representation of the outgoing hard parton. The color contractions between the Wilson lines are suppressed in \eq{Omega2}, but are normalized such that
$\langle 0 | Y_J^\dagger Y_b^\dagger Y_a^\dagger Y_a Y_b Y_J |0 \rangle = 1$. In \fig{Omegahad_pTJ} we show that $\Omega_\kappa^{\rm had}(R)$ in \textsc{Pythia}8\xspace and \textsc{Herwig++}\xspace is independent of $p_T^J$ over the large range of $p_T^J$ considered. (The $qg\to Zq$ and $gg\to Hg$ channels in \textsc{Herwig++}\xspace have a small downward trend in $p_T^J$.)
For $e^+e^-\to$ dijets, a boost allows one to derive important universality properties of the corresponding $\Omega$ parameter~\cite{Lee:2006nr}. Here for \eq{Omega2}, boosting by $-y_J$ along the beam axis and rotating by $-\phi_J$, as illustrated in \fig{yJboost}, the Wilson lines and energy flow operator transform as
\begin{align}
Y_{a,b} &\to Y_{a,b}
\,, \nonumber \\
Y_J(y_J,\phi_J) &\to Y_J(0,0)
\,, \nonumber \\
\hat {\mathcal E}_T(r,y,\phi) &\to \hat {\mathcal E}_T(r,y-y_J,\phi-\phi_J)
\,.\end{align}
Changing variables $y\to y+y_J$ and $\phi\to \phi+\phi_J$ then yields an expression depending only on $y$ and $\phi$, which thus shows that $\Omega_\kappa(R)$ in \eq{Omega2} is independent of $y_J$ and $\phi_J$. We therefore set $y_J=\phi_J=0$ in the following. Note that unlike for $e^+e^-\to$ dijets, the matrix element is not independent of $y$ and $\phi$, so these dependencies in the measurement $f(r,y,\phi)$ do not generically decouple. In \fig{Omegahad_yJ} we show that $\Omega_\kappa^{\rm had}$ obtained from Monte Carlo programs does not depend on the jet rapidity $y_J$. The behaviour with $p_T^J$ and $y_J$ shown in these plots does not depend on the value of $R$. In Figs.~\ref{fig:Omegahad_pTJ} and \ref{fig:Omegahad_yJ} we see again that in \textsc{Pythia}8\xspace the overall size of $\Omega_\kappa^{\rm had}(R)$ depends on the channel, being larger for the channels with a gluon jet. In contrast, $\Omega_\kappa^{\rm had}(R)$ in \textsc{Herwig++}\xspace is smaller and of similar sizes for all channels.
To discuss the $R$ dependence of $\Omega_\kappa(R)$, we switch to coordinates $\{y', \phi', r'\}$ measured with respect to the jet axis. This gives
\begin{align} \label{eq:Omegajetaxis}
\Omega_\kappa(R)
&= \int_0^1\! \mathrm{d} r' \int_{-\infty}^{\infty} \! \mathrm{d} y' \int_0^{2\pi}\! \mathrm{d} \phi' \,
f(r',\!y',\!\phi',\!R)
\nonumber \\ & \quad \times
\bigl\langle 0 \bigl\lvert \bar T[Y_J^\dagger Y_b^\dagger(0, \pi) Y_a^\dagger(0, 0)] \,\hat {\mathcal E}_\perp(r', y', \phi')\,
\nonumber \\ & \qquad \times
T[Y_a(0,0) Y_b(0,\pi) Y_J] \bigr\rvert 0 \bigr\rangle
\,,\end{align}
where the incoming beam Wilson lines $Y_{a,b}$ point in the $(y',\phi')=(0,0)$ and $(0,\pi)$ directions, and $r'=p_\perp/m_\perp$.
The measurement function in the original coordinates in \eq{Omega2} is given by
\begin{equation}
f(r,y,\phi,R) = (\cosh y-r \cos \phi)\, \theta\bigl[ b(y,\phi,r) < R^2 \bigr]
\,,\end{equation}
where we use $b(y,\phi,r)= 2(\cosh y- r \cos \phi)$ to define the jet boundary. In the primed coordinates it takes the form
\begin{align} \label{eq:fp}
&f(r',y',\phi',R)
\nonumber \\ & \qquad
= e^{-y'} \theta\Big[e^{2y'} - 2r'^2 \cos^2 \phi' + 1
\nonumber \\ & \qquad \quad
- \frac{2}{R^2}\sqrt{4 + R^4(r'^4 \cos^4 \phi' - r'^2 \cos^2 \phi')}\Big]
\,.\end{align}
Boosting along the \emph{jet axis} by $\ln(R/2)$ as in \fig{Rexpansion}, the Wilson lines and energy flow operator transform as
\begin{align}
Y_a(0,0) &\to Y_a(\ln \tfrac{R}{2},0)
\,, \nonumber \\
Y_b(0,\pi) &\to Y_b(\ln \tfrac{R}{2},\pi)
\,, \nonumber \\
Y_J &\to Y_J
\,, \nonumber \\
\hat {\mathcal E}_\perp(r',y',\phi') &\to \hat {\mathcal E}_\perp(r',y'\!+\!\ln \tfrac{R}{2},\phi')
\,.\end{align}
In these coordinates, the beam Wilson lines are an angle $\theta = 4\tan^{-1}(R/2) \simeq 2R$ apart.
We can now expand the result in $R$. To leading order in $R$, the measurement in \eq{fp} becomes
\begin{align}\label{eq:fR_expand}
f(r', y', \phi', R) = e^{-y'} \theta\bigl[y' + \ln(R/2) \bigr] \bigl[ 1 + \ord{R^2} \bigr]
\,.\end{align}
For the leading term in the $R\to 0$ limit, the beam Wilson lines fuse
\begin{align}\label{eq:YR_expand}
Y_a(\ln \tfrac{R}{2},0) Y_b(\ln \tfrac{R}{2},\pi) = Y_{\bar J}(-\infty,0) + {\cal O}(R^2)
\,,\end{align}
where $Y_{\bar J}$ is an incoming Wilson line along the direction opposite to the jet and in the appropriate conjugate color representation that forms a color singlet with the outgoing jet Wilson line $Y_J$. Since we now have two Wilson lines along the jet axis, we can boost along the jet axis to eliminate the $y'$ dependence. Integrating over $\phi'$ then yields
the result in \eq{Omega_R}, namely
\begin{align}\label{eq:Omega_R2}
\Omega_\kappa(R)
&= \frac{R}{2} \, \Omega_\kappa^{(1)} \!+\! \frac{R^3}{8}\, \Omega_\kappa^{(3)} \!+\! \frac{R^5}{32}\, \Omega_\kappa^{(5)} \!+\! {\cal O}\Bigl[\Bigr(\frac{R}{2}\Bigl)^7\Bigr]
\,,\end{align}
where the coefficient of the leading term comes from integrating the measurement function over $y'$,
\begin{align}
\int_{-\infty}^\infty\! \mathrm{d} y' \, e^{-y'} \theta\bigl[y' + \ln(R/2) \bigr] = \frac{R}{2}
\,.\end{align}
The leading nonperturbative parameter in \eq{Omega_R2} is given by a universal matrix element
\begin{align}
\Omega_\kappa^{(1)}
= c_e \int_0^1\! \mathrm{d} r' g_e(r')
\bigl\langle 0 \big| \bar T[Y_J^\dagger Y_{\bar J}^\dagger ]
\hat {\mathcal E}_\perp\!(r') T[Y_{\bar J} Y_J] \big|0 \bigr\rangle
.\end{align}
It depends on the color representation of the Wilson line (quark vs.~gluon) but not the full original color configuration. To extend our result to a more general jet measurement $e$, we included the parameters $c_e$ and $g_e(r')$, which in our case simply are given by $c_e=g_e(r')=1$. In general $c_e$ is the calculable coefficient for the observable $e$~\cite{Lee:2006nr} obtained here by integrating over our $y'$ variable. The calculable function $g_e(r')$ encodes the dependence on hadron mass effects~\cite{Mateu:2012nk}.
\begin{figure*}[t!]
\includegraphics[width=0.8\textwidth]{figs/R_expansion}
\caption{Changing coordinates and boosting by $\ln(R/2)$ along the jet direction in order to expand around small $R$.}
\label{fig:Rexpansion}
\end{figure*}
\begin{table}
\begin{tabular}{l|c|cccl}
\hline\hline
& $\kappa$ & $\Omega_\kappa^{(1)}$ & $\Omega_\kappa^{(3)}$ & $\Omega_\kappa^{(5)}$ & [GeV]\\
\hline
\textsc{Pythia}8\xspace AU2 & $qg\to q$ & $1.2$ & $1.5$ & $1.3$ \\
\textsc{Pythia}8\xspace 4C & $qg\to q$ & $1.1$ & $0.7$ & $2.0$ \\
\textsc{Herwig++}\xspace & $qg\to q$ & $1.2$ & $-0.9$ & $4.0$ \\
\hline
\textsc{Pythia}8\xspace AU2 & $q\bar q\to g$ & $2.1$ & $-0.9$ & $3.0$ \\
\textsc{Pythia}8\xspace 4C & $q\bar q\to g$ & $2.1$ & $-1.4$ & $3.4$ \\
\textsc{Herwig++}\xspace & $q\bar q\to g$ & $1.0$ & $0.3$ & $2.4$ \\
\hline
\textsc{Pythia}8\xspace AU2 & $gg\to g$ & $2.2$ & $1.5$ & $2.4$ \\
\textsc{Pythia}8\xspace 4C & $gg\to g$ & $2.1$ & $0.4$ & $3.0$ \\
\textsc{Herwig++}\xspace & $gg\to g$ & $1.0$ & $1.3$ & $1.5$ \\
\hline\hline
\end{tabular}
\caption{Fit coefficients for $\Omega_\kappa(R)$ in \eq{Omega_R} for different Monte Carlo programs and tunes which give the lines shown in Fig.~\ref{fig:Omegahad_R}.}
\label{tab:Omegakappa}
\end{table}
The expansions in \eqs{fR_expand}{YR_expand} can be carried out to higher orders in $R$, using Ref.~\cite{Marcantonini:2008qn} to expand the Wilson lines about the $\bar J$ direction, and lead to new nonperturbative matrix elements, collectively denoted as $\Omega_\kappa^{(3,5)}$ in \eq{Omega_R2}. Terms with an odd number of gauge field components that are transverse to the jet direction vanish due to parity invariance. Together with the overall factor of $R$, this implies that $\Omega_{\kappa}(R)$ only contains odd powers of $R$. The coefficients of the fits shown in \fig{Omegahad_R} are given in Table~\ref{tab:Omegakappa}. The leading coefficient in $R$, $\Omega_\kappa^{(1)}$, is the same for quark and gluon jets, while the higher coefficients are quite different for all three channels. The higher coefficients $\Omega_\kappa^{(3)}$ and $\Omega_\kappa^{(5)}$ strongly depend on the Monte Carlo program and tune. They are also correlated so their separation is not well constrained by the fit. The fact that all the coefficients are of similar size confirms that $R/2$ is indeed the appropriate expansion parameter.
\begin{widetext}
As an illustration of the utility of the operator formulation, we give explicit results for some $\Omega_\kappa^{(1,3)}$s. These results could be used to build models that follow the structure in QCD, or perhaps someday to compute these matrix elements on the lattice. For $\Omega_{qg \to q}^{(1)}$ and $\Omega_{qg \to q}^{(3)}$, and the case where $c_e=g_e(r')=1$, we have
\begin{align} \label{eq:O13qg}
\Omega_{qg \to q}^{(1)} &= \Omega_q^{(1)} = \int_0^1\!\!\mathrm{d} r\: \frac{1}{N_c} \big\langle 0 \big| \mathrm{tr} \big\{
Y_{\bar J}^\dagger Y_J \hat {\mathcal E}_\perp(r) Y_J^\dagger Y_{\bar J} \big\} \big| 0 \big\rangle
\,,
\\
\Omega_{qg \to q}^{(3)} &= \int_0^1\!\!\mathrm{d} r\: \frac{(-1)}{N_c} \Big\langle 0\, \Big| \mathrm{tr}\Big\{
\Big[Y_{\bar J}^\dagger Y_J \hat {\mathcal E}_\perp(r) Y_J^\dagger Y_{\bar J},
\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g {n_J} \!\cdot\! {\mathcal B}_{\bar J}
+ \frac{2}{({\bar n_J} \!\cdot\! \mathrm{i} \partial)^2}\,\mathrm{i} \partial_\perp \!\cdot\! g {\mathcal B}_{{\bar J}\perp}
- \frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial} \Big[\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial} g {\mathcal B}_{{\bar J}\perp}^\mu, g {\mathcal B}_{{\bar J}\perp\mu} \Big]\Big]
\nonumber \\ & \qquad
+\Big[\Big[Y_{\bar J}^\dagger Y_J \hat {\mathcal E}_\perp(r) Y_J^\dagger Y_{\bar J},
\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g {\mathcal B}_{{\bar J}\perp}^\mu\Big],\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g {\mathcal B}_{{\bar J}\perp\mu}\Big]
+\frac{1}{C_F} \Big[\Big[Y_{\bar J}^\dagger Y_J \hat {\mathcal E}_\perp(r) Y_J^\dagger Y_{\bar J},
\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g {\mathcal B}_{{\bar J}\perp}^{\mu A}\Big],\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g {\mathcal B}_{{\bar J}\perp\mu}^A\Big]
\nonumber \\ & \qquad
- \frac{(1-r^2)}{2} Y_{\bar J}^\dagger Y_J \hat {\mathcal E}_\perp(r) Y_J^\dagger Y_{\bar J}
\Big\} \Big|\,0 \Big\rangle
\nonumber
\,,\end{align}
where the Wilson lines $Y_J$, $Y_{\bar J}$, and ${\cal B}_{\bar J}^\nu = {\cal B}_{\bar J}^{\nu A} T^A= \frac{1}{g} [ Y_{\bar J}^\dagger \mathrm{i} D^\nu Y_{\bar J}]$ are all in the fundamental representation, and $\mathrm{tr}$ is a trace over $3$ and $\bar 3$ color indices. The path for $Y_J^\dagger Y_{\bar J}$ is $[-\infty,0]$ along $\bar n_J$, then $[0,\infty]$ along $n_J$. The measurement is normalized such that $\hat {\mathcal E}_\perp(r) = 2\pi \hat {\mathcal E}_\perp(r,0,0)$, which is equal to $\hat {\mathcal E}_\perp(r,y=0)$ of Ref.~\cite{Mateu:2012nk}. In \eq{O13qg} the inverse derivatives $1/(\bar n_J\!\cdot\!\mathrm{i}\partial)$ only act on the fields they are next to, and the fields on the right (left) side of the measurement $\hat {\mathcal E}_\perp(r)$ are (anti) time-ordered. For $\Omega_{q\bar q \to g}^{(1)}$, $\Omega_{gg \to g}^{(1)}$, and $\Omega_{q\bar q \to g}^{(3)}$, and the case where $c_e=g_e(r')=1$, we have
\begin{align} \label{eq:O13qqb}
\Omega_{q\bar q \to g}^{(1)} &= \Omega_g^{(1)}
= \int_0^1\!\!\mathrm{d} r\: \frac{1}{N_c^2\!-\! 1} \big\langle 0 \big| {\rm Tr} \big\{
{\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_{\perp}\!(r) {\mathcal Y}_J^T {\mathcal Y}_{\bar J} \big\} \big| 0 \big\rangle
= \Omega_{gg\to g}^{(1)} \,,
\\
\Omega_{q\bar q \to g}^{(3)} &=\!\int_0^1\!\!\!\mathrm{d} r\: \frac{(-1)}{N_c^2\!-\! 1} \Big\langle 0 \,\Big| {\rm Tr} \Big\{
\Big[ {\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_\perp\!(r) {\mathcal Y}_{J}^T {\mathcal Y}_{\bar J} ,
\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g n_J \!\cdot\! \tilde {\mathcal B}_{\bar J}
\!+\!\frac{2}{({\bar n_J} \!\cdot\! \mathrm{i} \partial)^2}\,\mathrm{i} \partial_\perp\! \!\cdot\! g \tilde {\mathcal B}_{{\bar J}\perp}
\!+\!\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial} \Big[ \frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial} g \tilde {\mathcal B}_{{\bar J}\perp}^\mu, g \tilde {\mathcal B}_{{\bar J}\perp\mu} \Big] \Big]
\nonumber \\ & \quad
+\Big[\Big[ {\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_\perp\!(r) {\mathcal Y}_{J}^T {\mathcal Y}_{\bar J} ,\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g \bar {\mathcal B}_{{\bar J}\perp}^\mu\Big],\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g \bar {\mathcal B}_{{\bar J}\perp\mu}\Big]
+ \frac{1}{T_F N_c} \Big[\Big[ {\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_\perp\!(r) {\mathcal Y}_{J}^T {\mathcal Y}_{\bar J} ,\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g \vec {\mathcal B}_{{\bar J}\perp}^\mu\Big],\frac{1}{{\bar n_J} \!\cdot\! \mathrm{i} \partial}\, g \vec {\mathcal B}_{{\bar J}\perp\mu}\Big]
\nonumber \\ & \qquad
-\frac{(1-r^2)}{2} {\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_\perp\!(r) {\mathcal Y}_{J}^T {\mathcal Y}_{\bar J}
\Big\} \Big|\, 0 \Big\rangle
\nonumber
\,,\end{align}
where the Wilson lines ${\mathcal Y}_J$ and ${\mathcal Y}_{\bar J}$ are in the adjoint representation, the gluon fields $\tilde {\mathcal B}_{\bar J}^{ab} = -i f^{Cab} {\mathcal B}_{\bar J}^C$ and $\bar {\mathcal B}_{\bar J}^{ab} = d^{Cab} {\mathcal B}_{\bar J}^C$ are matrices, and ${\rm Tr}$ is a trace over adjoint color indices. When ${\rm Tr}$ acts on the term with $\vec {\mathcal B}_{\bar J}^A={\mathcal B}_{\bar J}^A$ it simply contracts these color vectors to the appropriate sides of the color matrix ${\mathcal Y}_{\bar J}^T {\mathcal Y}_J \hat {\mathcal E}_\perp\!(r) {\mathcal Y}_{J}^T {\mathcal Y}_{\bar J}$.
\end{widetext}
\subsection*{Soft function contribution to ISR}
\begin{figure}[t!]
\hfill%
\includegraphics[width=0.94\columnwidth]{plots/DelM1_pT_ISRMPI_Zg}%
\hspace*{\fill}
\vspace{-2ex}
\caption{Same as in \fig{DelM1MPI} but for $q\bar q\to Zg$.
}
\label{fig:DelM1MPI2}
\vspace{-1ex}
\end{figure}
\begin{figure*}[t!]
\hfill%
\includegraphics[width=0.9\columnwidth]{plots/DelM1_yJ_ISRMPI_Zq}%
\hfill\hfill%
\includegraphics[width=0.9\columnwidth]{plots/DelM1_yJ_ISRMPI_Zg}%
\hspace*{\fill}
\vspace{-2ex}
\caption{Same as \fig{DelM1MPI} but for the $y_J$ dependence.}
\label{fig:DelM1MPI_yJ}
\vspace{4ex}
\vspace{-1ex}
\end{figure*}
\begin{figure*}[t!]
\hfill%
\includegraphics[width=0.9\columnwidth]{plots/DelM1_pT_ISRMPI_Hg}%
\hfill\hfill%
\includegraphics[width=0.9\columnwidth]{plots/DelM1_yJ_ISRMPI_Hg}%
\hspace*{\fill}
\vspace{-2ex}
\caption{Same as \fig{DelM1MPI} and \fig{DelM1MPI_yJ} but for the $gg\to Hg$ process.}
\label{fig:DelM1MPI_Hg}
\vspace{-1ex}
\end{figure*}
At $\ord{\alpha_s}$ the soft function contains the following term~\cite{Jouttenus:2011wh}
\begin{align} \label{eq:soft_interf}
S^{\rm pert}_\kappa(k_S) \supset \bigl[I_0(\alpha,\beta) +I_0(\beta,\alpha)\bigr] \frac{\alpha_s C_\kappa}{\pi}\, \frac{1}{\mu} \Bigl(\frac{\mu}{k_S}\Bigr)_+
.\end{align}
The color factor for this interference of soft ISR from the two beams is given by the color charge of the two incoming partons, $C_\kappa = - \mathbf{T}_a \cdot \mathbf{T}_b$. For the processes we consider, this is simply a number given in \eq{ISRcolor}, but in general this is a matrix in color space.
The $I_0$ in \eq{soft_interf} is given by the following integral
\begin{align} \label{eq:I0}
I_0(\alpha, \beta)
&= \frac{1}{\pi}\int_{-\pi}^{\pi}\!\mathrm{d}\phi \int\!\mathrm{d} y\,
\theta\bigl(e^{y_J -y} - \sqrt{\beta/\alpha}\bigr)
\\ & \quad \times
\theta\bigl(1/\alpha - 1 - e^{2(y_J -y)} + 2 e^{y_J -y} \cos\phi\bigr)
\,.\nonumber\end{align}
with parameters
\begin{align}
\alpha = (1- \tanh y_J)/(2\rho)
\,, \nonumber \\
\beta = (1+ \tanh y_J)/(2\rho)
\,.\end{align}
Here, $\rho(R,y_J)$ controls the jet size, which is chosen such that the jet area in $(y,\phi)$ space equals $\pi R^2$~\cite{Jouttenus:2011wh}. The total integral in \eq{I0} is an area in $(y,\phi)$ space, where the second theta function restricts the integral to the jet and the first theta function reduces to $\theta(y<0)$ and $\theta(y>0)$ for $I_0(\alpha,\beta)$ and $I_0(\beta,\alpha)$, respectively. Therefore, including the overal $1/\pi$ factor,
\begin{align}
I_0(\alpha,\beta) + I_0(\beta,\alpha) = R^2
\,,\end{align}
which yields the $R^2$ dependence shown in \eq{soft_interf_short}.
The $p_T$ dependence of the MPI and soft ISR contributions to the jet mass moment is discussed in \fig{DelM1MPI}. In \fig{DelM1MPI_yJ} we show in addition the $y_J$ dependence in the same way. The $y_J$ dependence of the MPI is essentially flat, except for perhaps a small reduction at large rapidities. Since soft ISR emissions are constant in rapidity, one would expect the soft ISR contribution to the moment to be independent of the jet rapidity at central rapidities. This agrees well with what is observed in \textsc{Herwig++}\xspace for $\lvert y_J\rvert \lesssim 1.5$, while for larger $y_J$ the soft ISR contribution reduces. As already observed before, \textsc{Pythia}8\xspace has a larger soft ISR and smaller MPI contribution than \textsc{Herwig++}\xspace. In addition, the $y_J$ dependence of the soft ISR differs noticably between \textsc{Pythia}8\xspace and \textsc{Herwig++}\xspace. Hence, measurements of the $y_J$ rapidity dependence can also provide constraints on the modelling of soft ISR in the Monte Carlo programs.
For completeness we have included the analogs of \fig{DelM1MPI}, for the $q\bar q \to Zg$ channel in \fig{DelM1MPI2}, for the rapidity dependence in \fig{DelM1MPI_yJ}, and for the $gg \to Hg$ channel in \fig{DelM1MPI_Hg}. Note that the size of the $R^4$ contribution from soft ISR for the gluon channel is very similar to $qg \to Zq$ (at central rapidities). This might be surprising since this is a purely gluonic process, but it is in agreement with the prediction from the color factors in \eq{ISRcolor}.
\newpage
\clearpage
\end{document} |
\section{Introduction}
The Hawking's discovery that black holes \cite{Haw1} have thermal radiations gave a fundamental tool to investigate
black hole thermodynamics in general relativity. In fact, in a asymptotically flat spacetime, to a static spherically symmetric black hole of
proper area $A$ can be associated an entropy $S_{BH}$ given by $S_{BH}=\frac{k_B A}{4 L_P^2}$, where
$k_B$ is the Boltzmann constant and $L_P$ the Planck length.
The situation is much more involved in a cosmological non static background.
As an example, only recently has been proved that
\cite{kal} the McVittie solution \cite{mac}
contains a black hole in an expanding universe. Moreover, it has been shown \cite{H1,H2}
that in an expanding universe, an important ingredient is provided by the apparent horizon for a black hole,
in particular for the definition of its entropy. Other examples of study of entropy of black holes embedded in an expanding universe can be found, for example in
\cite{E1,E2,E3}. However, to the best of my knowledge,
in all these papers one assumes for the entropy $S_{BH}$ the same expression of the static background, i.e.
$S_{BH}=\frac{k_B A}{4 L_P^2}$,
where $A$ denotes the proper area of the apparent horizon.
In this paper, by considering suitable theorems for the formation of trapped surfaces in expanding universes \cite{13,17},
we explore the possibility that the expression for the black hole entropy in expanding universes should contain an extra term depending on the Hubble flow.
In section 2 we present the black hole theorems suitable for an expanding universe.
In section 3 we study the consequences of our new
expression for $S_{BH}$.
Section 4 is devoted to a discussion of the
entropy bounds, while
section 5 we preliminarly study the inflationary paradigm within our proposal.
Finally, section 6 is devoted to some conclusions and final remarks.
\section{Black Holes Theorems in Flat Expanding Universes}
An important issue in general relativity concerns the conditions allowing a mass-energy concentration in a certain volume
to collapse in a black hole. On large scale, our universe
is well approximated by a spatially flat Friedmann metric, so in this paper we
mainly consider this case.
In the
case of spherical collapse in asymptotically flat background, if the condition
$2GM/c^2>R$ ($R$ is the radius of the sphere and $M$ the ADM total mass-energy) is fulfilled, a black hole arises. However $M$ is the total energy (together with the negative binding energy) not the one effectively concentred within $R$ and $R$ itself is not an appropriate measure of the proper volume inside $R$. A fundamental ingredient characterizing a black hole is its event horizon. Unfortunately,
the identification of the event horizon is a global property of the spacetime and its identification is a complicated task. A
more manageable (local) ingredient is provided by trapped surfaces. In fact in \cite{isr} it has been shown that the formation of trapped surfaces (TS)
caused by spherically symmteric mass-energy concentration satisfying the weak energy condition, unavoidably leads to a
black hole. In a series of papers \cite{13,17} one can find necessary conditions for the non formation of TS and sufficient conditions for the formation of TS expressed only in terms of proper quantities in the spherical case. In particular in
\cite{14,17} such conditions are done in Friedmann cosmologies.\\
In this paper, we consider the theorem shown in \cite{17} for open Friedmann flat cosmologies.
Consider, in a flat Friedmann cosmology with a background energy-density ${\overline{\rho}}_m$,
a spherical two surface $S$ of proper radius $L$ and proper area $A$ and
a perturbation of proper mass $\delta M>0$ within $S$.
Initial conditions are done in a three dimensional manifold $\Sigma$ that can be seen as a spacelike slice of the four dimensional manifold representing the solution of the Einstein's equations. The two-surface $S$ is embedded in $\Sigma$.
Suppose that the current matter perturbation $\delta J_{\mu}$
is vanishing on the boundary of $S$, i.e. that the matter is at rest on $S$ and that
the trace of the extrinsic curvature $k_{\alpha\beta}$ is constant on comoving foliation ($t=const.$), i.e. $k^{\alpha}_{\alpha}=const.$
As a consequence, if
\begin{equation}
\delta M\frac{G}{c^2}<\frac{L}{2}+A\sqrt{\frac{G\;{\overline{\rho}}_m}{6\pi c^2}},
\label{3}
\end{equation}
then $S$ is not trapped.
Note that $L$ and $A$
are respectively the proper length and the proper area with respect to
the perturbed configuration, i.e. backreaction is taken into account. Moreover,
the upper bound for the mass-excess depends on the energy-density of the background on which the
perturbation acts. The authors of \cite{17} do not quote necessary and sufficient conditions for the formation of TS. Fortunately, in
\cite{17} we can found a necessary condition for the formation of TS.
Under the same conditions of (\ref{3}), if
\begin{equation}
\delta M\frac{G}{c^2}>L+A\sqrt{\frac{G\;{\overline{\rho}}_m}{6\pi c^2}},
\label{3bis}
\end{equation}
then $S$ is trapped. Therefore, by introducing a real constant $\gamma\in[1,2]$, we can argue from
(\ref{3}) and (\ref{3bis}) that the necessary and sufficient condition for the non formation of TS is
\begin{equation}
\delta M\frac{G}{c^2}<\frac{\gamma L}{2}+A\sqrt{\frac{G\;{\overline{\rho}}_m}{6\pi c^2}}.
\label{3tris}
\end{equation}
In any case, with
respect to the static case, a further term appears proportional to $\sim A\sqrt{{\overline{\rho}}_m}$. This term can also be written
as a function
of the cosmological constant $H$. In fact, if all the energy-densities present in the universe $\sum_i{\rho}_i$
satisfy the Friedmann equation
$H^2=8/3\pi G\sum_i{\rho}_i$ (i.e. we have a spherical black hole embedded in a Friedmann expanding universe, see
\cite{wit, mac}), then inequality (\ref{3tris}) becomes:
\begin{equation}
\delta M\frac{G}{c^2}<\frac{\gamma L}{2}+\frac{AH}{4\pi c}.
\label{3z}
\end{equation}
The condition (\ref{3z}) indicates, according to physical intuition, that an expanding universe makes more difficult the formation of trapped surfaces and as a consequence the formation of black holes \cite{16}. The interesting feature of the inequality (\ref{3z}) it
is that is expressed only in terms of proper quantities that are in principle measurable.\\
The theorem quoted above is the starting point to investigate its consequences on the maximal entropy allowed in a certain spacetime region. It contains only well defined proper local quantities.
\section{Generalized Bekenstein-Hawking Entropy}
An interesting first consequence of the reasonings above is the modification of the Bekenstein-Hawking entropy.
In fact, if the maximum entropy contained in a spherical region of radius $R$ and energy $E$
is provided by the entropy of a black hole of the same
radius, then a correction term due to $H$ must be included. Otherwise,
a certain region of radius $R$ should contain 'more entropy' than the one used for a black hole, i.e. $S_{BH}=\frac{k_B A}{4 L_P^2}$,
but without forming, thanks to the inequality (\ref{3z}), a black hole: a contradiction.\\
In this section we explore this interesting aspect of our proposal.
\subsection{Bekenstein bound and entropy bound}
A first interesting consequence of our result is that
expression (\ref{3}) suggests a correction term caused by the degrees of freedom that are due to the
non-static nature of Friedmann spacetimes. To start with,
consider the old original
Bekenstein argument \cite{Bek} in an asymptotically flat spacetime in a static context.
This argument states that there exists an universal bound for the
entropy $S$ of a spherical object of radius $R$ and energy $E$ given by
$S\leq S_{max}=\frac{2\pi k_B RE}{\hbar c}$.\\
The derivation of this bound originated a dispute (see for example \cite{d1,d2,d3,d4}). This bound can be
violated in the case of strong gravitational fields as a system collapsed inside a black hole
(although the metric inside the event horizon becomes time dependent and calculations are not so obvious)
and in a cosmological context (see \cite{d5} for a proposal to general spacetimes in order
to save the holographic principle). This discussion, as we see below, although important it is not essential for our purposes to give a
more general expression for the black hole entropy in expanding universe.
In any case, an improvement of the Bekenstein original bound has been done by Susskind (see \cite{bb1,bb2}). The argument is the following.
Since $E\sim R$ and more precisely $E\leq E_{max}=c^4 R/(2G)$,
the Bekenstein bound results proportional to the proper area of the object, i.e.
$S_{max}= \frac{k_B A}{4 L_P^2}$. Hawking calculations \cite{Haw1} confirmed
this factor to be $1/4$ leading to the well known expression for the black hole entropy by showing that this bound is satured by
the black holes. This bound (spherical entropy bound) is weaker than the Bekenstein one. It is a consequence of the requirement
that a system be gravitationally stable, otherwise, a gravitational collapse with a final
black hole arises. Also this bound fails in practical
situations as the case of a collapsing object. A further improvement
is provided by the spacelike bound (see \cite{bb2} and references therein).
This bound implies that the entropy of the matter enclosed in a
compact spacelike surface of volume $V$ and area $A$ never exceeds the entropy of a black hole
of area $A$. Also in this form (see the discussion at section 4), counterexamples can be found.
The situation is much more involved in a cosmological dynamical context. There, because of the non static nature of the spacetime,
the bound becomes dynamical \cite{bb3,bb4}. In particular, in \cite{bb3} the holographic principle is discussed in
the context of
a closed radiation dominated Friedmann universe where the radiation is represented by a conformal field theory with a suitable central charge.
In this frame emerged a new bound satured when the Cardy formula is used. This bound depends on the Hubble flow
$H$ (see the discussions at subsection 3.3). The most interesting bound according to the holographic principle working also in cosmological spacetimes is the one quoted by Bousso \cite{d5,bb2}. There, the bounding area is not the one enclosed in a
region $B$ but rather the one expressed in terms of light-sheet of $B$ itself. This bound seems to work but its physical origin
is still obscure.
Summarizing, the reasonings above show that,
in order to save the holographic principle, the area appearing in the right hand side of the spacelike
entropy bound must be greater than the area enclosing the surface $A$. This is an important lesson of the Bousso idea.
In this paper we are mainly interested in a generalization of the Bekenstein-Hawking entropy in a cosmological context. In this regard, it is sufficient to consider the spacelike entropy bound. In fact, in this way we can obtain as an upper bound the correct expression for the
black hole entropy in asymptotically flat spacetime. Then, with the use of theorem (\ref{3z}) we can infer the new form of the black hole entropy suitable in expanding universes.
\subsection{Bekenstein-Hawking entropy in expanding universes}
Following the entropy bound argument, we propose a generilized expression for the Bekenstein-Hawking formula suitable
in a flat expanding universe with embedded a black hole.
To our purposes,
it is more appropriate to work with proper distances. There, $E$ will denote the proper energy within $S$ and $R$
the proper length $L$ of the sphere.
Thanks to (\ref{3z}), the maximal allowed energy $E_{max}$ in a spherical region of proper radius $L$ above which a black hole arises is
\begin{equation}
E_{max}=\frac{c^4}{G}\left[\frac{\gamma L}{2}+A\sqrt{\frac{G\;{{\rho}}_m}{6\pi c^2}}\right],
\label{bo1}
\end{equation}
where ${\rho}_m$ denotes the matter-energy content of the universe.
It is important to note that the proper quantities $L$ and $A$ in (\ref{bo1})
refer to the one of a black hole.
The identification
of the surface $A$ with the analogue of the event horizon in the Schwarzschild case
in expanding universes is not (see for example \cite{wit}) a simple task.
Since we are considering black holes in expanding universes,
the enclosing surface $A$, in light, for example, of the works \cite{H1,H2} can be well identified with the
apparent horizon of the black hole.
It is not
obvious how to define the proper volume or the proper length of a black hole, since the radial coordinate is time-like inside $V$.
As shown in \cite{vol}, one can define, for non rotating black holes,
an effective volume (named geometric volume)
as minus the volume excluded from a spatial slice by the black hole horizon, i.e. $V_h=4/3\pi R_h^3$
where $R_h=R(t,r_h)$ is the scale factor in a spherically symmetric expanding universe (i.e. the the angular part of the metric,
at $t=const$ and $R=const$, can be always written in the form $R^2(d{\theta}^2+{\sin}^2\theta d{\phi}^2)$)
and $r_h$
measures the comoving radius of the apparent horizon of the embedded black hole.
According to the definition of $V_h$, we can define the effective proper length of the black hole as $L_h=R_h$. In this context
$H=R_{,t}/R$. In fact, in the original derivation of theorem (\ref{3}), the authors
of \cite{14,15,17} work in isotropic coordinates
by factorizing $R$ as $R={\phi}^2(r)a(t)$. In this way we have that $H=R_{,t}/R=a_{,t}/a$, as happens for example for the McVittie solution
\cite{mac}.\\
The factor $\gamma$ can be fixed from the fact that in the static limit, for the Bekenstein bound we obtain:
\begin{equation}
S_{max}=S_{BH}=\frac{k_B\gamma A_h}{4 L_P^2},
\label{static}
\end{equation}
and thus $\gamma=1$.
As a consequence of the above reasonings,
with
$L_h=\sqrt{\frac{A_h}{4\pi}}$,
for the entropy $S_{BH}$ we obtain
\begin{equation}
S_{BH}=\frac{k_B A_h}{4 L_P^2}+\frac{k_BA_h^{\frac{3}{2}}}{c L_P^2}\sqrt{\frac{G{\rho}_m}{6}}.
\label{entr}
\end{equation}
The correction term takes into account
the degrees of freedom of a non-static (flat) Friedmann universe.
Expression (\ref{entr}) can also be written in terms of $H$ instead of ${\rho}_m$:
\begin{equation}
S_{BH}=\frac{k_B A_h}{4 L_P^2}\left(1+\frac{H}{c}\sqrt{\frac{A_h}{\pi}}\right).
\label{entr2}
\end{equation}
The formula (\ref{entr2}) says that our bound is greater than the ones given by the first term only:
an higher admissible upper bound for the density implies an higher upper bound for the entropy.
The usual term of the black hole entropy, i.e. the first one in the right hand side of (\ref{entr2}), derives on the term $L/2$ in
(\ref{3}). Consider now a black hole in an expanding universe with apparent horizon area $A_h$. If the usual expression
for the entropy $S_u$ where $S_u=\frac{k_B A_h}{4 L_P^2}$, thanks to (\ref{3z}) we could have an object with entropy $S$
such that $S_u<S<S_{BH}$, i.e. an object with an entropy greater than the one of a black hole of the same size but without forming a black hole, because in expanding universe more energy is required in a spherical surface A to form a TS. This is a logic contradiction
that is solved by adding to the black hole entropy the volume term $\sim A_h^{3/2}H$.
The bound $S<S_{BH}$ is an entropy bound. In subsection 3.3 a new interpretation of this bound, in light of the theorems used
in section 2, is done
solving some problems of the original formulation.\\
Generally, the correction term is negligible when $A_h<<c^2/H^2$: when the proper dimensions of an object become comparable with
the Hubble radius, this correction cannot be neglected. As a title of example,
suppose to have a spherical region of comoving radius $r_c$ with
a power low cosmology $a(t)=a_0 t^{\alpha},\;\alpha>0$ (in this case $R=L=a(t) r_c$).
The correction term in (\ref{entr2}) becomes dominant when
$r_c>>c/(a_{,t})=ct^{1-\alpha}/(a_0\alpha)$. As a consequence, if the strong energy condition is fulfilled ($\alpha<1$),
the correction term can dominate near the big bang singularity. Conversely, for inflationary cosmologies $\alpha>1$ with a
power law behaviour, the added term
becomes dominant also for relatively small comoving regions, provided that a sufficient cosmic time is
considered: this phenomenon also appears for a universe filled with a cosmological constant where asymptotically we have
$H\rightarrow c\sqrt{\Lambda/3}$. Also during the primordial quasi de Sitter inflationary era, this correction term can be relevant.
It is interesting to study the fate of the full expression
(\ref{entr2}) near the big bang. The first term is always vanishing at $t=0$. Concerning the added term,
if $H\sim 1/t$, then for a power law cosmologies with $a\sim t^{\alpha}$, for $\alpha\in (0, 1/3)$,
$S_{BH}\rightarrow\infty$. Conversely, for a spacetime with $\alpha>1/3$,
$S_{BH}\rightarrow 0$. Interestingly enough, for stiff matter, i.e. $\alpha=1/3$, then entropy reaches at the big bang a finite non-vanishing limit. The fact that only for superluminar acausal fluids the entropy diverges at the big bang
strongly supports the physical reasonability of our proposal.
As a further remark, we analyze the expressions (\ref{entr}) and (\ref{entr2}) from the point of view of the first law of thermodynamics.
To this purpose, note that
in (\ref{entr}) the term $A_h^{3/2}$ can be written also in terms of $V$.
Since we are in a spherically symmetric context, in (\ref{entr})
we have firstly explicitally written this term as a function of the proper area.
In practice, for a sphere in a spatially flat metric, the proper quantities $A_h$ and $V_h$ are not independent.
However, in ligth to the first law of thermodynamics, it seems more natural and useful to express the added term as a function of the
proper volume of the apparent horizon of the black hole.\\
To this purpose, we can follow a practical rule: if we have a behaviour
$L^a$ with $a$ a positive integer, then we must have $L^a\sim A^bV^q\sim L^{2b+3q}$, with $b,q$ positive integers.
In the case under consideration, the first term in (\ref{entr}) is trivially $A$ (i.e. $b=1, q=0$), while the second is
$V$ (i.e $b=0,q=1$). In practice surfaces and volumes do appear with integer exponents.
We obtain, after dropping the subscript $h$:
\begin{equation}
S_{BH}=\frac{k_B A}{4 L_P^2}+\frac{3k_B}{2c L_P^2}V H.
\label{termo}
\end{equation}
In the following we study some interesting consequences of this formula.
\subsection{Dynamical degrees of freedom and the Cardy-Verlinde formula}
By a first inspection of the expression (\ref{termo}), we note that the added term is proportional to $VH$. This reminds
the expression for the entropy arising in the context of a conformal field theory representing a radiation with a given central charge $C$
\cite{bb3}, by the help of the Cardy formula. In this context, the entropy is given, in $3+1$ dimensions, by
$S_H=k_B HV/(2cL_P^2)$. In the case considered in \cite{bb3} for a closed Friedmann cosmology,
the normalized dynamical Bekenstein bound is
$S\leq S_B=2\pi k_B ER/(3c\hbar)$ .
It is a simple matter to verify that with the normilized factor $1/3$, the
(dynamical) term proportional to
$H$ in (\ref{termo}) is exactly identical to the one found by Verlinde in \cite{bb3}. This is an interesting fact. Moreover,
the first term in our entropy (\ref{termo}) can solve the issue present for the Bekenstein bound when
one considers a closed universe near its maximum radius, without considering the covariant Bousso bound.
In this case, $H\simeq 0$ and $S_H\rightarrow 0$, but the entropy for a universe
near its maximum it is non-vanishing and so the term $\sim A$ in (\ref{termo}) allows the bound to be valid.
As pointed in \cite{bb3}, for a strongly self-gravitating universe ($R_u H>c$, being $R_u$ the radius of the universe),
the usual Bekenstein bound is still valid for regions larger than
the Hubble radius.
Conversely, for a (closed) universe with radius $R_u$ smaller than the Hubble one,
the added term only is not enough to make the bound
valid. In fact, our formula (\ref{termo}) implies that for a universe with a radius smaller than the Hubble one, the firs term
in (\ref{termo}) cannot be neglected.
As an intriguing suggestion, thanks to the reasonings above,
the term $\sim VH$ in (\ref{termo}) could be seen as due to the manifestation of
the Casimir energy (expansion energy) at the Hubble scale (cosmic apparent horizon):
the usual one $\sim A/4$ accounts for the remaining energy components of the universe.\\
However,
a more precise calculation must be done by considering that the formula (\ref{termo}) has been derivated in a flat universe, while the
arguments of \cite{bb3} do apply to a closed universe. For a closed universe \cite{14,15,17}, instead of (\ref{3}) the following
holds:
\begin{equation}
\delta M\frac{G}{c^2}<\frac{L}{2}+\frac{AH}{4\pi c}-\frac{3V}{8\pi a(t)^2}.
\label{closed}
\end{equation}
Note that in (\ref{closed}) the spatial dimensions are due to $a(t)$.
It is simple to verify that, under the condition that $\{L,V/(a(t)^2)\}>c/H$
(strong self-gravitating regime),
the first and the third therm on the right side of (\ref{closed}) are subdominant with respect to the 'Casimir like' one.
Hence, by considering the suitable
normalized dynamical Bekenstein bound in this case, we obtain
$S_{BH}=k_B HV/(2cL_P^2)$. It seems
difficult to believe that this is only a coincidence.
Also note that the first and third term in (\ref{closed})
are vanishing by taking $V=4/3\pi R^3a(t)^3, L=Ra(t)$.
Moreover, in the opposite regime (weak self-gravitating regime), or by considering
black holes smaller than the
Hubble radius, the leading term in
(\ref{closed}) is equivalent to the condition (\ref{3}), with $L=a(t)R(t)+o(1), V=4/3\pi R^3a(t)^3+o(1)$,
i.e. on sufficiently small scales, the negative curvature can be neglected.
We do not go further in this analogy, but this indicates that we are on the right road.
\subsection{Equation of state for black holes in a flat expanding universe}
In this section, we analyze some consequences of (\ref{termo}) in relation to the equation of state of the black hole.
By differentiating (\ref{termo}) we get
\begin{equation}
dS_{BH}=\frac{k_B}{4L_P^2} dA+\frac{3k_B}{2c L_P^2} V dH+
\frac{3k_B}{2c L_P^2}H dV.
\label{termo2}
\end{equation}
The first two terms in the right side of (\ref{termo2}) can be interpreted as representing $1/T$ times
the internal energy of the black hole.
In particular, the one involving $dH$ can be seen as a variation
of the internal energy
due to the expansion of the universe caused by the presence of some unspecified kind of matter.
In a physically viable usual cosmology (i.e. satisfying weak and strong energy conditions)
we have $dH\leq 0$. Hence, in an expanding universe this term lowers the internal energy.
The term proportional to $dV$ can be seen as a work
term $dW$ due to the Hubble flow: for an expanding universe ($H>0$),
when $dV>0$ we have $dW>0$, while when $dV<0$ obviously $dW<0$, a reasonable fact. The introduction of a term proportional to
$V$ in (\ref{termo2}) confirms that holographic principle struggles with cosmology and that more degrees of freedom are necessary, in particular the ones related to the non static nature of the spacetime.\\
Note that by using theorems present in \cite{13,17},
the expression (\ref{entr}) can also be generalized to a Friedmann universe
with negative curvature: in this case another positive term arises proportional to $\sim H^2A^2$ and in addiction more involved
expressions for $L$ and $A$ arise.
Another interesting consequence of the modified Bekenstein-Hawking formula is the possibility to write down, thanks to the volume
term $dV$ in (\ref{termo2}), an equation of state for the embedded black hole.
In fact, we can write:
\begin{equation}
\frac{P}{T} =\frac{3k_B H}{2c L_P^2},
\label{w1}
\end{equation}
where $P$ denotes the pressure. For a universe with negative curvature, a further positive term proportional to $V/a(t)^2$ also appears in
(\ref{w1}). The ratio $P/T$ measures the entropy variation with respect to a variation of the volume $V$. For a vanishing
$H$, it follows that $P=0$.
Hence, a non-vanishing
pressure for a black hole can be an indication of the non static nature of the universe.\\
By denoting with $R_H$ the Hubble radius,
we can write formula (\ref{w1}) in the following form:
\begin{equation}
P R_H L_P^2=\frac{3}{2}k_B T.
\label{ig}
\end{equation}
For an ideal gas we have $PV=N k_B T$, where $N$ is the particles number.
It is interesting that in the formula (\ref{ig}) explicitally emerges the apparent horizon.
After multiplying both members of
(\ref{ig}) for the proper volume $V$ of the apparent horizon of the black hole, we have:
\begin{equation}
P V=\left(\frac{3V}{2 R_H L_P^2}\right)
k_B T.
\label{ig2}
\end{equation}
Suppose now to decompose the effective
proper volume $V(t)$ of the apparent horizon in $n(t)$ elemenatary spherical cells of fixed proper radius $L_P$, i.e.
$V=4/3\pi n(t)L_P^3$. This can be justified, has remarked in section 5,
in light of the papers \cite{v1,v2} in the context of a quantum non-commutatice spacetime
at the Planck length, where a minimal volume arises as a consequence of well motivated space-time uncertainty relations.
Then expression (\ref{ig2}) becomes:
\begin{equation}
PV=\left(2\pi\frac{L_P}{ R_H}\right) n(t) k_B T.
\label{cb}
\end{equation}
The expression (\ref{cb}) generalizes the usual equation of state
suitable for ideal gases in the context of black hole thermodynamics in Friedmann flat expanding spacetimes. Formula
(\ref{cb}) can have a fine physical interpretation. When the universe is cold, as at present time, many quantum degrees of freedom
are frozen: this is described by the actual very low ratio of $L_P/R_H$.
But when the universe has been hot, more and more degrees of freedom have been
excited ($L_P/R_H\sim 1$).
At the Planck epoch, when $R_H\sim L_P$, we have $PV\sim n k_B T$, with $2\pi$ a geometric factor due to the
sphericity of the black hole (there is not reason to put a sphere into a rectangular box). A similar phenomenon happens for the
ordinary statistical mechanics. Note that $n(t)$ is an integer and so it is $R_H/L_P=N_P(t)$, provided that the length $R_H$ is
expressed in terms of the Planck length $L_P$.
Hence the equation (\ref{cb}) can be written in the
expressive form:
\begin{equation}
PV=\left(2\pi\frac{n}{ N_P}\right)k_B T,
\label{trek}
\end{equation}
a kind of Bohr-Sommerfield quantization rule for the black hole equation of state. We do not speculate further on this formula.
The presence of a volume term in (\ref{termo2}) could struggle with the holographic principle. However, note that
the work term $PdV$ of usual thermodynamics arises thanks to a volume dependence of the entropy, and is unavoidable.
Otherwise, no work would be associated to the expansion of $A$, that seems in a cosmological context rather unlikely.
It should also be noticed that the theorem (\ref{3}) remains valid if we substitute the energy-density
${\overline{\rho}}_m$ with a constant energy-density, i.e. in a de Sitter expanding universe. In such a case, we have
again the formula (\ref{termo}), but with $H(t)$ the constant de Sitter value
$H=\sqrt{\Lambda/3}$. In this case, the term involving $dH$ in (\ref{termo2}) is vanishing, and as a result the internal energy of a
black hole in a de Sitter universe is left unchanged with respect to the asymptotically flat case. Equation
(\ref{w1}) it gives $P/T\sim \sqrt{\Lambda/3}$.
\section{Entropy Bound and Its Possible Generalization}
First of all, note that our formula (\ref{termo}) reduces, in a static background ($H=0$),
to the usual black hole one. In this context, the
entropy bound $S_A(V)\leq k_B A/(4L_P^2)$ can be violated, where $S_A(V)$ denotes the entropy within a region of volume $V$
enclosed by $A$. The argument is the following. Suppose (see for example \cite{bb3}) to have a spherical star with entropy $S_u$ collapsing to form a black hole. For the generalized second law, the black hole will have an entropy $S_{BH}\geq S_u$.
By following the star inside
its own horizon, thanks to the spacelike singularity,
its surface area $A$ will approach zero
in a finite time but its entropy is at least $S_u$ and as a result the Bekenstein bound is violated.
This simple example, with the help of theorem (\ref{3tris}), can give some hint to a better formulation of the entropy spherical bound.
Consider the entropy bound in this formulation:
{\it In a given background, the entropy of a matter system enclosed in a compact hypersurface of
volume $V$ and area $A$ cannot exceed the entropy of the biggest black hole formed if
a proper mass excess where present within $V$
violating the (\ref{3})}.
Note that with respect to this formulation, the area $A$ of this new entropy bound does not refer necessarily
to the area of the hypersurface enclosing
the system. This is an important difference that is in agreement with the phylosophy of the covariant Bousso bound. In
the example above, the biggest black hole is given by the one
formed by the collapse of the star itself and the bound it is not given by $1/4$ the area of the collapsing object also within the horizon,
but rather from $1/4$ the area of the horizon in Planck units that thanks to the generalized second law cannot be violated. This also
applies to a collapsing spherical shell within an existing black hole of area $A_h$,
since there the 'biggest black hole' is the one just existing with horizon, thanks to the contribution of the infalling shell,
greater than $A_h$.
Other examples violating the bound can be built by considering a system with a huge entropy. Also in this case, the judicious use of
relation (\ref{3}) can help us to solve these situations.\\
Consider, in an asymptotically flat spacetime, a system made of a spherical ball of
gas made of $N$ non interacting elements of rest mass $m_n$. Its entropy can be approximated (also in a cosmological context see
\cite{bb5}) by $S_N=k_B N$. The (spacelike) entropy bound is violated iff:
\begin{equation}
N>\frac{A}{4L_P^2}.
\label{vb1}
\end{equation}
Suppose now that the ball is composed of photons with wavelength $\lambda$ and energy
$E=\frac{\hbar c}{2\pi\lambda}$. From (\ref{3}) we see that the condition under which a black hole do not form is
$N<\pi L \lambda/L_P^2$. By setting the obvious inequality $\lambda\leq L$ we obtain
$N<A/(4L_P^2)$. Therefore, also for the huge entropic photons ball, the (spacelike) entropy bound cannot be violated.
The situation is much more involved in a cosmological context. In fact, since the universe is filled with a non vanishing matter-energy
tensor $T_{\mu\nu}$, we can calculate the entropy content of a given closed region of proper area $A$ and volume $V$. In this context,
the dinamical equilibrium is obtained between the background density $\rho(t)$ and the Hubble flow $H(t)$: the Friedmann
equations dictate that an higher density
$\rho$ implies an higher Hubble flow ($H^2\sim\rho$). Hence black holes, as well know, do not exist in Friedmann universes, also
near the big bang where $\rho\rightarrow\infty$: to build a black hole, a perturbation with a positive mass excess violating
the (\ref{3}) must be imposed.\\
Consider a flat Friedmann universe together with a
spherical homogeneous hypersurface
at fixed time $t$.\\
For the volume and area we have $V=4/3\pi L^3,\;\;A=4\pi L^2$. The entropy $S_m$ of this region
can be written as \cite{bb2} $S_m=k_B\sigma V$, where $\sigma=S/a(t)^3$ and $S$ is the space and time constant comoving entropy depending generally on the mass-energy content within the comoving region enclosed by $V$.
The spacelike bound with the old proposal $S_ {BH}\sim A/4$ is violated iff:
\begin{equation}
L>\frac{3}{4\sigma L_P^2}.
\label{vb2}
\end{equation}
The inequality
(\ref{vb2}) is satisfied for very large volumes
and sufficiently near the big bang \cite{bb6} for causal usual fluids ($\alpha\geq 1/3$) below the particle horizon.
In our frame, the (\ref{vb2}) becomes:
\begin{equation}
\sigma V_s>\frac{A_h}{4L_P^2}+\frac{3}{2L_P^2}V_h H.
\label{vb3}
\end{equation}
Note that in (\ref{vb3}) do appear the proper dimensions of the black hole with mass excess violating the (\ref{3}) and not the one of
{\bf s}. On general grounds we expect that $L_h\geq 2G M/c^2$ and thus
\begin{equation}
L_h\geq L_s+2L_s^2\frac{H}{c}.
\label{vb4}
\end{equation}
Also by taking the equality relation for $L_h$ in (\ref{vb4}), in the (\ref{vb3}) do appear terms containing the volume $V_s$ that can
'struggle' with the volume term $\sigma V_s$. To this purpose, consider the more restrictive case by putting
$A_h=0$ in (\ref{vb3}) and $L_h=L_s$ in (\ref{vb4}), after denoting with $t_P$ the Planck time,
the (\ref{vb3}) becomes:
\begin{equation}
S>\frac{3\alpha}{2 c L_P^2 t_P^{\alpha}}t^{3\alpha-1},
\label{vb5}
\end{equation}
where we have fixed the scalar factor to be $a(t_P)=1$.
Note that, thanks to our added term only in the entropy formula,
the comoving length of the hypersurface (two sphere) {\bf s} does not appear explicitely
but only the constant entropy density $S$. However,
this dependence can be contained in $S$ and thus it
must be verified if this dependence agrees with Einstein's equations predictions
or if
(\ref{vb5}) it is satisfied for a region larger than the dimensions of the
observable universe. To be more quantitative, consider once again
$S_m=k_B N_s(t)=k_B \sigma(t)V_s(t), \sigma(t)=N_s/V_s=\rho/m(t) $, where $m$ denotes the rest mass of the particles filling the universe that can also be a time function. This case is more general than the adiabatic case that can be regained by setting
$N_s(t)=const$.
Hence, by using
the Einstein's equation $H^2=8/3\pi G{\rho}$, the inequality $S_m>k_B \frac{3}{2L_P^2}V_s H$ (restrictive hypothesis)
with $E=mc^2$ becomes:
\begin{equation}
H>\frac{4\pi G E}{c^3L_P^2}.
\label{vb6}
\end{equation}
The most favorable situation can obtained with photons of proper wavelength ${\lambda}_p$.
We finally get: $H>2c/{\lambda}_p$.
By taking ${\lambda}_p=c/H$ (Hubble radius of the universe), we obtain $1>2$,
a contradiction. The adiabatic case is obtained as a subcase with $N_s(t)=const$ and
$E\sim 1/a(t)$ for photons in an expanding universe.
This simple example shows as the added term works.
The use of the Hubble radius is justified from the fact that it is an apparent cosmological horizon.
It is interesting to note that, with the use of the old expression $S_{BH}\sim A/4$, we obtain that the bound is exactly saturated
at the Hubble radius $c/H$, but it is violated for larger regions within the particle horizon. Conversely, with our added term only,
the bound is exactly saturated at the particle horizon, a reassuring fact. This certainly implies that our suggested improvement of the
entropy bound certainly works and that perhaps also a stronger version ($L_s\sim L_h$) can be suitable.
Note that we have considered only the spherical case.
However, our proposal for a generalized bound can also be formulated for
non-spherical black holes. In fact, we can advantage of the isoperimetric Penrose inequality
(see \cite{v1,v2} and references therein) and its generalization \cite{v2} in a cosmological context to justify our expression
(\ref{termo}) also for non spherical configurations.
As a final consideration of this section, note that all the reasonings above are valid until
a classical description of the geometry is available. At the Planck scale, a classical description of the geometry is certainly
uncorrect. Suppose to explore the quantum world.
To this purpose, consider the original spacelike entropy bound applied to photons within an enclosed volume.
By using the Heisemberg uncertainty relation,
we obtain $A\geq 4\pi L_P^2$ and hence a minimal surface (and also volume) does appear.
This can be obtained in the context of a non-commutative geometry \cite{v1,v2}, also for non-spherical volumes,
leading to a quantum spacetime at the Planck length. The existence of a minimal surface and a minimal volume could be the true
reason for which a limit exists on the informations that a certain region can contain. Moreover, in an expanding universe, due to the Hubble flow, this minimal surface (and volume) becomes smaller by using theorem (\ref{3}) (see \cite{v2}), allowing
more information (entropy) to be stored within a given surface, according to our formula (\ref{termo}).
\section{Primordial Inflation from Black Hole Evaporation?}
The idea that black hole physics could inspirate an alternative mechanism to begin inflation is not new
(see for example \cite{infl1,infl2}). In particular in \cite{infl1} the apparent horizon is considered as a valid object capable to produce an
ingoing radiation similar to the Hawking one allowing inflation.\\
In \cite{infl2} micro black holes remnants can induce a matter dominated universe before the inflation.\\
In this subsection we shortly explore the possibility that primordial black hole evaporation
can provide a mechanism to begin primordial inflation. As a first step, suppose that the primordial inflation soon after the Planck era is
dominated by a foam of micro black holes.
As a title of example, we consider a single black hole with size $L$ that can be greater equal or less than
the Hubble radius $c/H$.
As shown in \cite{infl1}, this universe can be considered adiabatic and open since of the evaporation process. Contrary to the
mechanism shown in \cite{infl1}, the Hawking radiation born from black hole evaporation produces an outgoing flux.
Since for an adiabatic universe $TdS_{BH}=dQ$, $V=4\pi/3 L^3, A=4\pi L^2$ (i.e. $L$ is the proper radius of the apparent horizon)
and after deviding for $dt$ we obtain (see \cite{infl1}):
\begin{eqnarray}
& &\frac{\pi k_B}{L_P^2}2L L_{,t}+\frac{2\pi k_B}{cL_P^2}L^3H_{,t}+\frac{6\pi k_B}{cL_P^2}HL^2L_{,t}=
-P_{rad},\nonumber\\
& &P_{rad}=\frac{{\pi}^2k_B^4}{60c^2{\hbar}^3}A T^3,\;\;T=\frac{\hbar c}{2\pi k_B L}
|1-\frac{L_{,t}}{2H L} |. \label{i1}
\end{eqnarray}
The second and third term at the right hand side of (\ref{i1}) are a consequence of our modification of the original Bekenstein-Hawking
entropy.
In the following we investigate the viability that the equation (\ref{i1}) allows an inflationary quasi de Sitter phase. Soon after the Planck era the universe expanded very quickly. Therefore we may reasonably suppose that in such a phase $L_{,t}>0$, i.e. that the black holes are
influenced by the tremendous Hubble flow.
However, as the time
flow, at early times,
the expansion rate decreases and the evaporation mechanism
becomes more and more efficient. As a consequence, there existed
an early phase where
$L_{,t}\simeq 0, L_{,t}>0$ and $H_{,t}\simeq 0$. For $H_{,t}>0$ we have a blue-shifted expansion while for
$H_{,t}<0$ we have a red-shifted inflation. The Planck data \cite{plan} predicts a red-shifted inflation, while the more recent BICEP2
one \cite{Bic} could account also for
a blue-shifted inflation \cite{mad}. However the combination of Planck and BICEP2 data is still in good agreement with a red-shifted inflation.\\
First of all, note that in this phase certainly $|L_{,t}/L|<<H$ and hence
$T\sim 1/L$.
Suppose that at the beginning of the inflationary phase $H_I\geq c/L_I$,
where the pedix '$I$' denotes the value at the starting of inflation.
Also suppose that $|H_{,t}/H|>> |L_{,t}/L|$ (which implies a very small expansion rate for the apparent horizon of the micro black hole)
and hence the second term in the left
hand side of (\ref{i1}) dominates:
as a consequence from (\ref{i1}) we obtain:
\begin{equation}
H=H_I-\frac{c^2L_P^2}{240\pi L_I^4}\left(t-t_I\right)+o(1),
\label{i2}
\end{equation}
i.e. a red-shifted inflation. The ratio $L_P^2/L_I^4$ measures the non gaussianity level of the perturbations.
Note that the presence of the factor $L_P^2$ in (\ref{i2}) can hint a possible non-commutative
signature of the inflation. In the limit $L_P=0$ the evolution is purely de Sitter.
Since
we expect that $\Delta t=t_F-t_I\in[10^{-34},10^{-32}] s$ ($t_F$ denotes the end of the inflation), we must have
\begin{equation}
H_I>>\frac{c^2L_P^2\Delta t}{240\pi L_I^4}.
\label{hi}
\end{equation}
In the opposite regime, i.e. $|H_{,t}/H|\leq |L_{,t}/L|$ and $H_I<<c/L_I$, we obtain
\begin{equation}
L^3=L_I^3-\frac{cL_P^2}{720\pi}\left(t-t_I \right)+o(1).
\label{i3}
\end{equation}
We obtain a monotonically decreasing expression for $L$ during inflationary epoch. This solution struggles with the
reasonable hypothesis that $L_{,t}>0$ just before inflation. If we dismiss this assumption, solution
(\ref{i3}) can lead to a blue-shifted inflation by considering, for example $H\sim c/L$.
These preliminary crude estimations indicate that black hole inflation as a viable possibility. Apart from equation
(\ref{i1}), we must also consider the Einstein's equations. In particular, we can consider a universe filled with a foam of primordial black holes
that can be modelized as dust with density ${\rho}_{mbh}$ (see \cite{infl2}) together with the radiation ${\rho}_{rad}$. The Einstein's equations are thus $H^2=8/3\pi G\left({\rho}_{mbh}+{\rho}_{rad}\right)$ together with the continuity equations for
${\rho}_{mbh}$ and ${\rho}_{rad}$.
This is certainly matter of further investigations.
\section{Conclusions}
In this paper we have considered the Bekenstein-Hawking entropy formula in flat expanding universes in relation to known theorems of the
general relativity for the formation of trapped surfaces. In particular, we quoted the theorems present in
\cite{14,15,17}. Thanks to these theorems,
we are in the position to propose a generalization of the Bekenstein-Hawking entropy in a flat expanding
universe. The expression contains an extra term with respect to the usual expression taking into account the non static
nature of a cosmological background represented by the hubble flow $H$.\\
This new proposal for $S_{BH}$
allows us to write the equation of state of a black hole (equation (\ref{cb})). Interestingly enough, this equation of state depends on the ratio
$L_P/R_H$, that measures the deviation from the usual equation of state for an ideal fluid. This factor becomes of the order of unity
when $L_P\sim R_H$, i.e. near the Planck era.
Our approach can be also extended to Friedmann universes with negative and positive curvature with embedded a black hole
(although in the case of closed universes it is not yet clear how to define a black hole). The only difference with the flat case
concerns the expressions for the effective volume, proper length and proper area of the black hole, where geometrical factors do appear.
In this paper we have considered spherically symmetric configurations. Hence, the proper area $A$ and the proper effective volume
$V$ are not independent quantities. Nevertheless, the identification $A^{3/2}\sim V$ in (\ref{entr}) is certainly well motivated.
If where
possible to consider the expression (\ref{termo}) for non spherical configurations
(Kerr black holes embedded in Friedmann universes), then we could consider transformations with
$dA=0$ and $dV\neq 0$ that
leave the old entropy formula $S_{BH}\sim A/4$ unchanged. Conversely, in our context we have an effective term $\sim VH$ that is not
left unchanged by an isoperimetric transformation. Fortunately, the theorem leading to inequality (\ref{3bis})
has been generalized in \cite{15} to non-spherical configurations with the condition that $S$ be an equipotential convex surface,
allowing to write for $S_{BH}$ an expression closed to the formula (\ref{termo}), but with the difficulty
to introduce a good definition for the volume of a non spherical black hole. In any case, we point out this important conceptual difference
with the old proposal.
We also explore the possibility that primordial inflation can be enhanced by black hole evaporation. Simple preliminar computations show
that this possibility is compatible with our equations. In particular, it is thanks to our added term that red-shifted inflation can be easily
obtained.
This can be certainly matter for further investigations.
As a final remark, note that an higher value for the entropy with respect to the usual formula in expanding universes implies that it is possible
to store more informations (remember that $I=S/\ln(2)$) in a given region than the ones contained in its surface and thus that holographic principle is not enough for a suitable knowledge of our universe.
\section*{Acknowledgements}
I would like to thank Alessandra D'Angelo for interesting discussions and suggestions,
Luca Tomassini for many useful discussions on the quantum nature of the spacetime.
Finally, I also thank the interesting comments of the
anonymous referee that contributed to an improvement of this paper.
|
\section{Introduction}
Wolf - Rayet (WR) galaxies show broad emission lines in the optical spectrum
due to WR stars. Broadband studies of the spectrum, from radio to optical, of
the dwarf galaxy He 2-10 indicated the presence of $10^3$ to $10^4$ WR stars
\citep{allen1976}. The discovery of WR galaxies has been serendipitous until a
catalogue of 37 WR galaxies was first compiled by \citet{conti1991}. 139 new
members were added by \citet{schaerer1999}. Currently 846 WR galaxies are
known, following the SDSS Data Release 6 \citep[DR6;][]{adelman2008}. WR
features in a galaxy provide useful information about the star-formation
processes in the system like the age and strength of the starburst. The WR
phase is characterised by the ejection of the outer layers of evolved massive
stars by stellar winds. These galaxies, along with blue compact dwarf and
irregular star-forming galaxies, are collectively referred to as H{\sc ii}
galaxies.
\begin{table*}
\caption{General Properties of the Galaxies}
\begin{tabular}{@{}llllllllll@{}}
\hline
Galaxy & Right Ascension$^{(1)}$ & Declination$^{(1)}$ & Distance & $d_{25}$$^{(5)}$ & $m_B$$^{(5)}$ & 12+log$\frac{O}{H}$ & $v_{hel}$$^{(10)}$ & F$_{H\alpha}$ \\
& (J2000) & (J2000) & (Mpc) & (arc min) & (mag) & & $({\rm km s^{-1}})$ &$({\rm ergsec^{-1}cm^{-2}})$ \\
\hline\hline
NGC 4214 & $12^h15^m39.2^s$ & $+36^{\circ}19'37''.0$ & 2.94$^{(2)}$ & $8.91\pm1.32$& $10.17\pm0.21$ & $8.28\pm0.08$$^{(6)}$& $291\pm3$ & $1.47\times10^{-11}$$^{(8)}$\\
NGC 4449 & $12^h28^m11.9^s$ & $+44^{\circ}05'40''$& 3.9$^{(3)}$ & $5.63\pm0.29$& $9.65\pm0.62$& $8.31\pm0.07$$^{(7)}$& $202\pm7$& $2.03\times10^{-11}$$^{(9)}$\\
\hline
\end{tabular}
\\
\label{tab1:Galaxy Parameters}
\footnotesize{References: (1) NED; (2) \citet{maiz2002}; (3) \citet{annibali2008};
(4) \citet{deVaucouleurs1991}; (5) \citet{paturel1997}; (6) \citet{kobulnicky1996};
(7) \citet{martin1997}; (8) \citet{martin1998}; (9) \citet{hunter1993}; (10) \citet{bottinelli1990} }\\
\end{table*}
Some of these H{\sc ii} galaxies have been studied at radio wavelengths. A
sample of 26 blue compact dwarf (BCD) galaxies was studied at high radio
frequencies ($>$1 GHz) by \citet{klein1991}. For five galaxies, they found a
steep non-thermal spectrum ($\alpha < -0.7$) after separation of the thermal
free--free component of the radio emission. They suggested a new relation
between the radio spectral index and optical luminosity where they find that
the spectral index is steeper for luminous objects. The galaxies were found to
follow the radio - FIR correlation observed in normal disk galaxies. A high
resolution radio continuum study between 1.4 GHz and 15 GHz along with
H$\alpha$ observations of nine WR galaxies revealed flat ($\alpha > -0.4$)
radio spectra and in many cases inverted spectra between 5 and 15 GHz
\citep{beck2000}. They explain that this is due to WR galaxies harbouring
young starbursts and hence the spectrum is dominated by free-free emission from
the star-forming regions. However, a spectral study of seven such H{\sc ii}
galaxies at 325 MHz by \citet{deeg1993} revealed a range of spectral behaviour.
They noted that a variety of emission, absorption and energy-loss mechanisms
are responsible for the shape of the radio spectra. They do not find steep
non-thermal spectra in their sample, unlike \citet{klein1991}. A recent radio
continuum study of five BCD galaxies using the Giant Metrewave Radio Telescope
(GMRT) by \citet{ramya2011} also resulted in a range of spectral shapes at low
radio frequencies, which they suggest could be indicative of environmental
effects. They found that the extended radio emission is generally detected in
galaxies residing in a group environment whereas localised emission, which
seems to correlate with the H$\alpha$ emission, is detected in galaxies
evolving in a relatively isolated environment. Thus, the former shows a
steeper spectrum as compared to the latter. However, their sample is small and
it will be interesting to check for a larger number of H{\sc ii} galaxies.
It is important to increase the sample of H{\sc ii} galaxies studied at low
radio frequencies and disentangle the different physical mechanisms which play
a role in shaping their evolution and spectra. We are studying the low radio
frequency properties of a sample of seven WR galaxies using the GMRT. We plan
to combine this with the higher radio frequency data and model their spectra,
separate the thermal contribution to the spectrum and estimate the non -
thermal spectral index. It will be interesting to examine the influence of
environment on the morphology of the radio continuum emission and the low radio
frequency spectrum and compare the low radio frequency continuum properties of
these low-metallicity systems with disk star-forming galaxies.
In this paper, we present low-frequency study of two well-known WR galaxies of
our sample, namely, NGC 4214 and NGC 4449. Both galaxies are classified as
Magellanic irregular and are located in the Canes Venatici region. Some
properties of these galaxies are listed in Table \ref{tab1:Galaxy Parameters}.
In \textsection 2 we describe observations, data calibration and imaging
processes. The results are presented in \textsection 3 and discussed in
\textsection 4. The conclusion is presented in the last section. We use the
convention of spectral index $\alpha$ defined via $S\propto\nu^\alpha$.
\subsection{NGC 4214}
NGC 4214 is a nearby barred S-shaped Magellanic irregular galaxy having nearly
face-on orientation \citep{allsopp1979}. It has been identified as a starburst
galaxy with an inner starburst region and an older red disk and has been the
subject of extensive multi-wavelength research. NGC 4214 has two main regions
of star formation - the northern one (RA=12h15m39s, DEC=$36^{\circ}19'35''$,
J2000) has a shell-type morphology visible at high angular resolution near the
centre of the galaxy which is known in literature as NGC 4214-I. A smaller
second star forming complex located to the south (RA=12h15m40s,
DEC=$36^{\circ}19'09''$, J2000) of NGC 4214-I is known as NGC 4214-II. NGC
4214 is a member of the group LGG 291 \citep{garcia1993} which contains 14
galaxies. The two closest members are the dwarf galaxy DDO 113 and the
star-forming irregular galaxy NGC 4190 having similar radial velocities as NGC
4214. These galaxies are within angular separation of $\leq 30'$ from NGC 4214
and consequently hace been in the field of view. The distance to NGC 4214
reported in literature ranges from 2 Mpc to 7 Mpc. Throughout this paper, we
use a distance of 2.94 Mpc \citep{maiz2002} and thus 1$''$ corresponds to 14
pc. NGC 4214 is a gas-rich blue galaxy with a thick disk \citep{maiz1997}.
Diffuse X-ray emission from hot gas is detected in this galaxy
\citep*{ott2005}. Intense UV emission observed from this galaxy is believed to
be due to its low dust content, leading to inefficient processing of UV to
longer wavelengths \citep{fanelli1997}. \citet{mackenty2000} detected several
young ($<$10 Myr) star-forming complexes of various ionised gas morphologies
with sizes $\sim$10-200 pc in their HST observations. In addition, they
reported that the extended diffuse ionised gas in NGC 4214 contributes 40$\%$
of the total H$\alpha$ and O III $\lambda$5007$\AA$ emission of the galaxy.
The H{\sc i} gas in this galaxy is observed to be $\sim1.4$ times the Holmberg
radius in extent \citep{allsopp1979}.
NGC 4214 has been studied in the radio continuum by \citet{allsopp1979,
beck2000, mackenty2000} and recently by \citet{kepley2011}. \citet{beck2000}
studied the radio continuum emission from the star forming regions NGC 4214-I
and NGC 4214-II at 15, 8.3, 5 and 1.4 GHz using VLA at high angular resolution.
Interestingly, they found that the spectrum was rising between 5 and 15 GHz,
indicating the presence of optically thick free-free thermal emission.
\citet{kepley2011} estimated a magnetic field of $30 \mu$G in the central parts
and $10 \mu$G at the edges and do not detect significant polarisation on scales
larger than 200 pc.
\subsection{NGC 4449}
NGC 4449 is a typical type 1 Magellanic irregular galaxy with star formation
occurring throughout the galaxy at a rate almost twice that of the LMC
\citep{thronson1987, hunter1999}. The distance to the galaxy reported in the
literature ranges from 2.9 Mpc to 5.4 Mpc. Throughout this paper, we use a
distance of 3.9 Mpc given by \citet{annibali2008} and thus $1''$ corresponds to
19 pc. This galaxy has been studied extensively in the optical bands ( e.g.
\citet{crillon1969, kewley2002}), in the radio \citep{seaquist1978, klein1986,
hunter1991, klein1996}, in the H{\sc i} \citep{rogstad1967, vanWoerden1975} and
in CO \citep{tacconi1985, sasaki1990, bottner2003}. NGC 4449 appears to be
undergoing tidal interaction with the neighbouring dwarf galaxy DDO 125
\citep{hunter1998, theis2001}.
\citet{klein1996} have studied NGC 4449 using WSRT and detected a large
synchrotron halo ($\sim 7$ kpc) around the central star forming region at 610
MHz. At the distance of 3.9 Mpc, the halo extent is about 7.4 kpc
\citep{klein1996}. They found the radio spectrum of the synchrotron radiation
between 610 MHz and 24 GHz to be $\sim -0.5$ in the central regions and $\sim
-0.7$ in the halo regions. They fitted a combination of thermal and
non-thermal emission to the integrated spectrum between 610 MHz and 24 GHz and
found the global non-thermal spectral index to be $-0.7$. This galaxy is found
to be rich in H{\sc i}, with M$_{HI}$ = $2.4\times10^{9}{\rm M_\odot}$
\citep{epstein1964} and the H{\sc i} halo extends to $\sim$6 times the Holmberg
radius \citep{hunter1998}.
X-ray observations of NGC 4449 show a circular region of diffuse emission
$\sim$ 1$'$ north of the centre of the galaxy \citep{ott2005}.
\citet{chyzy2000} studied the polarised radio continuum emission of NGC 4449,
at 8.46 and 4.86 GHz and found a regular galaxy-scale magnetic field of
strength $6-8 \mu$G. They point out that the strong magnetic fields are
intriguing since the galaxy exhibits slow rotation which cannot support the
dynamo action required to amplify the magnetic field. They also detect the
large continuum halo at 8 GHz.
\section[]{Observations and Data Reduction}
Both the galaxies were observed using the GMRT; \citep{swarup1991, ananth2005}
in the radio continuum at 150, 240, 325 and 610 MHz. The 240 and 610 MHz
observations were simultaneous in the dual band mode. It is remarked that the
240 MHz data of both the galaxies were corrupted to the extent that even after
extensive flagging the final image quality was suspect. We hence did not use
240 MHz data in the analysis presented here. Since the galaxies are closely
located in the sky, they were observed in the same observing run with common
calibration scans. The target source scans alternated between the two galaxies
resulting in better {\it uv-}coverage for both the galaxies.
The 240 and 610 MHz observations were simultaneous in the dual band mode. All
the observations were done in the non-polar mode of observing known as the
Indian polar mode. The details of these observations are given in
Table~\ref{tab2:Observational Parameters}. We also analysed GMRT archival data
at 610 MHz for NGC 4214 from August 2007. These single frequency observations
have better sampling of short spacings and hence are presented in this paper.
\begin{table*}
\centering
\caption{Our Observations}
\label{tab2:Observational Parameters}
\begin{tabular}{@{}llll@{}}
\hline
Observing Band (MHz) & 150 & 325 & 610$^{\#}$ \\
\hline\hline
Observation date & 28/10/2012 & 23/12/2009 & 20/12/2008 \\
Pointing Centre - NGC 4214 & $\alpha = 12^h15^m39.2^s$ and $\delta = +36^{\circ}19'37''.0$ & & \\
Pointing Centre - NGC 4449 & $\alpha = 12^h28^m11.9^s$ and $\delta = +44^{\circ}05'40''.0$ & & \\
Receiver Bandwidth (MHz) & 16 & 32 & 16 \\
No. of Antennas$^{*}$ & 27 & 29 & 28 \\
Primary Calibrator(s) & 3C286 & 3C147 & 3C147, 3C286 \\
Flux density (Jy)$^{\dagger}$ & 31.05 & 52.69 & 38.12, 21.01 \\
Secondary calibrator(s) & 3C286 & 1123$+$055 & 1227$+$365 \\
Flux density (Jy)$^{\ddagger}$ & 31.05 & 6.7$\pm$0.15 & 1.8$\pm$0.03 \\
Telescope time (hrs) & 9 & 7 & 7 \\
\hline
\end{tabular}
\\
\footnotesize{$^{\#}$ Dual band observation (240 \& 610 MHz).
$^*$ Maximum number of operational antennas during the observation.
$\dagger$ Set by SETJY task : Using (1999.2) VLA or Reynolds (1934-638) coefficients.
$\ddagger$ Flux density and error from GETJY}\\
\end{table*}
\begin{table*}
\caption{Image parameters}
\label{tab3:Achieved Parameters}
\begin{tabular}{@{}lllllll@{}}
\hline
Galaxy Name & \multicolumn{3}{c}{NGC 4214} & \multicolumn{3}{c}{NGC 4449} \\
\hline
Observing Bands & 150 (MHz) & 325 (MHz) & 610 (MHz) & 150 (MHz) & 325 (MHz) & 610 (MHz) \\
\hline\hline
Beam Size(ROBUST=0) & $21''.7\times17''.8$ & $10''.5\times8''.6$ & $5''.4\times4''.8$ & $20''.5\times17''.6$ & $11''.3\times8''.7$ & $5''.3\times5''.0$ \\
Position Angle & $71^{\circ}.7$ & $61^{\circ}.2$ & $47^{\circ}.2$ & $62^{\circ}.9$ & $56^{\circ}.7$ & $0^{\circ}.5$ \\
RMS noise level (mJy beam$^{-1}$) & 3.2 & 0.34 & 0.09 & 2.3 & 1.4 & 1.1 \\ \\
\hline
Beam Size (ROBUST=5)& $30''.3\times24''.4$ & $32''.0\times13''.3$ & $8''.5\times6''.5$ & $28''.8\times23''.6$ & $15''.3\times11''.1$ & $7''.6\times7''.1$ \\
Position Angle & $77^{\circ}.0$ & $3^{\circ}.3$ & $42^{\circ}.6$ & $65^{\circ}.5$ & $59^{\circ}.1$ & $5^{\circ}.2$ \\
RMS noise level (mJy beam$^{-1}$) & 6.2 & 0.56 & 0.13 & 4.3 & 1.9 & 0.22 \\
\hline
Shortest spacing (k$\lambda$) & 0.04 & 0.06 & 0.20 &0.04 & 0.06& 0.20 \\
& $\sim 77$ m & $ \sim55$ m & $\sim 98$ m &$\sim 77$ m &$ \sim55$ m& $\sim 98$ m \\
Longest spacing (k$\lambda$) & 13 & 27 & 50 &13 &27& 50 \\
& $\sim 25$ km & $\sim 25$ km & $\sim 24$ km & $\sim 25$ km & $\sim 25$ km &$\sim 24$ km \\
Largest Visible structure$^{\#}$ & $\sim 52\arcmin$ & $\sim 34\arcmin$ & $\sim 10\arcmin$ & $\sim 52\arcmin$ & $\sim34\arcmin$ & $\sim10\arcmin$\\
\hline
\end{tabular}
\\
\footnotesize{Note: $^{\#}$ Corresponding to the shortest spacing present in our data.}
\end{table*}
The observations started and ended with a 15m observing run on an
amplitude/bandpass calibrator (3C 286 or 3C 147). The phase calibrator was
observed for about 5m after every 20m on a target galaxy. All the
observations except the last one in 2012 used the old, now decommissioned GMRT
hardware correlator. The 2012 run used the new GMRT software backend. These
observations used a baseband width of 16 MHz. The raw data obtained in the
local $lta$ format were converted into FITS and this data file was then
imported into the NRAO AIPS{\footnote{AIPS software is produced and maintained
by the National Radio Astronomy Observatory, a facility of the National Science
Foundation operated under cooperative agreement by Associated Universities,
Inc.}} software. The data were edited, calibrated and imaged using standard
tasks in AIPS. After an initial round of flagging, the single channel data of
the amplitude calibrator were calibrated and then the editing of bad data and
calibration were iteratively repeated. The data for antennas with relatively
large errors in antenna-based gain solutions were examined and flagged. Once
the gain calibration was satisfactory, the calibrated data were used to obtain
the bandpass calibration tables. Once the bandpass calibration became
satisfactory, i.e. the closure errors were less than one percent, the bandpass
gain tables were applied to the data and the channel data averaged to generate
continuum data sets. To avoid bandwidth smearing at low frequencies due to
large fields of view, we averaged 20 channels at 610 MHz, 7 channels at 325 MHz
and 4 channels at 150 MHz corresponding to channel widths of 1.8 MHz, 0.9 MHz
and 0.4 MHz, respectively. Thus a SPLAT data base with these channel widths
was generated for further analysis. The amplitude and phase calibrator data
were used to obtain the gains of the antennas which were then applied to the
target source. Bad data on the target source were edited and the final
calibrated visibility data were obtained.
These data were then imaged by creating multiple facets across the primary beam
to avoid wide-field effects in the image plane. 25 facets were used for
imaging data at 610 MHz whereas 49 facets were used at 325 and 150 MHz. This
image was used as the model for the first round of phase-only self-calibration.
On average two to three rounds of phase-only self-calibration were required,
followed by a final round of amplitude and phase self-calibration. This
procedure improved the images at all the frequencies. All facets were combined
using the task FLATN to produce a final map at each observing frequency.
Moreover, we made maps using a range of different weighting schemes with
robustness parameters, where +5 is a pure natural weight and -5 is a pure
uniform weight \citep{briggs1995}. This resulted in different beam sizes and
signal-to-noise ratios. We also made images after including a uv cutoff in the
data to examine the presence of features on different angular scales. The
final images were corrected for primary-beam attenuation before undertaking
flux density measurements.
The presence of a strong source close to NGC 4449 affected the image quality
and to improve the image, we removed this source using the task UVSUB in AIPS.
While this did improve the images, it was difficult to remove all the
artifacts. In particular the final image at 325 MHz was badly affected and
therefore, we do not present the image but include the flux density estimates
at this frequency, with the caveat that the value quoted might have a large
systematic error.
\section{Results}
The final images made at 610, 325 and 150 MHz for the two galaxies are shown in
Figures~\ref{fig3:4214.ROB0} and ~\ref{fig15:4449} and the image parameters
along with the results are listed in Table~\ref{tab3:Achieved Parameters}.
The flux densities for both the galaxies estimated at the different GMRT
frequencies are listed in Table~\ref{tab7:Flux Densities}, along with the
values obtained from literature. The spectrum is shown in
Figure~\ref{fignew10:4449}. It is to be noted that the systematic errors on
the GMRT flux denstities estimated using the noise statistics in a similar
region around the source were larger than the statistical errors.
\subsection{NGC 4214}
The images of NGC 4214 at 150, 325 and 610 MHz made with robust weighting= 0
are shown in Figure~\ref{fig3:4214.ROB0}.
The 325 MHz image, which has an rms noise of 0.34 mJy/beam, detects the
emission from both the central region and the diffuse halo .
The 150 MHz detects only the compact emission from the central parts of the
galaxy while the 610 MHz image has picked up emission from diffuse regions
immediately surrounding the central compact regions.
In Figure~\ref{fig3:sub2} the radio emission is observed to extend towards the
north from the main disk of the galaxy along a narrow ridge which connects to
the northern region. The emission at 325 MHz appears to be fragmented at
several locations in the main disk of the galaxy, some of which could be
imaging artifacts. We do not detect the diffuse extended emission at 150 MHz.
To understand this let us note that the 1 sigma rms noise in the image is 3.2
mJy/beam for a beam size of about $22''\times18''$, whereas the brightness at
150 MHz, arrived at after employing the brightness of the diffuse region at 325
MHz and a spectral index of $-$1, is expected to be 3.3 mJy beam$^{-1}$ and
which is below the 3 sigma sensitivity of our 150 MHz image. As shown in
Figure~\ref{fig3:4214.ROB0}, we detect the intense star forming regions NGC 4214-I (northern) and
NGC 4214-II (southern) in the centre of NGC 4214.
We also analysed the 1.4 GHz VLA archival data (AM236) observed in A
configuration and have detected the star-forming regions NGC 4214-I and NGC
4214-II in addition to another star-forming complex in the north-west of these
regions. Moreover at all the three bands, emission is detected from a region
located in the north around right ascension of $12^{h}15^{m}40^{s}$ and
declination of $36^{\circ}23'$.
NVSS images show extended emission from this galaxy. We made a low resolution
map at 325 MHz to match the resolution of the NVSS. These maps, were then used
to generate a spectral index distribution across the galaxy. The spectral index
across the galaxy varies from $\sim -1.5$ in the outer parts of the halo
emission to $\sim -0.5$ in the central parts of the galaxy.
NGC 4190 ($\alpha = 12^h13^m44.8^s$ , $\delta = +36^{\circ}38'03''$(J2000))
lying within the field of view of NGC 4214 is detected at 325 MHz with a flux
density of 8.1$\pm$1.0 mJy.
\subsection{NGC 4449}
The 150 and 610 MHz images made with robust=5 weighting which are sensitive to
the extended emission are shown in Figure~\ref{fig15:4449}. A spiral arm-like
extension from east to west in the north of the galaxy is seen
(Figure~\ref{fig15.sub1}). Diffuse radio emission is detected with two radio
peaks embedded within it. The southern compact source is the centre of the
galaxy whereas the northern source is the supernova remnant J1228+441. The
fragmented emission is likely due to imaging artifacts introduced by the sparse
longer baselines in the data.
The presence of a strong source located about 7$'$ to the southwest of the
galaxy resulted in enhanced artifacts in our images.
\begin{figure*}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm,angle=-90]{4214.150.ROB0.ps}\label{fig3:sub3}}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm,angle=-90]{4214.330.ROB0.ps}\label{fig3:sub2}}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm,angle=-90]{4214.610.ROB0.ps}\label{fig3:sub1}}
\caption{\small{The radio contour images of galaxy NGC 4214 with robust
weighting 0 are presented. The radio contours are overlaid on radio grey
scale. (a) 150 MHz, beam size = $21''.7\times17''.8$, p.a.= $71^{\circ}.7$, the
contours are 3.5$\times$(-3, 3, 4.2,6,8.4,12) mJy/beam;
(b) 325 MHz, beam size = $10''.4\times8''.6$, p.a.= $60^{\circ}.0$,
the contours are 0.34$\times$(-6, -3, 3, 4.2, 6, 8.4, 12) mJy/beam, ;
(c) 610 MHz, beam size = $5''.4\times4''.8$, p.a.= $47^{\circ}.2$, the contours are 96 $\times$(-3, 3, 6, 12, 24) $\mu$Jy/beam.}}
\label{fig3:4214.ROB0}
\end{figure*}
\begin{figure*}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm]{4449.150.ROB5.ps}\label{fig15.sub3}}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm]{4449.610.ROB5.ps}\label{fig15.sub1}}
\caption{The radio continuum maps of NGC 4449 made using robust
weighting of 5 with full uv range are presented here. The radio contours are
overlaid on its radio grey scale. (a) 150 MHz contours are 4.0 $\times$(-3,
3,5,7,9,11,13,15) mJy/beam, beam size = $28''.8\times23''.6$ , p.a. =
$65^{\circ}.4$ ; (b) 610 MHz contours are 0.21 $\times$(-6,6,8.4,12,16.9,
24,33.9,48) mJy/beam, beam size = $7''.6\times7''.1$, p.a. = $5^{\circ}.2$.}
\label{fig15:4449}
\end{figure*}
The double source seen to the south of NGC 4449 in both the 610 and 150 MHz
images is most likely a distant radio galaxy as suggested by \citet{chyzy2000}.
\begin{table}
\centering
\caption{Flux Densities for NGC 4214 and NGC 4449}
\label{tab7:Flux Densities}
\begin{tabular}{@{}cccc@{}}
\hline
$\nu$ (MHz) & NGC 4214 & NGC 4449 & References \\
& $S_{\nu}$ (mJy) & $S_{\nu}$ (mJy) & \\
\hline
150 & $104.1\pm15.6^{\#}$ & $639\pm95.8^{\#}$ & 1,1 \\
325 & $192.5\pm43.9^{\#}$ & $785\pm225^{\#}$ & 1,1 \\
609 & & $480\pm10$ & 2 \\
610 & $74.6\pm11.2^{\#}$ & $370\pm55.5^{\#}$ & 1,1 \\
1400 & $56.9\pm0.4$ & $272\pm0.4^{\#}$ & 3,3\\
1400 & $51.5\pm10.3^{\#}$ & & 4 \\
1400 & $38.3\pm7.7$ & & 5\\
1400 & $70\pm25$ & & 6\\
1415 & & $224\pm22$ & 7\\
1600 & $65\pm25$ & & 6\\
2380 & $36.0\pm3.0^{\#}$ & & 8\\
2700 & & $167^{\#}$ & 9\\
3300 & $\leq50$ & & 6\\
4750 & & $138\pm10$ & 10\\
4850 & $30.0\pm4.5$ & & 11 \\
4850 & $30.0\pm7.0^{\#}$ & & 12\\
4860 & $34.0\pm6.8$ & & 4 \\
5000 & & $135\pm16^{\#}$ & 13 \\
8460 & $20.5\pm0.5$ & & 14\\
8460 & $24.2\pm4.8^{\#}$ & & 4\\
10550 & & $87\pm6^{\#}$ & 2\\
10770 & & $92\pm10$ & 15\\
10770 & & $83\pm7$ & 10\\
24500 & & $67\pm10^{\#}$ & 10\\
\hline \\
\end{tabular}
\\
\footnotesize{ The estimated systematic errors are quoted here.\\
{\#} shows the datapoints plotted in Figure~\ref{fignew10:4449} \\
References:(1) This paper; (2) \citet{klein1996}; (3) NVSS; (4) \citet{kepley2011};
(5) \citet{condon2002}; (6) \citet{mas-Hesse1999} Radio data taken at Nancay Radio Telescope;
(7) \citet{hummel1980}; (8) \citet{dressel1978}; (9) \citet{haynes1975};
(10) \citet{klein1986}; (11) \citet{becker1991}; (12) \citet{gregory1991};
(13) \citet{sramek1975}; (14) \citet{schmitt2006}; (15) \citet{israel1983}}
\end{table}
NVSS detects extended emission enclosing the entire disk of NGC 4449. We used
the NVSS 1.4 GHz map of NGC 4449, which has an angular resolution of 45$''$ and
150 MHz (robust=5) image after smoothing the latter to the NVSS resolution to
generate a spectral index map of the radio emission. The spectral index across
the galaxy varies from $-$1.5 in the outer regions to $-$0.5 in the inner
regions.
\begin{figure*}
\subfigure[]{\includegraphics[height=6.5cm,width=6.5cm]{4214GALEXFUV.PS}\label{fig6:4214GALEXFUV}}
\subfigure[]{\includegraphics[height=6.5cm,width=6.5cm]{4214.2MASS.eps}\label{fig7:4214.2MASS}}
\caption{NGC 4214, 325 MHz radio contour are superposed on (a) GALEX-FUV gray
scale (b) 2MASS gray scale in which the labelled numbers represents the sources
in SDSS as: (1) star, (2) star cluster, (3) star, (4) 2MASX galaxy 2MASX
J1253795+362218 (5) NGC 4214:[HSS2004]X-rayS, (6) star, (7) star. The contour
level is same as in Figure~\ref{fig3:sub2}.}
\end{figure*}
\subsection{Estimating the thermal emission}
The continuum emission at radio frequencies is mainly due to two emission
processes - namely synchrotron (also referred to as nonthermal) emission from
relativistic electrons accelerated in an ambient galactic magnetic field and
thermal free--free emission from star forming H{\sc ii} regions. Owing to the
steep spectral index, of the nonthermal emission ($\rm {\alpha_{nt}}< -0.5$) compared to the
thermal emission which has a spectral index of $-0.1$, it is expected to
dominate at low ($\lesssim1$ GHz) radio frequencies. Relativistic electrons are
believed to be accelerated in supernova shock fronts by diffusive shock
acceleration mechanism and are injected into the interstellar medium (ISM) with
$\rm {\alpha_{nt}}\sim-0.5~{\rm to}-0.7$ \citep{bell1978, bogdan1983, biermann1993}. This
has been confirmed by observations of galactic supernova remnants
\citep[see][]{kothes2006, green2009}. The electrons, then, propagate away and
lose energy in the process mainly due to synchrotron and/or inverse-Compton
cooling and thereby changing $\rm {\alpha_{nt}}$ \citep[see e.g.][]{longair2011}. The study
of $\rm {\alpha_{nt}}$ can throw meaningful insight into the energy loss/gain mechanisms of
relativistic electrons giving rise to the nonthermal emission. However, due to
the presence of thermal emission, the global spectral index $\alpha$ is
contaminated and is flatter than actual, particularly in giant H{\sc ii}
regions harbouring recent star formation activity.
This, therefore, necessitates separating the thermal emission from the total
emission in these galaxies which have massive ongoing star formation
\citep{walter2001, reines2008}. We estimate the thermal emission from these
galaxies, employing H$\alpha$ emission, following \citet{tabatabaei2007}, since
H$\alpha$ and thermal free--free emission arise in the same ionised gas. The
thermal emission flux density ($S_{\rm \nu, th}$) at a frequency $\nu$ is given
by,
\begin{equation}
S_{\rm\nu,th} = \frac{2k T_{\rm e} \nu^2}{c^2}(1 - e^{-\tau})
\end{equation}
Here,
$k$ is the Boltzmann constant, $c$ is the velocity of light, $T_{\rm e}$ is the
temperature of thermal electrons assumed to be $10^4$ K and $\tau$ is the
free--free optical depth. $\tau$ is related to the emission measure ($EM$) as
\begin{equation}
\tau = 0.082 T_{\rm e}^{-1.35} \left(\frac{\nu}{\rm GHz} \right)^{-2.1}
(1+0.08)\left(\frac{EM}{\rm cm^{-6}pc}\right)
\end{equation}
The $EM$ is determined from the
H$\alpha$ intensity ($I_{\rm H\alpha}$) of the galaxy using equation (9) in
\citet{valls-gabaud1998}, \begin{equation} I_{\rm H\alpha} = 9.41\times10^{-8}
T_{\rm e4}^{-1.017}10^{-0.029/T_{\rm e4}} \frac{EM}{\rm cm^{-6} pc} ~~~{\rm
erg~cm^{-2}s^{-1}sr^{-1}} \end{equation} Here, $T_{\rm e4}$ is the electron
temperature in units of $10^4$ K.
Thermal emission was computed for NGC 4214 and NGC 4449 for each pixel from the
H$\alpha$ maps observed with the 2.3-meter BOK telescope at Kitt Peak National
Observatory, downloaded from the NED. The galaxies were observed at a centre
wavelength of 6850 \AA, with a filter width of 68.7 \AA. The counts per second
in these maps were converted to apparent magnitude $m_{\rm AB}$ using the
zero-point given in the FITS file header. The specific intensity ($f_{\nu}$) was
calculated from $m_{\rm AB}$ i.e. $f_\nu ({\rm erg~s^{-1}cm^{-2}Hz^{-1}}) =
10^{-(m_{\rm AB}+48.6)/2.5}$, which was converted to the desired flux units of
$\rm erg~s^{-1}cm^{-2}$ from, $f = f_\nu d\nu = f_\nu c (d\lambda/\lambda)^2$.
The thermal emission map determined based on the H$\alpha$ maps has an
angular resolution of $\sim2''$. The optically thin free--free emission was then
removed from the total flux density at all frequencies, and the remnant flux was
assumed to be due to synchrotron emission. Integrated radio spectrum of NGC 4214 and
NGC 4449 are displayed in Figure~\ref{4214integ.ps} and \ref{4449integ.ps} where solid, dot-dashed and dashed lines
represent, respectively, the total, the nonthermal, and the the thermal components.
A power law fitting gives us $\alpha_{\rm nt} = -0.63\pm0.04$ for NGC 4214 and $\alpha_{\rm nt} = -0.49\pm0.02$ for NGC 4449.
We note that the H$\alpha$ emission suffer due to extinction from dust present
in these galaxies. The dust extinction was estimated, using the optical depth,
$\tau_{\rm dust}$, of the obscuring dust at FIR wavelength of $\lambda160~\mu$m
\citep{tabatabaei2007}. We used the MIPS data of the Spitzer space telescope
to estimate the dust temperature and thus $\tau_{\rm dust}$. $T_{\rm dust}$ was
estimated by fitting a modified black body spectrum of the form, $f_{\rm IR}
\propto \nu^\beta B_{\rm IR}(T_{\rm dust})$ between $\lambda$70 and 160 $\mu$m,
where $B_{\rm IR}(T_{\rm dust})$ is the Planck function. Here, $f_{\rm IR}$ is
the flux at a FIR frequency $\nu_{\rm IR}$ and $\beta$ is the power-law index
of dust absorption efficiency assumed to be 2 \citep{draine84}. The optical
depth at H$\alpha$ wavelength was estimated by extrapolating $\tau_{\rm dust}$
using the standard dust model for our galaxy \citep{kruegel2003}. In dense
H{\sc ii} regions, the extinction was found to be $\sim20-30\%$, while in other
parts it was $<10\%$. Thus, the H${\alpha}$ flux and thereby the thermal
emission at radio frequencies determined by us in this study are lower limits.
\section{Discussion}
\subsection{NGC 4214}
\subsubsection{Morphology and correlation with emission at other wavebands}
\begin{figure*}
\subfigure[]{\includegraphics[height=7cm,width=7cm,angle=-90]{4214integ.ps}\label{4214integ.ps}}
\subfigure[]{\includegraphics[height=7cm,width=7cm,angle=-90]{new4449integ.ps}\label{4449integ.ps}}
\caption{Integrated radio spectrum of (a) NGC 4214 and (b) NGC 4449. In both
the spectra the data points $<$1.4 GHz are from this paper. The solid line
represents the total, the dot-dashed the nonthermal, and the dashed line the
thermal components. The fitting for the total and non-thermal was done
excluding the points at 325 and 150 MHz for NGC 4214 whereas for NGC 4449 data
point at 325 MHz was excluded. Thus by a power law fitting we get $\rm {\alpha_{nt}}$ = $-0.63\pm0.04$ for NGC
4214 and $\rm {\alpha_{nt}}$ = $-0.49\pm0.02$ for NGC 4449. The details of the plotted data
are given in Table \ref{tab7:Flux Densities}. }
\label{fignew10:4449}
\end{figure*}
In Figures~\ref{fig6:4214GALEXFUV} and ~\ref{fig7:4214.2MASS}, the radio
contours at 325 MHz are shown superposed on the GALEX\footnote{Galaxy Evolution
Explorer is an orbiting ultraviolet space telescope.} FUV (1350 - 1750\AA) and
2MASS NIR images. Scattered star--forming regions are seen to the west, south
and north of the intense star forming disk of NGC 4214 in the UV images,
whereas the NIR emission has a boxy morphology. Radio emission at 325 MHz is
detected from the entire FUV disk. The angular extent of the 325 MHz halo is
about 2.3 kpc. While radio emission is detected from a part of the northern
star forming region, no radio emission is seen to be associated with the
ultraviolet emission in the west and south of the disk of the galaxy.
We note that there are two more discrete radio continuum regions in the field,
similar to as pointed out earlier by \citet{kepley2011} - one in the northern
part of NGC 4214 and the other in the northeast. While FUV emission is
associated with the western part of the northern star forming region, no such
emission is associated with the source in the north-east. Our high-resolution
images (e.g. Figure~\ref{fig3:sub1}) show the source in the north-east to be a
double source and our results support the conclusion by \citet{kepley2011}
about it being a background radio galaxy.
The NUV and FUV emissions show the galaxy to be undergoing vigorous star
formation, which extends along a bridge-like feature to the north. Radio
emission at 325 MHz is also detected along the bridge-like extension and the
northern star forming region. The SFR estimated from the non-thermal emission
is $1.97\times10^{-8}~{\rm M_{\odot} yr^{-1}pc^{-2}}$. The UV emission appears
to be of irregular morphology, with the star formation triggered at various
locations along a thick and almost north-south ridge. While the extent along
the major axis of the galaxy is similar, the radio continuum appears to be boxy
in nature (Figure~\ref{fig7:4214.2MASS}).
The old stellar population is distributed in a spherical halo resembling a
large globular cluster as is also seen in the I band image of
\citet{fanelli1997}. The radio continuum appears to be more extended than the
NIR emission. No NIR emission is detected along the bridge-like extension or
from the northern star-forming region, indicating a fairly recent star
formation episode in these regions. The radio emission associated with this
northern region is more extended in the east compared to the FUV (see
Figure~\ref{fig6:4214GALEXFUV}). Two radio peaks are visible in this northern
region at 610 and 325 MHz, with the eastern peak coinciding with the 2MASS
source, 2MASX J12153795$+$3622218, which is classified as a background galaxy
(see Figure~\ref{fig7:4214.2MASS}). The western radio peak region shows the
presence of FUV emission and is a star-forming region in NGC 4214. At 150 MHz
the emission from the northern region is seen to be exceptionally bright in
comparison with the rest of the galaxy. The spectrum of this northern region
between 150 and 610 MHz is fitted by a single power-law with a spectral index
of $-1.1$. X-ray emission has been detected from this region by
\citet{ott2005}. The emission at 150 MHz shows a single peak coincident with
the 2MASS galaxy.
Thus, we suggest that the northern region consists of two distinct parts - the
western part which is associated with the star-forming region in NGC 4214 and
the eastern part which is associated with the background galaxy 2MASX
J1253795+362218 \citep*{srivastava2011}. The two peaks are easily identifiable
in our 610 MHz image in Figure~\ref{fig3:sub1}. The radio galaxy to the
north-east and the galaxy in the north are likely part of a larger distant
cluster.
An alternative scenario is provided by \citet{kepley2011} who suggest that the
entire northern part is a star-forming region of age $10-13.5$ Myr and the
emission at 20 cm is elongated due to a uniform magnetic field with a strength
of 7.6 $\mu$G.
The central star forming regions are well resolved and compare well with the
higher radio frequency maps by \citet{beck2000}. The molecular gas has been
reported confined to three distinct regions - NGC 4214-I, NGC 4214-II and and a
region $\sim$ 760 pc to the north-west \citep{walter2001}. Star-forming regions
are close to all the three clouds and radio continuum emission is detected from
these. Starbursting dwarf galaxies tend to be H{\sc i} - rich
(\citealt{thuan1981}; \citealt{taylor1995}). The H{\sc i} from NGC 4214 is
reported to be extended \citep{allsopp1979}. However, the radio continuum at
325 MHz is found to bear little resemblance to this.
\subsubsection {The global spectrum}
In this subsection, we discuss the global spectrum of NGC 4214 after combining
our three-frequency radio data with data from literature. The data are listed
in Table~\ref{tab7:Flux Densities} and the entries marked by a \# are plotted
in Figure~\ref{4214integ.ps}. The integrated spectral index down to 610 MHz is
fitted by a power-law spectrum of index $-0.45\pm0.03$. This is similar to the
value of $- 0.43\pm0.06$ determined by \citet{kepley2011} for $\nu >$1 GHz.
The presence of the halo emission detected at 325 MHz makes the galaxy similar
to normal star-forming disk galaxies and NGC 4449 \citep{klein1996}. We note
that the low frequency spectrum includes a contribution from the region located
to the north of the galaxy.
Integrated radio spectrum presented in Figure~\ref{4214integ.ps} reveals that the thermal
emission starts becoming significant at higher radio frequencies, with about
{40\%} of the total emission near 5 GHz and 22{\%} at 610 MHz being thermal in
origin. This is significantly higher than $\sim$10$\%$ thermal emission at 1
GHz found for normal star-forming galaxies. We determine the synchrotron flux
density by subtracting the thermal emission from the total flux density at all
frequencies of observation and represent it in Figure~\ref{4214integ.ps} by
dot-dash line.
It is pointed out that the flux density is high at 325 MHz due to contribution
from diffuse halo emission and is low at 150 MHz due to sensitivity issues,
hence, we excluded these points from the power-law fit to the non-thermal
spectrum and
and we estimated $\alpha_{\rm nt}$ to be $-0.63\pm0.04$ for NGC 4214. Thus a
combination of synchrotron spectrum with index $-0.63$ and free-free thermal
emission with index $-0.1$ explains well the observed spectrum between 610 MHz
and 8.7 GHz (Figure~\ref{4214integ.ps}).
From the spectral index map made between our 325 MHz and 1.4 GHz (NVSS), we
find that the spectrum of the halo emission in the outer parts of the galaxy is
$\sim -1.5$, indicative of an aged electron population.
\subsubsection{The compact star forming regions: NGC 4214-I and NGC 4214-II}
As shown in Figure~\ref{fig3:4214.ROB0}, we detect the intense star forming
regions in the centre of NGC 4214: NGC 4214-I and NGC 4214-II. These regions
are detected at 325 MHz as surrounded by diffuse extended. To study these
compact regions, we imaged the GMRT data by excluding baselines shorter than
3$k\lambda$, so that the extended emission was resolved (see
Figure~\ref{4214HIGRES}).
We study the spectra of these two regions from 325 MHz to 15 GHz after
including the high frequency data between about 4 to 15 GHz from
\citet{beck2000} (see Figures~\ref{4214HIGRES}, \ref{radio.spec}).
As noted in the previous section, the low sensitivity of the 150 MHz image did
not allow detection of the diffuse emissions from these regions. Besides, the
flux densities of NGC 4214-I and NGC 4214-II are also lower than predicted by a
non-thermal spectrum (see Figure~\ref{radio.spec}), implying that thermal
absorption could also be a reason for non-detection of the diffuse emissions.
Since it is difficult to disentangle the two effects with the current data, we
have excluded the 150 MHz point from our analysis. The spectra of both the
regions display similar behaviour but significantly different from the global
spectrum (see Figure~\ref{4214integ.ps}), especially at high radio frequencies.
The spectra shown in Figure~\ref{radio.spec} appear to consist of three
components: the synchrotron non-thermal emission dominating at the lowest
frequencies, the optically thin $(\tau < 1$) free-free thermal emission and the
optically thick ($\tau >1$) free-free thermal emission dominating at the higher
frequencies.
We note that the last component ($\tau >1$) has been studied by
\citet{beck2000} who suggested a turnover frequency for this component to be
$\sim 15$ GHz. We have tried to ensure that all the data points used in our
analysis (see Figure~\ref{radio.spec}) were sensitive to the largest angular
scale that we are probing for the two regions. We have excluded 8 GHz data
point \citep{beck2000} from our analysis on the ground that the largest angular
scale probed by them was 15$''$ and hence, it is likely that the flux is missed
out.
In the first step of the analysis, the optically thin thermal emission was
estimated by extrapolating from the H$\alpha$ map of the region. No correction
for internal extinction was made to the H$\alpha$ map. The estimated spectrum
is shown by the dashed lines in Figure~\ref{radio.spec}. We, then subtracted
this thermal emission from the total flux density at each frequency. In the
second step, the remaining emission was fitted by a combination spectrum of the
type:
\begin{figure*}
\begin{center}
\subfigure[]{\includegraphics[height=6.5cm, width=4.5cm]{4214.330.LOWRES.ps}\label{fig11.sub2}}
\subfigure[]{\includegraphics[height=6.5cm, width=4.5cm]{4214.610.LOWRES.ps}\label{fig11.sub1}}
\subfigure[]{\includegraphics[height=6.5cm, width=4.5cm]{4214.1400.LOWRES.ps}\label{fig11.sub3}}
\caption{Images showing NGC 4214-I and NGC 4214-II at (a) 325 MHz, contours are 0.3$\times$(3,4,5,6,7,8,9) mJy/beam;
(b) 610 MHz; contours are 0.1$\times$(3,4.2,6,8.4,12,16) mJy/beam and (c) 1.4 GHz image from VLA
archival data, contours are 0.16 $\times$(-3, 3, 4.2, 6, 8.4) mJy/beam. All
images are made by excluding baselines shorter than 3k$\lambda$. }
\label{4214HIGRES}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\subfigure[]{\includegraphics[height=7cm,width=7cm,angle=-90]{4214NW.ps}\label{radio.spec_sub2}}
\subfigure[]{\includegraphics[height=7cm,width=7cm,angle=-90]{4214SE.ps}\label{radio.spec_sub3}}
\caption{The spectra of star-forming regions (a) NGC 4214-I;
$\alpha_{nth}=-0.32\pm0.02$ (b) NGC 4214-II; $\alpha_{nth}=- 0.94\pm0.12$.
In both plots the solid line represents the final model values; the dashed
line shows the thermal spectrum estimated using $S\propto\nu^{-0.1}$; the
dot-dash line represents the non-thermal emission fitted by the best fitting
power law and the filled square data points are estimated by subtracting the
thermal and non-thermal contribution from the total value and shows the
variation of the flux density as $\nu^{2}$ relation (small dashed line). The
data points at frequencies less than 1 GHz are from the GMRT, the 1.4 GHz data
is from VLA archival data and the data points at frequencies $>$1.4 GHz are
from \citet{beck2000}.}
\label{radio.spec}
\end{center}
\end{figure*}
\begin{equation}
f(\nu)=a\nu^{\alpha_{nt}}+c\nu^{2}
\end{equation}
This simple model resulted in a reasonably good fit (solid line curve in
Figure~\ref{radio.spec}) to the data for both the complexes.
While the optically thick component was fixed to vary as $\nu^2$, the
non-thermal spectral index was a free parameter in the model. The best fits
gave $\alpha_{nth}$ =$- 0.32\pm0.02$ for NGC 4214-I and $\alpha_{nth}$ =$-
0.93\pm0.12$ for NGC 4214-II.
The thermal contribution due to the optically thin component for NGC 4214-I at
325 MHz, 610 MHz and 1.4 GHz are about 35\%, 40\% and 42\% whereas for NGC
4214-II, it is about 32\%, 49\% and 54\% respectively. \citet{kepley2011}
also estimate the thermal emission at 1.4 GHz in central parts to be about 50\%
of the total emission. The complex spectrum of NGC 4214-I and II indicate the
presence of $\tau >$1 ultra-compact HII regions, $\tau<$1 diffuse thermal and
non-thermal emission.
A detailed look at Figure~\ref{4214HIGRES} shows two radio peaks in NGC 4214-I,
the eastern peak is diffuse and intense at 325 MHz whereas the western peak is
strongest at 1.4 GHz. This indicates a steeper spectrum in the east and hence
possibly an older star-forming region. The H$\alpha$ image (see
Figure~\ref{ngc4214Halpha}) shows bright compact regions in both the eastern
and western half of NGC 4214-I surrounded by diffuse emission. The eastern
H$\alpha$ peaks are displaced to the south of the radio peak at 610 MHz whereas
the western H$\alpha$ peak coincides with the radio peak.
We recall that such displacement is commonly seen in the circumnuclear rings in
disk galaxies and is believed to be due to the presence of dust. Good
correlation between the location of the 610 MHz emission and the H$\alpha$
emission is also observed in NGC 4214-II (see Figure~\ref{ngc4214Halpha})
indicating lower dust extinction in this star forming region. We note that
this galaxy is overall a dust-deficient galaxy \citep{mackenty2000} with the
stellar clusters in NGC 4214-I of age $\sim$3~-~4 Myr, whereas the clusters in
NGC 4214-II are younger at about 2.5~-~3 Myr.
\begin{figure}
{\includegraphics[height=8.5cm,width=10cm]{610halpha.ps}}
\caption{Radio contours showing the 610 MHz emission
overlaid on the H$\alpha$ image of NGC 4214.
$1$ shows NGC 4214-I and $2$ shows NGC 4214-II}
\label{ngc4214Halpha}
\end{figure}
The radio luminosity at 610 MHz is $1.2\times10^{19}$ WHz$^{-1}$ for NGC 4214-I
and $5\times10^{18}$ WHz$^{-1}$ for NGC 4214-II for a resolution of about 200
pc. The luminosities of these star-forming regions (10$^{19}$ W~Hz$^{-1}$)
appear to be similar to those in circumnuclear rings of typical diameter
$500-1000$ pc in several normal disk galaxies observed with similar linear
resolution.
For example, \citet{kodilkar2011}, who have done a radio study of the
circumnuclear ring in NGC 2997 with a linear resolution of 200 pc, find that
the luminosity of the five star-forming clumps resolved in the ring at 1.4 GHz
is between $10^{19}$ to $10^{20}$ W~Hz$^{-1}$. Moreover, \citet{kodilkar2011}
estimate an equipartition magnetic field of about 30 $\mu$G for the
circumnuclear ring in NGC 2997 which is similar to that estimated for the
central regions of NGC 4214 by \citet{kepley2011}. Thus star-forming regions
in the centre of NGC 4214 appear to display properties similar to the centres
of normal disk galaxies. This is important since NGC 4214 is a dwarf WR
galaxy of lower metallicity and it is instructive to compare their star-forming
complexes with those in the better studied disk galaxies.
Finally, we comment on the possible corrections that can be included to improve
this study. Firstly, the correction for extinction in the H$\alpha$ signal due
to dust in NGC 4214 can be included.
We estimate a maximum variation of 20\% in the estimates of thermal emission
due to extinction.
Including dust extinction might lead to a synchrotron spectrum that is steeper
than the estimated $-0.32\pm0.02$ for NGC 4214-I. The effect of dust does not
seem to be observable in the H$\alpha$ image of NGC 4214-II (see
Figure~\ref{ngc4214Halpha}).
This would indicate that only small corrections will have to be made on the
estimated non-thermal spectrum. Secondly, the radio maps used to obtain this
interesting result are made with varying uv coverage of the fixed antenna
configuration of GMRT. Being low radio frequencies, the data are subject to
extensive flagging. Thus, there is a possibility that the images might be
sensitive to somewhat different angular extents.
\subsection{NGC 4449}
This galaxy has been extensively studied in various wavebands including low
radio frequencies. \citet{klein1996} studied the galaxy at 610 MHz with the
WSRT and detected an extended synchrotron halo around the optical galaxy.
\citet{chyzy2000} have detected the radio halo at 8.46 and 4.86 GHz and mapped
the galaxy in polarised emission. Our continuum observations trace the regions
of large scale regular magnetic field seen by \citet{chyzy2000}. Extended
radio emission was detected at all the observed GMRT bands of 150 MHz, 325 MHz
and 610 MHz.
\subsubsection{Comparing with emission at other wavebands}
The radio emission detected at 150 MHz and 610 MHz is shown in
Figure~\ref{fig15:4449}. We estimate the size of the 150 MHz halo to be about
4 kpc which is similar to the size of 3.5 kpc that \citet{chyzy2000} estimate
but smaller than the 7 kpc that \citet{klein1996} estimate. The spiral
arm-like feature in the north of the galaxy, clearly detected in our 610 MHz
images shown in Figures~\ref{fig25:4449.610GALEXFUV} and \ref{4449.6102MASS}
has highly ordered magnetic field \citep{chyzy2000, klein1996}. Polarization
study by \citet{chyzy2000} reveals intriguing galaxy-scale magnetic fields of
$6-8~\mu$G, even though the galaxy is a slow, chaotic rotator.
\begin{figure*}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm]{4449.610GALEXFUV.ps}\label{fig25:4449.610GALEXFUV}}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm]{4449.6102MASS.PS}\label{4449.6102MASS}}
\subfigure[]{\includegraphics[height=5.5cm,width=5.5cm]{610_0TO20_ROB5_4449.ps}\label{0TO20_ROB5_4449}}
\caption{NGC 4449, 610 MHz, Rob =5 radio contour is superposed on (a) GALEX-FUV
gray scale, (b) 2MASS gray scale. The contour level is
0.21$\times$(-6,6,12,24,33,48) mJy/beam (c) 610 MHz
made with ROB 5 and uvcutoff of 20 k$\lambda$ superposed on the DSS grey scale
image. Beam Size = $12''.30\times10''.40$, p.a. =$50^{\circ}.35$.
The contours are 0.41 $\times$(-3, 3, 6, 9, 12, 15, 18, 21) mJy/beam. }
\end{figure*}
NGC 4449 is estimated to be about five times more luminous than NGC 4214,
indicating vigorous star formation in NGC 4449. The star formation rates
estimated from the non-thermal emission is $3.35\times10^{-8}~{\rm
M_{\odot}yr^{-1} pc^{-2}}$. About 10 ULXs (X-ray luminosity $> 10^{39}{\rm erg
s^{-1}}$) have been detected in NGC 4449 \citep{ott2005} which confirms the
heightened level of activity in this galaxy. We have overlaid the radio
contours at 610 MHz on the GALEX FUV map in
Figure~\ref{fig25:4449.610GALEXFUV} and NIR image from 2MASS in
Figure~\ref{4449.6102MASS}. The FUV map show an almost north-south
distribution of vigorous star formation, whereas the 2MASS emission is boxy in
nature and centrally concentrated. This is similar to NGC 4214 discussed in
the previous section. While the 2MASS emission shown in
Figure~\ref{4449.6102MASS} is symmetric about the centre of the galaxy, the
ultraviolet and radio emission are asymmetric, with excess emission detected in
the north of the galaxy. The ultraviolet emission extends further south
compared to the radio continuum emission. The spiral arm-like feature seen in
the north of the galaxy shows a counterpart in the ultraviolet. However, it is
seen to skirt the NIR emission in the galaxy, indicating the recent
star formation there. The extent of disk in south is similiar to radio. The
X-ray emission is surrounded by an extended HI envelope \citep[see
e.g.][]{yun1994, hunter1998}.
\subsubsection {The global spectrum }
Our multi-frequency data alongwith the integrated flux density data available
in literature are listed in Table~\ref{tab7:Flux Densities}.
Figure~\ref{4449integ.ps} shows the global spectrum of NGC 4449 in the
frequency range of 150 MHz to 22 GHz. \citet{chyzy2000} have presented
spectral-index maps between 8.46 GHz and 4.86 GHz and find that the spectral
index changes from $\sim -0.3$ in strongly star forming regions to $\sim -1.1$
locally in the outer southern region. They also reported that the
radio-bright peaks coincide with strongly star-forming regions which show
increased thermal fractions (up to 80$\%$ at 8.46 GHz). \citet{klein1996},
estimated a thermal fraction at 1 GHz of $f_{th} = 10\pm4\%$ and a non-thermal
spectral index $\alpha_{nth}$ = $-$0.7$\pm$0.1.
Figure~\ref{4449integ.ps} reveals that about 20\% of the total emission near 5 GHz is thermal
in origin. The synchrotron spectral index for NGC 4449 using 150 MHz and 22 GHz
data is estimated to be $-0.49\pm0.02$. The 325 MHz point was excluded.
Interestingly, \citet{klein1996} record a flux density of 480 mJy at 610 MHz.
The spectral index between this and our 325 MHz data is $-$0.74. Thus a
combination of a synchrotron spectrum with index $-$0.49 and free-free thermal
emission with index $-$0.1 explains the observed spectrum between 150 MHz and
22 GHz as shown in Figure~\ref{4449integ.ps}. We estimate that about 9\% of
the total emission at 610 MHz is thermal in origin, which is similar to normal
disk galaxies but less than to what we estimate for NGC 4214 ($22\%$). One
possible reason could be the higher dust extinction in NGC 4449 leading to an
under estimation of the thermal fraction at radio frequencies from the
H$\alpha$ map. Alternatively, the continuous massive star formation in NGC
4449 might be arising enhanced non-thermal emission in the galaxy. The
spectral index is flatter than found in normal galaxies, suggesting continuous
star formation injecting energy into the relativistic plasma.
\subsubsection{The supernova remnant SNR J1228+441:}
The flux density of the supernova remnant SNR J1228+441 was estimated from
images made after excluding the short baselines. The SNR was unresolved in all
our images and is the most intense feature in NGC 4449 at all GMRT frequencies.
This SNR is five time more luminous than Cas A at 20cm \citep{chomiuk2009}. It
has been extensively monitored (e.g. \citet{lacey2007} and references
therein). \citet{lacey2007} report steepening of the spectral index from
$\alpha=-0.64\pm0.02$ in 1994 to $\alpha=-1.01\pm0.02$ in 2001-2002, showing
rapid evolution. \citet{reines2008} using high--resolution VLA data at several
frequencies find that the supernova remnant had a spectral index of $-1.8$
between 3.6 cm and 6 cm and an index of $-0.9$ between 1.3 cm and 3.6 cm
between 2001 and 2002, indicating a break in the spectrum. From our
observations at 325 and 610 MHz in 2008-2009, we estimate flux densities of
35.2$\pm$9.1 mJy and 11.2$\pm$2.9 mJy, respectively, and at 150 MHz in 2012, we
estimate a flux density of 32.4$\pm$5.1 mJy. The spectral index between 325
and 610 MHz is calculated to be $-1.8$ for epoch 2008-2009.
\subsection{Radio--FIR correlation}
\begin{figure*}
\subfigure[]{\includegraphics[height=7cm,width=7cm]{ngc4214_radio_fir.eps}\label{rf1.4214}}
\subfigure[]{\includegraphics[height=7cm,width=7cm]{ngc4449_radio_fir.eps}\label{rf1.4449}}
\subfigure[]{\includegraphics[height=5cm,width=5cm]{ngc4214_region.ps}\label{rf2.4214}}
\subfigure[]{\includegraphics[height=5cm,width=5cm]{ngc4449_region.ps}\label{rf2.4449}}
\caption{(a) The radio intensity of NGC 4214 at 325 MHz (solid black symbols)
and 610 MHz (gray symbols) vs. the far infrared intensity at $\lambda70~\mu$m
in units of Jy beam$^{-1}$. The circular and square symbols represents the
bright H{\sc ii} regions and diffuse emission respectively; (b) The radio
intensity of NGC 4449 at 150 MHz vs. the far infrared intensity at
$\lambda70~\mu$m in units of Jy beam$^{-1}$. The circular and square symbols
represents the bright H{\sc ii} regions and diffuse emission respectively; (c)
and (d) The white circles shows the the regions corresponding to the bright
H{\sc ii} regions and the black circles are the diffuse H{\sc ii} regions
overlayed on the H$\alpha$ image of the galaxy (c) NGC 4214 and (d) NGC 4449
after smoothing to 16$''$.}
\end{figure*}
The well known radio-far infrared (FIR) correlation is known to hold good over
five orders of magnitude in luminosity, with a dispersion less than a factor of
2. The radio emission is due to relativistic electrons being accelerated in
ambient galactic magnetic field, while the FIR emission is radiation from dust
heated by ultra-violet (UV) photons from massive ($\gtrsim10~\rm M_\odot$)
short lived ($\sim10^6$ yr) stars. The two regimes are thought to be connected
by star formation activity \citep{harwi75}. Small-scale amplification of the
magnetic field due to turbulent dynamo action couples the magnetic field ($B$)
and gas density ($\rho_{\rm gas}$) as $B\propto \rho_{\rm gas}^{\kappa}$
\citep[see e.g.,][]{cho00, grove03, chand53} and is believed to produce the
tightness seen in the correlation at local as well as on global scales
\citep[see e.g.,][]{nikla97, dumas11, basu12}. Here, $\kappa$ is the coupling
index predicted to lie in the range 0.4$-$0.6 by numerical simulations of
different ISM turbulence \citep{fiedl93, grove03, kim01, thomp06, murgi05}.
The slope of the radio--FIR correlation, $b$, defined via $S_{\rm radio} \propto
S_{\rm FIR}^b$ where $S_{\rm radio}$ is the flux density at a radio frequency
and $S_{\rm FIR}$ is the flux density of the FIR emission, is related to the
coupling index $\kappa$ giving rise to the tight radio--FIR correlation
\citep[see e.g.,][]{nikla97, dumas11, basu12}. However, note that this
prescription holds good if both the radio and FIR emission originate from the
same emitting volume. Owing to different propagation lengths of the cosmic ray
electrons (CREs) from the sites of generation, the slope of the correlation in
spatially resolved galaxies is expected to change \citep{basu12, berkh13}.
Here we study spatially resolved radio--FIR correlation at 16$''$ angular
resolution, corresponding to linear scale of $\sim0.23$ and $\sim0.3$ kpc for
the two galaxies NGC 4214 and NGC 4449, respectively. The correlation is
studied using the nonthermal radio emission at 325 and 610 MHz for NGC 4214 and
at 150 MHz for NGC 4449, and the FIR emission at $\lambda70~\mu$m observed
using the {\it Spitzer} space telescope and downloaded from the NED.
Figure~\ref{rf1.4214} shows the nonthermal radio flux density at 325 MHz (solid
black symbols) determined within regions of 16$''$ in size and 610 MHz (gray
symbols) with the FIR flux density at $\lambda70~\mu$m in units of Jy
beam$^{-1}$ for NGC 4214. The circular, square and triangular symbols represent
bright H{\sc ii}, diffuse H{\sc ii} and bridge emission respectively. The
different regions are shown in Figure~\ref{rf2.4214}. The data were fitted with
$S_{\rm radio} = a\times S_{\rm FIR}^b$ using an ordinary least-square bisector
method \citep{isobe90} in the log-log plane at both radio frequencies. The fits
to the data are shown by the dashed lines in Figure~\ref{rf1.4214}. The
emission at the two wavebands are found to be strongly correlated with
Pearson's correlation coefficient, $r=0.9$ at 325 MHz and $r=0.8$ at 610 MHz.
The slope at 325 MHz is found to be 0.47$\pm$0.04, flatter than that at 610 MHz
where the slope is 0.68$\pm$0.08.
The magnetic field in NGC 4214 has been estimated to be $\sim30~\mu$G towards
the bright H{\sc ii} regions and $\sim10~\mu$G in the diffuse regions
\citep{kepley2011}. The magnetic field in such bright HII regions is expected
to be tangled due to turbulence produced by star formation activity and by
propagation of CREs due to streaming instability at Alfv{\'e}n velocity
($v_{\rm A}$, $\sim 50\rm ~km~s^{-1}$). CREs emitting synchrotron radiation at
325 and 610 MHz would propagate $\sim1$ and $\sim$0.6 kpc within synchrotron
cooling timescales. Thus, at 325 MHz, the newly generated CREs and the older
population of CREs are well mixed, giving rise to comparatively higher radio
emission away from the H{\sc ii} regions. This makes the slope of the
radio--FIR correlation flatter at 325 MHz than at 610 MHz for the galaxy NGC
4214.
Figure~\ref{rf1.4449} shows the plot of nonthermal intensity at 150 MHz with
the $\lambda70~\mu$m intensity for the galaxy NGC 4449. The circles represent
bright H{\sc ii} regions tracing high star formation activity and the squares
represent regions of diffuse H$\alpha$ emission. Figure~\ref{rf2.4449} marks
the region in the smoothed 16$''$ H$\alpha$ image of the galaxy. The white
circles show the bright star forming regions and the black circles show the
diffuse H$\alpha$ regions. The slope of the radio--FIR correlation is found to
be 0.75$\pm$0.07 and is significant, with $r = 0.67$.
NGC 4449 shows a steeper slope compared to NGC 4214 for the radio--FIR
correlation when studied at 150 MHz indicating enhanced nonthermal emission in
regions of star-forming sites. Such an effect was also seen for this galaxy at
8.46 GHz \citep{chyzy2000}. The nonthermal spectral index is seen to be flat in
regions of dense star formation, with $\rm {\alpha_{nt}}\sim-0.5$ (estimated between the 150
and 610 MHz) and the overall nonthermal disc has $\rm {\alpha_{nt}} = -0.49\pm0.02$. This likely indicates
that the nonthermal emission in NGC 4449 predominantly originates from freshly
generated CREs which are emitting close to their sites of generation, making
the slope of the radio--FIR correlation steeper. Since the slope is similar to
that seen in star forming regions in normal galaxies ($\sim 0.8$;
\citet{basu12, basu2014}) it implies that the CREs have propagation length
scales lower than our linear resolution of $\sim 0.3$ kpc. This allows us to
put upper limits on the age of the CREs assuming they are propagating with an
Alfv{\'e}n velocity of $\sim50\rm ~km~s^{-1}$. It turns out that the bulk
CREs were produced less than $\sim 6\times 10^6$ ago, due to ongoing star
formation in NGC 4449. This is supported by the fact that the recent star
formation, as traced by the H$\alpha$ emission of NGC 4449 is about factor of
3--5 higher than that of NGC 4214 \citep{drozdovsky2002, ott2005, hunter1998}.
\section{Conclusions}
\begin{center}
\begin{table}
\caption{Results}
\label{final.summ}
\scalebox{0.70}{
\begin{tabular}{@{}lllllllll@{}}
\hline
Galaxy & $\alpha_{nth}$ & S$_{th}$ $\%$ & SFR & \multicolumn{3}{c}{b} \\
& & (610 MHz) & $10^{-8} M_{\odot} yr^{-1} pc^{-2}$ & 150 MHz & 325 MHz & 610 MHz \\
\hline
NGC 4214 & $-0.63(0.04)$ & 22 & $1.97$ & -- & 0.47(0.04) & 0.68(0.08) \\
NGC 4449 & $-0.49(0.02)$ & 9 & $3.35$ & 0.75(0.07) & --- & --- \\
\hline \hline
NGC 4214-I & $-0.32(0.02)$ & 40 & $2.1$ & \\
NGC 4214-II & $-0.94(0.12)$ & 49 & $2.5$ & \\
\hline
\end{tabular}
}
\\
\end{table}
\end{center}
In this paper, we have presented the low radio frequency images of the WR
galaxies NGC 4214 and NGC 4449 made using the GMRT. We detect a large radio
halo in NGC 4214 at 325 MHz whose morphology resembles the ultraviolet emission
seen by GALEX with an extent similar to the NIR emission from 2MASS. The
synchrotron spectral index ($\alpha_{nt}$) is $-0.63\pm0.04$. The thermal
fraction at 610 MHz is about 22\%. The spectra of two compact star--forming
complexes NGC 4214-I and NGC 4214-II are studied. We estimate that at 610 MHz
$\sim 40\%$ of emission from NGC 4214-I and $\sim 49\%$ from NGC 4214-II are
thermal in origin. The non-thermal spectral index of NGC 4214-I is estimated
to be $-0.32\pm0.02$ and of NGC 4214-II is $-0.94\pm0.12$. The combination of
compact star-forming regions and presence of an extended halo make this dwarf
galaxy similar to a normal star-forming galaxy. This galaxy follows the
radio-FIR relation. We find a significant correlation between the local 325
MHz and 610 MHz emission and the 70 $\mu$m emission for the entire galaxy, with
slopes of 0.47 and 0.68 respectively, which is within the range expected from
simulations of turbulence in the interstellar medium.
We detect a large radio halo around NGC 4449 at 150 MHz.
NGC 4449 is estimated to be five time more luminous than NGC 4214, indicating
vigorous star formation in the former. Both the galaxies are in a group
environment and hence their star formation properties are likely to be tidally
influenced by the other group members.
Separating the non-thermal from the thermal emission which is obtained from the
H$\alpha$ map of NGC 4449 results in $\alpha_{nt}$ of $-0.49\pm0.02$. For NGC
4449 we estimate a thermal fraction at 610 MHz of $\sim 9\%$ and find that the
150 MHz emission is well correlated with the 70 $\mu$m emission on small
scales, with a slope of 0.75. The non-thermal spectral index of both galaxies
is flatter than estimated for normal disk galaxies.
\section{Acknowledgements} {\small We thank the staff of GMRT who made these
observations possible. GMRT is run by the National Centre for Radio
Astrophysics (NCRA) of the Tata institute of Fundamental Physics. This research has
made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the
Jet Propulsion Laboratory, California Institute of Technology, under contract
with the National Aeronautics and Space Administration. This research has also
made use of the GALEX and {\it Spitzer Space Telescope} which are NASA mission
managed by the Jet Propulsion Laboratory. We also like to acknowledge the site
http://skyview.gsfc.nasa.gov/. SA acknowledges an INSA Senior scientist fellowship.
SS gratefully acknowledges the support of a
research grant SR/S2/HEP-08/2008 by the Department of Science $\&$ Technology
(DST), Govt. of India and support from NCRA. }
\bibliographystyle{mn2e} \footnotesize |
\section*{APPENDIX}\vspace{-.5\baselineskip}
\setcounter{section}{0}%
\setcounter{subsection}{0}%
\renewcommand\thesection{\Alph{section}}}
\begin{document}
\title{Time Versus Energy in the Averaged Optimal Coplanar Kepler Transfer towards Circular Orbits
\thanks{The second author was partially supported by \textit{Thales Alenia Space} and \textit{région Provence Alpes
Côte d'Azur}
}
}
\titlerunning{Averaged Optimal Coplanar Kepler Transfer}
\author{
Bernard Bonnard
\and
Helen C. Henninger
\and \\
Jana N{\v e}mcov\'a
\and
Jean-Baptiste Pomet
}
\authorrunning{B. Bonnard,
H. Henninger,
J. N{\v e}mcov\'a,
J.-B. Pomet}
\institute{
B. Bonnard \at
Institut de Mathématiques de Bourgogne, Université de Bourgogne, \\9 avenue Alain Savary, 21078 Dijon,
France. \\\email{<EMAIL>}
\\
On leave to: team McTAO, Inria Sophia Antipolis Méditerrannée.
\and
H. Henninger \at
team McTAO, Inria Sophia Antipolis Méditerrannée, \\ 2004 rte des lucioles, B.P. 92, 06902 Sophia Antipolis cedex, France.\\
\email{<EMAIL>}
\and
J. N{\v e}mcov\'a \at
Department of Mathematics,
Institute of Chemical Technology,\\
Technick{\'a} 5,
166 28 Prague 6,
Czech Republic. \\\email{<EMAIL>}
\and
J.-B. Pomet \at
team McTAO, Inria Sophia Antipolis Méditerrannée, \\ 2004 rte des lucioles, B.P. 92, 06902 Sophia Antipolis cedex, France.\\
\email{<EMAIL>}
}
\date{Received: date / Accepted: date}
\maketitle
\begin{abstract}
This article makes a study of the averaged optimal coplanar transfer towards circular orbits.
Our objective is to compare this problem when the cost minimized is transfer time to the same problem when the cost minimized is energy consumption.
While the minimum energy case leads to the analysis of a $2D-$ Riemannian metric using
the standard tools of Riemannian geometry, the minimum time case is
associated with a Finsler metric which is not smooth. Nevertheless a qualitative analysis of the geodesic flow is
given in this article to describe the optimal transfers of the time minimal case.
\end{abstract}
\keywords{
Averaging, Optimal control, Low thrust orbit transfer, Geodesic convexity, Riemann-Finsler Geometry
}
\section{Introduction}
We consider the controlled Kepler equation describing orbital transfers with low thrust engines,
that we normalize as
\begin{equation}
\label{eq:5}
\ddot{q} =-\frac{q}{\| q\|^{3}}+u;
\end{equation}
the control is constrained by $\| u\|\leq\varepsilon$, where
$\varepsilon$ is a small parameter. The phase space, or state space, is the one with coordinates $(q,\dot{q})$.
Let $K=\frac12\,\|\dot{q}\| ^{2}-1/\| q\|$ be the mechanical energy of the uncontrolled system and $X$ be the elliptic
domain:
\[
X=\{K<0,q\wedge\dot{q}\neq0\}\,.\]
For the free motion ($u = 0$), the solutions that lie in $X$ are ellipses ---or more precisely closed curves that project
on the $q$ component as ellipses--- and they form a foliation of $X$.
In this domain, we may chose coordinates $(x,l)$ where $x$ is made of independent first integrals
of the uncontrolled motion (so that $x$ describes the geometry of the ellipses) and the ``longitude'' $l$ defines the
position of the spacecraft on this ellipse; $(q,\dot q)$ can be expressed in terms of $(x,l)$ and vice versa.
Restricting to the coplanar case, where $q$ and $\dot q$ have dimension 2 and $x$ has dimension 3, the system can be written as
\begin{displaymath}
\dot{x}=\sum_{i=1,2}u_{i}F_{i}(x,l)\,,\ \ \ \
\dot{l} =\Omega(x,l)\,,
\end{displaymath}
where the control $u=(u_{1},u_{2})$ is the coordinates of the original acceleration $u$ in some frame
$F_{1},F_{2}$, e.g., the tangential/normal frame (the vector fields $F_{1},F_{2}$ are another basis of the distribution
spanned by $\partial/\partial \dot q_1, \partial/\partial \dot q_2$ in the original cartesian coordinates).
In these coordinates, the free motion is $\dot x=0,\dot{l} =\Omega(x,l)$; there may be a control term in $\dot l$ too but
we neglect it for clarity.
The energy minimization problem is the one of minimizing a quadratic criterion $\int \|u\|^2\mathrm{d}t$ for fixed
initial and final value of $x$, and free $l$; it was analyzed from the averaging point of view in a series of articles
\cite{Edel64,Edel65}, \cite{Geff-Epe97,Geff97th}, \cite{Bonn-Cai09forum}.
The Pontryagin maximum principle yields (for any type of cost: energy, final time or others) an Hamiltonian on the cotangent bundle of the state space with the property
that a minimizing trajectory must be the projection of an integral curve of the Hamiltonian vector field.
For energy minimization, this Hamiltonian is
\[
H(x,p,l)={\textstyle\frac12}(H_{1}(x,p,l)^{2}+H_{2}(x,p,l)^{2})\]
where $H_{i}(x,p,l)=\langle \,p,F_{i}(x,l)\rangle$ are the Hamiltonian lifts of
the vector fields $F_i$ and $p$ is the vector of costate variables of the same dimension as the state vector.
As the bound $\varepsilon$ tends
to zero, the time needed to reach a given orbit tends to infinity.
During this very long time, the variable $x$ move slowly because the control is small while variables like
$l$ move fast thanks to the term $\Omega$; this yields ill conditioned integration if numeric methods are used.
It may be shown that there is an average Hamiltonian
\[
H(x,p)=\frac{1}{2\pi}\int_{0}^{2\pi}\varpi(x,l)H(x,p,l)\mathrm{d}l\,,\]
with $\varpi$ some weight function to be determined, that eliminates the fast variable $l$ and whose Hamiltonian flow
gives a remarkably good approximation of the movement of $x$ in the original system if $\varepsilon$ is indeed small.
It sometimes leads to explicit formulas, and is anyway much better conditioned numerically because the fast
variable has been eliminated.
We shall recall briefly these facts but are more interested in studying qualitatively this new Hamiltonian.
We refer the reader to \cite[\S52]{Arno89} (although no control is considered there) for details on this approximation and its validity.
It turns out that it is quadratic definite positive with respect to $p$ and hence derives from a Riemannian metric on
$X$; furthermore, the coefficients of this metric can be
explicitly computed. In the coplanar case the geodesic flow is Liouville
integrable and the metric associated to a subproblem related to transfer from an arbitrary orbit (in $X$) to a circular
one is even flat: in suitable coordinates the minimizing
solutions are straight lines
\cite{Bonn-Cai09forum}. Moreover this result is still true if the
thrust is oriented only in the tangential direction \cite{Bonn-Cai-Duj06b}.
The same averaging technique can be applied in the minimum time case. The non averaged Hamiltonian reads
$\sqrt{H_{1}^{2}(x,p,l) + H_{2}(x,p,l)^{2}}$ and again an averaged Hamiltonian may be constructed:
\[
H(x,p)=\frac{1}{2\pi}\int_{0}^{2\pi}\varpi(x,l)\sqrt{H_{1}^{2}(x,p,l) + H_{2}(x,p,l)^{2}}\,\mathrm{d}l\,.\]
Like in the energy case, this Hamiltonian derives from a metric on $X$, i.e. the data of a norm on each tangent space to $X$;
however, unlike in the energy case and as observed in the article \cite{Bomb-Pom13}, these norms are not associated with inner products on
these tangent spaces ---this defines a Finsler metric \cite{Bao-Che-She00}, not necessarily Riemannian--- and are not everywhere smooth.
Technical problems involved in going from Riemannian to non
smooth Finsler geometry make the computations of time minimal transfer towards circular
orbits a complicated problem.
The objective of this article is to make a preliminary qualitative
description of the time minimum transfers and to compare them with the energy minimum ones:
section~\ref{sec:prelim} recalls the equations and the computation of the
average Hamiltonians; section~\ref{sec:energy} recalls the results from
\cite{Bonn-Cai09forum,Bonn-Cai-Duj06b} on the minimum energy problem; section~\ref{sec:Tmin} provides a new analysis of the minimum time problem, for transfers to
circular orbits, and in particular proves that the elliptic domain is geodesically convex in this case;
section~\ref{sec:compa} explains why that proof fails in the minimum energy problem, which is consistent with the
non-convexity mentioned in \cite{Bonn-Cai09forum}.
\section{Preliminaries}
\label{sec:prelim}
\subsection{Hamiltonian formalism, Pontryagin maximum principle}
The goal of this paper is to study some Hamiltonian systems associated to optimal control problems.
For the sake of self containedness, let us sketch the relation to the optimal control problems.
Consider the smooth control system $\dot{x} = f(x,u,t)$ for $x \in X$, an $n$-dimensional manifold, $t \in \mathbb{R}$ and $u \in B\subset\mathbb{R}^{\ell}$.\\
An \emph{optimal control problem} on $X$ associated with the control system
$\dot{x} = f(x,u,t)$ is, for instance, the problem of finding relative to the given points $x_0,x_T$
the trajectory $x(\cdot)$ and control $u(\cdot)$, and possibly the final time $T$ if it is not specified, such that
\begin{equation}\label{CntrlPrbm}
\begin{tabular}{ r l}
\(\dot{x}\)&\( = \;\,\, f(x,u,t),\qquad x \in X,\; (u_1,u_2,...,u_{\ell}) \in B\subset\mathbb{R}^{\ell}\)\\
\(x(0)\)&\( =\;\; x_0,\quad x(T) = x_T\)\\
\(\mathcal{J}\)&\( =\;\; \int_{0}^{T}\mathcal{L}(x(t),u(t))dt\rightarrow \mbox{Min}.\)
\end{tabular}
\end{equation}
We call ``minimum time'' the problem where $\mathcal{L}(x,u)=1$ and $T$ is free, and ``minimum energy'' the one
where $T$ is fixed and $\mathcal{L}(x,u)=\|u\|^2$.
The Hamiltonian of the optimal control problem (\ref{CntrlPrbm}) is the function
\[\mathcal{H}(x,p,u,p_0, t)= p_0\mathcal{L}(x,u) + \langle\, p,f(x,u,t)\rangle\,\]
where $p$ is a vector of costate variables (the adjoint vector) of the same dimension as the state variables $x(t)$, and $p_0$
is either $0$ or $-1$.
The Pontryagin maximum principle \cite{Pont-Bol-Gam-M62} (see also \cite[Chap. 6]{Bonn-Fau-Tre06} for applications to
the problems we consider here) is a powerful necessary condition for optimality, that states
the following: if $(x(\cdot),u(\cdot))$ is an optimal trajectory-control pair of the above optimal control problem on a
time interval $[0,T]$, then it can be lifted to a parameterized curve $t\mapsto(x(t),p(t))$ on the cotangent bundle
$T^\star X$ ($p$ is the adjoint vector, or the vector of costate variables) that satisfies, for almost all time and
either for $p_0=0$ or for $p_0=-1$,
\begin{align}\nonumber
&\dot x(t)=\frac{\partial \mathcal{H}}{\partial p}(x(t),p(t),u(t),p_0, t)=f(x(t),u(t),t)
\\\label{eq:48}
&\dot{p}(t)=-\frac{\partial \mathcal{H}}{\partial p}(x(t),p(t),u(t),p_0, t)
\end{align}
and, for almost all $t$, $\mathcal{H}(x(t),p(t),u(t),p_0, t)$ is the maximum of $\mathcal{H}(x(t), p(t), u,$ $p_0, t)$
with respect to $u\in B$.
The solutions where $p_0=0$ are called abnormal. Let us assume $p_0=-1$.
In the problems we consider here, we are in the nice situation where for all $(x,p,t)$, or almost all $(x,p,t)$, there is
a unique $u^\star(x,p,t)$ such that
$$
H(x,p,t)=\mathcal{H}(x,p,u^\star(x,p,t),-1, t)=\max_{u\in B}\mathcal{H}(x,p,u,-1, t)
$$
(the second equality is a property of $u^\star(x,p,t)$; the first equality is the definition of $H$ from $\mathcal{H}$
and $u^\star$).
In that case, one may sum up the above in the following way: if $(x(\cdot),u(\cdot))$ is an optimal trajectory, then
$x(.)$ may be lifted to a solution of the Hamiltonian vector field associated to $H$ on $T^\star X$:
\begin{equation}
\label{eq:49}
\dot{x} = \frac {\partial H}{\partial p}(x,p,t), \quad \dot{p} =- \frac {\partial H}{\partial x}(x,p,t).
\end{equation}
The situation is even nicer if $u^\star$ is a smooth function of $x,p,t$; if not, one must be careful about existence
and uniqueness of solutions of solutions to this differential equation.
We kept the above time-varying system because we will encounter time-periodic Hamiltonians that we average with respect to time,
or with respect to a variable that we may view as a new time.
\subsection{Coordinates}
First of all, we recall the equations describing the planar controlled Kepler
problem in the elliptic case (mechanical energy $K$ is negative).
If we chose as coordinates $(n,e_x,e_y,l)$ where $n$ is the mean movement ($n=\sqrt{1/a^{3}}=(-2K)^{3/2}$; $a$ is the semi-major axis), $(e_x,e_y)$
are the coordinates of the eccentricity vector in a fixed frame and $l$ is the ``longitude'', or the polar angle with
respect to a fixed direction, then the elliptic domain is given by $\{n>0,\,{e_x}^2+{e_y}^2<1\}$. The control system is
described by the Gauss equations, where $u_t,u_n$ are the coordinates of the control in the tangential-normal frame:
\begin{subequations}
\label{eq:equinox}
\begin{align}
\label{eq:6}
\displaybreak[0]
&\dot{n}=-3n^{2/3}\frac{\sqrt{1+2\,(e_x\cos l+e_y\sin l)
+{e_x}^{2}+{e_y}^{2}}}{\sqrt{1-{e_x}^{2}-{e_y}^{2}}}\,u_{t} \\
\nonumber &\dot{e}_x= n^{-1/3}\frac{\sqrt{1-{e_x}^{2}-{e_y}^{2}}}{\sqrt{1+2\,(e_x\cos l+e_y\sin l)
+{e_x}^{2}+{e_y}^{2}}}
\\
\displaybreak[0]
&\label{eq:7}\hspace{7em}\times\!\!\left[2\,(\cos l+e_x)\,u_t -\frac{\sin l+2\,e_y+2 e_x
e_y\cos l-({e_x}^{2}-{e_y}^{2})\sin l}{\sqrt{1-{e_x}^{2}-{e_y}^{2}}}\,u_n\right]
\\
\nonumber &\dot{e}_y= n^{-1/3}\frac{\sqrt{1-{e_x}^{2}-{e_y}^{2}}}{\sqrt{1+2\,(e_x\cos l+e_y\sin l)
+{e_x}^{2}+{e_y}^{2}}}
\\
\displaybreak[0]
&\label{eq:8}\hspace{7em}\times\!\!\left[2\,(\sin l+e_y)\,u_t
-\frac{\cos l+2\,e_x+({e_x}^{2}-{e_y}^{2}) \cos l+2 e_x e_y\sin l}{\sqrt{1-{e_x}^{2}-{e_y}^{2}}}\,u_n\right] \\
\label{eq:9}
&\dot{l}=n\,\frac{(1+e_x\cos l+e_y\sin l)^{2}}{(1-e^{2})^{3/2}}.
\end{align}
\end{subequations}
Instead of $e_x,e_y$, it will be more convenient to use the eccentricity $e$ and the argument
of the pericenter $\omega$ (not defined if $e=0$), defined by
\begin{equation}
\label{eq:10}
e_{x}=e\cos \omega,\ e_{y}=e\sin \omega\;.
\end{equation}
The equations become:
\begin{subequations}
\label{eq:42}
\begin{align}
\label{eq:1}
&\dot{n}=-\frac{3n^{2/3}}{\sqrt{1-e^{2}}}\left[\sqrt{1+2e\cos v+e^{2}}\,u_{t}\right]
\\
\label{eq:2}
&\dot{e}=\frac{\sqrt{1-e^{2}}}{\sqrt[3]{n}}\frac{1}{\sqrt{1+2e\cos v+e^{2}}}\left[2(e+\cos v)\,u_{t}-\sin
v\frac{1-e^{2}}{1+e\cos v}\,u_{n}\right]
\\
\label{eq:3}
&\dot{\omega}=\frac{\sqrt{1-e^{2}}}{e\sqrt[3]{n}}\frac{1}{\sqrt{1+2e\cos v+e^{2}}}\left[2\sin v \,u_{t}
+\frac{2e+\cos v+e^{2}\cos v}{1+e\cos v}\,u_{n}\right]
\\
\label{eq:4}
&\dot{l}=n\,\frac{(1+e\cos v)^{2}}{(1-e^{2})^{3/2}}.
\end{align}
\end{subequations}
The angle $v$ is the true anomaly
\begin{equation}
\label{eq:24}
v=l-\omega.
\end{equation}
In these coordinates, the elliptic domain is
\begin{equation}
\label{eq:47}
X=\{x=(n,e,\omega)\,,\ n>0,\,0\leq e<1,\,\omega\in S^{1}\}\;.
\end{equation}
\begin{remark}[Transfer towards a circular orbit]
\label{rmk:omega}
In the transfer ``towards a circular orbit'' (or merely if we do not take into account the direction of the semi-major
axis during the transfer), we may use these coordinates although they are singular at $e=0$, because the variable
$\omega$ may simply be ignored; this is possible because it is a cyclic variable, i.e. it does not influence the
evolution of the other variables $(n,e,v)$.
In the variables $(n,e)$, the elliptic domain is:
\begin{equation}
\label{eq:33}
\mathcal{X}=\{(n,e),\; 0<n<+\infty,\;-1<e<1\}\;.
\end{equation}
The fact that negative values of $e$ are allowed comes from identifying $(-e,\omega)$ with $(e,\omega+\pi)$, or, equivalently, considering that
$(e_x,e_y)$ (see \eqref{eq:10}) lies on a line of fixed arbitrary direction instead of a half-line.
This line may for instance be $\{e_y=0\}$, and $\mathcal{X}$ is then identified with
$\{(n,e_x,e_y),\,n>0,-1< e <1, e=e_{x},e_{y}=0\}$.
\end{remark}
Equations \eqref{eq:1}-\eqref{eq:4} read:
\begin{equation}
\label{eq:11}
\dot{x} =\sum_{1\leq i\leq 2} u_i\,F_i(x,l),\;\;\dot{l}=\Omega(x,l)
\end{equation}
where $u_1, u_2$ stand for $u_n,u_t$, $x = (n,e,\omega)$, the vectors $F_1,F_2$ are readily obtained from \eqref{eq:1}-\eqref{eq:3}, and
\begin{equation}
\label{eq:16}
\Omega(x,l)=n\,\frac{(1+e\cos(l-\omega))^{2}}{(1-e^{2})^{3/2}}\ .
\end{equation}
One way to introduce averaging is to use the so-called ``mean eccentric anomaly''. The eccentric anomaly is $E$,
related to $e$ and $v$ by
\begin{equation}
\label{eq:20}
\tan\frac v 2\ =\ \sqrt{\frac{1+e}{1-e}}\tan\frac E 2\,
\end{equation}
and the mean eccentric anomaly is $E-e\sin E$; the Kepler equation (third Kepler law) implies that, when the control is zero,
\[
E-e\sin E=n\,t\,,\]
$t=0$ being the time at the pericenter.
Introducing (see for instance \cite[sec. 3.6.3]{Bonn-Fau-Tre06})
$$x_{0}=(E-e\sin E)/n\,,$$
one has $\dot{x}_0 =1$ if $u=0$, i.e. the variable $x_0$ behaves like time modulo an additive constant;
this is an implementation of the flow-box theorem. In the coordinates $(x,x_0)$, the system becomes
\[\dot{x}=\sum_{i=1,2}u_{i}\widehat{F}_{i}(x,x_{0}),\;\;\;\,\dot{x}_0 =1+\sum_{i=1,2}u_iG_{i}(x,x_{0}).\]
Due to the implicit relation between $E$ and $x_0$, the practical derivation of such equations is complicated, but
they will be useful in formally identifying averaging with
respect to $l\in[0,2\pi]$ and averaging with respect to $t\in[0,2\pi/n]$.
We define the Hamiltonian lifts ($i=1,2$):
\begin{equation}
\label{eq:12}
H_{i}(x,p,l)=\langle p,F_{i}(x,p,l)\rangle\,,\ \ \ \widehat{H}_{i}(x,p,x_{0})=\langle p,\widehat{F}_{i}(x,p,x_{0})\rangle\,.
\end{equation}
\subsection{Averaging}
Using the previous equations and rescaling the control with $u=\varepsilon v$
to introduce the small parameter, the trajectories parameterized by
$x_{0}$ are solutions of
\[
\frac{dx}{dx_{0}}=\frac{\varepsilon\sum_{i=1,2}v_{i}\widehat{F}_{i}(x,x_{0})}{1+\varepsilon\sum_{i=1,2}v_{i}G_{i}(x,x_{0})},\]
which is approximated for small $\varepsilon$ by
\[
\frac{dx}{dx_{0}}=\varepsilon\sum_{i=1,2}v_{i}\widehat{F}_{i}(x,x_{0}).\]
For this system, we consider the following minimization problems:
\begin{equation*}
\begin{tabular}{ c l }
\(\bullet\) & \(\displaystyle \mbox{Energy}\;:\; \min_{v}\varepsilon^{2}\int_{0}^{x_{0}}\sum_{i=1,2}v_{i}^{2}\mathrm{d}t\)\\[0.3cm]
\(\bullet\) & \(\displaystyle \mbox{Time}\; :\; \min_{v}\;x_{0}, \| v\|\leq 1. \)\\
\end{tabular}
\end{equation*}
Applying the Pontryagin maximum principle leads to the following respective
Hamiltonians (normal case in the energy minimization problem),
\begin{equation}
\label{eq:44}
\begin{split}
H_{\mathrm{e}}(x,p,x_{0})=\sum_{i=1,2}\widehat{H}_{i}(x,p,x_{0})^{2}
\,,\ \ \ \
H_{\mathrm{t}}(x,p,x_{0})=\sqrt{\sum_{i=1,2}\widehat{H}_{i}(x,p,x_{0})^{2}}\,,
\end{split}
\end{equation}
where the lifts $\widehat{H}_{i}$, defined by \eqref{eq:12}, are periodic
with respect to $x_{0}$ with period $2\pi/n$.
\begin{remark}[Tangential thrust]
\label{rmk:tangential}
If the normal component $u_n$ is forced to be zero, there is a single term in the sums in
\eqref{eq:44}, and these equations become
\quad
$H_{\mathrm{e}}={\widehat{H}_{1}}^{2}$,
\quad
$H_{\mathrm{t}}=\left|\widehat{H}_{1}\right|$. The considerations in the present section are valid both in the full
control case and in the ``tangential thrust'' case.
\end{remark}
The respective averaged Hamiltonians are
\begin{gather}
\label{eq:13}
H_{\mathrm{e}}(x,p)=\frac{n}{2\pi}\int_{0}^{2\pi/n}H_{\mathrm{e}}(x,p,x_{0})dx_{0} \\
\label{eq:14}
H_{\mathrm{t}}(x,p)=\frac{n}{2\pi}\int_{0}^{2\pi/n}H_{\mathrm{t}}(x,p,x_{0})dx_{0}\,.
\end{gather}
(for ease of notation we use $H_\mathrm{e}, H_\mathrm{t}$ to represent both the Hamiltonians and the averaged Hamiltonians, although the inputs into these functions are different).
These may be re-computed in terms of $H_1,H_2$. Unlike when $\widehat{H}_1,\widehat{H}_2$ are used in the computation, using the Hamiltonian lifts $H_1, H_2$ allows for an explicit expression of
the averaged Hamiltonians $H_\mathrm{e}(x,p), H_\mathrm{t}(x,p)$. Making the change of
variables $x_0=\Xi(e,\omega,l)$ ---with $\Xi$ deduced from $x_{0}=(E-e\sin E)/n$, \eqref{eq:20} and
\eqref{eq:24}--- in the integral, and using the facts that $\partial\Xi/\partial l=1/\Omega(x,l)$ and
$$
\widehat{H}_\mathrm{e}(x,p,\Xi(e,\omega,l))=H_\mathrm{e}(x,p,l)\,,\;\;
\widehat{H}_\mathrm{t}(x,p,\Xi(e,\omega,l))=H_\mathrm{t}(x,p,l)\,,
$$
then, using \eqref{eq:16},
\begin{gather}
\label{eq:25}
H_{\mathrm{e}}(x,p)=\frac{(1-e^{2})^{3/2}}{2\pi}\int_{0}^{2\pi}\left(\sum_{i=1,2}H_{i}(x,p,l)^{2}\right) \frac{\mathrm{d}l}{(1+e\cos(l-\omega))^{2}} \\
\label{eq:15}
H_{\mathrm{t}}(x,p)=\frac{(1-e^{2})^{3/2}}{2\pi}
\int_0^{2\pi}\sqrt{\sum_i H_i(x,p,l)^2}\frac{\mathrm{d}l}{(1+e\cos(l-\omega))^{2}}\,.
\end{gather}
\begin{remark}
\label{rem-smallthrust} In the original system, the control is ``small'' (parameter $\varepsilon$).
The average system that we study in the next sections can be seen as a limit as $\varepsilon\to0$.
The smaller $\varepsilon$ is, the better the average system approximates the real system, but neither the results of this
paper not any analysis or simulation in the next sections depend on the size of $\varepsilon$, that is on the magnitude
of the thrust.
\end{remark}
\paragraph{Singularities.}
Let us explain how the non smoothness is a result of the averaging of singularities of a control
system.
Consider the time minimal control problem for a generic smooth system of the
form
\[\dot{x} =F_{0}(x)+\sum_{i=1,m}u_{i}F_{i}(x),\;\;\| u\|\leqq1.\]
Moreover assume for simplicity that the control distribution $D=\Span\{F_{1}, \ldots$, $F_{m}\}$
is involutive. From the maximum principle in this case, the
extremal control is defined by $u_{i}=\frac{H_{i}(x,p)}{\sqrt{\sum_i H_{i}^{2}(x,p)}}$
where $H_{i}(x,p)$ are the Hamiltonian lifts of $F_{i}(x)$. More
complicated extremals are related to the switching surface $\Sigma:$
$H_{i}=0.$ Observe that in the single-input case the control is given
by $u_{1}=\sign H_{1}(x,p)$ and meeting the surface $\Sigma$ transversally
corresponds to a regular switching. This can be generalized to the
multi-input case. More complicated singularities can occur in the
non transversal case, for instance in relation with singular trajectories
of the system (contained by definition in the surface $\Sigma$) \cite{Bonn-Sug12book}.
\section{The analysis of the averaged systems for minimum energy}
\label{sec:energy}
First of all we recall the results from the energy case \cite{Bonn-Cai09forum} . The energy minimization problem is expressed as
\[
\int_{0}^{l_{f}}\left(u_{1}^{2}(t)+u_{2}^{2}(t)\right)dt\rightarrow \mbox{Min},\]
where we fix the final cumulated longitude $l_{f}$ (this is slightly different from fixing the transfer time).
\subsection{The coplanar energy case}
In this case the averaged system can be computed explicitly by quadrature,
and we have the following proposition.
\begin{proposition}
In the coordinates $(n,e,\omega)$ the averaged Hamiltonian (up
to a positive scalar) is given by
\begin{equation}
\label{eq:30}
H_\mathrm{e} =\frac{1}{n^{5/3}}[18n^{2}p_{n}^{2}+5(1-e^{2})p_{e}^{2}+\frac{5-4e^{2}}{e^{2}}p_{\omega}^{2}]
\end{equation}
where the singularity $e=0$ corresponds to circular orbits. In particular
$(n,e,\omega)$ are orthogonal coordinates for the Riemannian metric
associated to $H$, namely
\[
g=\frac{1}{9n^{1/3}}\mathrm{d}n^{2} + \frac{2n^{5/3}}{5(1-e^{2})}\mathrm{d}e^{2} + \frac{2n^{5/3}}{5-4e^{2}}\mathrm{d}\omega^{2}.\]
\end{proposition}
Further normalizations are necessary to capture the main properties
of the averaged orbital transfer.
\begin{proposition}
In the elliptic domain we set
\[
r=\frac{2}{5}n^{5/6},\varphi=\arcsin e\]
and the metric is isometric to
\[
g=\mathrm{d}r^{2}+\frac{r^{2}}{c^{2}}(\mathrm{d}\varphi^{2}+G(\varphi)\mathrm{d}\omega^{2})\]
where $c=\sqrt{2/5}$ and $G(\varphi)=\frac{5\sin ^{2}\varphi}{1+4\cos ^{2}\varphi}.$
\end{proposition}
\subsection{Transfer towards circular orbits}
As noticed in Remark~\ref{rmk:omega}, for such transfers we may ignore the cyclic variable $\omega$ and allow negative $e$. In this case, the
elliptic domain is the $\mathcal{X}$ given by \eqref{eq:33}. The metric above then reduces to
\[
g=\mathrm{d}r^{2}+r^{2}d\psi^{2}\,,\ \ \text{with}\ \ \psi=\varphi/c\]
defined on the domain $\{(r,\psi),\,0<r<+\infty,-\frac\pi{2c}<\psi<\frac\pi{2c}\}$; it is
a polar metric isometric to the flat metric $dx^{2}+dz^{2}$
if we set $x=r\sin \psi$ and $z=r\cos \psi$. Flatness in the original coordinates can be checked
by computing the Gauss curvature. We deduce
the following theorem:
\begin{theorem}
\label{th:droites}
The geodesics of the averaged coplanar transfer towards circular orbits
are straight lines in the domain $\mathcal{X}$ (see \eqref{eq:33}) in suitable coordinates,
namely
\[
x=\frac{2^{3/2}}{5}n^{5/6}\sin (\frac{1}{c}\arcsin e),\;z=\frac{2^{3/2}}{5}n^{5/6}\cos (\frac{1}{c}\arcsin e)\]
with $c=\sqrt{2/5}.$ Since $c<1,$ the domain is not (geodesically)
convex and the metric is not complete.
\end{theorem}
\begin{remark}[Tangential thrust]
The properties of theorem \ref{th:droites} are still true when the thrust is only in the tangential direction except that the metric has a
singularity at $e=1$. The formula is
\[
g=\frac{1}{9n^{1/3}}\;\mathrm{d}n^{2}\;+\;\; \frac{(1+\sqrt{1-e^{2}})n^{5/3}}{4(1-e^{2})}\left[\frac{1}{\sqrt{1-e^{2}}} \;\, \mathrm{d}e^{2}\; + \;\; e^{2}\; \mathrm{d}\omega^{2}\right].\]
We may slightly twist the previous coordinates using $e=\sin\varphi\sqrt{1+\cos^{2}\varphi}$ to get the normal form
$\mathrm{d}r^{2}\;+\;\,(r^{2}/c_{t})(\;\mathrm{d}\varphi^{2}\; +\; G_{t}(\varphi)\;\mathrm{d}\omega^2),c_{t}=c^{2}=2/5,G_t(\varphi)=\sin^{2}\varphi(\frac{1-(1/2)\sin ^{2}\varphi}{1-\sin
^{2}\varphi})^{2}.$
\end{remark}
\section{The analysis of the averaged systems for minimum time}
\label{sec:Tmin}
\subsection{The Hamiltonian}
\label{sec:Tmin:Ham}
We compute $H_\mathrm{t}$ according to \eqref{eq:15}.
The functions $H_i$, $i=1,2$ depend on $n,e,\omega$, $p_n,p_e,p_\omega,l$.
Since we only consider transfer towards a circular orbit, we set $p_\omega=0$ and define $h_1, h_2$ by
\begin{equation}
\label{eq:18}
h_i(n,e,p_n,p_e,v)=H_i(n,e,\omega,p_n,p_e,0,\omega+v)\ .
\end{equation}
The right-hand side does not depend on the cyclic variable $\omega$, see Remark~\ref{rmk:omega}.
From here on we will use the subscripts $1$ and $2$ to denote respectively the tangential and normal directions, rather than $t$ and $n$, for ease of notation.
From \eqref{eq:42}, we get
\begin{subequations}
\label{eq:43}
\begin{align}
\label{eq:45}
&h_1=n^{-1/3}\left(-3n\,p_n\frac{\sqrt{1+2e\cos v+e^{2}}}{\sqrt{1-e^2}}+ 2 p_e
\frac{(e+\cos v)\sqrt{1-e^2}}{\sqrt{1+2e\cos v+e^{2}}} \right)
\\
&h_2=-\,n^{-1/3}\,p_e\, \frac{\sin v \,(1-e^2)^{3/2}}{(1+e\cos v)\sqrt{1+2e\cos v+e^{2}}}
\end{align}
\end{subequations}
Note that $\omega$ does not vary
in the integral; the integrand has period $2\pi$ with respect to either $l$ or $v$. This allows us to
make the change of variable $l=\omega+v$ in the integral in \eqref{eq:15}. In the full control case (both tangential and
normal control), the sum in \eqref{eq:15} contains two terms, and we obtain
\begin{equation}
\label{eq:17}
H_\mathrm{t}(n,e,p_n,p_e)=\frac{(1-e^{2})^{3/2}}{2\pi}
\!\int_0^{2\pi}
\!\!\!\!
\sqrt{\sum_{i=1}^2 h_i(n,e,p_n,p_e,v)^2}\,\frac{\mathrm{d}v}{(1+e\cos v)^{2}}\,,
\end{equation}
In the tangential thrust case it only
contains $h_1$ ---see remark~\ref{rmk:tangential}--- and we get (the superscript 1 in $H_\mathrm{t}^1$ denotes single input):
\begin{equation}
\label{eq:17t}
H_\mathrm{t}^1(n,e,p_n,p_e)=\frac{(1-e^{2})^{3/2}}{2\pi}
\int_0^{2\pi}\bigl|h_1(n,e,p_n,p_e,v)\bigr|\,\frac{\mathrm{d}v}{(1+e\cos v)^{2}}\ .
\end{equation}
In order to highlight some properties of these Hamiltonians, we perform a canonical change of coordinates
$(n,e,p_n,p_e)\mapsto(\lambda,\varphi,p_\lambda,p_\varphi)$:
$$
n=e^{3\lambda}\,,\ e=\sin\varphi\,,\ p_n=\frac{p_\lambda}{3n}\,,\ p_e=\frac{p_\varphi}{\cos\varphi}
$$
followed by taking $(\rho,\psi)$ as polar coordinated for the adjoint vector $(p_\lambda,p_\varphi)$; we shall never use again
the notations $\lambda$, $p_\lambda$, $p_\varphi $ and directly write the change as
\begin{equation}
\label{eq:19}
3\,n\,p_n=\rho\cos\psi\,,\ \ \ \ \sqrt{1-e^2}\,p_e=\rho\sin\psi\,,\ \ \ \ e=\sin\varphi
\,,\ {\textstyle -\frac\pi2<\varphi<\frac\pi2}\,.
\end{equation}
Equations \eqref{eq:17} and \eqref{eq:17t} then yield
\begin{gather}
\label{eq:21}
H_\mathrm{t}(n,\sin\varphi,\frac{\rho\cos\psi}{3n},\frac{\rho\sin\psi}{\cos\varphi})
=
\rho\,n^{-1/3}\,L(\psi,\varphi)
\\
\label{eq:21t}
H_\mathrm{t}^1(n,\sin\varphi,\frac{\rho\cos\psi}{3n},\frac{\rho\sin\psi}{\cos\varphi})
=
\rho\,n^{-1/3}\,M(\psi,\varphi)
\end{gather}
with $L$ and $M$ some functions $\mathcal{C}\to\RR$, where $\mathcal{C}$ is the cylinder
\begin{equation}
\label{eq:0069}
\mathcal{C}=\{(\psi,\varphi),\;\psi\in(\RR/2\pi\ZZ),\;\varphi\in\RR,\,-\frac\pi2<\varphi<\frac\pi2\}
=\RR/2\pi\ZZ\times (-\frac\pi2,\frac\pi2)\,.
\end{equation}
The expressions of $L$ and $M$ are, taking the eccentric anomaly $E$ as the variable of integration instead of $v$ (see \eqref{eq:20}; in particular,
${\mathrm{d}v}/(1+e\cos v)=\mathrm{d}E/\sqrt{1-e^2}$) and restricting the interval of integration from $[0,2\pi]$ to
$[0,\pi]$ because the integrand depends on $\cos E$ only:
\begin{align}
\label{eq:26}
&L(\psi,\varphi)
=
\frac{1}{\pi}\!\int_0^{\pi}\!\!
\sqrt{\widetilde{I}(\psi,\varphi,E)} \,\mathrm{d}E\,,
\\
\label{eq:22}
&\widetilde{I}(\psi,\varphi,E)= \alpha^{1,1}(\varphi,\cos E)
\cos^2\!\psi\,
+2\,\alpha^{1,2}(\varphi,\cos E)\,\cos\psi\sin\psi+\alpha^{2,2}(\varphi,\cos E)\sin^2\!\psi\,,
\\[1ex]
\nonumber &\ \ \alpha^{1,1}=1-\sin^2\!\varphi\cos^2\!E\,,
\hspace{3em}
\alpha^{1,2}=-2 \cos\varphi\, (1\!-\!\sin\varphi\cos E)\cos E\,,
\\
\label{eq:0046}
&\ \ \alpha^{2,2}=(1\!-\!\sin\varphi\cos E) \left(1-3\sin\varphi \cos E+3\cos^2\!E-\sin\varphi\cos^3 \!E\right)
\end{align}
and
\begin{align}
\label{eq:26t}
&M(\psi,\varphi)
=
\frac{1}{\pi}\!\int_0^{\pi}\!\!\left|\widetilde{J}(\psi,\varphi,E)\right|\,\mathrm{d}E\,,
\\
\label{eq:22t}
&\widetilde{J}(\psi,\varphi,E)= \sqrt{\frac{1-\sin\varphi\cos E}{1+\sin\varphi\cos E}}
\,
\Bigl( ( 2\cos\varphi\sin\psi - \sin\varphi\cos\psi )\cos E -\cos\psi\Bigr)\,.
\end{align}
In the sequel we take advantage of the double homogeneity with respect to $\rho$ and $n$ displayed in \eqref{eq:21} and \eqref{eq:21t}.
\subsection{Singularities of the Hamiltonian in the single-input and two-input cases}
According to \eqref{eq:21} and \eqref{eq:21t}, the Hamiltonians $H_\mathrm{t}$ and $H_\mathrm{t}^1$, have the same
degree of smoothness as, respectively the maps $L$ and $M$.
\begin{proposition}
\label{lem:unic0}
The maps $L:\mathcal{C}\to\RR$
and
$M:\mathcal{C}\to\RR$ are real analytic away from
\begin{equation}
\label{eq:38}
\mathcal{S}=\left\{\!(\psi,\varphi),\,
\tan\psi=\frac{1+\sin\varphi}{2\cos\varphi}\right\}
\cup
\left\{\! (\psi,\varphi)\,,\ \tan\psi=\frac{-1+\sin\varphi}{2\cos\varphi}\right\}.
\end{equation}
They are both continuously differentiable on $\mathcal{C}$, but their differentials are not locally Lipschitz-continuous
on the set $\mathcal{S}$; we have the following moduli of continuity of the differentials: in a neighborhood of a point
$\xi=(\psi,\varphi)\in\mathcal{S}$, in some corrdinates and for a ``small'' $\delta$,
\begin{align}
\label{eq:46}
\|\mathrm{d}L(\xi+\delta)-\mathrm{d}L(\xi)\|
&\leq k\,\|\delta\|\,\ln(1/\|\delta\|)
\\
\label{eq:46t}
\|\mathrm{d}M(\xi+\delta)-\mathrm{d}M(\xi)\|
&\leq k\,\|\delta\|^{1/2}
\end{align}
\end{proposition}
\begin{proof}
The set $\mathcal{S}$ is the set of points $(\psi,\varphi)$ such that $\widetilde{I}(\psi,\varphi,E)$ vanishes
for some value of $E$; hence the integrand in \eqref{eq:26} is real analytic on $\mathcal{C}\setminus \mathcal{S}$ and
so is $L$. The degree of regularity \eqref{eq:46} for $L$ at points in $\mathcal{S}$ is given in \cite{Bomb-Pom13}.
Let us now treat $M$.
It turns out that $\mathcal{S}$ is \emph{also} the border between the region
\begin{equation}
\label{eq:R1}
\mathcal{R}_1 = \{(\psi,\varphi) \in \mathcal{C}\;:\;\frac{-1 + \sin\varphi}{2\cos\varphi}< \tan\psi < \frac{ 1 +
\sin\varphi}{2\cos\varphi} \}
\end{equation}
where the sign of $\widetilde{J}(\psi,\varphi,E)$ does not depend on $E$ and the region
\begin{equation}
\label{eq:R2}
\mathcal{R}_2 = \{(\psi,\varphi) \in \mathcal{C}\;:\;\tan\psi <\frac{-1 + \sin\varphi}{2\cos\varphi} \mbox{ or } \frac{ 1 +
\sin\varphi}{2\cos\varphi}< \tan\psi \}
\end{equation}
where $\widetilde{J}(\psi,\varphi,E)$ vanishes for two distinct values of the angle $E$ where it changes sign; these two
values are given by $\cos E=R(\psi,\varphi)$ with
\begin{equation}
\label{eq:35}
R(\psi,\varphi)=
\frac {\cos\psi}{P(\psi,\varphi) }\,,\ \ \
P(\psi,\varphi)=2\cos\varphi\sin\psi - \sin\varphi\cos\psi
\end{equation}
(note that \eqref{eq:R1},\eqref{eq:R2} amount to $\mathcal{R}_1 = \{ |R(\psi,\varphi)|>1\}$, $\mathcal{R}_2 = \{
|R(\psi,\varphi)|<1\}$ and $\mathcal{S}$ is the locus where $R=\pm1$). Hence \eqref{eq:26t} yields
\begin{equation}
\label{eq:36}
M(\psi,\varphi)=
\begin{cases}
\frac{-\sign\cos\psi}{\pi}\!\int_0^{\pi} \widetilde{J}(\psi,\varphi,E)\,\mathrm{d}E&\mbox{on }\mathcal{R}_1\,,
\\[1ex]
\frac{\sign P(\psi,\varphi)}{\pi}\left(
\int_0^{\arccos R(\psi,\varphi)} \widetilde{J}(\psi,\varphi,E)\,\mathrm{d}
-
\int_{\arccos R(\psi,\varphi)}^\pi \widetilde{J}(\psi,\varphi,E)\,\mathrm{d}E
\right)&\mbox{on }\mathcal{R}_2\,.
\end{cases}
\end{equation}
It is therefore clear that $M$ is real analytic on $\mathcal{C}\setminus \mathcal{S}=\mathcal{R}_1\cup\mathcal{R}_2$.
The singularity of $M$ on $\mathcal{S}$ is not of the type treated in \cite{Bomb-Pom13}, but it is clear above that the restriction of $M$
to $\mathcal{R}_1$ has a real analytic continuation through $\mathcal{S}$ while its restriction to $\mathcal{R}_2$, on
the contrary, behaves like a square root in a neighborhood of $\mathcal{S}$, whence \eqref{eq:46t}.
\qed
\end{proof}
The properties of the differential of the Hamiltonian are important because it is the right-hand side of the Hamiltonian equation.
Studying these singularities more precisely is an interesting program that is not yet carried out.
\subsection{The Hamiltonian flow}
\label{flow-full}
Let us now study the solutions of the Hamiltonian equation associated with the minimum time problem in the
full control or single control (tangential thrust) cases, namely:
\begin{equation}
\label{eq:23}
\dot n=\frac{\partial H_\mathrm{t}}{\partial p_n}\,,\ \dot e=\frac{\partial H_\mathrm{t}}{\partial p_e}\,,\
{\dot p}_n=-\frac{\partial H_\mathrm{t}}{\partial n}\,,\ {\dot p}_e=-\frac{\partial H_\mathrm{t}}{\partial e}
\end{equation}
and
\begin{equation}\label{A4}
\dot{n} = \frac{\partial H_\mathrm{t}^1}{\partial p_n},\quad \dot{e} = \frac{\partial H_\mathrm{t}^1}{\partial p_e}, \quad \dot{p}_n = \frac{\partial H_\mathrm{t}^1}{\partial n},
\quad \dot{p}_e = \frac{\partial H_\mathrm{t}^1}{\partial e}.
\end{equation}
with $H_\mathrm{t}$ given by \eqref{eq:17} and $H_\mathrm{t}^1$ by \eqref{eq:17t}.
Specifically, we establish geodesic convexity of the elliptic domain
$\mathcal{X}$ (see \eqref{eq:33}), i.e. any two points in $\mathcal{X}$ can be joined by a extremal curve. This is
contained in the following result:
\begin{theorem}[geodesic convexity]
\label{th:convex} \label{Th1}
For any $(n^0,e^0)$ and $(n^1,e^1)$ in $\mathcal{X}$, there exist a time $T\geq0$ and
a solution $[0,T]\to\mathcal{X}$, $t\mapsto(n(t),e(t),p_n(t),p_e(t))$ of \eqref{eq:23} (resp. of \eqref{A4}) such that $(n(0),e(0))=(n^0,e^0)$
and $(n(T),e(T))=(n^1,e^1)$.
\end{theorem}
\medskip
In order to ease the proof, let us write \eqref{eq:23} and \eqref{A4} in other coordinates.
\begin{proposition}
\label{prop-varpsi}
In the coordinates $(n,\varphi,\psi,\rho)$ defined by \eqref{eq:19}, and after a time re-parametrization
\begin{equation}
\label{Time}
\mathrm{d}t=n^{1/3}\, \mathrm{d}\tau\,,
\end{equation}
equation \eqref{eq:23} (resp. equation \eqref{A4}) becomes
\begin{equation}\label{A5}
\frac{\mathrm{d}\psi}{\mathrm{d}\tau} = a(\psi,\varphi) \,,\ \ \
\frac{\mathrm{d}\varphi}{\mathrm{d}\tau} = b(\psi,\varphi) \,,\ \ \
\frac{\mathrm{d}n}{\mathrm{d}\tau} =-\,3n\,c(\psi,\varphi) \,,
\end{equation}
where $a,b,c$ are given by\footnote{lower indices stand for partial derivatives}:
\begin{align}
\nonumber
&a(\psi,\varphi)=-L(\psi,\varphi)\sin\psi-L_\varphi (\psi,\varphi)\cos\psi\,,
\\
\label{eq:0026}
&b(\psi,\varphi)=\ \ \;L(\psi,\varphi)\sin\psi+L_\psi (\psi,\varphi)\cos\psi\,,
\\
\nonumber
&c(\psi,\varphi)=\ \ \;L(\psi,\varphi)\cos\psi-L_\psi (\psi,\varphi)\sin\psi
\end{align}
(resp. given by:
\begin{align}
\nonumber
&a(\psi,\varphi)=-M(\psi,\varphi)\sin\psi-M_\varphi (\psi,\varphi)\cos\psi\,,
\\
\label{A2}
&b(\psi,\varphi)=\ \ \;M(\psi,\varphi)\sin\psi+M_\psi (\psi,\varphi)\cos\psi\,,
\\
\nonumber
&c(\psi,\varphi)=\ \ \;M(\psi,\varphi)\cos\psi-M_\psi (\psi,\varphi)\sin\psi \ \ \ \mbox{)}
\end{align}
and the evolution of $\rho$ is given by:
\begin{align}
\label{eq:34}
\rho(\tau)=&\,\rho(0)\left(\frac{n(0)}{n(\tau)}\right)^{-1/3}\frac{L(\psi(0),\varphi(0))}{L(\psi(\tau),\varphi(\tau))}
\\\label{39}
\mbox{(resp.}\ \ \ \rho(\tau) =&\, \rho(0)\left(\frac{n(0)}{n(\tau)}\right)^{-1/3}\frac{M(\psi(0),\varphi(0))}{M(\psi(\tau),\varphi(\tau))}
\ \ \ \mbox{).}
\end{align}
The ``time'' $\tau$ is related to the real time $t$ by
\begin{align}
\label{eq:0054}
t=&\,
\frac{n(\tau)^{1/3}\cos\psi(\tau)}{L(\varphi(\tau),\psi(\tau))} -
\frac{n(0)^{1/3}\cos\psi(0)}{L(\varphi(0),\psi(0))}\,.
\\\label{40}
\mbox{(resp.}\ \ \ t=&\,
\frac{n(\tau)^{1/3}\cos\psi(\tau)}{M(\psi(\tau),\varphi(\tau))} - \frac{n(0)^{1/3}\cos\psi(0)}{M(\psi(0),\varphi(0))}
\ \ \ \mbox{).}
\end{align}
\end{proposition}
\begin{proof}
From \eqref{eq:19} and \eqref{eq:23} (resp. \eqref{eq:19} and \eqref{A4}), one gets
\begin{align}
\nonumber
&\dot\psi=\frac1\rho\left(3 n \sin\psi \frac{\partial H}{\partial n} - \cos\varphi \cos\psi \frac{\partial H}{\partial
e}\right)
-\cos\psi\sin\psi\left(\frac1n \frac{\partial H}{\partial p_n}
+\frac{\sin\varphi}{\cos^2\!\varphi}\frac{\partial H}{\partial p_e}\right),
\\
\label{eq:27}
&\dot\varphi=\frac1{\cos\varphi}\frac{\partial H}{\partial p_e}\,,\ \ \ \ \ \dot n=\frac{\partial H}{\partial p_n}
\end{align}
where $H$ stands for $H_\mathrm{t}$ (resp. for $H_\mathrm{t}^1$).
Differentiating \eqref{eq:21} (resp. \eqref{eq:21t}) with respect to $n,\varphi,\rho,\psi$ and solving for
$\frac{\partial H_\mathrm{t}}{\partial n}, \frac{\partial H_\mathrm{t}}{\partial e}, \frac{\partial H_\mathrm{t}}{\partial p_n}, \frac{\partial
H_\mathrm{t}}{\partial p_e} $
(resp. for $\frac{\partial H_\mathrm{t}^1}{\partial n}$, $\frac{\partial H_\mathrm{t}^1}{\partial e}$, $\frac{\partial H_\mathrm{t}^1}{\partial p_n}$, $\frac{\partial
H_\mathrm{t}^1}{\partial p_e} $), we obtain the latter as linear combinations of $L(\psi,\varphi)$,
$L_\varphi(\psi,\varphi)$, $L_\psi(\psi,\varphi)$ (resp. of $M(\psi,\varphi), M_\varphi(\psi,\varphi), M_\psi(\psi,\varphi)$)
with coefficients depending on $n,\varphi,\rho,\psi$; substituting these expressions into \eqref{eq:27} gives
\begin{eqnarray*}
\dot\psi&=& n^{-1/3}\left(-L\sin\psi-L_\varphi\cos\psi\right)\,,
\\
\dot\varphi&=&n^{-1/3}\left(L\sin\psi+L_\psi\cos\psi\right)\,,
\\
\dot n&=& -3\,n^{2/3}\left(L\cos\psi-L_\psi\sin\psi\right)
\\[1ex]
\mbox{(resp.}\ \ \ \ \
\dot\psi&=& n^{-1/3}\left(-M\sin\psi-M_\varphi\cos\psi\right)\,,
\\
\dot\varphi&=&n^{-1/3}\left(M\sin\psi+M_\psi\cos\psi\right)\,,
\\
\dot n&=& -3\,n^{2/3}\left(M\cos\psi-M_\psi\sin\psi\right)\ \ \ \mbox{).}
\end{eqnarray*}
With the new time $\tau$ given by \eqref{Time}, one easily deduces \eqref{A5} and the expressions \eqref{eq:0026}
(resp. \eqref{A2}) of $a,b,c$.
Finally, \eqref{A5} and \eqref{eq:0026} (resp. \eqref{A5} and \eqref{A2}) imply
$
\frac{\mathrm{d}}{\mathrm{d}\tau} \!\left(\! \frac{n^{1/3}\,\cos\psi}{L(\psi,\varphi)} \!\right) =n^{-1/3}
$
(resp. $\frac{d}{d\tau}\left(\frac{n^{1/3}\cos\psi}{M(\psi, \varphi)} \right)= n^{1/3}$), that implies
\eqref{eq:0054} (resp. \eqref{40}) according to \eqref{Time}.\qed
\end{proof}
The first two equations in \eqref{A5} form an autonomous system of equations
in the two variables $(\psi,\varphi)\in\mathcal{C}$ that will be the core of our analysis; the third one may be
integrated and yields $n(\tau)$:
\begin{equation}
\label{eq:29}
n(\tau)=n(0)\exp\Bigl(-3
\int_0^\tau c(\psi(\sigma),\varphi(\sigma))\mathrm{d}\sigma\Bigr)\,.
\end{equation}
The variable $\rho$ (the magnitude of the adjoint vector) plays no role in the
evolution of the other variables, in particular the state $(n,e)$ ($e=\sin\varphi$); this is
a well-known consequence of the Hamiltonian being homogeneous of degree 1 with
respect to the adjoint vector and is anyway obvious from \eqref{A5}.
\bigskip \bigskip
Let us now gather some properties of the maps $a,b,c$, \emph{i.e.} of
the differential equation \eqref{A5}, that are valid both for $a,b,c$ given by \eqref{eq:0026} and for $a,b,c$ given by
\eqref{A2}; they contain all the information to prove Theorem~\ref{Th1}.
\begin{proposition}
\label{lem-check}The maps $a,b,c$ given by \eqref{eq:0026} satisfy the following properties
with $\overline\sigma=0$.
The maps $a,b,c$ given by \eqref{A2} satisfy the same properties with $\overline\sigma=\arctan\frac12$.
\begin{enumerate}
\item\label{ass-sym} \textbf{\boldmath Symmetries.} For all $(\psi,\varphi)$ in $\mathcal{C}$,
\begin{equation}
\label{eq:0036}
\begin{array}{l}
a(\psi+\pi,\varphi)=-a(\psi,\varphi) \,, \\
b(\psi+\pi,\varphi)=-b(\psi,\varphi) \,, \\
c(\psi+\pi,\varphi)=-c(\psi,\varphi) \,,
\end{array}
\hspace{3em}
\begin{array}{l}
a(-\psi,-\varphi)=-a(\psi,\varphi) \,, \\
b(-\psi,-\varphi)=-b(\psi,\varphi) \,, \\
c(-\psi,-\varphi)=c(\psi,\varphi) \,.
\end{array}
\end{equation}
\item\label{ass-cauchy} \textbf{\boldmath Uniqueness of solutions.}
The following differential equation on $\mathcal{C}$:
\begin{equation}
\label{eq:0024}
\dot\psi=a(\psi,\varphi)\,,\ \ \ \dot\varphi=b(\psi,\varphi)
\end{equation}
has, for any $(\psi^o\!,\varphi^o)\in\mathcal{C}$, a unique solution $t\mapsto(\psi(t),\varphi(t))$ such that
$(\psi(0),\varphi(0))=(\psi^o\!,\varphi^o)$, defined on a maximum open interval of definition $(\tau^-\!,\tau^+)$. In this interval,
$\tau^-<0<\tau^+$ where $\tau^{-}$ is such that either
$\tau^-=-\infty$ or $\varphi(\tau^-)=\pm\frac\pi2$, and $\tau^{+}$ is such that either $\tau^+=+\infty$ or
$\varphi(\tau^+)=\pm\frac\pi2$. This defines a flow $\Phi$ from an open subset of $\mathcal{C}\times\RR$ to
$\mathcal{C}$ such that the above unique solution is
\begin{equation}
\label{zeq:12}
t\mapsto\Phi(\psi^o,\varphi^o,t)\,.
\end{equation}
\item \label{ass-b} \textbf{\boldmath Sign and zeroes of $b$.}
There exists a continuous map
\begin{equation}
\label{zeq:13}
Z_b:\,[0,\frac\pi2]\to(-\frac\pi2,0]
\end{equation}
\emph{continuously differentiable on the open interval} $(0,\frac\pi2)$, such that
\begin{equation}
\label{zeq:19}
Z_b(0)=-\overline{\sigma}
\end{equation}
and
\begin{equation}
\label{zeq:7}
\left.\begin{array}{r}
b(\psi,\varphi)=0,\\ \varphi\geq0
\end{array}\right\}
\ \Leftrightarrow\
\begin{cases}
\text{either}&\psi=Z_b(\varphi),\\\text{or}&\psi=\pi+Z_b(\varphi),
\\
\mbox{or}&\varphi=0\ \ \text{and}\ \ \psi\in[-\overline{\sigma},\overline{\sigma}]\cup[\pi-\overline{\sigma},\pi+\overline{\sigma}]\,.
\end{cases}
\end{equation}
Furthermore,
\begin{equation}
\label{zeq:9}
\left.\begin{array}{r}
b(\psi,\varphi)>0,\\ \varphi\geq0
\end{array}\right\}
\ \Leftrightarrow\
\begin{cases}\text{either}
& \varphi>0\ \ \text{and}\ \ Z_b(\varphi)<\psi<\pi+Z_b(\varphi)\,, \\\mbox{or}&
\varphi=0\ \ \text{and}\ \ \overline{\sigma}<\psi<\pi+\overline{\sigma} \,,
\end{cases}
\end{equation}
\item \label{ass-a} \textbf{\boldmath Sign and zeroes of $a$.}
One has
\begin{equation}
\label{zeq:15}
\begin{array}{ll}
&0<\varphi<\frac\pi2\Rightarrow a(Z_b(\varphi),\varphi)>0\,,
\hphantom{\mbox{and, if }\overline{\sigma}>0}
\\[.6ex]
&a(0,0)=0\,,
\\[1ex]
\mbox{and, if }\overline{\sigma}>0\,,\ \ \ & -\overline{\sigma}\leq\psi<0\Rightarrow a(\psi,0)>0\,.
\end{array}
\end{equation}
\item\label{ass-saddle} \textbf{\boldmath Hyperbolic saddle point at $(0,0)$.} The maps $a$ and $b$ are smooth in a neighborhood
of $(0,0)$ and
\begin{equation}
\label{zeq:16}
a(0,0)=b(0,0)=0\,,\ \ \
\frac{\partial a}{\partial\psi}(0,0) \,\frac{\partial b}{\partial\varphi}(0,0)-\frac{\partial a}{\partial\varphi}(0,0) \,\frac{\partial b}{\partial\psi}(0,0)<0\,.
\end{equation}
\item \label{ass-c} \textbf{\boldmath Values of $c$ at equilibria.}
\begin{equation}
\label{zeq:18}
c(0,0)=1\,,\ \ \ c(\pi,0)=-1\,.
\end{equation}
\item \label{ass-US} \textbf{\boldmath Stable and unstable manifolds of $(0,0)$.}
There exists continuous maps
\begin{equation}
\label{zeq:1}
S:\,[0,\frac\pi2]\to[-\pi,0]
\,,\
U:\,[0,\frac\pi2]\to[0,\pi]
\,,
\end{equation}
\emph{continuously differentiable on the open interval} $(0,\frac\pi2)$, and a number $\overline{\sigma}$ with
\begin{equation}
\label{zeq:3}
U(0)=0, \ \ S(0)=-\overline{\sigma},\ \,\overline{\sigma}\geq0\,,
\end{equation}
such that
the stable and unstable manifolds of $(0,0)$ are described by
\begin{equation}
\label{zeq:10}
\begin{array}{l}
\mathcal{S}^0=\{(S(\varphi),\varphi),\,0\leq \varphi<\frac\pi2\}\cup\,[-\overline{\sigma},\overline{\sigma}]\!\times\!\{0\}
\,
\cup \,\{(-S(-\varphi),\varphi),\,-\frac\pi2<\varphi\leq 0\}
\\
\mathcal{U}^0=\{(U(\varphi),\varphi),\,0\leq \varphi<\frac\pi2\}\cup\{(-U(-\varphi),\varphi),\,-\frac\pi2<\varphi\leq 0\}
\end{array}
\end{equation}
Furthermore, the zeroes of $b$ are positioned with respect to the stable and unstable manifolds so that
the maps $S,U,Z_b$ satisfy:
\begin{equation}
\label{zeq:2}
0<\varphi<\frac\pi2\Rightarrow S(\varphi)<Z_b(\varphi)<0<U(\varphi)\,.
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof}
See Appendix~\ref{sec-proof-check}.\qed
\end{proof}
The following theorem is almost independent of the rest of the paper: it states that for any $a,b,c$
that satisfy the seven conditions established in Proposition~\ref{lem-check}, the differential equation \eqref{eq:0024}
has some properties (that will lead to geodesic convexity); the conditions are of course much more general than the two
cases considered in Proposition~\ref{lem-check}.
Theorem~\ref{th:convex} will be easily deduced from Theorem~\ref{prop:conv}.
\begin{theorem}
\label{prop:conv}
If $a,b,c$ satisfy the properties of Proposition~\ref{lem-check}, i.e. \eqref{eq:0036} through \eqref{zeq:18}, then,
for any $\varphi^0$ and $\varphi^1$ in the interval $(-\pi/2,\pi/2)$ and any $\bar\lambda\in\RR$,
there exists $\tfin\geq0$ and a solution $(\psi(.),\varphi(.)):[0,\tfin]\to\mathcal{C}$ of
\eqref{eq:0024} such that
\begin{equation}
\label{eq:0027}
\varphi(0)=\varphi^0\,,\ \ \ \varphi(\tfin)=\varphi^1\,,\ \ \ \int_0^{\tfin}c(\psi(\tau),\varphi(\tau))\mathrm{d}\tau=\bar\lambda\,.
\end{equation}
\end{theorem}
\noindent\textit{Proof of Theorem~\ref{prop:conv}.}
See Appendix~\ref{sec-proof-thm}.\qed
\noindent\textit{Proof of Theorem~\ref{th:convex}.}
Pick $n^0,e^0,n^1,e^1$; according to Proposition~\ref{lem-check}, Theorem~\ref{prop:conv} applies to $a,b,c$ defined
either by \eqref{eq:0026} or by \eqref{A2}. Take
$$
\varphi^0=\arcsin e^0\,,\ \ \varphi^1=\arcsin e^1\,,\ \ \bar\lambda=-\frac13\ln\frac{n^1}{n^0}
$$
and apply this theorem.
Use \eqref{eq:29} to get $n(\tau)$ and \eqref{eq:34} or \eqref{39} to get $\rho(\tau)$ (with some arbitrary $\rho(0)$,
for instance $\rho(0)=1$) and finally
\eqref{eq:19} to get $e(\tau),p_n(\tau),p_e(\tau)$ from $\psi(\tau),\varphi(\tau),n(\tau),\rho(\tau)$. Apply the time reparametrization ($\tau\leadsto t$) given by
\eqref{eq:0054} or \eqref{40}, $T$ being deduced from $\tfin$ in the same way.
According to Proposition~\ref{prop-varpsi},
the obtained $t\mapsto(n(t),e(t),p_n(t),p_e(t))$ satisfies the conclusions of Theorem~\ref{th:convex}.\qed
\subsection{Simulations}
\begin{figure}[ht!]
\centering
\includegraphics[width= 1\textwidth,height= .48\textwidth]{Simulations_Full_Zb_thick_and_arrows.jpg}
\caption{Numerical plot (obtained using Matlab) of the stable and unstable manifolds (bold)
and trajectories through a number of arbitrary initial values in $\mathcal{C}$ for the full control case. The other
curve shown is $\psi=Z_b(\varphi)$.
\label{Fig_1}
}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=1\textwidth,height= .5\textwidth]{SimTanNR.jpg}
\caption{Numerical plot (obtained using Matlab) of the stable and unstable manifolds (bold)
and trajectories through a number of arbitrary initial values in $\mathcal{C}$ for the tangential case. The other curve shown
is $\psi=Z_b(\varphi)$.
\label{Fig_2}
}
\end{figure}
A numerical simulation of the phase portrait of the differential equation \eqref{eq:0024} (or the first two equations
in \eqref{A5}) is displayed in Figure~\ref{Fig_1} in the ``full control case'' where $a$ and $b$ are given by
\eqref{eq:0026} and in Figure~\ref{Fig_2} in the ``tangential thrust case'' where $a$ and $b$ are given by \eqref{A2}.
This is supposed to be a phase portrait on the cylinder $\mathcal{C}$ (for instance, identify $\{\psi=\frac\pi2\}$ with $\{\psi=\frac{3\pi}{2}\}$).
The thick trajectories are the stable and unstable manifolds of $(0,0)$ and $(\pi,0)$; the other thick curve is the set
of zeroes of $b(\psi,\varphi)$ (i.e. the isocline $\{\dot\varphi=0\}$).
One may check visually the properties established in Proposition~\ref{lem-check}; in particular the unstable manifold of $(0,0)$
is, in both cases, a graph $\varphi\mapsto\psi$ while the stable manifold is also such a graph in the full control case (Figure~\ref{Fig_1})
but not in the tangential thrust case (Figure~\ref{Fig_1}) where it comprises a segment of the $\psi$-axis.
It can be seen that in both
cases, the cylinder $\mathcal{C}$ is divided into six
regions by these invariant manifolds: one region (called $F$ in Appendix~\ref{sec-proof-thm}) where all trajectories go
``up'' ($\varphi$ is monotone increasing), one (called $F^+$ in Appendix~\ref{sec-proof-thm}) where all
trajectories go ``down'', and four other regions (called $E$, $E^\sharp$, $E^+$ and $E^{+\sharp}$ in
Appendix~\ref{sec-proof-thm}) where all trajectories cross once the isocline $\{\dot\varphi=0\}$ so that they go up and
then down or down and then up.
This is exploited in the proof of Theorem~\ref{prop:conv}.
The generic figure \ref{fig:regions} is a drawing used to support that proof, that figures in an illustrative manner the
features contained in the assumptions of Theorem~\ref{prop:conv}, and that can also be observed in the numerical
simulations of the two cases that we are really interested in (Theorem~\ref{Th1}).
\section{Comparison between the minimum-energy and minimum-time cases from the convexity point of view}
\label{sec:compa}
In section \ref{sec:energy} we recalled some results from \cite{Bonn-Cai09forum} (and previous work by the same
authors); in particular, Theorem~\ref{th:droites} states that the elliptic domain is
\emph{not} geodesically convex for the energy minimization problem, i.e. some pairs of points in $\mathcal{E}$ cannot be
joined by a geodesic.
In that case, in suitable coordinates ($(n^{5/6},\sqrt{5/2}\varphi)$ as polar coordinates), geodesics are straight lines
hence geodesic convexity reduces to usual (affine) convexity, thus the
simplest way to see this non convexity is to determine the shape on the elliptic domain in these polar coordinates.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{image11b2}
\caption{The phase portrait, for energy minimization, in the same coordinates as Figure \ref{Fig_1} and \ref{Fig_2}. There are
two lines of non isolated equilibria. The darker zone is all the points that can be reached in positive time from
the line $\{\varphi=\varphi^0\}$, with $\varphi^0$ rather close to $-\frac\pi2$. The highest possible final
value is $\varphi$ is $\varphi^1_{\mathsf{max}}=\varphi^0+\sqrt{2/5}\,\pi$.}
\label{fig:L2}
\end{figure}
Here we try to explain why convexity holds in the minimum-time case and not in the minimum-energy case.
Using the coordinates from Theorem~\ref{th:droites} for the time-minimizing problem does not seem to shed any light.
Rather, we explain how the proof of convexity that we made in the minimum-time case fails when applied to
the minimum-energy case.
When $p_\omega=0$, the Hamiltonian in the minimum-energy case is given by \eqref{eq:30} and can be written as follows
\begin{equation}
\label{eq:39}
H=n^{-5/3}\left[2(3n\,p_{n})^{2}+5p_{\varphi}^{2}\right]
\end{equation}
in the coordinates
$(n,\varphi,p_n,p_\varphi)$ that result from the symplectic change of coordinates $e=\sin\varphi$, $p_\varphi=\sqrt{1-e^{2}}p_e$.
The Hamiltonian equations can be written
$$
\begin{array}{ll}
\dot n=12 \,n^{-2/3}\,(3n\,p_{n})\,, \hspace{2em}& {\textstyle\frac{\mathrm{d}}{\mathrm{d}t}} (3n\,p_{n}) =
5n^{-5/3}\left[2(3n\,p_{n})^{2}+5p_{\varphi}^{2}\right]\,,
\\
\dot \varphi = 10\,n^{-5/3} \,p_\varphi\,, & \hspace{3em}\dot{p}_\varphi=0\,.
\end{array}
$$
With the same polar coordinates as in \eqref{eq:19}\; (namely $\rho\cos\psi=3\,n\,p_n$, $\rho\sin\psi=-\,p_\varphi$),\; and
the time reparametrization $\mathrm{d}t=5\,n^{-5/3}\mathrm{d}\tau$,\; the state equations of these Hamiltonian equations have the form
\begin{equation}
\label{eq:40}
\mathrm{d}\psi/\mathrm{d}\tau=-\sin\psi\,(2+3\sin^2\!\psi)\,,\ \
\mathrm{d}\varphi/\mathrm{d}\tau=2\sin\psi\,.
\end{equation}
It is easy to describe the solutions of these equations on the cylinder $\mathcal{C}$ (see \eqref{eq:0069}).
There are two lines of equilibria at $\psi=0$ and $\psi=\pi$ and
\begin{equation}
\label{eq:41}
\varphi+\sqrt{\frac25}\arctan\left(\sqrt{\frac52}\tan\psi\right)
\end{equation}
is a first integral (it is smooth at $\psi=\frac\pi2$).
These solutions are drawn on Figure~\ref{fig:L2}. It is clear
that, on a solution, the maximum possible variation of the variable $\varphi$ is $\sqrt{2/5}\,\pi$;\; this implies that, if
$|\varphi^0|>(\sqrt{2/5}-\frac12)\pi$,\; there are some values of $\varphi$ that cannot be reached by any solution
starting from the line $\{\varphi=\varphi^0\}$.
\section{Conclusion and open problems}
We have studied the average minimum time problem as described in section~\ref{sec:Tmin:Ham}.
This is a reduced subproblem of the planar transfer problem: the state has dimension 2, whereas it would have
dimension 3 in the real planar problem (we have set $p_\omega=0$; this imposes that the cyclic variable is constant
along transfers) and dimension 5 in the full problem where the plane containing the orbits is not fixed.
In \cite{Bonn-Cai09forum,Bonn-Cai-Duj06b}, the energy problem in full dimension is treated; the planar case is integrable (but
only the reduced planar case is flat); the full problem is not integrable but extremals may still be computed
explicitly.
Studying minimum time in higher dimension is an interesting program.
Concerning the reduced problem considered here, the main contribution of the paper is to prove geodesic convexity of the
elliptic domain (any two points in the domain may be joined by an extremal trajectory). On the one hand, it is not clear
that this result holds true in higher dimension, and on the other hand, in the present small dimension, optimality
and/or uniqueness of the extremal trajectories has not been studied.
Finally the singularities of the Hamiltonian have been investigated roughly, mostly to ensure existence of a
Hamiltonian \emph{flow}. It would be interesting to better understand their nature and their role, in particular the
singularities they cause on the balls of small radius for the metric.
\bigskip
|
\section{The He-like sequence ($N=2$)}
\label{He}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.5 + 0.125~(Z/137.036)^2,\\
\nonumber a_1 &=& -0.625 -0.0989536~(Z/137.036)^2,\\
\nonumber a_0 &=& 0.368987,\\
\nonumber a_{-1} &=& -0.216282.
\label{he_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm H}) &=& 0.0277063,\\
s &=& 0.59129.
\label{rest_he}
\end{eqnarray}
The slope was computed from $E_a({\rm H})$ and $R_{cov}({\rm H})=0.604712$. Here, and in the analysis below, data for $E_a$ and $R_{cov}$ are taken from Ref. \onlinecite{RSC}.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n2.pdf}
\caption{\label{fig2} (Color online) The He-like systems ($N=2$). The difference between NIST values and the relativistic RPT versus atomic number. Error bars are very small and can not be seen in the scale of the figure. No inconsistencies were detected.}
\end{figure}
\end{center}
\textbf{Discussion:}
Let us comment on some features of the He-like case. In the range $2 \le Z \le 60$, the difference between our RPT series and the reported values is very well behaved. In the intermediate region, the maximum relative error is around $0.15\%$.
When $Z>56$, NIST-RPT rises. NIST ionization potentials in the $12 \le Z \le 100$ range come mainly from \textit{ab intio} QED calculations by Artemyev et al., \cite{HeQED} which include finite nuclear-size effects. Our perturbative treatment of relativity cannot reproduce their results for very large $Z$.
\section{Second row elements}
\subsection{The Be-like sequence ($N=4$)}
\label{Be}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.125 + 0.0390625~(Z/137.036)^2,\\
\nonumber a_1 &=& -0.548196 -0.0952831~(Z/137.036)^2,\\
\nonumber a_0 &=& 0.675679,\\
\nonumber a_{-1} &=& -0.400253.
\label{be_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Li}) &=& 0.0227050,\\
s &=& 0.246343.
\label{rest_be}
\end{eqnarray}
The slope was computed from $E_a({\rm Li})$ and $R_{cov}({\rm Li})=2.456644$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n4.pdf}
\caption{\label{fig4} (Color online) The Be-like isoelectronic sequence ($N=4$). The abrupt jump at $Z =50-51$ can be, however, accomodated within error bars.}
\end{figure}
\end{center}
\textbf{Discussion:}
The comparison between NIST and RPT values for the Be-like sequence shows an abrupt jump of $0.28$ a.u. at $Z=50-51$ (see Fig. \ref{fig4}). These are numbers based on Dirac-Fock calculations of $I_p$ computed by different groups. When $Z\le 50$, numbers come from Ref. [\onlinecite{Interpol_N42-50}], whereas for $Z \ge 51$ almost all reported numbers come from Rodrigues et al. \cite{Dirac-Fock1}. We notice that, in Ref. [\onlinecite{Interpol_N42-50}], a formula like Eq. (\ref{ip}) is used as a fit to correct the computed values.
The observed jump is consistent with the natural dispersion of points, as suggested by the reported error bars. Thus, no inconsistency is detected.
\subsection{The C-like sequence ($N=6$)}
\label{C}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.125 + 0.0390625~(Z/137.036)^2,\\
\nonumber a_1 &=& -0.945089 -0.241928~(Z/137.036)^2,\\
\nonumber a_0 &=& 0.0390625,\\
\nonumber a_{-1} &=& 0.643271.
\label{c_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm B}) &=& 0.0102761,\\
s &=& 0.279379.
\label{rest_c}
\end{eqnarray}
The slope was computed from $E_a({\rm B})$ and $R_{cov}({\rm B})=1.58737$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n6.pdf}
\caption{\label{fig6} (Color online) The C-like systems ($N=6$). An inconsistency at $Z=74$ and an abrupt jump at $Z=50-51$) are noticed.}
\end{figure}
\end{center}
\textbf{Discussion:}
Analysis of Fig. \ref{fig6} (C-like sequence) also shows an abrupt jump of nearly $0.43$ a.u. at $Z=50- 51$. As in the Be-like case, this jump is associated to changes in the calculation methodology.
However, the jump exactly equals the reported error bars. Thus, we recommend revision of this data.
In addition, an apparent deviation in the ionization potential is noticed at $Z=74$. This number, corresponding to a tungsten heavy ion (W$^{+68}$), was collected by NIST compilers Kramida and Reader with the help of a semi-empirical approach \cite{W-ions1}. Comparison with the average curve suggests that the reported value for W$^{+68}$ is overestimated in 1.577 a.u.
\subsection{The N-like sequence ($N=7$)}
\label{N}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.125 + 0.0078125~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.10915 -0.152346~(Z/137.036)^2,\\
\nonumber a_0 &=& 2.26192,\\
\nonumber a_{-1} &=& -0.356612.
\label{n_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm C}) &=& 0.0463657,\\
s &=& 0.400642.
\label{rest_n}
\end{eqnarray}
The slope was computed from $E_a({\rm C})$ and $R_{cov}({\rm C})=1.417295$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n7.pdf}
\caption{\label{fig7} (Color online) The N-like ions ($N=7$). Deviations at $Z=26$ and 74, and a strong dispersion of points in the region $Z=42-50$ are noticed. However, the latter is consistent with the relatively high error bars.}
\end{figure}
\end{center}
\textbf{Discussion:}
The N-like sequence (Fig. \ref{fig7}) presents noticeable jumps at $Z=26$ and, once more, at $Z=74$.
The ionization potential for $Z=26$ was taken from the paper by Sugar and Corliss.\cite{Sugar} According to our procedure, an overestimation of 0.051 a.u. is noticed.
The $Z=74$ case is again taken from Ref. [\onlinecite{W-ions1}]. It seems to be an inconsistent point, underestimated by $1.395$ a.u.
Data for $Z=42-50$ come from the Dirac-Fock calculations by Bi\'emont et al. \cite{Interpol_N42-50} The reported large uncertainties in the data, of around $1.5$ a.u., are consistent with the observed deviations.
Besides, the great dispersion for $Z \ge 96$, questions the consistency of Dirac-Fock calculations by Rodrigues et al.\cite{Dirac-Fock1} for highly charged ions. Notice, however, that deviations are within error bars, which are remarkably high (from $2.0$ to $20$ a.u.) for $Z > 74$.
\subsection{The F-like sequence ($N=9$)}
\label{F}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.125 + 0.0078125~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.4654 -0.181681~(Z/137.036)^2,\\
\nonumber a_0 &=& 2.79034,\\
\nonumber a_{-1} &=& 7.91836.
\label{f_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm O}) &=& 0.0536759,\\
s &=& 0.410681.
\label{rest_f}
\end{eqnarray}
The slope was computed from $E_a({\rm O})$ and $R_{cov}({\rm O})=1.417295$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n9.pdf}
\caption{\label{fig9} (Color online) The F-like sequence ($N=9$). There are remarkable inconsistencies at $Z=27$ and 74. The jump at $Z=50-51$ is within error bars.}
\end{figure}
\end{center}
\textbf{Discussion:}
In the F-like sequence (Fig. \ref{fig9}), high deviations are noticed for $Z=27$ and 74. $I_p$ for $Z=27$ comes from Ref. [\onlinecite{Sugar}], the very same paper by Sugar and Corliss cited in the N-like case. According to our Fig., it seems to be overestimated in 0.183 a.u.
On the other hand, $Z=74$ belongs to the case of tungsten ions (W$^{+65}$), and the $I_p$ value is, once more, taken from Ref. [\onlinecite{W-ions1}]. In this case, an underestimation of 1.071 a.u. is apparent.
Finally, the observed jump at $Z=50-51$ is consistent with the reported error bars, and there is also a great dispersion of the data for $Z > 100$, coming again from the Dirac-Fock calculations of Ref. [\onlinecite{Dirac-Fock1}]. It is also worth mentioning the large error bars accompanying the data for $Z \ge 80$.
\subsection{The Ne-like sequence ($N=10$)}
\label{Ne}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.125 + 0.0078125~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.63649 -0.195621~(Z/137.036)^2,\\
\nonumber a_0 &=& 4.13824,\\
\nonumber a_{-1} &=& 5.3556.
\label{ne_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm F}) &=& 0.124985,\\
s &=& 0.547149.
\label{rest_ne}
\end{eqnarray}
The slope was computed from $E_a({\rm F})$ and $R_{cov}({\rm F})=1.13384$.
\textbf{Discussion:}
In the Ne-like sequence (Fig. \ref{fig10}), deviations at $Z = 21$, 22, and 27 are noticed. Points are related to Ref. [\onlinecite{Sugar}]. We suggest correcting these values in -0.034, +0.012 and +0.036 a.u., respectively.
The $Z = 74$ case, \cite{W-ions1} is also clearly inconsistent. The ionization potential seems to be underestimated in 1.077 a.u.
The previously discussed abrupt jump at $Z =50- 51$, should be revised by the NIST team because the jump is greater than the data uncertainty.
Finally, the great dispersion of points for $Z \ge 101$, related to the Dirac-Fock calculations in Ref. \onlinecite{Dirac-Fock1}, are within uncertainty bars.
\section{Third row elements}
\subsection{The Mg-like sequence ($N=12$)}
\label{Mg}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.0138889~(Z/137.036)^2,\\
\nonumber a_1 &=& -0.931477 -0.0906944~(Z/137.036)^2,\\
\nonumber a_0 &=& 2.62996,\\
\nonumber a_{-1} &=& 10.0077.
\label{mg_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Na}) &=& 0.0201287,\\
s &=& 0.209421.
\label{rest_mg}
\end{eqnarray}
The slope was computed from $E_a({\rm Na})$ and $R_{cov}({\rm Na})=3.023562$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n12.pdf}
\caption{\label{fig12} (Color online) Mg-like systems ($N=12$). Inconsistencies at $Z=23$, 24, 25, 27 and 28, and a jump at $Z =50- 51$ are apparent.}
\end{figure}
\end{center}
\textbf{Discussion:}
The Mg-like sequence (Fig. \ref{fig12}) of ions shows clear deviations at $Z=23$, 24, 25, 27 and 28.\cite{Sugar} The figure suggests corrections of -0.015, -0.013, +0.026, -0.008 and -0.016 a.u. to these points, respectively.
The jump at $Z=50- 51$ was discussed above. It is clearly not consistent with error bars, thus we suggest revision of these data.
\subsection{The Si-like sequence ($N=14$)}
\label{Si}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.0138889~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.13489 -0.148767~(Z/137.036)^2,\\
\nonumber a_0 &=& 4.44303,\\
\nonumber a_{-1} &=& 12.1396.
\label{si_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n14.pdf}
\caption{\label{fig14} (Color online) The Si-like sequence ($N=14$). Deviations at $Z =21-26$ and Z=74 are remarkable.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Al}) &=& 0.0159006,\\
s &=& 0.239636.
\label{rest_si}
\end{eqnarray}
The slope was computed from $E_a({\rm Al})$ and $R_{cov}({\rm Al})=2.343260$.
\textbf{Discussion:}
It seems that there is a group of four points, $Z=23-26$, out of the general trend. Thus, we use a 7-points running average (instead of 5-points) in order to create a smooth curve. We suggest using the top of the error bars as values for $Z=21-22$, and the bottom of the error bars for $Z=23-26$.
The $Z=74$ point, on the other hand, should be corrected in -0.329 a.u.
\subsection{The P-like sequence ($N=15$)}
\label{P}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.19956 -0.109484~(Z/137.036)^2,\\
\nonumber a_0 &=& 5.24386,\\
\nonumber a_{-1} &=& 10.0615.
\label{p_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Si}) &=& 0.0510817,\\
s &=& 0.304873.
\label{rest_p}
\end{eqnarray}
The slope was computed from $E_a({\rm Si})$ and $R_{cov}({\rm Si})=2.154288$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n15.pdf}
\caption{\label{fig15} (Color online) The P-like systems ($N=15$). A big jump at $Z =50- 51$, and a noticeable deviation at $Z=74$ are apparent.}
\end{figure}
\end{center}
\textbf{Discussion:}
We suggest a revision of the data near $Z=50$, and correcting the $Z=74$ value in + 0.247 a.u.
\subsection{The S-like sequence ($N=16$)}
\label{S}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.26711 -0.114919~(Z/137.036)^2,\\
\nonumber a_0 &=& 4.83765,\\
\nonumber a_{-1} &=& 25.569.
\label{s_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n16.pdf}
\caption{\label{fig16} (Color online) The S-like ions ($N=16$). The only detected inconsistency is the deviation at $Z=74$.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm P}) &=& 0.0274519,\\
s &=& 0.286207.
\label{rest_s}
\end{eqnarray}
The slope was computed from $E_a({\rm P})$ and $R_{cov}({\rm P})=2.0598015$.
\textbf{Discussion:}
Although error bars are relatively high in this case, we suggest a correction of + 0.249 a.u. to the $Z=74$ value of $I_p$.
\subsection{The Cl-like sequence ($N=17$)}
\label{Cl}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~(Z/137.036)^2,\\
\nonumber a_1 &=& -1.34042 -0.121136~(Z/137.036)^2,\\
\nonumber a_0 &=& 5.81594,\\
\nonumber a_{-1} &=& 23.9212.
\label{cl_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm S}) &=& 0.0763283, \\
s &=& 0.344289.
\label{rest_cl}
\end{eqnarray}
The slope was computed from $E_a({\rm S})$ and $R_{cov}({\rm S})=1.965315$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n17.pdf}
\caption{\label{fig17} (Color online) Cl-like systems ($N=17$). The jump at $Z=50- 51$, and the deviation at $Z=74$ are noticeable.}
\end{figure}
\end{center}
\textbf{Discussion:}
Although error bars are relatively high in this case, we suggest a revision of the data near $Z=50$, which show a jump of 0.216 a.u., and using in the $Z=74$ case the top value of the error bar.
\subsection{The Ar-like sequence ($N=18$)}
\label{Ar}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n18.pdf}
\caption{\label{fig18} (Color online) The Ar-like systems ($N=18$). In this case, we distinguish only the deviation at $Z=74$.\cite{W-ions1}}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.40796 -0.12657~ (Z/137.036)^2,\\
\nonumber a_0 &=& 6.30212,\\
\nonumber a_{-1} &=& 29.2914.
\label{ar_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Cl}) &=& 0.132775, \\
s &=& 0.380045.
\label{rest_ar}
\end{eqnarray}
The slope was computed from $E_a({\rm Cl})$ and $R_{cov}({\rm Cl})=1.889726$.
\textbf{Discussion:}
Although error bars are relatively high, we suggest using the top of the error bar as the $Z=74$ value of $I_p$.
\section{Four row elements}
\subsection{The Ca-like sequence ($N=20$)}
\label{Ca}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.70312 -0.200236~ (Z/137.036)^2,\\
\nonumber a_0 &=& 7.96996,\\
\nonumber a_{-1} &=& 83.4702.
\label{ca_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n20.pdf}
\caption{\label{fig20} (Color online) The Ca-like systems ($N=20$). Only the $Z=74$ point is distinguished.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm K}) &=& 0.0184220, \\
s &=& 0.176307.
\label{rest_ca}
\end{eqnarray}
The slope was computed from $E_a({\rm K})$ and $R_{cov}({\rm K})=3.779452$.
\textbf{Discussion:}
The Ca-like sequence is the first with a rearrangement of the electronic spectrum with the increase of $Z$. For $Z\approx N$ the last two electrons occupy the 4s subshell, whereas for larger $Z$ they move to the 3d orbital. The observed jump at $Z=20-21$ is surely related to this fact.
On the other hand, we suggest using for $I_p$ at $Z=74$ the bottom of its error bar.
\subsection{The Sc-like sequence ($N=21$)}
\label{Sc}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.79098 -0.209669~ (Z/137.036)^2,\\
\nonumber a_0 &=& 7.18837,\\
\nonumber a_{-1} &=& 129.196.
\label{sc_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n21.pdf}
\caption{\label{fig21} (Color online) The Sc-like systems ($N=21$). Only the $Z=74$ point is distinguished.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ca}) &=& 0.000901924, \\
s &=& 0.107729.
\label{rest_sc}
\end{eqnarray}
The slope was computed from $E_a({\rm Ca})$ and $R_{cov}({\rm Ca})=3.288123$.
\textbf{Discussion:}
Very similar to the Ca sequence. We suggest correcting $I_p$ at $Z=74$ in -0.224 a.u.
\subsection{The Ti-like sequence ($N=22$)}
\label{Ti}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00462963~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.87158 -0.218161~ (Z/137.036)^2,\\
\nonumber a_0 &=& 8.87906,\\
\nonumber a_{-1} &=& 125.803.
\label{ti_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Sc}) &=& 0.00690650, \\
s &=& 0.175933.
\label{rest_ti}
\end{eqnarray}
The slope was computed from $E_a({\rm Sc})$ and $R_{cov}({\rm Sc})=3.004665$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n22.pdf}
\caption{\label{fig22} (Color online) Ti-like systems ($N=22$). Only the $Z=74$ point is distinguished.}
\end{figure}
\end{center}
\textbf{Discussion:}
Very similar to the previous sequences. We suggest correcting $I_p$ at $Z=74$ in -0.205 a.u.
\subsection{The V-like sequence ($N=23$)}
\label{V}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00154321 (Z/137.036)^2,\\
\nonumber a_1 &=& -1.95847 -0.154589~ (Z/137.036)^2,\\
\nonumber a_0 &=& 8.95387,\\
\nonumber a_{-1} &=& 160.08.
\label{v_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n23.pdf}
\caption{\label{fig23} (Color online) The V-like ions ($N=23$). No inconsistencies were detected.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ti}) &=& 0.00290230, \\
s &=& 0.156499.
\label{rest_v}
\end{eqnarray}
The slope was computed from $E_a({\rm Ti})$ and $R_{cov}({\rm Ti})=2.796794$.
\subsection{The Cr-like sequence ($N=24$)}
\label{Cr}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00154321~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.03885 -0.15826~ (Z/137.036)^2,\\
\nonumber a_0 &=& 10.9246,\\
\nonumber a_{-1} &=& 153.618.
\label{cr_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n24.pdf}
\caption{\label{fig24} (Color online) The Cr-like sequence ($N=24$). An apparent deviation at $Z=32$
is detected.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm V}) &=& 0.0192866, \\
s &=& 0.223850.
\label{rest_cr}
\end{eqnarray}
The slope was computed from $E_a({\rm V})$ and $R_{cov}({\rm V})=2.721206$.
\textbf{Discussion:}
In the Cr-like sequence, we shall distinguish the problematic point at $Z=32$, coming from the paper by Sugar and Musgrove \cite{Sugar1}. $I_p$ is overestimated in 0.181 a.u.
\subsection{The Mn-like sequence ($N=25$)}
\label{Mn}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00154321~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.12264 -0.162119~ (Z/137.036)^2,\\
\nonumber a_0 &=& 12.0502,\\
\nonumber a_{-1} &=& 168.234.
\label{mn_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n25.pdf}
\caption{\label{fig25} (Color online) The Mn-like systems ($N=25$). No inconsistency is detected.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Cr}) &=& 0.0244666, \\
s &=& 0.249254.
\label{rest_mn}
\end{eqnarray}
The slope was computed from $E_a({\rm Cr})$ and $R_{cov}({\rm Cr})=2.456644$.
\subsection{The Co-like sequence ($N=27$)}
\label{Co}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n27.pdf}
\caption{\label{fig27} (Color online) The Co-like sequence ($N=27$). The only detected inconsistency is at $Z=74$.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00154321~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.29511 -0.170117~ (Z/137.036)^2,\\
\nonumber a_0 &=& 11.9833,\\
\nonumber a_{-1} &=& 266.793.
\label{co_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Fe}) &=& 0.00554713, \\
s &=& 0.195878.
\label{rest_co}
\end{eqnarray}
The slope was computed from $E_a({\rm Fe})$ and $R_{cov}({\rm Fe})=2.343260$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is overestimated in 0.196 a.u.
\subsection{The Ni-like sequence ($N=28$)}
\label{Ni}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.0555556 + 0.00154321~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.3755 -0.173789~ (Z/137.036)^2,\\
\nonumber a_0 &=& 14.2165,\\
\nonumber a_{-1} &=& 258.787.
\label{ni_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Co}) &=& 0.0243196, \\
s &=& 0.266002.
\label{rest_ni}
\end{eqnarray}
The slope was computed from $E_a({\rm Co})$ and $R_{cov}({\rm Co})=2.229877$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n28.pdf}
\caption{\label{fig28} (Color online) The Ni-like systems ($N=28$). Inconsistencies are detected at $Z=42$, 74, and 79.}
\end{figure}
\end{center}
\textbf{Discussion:}
The Ni-like ions (Fig. \ref{fig28}) shows deviations at $Z=42$, 74 and 79, the latter coming from the paper by Tragin et al.\cite{Tragin}.
As suggested by the figure, $I_p$ at $Z=42$ is overestimated in 0.071 a.u. Corrections of -0.141 and -0.203 a.u. should be added to the points at $Z=74$ and 79, respectively.
\subsection{The Cu-like sequence ($N=29$)}
\label{Cu}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00634766~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.34713 -0.0620302~ (Z/137.036)^2,\\
\nonumber a_0 &=& 9.66505,\\
\nonumber a_{-1} &=& 96.9281.
\label{cu_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ni}) &=& 0.0424673, \\
s &=& 0.291495.
\label{rest_cu}
\end{eqnarray}
The slope was computed from $E_a({\rm Ni})$ and $R_{cov}({\rm Ni})=2.210980$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n29.pdf}
\caption{\label{fig29} (Color online) Cu-like systems ($N=29$). Small deviations at $Z=70$ and 79 are noticed.}
\end{figure}
\end{center}
\textbf{Discussion:}
In the Cu-like sequence (Fig. \ref{fig29}), slight deviations at $Z=70$ and 79 are apparent. The data come from Ref. [\onlinecite{Tragin}]. Points seem to be overestimated in 0.064 and 0.097 a.u., respectively.
\subsection{The Zn-like sequence ($N=30$)}
\label{Zn}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00634766~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.3844 -0.0677735~ (Z/137.036)^2,\\
\nonumber a_0 &=& 9.2367,\\
\nonumber a_{-1} &=& 131.194.
\label{zn_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n30.pdf}
\caption{\label{fig30} (Color online) The Zn-like systems ($N=30$). Inconsistencies are noticed at $Z=35$, 42, 45 and 74.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Cu}) &=& 0.0453693, \\
s &=& 0.285562.
\label{rest_zn}
\end{eqnarray}
The slope was computed from $E_a({\rm Cu})$ and $R_{cov}({\rm Cu})=2.305466$.
\textbf{Discussion:}
In Fig. \ref{fig30} (the Zn-like sequence), the data for $Z=35$ is taken from the Dirac-Fock calculation of Ref. [\onlinecite{Dirac-Fock1}], while the value for $Z=42$ is due to Refs. [\onlinecite{Sugar2}] and [\onlinecite{Litzen}]. These potentials should be corrected in +0.076 and -0.045 a.u., respectively.
On the other hand, $Z=45$ ionization potential was collected from the relativistic multireference many-body perturbation theory calculations of Vilkas et al. \cite{Vilkas}. It seems to be overestimated in 0.087 a.u.
The $Z=74$ ionization potential comes from Ref. [\onlinecite{W-ions1}], as before. It is 0.103 a.u. higher than the average curve.
\subsection{The Ge-like sequence ($N=32$)}
\label{Ge}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00634766~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.51951 -0.0952627~ (Z/137.036)^2,\\
\nonumber a_0 &=& 10.9568,\\
\nonumber a_{-1} &=& 185.117.
\label{ge_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ga}) &=& 0.0157966, \\
s &=& 0.240629.
\label{rest_ge}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n32.pdf}
\caption{\label{fig32} (Color online) Ge-like ions ($N=32$). Only the $Z=74$ point is distinguished.}
\end{figure}
\end{center}
The slope was computed from $E_a({\rm Ga})$ and $R_{cov}({\rm Ga})=2.324363$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is overestimated in 0.116 a.u.
\subsection{The As-like sequence ($N=33$)}
\label{As}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.55446 -0.0829917~ (Z/137.036)^2,\\
\nonumber a_0 &=& 12.6573,\\
\nonumber a_{-1} &=& 164.459.
\label{as_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ge}) &=& 0.0453118, \\
s &=& 0.288928.
\label{rest_as}
\end{eqnarray}
The slope was computed from $E_a({\rm Ge})$ and $R_{cov}({\rm Ge})=2.267671$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n33.pdf}
\caption{\label{fig33} (Color online) The As-like systems ($N=33$). No inconsistencies are detected.}
\end{figure}
\end{center}
\subsection{The Se-like sequence ($N=34$)}
\label{Se}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.59101 -0.0853527~ (Z/137.036)^2,\\
\nonumber a_0 &=& 11.7569,\\
\nonumber a_{-1} &=& 222.876.
\label{se_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm As}) &=& 0.0295464, \\
s &=& 0.271227.
\label{rest_se}
\end{eqnarray}
The slope was computed from $E_a({\rm As})$ and $R_{cov}({\rm As})=2.267671$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n34.pdf}
\caption{\label{fig34} (Color online) The Se-like sequence ($N=34$). Slight deviations at $Z =39$ and 42 are apparent.}
\end{figure}
\end{center}
\textbf{Discussion:}
In Fig. \ref{fig34} (the Se-like sequence), we detect inconsistencies at $Z=39$ and 42. $I_p$ at $Z=39$ seems to be overestimated in 0.033 a.u.
The number for $I_p$ in Mo$^{+18}$ (i.e. $Z=42$, $N=34$) comes from Refs. [\onlinecite{Khatoon}] and [\onlinecite{Sugar2}]. It seems to be 0.059 a.u. higher than the average curve.
\subsection{The Br-like sequence ($N=35$)}
\label{Br}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.63074 -0.0881566~ (Z/137.036)^2,\\
\nonumber a_0 &=& 13.0595,\\
\nonumber a_{-1} &=& 215.756.
\label{br_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n35.pdf}
\caption{\label{fig35} (Color online) Br-like systems ($N=35$). A smooth curve. No inconsistencies are detected.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Se}) &=& 0.0742704, \\
s &=& 0.312415.
\label{rest_br}
\end{eqnarray}
The slope was computed from $E_a({\rm Se})$ and $R_{cov}({\rm Se})=2.229877$.
\subsection{The Kr-like sequence ($N=36$)}
\label{Kr}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n36.pdf}
\caption{\label{fig36} (Color online) The Kr-like systems ($N=36$). No inconsistencies are detected.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.66728 -0.0905176~ (Z/137.036)^2,\\
\nonumber a_0 &=& 13.5057,\\
\nonumber a_{-1} &=& 234.611.
\label{kr_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Br}) &=& 0.123625, \\
s &=& 0.333943.
\label{rest_kr}
\end{eqnarray}
The slope was computed from $E_a({\rm Br})$ and $R_{cov}({\rm Br})=2.210980$.
\section{Fifth row elements}
\subsection{The Sr-like sequence ($N=38$)}
\label{Sr}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n38.pdf}
\caption{\label{fig38} (Color online) The Sr-like systems ($N=38$). An apparent deviation at $Z=74$\cite{W-ions1} is shown.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.86543 -0.136324~ (Z/137.036)^2,\\
\nonumber a_0 &=& 15.882,\\
\nonumber a_{-1} &=& 388.488.
\label{sr_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Rb}) &=& 0.0178511, \\
s &=& 0.166529.
\label{rest_sr}
\end{eqnarray}
The slope was computed from $E_a({\rm Rb})$ and $R_{cov}({\rm Rb})=4.062911$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is overestimated in 0.132 a.u.
\subsection{The Y-like sequence ($N=39$)}
\label{Y}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.90897 -0.139844~ (Z/137.036)^2,\\
\nonumber a_0 &=& 14.2754,\\
\nonumber a_{-1} &=& 504.63.
\label{y_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Sr}) &=& 0.00176347, \\
s &=& 0.120083.
\label{rest_y}
\end{eqnarray}
The slope was computed from $E_a({\rm Sr})$ and $R_{cov}({\rm Sr})=3.590480$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n39.pdf}
\caption{\label{fig39} (Color online) The Y-like systems ($N=39$). The only detected inconsistency is at $Z=74~$\cite{W-ions1}.}
\end{figure}
\end{center}
\textbf{Discussion:}
$I_p$ at $Z=74$ should be corrected in -0.148 a.u.
\subsection{The Zr-like sequence ($N=40$)}
\label{Zr}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00244141~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.94905 -0.142908~ (Z/137.036)^2,\\
\nonumber a_0 &=& 16.4347,\\
\nonumber a_{-1} &=& 476.155.
\label{zr_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n40.pdf}
\caption{\label{fig40} (Color online) The Zr-like sequence ($N=40$). The $Z=74$ point is distinguished.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Y}) &=& 0.0112782, \\
s &=& 0.179247.
\label{rest_zr}
\end{eqnarray}
The slope was computed from $E_a({\rm Y})$ and $R_{cov}({\rm Y})=3.325918$.
\textbf{Discussion:}
$I_p$ at $Z=74$ shows an overestimation of 0.147 a.u.
\subsection{The Nb-like sequence ($N=41$)}
\label{Nb}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -1.99199 -0.112623~ (Z/137.036)^2,\\
\nonumber a_0 &=& 17.6019,\\
\nonumber a_{-1} &=& 492.881.
\label{nb_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n41.pdf}
\caption{\label{fig41} (Color online) The Nb-like systems ($N=41$). The only detected inconsistency is the $Z=74$ point.\cite{W-ions1}}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Zr}) &=& 0.0156496, \\
s &=& 0.198128.
\label{rest_nb}
\end{eqnarray}
The slope was computed from $E_a({\rm Zr})$ and $R_{cov}({\rm Zr})=3.099151$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is overestimated in 0.116 a.u.
\subsection{The Mo-like sequence ($N=42$)}
\label{Mo}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.03196 -0.114317~ (Z/137.036)^2,\\
\nonumber a_0 &=& 18.7868,\\
\nonumber a_{-1} &=& 503.25.
\label{mo_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Nb}) &=& 0.0336625, \\
s &=& 0.229290.
\label{rest_mo}
\end{eqnarray}
The slope was computed from $E_a({\rm Nb})$ and $R_{cov}({\rm Nb})=2.947973$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n42.pdf}
\caption{\label{fig42} (Color online) The Mo-like ions ($N=42$). The only inconsistent point is $Z=74$\cite{W-ions1}.}
\end{figure}
\end{center}
\textbf{Discussion:}
$I_p$ at $Z=74$ should be corrected in -0.120 a.u.
\subsection{The Tc-like sequence ($N=43$)}
\label{Tc}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.07354 -0.116109~ (Z/137.036)^2,\\
\nonumber a_0 &=& 19.0013,\\
\nonumber a_{-1} &=& 556.886.
\label{tc_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n43.pdf}
\caption{\label{fig43} (Color online) The Tc-like ions ($N=43$). Only the $Z=74$ point is distinguished.\cite{W-ions1}.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Mo}) &=& 0.0274790, \\
s &=& 0.233850.
\label{rest_tc}
\end{eqnarray}
The slope was computed from $E_a({\rm Mo})$ and $R_{cov}({\rm Mo})=2.759000$.
\textbf{Discussion:}
Notice that $I_p$ at $Z=74$ is systematically overestimated (since the Ca sequence). This time in 0.123 a.u.
\subsection{The Ru-like sequence ($N=44$)}
\label{Ru}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n44.pdf}
\caption{\label{fig44} (Color online) The Ru-like systems ($N=44$). An inconsistency at $Z=74$ is apparent.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.11747 -0.118053~ (Z/137.036)^2,\\
\nonumber a_0 &=& 19.1396,\\
\nonumber a_{-1} &=& 621.055.
\label{ru_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Tc}) &=& 0.0202049, \\
s &=& 0.232165.
\label{rest_ru}
\end{eqnarray}
The slope was computed from $E_a({\rm Tc})$ and $R_{cov}({\rm Tc})=2.607822$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is overestimated in 0.136 a.u.
\subsection{The Rh-like sequence ($N=45$)}
\label{Rh}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n45.pdf}
\caption{\label{fig45} (Color online) Rh-like systems ($N=45$). There are apparent inconsistencies at $Z=49~$\cite{Dirac-Fock1} and $Z=74~$\cite{W-ions1}.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.15904 -0.119845~ (Z/137.036)^2,\\
\nonumber a_0 &=& 20.2974,\\
\nonumber a_{-1} &=& 640.445.
\label{rh_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ru}) &=& 0.0385732, \\
s &=& 0.232165.
\label{rest_rh}
\end{eqnarray}
The slope was computed from $E_a({\rm Ru})$ and $R_{cov}({\rm Ru})=2.570028$.
\textbf{Discussion:}
$I_p$ at $Z=49$ is underestimated in 0.072 a.u., whereas the $Z=74$ point is 0.136 a.u. higher than the average curve.
\subsection{The Pd-like sequence ($N=46$)}
\label{Pd}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.19902 -0.121538~ (Z/137.036)^2,\\
\nonumber a_0 &=& 20.4131,\\
\nonumber a_{-1} &=& 703.998.
\label{pd_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n46.pdf}
\caption{\label{fig46} (Color online) The Pd-like sequence ($N=46$). Inconsistencies at $Z=55$ and 74 are noticed.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Rh}) &=& 0.0417692, \\
s &=& 0.258127.
\label{rest_pd}
\end{eqnarray}
The slope was computed from $E_a({\rm Rh})$ and $R_{cov}({\rm Rh})=2.532233$.
\textbf{Discussion:}
In the Pd-like sequence (Fig. \ref{fig46}), in addition to $Z=74$, an apparent inconsistency at $Z=55$ comes from the experimental work of Churilov et al.\cite{Churilov} The $Z=55$ and 74 points are, respectively, 0.120 and 0.118 a.u. above the average curve.
\subsection{The Ag-like sequence ($N=47$)}
\label{Ag}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.41502 -0.161518~ (Z/137.036)^2,\\
\nonumber a_0 &=& 35.8484,\\
\nonumber a_{-1} &=& 446.393.
\label{ag_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n47.pdf}
\caption{\label{fig47} (Color online) The Ag-like systems ($N=47$). The average curve is drawn (blue) in order to identify problematic points. Inconsistencies at $Z=61$ and 74 are apparent. The $Z=95$) point also distinghuishes in spite of its huge error bar.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Pd}) &=& 0.0206460, \\
s &=& 0.263779.
\label{rest_ag}
\end{eqnarray}
The slope was computed from $E_a({\rm Pd})$ and $R_{cov}({\rm Pd})=2.456644$.
\textbf{Discussion:}
A qualitatively different picture appears in the Ag-like sequence, Fig. \ref{fig47}. The maximum of $|{\rm RPT-NIST}|$ diminished, as compared with the Pd-like sequence, but the dispersion of points has significantly increased.
The problematic point at $Z=61$ comes from Ref. [\onlinecite{Dirac-Fock1}]. It seems to be 0.191 a.u. below the average curve. $I_p$ at $Z=74$, on the other hand, is underestimated in 0.275 a.u.
Data for $Z\ge 77$ is the entire responsibility of Carlson et al.\cite{Carlson}, who employ a simple spherical shell model in order to compute the ionization potentials. The indicated error bars are very high. The $Z=95$ point, with a deviation of -1.134 a.u. should, however, be noticed.
\subsection{The Cd-like sequence ($N=48$)}
\label{Cd}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.46174 -0.163993~ (Z/137.036)^2,\\
\nonumber a_0 &=& 37.7704,\\
\nonumber a_{-1} &=& 449.257.
\label{cd_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Ag}) &=& 0.0478309, \\
s &=& 0.265692.
\label{rest_cd}
\end{eqnarray}
The slope was computed from $E_a({\rm Ag})$ and $R_{cov}({\rm Ag})=2.570028$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n48.pdf}
\caption{\label{fig48} (Color online) The Cd-like systems ($N=48$). The only inconsistent point is $Z=74$.}
\end{figure}
\end{center}
\textbf{Discussion:}
$I_p$ at $Z=74$ is underestimated in 0.431 a.u.
\subsection{The Sn-like sequence ($N=50$)}
\label{Sn}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n50.pdf}
\caption{\label{fig50} (Color online) The Sn-like systems ($N=50$). A deviation at $Z=74~$ \cite{W-ions1} is apparent.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.56241 -0.169401~ (Z/137.036)^2,\\
\nonumber a_0 &=& 37.191,\\
\nonumber a_{-1} &=& 688.884.
\label{sn_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm In}) &=& 0.0110208, \\
s &=& 0.205786.
\label{rest_sn}
\end{eqnarray}
The slope was computed from $E_a({\rm In})$ and $R_{cov}({\rm In})=2.683411$.
\textbf{Discussion:}
$I_p$ at $Z=74$ shows a huge deviation of -0.720 a.u. We may notice also that, apparently, there are crossing points at $Z=60$ and 64. Thus the group of points in $Z=61-63$ should be below the $x$ axis. However, we do not have a precise way of estimating these potentials.
\subsection{The Sb-like sequence ($N=51$)}
\label{Sb}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n51.pdf}
\caption{\label{fig51} (Color online) The Sb-like sequence ($N=51$). The $Z=74~$\cite{W-ions1} point is distinguished.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.61122 -0.172008~ (Z/137.036)^2,\\
\nonumber a_0 &=& 40.6766,\\
\nonumber a_{-1} &=& 628.32.
\label{sb_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Sn}) &=& 0.0408532, \\
s &=& 0.254716.
\label{rest_sb}
\end{eqnarray}
The slope was computed from $E_a({\rm Sn})$ and $R_{cov}({\rm Sn})=2.645617$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is underestimated in -0.678 a.u. We have stressed previously the problem of $Z=61-63$ ions for Sn-like case. A similar comment applies here.
\subsection{The Te-like sequence ($N=52$)}
\label{Te}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n52.pdf}
\caption{\label{fig52} (Color online) The Te-like systems ($N=52$). A marked deviation at $Z=74~$\cite{W-ions1} is noticed.}
\end{figure}
\end{center}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.00113932~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.65793 -0.174483~ (Z/137.036)^2,\\
\nonumber a_0 &=& 41.4216,\\
\nonumber a_{-1} &=& 699.342.
\label{te_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Sb}) &=& 0.0384262, \\
s &=& 0.252618.
\label{rest_te}
\end{eqnarray}
The slope was computed from $E_a({\rm Sb})$ and $R_{cov}({\rm Sb})=2.645617$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is 0.668 a.u. below the average curve.
\subsection{The I-like sequence ($N=53$)}
\label{I}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.000488281 ~(Z/137.036)^2,\\
\nonumber a_1 &=& -2.71192 -0.133452 ~(Z/137.036)^2,\\
\nonumber a_0 &=& 44.5066,\\
\nonumber a_{-1} &=& 670.563.
\label{i_coeff}
\end{eqnarray}
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n53.pdf}
\caption{\label{fig53} (Color online) The case of I-like systems ($N=53$). An inconsistency is detected at $Z=74$.}
\end{figure}
\end{center}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm Te}) &=& 0.0724331,\\
s &=& 0.278176.
\label{rest_i}
\end{eqnarray}
The slope was computed from $E_a({\rm Te})$ and $R_{cov}({\rm Te})=2.588925$.
\textbf{Discussion:}
$I_p$ at $Z=74$ is 0.668 a.u. below the average curve. Notice that underestimation of this point is systematical since the Ag sequence.
\subsection{The Xe-like sequence ($N=54$)}
\label{Xe}
RPT coefficients:
\begin{eqnarray}
\nonumber a_2 &=& 0.03125 + 0.000488281~ (Z/137.036)^2,\\
\nonumber a_1 &=& -2.75916 -0.134776~ (Z/137.036)^2,\\
\nonumber a_0 &=& 46.3287,\\
\nonumber a_{-1} &=& 694.4.
\label{xe_coeff}
\end{eqnarray}
Conditions at $Z=N-1$:
\begin{eqnarray}
\nonumber E_a({\rm I}) &=& 0.112416, \\
s &=& 0.293706.
\label{rest_xe}
\end{eqnarray}
The slope was computed from $E_a({\rm I})$ and $R_{cov}({\rm I})=2.570028$.
\begin{center}
\begin{figure}[!ht]
\includegraphics[width=0.9\linewidth,angle=0]{n54.pdf}
\caption{\label{fig54} (Color online) The Xe-like systems ($N=54$). Only the $Z=74$ point shows a clear inconsistency.}
\end{figure}
\end{center}
\textbf{Discussion:}
$I_p$ at $Z=74$ is underestimated in 0.799 a.u.
|
\section{Introduction}
\setcounter{equation}{0}
\label{sec.intro}
The problem of determining best possible expectation bounds
on linear functions of order statistics in terms of means and variances of
the
observations has a long history. Especially
for the sample range based on $n\geq 2$ independent identically
distributed (i.i.d.) random variables, the problem goes back to
Plackett (1947), Gumbel (1954) and Hartley and David (1954)
who derived
the inequality
\be
\label{eq.plackett}
\mbox{$
\E\big[\max_{1\leq i \leq n}\{X_i\}-\min_{1\leq i \leq n}\{X_i\}\big]\leq
n\sigma \sqrt{\frac{2}{2n-1}\Big(1-\frac{1}{{{2n-2 \choose n-1}}}\Big)},
$}
\ee
where $\sigma^2$ is the common variance of $X_i$. This bound
is best possible in the sense that for any given values of
$\mu\in\R$ and $\sigma\in(0,\infty)$ there exist $n$ i.i.d.\ random variables
with mean $\mu$ and variance $\sigma^2$ that attain the equality in
(\ref{eq.plackett}).
Since then, a lot of research has been developed in order to
drop the assumptions of independence and/or
identical distributions on the observations, and
also to extend the results to any
$L$-statistic of the form
$
L=\sum_{i=1}^n c_i X_{i:n},
$
where $c_i$ are given constants and $X_{1:n}\leq \cdots\leq X_{n:n}$
are the order statistics corresponding to the random vector
$(X_1,\ldots,X_n)$.
When the components $X_i$ are merely assumed to be i.d.\
(identically distributed but not necessarily independent)
with mean $\mu$ and variance $\sigma^2$,
the best possible bounds for $\E L$
were established by Rychlik (1993b). In particular, setting
$c_1=-1$, $c_n=1$ and $c_i=0$ for any other $i$
in Rychlik's result,
we get the optimal upper bound
for the expected range:
\be
\label{eq.Rychlik}
\E\left[X_{n:n}-X_{1:n}\right]\leq
\sigma \sqrt{2n}.
\ee
For a comprehensive review of related
results and extensions, the reader is
referred to Rychlik's (2001)
monograph; see also
David (1981), Rychlik (1998) and David and Nagaraja (2003).
Dropping both assumptions of independence and i.d.,
Arnold and Groeneveld (1979)
obtained the upper bound
\be
\label{eq.AG.general}
\mbox{$
\E\left(\sum_{i=1}^n c_i X_{i:n}\right)
\leq \overline{\mu} \sum_{i=1}^n c_i
+\sqrt{\sum_{i=1}^n (c_i-\overline{c})^2}
\sqrt{\sum_{i=1}^n \{(\mu_i-\overline{\mu})^2+\sigma_i^2\}},
$}
\ee
which is valid for any random vector with $\E X_i=\mu_i$
and $\Var X_i=\sigma_i^2$,
where $\overline{\mu}=\frac{1}{n}\sum_{i=1}^n \mu_i$,
$\overline{c}=\frac{1}{n}\sum_{i=1}^n c_i$.
For other inequalities related to (\ref{eq.AG.general})
the reader is referred to Nagaraja (1981), Aven (1985),
Lef\`{e}vre (1986), Papadatos (2001a) and
Kaluszka, Okolewski and Szymanska (2005); see also
the monograph by Arnold and Balakrishnan (1989).
Applied to the range, (\ref{eq.AG.general}) yields
the inequality
\be
\label{eq.AG}
\mbox{$
\E[ X_{n:n}-X_{1:n} ]
\leq AG_n:=
\sqrt{2\sum_{i=1}^n \{(\mu_i-\overline{\mu})^2+\sigma_i^2\}},
$}
\ee
which, in the homogeneous case $\mu_i=\mu$, $\sigma_i^2=\sigma^2$,
reduces to (\ref{eq.Rychlik}).
However, the upper bound in (\ref{eq.AG})
is not tight under general mean-variance information, and the
purpose of the present work is to replace the RHS of
(\ref{eq.AG}) by its best possible value.
Recently, Bertsimas, Natarajan and Teo (2004, 2006) applied
convex optimization techniques in order to replace
the RHS of (\ref{eq.AG.general}) by its tight counterpart
in some particular cases of interest.
They obtained, among other things, the best possible upper bound for
the expected maximum under any mean-variance information and any
dependence structure, namely,
\be
\label{eq.BNT}
\mbox{$
\E X_{n:n} \leq BNT_n:=-\frac{n-2}{2}y_0
+\frac{1}{2}\sum_{i=1}^n \mu_i
+\frac{1}{2} \sum_{i=1}^n \sqrt{(\mu_i-y_0)^2+\sigma_i^2},
$}
\ee
where $y_0$ is the unique solution to the equation
\be
\label{eq.equation}
\mbox{$
\sum_{i=1}^n {
\frac{y_0-\mu_i}{\sqrt{(\mu_i-y_0)^2+\sigma_i^2}}}=n-2.
$}
\ee
The equality in (\ref{eq.BNT}) is attained by the extremely dependent
random vector with
\[
\Pr[X_1=y_0-\alpha_1,
\ldots,X_j=y_0+\alpha_j,\ldots,
X_n=y_0-\alpha_n]=p_j, \ \ j=1,\ldots,n,
\]
where
\[
\mbox{$
\alpha_j=\sqrt{(\mu_j-y_0)^2+\sigma_j^2},
\ \
p_j=\frac{1}{2}\Big(1-\frac{y_0-\mu_j}{\sqrt{(\mu_j-y_0)^2+\sigma_j^2}}\Big),
\ \ j=1,\ldots,n.
$}
\]
Note that $p_j>0$ and, by (\ref{eq.equation}), $\sum_{j=1}^n p_j=1$.
In the present work we extend the techniques
of Lai and Robbins (1976) and of
Bertsimas, Natarajan and Teo (2006), in order to obtain the best
possible upper bound for the expected range. Also, we characterize
the {\it extremal random vectors}, i.e.\
the vectors that attain the equality in the bound, and we
provide simple conditions (on $\mu_i$ and $\sigma_i$)
under which the $AG_n$ bound of (\ref{eq.AG}) is already sharp.
The main result is given in Theorem \ref{theo.6.1}.
Particular cases of interest are presented as examples.
\section{An upper bound for the expected range}
\label{sec.main.inequality}
\setcounter{equation}{0}
Let ${\bXX}=(X_1,\ldots,X_n)$
be an arbitrary random vector
with $\E {\bXX}=\bmu:=(\mu_1,\ldots,\mu_n)$
and
\\
$(\Var X_1,\ldots,\Var X_n)=(\sigma_1^2,\ldots,\sigma_n^2)$
where $0<\sigma_i<\infty$ for all $i$.
For notational simplicity
we write
${\bsigma}=(\sigma_1,\ldots,\sigma_n)$,
${\bsigma}^2=(\sigma_1^2,\ldots,\sigma_n^2)$
and $\Var{\bXX}={\bsigma}^2$; that is,
$\Var {\bXX}:=\mbox{diag}({\bSigma})$
where ${\bSigma}$
is the dispersion matrix of
${\bXX}$.
The class of random vectors satisfying the above moment requirements
will be denoted by
\be
\label{eq.F}
{\cal F}_n(\bmu,{\bsigma}):=\{{\bXX}:\E{\bXX}=\bmu,
\Var{\bXX}={\bsigma}^2\}.
\ee
In particular, $X\in{\cal F}_1(\mu,\sigma)$ means that $\E X=\mu$ and
$\Var X=\sigma^2$.
Let $X_{1:n}\leq \cdots \leq X_{n:n}$ be the order statistics
corresponding to ${\bXX}$ and set $R_n=X_{n:n}-X_{1:n}$ for the range.
Our main interest is in calculating
\be
\label{eq.problem}
\inf_{{\footnotesize{\bXX}}
\in{\cal F}_n({\footnotesize\bmu},
{\footnotesize{\bsigma}})} \E R_n,
\ \ \ \ \ \
\sup_{{\footnotesize{\bXX}}\in{\cal F}_n({\footnotesize\bmu},
{\footnotesize{\bsigma}})} \E R_n,
\ee
for any given $\bmu\in\R^n$ and ${\bsigma}\in\R_+^n$.
However, the result is known for the infimum:
\[
\inf_{{\footnotesize{\bXX}}
\in{\cal F}_n({\footnotesize\bmu},
{\footnotesize{\bsigma}})}
\E R_n
= \max_i \{\mu_i\} -\min_i \{\mu_i\}.
\]
Indeed, since $R_n=R_n({\bXX})$ is a convex function of ${\bXX}$ we have
$\E R_n({\bXX})\geq R_n(\bmu)=\max_i\{\mu_i\}-\min_i\{\mu_i\}$
from Jensen's inequality.
Bertsimas, Doan, Natarajan and Teo (2010) showed that
this lower bound is best possible even
for the narrowed class of random vectors with given mean vector
$\bmu$ and (any) given non-negative defined
dispersion matrix \mbox{${\bSigma}$\hspace{.3ex}.}
For
clarity of the presentation
we provide here the construction
of
Bertsimas, Doan, Natarajan and Teo (2010).
Define
\[
{\bXX}_{\epsilon}=\bmu+\frac{I_{\epsilon}}{\sqrt{\epsilon}}
\,
\bVV
\,
{\bSigma}^{1/2}, \ \ \
0<\epsilon<1,
\]
where
${\bVV}=(V_1,\ldots,V_n)$ with $V_i$ being i.i.d.\
with zero mean and variance one
and $I_{\epsilon}$ is a
Bernoulli random
variable, independent of ${\bVV}$, with probability of
success equal to $\epsilon$.
Then it is easy to verify that for all $\epsilon\in(0,1)$,
${\bXX}_\epsilon$ has mean $\bmu$
and dispersion matrix \mbox{${\bSigma}$\hspace{.3ex}.}
Let
$A\subseteq \R^n$ be the finite collection of
vectors of the form
$\bbee(i)-\bbee(j)$, $i\neq j$,
$i,j\in\{1,\ldots,n\}$, where ${\bbee}(i)=(0,\ldots,1,\ldots,0)$
is the unitary vector of the $i$-th axis.
With $\bx^{\mbox{t}}$ denoting the transpose
of any $1\times n$ random vector $\bx$ we have
\[
\mbox{$
R_n({\bXX}_{\epsilon})
=\max_{\bbalpha\in A}\{
\balpha{\bXX}_{\epsilon}^{\mbox{t}}\}
\leq
\max_{\bbalpha\in A}\{ \balpha\bmu^{\mbox{t}}\}
+ \frac{I_{\epsilon}}{\sqrt{\epsilon}}
\max_{\bbalpha\in A}\{ \balpha\ {\bSigma}^{1/2}{\bVV}^{\mbox{t}}\}.
$}
\]
Clearly, $\max_{\bbalpha\in A}\{ \balpha\bmu^{\mbox{t}}\}
=\max_i\{\mu_i\}-\min_i\{\mu_i\}$,
while
\[
\mbox{$
\E\Big(\frac{I_{\epsilon}}{\sqrt{\epsilon}}
\max_{\bbalpha\in A}\{ \balpha \,
{\bSigma}^{1/2}{\bVV}^{\mbox{t}}\}\Big)
=
\sqrt{\epsilon}\, \E\Big(\max_{\bbalpha\in A}\{
\balpha
\,
{\bSigma}^{1/2}{\bVV}^{\mbox{t}}\}\Big)
\leq
\sqrt{\epsilon}\sum_{\bbalpha\in A}
\E\Big|\balpha
\,
{\bSigma}^{1/2}{\bVV}^{\mbox{t}}\Big|
=\gamma\sqrt{\epsilon},
$}
\]
where $\gamma\geq 0$ is a finite constant independent of $\epsilon$.
It follows that
\[
\E R_{n}({\bXX}_\epsilon)\leq
\max_i\{\mu_i\}-\min_i\{\mu_i\}+\gamma\sqrt{\epsilon}
\]
and thus,
\[
\lim_{\epsilon\searrow 0}\E R_n({\bXX}_{\epsilon})
= \max_i\{\mu_i\}-\min_i\{\mu_i\}.
\]
Hence, the best possible lower bound for $\E R_n$
is $\max_i\{\mu_i\}-\min_i\{\mu_i\}$.
Regarding the supremum in
(\ref{eq.problem}), we shall make
use of the following definition.
\begin{DEFI}
\label{def.3.1}
A random vector ${\bXX}\in {\cal F}_n(\bmu,{\bsigma})$
of dimension $n\geq 2$
will be called {\it extremal random vector} (for the range)
if
$\E R_n({\bXX})=\sup \E R_n$, where the supremum is
taken over ${\cal F}_n(\bmu,{\bsigma})$.
The class of extremal random vectors is denoted
by ${\cal E}_n(\bmu,{\bsigma})$.
\end{DEFI}
To the best of our knowledge,
the value of the supremum and the nature of the
set ${\cal E}_n(\bmu,{\bsigma})$
have not been analysed elsewhere; it is not even
known whether ${\cal E}_n(\bmu,{\bsigma})$
in nonempty for general $\bmu$ and ${\bsigma}$.
In the present article we shall address both issues.
We start with a deterministic inequality
which is the range analogue of the inequality
given by Lai and Robbins (1976):
\begin{LEM}
\label{lem.1}
For any ${\bXX}\in \R^n$, $c\in\R$ and $\lambda>0$,
\be
\label{eq.deterministic}
\mbox{$
R_n\leq -(n-2)\lambda+\frac{\lambda}{2}\sum_{i=1}^n
\left\{\left|\frac{X_i-c}{\lambda}-1\right|
+\left|\frac{X_i-c}{\lambda}+1\right|\right\}.
$}
\ee
The equality in (\ref{eq.deterministic}) is attained
if and only if
\be
\label{eq.equality}
X_{1:n}\leq c-\lambda\leq X_{2:n}\leq \cdots \leq X_{n-1:n}
\leq c+\lambda\leq X_{n:n}.
\ee
\end{LEM}
\noindent
The Lemma entails that the use of two decision variables
is sufficient for properly handling $R_n$. Also, it
suggests the investigation of
$\sup\E\big\{|X-1|+|X+1|\big\}$ when $X$
is a random variable with given mean and variance:
\begin{LEM}
\label{lem.2}
For any $X\in{\cal F}_1(\mu,\sigma)$
($0<\sigma<\infty$),
\be
\label{eq.in}
\E\big\{|X-1|+|X+1|\big\}\leq U(\mu,\sigma),
\ee
where
\be
\label{eq.u}
U(\mu,\sigma):=\left\{
\begin{array}{lll}
2\sqrt{\mu^2+\sigma^2}, \medskip & \mbox{if}
& \mu^2+\sigma^2\geq 4,
\\
2+\frac{1}{2}(\mu^2+\sigma^2), \medskip
& \mbox{if} & 2|\mu|<\mu^2+\sigma^2< 4,
\\
|\mu|+1+\sqrt{(|\mu|-1)^2+\sigma^2},
& \mbox{if} & \mu^2+\sigma^2\leq 2|\mu|< 4.
\end{array}
\right.
\ee
The equality in (\ref{eq.in}) is attained by a unique random
variable $X^*\in{\cal F}_1(\mu,\sigma)$. Depending on
$(\mu,\sigma)$, $X^*$ assumes two or three
supporting values.
More
precisely:
\noindent
(a) For $\mu^2+\sigma^2\geq 4$,
\[
\mbox{$
\Pr\big[X^*=\sqrt{\mu^2+\sigma^2}\big]=\frac{1}{2}
\Big(1+\frac{\mu}{\sqrt{\mu^2+\sigma^2}}\Big)
=1- \Pr\big[X^*=-\sqrt{\mu^2+\sigma^2}\big].
$}
\]
\noindent
(b) For $2|\mu|<\mu^2+\sigma^2< 4$,
\[
\mbox{$
\Pr[X^*=0]=1-\frac{\mu^2+\sigma^2}{4}, \ \
\Pr[X^*=-2]=\frac{\mu^2+\sigma^2-2\mu}{8}, \ \
\Pr[X^*=2]=\frac{\mu^2+\sigma^2+2\mu}{8}.
$}
\]
\noindent
(c) For $\mu^2+\sigma^2\leq 2\mu$ (and hence, $0<\mu<2$),
\[
\mbox{$
\Pr\big[X^*=1+\sqrt{(\mu-1)^2+\sigma^2}\big]=\frac{1}{2}
\Big(1+\frac{\mu-1}{\sqrt{(\mu-1)^2+\sigma^2}}\Big)
=1- \Pr\big[X^*=1-\sqrt{(\mu-1)^2+\sigma^2}\big].
$}
\]
\noindent
(d) For $\mu^2+\sigma^2\leq -2\mu$ (and hence, $-2<\mu<0$),
\[
\mbox{$
\Pr\big[X^*=-1+\sqrt{(\mu+1)^2+\sigma^2}\big]=\frac{1}{2}
\Big(1+\frac{\mu+1}{\sqrt{(\mu+1)^2+\sigma^2}}\Big)
=1- \Pr\big[X^*=-1-\sqrt{(\mu+1)^2+\sigma^2}\big].
$}
\]
\end{LEM}
\begin{REM}
Isii (1963) presented general results
that include
inequalities of the form of Lemma \ref{lem.2}; see also
Karlin and Studden (1966).
The univariate mean-variance
inequality in Isii's paper can be stated as
follows: If $h:\R\to\R$ is a Borel function, $\mu\in\R$ and $\sigma>0$ then
\[
\sup_{X\in{\cal F}_1(\mu,\sigma)}
\E h(X) =
\inf_{\alpha_0,\alpha_1,\alpha_2}
\big\{\alpha_0+\alpha_1\mu+\alpha_2 (\mu^2+\sigma^2):\alpha_0+\alpha_1 x
+\alpha_2 x^2\geq h(x) \mbox{ for all } x\big\}.
\]
Isii showed that the above infimum is attained by some
${\balpha}^*=(\alpha_0^*,\alpha_1^*,\alpha_2^*)\in A$,
where
\[
A=\{(\alpha_0,\alpha_1,\alpha_2):\alpha_0+\alpha_1 x
+\alpha_2 x^2\geq h(x) \mbox{ for all } x\in\R\}\subseteq \R^3,
\]
provided that the infimum is finite.
However, usually it is not an easy task to specify the subset $A$
and the extremal point(s) ${\balpha}^*$. Lemma \ref{lem.2}
shows that this is possible for $h(x)=|x-1|+|x+1|$ and,
more importantly, characterizes the case of equality.
\end{REM}
\noindent
The following corollary is a straightforward consequence of
Lemma \ref{lem.2}.
\begin{COR}
\label{cor.1}
Let $X\in{\cal F}_1(\mu,\sigma)$
($0<\sigma<\infty$).
Fix $c\in\R$ and $\lambda>0$.
Then,
\be
\label{eq.in22}
\mbox{$
\E\big\{|(X-c)-\lambda|+|(X-c)+\lambda|\big\}\leq \lambda
\,
U\big(\frac{\mu-c}{\lambda},\frac{\sigma}{\lambda}\big),
$}
\ee
with
$U(\cdot,\cdot)$ given by (\ref{eq.u}).
The equality in (\ref{eq.in22}) is attained
by a unique two or three-valued random variable.
Setting
\[
\mbox{
$\xi=\mu-c$,
\ \
$\theta=\sqrt{(\mu-c)^2+\sigma^2}$,
\ \
$\alpha=\sqrt{(\xi-\lambda)^2+\sigma^2}$,
\ \
$\beta=\sqrt{(\xi+\lambda)^2+\sigma^2}$,
}
\]
the distribution that attains the equality is described
by the
following table:
\bigskip
\noindent
{\small
\begin{tabular}{l|ccc}
\relax
\begin{tabular}{l}
No: \ \ Condition on $\mu$, $\sigma$, $c$, $\lambda$
\\
Tight Upper Bound $\lambda \, U\big(\frac{\mu-c}{\lambda},
\frac{\sigma}{\lambda}\big)$
\end{tabular}
&
\begin{tabular}{c} value $x^{-}$ \\
probability $p^{-}$
\end{tabular}
&
\begin{tabular}{c} value $x^{o}$ \\
probability
$p^{o}$
\end{tabular}
&
\begin{tabular}{c} value $x^{+}$ \\
probability
$p^{+}$
\end{tabular}
\relax
\\
\hline
\hline
\relax
\begin{tabular}{c}
{\bf 1}: \ \ $(\mu-c)^2+\sigma^2\geq 4\lambda^2$
\vspace{.5ex}
\\
\hspace{5ex}
$2\sqrt{(\mu-c)^2+\sigma^2}$
\end{tabular}
&
\begin{tabular}{c}
$c-\theta$
\\
$\frac{1}{2}\big(1-\frac{\xi}{\theta}\big)$
\end{tabular}
&
\begin{tabular}{c}
\\
\end{tabular}
&
\begin{tabular}{c}
$c+\theta$
\\
$\frac{1}{2}\big(1+\frac{\xi}{\theta}\big)$
\end{tabular}
\vspace{1ex}
\\
\hline
\relax
\begin{tabular}{l}
{\bf 2}: \ \ $2\lambda|\mu-c|<(\mu-c)^2+\sigma^2< 4\lambda^2$
\vspace{.5ex}
\\
\hspace{5ex}
$2\lambda+\frac{1}{2\lambda}\left[(\mu-c)^2+\sigma^2\right]$
\end{tabular}
&
\begin{tabular}{c}
$c-2\lambda$
\\
$\frac{1}{8\lambda^2}\left[\theta^2-2\lambda\xi\right]$
\end{tabular}
&
\begin{tabular}{c}
$c$
\\
$1-\frac{1}{4\lambda^2}\theta^2$
\end{tabular}
\vspace{1ex}
&
\begin{tabular}{c}
$c+2\lambda$
\\
$\frac{1}{8\lambda^2}\left[\theta^2+2\lambda\xi\right]$
\end{tabular}
\\
\hline
\relax
\begin{tabular}{l}
{\bf 3}: \ \ $(\mu-c)^2+\sigma^2\leq2\lambda (\mu-c)$
\vspace{.5ex}
\\
\hspace{3ex}
$\mu-c+\lambda+\sqrt{(\mu-c-\lambda)^2+\sigma^2}$
\end{tabular}
&
\begin{tabular}{c}
\\
\end{tabular}
&\begin{tabular}{c}
$c+\lambda-\alpha$
\\
$\frac{1}{2}\big(1-\frac{\xi-\lambda}
{\alpha}\big)$
\end{tabular}
\vspace{1ex}
&
\begin{tabular}{c}
$c+\lambda+\alpha$
\\
$\frac{1}{2}\big(1+\frac{\xi-\lambda}
{\alpha}\big)$
\end{tabular}
\\
\hline
\relax
\begin{tabular}{l}
{\bf 4}: \ \ $(\mu-c)^2+\sigma^2\leq2\lambda (c-\mu)$
\vspace{.5ex}
\\
\hspace{3ex}
$c-\mu+\lambda+\sqrt{(c-\mu-\lambda)^2+\sigma^2}$
\end{tabular}
&
\begin{tabular}{c}
$c-\lambda-\beta$
\\
$\frac{1}{2}\big(1-\frac{\xi+\lambda}
{\beta}\big)$
\end{tabular}
&
\begin{tabular}{c}
$c-\lambda+\beta$
\\
$\frac{1}{2}\big(1+\frac{\xi+\lambda}
{\beta}\big)$
\end{tabular}
&
\begin{tabular}{c}
\\
\end{tabular}
\end{tabular}
}
\end{COR}
\noindent
\begin{Proof} Write $|(X-c)-\lambda|+|(X-c)+\lambda|
=\lambda\Big\{\big|\frac{X-c}{\lambda}-1\big|
+\big|\frac{X-c}{\lambda}+1\big|\Big\}$.
Since $Y=\frac{X-c}{\lambda}\in{\cal F}_1\big(\frac{\mu-c}{\lambda},
\frac{\sigma}{\lambda}\big)$,
Lemma \ref{lem.2}
yields (\ref{eq.in22})
as follows:
\[
\mbox{$
\E\big\{|(X-c)-\lambda|+|(X-c)+\lambda|\big\}=\lambda
\E\big\{|Y-1|+|Y+1|\big\}\leq
\lambda
\,
U\big(\frac{\mu-c}{\lambda},\frac{\sigma}{\lambda}\big).
$}
\]
Since $\lambda>0$, Lemma \ref{lem.2}
asserts that the equality is attained by a unique random variable
$Y^{*}\in{\cal F}_1\big(\frac{\mu-c}{\lambda}, \frac{\sigma}{\lambda}\big)$.
Thus, $X^*=c+\lambda Y^*$ is
the unique random variable in ${\cal F}_1(\mu,\sigma)$
that attains the equality in (\ref{eq.in22}). Substituting the
probability function
of $Y^*$ in the four distinct cases of Lemma \ref{lem.2} we obtain
the probabilities and supporting points as in the table.
$\Box$
\medskip
\end{Proof}
\noindent
It is important to observe that, whatever the values of
$\mu,\sigma,c,\lambda$ are,
the supporting points satisfy the relation
$x^{-}<c-\lambda<x^{o}<c+\lambda<x^{+}$.
We can now obtain the proposed upper bound for the expected range.
\begin{THEO}
\label{th.main}
If $\E {\bXX}=\bmu$ and $\Var {\bXX}={\bsigma}^2$ then
\be
\label{eq.main}
\E R_n\leq \inf_{c\in\R,\lambda>0}
\Big\{-(n-2)\lambda
+\frac{\lambda}{2}\sum_{i=1}^n
U\Big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\Big)\Big\},
\ee
where the function $U(\cdot,\cdot):\R\times(0,\infty)\to(2,\infty)$ is given by (\ref{eq.u}).
\end{THEO}
\noindent
\begin{Proof}
Fix $c\in\R$, $\lambda>0$. We take expectations in
(\ref{eq.deterministic}) and then use (\ref{eq.in22})
to get
\begin{eqnarray*}
\E R_n
\hspace{-1ex}
& \leq &
\hspace{-1ex}
\mbox{$
-(n-2)\lambda + \frac{1}{2}\sum_{i=1}^n
\E\big\{|(X_i-c)-\lambda|+|(X_i-c)+\lambda|\big\}
$}
\\
\hspace{-1ex}
&\leq &
\hspace{-1ex}
\mbox{$
-(n-2)\lambda + \frac{\lambda}{2}\sum_{i=1}^n
U\big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\big).
$}
\end{eqnarray*}
Since for all $c\in\R$
and $\lambda>0$ the last quantity is an upper bound for $\E R_n$,
its infimum is an upper bound too.
$\Box$
\medskip
\end{Proof}
\begin{REM}
\label{rem.2}
It is not clear at this stage whether
the upper bound (\ref{eq.main}) is tight, and it is not
an obvious task to find $c=c_0$ and $\lambda=\lambda_0$ (if exist) that
realize the infimum in the RHS of (\ref{eq.main}). However,
the substitution of any (convenient)
arguments $c$
and $\lambda$
in the function
\be
\label{eq.phi}
\mbox{$
\phi_n(c,\lambda):=-(n-2)\lambda + \frac{\lambda}{2}\sum_{i=1}^n
U\big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\big)
$}
\ee
will produce an upper bound for $\E R_n$. For example,
one can choose $c=\overline{\mu}$ and $\lambda=\frac{1}{4}AG_n$
(see (\ref{eq.AG})).
A simple way to produce a closed-form upper bound is the following:
First observe that
\[
\mbox{$
\lambda
\,
U\Big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\Big)
\leq 2\lambda+\frac{1}{2\lambda}\big[(\mu_i-c)^2+\sigma_i^2\big],
$}
\]
because the RHS is an upper bound for the expectation
$\E\big\{|(X_i-c)-\lambda|+|(X_i-c)+\lambda|\big\}$ (since
$|(X_i-c)-\lambda|+|(X_i-c)+\lambda|\leq 2\lambda+\frac{1}{2\lambda}(X_i-c)^2$
and $\E\big\{2\lambda+\frac{1}{2\lambda}(X_i-c)^2\big\}
=2\lambda+\frac{1}{2\lambda}[(\mu_i-c)^2+\sigma_i^2]$),
while the LHS
is the least upper bound for the same expectation as
$X_i$ varies in
${\cal F}_1(\mu_i,\sigma_i)$. It follows that
\[
\mbox{$
\phi_n(c,\lambda)\leq \overline{\phi}_n(c,\lambda)
:=2\lambda+\frac{1}{4\lambda}\sum_{i=1}^n\big\{(\mu_i-c)^2+\sigma_i^2\big\}.
$}
\]
Minimizing $\overline{\phi}_n(c,\lambda)$ is a simple fact: it suffices
to take $c=\overline{\mu}$ and $\lambda=\frac{1}{4}AG_n$ as before.
Observing that
$\sum_{i=i}^n\big\{(\mu_i-\overline{\mu})^2+\sigma_i^2\big\}=\frac{1}{2}AG_n^2$
we get
\[
\E R_n \leq
\inf_{c\in\R, \lambda>0}\phi_n(c,\lambda)\leq \inf_{c\in\R, \lambda>0}
\overline{\phi}_n(c,\lambda)=
\overline{\phi}_n\Big(\overline{\mu},\frac{1}{4}AG_n\Big)
=\frac{1}{2}AG_n+\frac{1}{2}AG_n=AG_n.
\]
Now it became clear that the bound in (\ref{eq.main}) is reasonable, since it
outperforms the bound in (\ref{eq.AG}) for any given values of $\bmu$ and
${\bsigma}$. As a result, the $AG_n$ bound need no be tight;
e.g., the infimum of $\phi_n(c,\lambda)$ need no be attained at
$(c,\lambda)=\big(\overline{\mu},\frac{1}{4}AG_n\big)$.
We shall prove in the sequel that the new bound is always tight,
and (for $n\geq 3$) the infimum in the RHS of (\ref{eq.main}) is attained
by a unique value $(c_0,\lambda_0)$.
\end{REM}
\begin{REM}
\label{rem.2.3}
Fixing $\mu$ in (\ref{eq.u}) and taking limits for $\sigma\searrow 0$ we see that
\[
\lim_{\sigma\searrow 0} U(\mu,\sigma)=2\max\{|\mu|,1\}
=|\mu-1|+|\mu+1|, \ \ \mu\in\R.
\]
Let us now set $\sigma_{n:n}=\max\{\sigma_1,\ldots,\sigma_n\}$
and fix $\bmu=(\mu_1,\ldots,\mu_n)$. Then,
\[
\lim_{\sigma_{n:n} \searrow 0} \phi_n(c,\lambda)
=-(n-2)\lambda +\frac{\lambda}{2}\sum_{i=1}^n
\Big\{\Big|\frac{\mu_i-c}{\lambda}-1\Big|
+\Big|\frac{\mu_i-c}{\lambda}+1\Big|\Big\}, \ \
\bmu\in\R^n, \ c\in\R, \ \lambda>0.
\]
Let $\mu_{1:n}\leq \cdots \leq \mu_{n:n}$ be the ordered values of
$\mu_1,\ldots,\mu_n$, and assume that
the $\mu$'s are not all equal, that is,
$\mu_{1:n}<\mu_{n:n}$.
Substituting in the above limit
$c=c_0=\frac{\mu_{1:n}+\mu_{n:n}}{2}$,
$\lambda=\lambda_0=\frac{\mu_{n:n}-\mu_{1:n}}{2}>0$,
we obtain
\[
\lim_{\sigma_{n:n} \searrow 0} \phi_n(c_0,\lambda_0)
=-(n-2)\lambda_0 +\frac{\lambda_0}{2}\sum_{i=1}^n
\Big\{\Big|\frac{\mu_i-c_0}{\lambda_0}-1\Big|
+\Big|\frac{\mu_i-c_0}{\lambda_0}+1\Big|\Big\}
=\mu_{n:n}-\mu_{1:n}.
\]
Note that the last equality follows from (\ref{eq.deterministic}) and
(\ref{eq.equality}), applied to ${\bXX}=\bmu$
(with $R_n(\bmu)=\mu_{n:n}-\mu_{1:n}$), observing that
for the particular choice of $(c_0,\lambda_0)$,
\[
\mu_{1:n}\leq c_0-\lambda_0\leq \mu_{2:n}\leq \cdots\leq
\mu_{n-1:n}\leq c_0+\lambda_0\leq \mu_{n:n}.
\]
For any ${\bXX}\in{\cal F}_n(\bmu,{\bsigma})$
it is true that
$\mu_{n:n}-\mu_{1:n}\leq \E R_n({\bXX})\leq
\inf_{c\in\R,\lambda>0}\big\{\phi_n(c,\lambda)\big\}$.
Therefore,
\begin{eqnarray*}
\mu_{n:n}-\mu_{1:n}
\leq
\lim_{\sigma_{n:n}\searrow 0} \E R_n({\bXX})
\leq
\lim_{\sigma_{n:n}\searrow 0}
\Big\{\inf_{c\in\R,\lambda>0}\phi_n(c,\lambda) \Big\}
\leq \lim_{\sigma_{n:n}\searrow 0} \phi_n(c_0,\lambda_0)
=\mu_{n:n}-\mu_{1:n},
\end{eqnarray*}
and we conclude that
\be
\label{eq.2.20}
\lim_{\sigma_{n:n}\searrow 0} \Big\{\inf_{c\in\R,\lambda>0}\phi_n(c,\lambda) \Big\}
=\mu_{n:n}-\mu_{1:n}.
\ee
The limit (\ref{eq.2.20}) continue to hold even if all
$\mu_i$'s are equal. Then
$\mu_{1:n}=\mu_{n:n}$ and
the inequality
$
\inf_{c\in\R,\lambda>0}\big\{\phi_n(c,\lambda)\big\}\leq AG_n$
(see Remark \ref{rem.2}) shows that
\[
\mbox{$
0
\leq
\inf_{c\in\R,\lambda>0}
\big\{\phi_n(c,\lambda)\big\}
\leq AG_n=\sqrt{2\sum_{i=1}^n\sigma_i^2}\leq
\sigma_{n:n}\sqrt{2n}\to 0, \ \ \mbox{ as }
\sigma_{n:n}\searrow 0.
$}
\]
From these considerations it is again clear that the $AG_n$
bound
is not tight in general; for example,
\[
\lim_{\sigma_{n:n}\searrow 0} AG_n=
\mbox{$
\sqrt{2\sum_{i=1}^n(\mu_i-\overline{\mu})^2} >\mu_{n:n}-\mu_{1:n}
$}
\]
whenever ($n\geq 3$ and) $\mu_{1:n}+\mu_{n:n}\neq 2\overline{\mu}$.
The $AG_n$ bound need no be tight even for equal $\mu_i$'s; see Theorem
\ref{theo.AG} and Example
\ref{ex.2}, below.
\end{REM}
\section{When is the Arnold-Groeneveld bound tight?}
\label{sec.AG}
\setcounter{equation}{0}
Arnold and Groeneveld (1979), Rychlik (1993b) and Papadatos (2001a)
showed that if $\mu_i=\mu$ and $\sigma_i=\sigma$
for all $i$, the $AG_n$ bound of
(\ref{eq.AG}), which reduces to (\ref{eq.Rychlik}), is
attainable.
In the present section we provide
an exact characterization of the attainability
of the $AG_n$ bound
under any mean-variance information.
The proof of Theorem \ref{theo.AG}, below, is based on the construction of particular
bivariate probability distributions supported in a subset of
$\{1,\ldots,n\}^2$.
A distribution of this kind corresponds to a
$n\times n$ matrix with nonnegative
elements having sum $1$; a probability matrix.
Matrices of this form with integer-valued entries
have been extensively studied; for a recent review
see Barvinok (2012).
The actual question, related to our problem, is
whether there exist probability
matrices with given marginals and vanishing trace.
The following notation and terminology will be used in the sequel.
\begin{DEFI}
\label{def.3.2}
A $n\times m$ matrix $Q=(q_{ij})$ ($n\geq 1$, $m\geq 1$)
is called {\it a probability matrix} if it has nonnegative elements
summing to $1$.
In particular, a $n$-variate {\it probability vector}
${\bpp}=(p_1,\ldots,p_n)$ is
a probability matrix with dimension $1\times n$, and
$X\sim {\bpp}$ is a convention for
$\Pr[X=i]=p_i$ for all $i$.
The {\it marginals} of $Q$, say ${\bpp}$, ${\bqq}$,
are the probability vectors obtained by summing
the rows and columns of $Q$, respectively; and
${\cal M}({\bpp},{\bqq})$ denotes the class of probability
matrices with given marginals ${\bpp}$, ${\bqq}$.
Moreover, $(X,Y)\sim Q$ is a convention for
$\Pr[X=i,Y=j]=q_{ij}$ for all $i,j$.
\end{DEFI}
We now state a characterization for the $AG_n$ bound.
\begin{THEO}
\label{theo.AG}
Assume that $\E {\bXX}=\bmu$ and $\Var{\bXX}={\bsigma}^2$.
Then the equality in (\ref{eq.AG}) is attainable if and only if
both conditions (i) and (ii) below are satisfied.
\be
\label{eq.3.1}
\begin{array}{ll}
\mbox{(i)}
\hspace*{3ex}
&
\displaystyle
|\mu_i-\overline{\mu}|
\leq
\displaystyle
\frac{\sqrt{2}\big[(\mu_i-\overline{\mu})^2+\sigma_i^2\big]}
{\sqrt{\sum_{j=1}^n
\big\{(\mu_j-\overline{\mu})^2+\sigma_j^2\big\}}},
\vspace{1ex}
\\
\mbox{(ii)} \hspace*{3ex}
&
\displaystyle
\frac{(\mu_i-\overline{\mu})^2+\sigma_i^2}{\sum_{j=1}^n
\big\{(\mu_j-\overline{\mu})^2+\sigma_j^2\big\}}
\leq
\frac{1}{2},
\end{array}
\hspace{3ex}
i=1,\ldots,n.
\ee
Provided that (i) and (ii) are fulfilled,
any extremal random vector
${\bXX}\in{\cal E}_n(\bmu,{\bsigma})$
has the representation
\be
\label{eq.extremal}
\mbox{$
{\bXX}=g(X,Y):=\overline{{\mu}}\, \bone
+{\displaystyle\frac{{\bbee}(X)-{\bbee}(Y)}{\sqrt{2}}}
\sqrt{\sum_{j=1}^n
\big\{(\mu_j-\overline{\mu})^2+\sigma_j^2\big\}},
$}
\ee
where ${\bone}=(1,\ldots,1)\in\R^n$,
${\bbee}(i)=(0,\ldots,1,\ldots,0)$,
and $(X,Y)$ is a discrete random pair
satisfying $\Pr[X=Y]=0$, with
marginal distributions
\be
\label{eq.marginals}
\begin{array}{l}
\displaystyle
p_i^{+}
=\Pr[X=i]
=
\frac{(\mu_i-\overline{\mu})^2
+\sigma_i^2+\frac{1}{2}(\mu_i-\overline{\mu})AG_n}{\sum_{j=1}^n
\big\{(\mu_j-\overline{\mu})^2+\sigma_j^2\big\}},
\vspace{1ex}
\\
\displaystyle
p_i^{-}=
\Pr[Y=i]
=
\frac{(\mu_i-\overline{\mu})^2
+\sigma_i^2-\frac{1}{2}(\mu_i-\overline{\mu})AG_n}{\sum_{j=1}^n
\big\{(\mu_j-\overline{\mu})^2+\sigma_j^2\big\}},
\end{array}
\hspace{3ex} i=1,\ldots,n.
\ee
Moreover, if the inequalities in (\ref{eq.3.1}) are strict for all
$i$, we can find infinitely many extremal
random vectors; and if (\ref{eq.3.1}) is satisfied
and for some
$i$
we have equality in (ii), then the extremal random vector
is unique.
\end{THEO}
\begin{REM}
\label{rem.3.1}
Let $(\mu_1,\mu_2,\mu_3)=(-1,0,1)$,
$(\sigma_1^2,\sigma_2^2,\sigma_3^2)=(1,3,2)$, so that
(\ref{eq.3.1}) holds. However,
(\ref{eq.3.1})(ii) is satisfied with strict
inequalities for all $i$, while this is not true for
(\ref{eq.3.1})(i).
We find $AG_3=4$ and
${\bpp}^+=\big(0,\frac{3}{8}, \frac{5}{8}\big)$,
${\bpp}^-=\big(\frac{1}{2},\frac{3}{8}, \frac{1}{8}\big)$.
It is easily seen that the distribution of $(X_1,X_2,X_3)$
(given in (\ref{eq.extremal}))
is uniquely defined: it assigns probabilities $\frac{2}{8},
\frac{2}{8},\frac{3}{8},\frac{1}{8}$, to the points
$(-2,2,0)$, $(-2,0,2)$, $(0,-2,2)$, $(0,2,-2)$, respectively.
It follows that a random vector that attains
the $AG_n$ bound
can be unique even if (\ref{eq.3.1})(ii) is satisfied
with strict
inequalities
for all $i$.
\end{REM}
\begin{EXAMPLE}
\label{ex.1} {\it The homogeneous case $\mu_i=\mu$, $\sigma_i=\sigma>0$.} Conditions
(\ref{eq.3.1}) are obviously satisfied with strict inequalities
(for $n\geq 3$) and the $AG_n$ bound is sharp (see also (\ref{eq.Rychlik})):
\[
\sup \E R_n=AG_n=\sigma\sqrt{2 n}.
\]
Moreover, $p_i^+=p_i^-=\frac{1}{n}$ and
from Theorem \ref{theo.AG} we see that
infinitely many random vectors
attain the equality. The totality of them is characterized
by (\ref{eq.extremal}) via the probability matrices $Q$ of $(X,Y)$.
Recall that $X$ and $Y$ are, respectively, the positions where
$\mu+\sigma\sqrt{n/2}$ and $\mu-\sigma\sqrt{n/2}$ appears in
the extremal vector $(X_1,\ldots,X_n)$; the rest entries are equal to
$\mu$. Thus,
$Q$ has uniform marginals and vanishing principal diagonal.
A famous theorem of Birkhoff on {\it magic matrices}
asserts that any matrix with nonnegative elements
having row/column sums equal to $1$
is a convex combination of
{\it permutation matrices},
i.e., matrices with entries $0$ or $1$, having exactly one $1$
in each row and in each column; see Theorem 2.54 in
Giaquinta and Modica (2012).
From Birkhoff's result it is evident that the probability matrix $Q$ of $(X,Y)$,
corresponding to any extremal random vector
${\bXX}=\mu {\bone}+\sigma[{\bbee}(X)-{\bbee}(Y)]\sqrt{n/2}$,
can be written as
\[
\mbox{$
Q=\sum_{i=1}^k{\lambda_i}D_i, \ \ \lambda_i\geq0, \ \ \sum_{i=1}^k {\lambda_i}=\frac1n,
$}
\]
where the $D_i$'s are {\it derangement matrices}, i.e.\ permutation matrices with
vanishing diagonal entries. It is well-known that there exist
$n!\sum_{k=0}^{n}\frac{(-1)^k}{k!}\approx
e^{-1}n!$ different derangement matrices; they coincide with
the extremal points of the convex polytope
$\big\{D=(d_{ij}):\sum_i d_{ij}=\sum_j d_{ij}=1,
\ d_{ij}\geq 0, \ d_{ii}=0
\mbox{ for all }i,j\big\}$.
In general, a convex polytope has a finite (often quite large) number of
extremal points, but
it is rather difficult to evaluate them exactly,
since their total number depends on the marginals in
an ambiguous way (cf.\ Example \ref{ex.2}, below).
\end{EXAMPLE}
\begin{EXAMPLE}
\label{ex.2} {\it The case $\mu_i=\mu$.} Assume $0<\sigma_1\leq\cdots\leq \sigma_n$
without loss of generality. From Theorem \ref{theo.AG} we
see that if the larger variance does not dominate the sum of the
other variances then the $AG_n$ bound is tight:
\[
\sup \E R_n=AG_n=\sqrt{2\mbox{$\sum_{i=1}^n\sigma_i^2$}}, \ \
\mbox{ whenever } \ \ \mbox{$\sigma_n^2\leq \sum_{i=1}^{n-1}\sigma_i^2$}.
\]
Moreover, if $\sigma_n^2= \sum_{i=1}^{n-1}\sigma_i^2$, the
equality is uniquely attained by the random vector
${\bXX}$
taking values
\begin{eqnarray*}
\mbox{$
{\bxx}_i=\big(\mu,\ldots,\mu,\mu+ \frac{AG_n}{2},\mu,\ldots,\mu;\mu-\frac{AG_n}{2}\big)
$},
& \mbox{ with probability }p_i,
\\
\mbox{$
{\byy}_i=\big(\mu,\ldots,\mu,\mu- \frac{AG_n}{2},\mu,\ldots,\mu;\mu+\frac{AG_n}{2}\big)
$},
& \mbox{ with probability }p_i,
\end{eqnarray*}
where $p_i=\frac{\sigma_i^2}{\sum_{j=1}^n\sigma_j^2}$, $i=1, \ldots, n-1$.
Of course, if
$\sigma_n^2< \sum_{i=1}^{n-1}\sigma_i^2$ then there exist
infinitely many extremal random vectors. They have the form $g(X,Y)$
(see (\ref{eq.extremal})), with $\Pr[X=Y]=0$, $X\sim{\bpp}$, $Y\sim{\bpp}$,
where ${\bpp}=(p_1,\ldots,p_{n-1}, p_n)$.
However, if
$\sigma_n^2>\sum_{i=1}^{n-1}\sigma_i^2$ then the $AG_n$
is no longer tight: The infimum
in (\ref{eq.main}) is attained at $c_0=\mu$,
$\lambda_0=\frac{1}{2}\sqrt{\sum_{i=1}^{n-1}\sigma_i^2}<\frac14 AG_n$,
and we get the inequality
\[
\mbox{$
\E R_n \leq \phi_n(c_0,\lambda_0)
=\sigma_n+\sqrt{\sum_{i=1}^{n-1}\sigma_i^2}
\ \ \ \ \ \
\big(\sigma_n^2>\sum_{i=1}^{n-1}\sigma_i^2\big).
$}
\]
From $\sqrt{x}+\sqrt{y}<\sqrt{2(x+y)}$ for
$x\neq y$
we conclude that this bound
is strictly better than $AG_n$. Moreover,
the new bound is tight; one can verify
that the equality is (uniquely) attained by the random vector
${\bXX}$
taking values
\begin{eqnarray*}
\mbox{$
{\bxx}_i=\big(\mu,\ldots,\mu,\mu+ \sqrt{\sum_{i=1}^{n-1}\sigma_i^2},\mu,\ldots,\mu;
\mu-\sigma_n\big)
$},
& \mbox{ with probability }p_i,
\\
\mbox{$
{\byy}_i=\big(\mu,\ldots,\mu,\mu-
\sqrt{\sum_{i=1}^{n-1}\sigma_i^2},\mu,\ldots,\mu;
\mu+\sigma_n\big)
$},
& \mbox{ with probability }p_i,
\end{eqnarray*}
where $p_i=\frac{\sigma_i^2}{2\sum_{j=1}^{n-1}\sigma_j^2}$, $i=1, \ldots, n-1$.
Thus, the tight upper bound on the expected
range from dependent observations with equal means
admits a simple closed form:
\be
\label{eq.tight}
\sup \E R_n=\left\{
\begin{array}{lll}
\vspace{.5ex}
\sqrt{2\sum_{i=1}^n \sigma_i^2}, & \mbox{if} & 2\max_i\{\sigma_i^2\}
\leq \sum_{i=1}^n\sigma_i^2,
\\
\max_i\{\sigma_i\} +\sqrt{\sum_{i=1}^{n}\sigma_i^2-\max_i \{\sigma_i^2\}},
& \mbox{if} & 2\max_i\{\sigma_i^2\}
\geq \sum_{i=1}^n\sigma_i^2.
\end{array}
\right.
\ee
Assuming that one variance tends to infinity (and keeping all
other variances bounded),
the limit $\lim_{\sigma_i\to\infty}\frac{\sup \E R_n}{AG_n}
=\frac{1}{\sqrt{2}}\approx .707$ says that we
can gain of an up to $30\%$ improvement over the $AG_n$ bound.
\end{EXAMPLE}
The following lemma will play an important role
in verifying existence of extremal random vectors.
\begin{LEM}
\label{lem.3}
Let ${\bpp}=(p_1,\ldots,p_n)$ and ${\bqq}=(q_1,\ldots,q_n)$
be two probability vectors.
A necessary and sufficient condition for the existence of
a random pair $(X,Y)$ with
\be
\label{eq.bivariate}
\Pr[X=Y]=0, \ X\sim{\bpp}, \ Y\sim{\bqq}
\ee
is the following:
\be
\label{eq.condition}
\max_{1\leq i\leq n}\big\{p_i+q_i\big\}\leq 1.
\ee
If the equality holds in (\ref{eq.condition}), the random pair
$(X,Y)$ is uniquely defined. If strict inequality holds in
(\ref{eq.condition}) and, furthermore,
$\min_{i}\{p_i\}>0$, $\min_i\{q_i\}>0$, then
there exist infinitely many random pairs
satisfying (\ref{eq.bivariate}).
\end{LEM}
\section{Convexity}
\label{sec.4}
\setcounter{equation}{0}
The purpose of the present section
is to verify that for any given values of
$\bmu$, ${\bsigma}$,
the function $\phi_n(c,\lambda)$ of
(\ref{eq.phi}) is convex.
For convenience we set $T:=\R\times(0,\infty)$
for the domain of both functions $U$ (of (\ref{eq.u}))
and $\phi_n$.
We begin with a simple lemma.
\begin{LEM}
\label{lem.4.1}
The function $U(x,y):T\to(2,\infty)$ of
(\ref{eq.u})
has continuous partial derivatives, that is,
$U\in C^1(T)$.
\end{LEM}
\noindent
We also need another simple lemma;
see, e.g., Giaquinta and Modica (2012).
\begin{LEM}
\label{lem.4.2}
Let $K$ be a convex subset of $\R^n$ and $f:K\to \R$.
For ${\bxx}$ and ${\byy}$ in $K$
consider the function $g:[0,1]\to \R$
given by
\[
g(t):=f\big({\bxx}+t({\byy}-{\bxx})\big), \ \ 0\leq t\leq 1.
\]
Then, $f$ is convex if and only if
$g$ is convex for any choice of
${\bxx}$ and ${\byy}$ in $K$.
\end{LEM}
\noindent
Also, we shall make use of the following lemma.
\begin{LEM}
\label{lem.4.3}
Consider a finite interval $[\alpha,\beta]$, a partition
\[
\alpha=t_0<t_1<\cdots<t_k<t_{k+1}=\beta
\]
and the convex functions
$g_i:[\alpha,\beta]\to\R$ ($i=1,\ldots,k+1$).
Assume that
\be
\label{eq.4.3}
g_i(t_i)=g_{i+1}(t_i) \ \mbox{ and } \
g_i'(t_i-)\leq g'_{i+1}(t_i+), \ \ i=1,\ldots,k,
\ee
where $g'(t-)$ and $g'(t+)$ denote, respectively,
the left and right hand side derivatives of $g$ at $t$.
Then, the function
\be
\label{eq.4.4}
g(t):=\left\{
\begin{array}{cl}
g_1(t),
&
\alpha\leq t\leq t_1,
\\
g_2(t),
&
t_1\leq t\leq t_2,
\\
\vdots
\\
g_k(t),
&
t_{k-1}\leq t\leq t_k,
\\
g_{k+1}(t),
&
t_{k}\leq t\leq \beta,
\end{array}
\right.
\ee
is convex.
\end{LEM}
\noindent
\begin{Proof} Since all $g_i$ have non-decreasing left and right hand side derivatives,
it is easily seen that
the same is true for $g$.
$\Box$
\medskip
\end{Proof}
\noindent
Now we can verify the following result.
\begin{PROP}
\label{prop.4.1}
The function $U:T\to(2,\infty)$ in (\ref{eq.u}) is convex.
\end{PROP}
\noindent
Finally, we shall make use of the following property,
which seems to be of some independent interest.
\begin{LEM}
\label{lem.4.4}
Let $f(x,y):T\to\R$ and for fixed $x_0\in\R$, $y_0>0$, consider the function
$h(c,\lambda):T\to\R$ with
\[
\mbox{$
h(c,\lambda):=\lambda\, f\big(\frac{x_0-c}{\lambda},\frac{y_0}{\lambda}\big),
\ \ \ (c,\lambda)\in \R\times(0,\infty).
$}
\]
(i) If $f$ is convex then $h$ is convex for all choices of $x_0\in\R$, $y_0>0$.
\\
(ii) If $h$ is convex for a particular choice of $x_0\in\R$, $y_0>0$, then $f$ is
convex.
\end{LEM}
\noindent
We can now state and prove the final conclusion of the present section:
\begin{THEO}
\label{th.4.1}
For any given $\bmu$ and ${\bsigma}$,
the function $\phi_n(c,\lambda)$ in (\ref{eq.phi}) is convex
and belongs to $C^1(T)$, $T=\R\times (0,\infty)$.
\end{THEO}
\noindent
\begin{Proof} The fact that $\phi_n\in
C^1(T)$ follows by an obvious application of Lemma \ref{lem.4.1}.
Also, the function $U(x,y)$ in (\ref{eq.u})
is convex by
Proposition \ref{prop.4.1}. Hence, by Lemma \ref{lem.4.4},
the same is true for the
function
$h_i(c,\lambda)=\frac{1}{2}\lambda\, U\big(\frac{\mu_i-c}{\lambda},
\frac{\sigma_i}{\lambda}\big)$ ($i=1,\ldots,n)$. Since
$h(c,\lambda)=-(n-2)\lambda$ is trivially convex,
$\phi_n(c,\lambda)$ is a sum of convex functions.
$\Box$
\end{Proof}
\section{Attainability of the infimum in (\ref{eq.main}) at a unique point}
\label{sec.5}
\setcounter{equation}{0}
From now on we assume that $n\geq 3$. The simple
(but interesting)
case $n=2$
is deferred to the last section, noting that
the optimal upper bound for
$\E R_2$ is closely related to
the bound $BNT_2$ of (\ref{eq.BNT}).
In the present section we shall prove that
the minimum value of
$\phi_n(c,\lambda)$ is achieved
at a unique
point $(c_0,\lambda_0)\in T$.
Of course, since $\phi_n$ is differentiable,
a minimizing point
(if exists) has to satisfy the system
of equations
\be
\label{eq.5.1}
\frac{\partial}{\partial c}\phi_n(c,\lambda)=0,
\ \ \ \
\frac{\partial}{\partial \lambda}\phi_n(c,\lambda)=0.
\ee
However, due to the complicated form of the derivatives
(see (\ref{eq.4.1}), (\ref{eq.4.2})),
it is not a trivial fact to solve $(\ref{eq.5.1})$, or even
to verify its consistency analytically. On the other hand,
as we shall see in the sequel,
it is important to know the existence (and uniqueness) of a
minimizing point; it will be used
in an essential way in the construction of
extremal random vectors, concluding
tightness of the bound (\ref{eq.main}).
The attainability of the infimum can be seen as follows:
\noindent
Set $\epsilon_0:=\frac{1}{4}\min_i\{\sigma_i\}>0$.
For $c\in\R$ and $\lambda\in (0,\epsilon_0]$,
$(\mu_i-c)^2+\sigma_i^2\geq 4\lambda^2$
($i=1,\ldots,n$).
Thus,
$\lambda\, U(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda})
=2\sqrt{(\mu_i-c)^2+\sigma_i^2}$
for all $i$, and
\begin{eqnarray*}
&&
\hspace{-15ex}
\mbox{$
\phi_n(c,\lambda)=-(n-2)\lambda
+\sum_{i=1}^n \sqrt{(\mu_i-c)^2+\sigma_i^2}
$}
\\
&&
\mbox{$
\geq -(n-2)\epsilon_0+\sum_{i=1}^n \sqrt{(\mu_i-c)^2+\sigma_i^2}
=\phi_n(c,\epsilon_0).
$}
\end{eqnarray*}
The function $c\mapsto \sum_{i=1}^n \sqrt{(\mu_i-c)^2+\sigma_i^2}$
is strictly convex, tending to $\infty$ as $|c|\to\infty$;
thus, its minimum is attained at a unique $c=c_1$.
From $\phi_n(c,\epsilon_0)\geq \phi_n(c_1,\epsilon_0)$ we get
\[
\phi_n(c,\lambda)
\geq
\phi_n(c_1,\epsilon_0)
=
-(n-2)\epsilon_0 +\sum_{i=1}^n \sqrt{(\mu_i-c_1)^2+\sigma_i^2},
\ \
c\in\R, \ 0<\lambda\leq \epsilon_0.
\]
We now chose $\lambda_1:=\frac{1}{2}\min_i\{\sigma_i\}$,
so that
$\lambda_1>\epsilon_0$ and
$(\mu_i-c_1)^2+\sigma_i^2\geq 4\lambda_1^2$
for all $i$.
Therefore,
$\lambda_1\, U(\frac{\mu_i-c_1}{\lambda_1},\frac{\sigma_i}{\lambda_1})
=2\sqrt{(\mu_i-c_1)^2+\sigma_i^2}$
($i=1,\ldots,n$), and it follows that
$\phi_n(c_1,\lambda_1)=-(n-2)\lambda_1+
\sum_{i=1}^n \sqrt{(\mu_i-c_1)^2+\sigma_i^2}$.
Since $\lambda_1>\epsilon_0$ and $n\geq 3$,
the inequality $-(n-2)\epsilon_0>-(n-2)\lambda_1$
leads to $\phi_n(c_1,\epsilon_0)>\phi_n(c_1,\lambda_1)$.
Moreover,
$U(x,y)\geq U(0,y)=2+\int_{0}^y \min\{t,2\} dt >2$
for all $x\in\R$ and $y>0$.
We thus obtain
$
\phi_n(c,\lambda)=-(n-2)\lambda+\frac{\lambda}{2}\sum_{i=1}^n
U(\frac{\mu_i-c}{\lambda}
,\frac{\sigma_i}{\lambda})>-(n-2)\lambda+n\lambda=2\lambda
$
for all $c$ and $\lambda>0$.
Setting $M_0:=\frac{1}{2}\phi_n(c_1,\epsilon_0)>\epsilon_0$
we see that
\[
\phi_n (c,\lambda)\geq \phi_n(c_1,\epsilon_0)>\phi_n(c_1,\lambda_1)
\
\mbox{ for all } c\in\R, \ \lambda\in(0,\epsilon_0]\cup[M_0,\infty).
\]
Assume now that $\lambda\in(\epsilon_0,M_0)$ with $\epsilon_0$, $M_0$
as above. From the obvious inequality $U(x,y)\geq 2\max\{|x|,1\}\geq 2|x|$
we get
\[
\mbox{$
\phi_n(c,\lambda)
\geq
-(n-2)\lambda
+\sum_{i=1}^n |\mu_i-c|
\geq
-(n-2) M_0 +\sum_{i=1}^n |\mu_i-c|.
$}
\]
The last inequality shows that $\phi_n(c,\lambda)\to \infty$ as $|c|\to\infty$,
uniformly in $\lambda\in(\epsilon_0,M_0)$;
thus, we can find a constant $C_0$ such that
\[
\phi_n(c,\lambda)\geq \phi_n(c_1,\epsilon_0)
\
\mbox{ for all } |c|\geq C_0, \ \lambda\in(\epsilon_0,M_0).
\]
Since $\phi_n(c_1,\epsilon_0)>\phi_n(c_1,\lambda_1)$,
we arrived at the conclusion
\[
\phi_n(c,\lambda)
>\phi_n(c_1,\lambda_1)
\
\mbox{ for all } (c,\lambda)
\mbox { with }
|c|\geq C_0 \mbox{ or } \lambda\leq \epsilon_0
\mbox{ or } \lambda\geq M_0.
\]
This inequality shows that any minimizing point $(c_0,\lambda_0)$
of (the continuous function)
$\phi_n(c,\lambda)$
over the compact rectangle
$R:=[-C_0,C_0]\times[\epsilon_0,M_0]$
must lie in the interior of $R$.
The convexity of $\phi_n$
implies that
its global
minimum
is attained at $(c_0,\lambda_0)$.
On the other hand, the differentiability of $\phi_n$
shows that
$(c,\lambda)=(c_0,\lambda_0)$ is a solution to (\ref{eq.5.1});
and the convexity of $\phi_n$ implies that any such solution
is a minimizing point.
Let us now define
\be
\label{eq.5.2}
T_0:=\big\{(c,\lambda)\in T: (c,\lambda) \mbox{ is a solution to }
\mbox{(\ref{eq.5.1})}
\big\},
\ee
so that $T_0\neq \emptyset$.
The minimizing points of the convex function $\phi_n$
are exactly the points
of $T_0$; thus, $T_0$ is a convex compact
subset of $T$, and we have shown the following
\begin{PROP}
\label{prop.5.1}
If $n\geq 3$ then for any given values of $\bmu$ and ${\bsigma}$,
the system (\ref{eq.5.1}) is consistent,
and the set of solutions, $T_0$,
is a convex compact subset of $T$.
Moreover, for any $(c_0,\lambda_0)\in T_0$,
\[
\phi_n(c,\lambda)\geq \phi_n(c_0,\lambda_0)
\ \
\mbox{ for all } (c,\lambda)\in T=\R\times(0,\infty),
\]
with equality if and only if
$(c,\lambda)\in T_0$.
\end{PROP}
We now proceed to show that $T_0$ is a singleton.
Let as fix $c=c_1\in\R$.
For this particular value $c_1$ we consider the function
\[
\mbox{$
\psi_n(\lambda):=\phi_n(c_1,\lambda)=-(n-2)\lambda
+\sum_{i=1}^{n} u_i(\lambda), \ \ \lambda>0,
$}
\]
where
\[
\mbox{$
u_i(\lambda):=\frac{1}{2}\lambda\,
U(\frac{\mu_i-c_1}{\lambda},\frac{\sigma_i}{\lambda}), \ \
\lambda >0 \ \ (i=1,\ldots,n).
$}
\]
The function $u_i$ can be written more precisely as follows:
\[
u_i(\lambda)=\left\{
\begin{array}{cl}
\sqrt{(\mu_i-c_1)^2+\sigma_i^2}, & 0<\lambda\leq t_i,
\\
\lambda+\frac{1}{4\lambda}
\big[(\mu_i-c_1)^2+\sigma_i^2\big], & t_i\leq \lambda< \gamma_i,
\\
\frac{1}{2}\Big\{|\mu_i-c_1|+\lambda+
\sqrt{(|\mu_i-c_1|-\lambda)^2+\sigma_i^2}\Big\}, & \lambda\geq
\gamma_i,
\end{array}
\right.
\]
where
$t_i=t_i(c_1)$ and $\gamma_i=\gamma_i(c_1)$ are given by
\be
\label{eq.5.3}
\mbox{$
t_i:=\frac{1}{2}\sqrt{(\mu_i-c_1)^2+\sigma_i^2},
\
\gamma_i:=
\frac{(\mu_i-c_1)^2+\sigma_i^2}{2|\mu_i-c_1|},
\
\
0<t_i<\gamma_i \leq \infty
\ \ (i=1,\ldots,n).
$}
\ee
Each function $u_i$ is continuously
differentiable with derivative
\be
\label{eq.5.4}
u'_i(\lambda)=\left\{
\begin{array}{cl}
0, & 0<\lambda\leq t_i,
\\
1-\frac{(\mu_i-c_1)^2+\sigma_i^2}{4\lambda^2},
& t_i\leq \lambda< \gamma_i,
\\
\frac{1}{2}\Big(1+\frac{\lambda-|\mu_i-c_1|}
{\sqrt{(\lambda-|\mu_i-c_1|)^2+\sigma_i^2}}\Big),
& \lambda\geq
\gamma_i,
\end{array}
\right.
\ \
i=1,\ldots,n.
\ee
Obviously, $u_i(\lambda)$ is constant (equal to $2 t_i$)
in the interval $(0,t_i]$ and then it is strictly increasing;
its non-decreasing continuous derivative $u_i'(\lambda)$
satisfies $0\leq u_i'(\lambda)<1$ for all $\lambda$, and
$\lim_{\lambda\to \infty} u_i'(\lambda)=1$.
It follows that
\[
\mbox{$
\psi_n'(\lambda)=-(n-2)+\sum_{i=1}^n u_i'(\lambda)
$}
\]
is non-decreasing and, thus, $\psi_n$ is convex.
Let $t_{1:n}, \ldots, t_{n:n}$ be the ordered
values of $t_1,\ldots,t_n$. Noting that $n\geq 3$ and
$0<t_{1:n}\leq \cdots \leq t_{n:n}<\infty$,
we see that $\psi_n'(\lambda)=-(n-2)<0$ for
$\lambda\leq t_{1:n}$, and the function $\psi_n$
is strictly decreasing in the interval $(0,t_{1:n}]$.
Also, $\psi_n(\lambda)$ is strictly convex in the interval
$(t_{1:n},\infty)$,
because
$\psi_n'(\lambda)$
is strictly increasing in
that interval. Observe that $\psi_n$ is eventually
strictly increasing:
$\lim_{\lambda\to\infty}
\psi_n'(\lambda)=-(n-2)+\sum_{i=1}^n \lim_{\lambda\to\infty}u_i'(\lambda)=2$.
It follows that
$\psi_n(\lambda)$ attains its minimum value at a unique point
$\lambda=\lambda_1
>t_{1:n}$;
clearly, $\lambda_1=\lambda_1(c_1)$ is the unique
solution to the equation $\psi_n'(\lambda)=0$, $0<\lambda<\infty$.
\begin{LEM}
\label{lem.5.1}
Let $n\geq 3$ and fix an arbitrary $c_1\in\R$.
The function
$\psi_n:(0,\infty)\to(0,\infty)$, with
$\psi_n(\lambda):=\phi_n(c_1,\lambda)$,
attains its minimum value at a
unique point $\lambda_1=\lambda_1(c_1)$.
The minimizing point $\lambda_1$
is the unique solution of the equation
\be
\label{eq.5.5}
\mbox{$
\sum_{i=1}^n u_i'(\lambda)=n-2,
\ \ \ \ \
t_{n-1:n}<\lambda<\sum_{i=1}^n t_i,
$}
\ee
where $t_i=t_i(c_1)$ are as in (\ref{eq.5.3}),
$0<t_{1:n}\leq \cdots\leq t_{n:n}$ are the ordered values
of $t_i$ in (\ref{eq.5.3}), and the functions
$u_i'(\lambda)$ are given by (\ref{eq.5.4}).
\end{LEM}
\noindent
\begin{Proof}
It remains to verify that the unique solution,
$\lambda=\lambda_1$, of
$\sum_{i=1}^n u_i'(\lambda)=n-2$
lies in the interval $(t_{n-1:n},\sum_{i=1}^n t_i)$.
First observe that if $\lambda\leq t_{n-1:n}$,
then we can find
two indices $s\neq r$ with $\lambda\leq t_s$ and
$\lambda\leq t_r$. Since $u_r'(\lambda)=u_s'(\lambda)=0$,
the sum $\sum_{i=1}^n u_i'(\lambda)$ contains
at most $n-2$ strictly positive
terms $u_i'(\lambda)$; from $u_i'(\lambda)<1$
it follows that $\sum_{i=1}^n u_i'(\lambda)<n-2$.
This shows that $\lambda_1>t_{n-1:n}$.
On the other hand, we observe that
$\lim_{\lambda\searrow 0} \psi_n(\lambda)=2\sum_{i=1}^n t_i$.
Thus, $\psi_n(\lambda_1)\leq 2\sum_{i=1}^n t_i$
(because $\lambda=\lambda_1$ minimizes $\psi_n(\lambda)$).
However, we know that
$\phi_n(c,\lambda)>2\lambda$ for all $(c,\lambda)\in T$;
thus,
$2\sum_{i=1}^n t_i\geq \psi_n(\lambda_1)=\phi_n(c_1,\lambda_1)>2\lambda_1$.
$\Box$
\medskip
\end{Proof}
\begin{REM}
\label{rem.5.1} Fix a point $(c_1,\lambda_1)\in T_0$ and
define the following (possibly empty) sets of indices:
\be
\label{eq.6.2}
\begin{array}{l}
I_1:=\big\{ i\in\{1,\ldots,n\}: (\mu_i-c_1)^2+\sigma_i^2\geq
4\lambda_1^2\big\}=\{i: \lambda_1\leq t_i\},
\\
I_2:=\big\{ i\in\{1,\ldots,n\}:
2\lambda_1|\mu_i-c_1|<(\mu_i-c_1)^2+\sigma_i^2<4\lambda_1^2\big\}
=\{i: t_i<\lambda_1<\gamma_i\},
\\
I_3:=\big\{ i\in\{1,\ldots,n\}: (\mu_i-c_1)^2+\sigma_i^2\leq
2\lambda_1 (\mu_i-c_1)\big\}=\{i: \lambda_1\geq \gamma_i \mbox{ and } \mu_i>c_1\},
\\
I_4:=
\big\{ i\in\{1,\ldots,n\}: (\mu_i-c_1)^2+\sigma_i^2\leq
2\lambda_1 (c_1-\mu_i)\big\}=\{i: \lambda_1\geq \gamma_i \mbox{ and } \mu_i<c_1\}.
\end{array}
\ee
By definition, $I_i\cap I_j=\emptyset$ for $i\neq j$ and
$I_1\cup I_2\cup I_3\cup I_4=\{1,\ldots,n\}$. Since $(c_1,\lambda_1)\in T_0$
it follows that $\lambda_1$ must solves (\ref{eq.5.5}) (for this
particular value of $c_1$), that is,
\[
\sum_{i\in I_2} \Big\{1-\frac{(\mu_i-c_1)^2+\sigma_i^2}{4\lambda_1^2}\Big\}
+
\sum_{i\in I_3\cup I_4}
\frac{1}{2}\Big\{1+\frac{\lambda_1-|\mu_i-c_1|}
{\sqrt{(\lambda_1-|\mu_i-c_1|)^2+\sigma_i^2}}\Big\}
=n-2,
\]
where an empty sum should be treated as zero. Observe that
all summands are (strictly positive and) strictly less than $1$; thus,
$N(I_2)+N(I_3)+N(I_4)\geq n-1$, and it follows that
$N(I_1)\leq 1$, where $N(I)$ denotes the cardinality of $I$.
Furthermore, $(c,\lambda)=(c_1,\lambda_1)$ is a
solution to $\frac{\partial}{\partial c}\phi_n(c,\lambda)=0$.
Using $\frac{\partial}{\partial c} \phi_n(c,\lambda)=
-\frac{1}{2}\sum_{i=1}^n U_1\big(\frac{\mu_i-c}{\lambda},
\frac{\sigma_i}{\lambda}\big)$
and the explicit form of $U_1$, given by (\ref{eq.4.1}), we obtain
\be
\label{eq.5.6}
\mbox{$
\sum_{i\in I_1} \frac{\mu_i-c_1}{\sqrt{(\mu_i-c_1)^2+\sigma_i^2}}
+
\sum_{i\in I_2} \frac{\mu_i-c_1}{2\lambda_1}
+
\sum_{i\in I_3\cup I_4}
\frac{
\mbox{\footnotesize sign}(\mu_i-c_1)}{2}\Big\{1+\frac{|\mu_i-c_1|-\lambda_1}
{\sqrt{(|\mu_i-c_1|-\lambda_1)^2+\sigma_i^2}}\Big\}
=0.
$}
\ee
This equality shows that $N(I_3)\leq n-1$ and $N(I_4)\leq n-1$;
for if, e.g., $N(I_3)=n$ then we would have $I_1=I_2=I_4=\emptyset$ and,
since $\mu_i>c_1$ whenever $i\in I_3$, the above equation leads
to the (obviously impossible) relation
\[
\sum_{i=1}^n
\frac{1}{2}\Big\{1+\frac{(\mu_i-c_1)-\lambda_1}
{\sqrt{((\mu_i-c_1)-\lambda_1)^2+\sigma_i^2}}\Big\}
=0.
\]
We have thus concluded the following key-property
of a minimizing point:
\be
\label{eq.5.7}
\mbox{If $(c_1,\lambda_1)\in T_0$ then
$\max\big\{N(I_3),N(I_4)\big\}\leq n-1$ and
$N(I_1)\leq 1$.}
\ee
Most cases
suggested by (\ref{eq.5.7}) may
appear for some values
of $\bmu,{\bsigma}$
(one of the rare exceptions is $N(I_1)=N(I_2)=0$,
$\max\{N(I_3),N(I_4)\}=n-1$).
Note that Theorem \ref{theo.AG} is, in fact, concerned
with the particular situation
where $N(I_2)=n$ (thus, $N(I_1)=N(I_3)=N(I_4)=0$).
It is, essentially, the unique
situation in which the
$AG_n$ bound is tight (plus boundary subcases).
Due to (\ref{eq.5.7}),
it seems that this particular (but plausible) case is
quite restricted.
Behind the tedious calculations, the rough meaning of
the argument the led to (\ref{eq.5.7}),
is the following: For a particular $(c,\lambda)$
to be optimal (i.e., to minimize $\phi_n$) it is
necessary that $c$ is not ``too far away'' from the
$\mu_i$'s and $\lambda$ is not ``too small'' or ``too large''
compared to $\frac{1}{2}\sum_{i=1}^n {\sigma_i}$.
In particular, (\ref{eq.5.6}) shows that
an optimal $c$ can never lie outside the interval
$\big[\min_i\{\mu_i\}, \max_i\{\mu_i\}\big]$, and it is located
in an interior point when the $\mu_i$'s are not all equal;
of course this fact is intuitively obvious.
\end{REM}
\begin{LEM}
\label{lem.5.2}
If the set $T_0$ of (\ref{eq.5.2})
contains two different elements, then
it must be a compact line segment
which is not parallel to the $\lambda$-axis.
That is, $T_0$ has to be of the form
$T_0=[{\bxx},{\byy}]=\{{\bxx}+t({\byy}-{\bxx}), 0\leq t\leq 1\}$,
for some ${\bxx}=(c_1,\lambda_1)\in T$ and ${\byy}=(c_2,\lambda_2)\in T$
with $c_1\neq c_2$.
\end{LEM}
\begin{LEM}
\label{lem.5.3}
Let $A_0=(c_0,\lambda_0)\neq A_1=(c_1,\lambda_1)$ be two points in $T$.
Fix $\mu\in\R$, $\sigma>0$
and consider the points
$B_0=\big(\frac{\mu-c_0}{\lambda_0},\frac{\sigma}{\lambda_0}\big)\in T$,
$B_1=\big(\frac{\mu-c_1}{\lambda_1},\frac{\sigma}{\lambda_1}\big)\in T$.
Let $A=(c,\lambda)$ and $B=\big(\frac{\mu-c}{\lambda},\frac{\sigma}{\lambda}\big)$.
As the point $A$ is moving linearly in the line segment $[A_0,A_1]$ (from $A_0$ to $A_1$),
the point $B=B(A)$ is moving continuously in the line segment
$[B_0,B_1]$ (from $B_0$ to $B_1$).
\end{LEM}
We are now ready to state
the conclusion of the present section.
\begin{THEO}
\label{theo.5.1}
If $n\geq 3$ then for any given values of $\bmu$ and ${\bsigma}$,
there exists a unique solution $(c,\lambda)=(c_0,\lambda_0)$ of (\ref{eq.5.1}), and
\be
\label{eq.5.8}
\phi_n(c,\lambda)\geq \phi_n(c_0,\lambda_0) \ \ \
\mbox{ for all }
(c,\lambda)\in \R\times(0,\infty),
\ee
with equality if and only if $(c,\lambda)=(c_0,\lambda_0)$.
\end{THEO}
\begin{REM}
\label{rem.5.2} For $n=2$, Theorem \ref{theo.5.1}
(as well as several conclusions of the present section)
is no longer true. It is again true that the convex function
$\phi_2(c,\lambda)$ attains its minimum value,
$\rho_2=\sqrt{(\mu_2-\mu_1)^2+(\sigma_1+\sigma_2)^2}$,
at the solutions of the system
\eqref{eq.5.1}, but now $T_0$ is not a singleton:
it contains points arbitrarily close to the boundary
of the domain of $\phi_2$.
More precisely, one can verify that for $n=2$,
the exact set of minimizing points
is the line segment
$T_0=\{(c_0,\lambda); 0<\lambda\leq \lambda_0\}$,
where
\[
\mbox{
$c_0=\frac{\sigma_1}{\sigma_1+\sigma_2}\mu_2+\frac{\sigma_2}{\sigma_1+\sigma_2}\mu_1$,
\ \ \
$\lambda_0=\frac{\rho_2}{2(\sigma_1+\sigma_2)}\min\{\sigma_1,\sigma_2\}$.
}
\]
However, the set
${\cal E}_2(\mu_1,\mu_2,\sigma_1,\sigma_2)$ is a singleton, and
this fact can be seen directly (see
Section \ref{sec.last});
thus, the above calculation is completely unnecessary.
Also, it is worth pointing out that, for $n=2$, $N(I_1)=2=n$
(compare with \eqref{eq.5.7}).
\end{REM}
\section{Tightness and characterization of extremal random vectors}
\label{sec.6}
\setcounter{equation}{0}
Let $n\geq 3$, $\bmu$, ${\bsigma}$ be fixed (with $0<\sigma_i<\infty$
for all $i$). Let $(c,\lambda)$ be the unique solution of (\ref{eq.5.1}).
With the help of $(c,\lambda)$ we shall
give a complete description of the set ${\cal E}_n(\bmu,{\bsigma})$
of extremal random vectors in
${\cal F}_n(\bmu,{\bsigma})$.
These are the random vectors ${\bXX}$ satisfying
$\E {\bXX}=\bmu$, $\Var {\bXX}={\bsigma}^2$ and
$\E R_n({\bXX})=\rho_n=\rho_n(\bmu,{\bsigma})$,
where
\be
\label{eq.6.1}
\mbox{$
\rho_n:=\phi_n(c,\lambda)=-(n-2)\lambda+\frac{\lambda}
{2}\sum_{i=1}^n
U\big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\big);
$}
\ee
recall that $U(\cdot,\cdot)$ is given by (\ref{eq.u}).
The construction, though more complicated, follows parallel
arguments as for the attainability of the $AG_n$ bound
(Theorem \ref{theo.AG}).
We start by considering the partition $I_1,\ldots,I_4$
of
$\{1,\ldots,n\}$ as in \eqref{eq.6.2},
and the
corresponding cardinalities $n_1,\ldots,n_4$.
The main difference from Remark \ref{rem.5.1} is that, now,
each $I_j$ has been stabilized, because $(c,\lambda)$ is unique; thus,
one has to substitute $c_1=c$ and $\lambda_1=\lambda$
in \eqref{eq.6.2}.
Clearly some of the sets $I_j$ may be empty; then $n_j=0$. The situation
with all $I_j$ being nonempty may also appear; this is the case, e.g.,
for $\bmu=(4,0,4,0)$, ${\bsigma}=(10,5,1,1)$.
From Remark \ref{rem.5.1} (see (\ref{eq.5.7})) we know that
$n_1$ $n_2$, $n_3$, $n_4$ (with $n_j\geq 0$, $\sum n_j=n$)
cannot be completely arbitrary; they have to satisfy the restrictions:
\be
\label{eq.6.8old}
n_3=N(I_3)\leq n-1, \ \ n_4=N(I_4)\leq n-1, \ \ n_1=N(I_1)\leq 1.
\ee
Other impossible cases are given by $n_3=1,n_4=n-1$ and $n_3=n-1,n_4=1$;
this is a by-product of Lemma \ref{lem.6.1}, below.
For notational simplicity it is helpful
to consider the following numbers $\xi_i$,
$\theta_i$:
\be
\label{eq.6.3}
\hspace{-.2ex}
\xi_i:=\left\{
\begin{array}{ll}
\mu_i-c,
&
i\in I_1\cup I_2;
\\
\mu_i-c-\lambda,
&
i\in I_3;
\\
c-\mu_i-\lambda,
&
i\in I_4;
\end{array}
\right.
\theta_i:=\left\{
\begin{array}{ll}
\sqrt{(\mu_i-c)^2+\sigma_i^2},
&
i\in I_1\cup I_2;
\\
\sqrt{(\mu_i-c-\lambda)^2+\sigma_i^2},
&
i\in I_3;
\\
\sqrt{(c-\mu_i-\lambda)^2+\sigma_i^2},
&
i\in I_4.
\end{array}
\right.
\ee
We note that $|\xi_i|<\theta_i$ for all $i$
and $2\lambda|\xi_i|<\theta_i^2<4\lambda^2$ for all $i\in I_2$ (if any).
Following Corollary \ref{cor.1}
we define
the
probabilities
\be
\label{eq.6.4}
\begin{array}{llll}
\vspace{.5ex}
p_i^-:=\frac{1}{2}\big(1-\frac{\xi_i}{\theta_i}\big),
&
p_i^{o}:=0,
&
p_i^+:=\frac{1}{2}\big(1+\frac{\xi_i}{\theta_i}\big),
&
i\in I_1;
\\
\vspace{.5ex}
p_i^-:=\frac{1}{8\lambda^2}\left[\theta_i^2-2\lambda \xi_i\right],
&
p_i^{o}:=1-\frac{\theta_i^2}{4\lambda^2},
&
p_i^+:=\frac{1}{8\lambda^2}\left[\theta_i^2+2\lambda \xi_i\right],
&
i\in I_2;
\\
\vspace{.5ex}
p_i^-:=0,
&
p_i^{o}:=\frac{1}{2}\big(1-\frac{\xi_i}
{\theta_i}\big),
&
p_i^+:=\frac{1}{2}\big(1+\frac{\xi_i}
{\theta_i}\big),
&
i\in I_3;
\\
p_i^-:=\frac{1}{2}\big(1+\frac{\xi_i}
{\theta_i}\big),
&
p_i^{o}:=\frac{1}{2}\Big(1-\frac{\xi_i}
{\theta_i}\big),
&
p_i^+:=0,
&
i\in I_4,
\end{array}
\ee
and the corresponding (univariate) supporting points
\be
\label{eq.6.5}
\begin{array}{llll}
x_i^-:=c-\theta_i,
&&
x_i^+:=c+\theta_i,
& i\in I_1;
\\
x_i^-:=c-2\lambda,
&
x_i^{o}:=c,
&
x_i^+:=c+2\lambda,
&
i\in I_2;
\\
&
x_i^{o}:=c+\lambda-\theta_i,
&
x_i^+:=c+\lambda+\theta_i,
&
i\in I_3;
\\
x_i^-:=c-\lambda-\theta_i,
&
x_i^{o}:=c-\lambda+\theta_i,
&
&
i\in I_4.
\end{array}
\ee
By definition, each
${\bpp}_i:=\big(p_i^-,p_i^{o},p_i^+\big)$ is a
probability vector.
Clearly, one could assign an arbitrary value to a missing point,
since its corresponding probability
is $0$. The most convenient choice is to assign the respective
values $c-2\lambda$, $c$, $c+2\lambda$, whenever $x_i^{-}$, $x_i^{o}$, $x_i^{+}$
is not specified from (\ref{eq.6.5}). With this convention,
\be
\label{eq.6.6}
x_i^{-}<c-\lambda<x_i^{o}<c+\lambda<x_i^{+}, \ \ \ i=1,\ldots,n.
\ee
\noindent
Let $X_i$ be a random variable
which assumes values $x_i^{-},x_i^{o},x_i^{+}$
with respective probabilities $p_i^{-},p_i^{o},p_i^{+}$.
Corollary \ref{cor.1} asserts that
(the distribution of) $X_i$
is
characterized be the
fact that maximizes the expectation of
$|(X-c)-\lambda|+|(X-c)+\lambda|$ as
$X$ varies in ${\cal F}_1(\mu_i,\sigma_i)$.
The following lemma provides the most
fundamental tool
for the main result.
\begin{LEM}
\label{lem.6.1}
The probabilities $p_i^+$, $p_i^-$ in (\ref{eq.6.4}) satisfy the relation
\be
\label{eq.6.7}
\sum_{i=1}^n {p_i^+}=\sum_{i=1}^n p_i^{-}=1.
\ee
\medskip
\end{LEM}
Lemma \ref{lem.6.1} enables us to define the $n$-variate
probability vectors
\be
\label{eq.6.9}
{\bpp}^{+}=(p_1^{+},\ldots,p_n^+), \ \
{\bpp}^{-}=(p_1^{-},\ldots,p_n^-).
\ee
By definition,
${\bpp}^+$ has its zero elements at exactly the positions $i$ where
$i\in I_4$
(if $I_4= \emptyset$, all $p_i^+$'s are positive), and
${\bpp}^-$ has its zero elements at exactly the positions $i$ where
$i\in I_3$
(if any).
\begin{PROP}
\label{prop.6.1}
Assume we are given $n\geq 3$, $\bmu$, ${\bsigma}$.
Then, (i) and (ii) are
equivalent:
\vspace{-1ex}
\begin{itemize}
\item[(i)] We can find a random vector
${\bXX}\in{\cal F}_n(\bmu,{\bsigma})$ such that
\vspace{-1ex}
$\E R_n=\rho_n$.
\item[(ii)] There exists a $n\times n$ probability matrix
$Q\in {\cal M}({\bpp}^{+},{\bpp}^{-})$
such that $q_{ii}=0$ for all
$i\in \{1,\ldots,n\}$.
\end{itemize}
Moreover,
with ${\cal L}({\bXX})$ denoting the probability law of the random vector
${\bXX}$, the correspondence ${\cal L}({\bXX})
\rightleftarrows
Q$ is a bijection; the
explicit formula for the
transformation $Q=(q_{ij})\mapsto {\cal L}({\bXX})$
is given by
\be
\label{eq.6.10}
\begin{array}{c}
\Pr[{\bXX}={\bxx}_{ij}]=q_{ij}, \mbox{ where }
{\bxx}_{ij}:=\big(x_1^{o},\ldots,x_{i-1}^{o},x_i^+,x_{i+1}^{o},\ldots,
x_{j-1}^{o},x_j^-,x_{j+1}^{o},\ldots,x_n^{o}\big),
\\
\mbox{~}\hspace{50ex} \ \
i\neq j, \ \ i,j=1,\ldots,n.
\end{array}
\ee
\end{PROP}
The main result of the present work reads as follows:
\begin{THEO}
\label{theo.6.1}
Let $n\geq 3$, $\mu_i\in\R$, $\sigma_i>0$ ($i=1,\ldots,n$). Then,
\noindent
(a)
\be
\label{eq.6.17}
\sup \E R_n =\rho_n,
\ee
where the supremum is taken over
${\bXX}\in{\cal F}_{n}(\bmu,{\bsigma})$
and $\rho_n=\rho_n(\bmu,{\bsigma})$ is given
by (\ref{eq.6.1}), with $(c,\lambda)=(c(\bmu,{\bsigma}),
\lambda(\bmu,{\bsigma}))$ being the
unique solution to the system of equations
\medskip
(\ref{eq.5.1}).
\noindent
(b)
The set ${\cal E}_n(\bmu,{\bsigma})$
is nonempty.
Any
extremal ${\bXX}\in{\cal E}_n(\bmu,{\bsigma})$
is produced by (\ref{eq.6.10}),
with $x_i^-$, $x_i^o$, $x_i^+$ as in (\ref{eq.6.5}),
and corresponds uniquely to a $n\times n$
probability matrix $Q\in {\cal M}({\bpp}^+, {\bpp}^-)$
with zero diagonal entries, where ${\bpp}^+$,
${\bpp}^-$ are given by (\ref{eq.6.9}).
\end{THEO}
\noindent
\begin{Proof} From Theorem \ref{th.main}
we know that $\E R_n\leq \rho_n$ and it suffices to
prove (b). In view of Proposition
\ref{prop.6.1}, it remains to verify that the class
of $n\times n$ probability matrices with zero diagonal entries
and marginals ${\bpp}^+$, ${\bpp}^-$
is nonempty.
However,
this fact follows immediately from Lemma \ref{lem.3},
because $\max_i\{p_i^++p_i^-\}\leq 1$ (see (\ref{eq.6.4})),
and the proof is complete.
$\Box$
\end{Proof}
\begin{REM}
\label{rem.6.1}
Since $\E R_n({\bXX})=\rho_n$ for any
${\bXX}\in{\cal E}_n(\bmu,\bsigma)$,
\[
\mbox{$
\rho_n=\sum_{i=1}^n \{ (x_i^+-c)p_i^+ +(c-x_i^-)p_i^-\}.
$}
\]
\end{REM}
\begin{COR}
\label{cor.6.1}
If $I_1\neq \emptyset$ (see (\ref{eq.6.2})) then $I_1=\{k\}$
for some $k\in\{1,\ldots,n\}$, and
the equality in
(\ref{eq.6.17}) characterizes the random vector ${\bXX}$
with probability law
\be
\label{eq.6.19}
\Pr[{\bXX}={\bxx}_{ik}]=p_i^+, \ \
\Pr[{\bXX}={\bxx}_{ki}]=p_i^-, \ \ i\neq k, \ i=1,\ldots,n.
\ee
\end{COR}
\noindent
\begin{Proof} From (\ref{eq.6.8old})
we know that $N(I_1)\leq 1$, and thus,
$I_1=\{k\}$ for some $k$. Since
$k\in I_1$, (\ref{eq.6.4}) shows that
$\max_i\{p_i^++p_i^-\}=p_k^++p_k^{-}=1$.
[Note that, by Lemma \ref{lem.6.1},
$\sum_{i\neq k}p_i^{-}+\sum_{i\neq k}p_i^{+}=(1-p_k^-)+(1-p_k^+)=1$
and, hence, (\ref{eq.6.19}) defines a probability law.]
Lemma \ref{lem.3} implies uniqueness
of $Q$, hence of ${\cal L}({\bXX})$
(see (\ref{eq.6.10})).
It is easily seen that the matrix $Q$, obtained
by (\ref{eq.6.19}) through (\ref{eq.6.16}), is
indeed the unique probability matrix
with vanishing diagonal entries and marginals ${\bpp}^+$,
${\bpp}^-$.
\medskip
$\Box$
\end{Proof}
Corollary \ref{cor.6.1}
implies uniqueness (denoted by (U)) for the second counterpart
of the bound (\ref{eq.tight}) in Example \ref{ex.2}.
It should be noted that the converse of Corollary \ref{cor.6.1}
does not hold; that is, the condition $I_1\neq \emptyset$
is not necessary for concluding uniqueness
of the extremal
random vector ${\bXX}$.
A particular example was given by Remark \ref{rem.3.1}.
Clearly, the most interesting situations in practice
arise when $I_1=\emptyset$. In such
cases it is fairly expected that there will be infinitely
many extremal vectors, as in Theorem \ref{theo.AG}.
This is, indeed, true in general, but not always.
Lemma \ref{lem.3} guarantees infiniteness
(denoted by (I)) only
if all $p_{i}^+$, $p_i^-$ are nonzero, and this corresponds
to the quite restricted case where $I_2=\{1,\ldots,n\}$.
Of course, given the existence of two extremal vectors,
one can deduce (I) by considering convex combinations of the
corresponding matrices; cf.\ Example \ref{ex.1}.
If $I_1=\emptyset$,
the complete distinction between (U) and
(I) depends upon the values of $n$, $n_3=N(I_3)$ and
$n_4=N(I_4)$ (see (\ref{eq.6.2}) and (\ref{eq.6.8old}));
and if $n_3=n_4=0$ we already know that (I) results.
We briefly discuss all remaining situations where $I_1=\emptyset$:
If $n_2=N(I_2)=0$ and $n_3\geq 2$, $n_4\geq 2$,
it is obvious that (I)
holds; note that $n_3=1,n_4=n-1$ and
$n_3=n-1,n_4=1$ are impossible
by
Lemma \ref{lem.6.1}.
If $n_2=n_3=n_4=1$ or $n_2=2,n_3=1, n_4=0$ or $n_2=2,n_3=0,n_4=1$
then we are in (U), while
(I) results if $n_2=n_3=1, n_4\geq 2$ or $n_2=n_4=1, n_3\geq 2$.
If $n_2=1, n_3\geq 2, n_4\geq 2$
then we get (I), as well as in
all remaining cases where
$n_2\geq 2, n_3\geq 0, n_4\geq 0$.
The final conclusion is as follows: If $I_1=\emptyset$,
the situations where the extremal distribution is uniquely
defined are described by $n_2=n_3=n_4=1$ or $n_2=2,n_3=1,n_4=0$ or
$n_2=2,n_3=0,n_4=1$
(and thus, $n=3$); this provides an explanation
to Remark \ref{rem.3.1}. However, we note that
knowledge of the values $n_j$ actually requires knowledge
of the region where the optimal $(c,\lambda)$
appears,
and this may be, or may not be, an easy task
for particular $\bmu$, ${\bsigma}$.
\begin{REM}
\label{rem.6.2}
The range
$R_n({\bXX})$ of an extremal vector ${\bXX}$ need not be
a degenerate random variable. An example is provided
by $\bmu=(-2,0,2)$, ${\bsigma}=(1,3,1)$. Then,
$n_1=0$, $n_2=n_3=n_4=1$ and it can be shown that
\[
\mbox{$
\lambda\approx 1.737, \ \ \
\rho_3=\frac{64 \lambda^3-72\lambda-81}{4\lambda(4\lambda^2-9)}\approx 6.066
$}
\]
($\lambda$ is the unique solution of
$4\lambda^2(2-\lambda)=(4\lambda^2-9)\sqrt{\lambda^2-4\lambda+5}$, and this
reduces to a four-degree polynomial equation).
The range $R_3$ of the unique extremal vector assumes
values $2\lambda+\frac{\lambda(8\lambda-9)}{4\lambda^2-9}\approx5.542$
and $\frac{2\lambda(8\lambda-9)}{4\lambda^2-9}\approx6.245$ with respective
probabilities $\frac{9}{4\lambda^2}\approx .254$ and
$1-\frac{9}{4\lambda^2}\approx .746$.
However, the improvement over the bound $AG_3=\sqrt{38}\approx 6.164$
is negligible. As a general observation, even for small $n$,
the value of $\rho_n$ is difficult to evaluate when more
than two index sets
$I_j$ are nonempty.
\end{REM}
\begin{EXAMPLE}
\label{ex.6.1} {\it Homoscedastic observations from two balanced groups.}
Let $n=2k$, $\sigma_i^2=\sigma^2$
and $\mu_i=-\mu$ or $\mu$ according to $i\leq k$
or $i>k$, respectively ($\mu\geq 0$).
The Arnold-Groeneveld bound
(\ref{eq.AG}) takes here the form
\[
\E R_{2k}\leq AG_{2k}=2\sqrt{k(\mu^2+\sigma^2)},
\]
and it is tight if $\mu\leq \frac{\sigma}{\sqrt{k-1}}$
(in particular, if $n=2$ or $\mu=0$). Also, we
know from Theorem \ref{theo.AG} the nature of the random vectors
that attain the equality.
However, for $\mu\geq \frac{\sigma}{\sqrt{k-1}}$ one finds
$N(I_3)=N(I_4)=k$, and the tight bound
of Theorem \ref{theo.6.1} becomes
\[
\mbox{$
\E R_{2k}\leq \rho_{2k}=2\mu+2\sigma\sqrt{k-1} \ \ \ \big(\mu\geq \frac{\sigma}{\sqrt{k-1}}\big);
$}
\]
note that $\rho_{2k}$ is equal to $AG_{2k}$ only in the boundary case
$\sigma=\mu\sqrt{k-1}$. For $\mu\geq \frac{\sigma}{\sqrt{k-1}}$ the
nature
of extremal random vectors is different:
They assume values
\[
{\byy}_{ij}=\big(-x,\ldots,-x,-y,-x,\ldots,-x \ ; \ x,\ldots,x,y,x,\ldots,x\big), \ \ i,j=1,\ldots,k,
\]
where $-y$ is located at the $i$-th place and
$y$ is located at the $(k+j)$-th place of the vector.
Here, $0\leq x=\mu-\frac{\sigma}{\sqrt{k-1}}<y=\mu
+\sigma\sqrt{k-1}$. The respective probabilities $p_{ij}=\Pr[{\bXX}={\byy}_{ij}]$, $i,j=1,\ldots,k$, correspond to a probability matrix $P_{k\times k}$
with uniform marginals. Both limits
\[
\mbox{
$\lim_{\mu\to\infty}\frac{\rho_{2k}}{AG_{2k}}=\frac{1}{\sqrt{k}}$
\ ($k$, $\sigma$ fixed), \ \ \
$\lim_{k\to\infty}\frac{\rho_{2k}}{AG_{2k}}
=\frac{\sigma}{\sqrt{\mu^2+\sigma^2}}$
\
($\mu$, $\sigma$ fixed)
}
\]
show that, under some circumstances,
the improvement that is achieved
by using $\rho_n$ instead of $AG_n$
can become arbitrarily large.
\end{EXAMPLE}
\begin{EXAMPLE}
\label{ex.6.2} {\it Homoscedastic data with a single outlier.}
Let $\sigma_i^2=\sigma^2$ for all $i$, $\mu_i=0$
($i=1,\ldots,n-1$) and $\mu_n=\mu\geq 0$. Theorem \ref{theo.AG}
asserts that the bound
\[
\mbox{$
\E R_n\leq AG_n=\sqrt{2\frac{n-1}{n}\mu^2+2n\sigma^2}
$}
\]
is not tight for $n\geq 3$ and $\mu>\frac{n}{\sqrt{n-1}}\sigma$.
The tight bound has the form
\[
\E R_n\leq \rho_n=\sqrt{(n-1)(c^2+\sigma^2)}+\sqrt{(\mu-c)^2+\sigma^2},
\]
where $c$ is the unique root of the equation
\[
\mbox{
$\frac{c\sqrt{n-1}}{\sqrt{c^2+\sigma^2}}=\frac{\mu-c}{\sqrt{(\mu-c)^2+\sigma^2}}$, \ \ \
$0<c<\min\big\{\frac{\mu}{n},\frac{\sigma}{\sqrt{n-2}}\big\}$.
}
\]
Although $\rho_n<AG_n$ (for $\mu\sqrt{n-1}>n\sigma$),
it is not easy to make direct comparisons. However,
$c^2<\frac{\sigma^2}{n-2}$ and $(\mu-c)^2<\mu^2$, so that
$\rho_n<\rho_n'=\frac{n-1}{\sqrt{n-2}}\sigma+\sqrt{\mu^2+\sigma^2}$.
Hence, for the (non-tight) upper bound $\rho_n'$,
\[
\mbox{$
\lim_{\mu\to\infty}\frac{\rho_n'}{AG_n}=\frac{\sqrt{n}}{\sqrt{2n-2}}
\ \ \ \ (n\geq3, \ n,\sigma \mbox{ fixed}).
$}
\]
\end{EXAMPLE}
\begin{REM}
\label{rem.6.3}
Example \ref{ex.6.2} and Remark \ref{rem.6.2}
entail that $\rho_n$ may have a rather
complicated form when the $\mu_i$'s
are not all equal. On the other hand,
$\rho_n$ becomes quite plausible in the case of
equal $\mu_i$'s; see Example \ref{ex.2}.
This particular case is useful in concluding
some facts about the behavior of $\rho_n$ in
general. Indeed,
taking into account the obvious relation
$U(x,y)\geq U(0,y)$, we see that
for any given $\bmu$ and ${\bsigma}$,
\[
\mbox{$
\rho_n=\phi_n(c,\lambda)\geq
-(n-2)\lambda+\frac{\lambda}{2}\sum_{i=1}^n
U\big(0,\frac{\sigma_i}{\lambda}\big)
=\widehat{\phi}_n(0,\lambda)\geq\widehat{\rho}_n:=\inf_{x\in \R, y>0}\widehat{\phi}_n(x,y),
$
}
\]
where $\widehat{\rho}_n$ is the upper bound of Theorem
\ref{th.main}, calculated under $\mu_i=\mu$ for all $i$,
and for the given ${\bsigma}$.
Since $\widehat{\rho}_n =\min_{y>0}\widehat{\phi}_n(0,y)$
admits a simple closed form, see (\ref{eq.tight}),
we get the following lower bound:
\[
\rho_n\geq
\widehat{\rho}_n=
\left\{
\begin{array}{lll}
\sqrt{2\sum_{i=1}^n \sigma_i^2}, & \mbox{if} & 2\max_i\{\sigma_i^2\}
\leq \sum_{i=1}^n\sigma_i^2,
\\
\max_i\{\sigma_i\} +\sqrt{\sum_{i=1}^{n}\sigma_i^2-\max_i \{\sigma_i^2\}},
& \mbox{if} & 2\max_i\{\sigma_i^2\}
\geq \sum_{i=1}^n\sigma_i^2,
\end{array}
\right.
\mbox{any }\bmu,{\bsigma}.
\]
Since $U(x,y)>U(0,y)$ for $x\neq 0$,
the equality holds only if all the $\mu_i$'s are equal.
Despite its weakness, this lower bound provides an idea of
what can be expected
for the actual size of $\rho_n$.
It is also helpful in giving some light to
the observation that, provided
the means are small compared to the variances,
the $AG_n$ bound tends to be tight.
More precisely,
assume that $\min_i \{\sigma_i^2\} \to \infty$ and
$(\sum_{i=1}^n (\mu_i-\overline{\mu})^2)/(\sum_{i=1}^n \sigma_i^2)\to 0$
(in particular, $\max_i |\mu_i-\overline{\mu}|\leq C<\infty$ suffices for this).
Then, the homogeneity assumption
$\max_i\{\sigma_i^2\}\leq (n-1)\min_i\{\sigma_i^2\}$
is sufficient for the asymptotic tightness of the $AG_n$
bound (for fixed $n\geq 3$).
Indeed, from this assumption we get
$\sum_{i=1}^n \sigma_i^2\geq \max_{i}\{\sigma_i^2\}+(n-1)\min_i \{\sigma_i^2\}
\geq 2\max\{\sigma_i^2\}$, and thus,
$\widehat{\rho}_n=\sqrt{2\sum_{i=1}^n\sigma_i^2}$.
Hence,
\[
1\geq
\Big(\frac{\rho_n}{AG_n}\Big)^2
\geq
\Big(\frac{\widehat{\rho}_n}{AG_n}\Big)^2
=
\frac{\sum_{i=1}^n\sigma_i^2}
{\sum_{i=1}^n\{(\mu_i-\overline{\mu})^2+\sigma_i^2\}}
=
\frac{1}
{1+(\sum_{i=1}^n (\mu_i-\overline{\mu})^2)/(\sum_{i=1}^n\sigma_i^2)}\to 1.
\]
Therefore, under the above circumstances,
the improvement achieved by using
$\rho_n$ instead of $AG_n$ becomes negligible.
\end{REM}
\section{The case $\bbb{n}\, \bbb{=}\, \bbb{2}$ and
further remarks}
\setcounter{equation}{0}
\label{sec.last}
For $n=2$ the bound $\rho_2$ admits a closed form. More precisely,
from Theorem \ref{th.main},
\be
\label{eq.7.1}
\E R_2\leq \rho_2, \ \
\mbox{ where }
\rho_2:=\inf_{c\in\R,\lambda>0}\phi_2(c,\lambda)
=\sqrt{(\mu_1-\mu_2)^2+(\sigma_1+\sigma_2)^2};
\ee
see Remark \ref{rem.5.2}. The inequality \eqref{eq.7.1}
is tight, since the equality
is attained
by
(and characterizes)
the random pair $(X_1,X_2)$ with distribution
given by
\be
\label{eq.7.2}
\begin{array}{c}
\vspace{1ex}\mbox{$
\Pr\big[
X_1=\frac{\sigma_2\mu_1+\sigma_1\mu_2}{\sigma_1+\sigma_2}
+\frac{\sigma_1}
{\sigma_1+\sigma_2}\rho_2,
\
X_2=\frac{\sigma_2\mu_1+\sigma_1\mu_2}{\sigma_1+\sigma_2}
-\frac{\sigma_2}
{\sigma_1+\sigma_2}\rho_2
\big]
=\frac{1}{2}\big(1+\frac{\mu_1-\mu_2}{\rho_2}\big),
$}
\\
\mbox{$
\Pr\big[
X_1=\frac{\sigma_2\mu_1+\sigma_1\mu_2}{\sigma_1+\sigma_2}
-\frac{\sigma_1}
{\sigma_1+\sigma_2}\rho_2,
\
X_2=\frac{\sigma_2\mu_1+\sigma_1\mu_2}{\sigma_1+\sigma_2}
+\frac{\sigma_2}
{\sigma_1+\sigma_2}\rho_2
\big]
=\frac{1}{2}\big(1+\frac{\mu_2-\mu_1}{\rho_2}\big).
$}
\end{array}
\ee
Therefore, ${\cal E}_2(\bmu,{\bsigma})$ is a singleton.
Also, $AG_2=\sqrt{(\mu_1-\mu_2)^2+2\sigma_1^2+2\sigma_2^2}$, and it is
worth pointing out that the bound $AG_2$ is tight if and
only if $\sigma_1=\sigma_2$. Another observation is that
the extremal random vector for the expected range
coincides with the (unique) extremal random vector for the expected
maximum (see \eqref{eq.BNT}). However,
this is not a coincidence. In view
of the obvious relationship
\be
\label{eq.max12}
\mbox{$
R_2=|X_1-X_2|=2\max\{X_1,X_2\}-X_1-X_2
=2 X_{2:2}-X_1-X_2,
$}
\ee
a bound for the maximum
can be translated to a bound for the range, and vice-versa
(provided that the expectations, $\mu_1,\mu_2$, of $X_1,X_2$,
are known). In this sense, the bound
$\rho_2$ turns to be a particular case of the results given by
Bertsimas, Natarajan and Teo (2004, 2006), namely
\[
\hspace{-1ex}
\mbox{$
\rho_2=\sup \E R_2
=\sup \E\{ 2 X_{2:2}-X_1-X_2\}
=2\sup \E X_{2:2} -\mu_1-\mu_2=2BNT_2-\mu_1-\mu_2,
$}
\]
and the equality characterizes the same extremal distribution
as for the maximum. Consequently, it is of some interest to observe that
the bound $BNT_2$ admits a closed form, namely,
\[
\mbox{$
BNT_2=\frac12(\mu_1+\mu_2)
+\frac12\sqrt{(\mu_1-\mu_2)^2+(\sigma_1+\sigma_2)^2}.
$}
\]
Note also that
the $BNT_2$--bound improves
the corresponding Arnold-Groeneveld bound (\ref{eq.AG.general})
for the expected maximum only in the case where $\sigma_1\neq \sigma_2$.
It is also worth pointing out that a particular application
of the main result in Papadatos (2001a) yields an even better
(than $BNT_2$, $AG_2$ and $\rho_2$) bound. Indeed, setting
$\rho:=\Corr(X_1,X_2)$, it follows from Papadatos' results
that for any $(X_1,X_2)\in{\cal F}_2(\bmu,{\bsigma})$,
\be
\label{eq.7.3}
\E R_2\leq \gamma_2:=\sqrt{(\mu_1-\mu_2)^2
+(\sigma_1+\sigma_2)^2-2(1+\rho)\sigma_1\sigma_2}.
\ee
Obviously, $\gamma_2\leq \rho_2$ with equality if and only if
$\rho=-1$. This inequality explains the fact that the extremal
random pair $(X_1,X_2)$
(that attains the bounds $\rho_2$ and $BNT_2$)
has correlation $\rho=-1$; see (\ref{eq.7.2}).
The preceding inequalities have some interest because they provide a basis
for the investigation of the dependence structure of an ordered pair.
This kind of investigation is particularly useful for its application to
reliability systems; see Navarro and Balakrishnan (2010). On the other hand,
in view of the obvious facts $X_{1:2}+X_{2:2}=X_1+X_2$ and
$X_{1:2}X_{2:2}=X_1X_2$,
we get the relation
\be
\label{eq.cov12}
\mbox{$
\Cov[X_{1:2},X_{2:2}]=\rho\sigma_1\sigma_2-\frac14 (\mu_2-\mu_1)^2+\frac14 (\E R_2)^2,
\ \ (X_1,X_2)\in {\cal F}_2(\bmu,{\bsigma}),
$}
\ee
where $\rho=\Corr(X_1,X_2)$.
Thus, any bound (upper or lower) for $\E R_2$ can be translated
to a bound for $\Cov(X_{1:2},X_{2:2})$ as well as for $\E X_{2:2}$; see
Papathanasiou (1990), Balakrishnan and Balasubramanian (1993).
Therefore,
it is of some interest to know whether the bound in \eqref{eq.7.3} is tight
for given $\rho$. This is indeed the case but,
to the best of our knowledge, this elementary fact does not
seem to be well-known, and we shall provide a simple proof here.
To this end, let $\bmu=(\mu_1,\mu_2)$, ${\bsigma}=(\sigma_1,\sigma_2)$
(with $\sigma_1>0$, $\sigma_2>0$), $-1\leq \rho\leq 1$, and define the section
\[
{\cal F}_2(\bmu,{\bsigma};\rho):=
\big\{(X_1,X_2)\in {\cal F}_2(\bmu,{\bsigma})
:\Corr(X_1,X_2)=\rho\big\}.
\]
Then we have the following.
\begin{THEO}
As $(X_1,X_2)$ varies in ${\cal F}_2(\bmu,{\bsigma};\rho)$,
\label{theo.7.1}
\be
\label{eq.7.4}
\inf \E R_2=|\mu_2-\mu_1|, \ \ \
\sup \E R_2=\sqrt{(\mu_2-\mu_1)^2+(\sigma_1+\sigma_2)^2-2(1+\rho)\sigma_1\sigma_2}.
\ee
\end{THEO}
\begin{REM}
\label{rem.7.1} From the proof it follows that (the probability law of)
the extremal vector $(X_1,X_2)$
$\in{\cal F}_2(\bmu,{\bsigma};\rho)$
that attains the equality in \eqref{eq.7.3}
is unique if and only if
either (i) $\rho=-1$ or (ii) $\sigma_1\neq \sigma_2$ and $\rho=1$.
With this in mind, let us keep $\mu_1,\mu_2,\sigma_1,\sigma_2$ constant,
and write
$\gamma_2=\gamma_2(\rho)$ for the quantity defined by \eqref{eq.7.3}.
Then, $\gamma_2(\rho)$ is strictly decreasing in $\rho$
(recall that $\sigma_1>0$, $\sigma_2>0$), attaining its maximum value
at $\rho=-1$. By definition, $\gamma_2(-1)=\rho_2$ (see \eqref{eq.7.1}),
and thus,
for the equality $\E R_2=\rho_2$ it is necessary that $\rho=-1$.
This observation verifies that the unique distribution that attains
the equality
in \eqref{eq.7.1} is the $BNT_2$--distribution, given by (\ref{eq.7.2}).
\end{REM}
In view of \eqref{eq.max12}, \eqref{eq.cov12},
the following result is straightforward
from Theorem \ref{theo.7.1}.
\begin{COR}
\label{cor.7.1}
Let $(X_1,X_2)\in{\cal F}_2(\mu_1,\mu_2,\sigma_1,\sigma_2;\rho)$ with
$\sigma_1>0,\sigma_2>0$. Then,
\[
\begin{array}{l}
\vspace{1.3ex}
\max\{\mu_1,\mu_2\}\leq \E\big\{\hspace{-.5ex}\max\{X_1,X_2\}\hspace{-.2ex}\big\}\leq
\frac{1}{2}(\mu_1+\mu_2)+\frac{1}{2}
\sqrt{(\mu_2-\mu_1)^2+\sigma_1^2+\sigma_2^2-2\rho\sigma_1\sigma_2},
\\
\rho\sigma_1\sigma_2\leq \Cov\hspace{-.4ex}\big[\min\{X_1,X_2\},\max\{X_1,X_2\}\big]
\leq
\frac{1}{4}
\big(\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2\big).
\end{array}
\]
All bounds are best possible.
\end{COR}
\noindent
It is worth pointing out that, as Corollary \ref{eq.7.1} shows,
the covariance of an ordered pair can never be
smaller than the covariance of the observations and, in particular,
an ordered pair formed from non-negatively correlated observations
is non-negatively correlated. While these facts,
as well as the lower covariance bound
of an ordered pair,
are well-known (see eq.'s (2.9), (2.11)
in Navarro and Balakrishnan (2010)),
the upper bound seems to be of some interest.
There are some propositions and questions for further
research. An obvious one is in extending the main result
of Theorem \ref{theo.6.1} and of (\ref{eq.BNT}) to more general
$L$-statistics.
Recall that the tight bound for any $L$-statistic under the i.d.\
assumption is known from the work of Rychlik (1993b). However,
Rychlik's result is not applicable
if arbitrary multivariate distributions are
allowed for the data.
A second one concerns extension to other $L$-statistics
of the bounds given in
Corollary \ref{cor.7.1} and Theorem \ref{theo.7.1}
for $n\geq 3$, noting that these bounds have a different nature, because
they use covariance information from the data.
It is particularly interest to know the tight bounds for the
the expected range and the expected maximum under mean-variance-covariance
information on the observations.
Non-tight bounds of this form are
given, e.g., in Aven (1985), Papadatos (2001a).
It is worth pointing out that
some sophisticated optimization techniques
(semidefinite programming)
have been fruitfully applied to this kind of problems, especially
for the maximum and the range.
The interested reader is referred to Natarajan and Teo (2014),
where some financial applications of the range bounds are also included.
However, note that one would hardly discover the simple formula (\ref{eq.7.3})
from the (reduced) semidefinite program in Natarajan and Teo's Section 4.
A lot of research has been devoted in deriving
distribution and expectation bounds for
$L$-statistics based on random vectors
with given marginals; see Arnold (1980, 1985, 1988), Caraux and Gascuel (1992),
Gascuel and Caraux (1992),
Meilijson and Nadas (1979), Papadatos (2001b), Rychlik (1992b,
1993a, 1994, 1995, 1998, 2007), Gajek and Rychlik (1996, 1998).
The results by Lai and Robbins (1976), Nagaraja (1981)
and Arnold and Balakrishnan (1989)
show that some deterministic
inequalities play an important role in
the derivation of tight bounds for $L$-statistics;
see Rychlik (1992a).
On the other hand, the deterministic inequality (\ref{eq.deterministic})
can be viewed as a range analogue of the inequality
from Lai and Robbins (1976). Noting that
the Lai-Robbins
inequality yields the tight bound for the expected
maximum under completely known marginal distributions
(see Bertsimas, Natarajan and Teo (2006),
Meilijson and Nadas (1979)), it would not be
surprising if (\ref{eq.deterministic}) could
produce
the best possible bound
for the expected range. Thus, a natural question is
whether it is true that for all multivariate vectors
with given marginal distributions $F_1,\ldots,F_n$
and finite
first moment,
\[
\mbox{$
\sup \E R_n =\inf_{c\in \R, \lambda>0}\big\{ \hspace{-.5ex}
-(n-2)\lambda+\frac{1}{2}\sum_{i=1}^n
\E\big[ \big|(X_i-c)-\lambda\big|+\big|(X_i-c)+\lambda\big|\big]\big\}.
$}
\]
Note that the RHS is an upper bound for the LHS, and
depends only on $F_1,\ldots,F_n$.
{
\begin{appendix}
\section{\normalsize Appendix: Proofs}
\label{app}
\small
\noindent
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.1}:}
Fix $c\in\R$ and $\lambda>0$ and set
$y_1=c-\lambda$, $y_2=c+\lambda$, so that $y_1<y_2$.
Observe that $R_n=X_{n:n}-X_{1:n}$ and
\[
\mbox{$
\sum_{i=1}^n \big\{|X_i-y_1|+|X_i-y_2|\big\}
=
\sum_{i=1}^n \big\{|X_{i:n}-y_1|+|X_{i:n}-y_2|\big\}.
$}
\]
Hence,
\begin{eqnarray*}
&&
\hspace{-15ex}
\mbox{$
\sum_{i=1}^n \big\{|X_i-y_1|+|X_i-y_2|\big\}
-(n-2)(y_2-y_1)-2R_n
$}
\\
&&
\mbox{$
=
\sum_{i=2}^{n-1}
\big\{|X_{i:n}-y_1|+|X_{i:n}-y_2|-(y_2-y_1)\big\}
$}
\\
&&
\hspace{5ex}
\mbox{$
+\big\{|X_{1:n}-y_1|+|X_{n:n}-y_1|-(X_{n:n}-X_{1:n})\big\}
$}
\\
&&
\hspace{5ex}
\mbox{$
+\big\{|X_{1:n}-y_2|+|X_{n:n}-y_2|-(X_{n:n}-X_{1:n})\big\}.
$}
\end{eqnarray*}
For each $i\in\{2,\ldots,n-1\}$ we have
\[
y_2-y_1=|y_2-y_1|=\big|(X_{i:n}-y_1)-(X_{i:n}-y_2)\big|\leq
\big|X_{i:n}-y_1\big|+\big|X_{i:n}-y_2\big|,
\]
with equality if and only if $y_1\leq X_{i:n}\leq y_2$.
Since the sum $\sum_{i=2}^{n-1}
\big\{|X_{i:n}-y_1|+|X_{i:n}-y_2|-(y_2-y_1)\big\}$
contains only non-negative terms,
it follows that
\[
\mbox{$
\sum_{i=2}^{n-1}
\big\{|X_{i:n}-y_1|+|X_{i:n}-y_2|-(y_2-y_1)\big\}\geq 0,
$}
\]
with equality if and only if $y_1\leq X_{2:n}\leq\cdots\leq X_{n-1:n}\leq y_2$.
Also, for $y=y_1$ or $y_2$,
\[
X_{n:n}-X_{1:n}=\big| (X_{n:n}-y)-(X_{1:n}-y)\big|\leq
\big| X_{1:n}-y\big|+
\big| X_{n:n}-y\big|
\]
with equality if and only if $X_{1:n}\leq y\leq X_{n:n}$.
Therefore,
\[
\mbox{$
-2R_n-(n-2)(y_2-y_1)+
\sum_{i=1}^n \big\{|X_i-y_1|+|X_i-y_2|\big\}\geq 0
$}
\]
with equality if and only if
$
X_{1:n}\leq y_1
\leq X_{2:n}\leq \cdots\leq X_{n-1:n}\leq y_2\leq X_{n:n}.
\ \ \
\Box
$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.2}:}
In case $\mu^2+\sigma^2\geq 4$
it suffices to use the inequality
\[
\mbox{$
|X-1|+|X+1|\leq \sqrt{\mu^2+\sigma^2}+\frac{X^2}{\sqrt{\mu^2+\sigma^2}},
$}
\]
where the equality holds if and only if
$X\in\{-\sqrt{\mu^2+\sigma^2},\sqrt{\mu^2+\sigma^2}\}$.
Taking expectations we get
\[
\mbox{$
\E\big\{|X-1|+|X+1|\big\}\leq
\sqrt{\mu^2+\sigma^2}+\frac{\E X^2}{\sqrt{\mu^2+\sigma^2}}
=2\sqrt{\mu^2+\sigma^2}.
$}
\]
For equality $X$ has to assume
the values $\pm\sqrt{\mu^2+\sigma^2}$.
Set $p=\Pr[X=\sqrt{\mu^2+\sigma^2}]$ so that $1-p=\Pr[X=-\sqrt{\mu^2+\sigma^2}]$.
The relation $\E X^2=\mu^2+\sigma^2$
is satisfied for any value of
$p\in[0,1]$, while the condition $\E X=\mu$ specifies $p$ to be as in (a).
Next, we assume that $2|\mu|<\mu^2+\sigma^2<4$ and use the inequality
\[
\mbox{$
|X-1|+|X+1|\leq 2+\frac{1}{2}X^2,
$}
\]
in which the equality holds if and only if $X\in\{-2,0,2\}$.
Taking expectations we again conclude (\ref{eq.in}) with
$U(\mu,\sigma)$ given by the second line of (\ref{eq.u}).
It is easy to see that the unique random variable in ${\cal F}_1(\mu,\sigma)$
that assumes values in the set $\{-2,0,2\}$ is the one given by (b).
Next, suppose that $\mu^2+\sigma^2\leq 2\mu$, and hence, $0<\mu<2$.
Working as before, it suffices to take expectations in the inequality
\[
\mbox{$
|X-1|+|X+1|\leq 2
+\frac{(X-1+\sqrt{(\mu-1)^2+\sigma^2})^2}{2\sqrt{(\mu-1)^2+\sigma^2}},
$}
\]
in which the equality holds if and only if $X\in\{x_1,x_2\}$,
where $x_1=1-\sqrt{(\mu-1)^2+\sigma^2}$, $x_2=1+\sqrt{(\mu-1)^2+\sigma^2}$.
Note that $0<(\mu-1)^2+\sigma^2=1-[2\mu-(\mu^2+\sigma^2)]\leq 1$; thus,
$0\leq x_1<1<x_2\leq 2$. Now it is easily seen that the unique
random variable in ${\cal F}_1(\mu,\sigma)$
that assumes values in the set $\{x_1,x_2\}$ is the one given by (c).
Observing that $|X-1|+|X+1|$ is even,
the case $\mu^2+\sigma^2\leq -2\mu$ is reduced
to the previous one by considering $-X\in{\cal F}_1(-\mu,\sigma)$.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.3}:}
For $n=1$ both (\ref{eq.bivariate}) and
(\ref{eq.condition}) are invalid, so we have nothing to prove.
For $n=2$ the result is trivial (we have uniqueness
if (\ref{eq.bivariate}) is satisfied; we have equality
in (\ref{eq.condition}) whenever it is fulfilled).
Assume $n\geq 3$
and consider the set of all probability matrices with the given marginals,
\[
\mbox{$
{\cal M}({\bpp},{\bqq})=\big\{Q=(q_{ij})\in\R^{n\times n}:q_{ij}\geq 0, \,
\sum_{i=1}^n q_{ij}=q_j, \, \sum_{j=1}^n q_{ij}=p_i \mbox{ for all } i,j\big\}.
$}
\]
The set ${\cal M}({\bpp},{\bqq})$ is nonempty since, e.g.,
it contains the matrix $Q=(p_i q_j)$. Also, the function
$f(Q):=\mbox{trace}(Q)=\sum_{i=1}^n q_{ii}$ is continuous with respect
to the total variation distance,
$d(Q,\widetilde{Q})=\sum_{i,j}|q_{ij}-\widetilde{q}_{ij}|$
(or any other equivalent metric on $\R^{n\times n}$).
Moreover, ${\cal M}({\bpp},{\bqq})$ is a compact subset of
$\R^{n\times n}$, since it is obviously closed, and it is
contained in a ball with center the null matrix $O_{n\times n}$
and (total variation) radius $1$.
It follows that $f(Q)$ attains its minimum value for some
$Q^*\in {\cal M}({\bpp},{\bqq})$.
Let $(X,Y)\sim Q^*=(q_{ij}^*)$ where $Q^*\in{\cal M}({\bpp},{\bqq})$ is
a minimizing matrix.
Then, $X\sim {\bpp}$, $Y\sim{\bqq}$ and $f(Q^*)=\Pr[X=Y]$.
A simple argument shows that
the principal diagonal of any minimizing matrix $Q^*$ can contain
at most one nonzero entry.
Indeed, if $q^*_{ii}>0$ and $q^*_{jj}>0$ with $i\neq j$,
set
$\gamma=\min\{q^*_{ii},q^*_{jj}\}>0$,
and consider the matrix $\widetilde{Q}=(\widetilde{q}_{ij})$
which differs from $Q^*$ only in the following four
entries:
$\widetilde{q}_{ii}=q^*_{ii}-\gamma$,
$\widetilde{q}_{jj}=q^*_{jj}-\gamma$,
$\widetilde{q}_{ij}=q^*_{ij}+\gamma$,
$\widetilde{q}_{ji}=q^*_{ji}+\gamma$.
Since the row/column sums are unaffected and the elements of
$\widetilde{Q}$ are nonnegative,
it is clear that $\widetilde{Q}\in{\cal M}({\bpp},{\bqq})$ and
we arrived at the contradiction
$f(\widetilde{Q})=f(Q^*)-2\gamma<f(Q^*)$.
Therefore, all diagonal entries of a minimizing matrix
$Q^*$ have to be zero,
with the possible exception of at most one of them.
{\it Sufficiency}:
Assume that (\ref{eq.condition}) is satisfied, and suppose that
$\min_{Q} f(Q)=f(Q^*)=\theta>0$.
Let $q^*_{kk}=\theta$ and thus, $q^*_{ii}=0$ for all $i\neq k$.
Then,
\[
\Pr[\{X=k\} \cup \{Y=k\}]=\Pr[X=k]+\Pr[Y=k]-\Pr[X=k,Y=k]=p_k+q_k-\theta.
\]
Since $1-p_k-q_k\geq 0$ (from (\ref{eq.condition})) we thus obtain
\[
\Pr[X\neq k, Y\neq k]=1-\Pr[\{X=k\} \cup \{Y=k\}]=\theta+(1-p_k-q_k)\geq
\theta>0.
\]
On the other hand, since $q^*_{ii}=0$ for all $i\neq k$, we have
\[
\Pr[X\neq k, Y\neq k]=\sum_{(i,j):\ i\neq k, j\neq k, i\neq j} q^*_{ij}.
\]
The above probability is at least $\theta$, and thus,
strictly positive.
It follows that the sum contains at least
one positive term. Hence, we can find
two indices $r,s$ with $r\neq k$,
$s\neq k$, $r\neq s$, such that $q^*_{rs}>0$.
Set
$\delta=\min\{\theta,q^*_{rs}\}>0$ and
consider the matrix $\widetilde{Q}=(\widetilde{q}_{ij})$
which differs from $Q^*$ only in the elements
$\widetilde{q}_{kk}=q^*_{kk}-\delta=\theta-\delta$,
$\widetilde{q}_{rs}=q^*_{rs}-\delta$,
$\widetilde{q}_{rk}=q^*_{rk}+\delta$,
$\widetilde{q}_{ks}=q^*_{ks}+\delta$.
Since the row/column sums are unaffected and the elements of
$\widetilde{Q}$ are nonnegative,
it is clear that $\widetilde{Q}\in{\cal M}({\bpp},{\bqq})$, and
this results to the contradiction
$f(\widetilde{Q})=\theta-\delta<\theta$.
Thus, $f(Q^*)=\Pr[X=Y]=0$; this proves the existence of random vectors
satisfying (\ref{eq.bivariate}).
{\it Necessity}: This is entirely obvious. For, if a random vector $(X,Y)$
satisfies (\ref{eq.bivariate}) then $(X,Y)\sim Q$ for some
$Q\in{\cal M}({\bpp},{\bqq})$ with $q_{ii}=0$ for all $i$.
Thus, for any $i$,
\[
p_i+q_i=p_i+q_i-q_{ii}=\Pr[X=i]+\Pr[Y=i]-q_{ii}=\Pr[\{X=i\}\cup\{Y=i\}]\leq 1.
\]
{\it Uniqueness}: Assume that $\max_{i}\big\{p_i+q_i\}=1$ and choose
$k$ with $p_k+q_k=1$. If $(X,Y)\sim Q$ satisfies (\ref{eq.bivariate}), we have
$\Pr[\{X= k\}\cup \{Y=k\}]=p_k+q_k-q_{kk}\geq p_k+q_k-\Pr[X=Y]=p_k+q_k=1$.
It follows that $Q$ can have non-zero entries only in its $k$-th row
and in its $k$-th column. Thus, $q_{ik}=p_i$ for all $i\neq k$,
$q_{kj}=q_j$ for all $j\neq k$ and $q_{ij}=0$ otherwise;
hence, $Q$ is uniquely determined from
${\bpp}$, ${\bqq}$. Note that $k$ need not be unique, but $Q$ is always unique.
For example, if ${\bpp}=(1-p,p,0,\ldots,0)$ and ${\bqq}=(p,1-p,0,\ldots,0)$
with $0\leq p\leq 1$, we obtain the unique solution to
(\ref{eq.bivariate}) as $\Pr[X=2,Y=1]=p=1-\Pr[X=1,Y=2]$. In fact,
one can easily verify that this example describes the most general case
(modulo the positions of $p, 1-p$) where the relation $p_k+q_k=1$ can hold for
more than one index $k$.
{\it Non-uniqueness}: Suppose that all $p_i$ and $q_i$ are positive
and that (\ref{eq.condition}) holds as a strict inequality,
that is,
$p_i+q_i<1$ for all $i$.
[The last assumption is possible only if
$n\geq 3$.]
Set $\beta=\frac{1}{n^2}\big[1-\max_i\big\{p_i+q_i\big\}\big]>0$,
$\delta=\min_{i,j}\{p_iq_j\}>0$ and
$\epsilon=\min\{\beta,\delta\}>0$. Define
\[
{\cal M}_{\epsilon}({\bpp},{\bqq}):=\big\{ Q\in {\cal M}({\bpp},{\bqq})
: q_{ij}\geq \epsilon \mbox{ for all } i,j \mbox{ with } i\neq j\big\}.
\]
Observe that
${\cal M}_{\epsilon}({\bpp},{\bqq})$ is a nonempty (since it contains $Q=(p_i q_j)$)
compact subject of $\R^{n\times n}$.
Applying the same arguments as in the beginning of the proof
we see that the continuous function $f(Q)=\mbox{trace}(Q)$
attains its minimum value at
a matrix $Q^*_{\epsilon}=(q_{ij}^*)\in{\cal M}_{\epsilon}({\bpp},{\bqq})$;
$Q^*_{\epsilon}$ has at most one nonzero diagonal entry while,
by the definition of ${\cal M}_{\epsilon}({\bpp},{\bqq})$, all off-diagonal
entries are at least $\epsilon$. Let $(X,Y)\sim Q^*_{\epsilon}$.
Assuming $\Pr[X=Y]=\theta>0$ we can find a unique index $k$ such that
$q^*_{kk}=\theta$;
then, $\Pr[\{X=k\}\cup \{Y=k\}]=p_k+q_k-\theta$. Since
$q_{ii}^*=0$ for $i\neq k$, we have
\begin{eqnarray*}
&&
\sum_{(i,j):\ i\neq k,j\neq k,i\neq j}q_{ij}^*=\Pr[X\neq k,Y\neq k]=\theta+(1-p_k-q_k)
\geq \theta+\big[1-\max_{i}\big\{p_i+q_i\big\}\big] \\
&&
\hspace{16ex}
=\theta+n^2 \beta \geq \theta +n^2\epsilon>n^2\epsilon.
\end{eqnarray*}
This sum contains $(n-1)(n-2)<n^2$ terms and
the inequality shows that at least one of them
is greater than $\epsilon$.
Thus, we can find
two indices $r,s$ with $r\neq k$,
$s\neq k$, $r\neq s$, such that $q^*_{rs}>\epsilon$;
say $q^*_{rs}=\epsilon+\gamma$ with $\gamma>0$.
Set
$\lambda=\min\{\theta,\gamma\}>0$ and
consider the matrix $\widetilde{Q}_{\epsilon}=(\widetilde{q}_{ij})$,
which differs from $Q^*_{\epsilon}$ at exactly
the four elements
$\widetilde{q}_{kk}=q^*_{kk}-\lambda=\theta-\lambda\geq 0$,
$\widetilde{q}_{rs}=q^*_{rs}-\lambda=\epsilon+(\gamma-\lambda)\geq \epsilon$,
$\widetilde{q}_{rk}=q^*_{rk}+\lambda$,
$\widetilde{q}_{ks}=q^*_{ks}+\lambda$.
It is clear that $\widetilde{Q}_{\epsilon}
\in{\cal M}_{\epsilon}({\bpp},{\bqq})$ and, once again,
it contradicts the definition of $Q_{\epsilon}^*$:
$f(\widetilde{Q}_{\epsilon})=\theta-\lambda<\theta=f(Q_{\epsilon}^*)$.
Thus, $f(Q^*_{\epsilon})=\Pr[X=Y]=0$. This shows the existence
of random vectors $(X,Y)$
satisfying (\ref{eq.bivariate}) with the additional
property $\Pr[X=i,Y=j]\geq \epsilon>0$ for all $i\neq j$, provided that
$\epsilon>0$ is sufficiently small. Given a probability matrix
$Q^*_{\epsilon}=(q_{ij}^*)$ of this form, it is easy to construct
a second solution, $Q=(q_{ij})$, to (\ref{eq.bivariate}); e.g., set
$q_{12}=q_{12}^*-\epsilon/2$, $q_{13}=q_{13}^*+\epsilon/2$,
$q_{21}=q_{21}^*+\epsilon/2$, $q_{23}=q_{23}^*-\epsilon/2$,
$q_{31}=q_{31}^*-\epsilon/2$, $q_{32}=q_{32}^*+\epsilon/2$,
and leave the rest entries unchanged.
Finally, it is easy to see that if $Q_0$, $Q_1$ both solve
(\ref{eq.bivariate}), the same is true for
$Q_t=tQ_1+(1-t)Q_0$, $0\leq t\leq 1$, and the proof is complete.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Theorem \ref{theo.AG}:}
Assume that $\E R_n=AG_n$ for some random vector
${\bXX}$ with
$\E{\bXX}=\bmu$ and $\Var{\bXX}=\bsigma^2$.
Set $c=\overline{\mu}$,
$\lambda=\frac{1}{4}AG_n>0$
and take expectations in (\ref{eq.deterministic})
to get (cf.\ Remark \ref{rem.2})
\begin{eqnarray*}
\mbox{$
AG_n=\E R_n \leq
\frac{-(n-2)AG_n}{4}+
\frac{AG_n}{8}\sum_{i=1}^{n}
\E
\Big\{\Big|\frac{X_i-\overline{\mu}}{AG_n/4}-1\Big|
+ \Big|\frac{X_i-\overline{\mu}}{AG_n/4}+1\Big|
\Big\}.
$}
\end{eqnarray*}
Next, from $|y-1|+|y+1|\leq 2 +\frac{1}{2}y^2$ with equality if and only
if $y\in\{-2,0,2\}$ we get
\[
\mbox{$
\sum_{i=1}^{n}
\E
\Big\{\Big|\frac{X_i-\overline{\mu}}{AG_n/4}-1\Big|
+ \Big|\frac{X_i-\overline{\mu}}{AG_n/4}+1\Big|
\Big\}\leq 2n+\frac{1}{2}\sum_{i=1}^n \E
\Big\{ \Big(\frac{X_i-\overline{\mu}}{AG_n/4}\Big)^2
\Big\}=2n+4.
$}
\]
Since $\frac{-(n-2)AG_n}{4}+\frac{AG_n}{8} (2n+4)=AG_n$,
it follows that the preceding inequalities are, in fact, equalities.
Therefore, $\E R_n=AG_n$ is equivalent to
(\ref{eq.equality})
(with $c=\overline{\mu}$, $\lambda=\frac{1}{4}AG_n$) and
$\frac{X_i-\overline{\mu}}{AG_n/4} \in\{-2,0,2\}$,
$i=1,\ldots,n$
(of course, it suffices to hold with probability $1$).
Hence, $\E R_n=AG_n$ if and only if
\be
\label{3.7}
\begin{array}{ll}
\mbox{(a)} &X_{1:n}\leq \overline{\mu}-\frac{1}{4}AG_n\leq X_{2:n}\leq
\cdots\leq X_{n-1:n}\leq \overline{\mu}+\frac{1}{4}AG_n\leq X_{n:n}
\vspace{-1ex}
\\
\hspace{-4ex}\mbox{and}
\vspace{-1ex}
&
\\
\mbox{(b)} &
X_i\in\big\{\overline{\mu}-\frac{AG_n}{2},\overline{\mu},
\overline{\mu}+\frac{AG_n}{2}\big\},
\ \
i=1,\ldots,n,
\end{array}
\ee
with probability $1$.
Therefore, the (essential) support of any extremal random vector is a subset of
\[
\mbox{$
S:=\big\{\big(\overline{\mu},\ldots,\overline{\mu},\overline{\mu}+\frac{AG_n}{2},
\overline{\mu},\ldots, \overline{\mu},
\overline{\mu}-\frac{AG_n}{2},\overline{\mu},\ldots,\overline{\mu}\big)\big\},
$}
\]
where the plus and minus signs can appear at any two (different) places.
Clearly,
$S$ has $n(n-1)$ elements and can be written as
\[
\mbox{$
S= \big\{\overline{\mu}\, {\bone} +\frac{{\bee}(i)-{\bee}(j)}{2}
AG_n:(i,j)\in\{1,\ldots,n\}^2, i\neq j\big\}.
$}
\]
Let $S':=\big\{(i,j) \in \{1,\ldots,n\}^2:i\neq j\big\}$.
The function $g:S'\to S$,
that sends $(i,j)$ to
$g(i,j)=\overline{{\mu}}\, {\bone} +\frac{{\bee}(i)-{\bee}(j)}{2}
AG_n$, is a bijection. It follows that $(X,Y):=g^{-1}({\bXX})$
is a random pair with values in a subset of $S'$, and
${\bXX}=g(X,Y)$; this verifies the representation (\ref{eq.extremal}).
For $i\in\{1,\ldots,n\}$ we set
\[
\mbox{$
p_i^+:=\Pr[X_i=\overline{\mu}+\frac{AG_n}{2}]=\Pr[X=i], \ \
p_i^-=\Pr[X_i=\overline{\mu}-\frac{AG_n}{2}]=\Pr[Y=i],
$}
\]
so that $\Pr[X_i=\overline{\mu}]=1-p_i^+-p_i^-$. From
$\E X_i=\mu_i$ we get $p_i^+-p_i^-=\frac{2(\mu_i-\overline{\mu})}{AG_n}$ and
from $\E\big\{ (X_i-\overline{\mu})^2\big\}=(\mu_i-\overline{\mu})^2+\sigma_i^2$
we obtain $p_i^++p_i^-=\frac{4[(\mu_i-\overline{\mu})^2+\sigma_i^2]}{AG_n^2}$.
Hence,
\[
\mbox{$
p_i^+=\frac{2[(\mu_i-\overline{\mu})^2+\sigma_i^2]+(\mu_i-\overline{\mu})AG_n}{AG_n^2},
\ \ \ \
p_i^-=\frac{2[(\mu_i-\overline{\mu})^2+\sigma_i^2]-(\mu_i-\overline{\mu})AG_n}{AG_n^2},
$}
\]
and (\ref{eq.marginals}) follows.
Therefore, we can find a random vector ${\bXX}$ with $\E {\bXX}=\bmu$,
$\Var{\bXX}={\bsigma}^2$ and $\E R_n=AG_n$ if and only if
the above construction of a random pair $(X,Y)$, with $\Pr[X=Y]=0$, is possible.
According to Lemma \ref{lem.3}, this is equivalent to
$\max_{i}\big\{p_i^++p_i^-\big\}\leq 1$, which gives
(\ref{eq.3.1})(ii) (it also guarantees
that $\Pr[X_i=\overline{\mu}]=1-p_i^+-p_i^-\geq 0$),
while (\ref{eq.3.1})(i) follows from
$p_i^+\geq 0$ and $p_i^-\geq 0$.
Finally, the inequalities
(\ref{eq.3.1}) are strict for all $i$ if and only if
$p_i^++p_i^-<1$, $p_i^+>0$ and $p_i^->0$ for all $i$. Lemma
\ref{lem.3} shows that there exist infinitely many vectors
$(X,Y)$
in this case.
Also, if (\ref{eq.3.1}) is satisfied and we have equality in
(\ref{eq.3.1})(ii) for some $i$, uniqueness follows again
from Lemma \ref{lem.3}.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.4.1}:}
The functions $f_i:T\to(0,\infty)$ ($i=1,2,3,4$) given by
$f_1(x,y):=2\sqrt{x^2+y^2}$,
$f_2(x,y):=2+\frac{1}{2}(x^2+y^2)$,
$f_3(x,y):=x+1+\sqrt{(x-1)^2+y^2}$ and
$f_4(x,y):=1-x+\sqrt{(x+1)^2+y^2}$ are
obviously $C^{\infty}(T)$. The function
$U$ can be defined as the restriction of
$f_1$ in $A_1:=\{(x,y)\in T: x^2+y^2\geq 4\}$,
of $f_2$ in $A_2:=\{(x,y)\in T:2|x|\leq x^2+y^2\leq 4\}$,
of
$f_3$ in $A_3:=\{(x,y)\in T: x^2+y^2\leq 2x\}$
and of $f_4$ in $A_4:=\{(x,y)\in T: x^2+y^2\leq -2x\}$.
Observe that $A_3$ and $A_4$ are the closed (with respect to $T$)
semidisks
$T\cap D((1,0),1)$, $T\cap D((-1,0),1)$;
also, $A_2=T\cap [D((0,0),2)\smallsetminus A_3^{o} \cup A_4^{o}]$,
and $A_1=T\smallsetminus A_2^{o}\cup A_3\cup A_4$.
Therefore, $A_1\cap A_3=\emptyset$,
$A_1\cap A_4=\emptyset$, $A_3\cap A_4=\emptyset$,
$\partial A_1=A_1\cap A_2=\{(x,y)\in T:x^2+y^2=4\}$,
$\partial A_3=A_2\cap A_3=\{(x,y)\in T:(x-1)^2+y^2=1\}$,
$\partial A_4=A_2\cap A_4=\{(x,y)\in T:(x+1)^2+y^2=1\}$ and
$\partial A_2=\partial A_1 \cup \partial A_3\cup \partial A_4$.
It is easy to check that both partial derivatives of
$f_1$ and $f_2$ coincide at $\partial A_1$, that both partial
derivatives of $f_2$ and $f_3$ coincide at $\partial A_3$
and that both partial derivatives
of $f_2$ and $f_4$ coincide at $\partial A_4$.
We conclude that for $(x,y)\in T$,
\be
\label{eq.4.1}
U_1(x,y):=\frac{\partial}{\partial x} U(x,y)=\left\{
\begin{array}{cll}
\frac{x}{\sqrt{x^2+y^2}},
&
\mbox{ if }
& x^2+y^2\geq 4,
\\
x,
&
\mbox{ if }
& 2|x|\leq x^2+y^2\leq 4,
\\
\frac{x-1}{\sqrt{(x-1)^2+y^2}}+1,
&
\mbox{ if }
&
(x-1)^2+y^2\leq 1,
\\
\frac{x+1}{\sqrt{(x+1)^2+y^2}}-1,
&
\mbox{ if }
& (x+1)^2+y^2\leq 1,
\end{array}
\right.
\ee
and
\be
\label{eq.4.2}
U_2(x,y):=\frac{\partial}{\partial y} U(x,y)=\left\{
\begin{array}{cll}
\frac{y}{\sqrt{x^2+y^2}},
&
\mbox{ if }
& x^2+y^2\geq 4,
\\
y,
&
\mbox{ if }
&
2|x|\leq x^2+y^2\leq 4,
\\
\frac{y}{\sqrt{(x-1)^2+y^2}},
&
\mbox{ if }
&
(x-1)^2+y^2\leq 1,
\\
\frac{y}{\sqrt{(x+1)^2+y^2}},
&
\mbox{ if }
&
(x+1)^2+y^2\leq 1,
\end{array}
\right.
\ee
and the above functions are obviously continuous.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Proposition \ref{prop.4.1}:}
Fix ${\bxx}$ and ${\byy}$ in $T$.
The set $\partial A_2$ (where $U$ changes type) is a union of three disjoint
semicircles, and the line segment $[{\bxx},{\byy}]=
\{{\bxx}+t({\byy}-{\bxx})$, $0\leq t\leq 1\}$
can have at most six common points with
$\partial A_2=\{(x,y)\in T:x^2+y^2=4 \mbox { or } (x-1)^2+y^2=1 \mbox
{ or } (x+1)^2+y^2=1\}$;
for the definition of $A_2$ see the proof of Lemma
\ref{lem.4.1}.
Consider now the function $g:[0,1]\to\R$ with
$g(t):=U({\bxx}+t({\byy}-{\bxx}))$, $0\leq t\leq 1$,
which is continuously differentiable from Lemma \ref{lem.4.1}.
Also, $g$ is of the form (\ref{eq.4.4}) with $k\in\{0,\ldots,6\}$,
where $g_i(t)=f_j({\bxx}+t({\byy}-{\bxx}))$, $0\leq t\leq 1$,
for some
$j=j(i)\in\{1,2,3,4\}$ (the functions $f_j:T\to (0,\infty)$
are defined in the proof of Lemma \ref{lem.4.1}).
It is easy to verify that each $f_j$ has
nonnegative definite Hessian matrix and, thus, is convex.
Lemma \ref{lem.4.2} asserts that
$g_i(t):[0,1]\to (0,\infty)$ ($i=1,\ldots,k+1$)
is convex.
Since $g$ is continuously differentiable, (\ref{eq.4.3}) is automatically satisfied,
and
we conclude from Lemma \ref{lem.4.3} that $g$ is convex.
Therefore, $g$ is convex for any choice of ${\bxx}$ and ${\byy}$ in $T$,
and a final application of Lemma \ref{lem.4.2} completes the proof.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.4.4}:}
(i) Fix $x_0\in\R$, $y_0>0$ and let
$\alpha\in(0,1)$, $c_1,c_2\in\R$, $\lambda_1,\lambda_2>0$.
Write $\beta_1=\frac{\alpha \lambda_1}
{\alpha \lambda_1+(1-\alpha)\lambda_2}>0$,
$\beta_2=\frac{(1-\alpha) \lambda_2}
{\alpha \lambda_1+(1-\alpha)\lambda_2}>0$,
so that $\beta_1+\beta_2=1$. We have
\begin{eqnarray*}
&&
\hspace{-3ex}
\mbox{$
\frac{h\big(\alpha c_1+(1-\alpha) c_2,\alpha \lambda_1+(1-\alpha) \lambda_2\big)}
{\alpha \lambda_1+(1-\alpha)\lambda_2}
=f\big(\frac{x_0-[\alpha c_1+(1-\alpha) c_2]}
{\alpha \lambda_1+(1-\alpha)\lambda_2},\frac{y_0}{\alpha \lambda_1+(1-\alpha)\lambda_2}\big)
$}
\\
&&
\hspace{5ex}
\mbox{$
=f\big(\beta_1\big(\frac{x_0- c_1}
{\lambda_1}\big)
+\beta_2 \big(\frac{x_0- c_2}
{\lambda_2}\big)
,
\beta_1\big(\frac{y_0}
{\lambda_1}\big)
+\beta_2 \big(\frac{y_0}
{\lambda_2}\big)\big)
$}
\\
&&
\mbox{$
\hspace{5ex}
\leq \beta_1 f\big(\frac{x_0- c_1}{\lambda_1},\frac{y_0}{\lambda_1}\big)
+\beta_2 f\big(\frac{x_0- c_2}{\lambda_2},\frac{y_0}
{\lambda_2}\big)
=\frac{\alpha h(c_1,\lambda_1)+(1-\alpha) h(c_2,\lambda_2)}
{\alpha \lambda_1+(1-\alpha)\lambda_2},
$}
\end{eqnarray*}
showing that $h$ is convex.
\noindent
(ii) Suppose that for a particular $(x_0,y_0)\in T$, the function
$h_0(c,\lambda)=\lambda f\big(\frac{x_0-c}{\lambda},\frac{y_0}{\lambda}\big)$
is convex. Set $x=\frac{x_0-c}{\lambda}$, $y=\frac{y_0}{\lambda}>0$, so that
\[
c=x_0-y_0\frac{x}{y}, \ \ \lambda=\frac{y_0}{y}, \
\
\
y_0f(x,y)=y\, h_0\Big(x-y_0\frac{x}{y}, \frac{y_0}{y}\Big),
\ (x,y)\in T.
\]
Let $\alpha\in(0,1)$, $x_1,x_2\in\R$ and $y_1,y_2>0$.
Let us now write
$\beta_1=\frac{\alpha y_1}{\alpha y_1+(1-\alpha) y_2}>0$,
$\beta_2=\frac{(1-\alpha) y_2}{\alpha y_1+(1-\alpha) y_2}>0$,
so that $\beta_1+\beta_2=1$.
It follows that
\begin{eqnarray*}
&&
\hspace{-5ex}
y_0 f\big(\alpha x_1+(1-\alpha)x_2,\alpha y_1+(1-\alpha)y_2\big)
\\
&&
=
\big[\alpha y_1+(1-\alpha)y_2\big] \,
h_0\Big(\beta_1 \Big(x_1-y_0\frac{x_1}{y_1}\Big)+\beta_2 \Big(x_2-y_0\frac{x_2}{y_2}\Big),
\beta_1 \Big(\frac{y_0}{y_1}\Big)+\beta_2 \Big(\frac{y_0}{y_2}\Big)
\Big)
\\
&&
\leq
\big[\alpha y_1+(1-\alpha)y_2\big] \,
\Big\{
\beta_1 h_0\Big(x_1-y_0\frac{x_1}{y_1},\frac{y_0}{y_1}\Big)
+
\beta_2 h_0\Big(x_2-y_0\frac{x_2}{y_2},\frac{y_0}{y_2}\Big)
\Big\}
\\
&&
=\alpha y_1 h_0\Big(x_1-y_0\frac{x_1}{y_1},\frac{y_0}{y_1}\Big)
+(1-\alpha)y_2 h_0\Big(x_2-y_0\frac{x_2}{y_2},\frac{y_0}{y_2}\Big)
\\
&&
=y_0 \, \big[\alpha f(x_1,y_1)+(1-\alpha)f(x_2,y_2)\big],
\end{eqnarray*}
and the proof is complete.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.5.2}:}
If $(c_0,\lambda_0)\in T_0$ then,
by Proposition \ref{prop.5.1}, $\phi_n(c,\lambda)\geq \phi_n(c_0,\lambda_0)$
for all $(c,\lambda)\in T$. On the other hand, for this $c_0$
we can define the function $\psi_n(\lambda)=\phi_n(c_0,\lambda)$;
by Lemma \ref{lem.5.1}, the function $\psi_n(\lambda)$
is minimized at a unique $\lambda=\lambda_1=\lambda_1(c_0)$.
Thus,
\[
\psi_n(\lambda_0)=\phi_n(c_0,\lambda_0)\leq \phi_n(c_0,\lambda_1)
=\psi_n(\lambda_1)\leq \psi_n(\lambda_0);
\]
the first inequality follows from $(c_0,\lambda_0)\in T_0$
and the second from the definition of $\lambda_1$. Therefore,
$\psi_n(\lambda_0)=\psi_n(\lambda_1)$, so that
$\lambda=\lambda_0$ is a minimizing point for $\psi_n(\lambda)$.
By uniqueness, $\lambda_1=\lambda_0$.
Thus, $\lambda_0=\lambda_1(c_0)$, where
$\lambda_1(\cdot):\R\to(0,\infty)$ is a well-defined function;
it is described (implicitly) in Lemma \ref{lem.5.1}.
Hence, if $(c_0,\lambda_0)\neq (c_2,\lambda_2)$
are any two points in $T_0$
then $c_0\neq c_2$; indeed, $c_0=c_2$ implies
$\lambda_0=\lambda_1(c_0)=\lambda_1(c_2)=\lambda_2$,
contradicting the assumption $(c_0,\lambda_0)\neq (c_2,\lambda_2)$.
Let $L$ be the straight line that passes through the points
$(c_0,\lambda_0)$ and $(c_2,\lambda_2)$.
We now verify that if $(c_3,\lambda_3)\in T_0$ then
$(c_3,\lambda_3)\in L$. Indeed, if $(c_3,\lambda_3)\in
T_0\smallsetminus L$ then the convex hull $B$ of the
triangle
$\{(c_0,\lambda_0)$, $(c_2,\lambda_2)$, $(c_3,\lambda_3)\}$
must be a subject of $T_0$, because $T_0$ is convex.
Since, however, $(c_3,\lambda_3)\notin L$, the set
$B$ contains a line segment of positive length, parallel to the
$\lambda$-axis and, by the previous argument,
this is impossible.
It follows that $T_0\subseteq L\cap T$, and since $T_0$
is compact and convex, it must be a compact line segment.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.5.3}:}
By assumption, $A$ is moving linearly in the line
segment $[A_0,A_1]$ from $A_0$ to $A_1$, thus we may write
$A=A(t):=(c(t), \lambda(t))$ where
$c(t)=c_0+t(c_1-c_0)$, $\lambda(t)=\lambda_0+t(\lambda_1-\lambda_0)$,
$0\leq t\leq 1$.
Then $B=B(t)=\big(\frac{\mu-c(t)}{\lambda(t)},\frac{\sigma}{\lambda(t)}\big)$,
so that $B(0)=B_0$, $B(1)=B_1$ and $B(t)$ is continuous in $t$.
It follows that for all $t\in[0,1]$,
\[
\det[B_0, B(t), B_1]:=
\left|
\begin{array}{ccc}
\frac{\mu-c_0}{\lambda_0} & \frac{\sigma}{\lambda_0} & 1 \\
\frac{\mu-c(t)}{\lambda(t)} & \frac{\sigma}{\lambda(t)} & 1 \\
\frac{\mu-c_1}{\lambda_1} & \frac{\sigma}{\lambda_1} & 1
\end{array}
\right|=
\frac{\sigma}{\lambda_0\lambda(t)\lambda_1}
\left|
\begin{array}{ccc}
c_0 & \lambda_0 & 1 \\
c(t) & \lambda(t) & 1 \\
c_1 & \lambda_1 & 1
\end{array}
\right|=0.
\ \ \Box
\]
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Theorem \ref{theo.5.1}:}
According to Proposition \ref{prop.5.1}, it remains to
verify that $T_0$ in (\ref{eq.5.2}) is a singleton.
Assume, in contrary, that $T_0$ contains two points
$(c_0,\lambda_0)\neq (c_1,\lambda_1)$.
From Lemma \ref{lem.5.2} we know that $c_0\neq c_1$,
and that all points
$(c,\lambda)\in T_0$ can be written
as $(c,\lambda)=(c,\alpha c+\beta)$, $c_2\leq c\leq c_3$,
for some $\alpha,\beta, c_2,c_3\in\R$ with $c_2<c_3$. Therefore,
we can write
$\lambda(c)=\alpha c+\beta$, $c_2\leq c\leq c_3$, and
\[
T_0=\{(c,\alpha c+\beta), \ c_2\leq c\leq c_3\},
\ \ \alpha, \beta, c_2, c_3\in\R, \
c_2<c_3.
\]
Note that the parameters $\alpha$, $\beta$, $c_2$, $c_3$ have to fulfill
additional restrictions so that $\lambda(c)>0$ for all $c\in[c_2,c_3]$;
namely, $\alpha c_2+\beta>0$ and $\alpha c_3+\beta>0$.
Consider now the points $A(c):=(c,\lambda(c))$ and
$B_i(c):=\big(\frac{\mu_i-c}{\lambda(c)},\frac{\sigma_i}{\lambda(c)}\big)$,
$i=1,\ldots,n$, $c_2\leq c\leq c_3$. As $c$ varies in $[c_2,c_3]$,
the point $A=A(c)$ is moved from $A(c_2)$ to $A(c_3)$,
generating the
line
segment $[A(c_2),A(c_3)]=T_0\subset T$.
It follows from Lemma \ref{lem.5.3}
that each point $B_i=B_i(c)$, $i=1,\ldots,n$, produces
a line segment too; that is, $B_i$ generates its corresponding
segment $L_i:=[B_i(c_2), B_i(c_3)]\subset T$.
Consider now the region
$A_2=\{(x,y)\in T : 2|x|\leq x^2+y^2\leq 4\}\subset T$.
The function $U(x,y)$ (see (\ref{eq.u})) changes types (and it is
not even $C^{2}$)
only at the boundary points of $A_2$,
i.e., at those
$(x,y)\in T$ that belong to the set
\[
C :=\{x^2+y^2=4\}\cup \{(x-1)^2+y^2=1\}\cup \{(x+1)^2+y^2=1\}\subset \R^2.
\]
The set $\partial A_2=C\cap T$ is a union of
three (disjoint) semicircles, and thus, any line segment
can have
at most six common points with it.
It follows that only of finite number of points
of the set $\cup_{i=1}^n L_i=\cup_{i=1}^n\cup_{c_2\leq c \leq c_3} B_i(c)$
can intersect $\partial A_2$. Let $\Gamma_1,\ldots, \Gamma_k$
be all these points.
Each $\Gamma_j$ belongs to some $L_i$;
that is, for any $j\in\{1,\ldots,k\}$
we can find an index $i=i(j)\in\{1,\ldots,n\}$,
and then a unique number $t=t_{ij}\in[c_2,c_3]$ such that
$B_i(t)=\Gamma_j$.
Clearly, for a particular index $j$,
the maximal number of different $t$'s that can be
found (satisfying $B_i(t)=\Gamma_j$ for some $i$)
is $n$, because $B_i(t_1)\neq B_i(t_2)$ if $t_1\neq t_2$.
Therefore, the set
\[
N:
=
\{t\in[c_2,c_3]: B_i(t)=\Gamma_j \mbox{ for some } i \mbox{ and } j\}
\]
is finite, say $N=\{t_1,\ldots,t_m\}$ with $c_2\leq t_1<\cdots<t_m\leq c_3$.
Fix now an interval $[t,s]\subseteq (c_2,c_3)$, of positive length,
such that $[t,s]\cap N=\emptyset$.
Since $[t,s]$ has no common points with $N$,
it is clear that the line segment $J_i:=[B_i(t),B_i(s)]\subseteq L_i$ does not
intersect $\partial A_2$, and this is true for all $i\in\{1,\ldots, n\}$.
In this way we obtain a subset $T_1$ of $T_0$, namely
\[
T_1:=\{(c,\alpha c+\beta), \ t\leq c\leq s\},
\ \
\mbox{ with } c_2<t<s<c_3.
\]
The boundary of $A_2$
divides $T$ into four disjoint open regions, namely
\[
\begin{array}{ll}
G_1:=\{(x,y)\in T: x^2+y^2>4\}, & G_2:=\{(x,y)\in T: 2|x|<x^2+y^2<4\},
\\
G_3:=\{(x,y)\in T: (x-1)^2+y^2<1\}, & G_4:=\{(x,y)\in T: (x+1)^2+y^2<1\}.
\end{array}
\]
Compared to $T_0$, the set $T_1$ has the additional
property that, as $c$ varies, every line segment
$\{B_i(c), t\leq c\leq s\}$ stays in the same open region.
This means that the sets of indices $I_1$, $I_2$, $I_3$, $I_4$,
defined in Remark \ref{rem.5.1}, do not depend on
$c$.
Recall that
\begin{eqnarray*}
&&
B_i(c)\in G_1
\Leftrightarrow
(\mu_i-c)^2+\sigma_i^2 >4\lambda^2
\Rightarrow i\in I_1,
\\
&&
B_i(c)\in G_2
\Leftrightarrow
2\lambda|\mu_i-c|<(\mu_i-c)^2+\sigma_i^2 <4\lambda^2
\Rightarrow i\in I_2,
\\
&&
B_i(c)\in G_3
\Leftrightarrow
(\mu_i-c)^2+\sigma_i^2 <2\lambda (\mu_i-c)
\Rightarrow i\in I_3,
\\
&&
B_i(c)\in G_4
\Leftrightarrow
(\mu_i-c)^2+\sigma_i^2 <-2\lambda (\mu_i-c)
\Rightarrow i\in I_4,
\end{eqnarray*}
where $\lambda=\lambda(c)=\alpha c+\beta$.
Consider now the function $g_n:(t,s)\to\R$ with
\[
g_n(c):=\phi_n(c,\lambda(c))=\phi_n(c,\alpha c+\beta), \ \ t<c<s.
\]
The explicit form of $g_n$ is quite complicated:
\begin{eqnarray*}
g_n(c)&=&
\mbox{$
-(n-2)(\alpha c+\beta) + \sum_{i\in I_1}\sqrt{(\mu_i-c)^2+\sigma_i^2}
$}
\\
&&
\mbox{$
+
\sum_{i\in I_2} \Big\{(\alpha c+\beta)
+\frac{1}{4(\alpha c+\beta)}
\big[(\mu_i-c)^2+\sigma_i^2\big]
\Big\}
$}
\\
&&
\mbox{$
+
\sum_{i\in I_3}
\frac{1}{2}\Big\{
\mu_i-c+(\alpha c+\beta)+
\sqrt{\big[\mu_i-c-(\alpha c+\beta)\big]^2+\sigma_i^2}
\Big\}
$}
\\
&&
\mbox{$
+
\sum_{i\in I_4}
\frac{1}{2}\Big\{
c-\mu_i+(\alpha c+\beta)+
\sqrt{\big[c-\mu_i-(\alpha c+\beta)\big]^2+\sigma_i^2}
\Big\}.
$}
\end{eqnarray*}
Since, however, the sets $I_j$ do not depend on $c$, it is obvious that
$g_n\in C^{\infty}(t,s)$. By assumption,
$(c,\lambda(c))$ minimizes $\phi_n(c,\lambda)$ for all $c\in(t,s)$, and this
means that $g_n(c)$ is constant, implying that
$g_n''(c)=0$, $t<c<s$.
A straightforward computation
shows that for all $c\in(t,s)$,
\begin{eqnarray*}
g''_n(c)&=&
\mbox{$
\sum_{i\in I_1}\frac{\sigma_i^2}{\left[(\mu_i-c)^2+\sigma_i^2\right]^{3/2}}
+\frac{1}{2\left[\lambda(c)\right]^3}
\sum_{i\in I_2}
\Big\{
\alpha^2 \sigma_i^2 +(\alpha \mu_i+\beta)^2
\Big\}
$}
\\
&&
\mbox{$
+
\frac{(\alpha+1)^2}{2}
\sum_{i\in I_3}
\frac{\sigma_i^2}{\left[\left(\beta+(\alpha+1)c-\mu_i\right)^2
+\sigma_i^2\right]^{3/2}}
+
\frac{(\alpha-1)^2}{2}
\sum_{i\in I_4}
\frac{\sigma_i^2}{\left[\left(\beta+(\alpha-1)c+\mu_i\right)^2
+\sigma_i^2\right]^{3/2}}.
$}
\end{eqnarray*}
Obviously, all summands are nonnegative.
If $\alpha\neq 0$, the only two possibilities
which are compatible with $g_n''(c)=0$ are the following:
(i) either $I_1=I_2=I_4=\emptyset$ (and thus, $N(I_3)=n$)
and $\alpha=-1$ or
(ii) $I_1=I_2=I_3=\emptyset$ (and $N(I_4)=n$) and $\alpha=1$.
However, because of (\ref{eq.5.7}), neither (i) nor (ii)
is allowed for a minimizing point $(c,\lambda)$, and in particular
for $(c,\lambda(c))$. Finally, if $\alpha=0$ then we must have
$I_1=I_3=I_4=\emptyset$ and, therefore, $N(I_2)=n$. The
condition
$\lambda(c)>0$ now yields $\beta>0$; thus,
$g_n''(c)=\frac{n}{2\beta}>0$
and $g_n(c)$ could not be a
constant function in the interval $t<c<s$.
The resulting contradiction implies that the set $T_0$ cannot contain
two
distinct elements, and the proof is complete.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Lemma \ref{lem.6.1}:}
From (\ref{eq.phi}), (\ref{eq.4.1}),
(\ref{eq.4.2}), and in view of (\ref{eq.6.3}), (\ref{eq.6.4}),
\begin{eqnarray*}
\mbox{$
\frac{\partial}{\partial \lambda} \phi_n(c,\lambda)
$}
\hspace{-1ex}
&=&
\hspace{-1ex}
\mbox{$
-(n-2)
+
\frac12
\sum_{i\in I_2} \Big\{
2-\frac{1}{2\lambda^2}
\big[(\mu_i-c)^2+\sigma_i^2\big]
\Big\}
$}
\\
&&
\mbox{$
+
\frac12
\sum_{i\in I_3}
\Big\{
1-\frac{\mu_i-c-\lambda}{\sqrt{\left(\mu_i-c-\lambda\right)^2+\sigma_i^2}}
\Big\}
+
\frac12
\sum_{i\in I_4}
\Big\{
1-\frac{c-\mu_i-\lambda}{\sqrt{\left(c-\mu_i-\lambda\right)^2+\sigma_i^2}}
\Big\}
$}
\\
\hspace{-1ex}
&=&
\hspace{-1ex}
-(n-2)
+
\sum_{i\in I_2\cup I_3\cup I_4} p_i^{o}.
\end{eqnarray*}
Since $\frac{\partial}{\partial\lambda}\phi_n(c,\lambda)=0$ and
$p_i^{o}=0$ for $i\in I_1$, it follows that
$\sum_{i=1}^n p_i^{o} =n-2$.
Taking into account the fact that
$p_i^{o}=1-p_i^+-p_i^-$, we obtain
$\sum_{i=1}^n p_i^- + \sum_{i=1}^n p_i^+ =2.$
Similarly, we have
\begin{eqnarray*}
\mbox{$
\frac{\partial}{\partial c} \phi_n(c,\lambda)
$}
\hspace{-1ex}
&=&
\hspace{-1ex}
\mbox{$
-\sum_{i\in I_1} \frac{\mu_i-c}{\sqrt{(\mu_i-c)^2+\sigma_i^2}}
-
\sum_{i\in I_2} \frac{\mu_i-c}{2\lambda}
$}
\\
&&
\mbox{$
-
\sum_{i\in I_3}
\frac12
\Big\{
1+\frac{\mu_i-c-\lambda}{\sqrt{\left(\mu_i-c-\lambda\right)^2+\sigma_i^2}}
\Big\}
+
\sum_{i\in I_4}
\frac12
\Big\{
1+\frac{c-\mu_i-\lambda}{\sqrt{\left(c-\mu_i-\lambda\right)^2+\sigma_i^2}}
\Big\}
$}
\\
\hspace{-1ex}
&=&
\hspace{-1ex}
-\sum_{i\in I_1} (p_i^+-p_i^-)
-\sum_{i\in I_2} (p_i^+-p_i^-)
-\sum_{i\in I_3} p_i^+
+ \sum_{i\in I_4} p_i^-,
\end{eqnarray*}
that is, $\frac{\partial}{\partial c}\phi_n(c,\lambda)
=-\sum_{i\in I_1\cup I_2\cup I_3} p_i^+ + \sum_{i\in I_1\cup I_2\cup I_4}p_i^-$.
From the fact
that $p_i^+=0$ for $i\in I_4$ and $p_i^-=0$ for $i\in I_3$
(see (\ref{eq.6.4})),
the relation $\frac{\partial}{\partial c}\phi_n(c,\lambda)=0$
implies the equality $\sum_{i=1}^n p_i^{+}=\sum_{i=1}^n p_i^{-}$,
and (\ref{eq.6.7}) follows.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Proposition \ref{prop.6.1}:}
[$\mbox{(ii)}\Rightarrow\mbox{(i)}$].
Suppose we are
given a probability matrix $Q$ satisfying
(ii).
By assumption,
$Q$ has vanishing principal diagonal.
Define
${\bXX}=(X_1,\ldots,X_n)$
as in (\ref{eq.6.10}).
Since $\sum_{(i,j):\ i\neq j} q_{ij}=\sum_{i,j}q_{ij}=1$,
this procedure maps $Q$ to a well-defined probability
law ${\cal L}({\bXX})$ on $\R^n$,
and the
map $Q\mapsto {\cal L}({\bXX})$ is, obviously, one to one.
Due to (\ref{eq.6.6}), the order
statistics of ${\bXX}$
satisfy
\be
\label{eq.6.11}
X_{1:n}<c-\lambda<X_{2:n}\leq \cdots \leq X_{n-1:n}<c+\lambda<X_{n:n}
\ \ \mbox{ with probability } 1.
\ee
Thus, from Lemma \ref{lem.1} it follows
that, with probability $1$,
\be
\label{eq.6.12}
\mbox{$
R_n=-(n-2)\lambda+\frac{1}{2}
\sum_{i=1}^n \big\{ \big|(X_i-c)-\lambda\big|
+\big|(X_i-c)+\lambda\big|\big\}.
$}
\ee
The assumptions $Q\in{\cal M}({\bpp}^{+},{\bpp}^{-})$
and $q_{ii}=0$
for all $i$ now
show that for any fixed $j$,
$\Pr[X_j=x_j^{-}]=\sum_{i\neq j} q_{ij}=\sum_{i=1}^n q_{ij}=p_j^{-}$.
Similarly
we conclude that for any fixed $i$,
$\Pr[X_i=x_i^{+}]=\sum_{j\neq i} q_{ij}=\sum_{j=1}^n q_{ij}=p_i^{+}$.
Thus, $\Pr[X_i=x_i^{o}]=1-p_i^{-}-p_i^{+}=p_i^{o}$, and the marginal
$X_i$ of ${\bXX}$ is the extremal random variable in
${\cal F}_1(\mu_i,\sigma_i)$.
That is, it has mean $\mu_i$, variance $\sigma_i^2$, and maximizes
$\E\big\{ |(X-c)-\lambda\big|
+\big|(X-c)-\lambda|\big\}$ as $X$ varies in ${\cal F}_1(\mu_i,\sigma_i)$.
Since this holds for all $i$, taking expectations in (\ref{eq.6.12})
we see that
\be
\label{eq.6.13}
\mbox{$
\E R_n=-(n-2)\lambda+\frac{\lambda}{2}
\sum_{i=1}^n U\big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\big)
=\rho_n,
$}
\ee
completing the proof.
\noindent
[$\mbox{(i)}\Rightarrow\mbox{(ii)}$].
Assumptions ${\bXX}\in{\cal F}_n(\bmu,{\bsigma})$ and $\E R_n=\rho_n$
imply that (repeat the proof of Theorem \ref{th.main})
\begin{eqnarray*}
\rho_n
\hspace{-1ex}
&=&
\hspace{-1ex}
\mbox{$
\E R_n
\leq
\E
\big\{-(n-2)\lambda
+\frac{1}{2}\sum_{i=1}^{n}\big[
\big| (X_i-c)-\lambda\big|
+
\big| (X_i-c)+\lambda\big|\big]
\big\}
$}
\\
\hspace{-1ex}
&=&
\hspace{-1ex}
\mbox{$
-(n-2)\lambda
+\frac{1}{2}\sum_{i=1}^{n}\E\big\{
\big| (X_i-c)-\lambda\big|
+
\big| (X_i-c)+\lambda\big|\big\}
$}
\\
\hspace{-1ex}
&&
\hspace{-1ex}
\mbox{$
\leq
-(n-2)\lambda
+\frac{\lambda}{2}\sum_{i=1}^{n}
U\big(\frac{\mu_i-c}{\lambda},\frac{\sigma_i}{\lambda}\big)
=\rho_n.
$}
\end{eqnarray*}
Thus, all displayed inequalities are attained as equalities.
In view of
Lemma \ref{lem.1} and
Corollary \ref{cor.1}, this
can happen only if the law ${\cal L}({\bXX})$ of the
given random vector ${\bXX}=(X_1,\ldots,X_n)$ satisfies
\be
\label{eq.6.14}
\begin{array}{ll}
\hspace{-1.5ex}
\mbox{(a)} &
\ \Pr\big[X_{1:n}\leq c-\lambda \leq X_{2:n}\leq
\cdots\leq X_{n-1:n}\leq c+\lambda\leq X_{n:n}\big]=1
\\
\hspace{-1.5ex}\mbox{and}
&
\\
\hspace{-1.5ex}
\mbox{(b)}
&
\begin{array}{l}
X_i
\mbox{ is extremal in } {\cal F}_1(\mu_i,\sigma_i)
\mbox{ for all } i,
\mbox{ or, equivalently, }
\\
\Pr[X_i=x_i^-]=p_i^{-},
\
\Pr[X_i=x_i^{o}]=p_i^{o},
\
\Pr[X_i=x_i^+]=p_i^{+},
\
i=1,\ldots,n.
\end{array}
\end{array}
\ee
Taking into account (\ref{eq.6.6}) we conclude that
(\ref{eq.6.14}) can happen only if
the (essential) support of ${\bXX}$ is contained in the set
\be
\label{eq.6.15}
S:=\big\{{\bxx}_{ij}, \ i\neq j, \ i,j=1,\ldots,n\big\},
\ee
with ${\bxx}_{ij}$ as in (\ref{eq.6.10}).
We can thus define the $n\times n$ matrix $Q$ as follows:
\be
\label{eq.6.16}
Q:=(q_{ij}), \mbox{ with } q_{ii}:=0, \ q_{ij}:=\Pr[{\bXX}={\bxx}_{ij}],
\ i\neq j, \ i,j=1,\ldots,n.
\ee
By definition, $Q$ has vanishing principal diagonal
and nonnegative entries,
and the relation
$\Pr[{\bXX}\in S]=1$ implies that $Q$ is a probability matrix.
By the assumption ${\bXX}\in{\cal F}_n(\bmu,{\bsigma})$
and $\E R_n=\rho_n$,
the marginal $X_i$ of ${\bXX}$ has to fulfill (\ref{eq.6.14})(b),
that is,
$\sum_{j=1}^n q_{ij}=\sum_{j\neq i} q_{ij}
=\sum_{j} \Pr[{\bXX}={\bxx}_{ij}] =\Pr[X_i=x_{i}^{+}]=
p_i^+$; similarly,
$\sum_{i=1}^n q_{ij}=p_j^-$. Therefore, we have
constructed a matrix $Q\in{\cal M}({\bpp}^{+},{\bpp}^-)$
with $q_{ii}=0$ for all $i$. Clearly, if
two random vectors ${\bXX}$, $\bYY$, with
${\cal L}(\bYY)\neq {\cal L}({\bXX})$,
satisfy the assumptions in (i),
the corresponding matrices (obtained through
(\ref{eq.6.16})) will be distinct. Consequently,
the above procedure determines a one to one mapping
${\cal L}({\bXX})\mapsto Q$,
completing the proof.
$\Box$
\medskip
\end{PR}
\noindent
\begin{PR}{\small\bf Proof of Theorem \ref{theo.7.1}:}
For the infimum, a proof (for any $n\geq 2$) is
given in the beginning of Section \ref{sec.main.inequality},
following the arguments of Bertsimas, Doan, Natarajan and Teo
(2010).
Regarding the supremum: The key-observation is that
\eqref{eq.7.3} is a special application
of the Cauchy-Schwarz inequality,
\[
\mbox{$
\E R_2=\E |X_1-X_2|\leq \sqrt{\E\big[(X_1-X_2)^2\big]}
=\sqrt{(\mu_1-\mu_2)^2+\sigma_1^2+\sigma_2^2-2\rho\sigma_1\sigma_2}=\gamma_2.
$}
\]
This means that, in order to justify the equality, we have to construct
a vector $(X_1,X_2)\in{\cal F}_2(\bmu,{\bsigma};\rho)$
such that the random variable $|X_1-X_2|$ is degenerate.
Let $\delta:=\Var[X_1-X_2]=\sigma_1^2+\sigma_2^2-
2\rho\sigma_1\sigma_2\geq 0$. We
distinguish cases $\delta>0$, $\delta=0$.
Assume $\delta>0$, so that $\gamma_2>0$.
First, we consider
a $0$--$1$ Bernoulli random variable $I_p$
with probability of success
$p:=\frac{1}{2}\big(1+\frac{\mu_1-\mu_2}{\gamma_2}\big)$.
Next, we consider another random variable
$T$ with mean
$\mu_T:=\mu_1\sigma_2^2+\mu_2\sigma_1^2-\rho\sigma_1\sigma_2(\mu_1+\mu_2)$
and variance
$\sigma_T^2:=\delta\sigma_1^2\sigma_2^2(1-\rho^2)\geq 0$,
stochastically independent
of $I_p$.
Finally, we define
\[
\mbox{$
(X_1,X_2):=\frac{1}{\delta}\
\big[\gamma_2(\sigma_1^2-\rho\sigma_1\sigma_2)(2I_p-1)+T,
\
\gamma_2(\rho\sigma_1\sigma_2-\sigma_2^2)(2I_p-1)+T\big].
$}
\]
It is easily seen that $(X_1,X_2)\in{\cal F}_2(\bmu,{\bsigma};\rho)$
and $|X_1-X_2|=\gamma_2$ with probability 1.
Let us now assume $\delta=0$.
This implies that $X_1-X_2=\mu_1-\mu_2$ with probability $1$, and hence,
$\sigma_1=\sigma_2$ and $\rho=1$.
Let $\sigma^2>0$ be the common variance and consider
the pair $(X_1,X_2):=(\mu_1+T,\mu_2+T)$, where $T$
is any random variable with mean zero and variance $\sigma^2$.
It follows that $(X_1,X_2)$ satisfies the moment requirements and
$|X_1-X_2|=|\mu_1-\mu_2|=\gamma_2$ with probability 1. This completes the proof.
$\Box$
\end{PR}
\end{appendix}
}
|
\section{Introduction}
\section{Introduction}
Quantum correlations in composite systems transcend entanglement \cite{modirev}.
A bipartite quantum state $\rho_{AB}$ can be defined as nonclassical or
nonclassically correlated if it cannot be expressed as a convex
mixture of local basis states of subsystems $A$ and $B$ \cite{nolocal}. Consequently, all
inseparable (entangled) states as well as the majority of separable states are nonclassical.
General nonclassical correlations, however, can be mapped to entanglement in a very precise sense, which provides an insightful framework for their characterization and operational interpretation. Specifically, it was proven in \cite{Piani_11,streltsov,PianiAdesso} and very recently experimentally observed in \cite{sciarrino} that all nonclassical states of a finite-dimensional system can be
turned into states with distillable entanglement between the system and a set of ancillae by an {\it activation
protocol}. Focusing on a bipartite setting, the protocol runs as follows. The subsystems $A$ and $B$ are first subject to arbitrary local unitary
transformations $U_{A,B}$; then, each system $j=A,B$ interacts via a controlled-NOT (CNOT) operation
$U_{jj'}^{CNOT}$ (i.e.~a so-called premeasurement interaction) with an auxiliary system $j'$, $j=A,B$, initialized
in a pure state $|0\rangle_{j'}$
The activation protocol then possesses two key properties: i) for all classical states
$\rho_{AB}$ at the input of the protocol, there exist local
unitaries $U_{A,B}$ for which the output state $\rho_{ABA'B'}$ is
separable across the $AB|A'B'$ splitting, and ii) for all
nonclassical states $\rho_{AB}$ and for all local
unitaries, the output state is entangled across the
$AB|A'B'$ splitting.
Let us stress that both criteria i) and ii) must be met by any scheme in order
to be a valid activation protocol. In particular, they allow us to define faithful measures of nonclassical correlations for the input state $\rho_{AB}$ in terms of the output $AB|A'B'$ entanglement, minimized over $U_{A,B}$. One such measure, when the output entanglement is quantified by the negativity \cite{Vidal_02}, has been termed negativity of quantumness \cite{Piani_11,Takafumi_13}, and has been experimentally investigated in \cite{isabela,sciarrino}
In this paper we study activation of nonclassical correlations in
multimode bipartite Gaussian states $\rho_{AB}$ of
continuous variable systems \cite{ourreview}. Nonclassical
correlations of Gaussian states have been studied extensively both
theoretically and experimentally
\cite{Adesso_10,GiordaParis,GAMID,gdexp} but their interplay with
entanglement has not been pinned down so far in terms of the
activation framework. Attempts to devise activation-like protocols
for Gaussian states have been explored \cite{maurosrep}. However,
these differed significantly from the original prescription in
that nonunitary operations were employed between system and
ancillae, so that the entanglement generation was obtained as a
dynamical feature, and conditions i) and ii) were not generally
verified.
Here we consider a general Gaussian activation protocol in
which $U_{A,B}$ are Gaussian unitaries and the CNOT gates are
replaced with a global Gaussian unitary on subsystems $A,B,A', B'$. In Section~\ref{sec_1} we then prove that any such protocol satisfying condition i) will unavoidably violate condition ii), which implies that activation of Gaussian nonclassical correlations by Gaussian operations is impossible. This fact establishes a new no-go theorem for Gaussian quantum information processing, which can be enlisted alongside other well known no-go results such as the the no-distillation theorem, according to which distilling entanglement from Gaussian states by using only Gaussian operations is impossible \cite{nogo}. We then show in Section~\ref{sec_2} how, by using non-Gaussian operations which properly extend the CNOT to infinite dimensions, one can construct the continuous variable counterpart of the activation protocol of \cite{Piani_11}, verifying criteria i) and ii). This allows us to define the negativity of quantumness for Gaussian states and to calculate it for relevant examples in Section~\ref{sec_3}.
This work provides an operational setting to understand and manipulate nonclassical correlations in paradigmatic infinite-dimensional systems. We draw our conclusions in Section~\ref{sec_4}, while some technical derivations (which can be of independent interest) are deferred to the Appendices.
\section{Gaussian no-activation theorem}\label{sec_1}
Gaussian states are quantum states of systems with an infinite-dimensional Hilbert
space (continuous variable systems), e.g.~a collection of harmonic oscillators, which possess a
Gaussian-shaped Wigner function in phase space \cite{ourreview}. $L$ modes are described by a
vector ${\bf r}=\left(x_{1},p_{1},\ldots,x_{L},p_{L}\right)^{T}$
of quadrature operators $x_{j},p_{j}$ satisfying the canonical
commutation rules expressible in terms of elements of the vector
${\bf r}$ as $[r_{j},r_{k}]=i\Omega_{jk}$, $j,k=1,\ldots,L$ with
$\Omega=\oplus_{j=1}^{L}i\sigma_{y}$, where $\sigma_{y}$ is the
Pauli $y$-matrix. An $L$-mode Gaussian state $\rho$ is fully
characterized by a $2L\times 1$ vector $\langle {\bf r}\rangle$ of
the first moments with elements $\langle
r_{i}\rangle=\mbox{Tr}(\rho r_{i})$ and by its $2L\times2L$
covariance matrix (CM) $\gamma$ with elements $\gamma_{ij}=\langle
\Delta r_{i}\Delta r_{j}+\Delta r_{j}\Delta r_{i}\rangle/2$,
$i,j=1,\ldots,L$, where $\Delta r_{i}=r_{i}-\langle r_{i}\rangle$.
Gaussian unitaries are generated by Hamiltonians that are
quadratic in the quadrature operators and they preserve the Gaussian
characteristic of quantum states. An $L$-mode Gaussian unitary $U(S)$ is represented in phase space by a $2L\times2L$ real
symplectic transformation $S$ satisfying the condition $S\Omega
S^{T}=\Omega$, which transforms a CM $\gamma$ to $S\gamma S^{T}$.
Here we are interested in the question of whether an
activation protocol exists satisfying conditions i) and ii) which
would rely solely on Gaussian states and Gaussian unitaries.
We therefore assume the state $\rho_{AB}$ to be a Gaussian state of
$(N+M)$ modes with CM $\gamma_{AB}$ and
the state $\rho_{A'B'}$ of the ancilla to be also a Gaussian state
with CM $\gamma_{A'B'}$. The local unitaries $U_{A,B}$ of the
original discrete protocol are replaced with local Gaussian
unitaries $U_{A}(S_{A})$ and $U_{B}(S_{B})$ represented by the
symplectic matrices $S_{A}$ and $S_{B}$, respectively. Likewise,
the global operation $U_{AA'}^{CNOT}\otimes U_{BB'}^{CNOT}$ on the
whole system $ABA'B'$ is replaced with one global Gaussian unitary
$U(S)$ represented by a symplectic matrix $S$.
Let us recall the definition of a fully classical state
\cite{nolocal,Piani_11,PianiAdesso}. Suppose $\rho_{AB}$ is a
bipartite state containing two subsystems $A$ and $B$ with $N$ and
$M$ modes respectively, and let
$\mathcal{B}_j=\{\ket{{\mathcal{B}_j(\bf n }_{j})}\}$ be a basis
of subsystem $j$, with ${\bf n}_A=(n_{A_1},\ldots,n_{A_N})$, ${\bf
n}_B=({n}_{B_1},\ldots,n_{B_M})$ and $n_{j_i}\in\mathbb{N}_0$. If
there exists a basis $\mathcal{B}$ consisting of the tensor
products of all elements of $\mathcal{B}_A$ with all elements of
$\mathcal{B}_B$, then $\rho_{AB}$ is a classical state if it is
diagonal with respect to $\mathcal{B}$. It has been shown in
\cite{Adesso_10,Rahimi-Keshari_13} that a two-mode Gaussian state
is classical if and only if it is a product state, i.e., its CM is
represented by a direct sum $\gamma_{A}\oplus\gamma_{B}$ of local
CMs $\gamma_{A,B}$. One can prove that this statement remains
valid for the generic case of bipartite $(N+M)$-mode Gaussian
states (see Appendix~\ref{secapp_1} for the proof).
Therefore, all non-product bipartite Gaussian states (including
separable ones) are nonclassical. According to condition i) in any
Gaussian activation protocol with an input Gaussian product state
there must exist local Gaussian unitaries $U_{A,B}$ for which one
gets a separable state $\rho_{ABA'B'}$ across the $AB|A'B'$
splitting at the output of the protocol. We will show however,
that this implies that for {\it all} separable Gaussian states
including nonclassical ones, there exist local Gaussian unitaries
$U_{A,B}$ for which the output state is separable. That is,
condition ii) is not satisfied. Thus, any Gaussian activation
protocol described above cannot meet simultaneously criteria i)
and ii), and hence does not exist.
\begin{figure}[tb]
\includegraphics[width=8.5cm]{No-gofigurecolor.pdf}
\caption{(Color online) Pictorial representation of the no-activation theorem.
$\rho_{AB}$ is a separable Gaussian state prepared
from a Gaussian product state $\rho_{A}\otimes\rho_{B}$ by
correlated displacements $D_{A}$ and $D_{B}$ distributed according
to a Gaussian distribution with correlation matrix $P$,
Eq.~(\ref{P}). $U_{A}(S_{A})$ and $U_{B}(S_{B})$ are local
Gaussian unitaries which are adjusted such that without the
displacements $D_{A}$ and $D_{B}$ the activation protocol produces
from the product state $\rho_{A}\otimes\rho_{B}$ an output state
which is separable across the $AB|A'B'$ splitting. As the unitaries
$U_{A}(S_{A})$, $U_{B}(S_{B})$ and $U(S)$ induce a linear
transformation of quadrature operators of the input modes, the
displacements $D_{A}$ and $D_{B}$ can be relocated behind the
global transformation $U(S)$ (dotted arrow). The new
displacements $D_{A},D_{B}, D_{A'}$ and $D_{B'}$,
Eq.~(\ref{displacement}), cannot turn a separable state into an
entangled state and therefore the protocol transforms the
separable state $\rho_{AB}$ into a state which is separable across the
$AB|A'B'$ cut (thick dashed line). See text for details.}
\label{fig1}
\end{figure}
The proof of this no-go theorem is depicted in
Fig.~\ref{fig1}. It follows from the
decomposability of any Gaussian separable state into a product
state and noise \cite{Werner_01}, and the linearity of symplectic
transformations. Namely, for any separable Gaussian state
with CM $\gamma_{AB}$ there exist local CMs $\gamma_{A,B}$ such
that
\begin{equation}\label{P}
P=\gamma_{AB}-\gamma_{A}\oplus\gamma_{B}\geq0.
\end{equation}
In other words, any separable Gaussian state with CM $\gamma_{AB}$
can be prepared from a suitable product state with CM
$\gamma_{A}\oplus\gamma_{B}$ by the addition of noise, represented
by a positive-semidefinite matrix $P$, i.e.,
$\gamma_{AB}=\gamma_{A}\oplus\gamma_{B}+P$. The noise can be
created by displacing the vector of quadratures ${\bf
r}=\left(x_{A_1},p_{A_1},\ldots,x_{A_N},p_{A_N},x_{B_1},p_{B_1},\ldots,x_{B_M},p_{B_M}\right)^{T}$
of the product state as ${\bf r}\rightarrow {\bf r}+{\cal V}{\bf
R}$. Here ${\cal V}$ is a $2(N+M)\times K$ matrix given by the
first $K$ columns of the matrix $V$ bringing the matrix $P$ to the
diagonal form
$V^{T}PV=\mbox{diag}\left(\lambda_1,\lambda_2,\ldots,\lambda_K,0,0,\ldots,0\right)$,
where $\lambda_1,\ldots,\lambda_K$ denote $K\leq2(N+M)$ strictly
positive eigenvalues of the matrix (\ref{P}), and ${\bf
R}=\left(R_1,\ldots,R_K\right)^{T}$ is the vector of classical
displacements uncorrelated with the vector of quadratures ${\bf
r}$ and distributed according to the Gaussian distribution with
zero means and the diagonal correlation matrix
$\mbox{diag}\left(\lambda_1,\lambda_2,\ldots,\lambda_K\right)$.
Let us now consider a separable state with CM $\gamma_{AB}$ at the
input of a Gaussian activation protocol and let
$\gamma_A\oplus\gamma_B$ be a CM of the product state from which
the state can be prepared using the aforementioned algorithm.
Assume that the local symplectic matrices $S_{A,B}$ are chosen
such that the CM $\gamma_{ABA'B'}^{(0)}\equiv S\left(S_{A}\oplus
S_{B}\oplus\openone_{A'B'}\right)\gamma_A\oplus\gamma_B\oplus\gamma_{A'B'}\left(S_{A}^{T}\oplus
S_{B}^{T}\oplus\openone_{A'B'}\right)S^{T}$ of the output state,
where $\gamma_{A'B'}$ is the CM of the state of the ancilla, is
separable across the $AB|A'B'$ splitting. Hence, for the original
separable state with CM $\gamma_{AB}$, the output of the
activation protocol is obtained by displacing the vector of
quadratures ${\bf r}^{(0)}$ for the state with CM
$\gamma_{ABA'B'}^{(0)}$ by
\begin{equation}\label{displacement}
{\bf r}^{(0)}\rightarrow {\bf r}^{(0)}+S\left(\begin{array}{c}
\left(S_A\oplus
S_B\right){\cal V}{\bf R} \\
\mathbb{O} \\
\end{array}\right),
\end{equation}
where $\mathbb{O}$ is a $2T\times 1$ zero vector with $T$ being
the number of modes of the ancilla $A'B'$. However, for a
separable state with CM $\gamma_{ABA'B'}^{(0)}$, where $AB$ is
separable from $A'B'$, the local displacements
(\ref{displacement}) cannot create a state in which the system
$AB$ is entangled with the system $A'B'$. Consequently, for any
separable state (even nonclassical) it is always possible to find
local Gaussian unitaries for which the output is separable, thus
accomplishing the proof of the no-go theorem.
Therefore, Gaussian operations are unable to activate nonclassical correlations of Gaussian separable states into entanglement in the worst-case scenario: assuming condition i) holds, then for any Gaussian separable state there exist local
Gaussian unitaries for which the output of the activation protocol
remains a separable Gaussian state. This indicates that a
non-Gaussian element, like a non-Gaussian global unitary $U$ or a
non-Gaussian state of the ancilla, is
necessary for faithful activation of nonclassical correlations in Gaussian
states. In the following we design such an activation
protocol involving a non-Gaussian CNOT gate in the Fock basis and
an ancillary system in a Gaussian state.
\section{Non-Gaussian activation protocol}\label{sec_2}
The main benefit of the activation protocol is that it allows one
to quantify the amount of nonclassical correlations in a given
quantum state as the potential to create entanglement in the
activation protocol \cite{Piani_11,streltsov,PianiAdesso}. More precisely, if
$E_{AB|A'B'}(\rho_{ABA'B'})$ denotes an entanglement measure
quantifying the amount of entanglement between systems $AB$
and $A'B'$ in a quantum state $\rho_{ABA'B'}$, then we can define
a measure of nonclassical correlations on the input state $\rho_{AB}$ as
\begin{eqnarray}\label{QE}
Q_{E}(\rho_{AB})=\mathop{\mbox{min}}_{U_{A},U_{B}}E_{AB|A'B'}(\rho_{ABA'B'}),
\end{eqnarray}
where the minimization is carried out over all local unitaries $U_{A}$ and $U_{B}$ on subsystems $A$ and $B$.
It has been proven in \cite{PianiAdesso} that $Q_E(\rho_{AB}) \geq E(\rho_{AB})$, with equality if $\rho_{AB}$ is pure.
From now on we assume that systems $A$ and $B$ each contain one mode. The non-Gaussian activation protocol
is obtained as a direct generalization of the finite-dimensional protocol
\cite{Piani_11}. At the input we allow for generally non-Gaussian
states $\rho_{AB}$ of continuous variable systems, local unitaries $U_A$ and $U_B$, and the global Gaussian unitary $U(S)$ of the preceding protocol is replaced with the
tensor product $V\equiv U_{AA'}^{CNOT}\otimes U_{BB'}^{CNOT}$ of the infinite-dimensional
generalizations of CNOT gates in the Fock basis,
\begin{eqnarray}\label{FockCNOT}
U_{jj'}^{CNOT}|m,n\rangle_{jj'}=|m,m+n\rangle_{jj'},\quad j=A,B,
\end{eqnarray}
where $\ket{m,n}_{jj'}\equiv\ket{m}_{j}\otimes\ket{n}_{j'}$,
$m,n=0,1,\ldots$, and $\ket{k}_l$ is the $k$th Fock state of mode
$l$. We also assume the initial state $\rho_{A'B'}$ of the
ancilla $A'B'$ to be the vacuum state
$\ket{0}_{A'}\bra{0}\otimes\ket{0}_{B'}\bra{0}$. Hence, the final
output state can be expressed as
\begin{eqnarray}\label{rhoout}
\rho_{ABA'B'}&=&V(\tilde{\rho}\otimes\ket{0}_{A'}\bra{0}\otimes\ket{0}_{B'}\bra{0})V^{\dagger},
\end{eqnarray}
where
\begin{eqnarray}\label{tilderho}
\tilde{\rho}&\equiv& (U_A\otimes
U_B)\rho_{AB}(U^{\dagger}_A\otimes U^{\dagger}_B).
\end{eqnarray}
By following arguments similar to the finite-dimensional case
\cite{Piani_11}, one can show that the non-Gaussian activation
protocol defined above satisfies both criteria i) and ii). For
condition i) we assume that $\rho_{AB}$ is classically correlated
and hence there exist local unitaries $U_A$ and $U_B$ such that
the density matrix $\tilde{\rho}$, Eq.~(\ref{tilderho}), takes the
form
$\tilde{\rho}=\sum_{n,m=0}^{\infty}p_{n,m}\ket{n,m}_{AB}\bra{n,m}$.
Making use of Eqs.~(\ref{FockCNOT}) and (\ref{rhoout}) it then
follows that the output state of the protocol is the following
convex mixture of product states, $
\rho_{ABA'B'}=\sum_{n,m=0}^{\infty}p_{n,m}\ket{n,m}_{AB}\bra{n,m}\otimes\ket{n,m}_{A'B'}\bra{n,m}
$,
and is thus a separable state across the $AB|A'B'$ splitting
as required.
For the proof of condition ii) we now suppose that the density matrix $\rho_{AB}$ is
nonclassical and show that then the density matrix $\rho_{ABA'B'}$ given in Eq.~(\ref{rhoout}) is entangled across
the $AB|A'B'$ cut for all local unitaries $U_A$ and $U_B$. To prove the presence of entanglement
in $\rho_{ABA'B'}$ we will use the negativity $\mathcal{N}$ defined in \cite{Vidal_02} as
\begin{equation}\label{N1}
\mathcal{N}(\rho_{ABA'B'})=\frac{1}{2}(\|\rho^{T_{AB}}_{ABA'B'}\|_1-1).
\end{equation}
Here $\|.\|_1$ denotes the trace norm, $\rho^{T_{AB}}_{ABA'B'}$ is the partial transpose \cite{Peres_96} of the state $\rho_{ABA'B'}$
with respect to subsystem $AB$, and a strictly positive value of negativity implies that the state $\rho_{ABA'B'}$ is (distillable) entangled with respect to the
$AB|A'B'$ splitting. The specific feature of the present activation protocol is that the output state $\rho_{ABA'B'}$ is a so-called maximally
correlated state and therefore, following results in \cite{Piani_11,Takafumi_13}, the output negativity can be expressed as
\begin{equation}\label{N2}
\mathcal{N}(\rho_{ABA'B'})=\frac{1}{2}\sum_{\mathbf{m\neq
n}=\mathbf{0}}^{\infty}|\tilde{\rho}_{\mathbf{m},\mathbf{n}}| = \frac{1}{2}\left(\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}|\tilde{\rho}_{\mathbf{m},\mathbf{n}}|-1\right),
\end{equation}
where $\mathbf{m}=(m_{1},m_{2})$, $\mathbf{n}=(n_{1},n_{2})$ and $\tilde{\rho}_{\mathbf{m},\mathbf{n}}=_{AB}\langle m_{1}m_{2}|\tilde{\rho}|n_{1}n_{2}\rangle_{AB}$ are elements of the density matrix $\tilde{\rho}$, Eq.~(\ref{tilderho}), in the Fock basis.
Since our input state $\rho_{AB}$ is nonclassical, the state $\tilde{\rho}$ is also nonclassical for any choice of unitaries $U_A$ and $U_B$.
Thus, there must be at least one non-zero off-diagonal element $\tilde{\rho}_{\mathbf{m},\mathbf{n}}$ for every choice of $U_A$ and $U_B$. Hence, Eq.~(\ref{N2})
implies $\mathcal{N}(\rho_{ABA'B'})>0$ and the output state $\rho_{ABA'B'}$ is entangled for any nonclassical input state. This completes the proof of our non-Gaussian activation protocol.
\section{Examples}\label{sec_3}
The optimization in Eq.~(\ref{QE}) is generally carried out over all local
unitary operations $U_{A}$ and $U_{B}$, including non-Gaussian ones,
which is not a tractable task. Here we consider input Gaussian states with CM in standard form \cite{Simon_00}, and consider the non-optimized output entanglement $E_{AB|A'B'}(\rho_{ABA'B'})$ obtained when the local unitaries $U_{A,B}$ are selected to be identity matrices. Therefore, the state
(\ref{tilderho}) remains a Gaussian state in standard form with
the following CM
\begin{eqnarray}\label{tildegamma}
\tilde{\gamma}=\left(\begin{array}{cc}
A & C \\
C^T & B\\
\end{array}\right),
\end{eqnarray}
where $A=\mbox{diag}(a,a)$, $B=\mbox{diag}(b,b)$ and $C=\mbox{diag}(c_1,c_2)$ are diagonal matrices.
In what follows we determine the non-optimized
quantity for some classes of two-mode Gaussian states by considering the negativity (\ref{N1}) as an entanglement measure $E$, and using Eq.~(\ref{N2}).
The corresponding measure of nonclassical correlations $Q_{\mathcal{N}}(\rho_{AB})$
is called the negativity of quantumness \cite{Piani_11} accordingly. Although our choice of local unitaries $U_{A,B}$
gives in general an upper bound on $Q_{\mathcal{N}}$, we find that it coincides with the
true measure on pure states, leading us to conjecture that our choice is optimal for calculating the negativity of quantumness of all two-mode Gaussian
states in standard form. Verifying this conjecture numerically is beyond the scope of this work.
\subsection{Pure states} A closed form of the output negativity can be found for pure two-mode Gaussian
states. The density matrix $\tilde{\rho}$ amounts to that of the two-mode squeezed vacuum
state, with $\tilde{\rho}_{\mathbf{m},\mathbf{n}}=[1-\tanh^2(r)](\tanh r)^{m_1+n_1}\delta_{m_1,m_2}\delta_{n_1,n_2}$,
where $r\geq0$ is the squeezing parameter. Hence, by a direct substitution into Eq.~(\ref{N2})
we get
\begin{equation}\label{Np}
\mathcal{N}_{\rm p}=\mbox{$\frac12(e^{2r}-1)$}.
\end{equation}
Consequently, as the output negativity $\mathcal{N}(\rho_{ABA'B'})$ is equal to the negativity
of the input state $\rho_{AB}$, it coincides with the true optimized
negativity of quantumness $Q_{\mathcal{N}}(\rho_{AB})$ \cite{PianiAdesso}, and our
choice of local unitaries is thus optimal for pure states.
The negativity (\ref{Np}) is depicted by a solid red line in
Fig.~\ref{fig2}.
\begin{figure}[tb]
\includegraphics[width=8.5cm]{LboundsInsetnew.pdf}
\caption{(Color online) Negativity of quantumness $\mathcal{N}_{\rm p}$ [Eq.~(\ref{Np})] (solid red
line) and its lower bound $\mathcal{L}_{\rm p}$ [Eq.~(\ref{Lp})] (dash-dotted brown line)
for pure squeezed vacuum states, plotted as a function of the local mean number of thermal photons
$\langle n\rangle=\sinh^2(r)$. Upper bound on the negativity of quantumness $\mathcal{N}_{\rm m}$ [Eq.~(\ref{Nm})] (dashed blue line)
and its lower bound $\mathcal{L}_{\rm m}$ [Eq.~(\ref{Lm})] (dotted black line) for separable mixed
states obtained as unbiased mixtures of coherent states, plotted as a function of the local mean number of thermal photons $\langle n\rangle=\sigma^2$. The inset shows a close-up for $\langle n \rangle \ll 1$, where the lower bounds become tight.}\label{fig2}
\end{figure}
\subsection{Unbiased mixtures of coherent states}
These Gaussian states are of the form
\begin{equation}\label{mixture}
\rho_{AB}=\int_{\mathbb{C}} P(\alpha)|\alpha\rangle_{A}\langle\alpha|\otimes|\alpha\rangle_{B}\langle\alpha|d^{2}\alpha,
\end{equation}
and can be prepared by splitting a thermal state with
mean number of thermal photons $2\sigma^2$ on a balanced beam
splitter. Here $\alpha\in\mathbb{C}$,
$P(\alpha)=\mbox{exp}(-|\alpha|^2/\sigma^{2})/(\pi\sigma^2)$ and
$d^{2}\alpha=d(\mbox{Re}\alpha)d(\mbox{Im}\alpha)$. The states are
already in standard form with a CM (\ref{tildegamma}) specified by
$a=b=\sigma^2+1/2$ and $c_1=c_2=\sigma^2$. Making use of
the components of a coherent state in Fock basis $\langle
m|\alpha\rangle=\mbox{exp}(-|\alpha|^2/2)\alpha^{m}/\sqrt{m!}$ we
get the following matrix elements of the state (\ref{mixture}),
\begin{equation}\label{mixtureFock}
\tilde{\rho}_{\mathbf{m},\mathbf{n}}=\frac{(m_1+m_2)!}{\sqrt{m_1!m_2!n_1!n_2!}}\frac{\delta_{m_1+m_2,
n_1+n_2}}{s(m_1+m_2)},
\end{equation}
where $s(j)=\sigma^2\left(1/\sigma^2+2\right)^{j+1}$. By
substitution of the latter expression into Eq.~(\ref{N2}) we get
after some algebra
\begin{equation}\label{Nm}
\mathcal{N}_{\rm m}=\frac{1}{2}\left\{\sum_{M=0}^{\infty}\frac{1}{s(M)}\left[\sum_{J=0}^{M}\sqrt{{M \choose J}}\right]^2-1\right\}.
\end{equation}
The negativity (\ref{Nm}) is depicted by a dashed blue line in
Fig.~\ref{fig2}, and is generally smaller than the one of pure states calculated in (\ref{Np}). Both classes of Gaussian states have a nonzero negativity of quantumness which increases with $\langle n \rangle >0$; this is in agreement with earlier studies of nonclassical correlations based on entropic measures of quantum discord \cite{Adesso_10,GiordaParis}.
\subsection{Standard-form two-mode Gaussian states}
In general we need the Fock
basis elements $\tilde{\rho}_{\mathbf{m},\mathbf{n}}$ for an
arbitrary two-mode Gaussian state with zero first moments. Combining the
results of Refs.~\cite{Dodonov_84,Dodonov_94,Fiurasek_review01} we
can express them as
\begin{equation}\label{rhoHermite}
\tilde{\rho}_{\mathbf{m},\mathbf{n}}=\frac{H_{m_1,m_2,n_1,n_2}^{(R)}(0)}{\sqrt{\text{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}\sqrt{m_1!m_2!n_1!n_2!}},
\end{equation}
where $\tilde{\gamma}$ is the CM of the state, $\openone$ is the
$4\times 4$ identity matrix, and $H_{m_1,m_2,n_1,n_2}^{(R)}(0)$ is
the four-dimensional Hermite polynomial \cite{Bateman_53} at the
origin; see Appendix~\ref{secapp_2} for a complete derivation of
Eq.~(\ref{rhoHermite}). Here
\begin{equation}\label{RR}
R=WO\left[\left(\tilde{\gamma}+\frac{1}{2}\openone\right)^{-1}-\openone\right]O^{\dagger}V
\end{equation}
is the symmetric matrix defining the polynomial, where
\begin{equation}
W=\left(\begin{array}{cccc}
1 & 0 & 0 & 0\\
0 & 0 & 1 & 0\\
0 & 1 & 0 & 0\\
0 & 0 & 0 & 1\end{array}\right),\quad V=\left(\begin{array}{cccc}
0 & 0 & 1 & 0\\
1 & 0 & 0 & 0\\
0 & 0 & 0 & 1\\
0 & 1 & 0 & 0\end{array}\right),
\end{equation}
and
\begin{eqnarray}\label{OO}
O=\bigoplus_{j=1}^{2}\frac{1}{\sqrt{2}}\left(\begin{array}{cc}
1 & i\\
1 & -i \end{array}\right).
\end{eqnarray}
For the standard-form CM $\tilde{\gamma}$, Eq.~(\ref{tildegamma}),
we get in particular
\begin{equation}\label{Rst}
R=\left(\begin{array}{cc}
R_{1}-R_{2} & R_{1}+R_{2}-\openone_2 \\
R_{1}+R_{2}-\openone_2 & R_{1}-R_{2}\\
\end{array}\right)
\end{equation}
with $\openone_2$ being the $2\times 2$ identity matrix,
\begin{equation}\label{Rj}
R_{j}=\displaystyle\frac{1}{2d_{j}}\left(\begin{array}{cc}
b+\frac{1}{2} & -c_{j} \\
-c_{j} & a+\frac{1}{2}\\
\end{array}\right),
\end{equation}
and $d_{j}=(a+1/2)(b+1/2)-c_{j}^2$ ($j=1,2$). One can then
evaluate the negativity (\ref{N2}) by performing a numerical
summation of the absolute values of the elements
(\ref{rhoHermite}). The higher-order Hermite polynomials can be
calculated from the lower-order ones by using e.g.~the recurrence
formula derived in Appendix~\ref{secapp_2}.
We remark that the compact expression in equation (\ref{rhoHermite}) is of independent interest and can be useful for the characterization of hybrid information processing involving conversion between continuous and discrete variable entanglement \cite{hybridrev}, or particularly for studies of Bell nonlocality of arbitrary two-mode Gaussian states by means of dichotomic pseudospin measurements \cite{chen}, whose expectation value can be conveniently evaluated at the Fock space level.
In the context of the present paper, apart from the utility for numerical evaluation of the output negativity (\ref{N2}), equation (\ref{rhoHermite}) also enables us to derive a simple analytical lower bound on the output negativity. The bound results from the following chain
of inequalities
\begin{eqnarray}\label{inequalities}
\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}|\tilde{\rho}_{\mathbf{m},\mathbf{n}}|&=&
\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}\frac{|H_{m_1,m_2,n_1,n_2}^{(R)}(0)|}
{\sqrt{\text{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}\sqrt{m_1!m_2!n_1!n_2!}}\nonumber\\
&\geq&\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}\frac{|H_{m_1,m_2,n_1,n_2}^{(R)}(0)|}
{\sqrt{\text{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}{m_1!m_2!n_1!n_2!}}\nonumber\\
&\geq&\vline\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}\frac{H_{m_1,m_2,n_1,n_2}^{(R)}(0)}
{\sqrt{\text{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}{m_1!m_2!n_1!n_2!}}\vline\nonumber\\
&=&\frac{e^{-\frac{1}{2}\sum_{i,j=1}^{4}R_{ij}}}
{\sqrt{\mbox{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}},
\end{eqnarray}
where the first inequality follows from the inequality $1/\sqrt{n!}\geq1/{n!}$ which holds for any $n\geq0$, the second
inequality is a consequence of the triangular inequality for absolute values, and the last equation follows
from the expression for the generating function of the four-dimensional Hermite polynomials at the origin \cite{Bateman_53},
\begin{equation}\label{generatingfunction}
e^{-\frac{1}{2}h^T R h}=\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}{\frac{{\alpha_1^*}^{m_1}{\alpha_2^*}^{m_2}\alpha_1^{n_1}\alpha_2^{n_2}}{m_1!m_2!n_1!n_2!}}
H_{m_1,m_2,n_1,n_2}^{(R)}(0),
\end{equation}
where $h=(\alpha_1^*,\alpha_2^*,\alpha_1,\alpha_2)^T$ and $R$ is the matrix (\ref{RR}).
A comparison between the right-hand side (RHS) of the previous equation and the expression of the
Husimi $Q$-quasiprobability distribution
$\Phi_{\mathcal{A}}(\alpha_1,\alpha_2)=\bra{\alpha_1\alpha_2}\tilde{\rho}\ket{\alpha_1\alpha_2}/\pi^2$ in the Fock
basis further yields
\begin{equation}\label{PhiAR}
\frac{e^{-\frac{1}{2}h^T R h}}{\sqrt{\mbox{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}}=\pi^2e^{|\alpha_1|^2+|\alpha_{2}|^2}\Phi_{\mathcal{A}}(\alpha_1,\alpha_2)
\end{equation}
as can be easily seen from the results of Appendix~\ref{secapp_2}.
Therefore, the last expression in the chain of inequalities (\ref{inequalities}) can be written in the following compact form
\begin{equation}\label{PhiA11}
\frac{e^{-\frac{1}{2}\sum_{i,j=1}^{4}R_{ij}}}
{\sqrt{\mbox{det}\left(\tilde{\gamma}+\frac{1}{2}\openone\right)}}=\left(\pi e\right)^2\Phi_{\mathcal{A}}(1,1).
\end{equation}
Now, making use of the inequalities (\ref{inequalities}) and equality (\ref{PhiA11}) one finds that the sum in (\ref{N2}) is lower-bounded as
\begin{equation}\label{sumbound}
\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}|\tilde{\rho}_{\mathbf{m},\mathbf{n}}|\geq\left(\pi e\right)^2\Phi_{\mathcal{A}}(1,1),
\end{equation}
which finally gives the following bound on the output negativity (\ref{N2})
\begin{equation}\label{L}
\mathcal{N}(\rho_{ABA'B'})\geq\frac{1}{2}\left[\left(\pi e\right)^2\Phi_{\mathcal{A}}(1,1)-1\right].
\end{equation}
The bound (\ref{L}) can be evaluated for any zero-mean two-mode Gaussian state with CM $\tilde{\gamma}$ by
calculating the matrix (\ref{RR}) and substituting it into the formula (\ref{PhiA11}). To test
the tightness of the bound we calculate it for the previous examples of pure states and mixtures of coherent states, and compare the obtained lower bounds with the exact values of the negativities (\ref{Np}) and (\ref{Nm}), respectively.
The CM $\tilde{\gamma}$ is in the standard form (\ref{tildegamma}) in both cases and therefore one can evaluate
easily the matrix (\ref{RR}) using Eqs.~(\ref{Rst}) and (\ref{Rj}) which gives, after substitution into Eq.~(\ref{PhiA11}),
\begin{equation}\label{PhiA11p}
\left(\pi e\right)^2\Phi_{\mathcal{A}}^{\rm p}(1,1)=\frac{e^{2\tanh r}}
{\cosh^2(r)}
\end{equation}
for pure states, and
\begin{equation}\label{PhiA11m}
\left(\pi e\right)^2\Phi_{\mathcal{A}}^{\rm m}(1,1)=\frac{e^{\frac{4\sigma^{2}}{2\sigma^{2}+1}}}{2\sigma^{2}+1}
\end{equation}
for unbiased mixtures of coherent states. The corresponding negativities then satisfy
\begin{equation}\label{Lp}
\mathcal{N}_{\rm p}\geq\frac{1}{2}\left[\frac{e^{2\tanh r}}
{\cosh^2(r)}-1\right]\equiv\mathcal{L}_{\rm p}
\end{equation}
and
\begin{equation}\label{Lm}
\mathcal{N}_{\rm m}\geq\frac{1}{2}\left(\frac{e^{\frac{4\sigma^{2}}{2\sigma^{2}+1}}}{2\sigma^{2}+1}-1\right)\equiv\mathcal{L}_{\rm m}.
\end{equation}
The bounds $\mathcal{L}_{\rm p}$ and $\mathcal{L}_{\rm m}$ as well as the negativities $\mathcal{N}_{\rm p}$, Eq.~(\ref{Np}), and $\mathcal{N}_{\rm m}$, Eq.~(\ref{Nm}), are depicted in Fig.~\ref{fig2}. The figure shows that both bounds are tight in the region of small $\langle n \rangle$ (see the inset), which also proves that Eq.~(\ref{Nm}) amounts to the exact value of the negativity of quantumness for mixtures of coherent states with small mean number of thermal photons in each mode. Both lower bounds are then shown to increase with increasing $\langle n\rangle$ and the gap between the bounds ${\cal L}_{\rm p, m}$ and the numerically evaluated values of the output negativities ${\cal N}_{\rm p, m}$ gets larger. Further analysis reveals however that the lower bounds $\mathcal{L}_{\rm p}$ and $\mathcal{L}_{\rm m}$ are nonmonotonic for larger $\langle n \rangle$; they both attain a maximum at
$\langle n\rangle \approx 0.62$ and $\langle n\rangle \approx 0.52$, respectively, and then both monotonically decrease for larger values of $\langle n\rangle$; eventually, both lower bounds become trivial as they enter the region of negative values, namely $\mathcal{L}_{\rm p}<0$ for
$\langle n\rangle\gtrsim5.26$ and $\mathcal{L}_{\rm m}<0$ for $\langle n\rangle\gtrsim 1.97$.
As a final remark, note that the sum in negativity (\ref{N2}) just amounts to the so-called $\ell_1$-norm of the density matrix $\tilde{\rho}$ \cite{Takafumi_13}, i.e.,
$\sum_{\mathbf{m,n}=\mathbf{0}}^{\infty}|\tilde{\rho}_{\mathbf{m},\mathbf{n}}|=\|\tilde{\rho}\|_{\ell_1}$.
The results of the present Section thus also describe how to calculate numerically the $\ell_1$-norm for an arbitrary two-mode Gaussian state with zero means and the inequality (\ref{sumbound}) gives a simple analytical lower bound $\|\tilde{\rho}\|_{\ell_1}\geq\left(\pi e\right)^2\Phi_{\mathcal{A}}(1,1)$ on such a norm.
\section{Conclusions}\label{sec_4}
We have shown that a protocol capable of activating
nonclassical correlations in bipartite Gaussian states based
solely on Gaussian operations cannot exist.
We have also
constructed a non-Gaussian activation protocol and we have
investigated quantitatively its performance using the negativity
of quantumness as a figure of merit. Our analysis suggests that
optimal performance of the protocol is achieved if the input
Gaussian state is in the standard form. Restricting to the local
Gaussian unitaries the conjecture can be proved or disproved with
the help of Eq.~(\ref{rhoHermite}) by numerical minimization of
the negativity (\ref{N2}) with respect to the unitaries, which is
left for further research.
We believe that our results
will stimulate further exploration of the negativity of
quantumness and its interplay with other nonclassicality
indicators \cite{Adesso_10,GAMID} in the context of Gaussian
states.
\acknowledgments
L.~M. acknowledges the Project No. P205/12/0694 of GA\v{C}R and the European Social Fund and MSMT
under project No. EE2.3.20.0060. D.~M. acknowledges the support of
the Operational Program Education for Competitiveness Project No.
CZ.1.07/2.3.00/20.0060 co-financed by the European Social Fund and
Czech Ministry of Education. G.~A. acknowledges the Brazilian
agency CAPES [Pesquisador Visitante Especial-Grant No.~108/2012]
and the Foundational Questions Institute [Grant No.
FQXi-RFP3-1317]. G.~A. would also like to thank M. Barbieri and M. Piani for discussions.
|
\section*{Introduction}
\label{introduction}
Let $K$ be a field. Naturally attached to a finite distributive lattice $L$ is a $K$-algebra $K[L]$ which nowadays is called the Hibi ring of $L$. This $K$-algebra was introduced by the third author in 1987, see \cite{H}. In that paper it is shown that $K[L]$ is a normal Cohen--Macaulay domain and that $K[L]$ is Gorenstein if and only if the poset $P$ of join-irreducible elements of $L$ is pure.
Let $R$ be an arbitrary standard graded Cohen--Macaulay $K$-algebra with canonical module $\omega_R$. Then $R$ is Gorenstein if and only if $\omega_R$ is a cyclic module, and hence generated in a single degree. The condition on $\omega_R$ may be weakened in different ways. If one only requires that the generators of $\omega_R$ are of the same degree, then $R$ is called a level ring, and if one requires that there is only one generator of least degree, then we call $R$ a pseudo-Gorenstein ring.
In this paper we intend to characterize the pseudo-Gorenstein and level Hibi rings $K[L]$ in terms of $P$. For that purpose we use the basic fact, observed in \cite{H}, that a $K$-basis of $K[L]$ can be described in terms of order reversing functions $\hat{P}\to \ZZ_{\geq 0}$, and that a $K$-basis of the canonical module $\omega_L$ of $K[L]$ can be described in terms of strictly order reversing functions $\hat{P}\to \ZZ_{\geq 0}$. Here $\hat{P}=P\union\{-\infty,\infty\}$ with $-\infty<x<\infty$ for all $x\in P$, see Section~\ref{pseudo} for details. Since the property of $K[L]$ to be level, Gorenstein or pseudo-Gorenstein does not depend on the field $K$, we simply say that $L$ is level, Gorenstein or pseudo-Gorenstein if $K[L]$ has this property.
In Section~\ref{1} we briefly list conditions which are equivalent to pseudo-Gorenstein and describe the relation of this notion to that of level and Gorenstein. Since pseudo-Gorenstein rings can be identified by the property that the leading coefficient of the numerator polynomial of the Hilbert series is equal to $1$, pseudo-Gorenstein rings are much easier accessible than level rings. Theorem~\ref{classification} gives a full characterization of Hibi rings which are pseudo-Gorenstein. Indeed, it is shown that $L$ is pseudo-Gorenstein if and only if $\depth(x)+\height(x)=\rank \hat{P}$ for all $x\in P$. This is equivalent to say, that for any given $x\in P$ there exists a chain of maximal length in $\hat{P}$ passing through $x$. It may be of interest to notice that this property of the chains in $P$ has its analogue in the fact that in an affine domain $\dim R=\dim R/\pp+\height \pp$ for all $\pp\in \Spec R$.
Though Theorem~\ref{classification} characterizes Hibi rings which are pseudo-Gorenstein, it may nevertheless be difficult to apply this characterization efficiently, even for planar lattices. In Section~\ref{hyper} we introduce hyper-planar lattices which represent a natural extension of planar lattices to higher dimensions. They are defined by the property that their poset $P$ of join-irreducible elements admits a canonical chain decomposition, that is, a decomposition into pairwise disjoint maximal chains. In general such a decomposition is not unique. For hyper-planer lattices we introduce a regularity condition with the effect that the height of an element in $P$ is the same as the height of the element in the chain to which it belongs. Apart from a few exceptions we keep this regularity hypothesis on hyper-planar lattices throughout the rest of the paper. In Theorem~\ref{equallength} it is shown that a regular hyper-planar lattice is pseudo-Gorenstein if and only if all chains in a canonical chain decomposition of $P$ have the same length. For simple planar lattices it is shown in Theorem~\ref{viviana} that the regularity condition is in fact indispensable. Indeed, it is proved that a simple planar lattice is pseudo-Gorenstein if and only if it is regular and the (two) chains in a canonical chain decomposition of $P$ have the same length. Unfortunately this result cannot be extended to hyper-planar lattices as we show by an example.
The study of level Hibi rings is more difficult. There is a very nice sufficient condition on $P$ that guarantees that $L$ is level. In \cite[Theorem~3.3]{M} Miyazaki showed that $L$ is level if for all $x\in P$ all chains in $\hat{P}$ ascending from $x$ have the same length, and he showed by an example that this condition is not necessary. Let $P^\vee$ be the dual poset of $P$, i.e., the poset on the same set as $P$ but with all order relation reversed, and let $L^\vee$ be the distributive lattice whose poset of join-irreducible elements is $P^\vee$. Then it is easily seen that $L$ is level if and only if $L^\vee$ is level. Therefore it follows from Miyazaki's theorem (as remarked by him in his paper), that $L$ is also level if for all $x\in P$, all chains in $\hat{P}$ descending from $x$ have the same length. Thus we call a finite poset $P$ a {\em Miyazaki poset} if for all $x\in P$ all chains in $\hat{P}$ ascending from $x$ have the same length or all chains in $\hat{P}$ descending from $x$ have the same length. Unfortunately $L$ may be level, though $P$ is not a Miyazaki poset, and this may happen even for regular planar lattices. On the other hand, in Theorem~\ref{alsoviviana} it is shown that if $L$ is an arbitrary finite distributive lattice which is level, then for all $x,y\in P$ with $x\gtrdot y$ we must have that $\height(x)+\depth(y) \leq \rank \hat{P}+1$. Here we use the notation ``$x\gtrdot y$" to express that $x$ covers $y$, that is, $x>y$ and for any $z\in P$ either $z>x$ or $z<y$. At present we do not know whether these inequalities for all covering pairs in $P$ actually characterize the levelness of $L$. On other hand, it is shown in Theorem~\ref{hibi} that a regular planar lattice $L$ is level if and only if these inequalities hold for all covering pairs in $P$. We apply this result to give in Theorem~\ref{butterfly} an explicit description of those distributive lattices $L$ whose poset $P$ of join-irreducible elements has a special shape which we call a butterfly. In this particular case it turns out that $L$ is level if and only if the initial ideal $\ini(I_L)$ defines a level ring where $I_L$ is the defining ideal of the Hibi ring $K[L]$. One may wonder whether the regularity condition in Theorem~\ref{viviana} is really needed. In the case of a planar lattice with only one inside corner the regularity hypothesis may indeed be dropped, as shown in Theorem~\ref{diagonalposet}.
In the last section of this paper we study the pseudo-Gorenstein and level property of generalized Hibi rings. For a fixed field $K$, a poset $P$ and any
integer $r$ one defines the so-called generalized Hibi ring $\MR_r(P)$ which is naturally attached to $r$-multichains of poset ideals in $P$, see \cite{EHM}.
For $r=2$ one obtains the ordinary Hibi rings. In Theorem~\ref{new} it is shown that $R_2(P)$ is pseudo-Gorenstein if and only if $R_r(P)$ is
pseudo-Gorenstein for some $r\geq 2$, and that $R_2(P)$ is level if $R_r(P)$ is level for some $r\geq 2$.
\section{Pseudo-Gorenstein rings}
\label{1}
Let $K$ be a field and $R$ a Cohen-Macaulay standard graded $K$-algebra of dimension $d$ with canonical module $\omega_R$. We choose a presentation $R\iso S/I$ where $S=K[x_1,\ldots,x_n]$ is a polynomial ring and $I\subset \mm^2$ with $\mm=(x_1,\ldots,x_n)$. Furthermore, let $\FF$ be the graded minimal free resolution of $S/I$.
It is a simple exercise to see that the following conditions are equivalent:
\begin{enumerate}
\item[(i)] Let $a=\min\{i\:\; (\omega_R)_i\neq 0\}$. Then $\dim_K(\omega_R)_a=1$;
\item[(ii)] Let $y_1,\ldots,y_d$ be a maximal regular sequence of linear forms in $R$ and set $\bar{R}=R/(y_1,\ldots,y_d)R$. Furthermore, let $b=\max\{i\:\; \bar{R}_i\neq 0\}$. Then
$\dim_K\bar{R}_b=1$.
\item[(iii)] Let $H_R(t)=P(t)/(1-t)^d$ be the Hilbert series of $R$. Then the leading coefficient of $P(t)$ is equal to $1$.
\item[(iv)] The highest shift $c$ in the resolution $\FF$ appears in $F_{n-d}$ and
\[
\beta_{n-d, c}(S/I)=1.
\]
\end{enumerate}
We call $R$ {\em pseudo-Gorenstein} if one (or all) of the above equivalent conditions hold. The ring $R$ is called {\em level} if $\omega_R$ is generated in a single degree. It is clear from (i) that a pseudo-Gorenstein ring is level if and only if it is Gorenstein.
Let $<$ be a monomial order and assume that $S/\ini_<(I)$ is Cohen-Macaulay. It follows from (iv) that $S/\ini_<(I)$ is pseudo-Gorenstein if and only if $S/I$ is pseudo-Gorenstein. In particular, if $I\subset S$ is a toric ideal such that $S/I$ is Gorenstein and $\ini_<(I)$ is a squarefree monomial ideal, then $S/\ini_<(I)$ is pseudo-Gorenstein. Here we use that $S/\ini_<(I)$ is Cohen--Macaulay, if $I$ is a toric ideal and $\ini_<(I)$ is squarefree, see \cite[Corollary~6.6.18]{V}.
\section{The canonical module of a Hibi ring}
\label{pseudo}
Let $(L,\wedge,\vee)$ be a finite distributive lattice. An element $\alpha\in L$ is called {\em join-irreducible} if $\alpha\neq \min L$ and whenever $\alpha=\beta\vee \gamma$, then $\alpha=\beta$ or $\alpha=\gamma$. Let $P$ be the subposet of join-irreducible elements of $L$. By a well-known theorem of Birkhoff \cite{B}, one has that $L\iso \MI(P)$, where $\MI(P)$ is the lattice of poset ideals of $P$ with the partial order given by inclusion and with union and intersection as join and meet operation. Poset ideals of $P$ are subsets $\alpha$ of $P$ with the property that if $x\in \alpha$ and $y\leq x$, then $y\in \alpha$. In particular, $\emptyset$ is a poset ideal of $P$.
Given a field $K$. The Hibi ring of $L$ over $K$ is the $K$-algebra $K[L]$ generated by the elements $\alpha\in L$ and with the defining relations $\alpha\beta-(\alpha\wedge \beta)(\alpha\vee \beta)$. Identifying $L$ with $\MI(P)$, it is shown in \cite{H} that $K[L]$ is isomorphic to the toric ring generated over $K$ by the elements $u_\alpha$ with $\alpha \in L$, where $u_\alpha$ is the monomial $s\prod_{x\in \alpha}t_x$ in the polynomial ring $K[s,t_x:x\in P]$.
Let $\hat{P}$ be the poset $P\union \{\infty,-\infty\}$ with $-\infty <x<\infty$ for all $x\in P$. A map $v\: \hat{P}\to \ZZ_{\geq 0}$ is called {\em order reversing} if $v(x)\leq v(y)$ for all $x,y\in\hat{P}$ with $x\geq y$, and $v$ is a called {\em strictly order reversing} if $v(x)<v(y)$ for all $x,y\in\hat{P}$ with $x>y$.
We denote by $\MS(\hat{P})$ the set of all order reversing functions $v$ on $\hat{P}$ with $v(\infty)=0$, and by $\MT(\hat{P})$ the set of all strictly order reversing functions $v$ on $\hat{P}$ with $v(\infty)=0$.
It is shown in \cite{H} that the toric ring $K[L]$ has a $K$-basis consisting of the monomials
\begin{eqnarray}
\label{reversing}
s^{v(-\infty)}\prod_{x\in P}t_x^{v(x)}, \quad v\in \MS(\hat{P}),
\end{eqnarray}
and that the monomials
\begin{eqnarray}
\label{strict}
s^{v(-\infty)}\prod_{x\in P}t_x^{v(x)}, \quad v\in \MT(\hat{P})
\end{eqnarray}
form a $K$-basis of the canonical ideal $\omega_L\subset K[L]$. The (finite) set of elements $v\in \MT(\hat{P})$ which correspond to a minimal set of generators of $\omega_L$ will be denoted by $\MT_0(\hat{P})$.
It follows from this description of $\omega_L$ that the property of $K[L]$ to be pseudo-Gorenstein, Gorenstein or level does not depend on $K$. Thus we call $L$ itself pseudo-Gorenstein, Gorenstein or level if $K[L]$ has this property
\medskip
Note that $K[L]$ is standard graded with $\deg s^{v(-\infty)}\prod_{x\in P}t_x^{v(x)}=v(-\infty)$ for each of the monomials in (\ref{reversing}).
\medskip
Before proceeding we recall some basic concepts and notation regarding finite posets.
Let $Q$ be an arbitrary poset. A nonempty subposet $C$ of $P$ which is totally ordered is called a {\em chain} in $P$. The {\em length} of $C$ is defined to be $|C|-1$, and denoted $\ell(C)$. The rank of $Q$, denoted $\rank Q$, is defined to be the maximal length of a chain in $Q$. Let $x\in Q$. Then $\height_Q(x)$ (resp. $\depth_Q(x)$) is defined to be the maximal length of a chain descending (resp.\ ascending) from $x$ in $Q$. In the case that $Q=\hat{P}$ for some poset $P$, we omit the lower index and simply write $\height(x)$ and $\depth(x)$.
Let $x,y \in P$. It is said that $x$ {\em covers} $y$, denoted $x\gtrdot y$, if $x>y$ and there exists no $z\in P$ such that $x>z>y$.
Let as before $L$ be a finite distributive lattice. Then $L$ is called {\em simple} if there exist no elements $\alpha,\beta\in L$ with the property
$\alpha\gtrdot \beta$ and such that for each $\gamma\in L$ with $\gamma \neq\alpha,\beta$, we have $\gamma>\alpha$ or $\gamma<\beta$. Let $P$ be the poset of join-irreducible elements of $L$. Then $L$ is simple if and only if there exists no element $x\in P$ which is comparable with all elements in $P$.
\medskip
It is observed in \cite{EHS} that $\min\{v(-\infty)\:\; v \in \MT(\hat{P})\}=\rank \hat{P}$.
Thus it follows that $L$ is pseudo-Gorenstein if and only if there exists precisely one $v\in \MT(\hat{P})$ with $v(-\infty)=\rank \hat{P}$, and that $L$ is level if and only if for any $v\in \MT(\hat{P})$ there exists $v'\in \MT(\hat{P})$ with $v'(-\infty)=\rank \hat{P}$ and such that $v-v'\in \MS(\hat{P})$.
\medskip
In the following result we characterize pseudo-Gorenstein distributive lattices in terms of their poset of join-irreducible elements.
\begin{Theorem}
\label{classification}
The distributive lattice $L$ is pseudo-Gorenstein if and only if
\[\depth(x)+\height(x)=\rank \hat{P}\quad \text{for all}\quad x\in P.
\]
\end{Theorem}
\begin{proof}
Suppose first that $\depth(x)+\height(x)=\rank \hat{P}$ for all $x\in P$. This implies that for any $x\in \hat{P}$ there exists a chain $C$ of length equal to $\rank \hat{P}$ with $x\in C$. Now let $v$ be any strictly order reversing function on $\hat{P}$ with $v(\infty)=0$ and $v(-\infty)=\rank \hat{P}$. Then for any $y\in C$ we must have $v(y)=\depth(y)$. In particular, $v(x)=\depth(x)$. This shows that $v$ is uniquely determined, and proves that $L$ is pseudo-Gorenstein.
Conversely, suppose that $L$ is pseudo-Gorenstein. For all $x\in\hat{P}$ we set $v(x)=\depth(x)$ and $v'(x)=\rank \hat{P}-\height(x)$. Then both, $v$ and $v'$, are strictly order reversing functions on $\hat{P}$ with $v(\infty)=v'(\infty)=0$ and $v(-\infty)=v'(-\infty)=\rank \hat{P}$. Since $L$ is pseudo-Gorenstein we have $v=v'$. This implies that $\depth(x)+\height(x)=\rank \hat{P}$ for all $x\in P$.
\end{proof}
\section{Hyper-planar lattices}
\label{hyper}
Let $L$ be a finite distributive lattice and $P$ its poset of join-irreducible elements. We call $L$ a {\em hyper-planar lattice}, if $P$ as a set is the disjoint union of chains $C_1,\ldots, C_d$, where each $C_i$ is a maximal chain in $P$. We call such a chain decomposition {\em canonical}. Of course in general an element $x\in C_i$ may be comparable with an element $y\in C_j$ for some $j\neq i$. If this is the case and if $x\gtrdot y$, then we call the chain $x\gtrdot y$ (of length one) a {\em diagonal} of $P$ (with respect to the given canonical chain decomposition). For example, the poset depicted in Figure~\ref{counter} has two diagonals. If $d=2$, we recover the simple planar lattices.
A canonical chain decomposition of the poset $P$ of join-irreducible elements for a hyper-planar lattice $L$ is in general not uniquely determined. However we claim that if $C_1\union C_2\union\cdots \union C_s$ and $D_1\union D_2\union \cdots \union D_t$ are canonical chain decompositions of $P$, then $s=t$.
Indeed, let $\max(Q)$ denote the set of maximal elements of a finite poset $Q$. Then
\begin{eqnarray}
\label{max}
\max(P)&=&\max(C_1)\union \max(C_2)\union\cdots \union \max(C_s)\\
&=&\max(D_1)\union \max(D_2)\union\cdots \union \max(D_t).\nonumber
\end{eqnarray}
Let $\max(C_i)=\{x_i\}$ for $i=1,\ldots,s$ and $\max(D_i)=\{y_i\}$ for $i=1,\ldots,t$. Then the elements $x_i$ as well as the elements $y_i$ are pairwise distinct, and it follows from (\ref{max}) that
\[
\{x_1,x_2,\ldots,x_s\}=\{y_1,y_2,\ldots,y_t\},
\]
Hence we see that $s=t$.
\medskip
One would even expect that the
\begin{eqnarray}
\label{equal}
\{\ell(C_1),\ell(C_2),\ldots, \ell(C_s)\} =\{\ell(D_1),\ell(D_2),\ldots,\ell(D_t)\},
\end{eqnarray}
as multisets. This however is not the case. For the poset $P$ displayed in Figure~\ref{different} we have the following two canonical chain decompositions
\[
C_1=a<b<c<d<e<f, \quad C_2=g<h<i<j<k<l,
\]
and
\[
D_1=a<b<i<e<f, \quad D_2=g<h<c<d<j<k<l.
\]
Thus we see that $\ell(C_1)=\ell(C_2)=5$, while $\ell(D_1)=4$ and $\ell(D_2)=6$.
Also, note that the chain $i\gtrdot b$ is a diagonal of $P$, depicted in Figure~\ref{different}, with respect to the canonical chain decomposition $C_1\cup C_2$, but not is not a diagonal of $P$ with respect to the canonical chain decomposition $D_1\cup D_2$.
\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-5,0)(4,6)
\rput(-1.5,0){
\psline(0,0)(0,5)
\rput(0,0){$\bullet$}
\rput(0,1){$\bullet$}
\rput(0,2){$\bullet$}
\rput(0,3){$\bullet$}
\rput(0,4){$\bullet$}
\rput(0,5){$\bullet$}
\psline(2,0)(2,6)
\rput(2,0){$\bullet$}
\rput(2,1){$\bullet$}
\rput(2,2.5){$\bullet$}
\rput(2,4){$\bullet$}
\rput(2,5){$\bullet$}
\rput(2,6){$\bullet$}
\psline(0,1)(2,2.5)
\psline(0,4)(2,2.5)
\psline(0,3)(2,4)
\psline(0,2)(2,1)
\rput(-0.3,0){$a$}
\rput(-0.3,1){$b$}
\rput(-0.3,2){$c$}
\rput(-0.3,3){$d$}
\rput(-0.3,4){$e$}
\rput(-0.3,5){$f$}
\rput(2.3,0){$g$}
\rput(2.3,1){$h$}
\rput(2.3,2.5){$i$}
\rput(2.3,4){$j$}
\rput(2.3,5){$k$}
\rput(2.3,6){$l$}
}
\end{pspicture}
\end{center}
\caption{}\label{different}
\end{figure}
In order to guarantee that equality (\ref{equal}) is satisfied we have to add an extra condition on the hyper-planar lattice:
let $L$ be a hyper-planar lattice whose poset of join-irreducible elements is $P$. In what follows this will be our standard assumption and notation.
We say that $L$ is a {\em regular hyper-planar lattice}, if for any canonical chain decomposition $C_1\cup C_2\cup\ldots\cup C_d$ of $P$, and for all $x<y$ with $x\in C_i$ and $y\in C_j$ it follows that $\height_{C_i}(x)<\height_{C_j}(y)$.
\medskip
\begin{Lemma}
\label{regular}
Let $L$ be a regular hyper-planar lattice and $C_1\cup \ldots \cup C_d$ be a canonical chain decomposition of $P$. Then for all $i$ and $x\in C_i$ we have $\height_{C_i}(x)=\height_P(x)$.
\end{Lemma}
\begin{proof}
We proceed by induction on $\height_P(x)$. If $\height_P(x)=0$, then there is nothing to show. Now assume that $\height_P(x)>0$ and let $y\in P$ covered by $x$ with $\height_P(y)=\height_P(x)-1$. Say, $y\in C_j$. Since $\height_P(y)=\height_P(x)-1$ we may apply our induction hypothesis and obtain
\[
\height_P(x)-1=\height_P(y)=\height_{C_j}(y)<\height_{C_i}(x)\leq \height_P(x).
\]
This yields the desired conclusion.
\end{proof}
\begin{Corollary}
\label{ell}
Let $L$ be a regular hyper-planar lattice, and assume that besides of $C_1\cup \ldots \cup C_d$ there is still another canonical chain decomposition $D_1\union D_2\union \cdots \union D_d$ of $P$. Then
\begin{enumerate}
\item[{\em (a)}] $\{\ell(C_1),\ell(C_2),\ldots, \ell(C_d)\} =\{\ell(D_1),\ell(D_2),\ldots,\ell(D_d)\}$,
as multisets.
\item[{\em (b)}]
$\rank P=\max\{\ell(C_1),\ldots,\ell(C_d)\}$.
\item[{\em (c)}]
$\height (x)+\depth (x)=\rank \hat{P}$ for all $C_i$ with $\ell(C_i)=\rank P$ and all $x\in C_i$.
\end{enumerate}
\end{Corollary}
\begin{proof}
Let $\max(C_i)=\{x_i\}$ and $\max(D_i)=\{y_i\}$ for $i=1,\ldots,t$. We have seen in the discussion before Lemma~\ref{regular} that
\[
\{x_1,x_2,\ldots,x_d\}=\{y_1,y_2,\ldots,y_d\},
\]
Therefore, \[\{\height_P(x_1),\height_P(x_2),\ldots,\height_P(x_d)\}=\{\height_P(y_1),\height_P(y_2),\ldots,\height_P(y_d)\},\] as multi-sets. By
Lemma~\ref{regular}, $\height_{P}(x_i)=\ell(C_i)$ and $\height_{P}(y_i)=\ell(D_i)$. This together with the observation that $\rank P=\max\{\height_P(x_1),\height_P(x_2),\ldots,\height_P(x_d)\}$ proves (a) and (b).
In order to prove (c) we observe that
\begin{eqnarray*}
\rank \hat{P}&=&\ell(\hat{C_i})=\height_{\hat{C_i}}(x)+\depth_{\hat{C_i}}(x)\\
&\leq & \height (x)+\depth (x)\leq \rank \hat{P}.
\end{eqnarray*}
\end{proof}
Now we are able to characterize the regular hyper-planar lattices which are pseudo-Gorenstein.
\begin{Theorem}
\label{equallength}
Let $L$ be a regular hyper-planar lattice and $C_1\cup \ldots \cup C_d$ be a canonical chain decomposition of $P$. Then $L$ is pseudo-Gorenstein if and only if all $C_i$ have the same length.
\end{Theorem}
\begin{proof}
Suppose all $C_i$ have the same length. Then Corollary~\ref{ell} implies that $\ell(C_i)=\rank \hat{P}$ for all $i$. Let $x\in P$. Then $x\in C_i$ for some $i$, and hence $\height (x)+\depth (x)=\rank \hat{P}$, by Corollary~\ref{ell}. Therefore, by Theorem~\ref{classification}, $L$ is pseudo-Gorenstein.
Conversely, suppose that not all $C_i$ have the same length. Then Corollary~\ref{ell} implies that there exists one $C_i$ with $\ell(C_i)<\rank P$. As in the proof of Theorem~\ref{classification} we consider the strictly order reversing function $v(x)=\depth (x)$ and $v'(x)=\rank \hat{P}-\height (x)$. Let $x=\max(C_i)$. Then $v(x)=1$ and, since $L$ is regular, $v'(x)=\rank \hat{P}-(\ell(C_i)+1)>\rank \hat{P}-\rank P-1=1$. This shows that $L$ is not pseudo-Gorenstein.
\end{proof}
In \cite{M} Miyazaki showed that $L$ is level, if for all $x\in P$ all maximal chains ascending from $p$ have the same length, or all maximal chains descending from $x$ have the same length. We call a poset $P$ with this property a {\em Miyazaki poset}.
\begin{Corollary}
\label{easy}
Let $L$ be a regular hyper-planar lattice and let $C_1\union C_2\union \cdots\union C_d$ be a canonical chain decomposition of $P$. We assume that all $C_i$ have the same length. Then the following conditions are equivalent:
\begin{enumerate}
\item[{\em (a)}] $L$ is Gorenstein;
\item[{\em (b)}] $L$ is level;
\item[{\em (c)}] $P$ is a Miyazaki poset.
\end{enumerate}
\end{Corollary}
\begin{proof}
By Theorem~\ref{equallength}, $L$ is pseudo-Gorenstein. Thus (a) and (b) are equivalent, as noticed in Section 1. The implication (c)\implies (b) follows by Miyazaki \cite{M}, and (a)\implies (c) follows by Hibi's theorem \cite{H} which says that $L$ is Gorenstein if and only if $P$ is pure.
\end{proof}
For simple planar lattices Theorem~\ref{equallength} can be improved as follows.
\begin{Theorem}
\label{viviana}
Let $L$ be a simple planar lattice, and let $C_1\union C_2$ be a canonical chain decomposition of $P$.
Then the following conditions are equivalent:
\begin{enumerate}
\item[{\em (a)}] $L$ is pseudo-Gorenstein;
\item[{\em (b)}] $L$ is regular and the chains $C_1$ and $C_2$ have the same length.
\end{enumerate}
\end{Theorem}
\begin{proof}
It suffices to prove (a)\implies (b) because the implication (b) \implies (a) is a special case of Theorem~\ref{equallength}. Now for the proof of the implication (a)\implies (b) we use the characterization of pseudo-Gorenstein lattices via the Hilbert series of $K[L]$, as given in (iii), Section 1. Namely, if $H_{K[L]}(t)=P(t)/(1-t)^d$ with $d=\dim K[L]$, then $L$ is pseudo-Gorenstein if and only if the leading coefficient of $P(t)$ is one. Note that $L$ may be identified with a two-sided ladder inside an $m\times n$ rectangle with $n\leq m$ as shown in Figure~\ref{ladder}.
\texttt{}\begin{figure}[hbt]
\begin{center}
\psset{unit=1cm}
\begin{pspicture}(-5,0)(4,6)
\rput(-4,0){
\rput(-0.3,-0.3){$(0,0)$}
\psline(0,0)(2,0)
\psline(2,0)(2,1)
\psline(2,1)(3,1)
\psline(3,1)(3,2)
\psline(3,2)(6,2)
\psline(6,2)(6,3)
\psline(6,3)(7,3)
\psline(7,3)(7,5)
\psline[linestyle=dotted](2,0)(7,0)
\psline[linestyle=dotted](7,0)(7,3)
\rput(3.5,3){$L$}
\rput(7.3,5.3){$(m,n)$}
\psline(0,0)(0,3)
\psline(0,3)(2,3)
\psline(2,3)(2,4)
\psline(2,4)(3,4)
\psline(3,4)(3,5)
\psline(3,5)(7,5)
\psline[linestyle=dotted](0,3)(0,5)
\psline[linestyle=dotted](0,5)(3,5)
}
\end{pspicture}
\end{center}
\caption{}\label{ladder}
\end{figure}
Now according to \cite{ERQ} the leading coefficient of $P(t)$ is the number of maximal cyclic sublattices of $L$. Thus $L$ is pseudo-Gorenstein if and only if $L$ admits precisely one maximal cyclic sublattice. By a cyclic sublattice of $L$, we mean a sublattice inside the two-sided ladder in Figure~\ref{ladder} consisting of a chain of squares and edges as in the example shown in Figure~\ref{cyclic}. We call the cyclic sublattice maximal if it has the maximal number of squares among all cyclic sublattices contained in $L$.
\texttt{}\begin{figure}[hbt]
\begin{center}
\psset{unit=1cm}
\begin{pspicture}(-5,0)(4,6)
\rput(-3,0){
\psline(0,0)(0,1)
\psline(0,0)(1,0)
\psline(0,1)(1,1)
\psline(1,0)(1,1)
\rput(0,0){$\bullet$}
\rput(0,1){$\bullet$}
\rput(1,0){$\bullet$}
\rput(1,1){$\bullet$}
\psline(1,1)(1,2)
\psline(1,2)(1,3)
\rput(1,2){$\bullet$}
\rput(1,3){$\bullet$}
\psline(1,3)(2,3)
\psline(2,3)(2,4)
\psline(2,4)(1,4)
\psline(1,3)(1,4)
\rput(2,3){$\bullet$}
\rput(2,4){$\bullet$}
\rput(1,4){$\bullet$}
\psline(2,4)(3,4)
\rput(3,4){$\bullet$}
\psline(3,4)(4,4)
\psline(4,4)(4,5)
\psline(4,5)(3,5)
\psline(3,5)(3,4)
\rput(3,5){$\bullet$}
\rput(4,4){$\bullet$}
\rput(4,5){$\bullet$}
\psline(4,5)(5,5)
\psline(5,5)(5,6)
\psline(5,6)(4,6)
\psline(4,5)(4,6)
\rput(4,6){$\bullet$}
\rput(5,6){$\bullet$}
\rput(5,5){$\bullet$}
}
\end{pspicture}
\end{center}
\caption{}\label{cyclic}
\end{figure}
In the first step of our proof we show by induction on $n+m$ that if $L$ is pseudo-Gorenstein, then $m=n$ and the lower inside corners of $L$ are below the diagonal connecting $(0,0)$ with $(n,n)$, while the upper inside corners are above this diagonal. In other words, if $(i,j)$ is a lower inside corner of $L$, then $i\geq j$, while for an upper inside corner $(i,j)$ we have $i\leq j$. If this is the case, then we say that the inside corners do not cross the diagonal.
If $m+n=2$, then there is nothing to prove. Let $L'$ be the maximal sublattice of $L$ (again viewed as a ladder) with the bottom and top elements $(0,0)$ and $(m-1,n-1)$, respectively. Now, we consider two cases. Suppose first that the integral points of the square $[(m-2,n-2),(m-1,n-1)]$ belong to $L'$. In this case, $L'$ is simple since $L$ is simple. We claim that $L'$ is pseudo-Gorenstein. Indeed, suppose this is not the case. Then $L'$ has at least two different maximal cyclic sublattices. Then because $L$ is simple, the square $[(m-1,n-1),(m,n)]$ is a subset of $L$. Therefore, each of these maximal sublattices may be extended in $L$ with the square $[(m-1,n-1),(m,n)]$, contradicting our assumption that $L$ is pseudo-Gorenstein. Thus, the induction hypothesis implies that $m-1=n-1$, so that $m=n$. Also, the induction hypothesis implies that the inside corners of $L'$ do not cross the diagonal, and hence this is also the case for $L$. Now, suppose that $L'$ does not contain the square $[(m-2,n-2),(m-1,n-1)]$. Then we may assume that for some $j<n-1$ the ladder $L$ contains the squares $[(m-1,k),(m,k+1)]$ for all $k$ with $j\leq k\leq n-1$, but does not contain the squares $[(i,k),(i+1,k)]$ for all $i$ and $j$ with $1\leq i\leq m-2$ and $j\leq k\leq n-1$. Then $K[L']$ has the same $h$-vector as $K[L'']$ where $L''$ is a sublattice of $L$ which viewed as a ladder is contained in $[(0,0),(m-1,j)]$. Note that $j\neq m-1$, because $j<n-1\leq m$. Therefore, since $L''$ is simple, our induction hypothesis implies that $L''$ is not pseudo-Gorenstein. Hence there exist at least two maximal cyclic sublattices in $L''$, and each of these cyclic sublattices may be extended to maximal cyclic sublattices in $L$, which contradicts the fact that $L$ is pseudo-Gorenstein. Thus this second case is not possible.
Now we are ready to prove (b): Let $C_1=x_1<x_2<\cdots <x_m$ and $C_2=y_1<y_2<\cdots<y_n$. Then $L$ viewed as a two-sided ladder contains the points $(0,0)$ (corresponding to the poset ideal $\emptyset$ of $L$), and $(m,n)$ (corresponding to the poset ideal $L$). Since $L$ is pseudo-Gorenstein, it follows that $m=n$, as we have seen before. Being regular is equivalent to the condition that the inside corners of a ladder $L$ do not cross the diagonal connecting $(0,0)$ and $(n,n)$. In fact, the join-irreducible elements of $L$ establishing the chain $C_2$ can be identified with the vertices of the ladder (as displayed in Figure~\ref{ladder}) which are located on the vertical border lines of the upper border and are different from the inside corners and different from $(0,0)$, while the join-irreducible elements of $L$ forming the chain $C_1$ can be identified with the vertices of the ladder which are located on the horizontal border lines of the lower border of $L$ and are different from the inside corners and different from $(0,0)$. After this identification let $x=(i,j)\in C_2$ be and $y=(k,l)\in C_1$, Then $\height_{C_2}(x)= j$ and $\height _{C_1}(y)=k$. Assume now that $x>y$. Then this implies that $i\geq k$. Since the inside corners of $L$ do not cross the diagonal we have $j>i$, and thus $\height_{C_2}(x)=j>i\geq k=\height _{C_1}(y)$. Similarly, one shows that $\height_{C_2}(x)<\height_{C_1}(y)$, if $x<y$. This completes the proof.
\end{proof}
Theorem~\ref{viviana} is not valid if the hyper-planar lattice is not planar, as the example displayed in Figure~\ref{notvalid} demonstrates. Indeed, the lattice $L$ corresponding to $P$ is pseudo-Gorenstein, but in this example we only have one canonical chain decomposition, and the chains of this decomposition have different lengths. Moreover, $L$ is not regular.
\texttt{}\begin{figure}[hbt]
\begin{center}
\psset{unit=1cm}
\begin{pspicture}(-5,0)(4,2)
\rput(-1.7,0){
\psline(0,0)(0,2)
\rput(0,0){$\bullet$}
\rput(0,1){$\bullet$}
\rput(0,2){$\bullet$}
\psline(2,0)(2,2)
\rput(2,0){$\bullet$}
\rput(2,1){$\bullet$}
\rput(2,2){$\bullet$}
\psline(1,0)(1,2)
\rput(1,0){$\bullet$}
\rput(1,2){$\bullet$}
\psline(0,1)(1,0)
\psline(1,2)(2,1)
}
\end{pspicture}
\end{center}
\caption{}\label{notvalid}
\end{figure}
We also would like to remark that in Theorem~\ref{viviana}(b) the condition ``regular" is required. Indeed the poset shown in Figure~\ref{counter} is the poset of join-irreducible elements of a non-regular simple planar lattice $L$ for which $L$ is not pseudo-Gorenstein.
\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-5,0)(4,6)
\rput(-1.5,0){
\psline(0,0)(0,4)
\rput(0,0){$\bullet$}
\rput(0,1){$\bullet$}
\rput(0,2){$\bullet$}
\rput(0,3){$\bullet$}
\rput(0,4){$\bullet$}
\psline(2,0)(2,4)
\rput(2,0){$\bullet$}
\rput(2,1.5){$\bullet$}
\rput(2,2.5){$\bullet$}
\rput(2,3.25){$\bullet$}
\rput(2,4){$\bullet$}
\psline(0,1)(2,1.5)
\psline(0,3)(2,2.5)
}
\end{pspicture}
\end{center}
\caption{}\label{counter}
\end{figure}
\section{Level distributive lattices}
\label{level}
Throughout this section $L$ will be a finite distributive lattice and $P$ its poset of join-irreducible elements. In the previous section we recalled the fact that $L$ is level if $P$ is a Miyazaki poset. In his paper \cite{M} Miyazaki mentioned the fact that his condition on $P$ is only a sufficient condition. One may ask whether for hyper-planar lattices a stronger result is possible.
We begin with a necessary condition for levelness which is valid for any distributive lattice.
\begin{Theorem}
\label{alsoviviana}
Suppose $L$ is level. Then
\begin{eqnarray}
\label{inequality}
\height(x)+\depth(y) \leq \rank \hat{P}+1
\end{eqnarray}
for all $x,y\in P$ with $x\gtrdot y$.
\end{Theorem}
\begin{proof}
Let $x,y\in P$ such that $x$ covers $y$ and suppose that $\height(x)+\depth(y) > \rank \hat{P}+1$. We have to show that $L$ is not level.
Our assumption implies that
\[
\height(x)+\depth(y) > \rank \hat{P}+1\geq \height(x)+\depth(x)+1,
\]
and hence
\[
\depth(y)>\depth(x)+1.
\]
We show that there exists $w\in \MT_0(\hat{P})$ with $w(-\infty)>\rank \hat{P}$. This then proves that $L$ is not level.
Let $\depth(y)-\depth(x)-1=\alpha.$ Then $\alpha>0$. We define $v\: \hat{P}\to {\ZZ}_{\geq 0}$ as follows:
\[
v(z)= \left\{ \begin{array}{ll}
\depth(z)+\alpha, & \;\textnormal{if $x\geq z$, $z\neq y$}, \\ \depth(z), & \;\text{otherwise.}
\end{array} \right.
\]
Then $v\in \MT(\hat{P})$. If $v\in \MT_0(\hat{P})$, then we are done, since
\[
v(-\infty)= \depth(-\infty) +\alpha = \rank \hat{P} +\alpha\geq \rank \hat{P}+1.
\]
The last inequality follows from the fact that $\alpha>0$.
On the other hand, if $v\not\in \MT_0(\hat{P})$, then there exists $w\in\MT_0(\hat{P})$ with $v-w\in \MS(\hat{P})$. It follows that
\[
0\leq v(x)-w(x)\leq v(y)-w(y)=\depth(y)-w(y)\leq 0.
\]
Hence
\[
w(x)=v(x)= \depth(x)+\alpha= \depth(x)+\depth(y)-\depth(x)-1=\depth(y)-1.
\]
Let
\[
x=z_0>z_1>\cdots > z_k=-\infty
\]
be a chain whose length is $\height(x)$. Then
\[
w(x)<w(z_1)<\cdots < w(z_k)=w(-\infty),
\]
which implies that
\[
w(-\infty)\geq w(x)+\height(x)=(\depth(y)-1)+\height(x) > \rank \hat{P}.
\]
\end{proof}
In his paper \cite{M} Miyazaki remarked that for all $z\in P$ all chains ascending from $z$ have the same length if and only if for all $x,y\in P$ with $x\gtrdot y$, we have $\depth(y) = \depth(x)+1$.
Therefore, $P$ is a Miyazaki poset if and only if
\[
\text{ $\depth(y) = \depth(x)+1$ for all $x,y\in P$ with $x\gtrdot y$,}
\]
or
\[
\text{$\height(x) = \height(y)+1$ for all $x,y\in P$ with $x\gtrdot y$.}
\]
In either case $L$ is level.
\begin{Corollary}
Suppose $L$ is pseudo-Gorenstein and $P$ satisfies the inequality (\ref{inequality}) for all $x,y\in P$ with $x\gtrdot y$. Then $\depth(y) = \depth(x)+1$ and $\height(x) = \height(y)+1$ for all $x,y\in P$ with $x\gtrdot y$. In particular, $L$ is level and hence Gorenstein.
\end{Corollary}
\begin{proof}
For all $x,y\in P$ with $x\gtrdot y$, we have $\height (x)+\depth(x) = \rank \hat{P}$, since $L$ is pseudo-Gorenstein. Thus, by the inequality (\ref{inequality}), we have $\height (x)+\depth(y)\leq \height (x)+\depth(x)+1$, and hence $\depth(y)\leq \depth(x)+1$. On the other hand, clearly, we have $\depth(y)\geq \depth(x)+1$, which implies the first desired formula. The formula regarding height is similarly obtained. Therefore, $L$ is level, because $P$ is a Miyazaki poset.
\end{proof}
As mentioned before, if $P$ is not a Miyazaki poset, then $L$ may nevertheless be level, and this may happen even if $L$ is a regular simple planar lattices.
Figure~\ref{butterflyposet} shows a poset which is not a Miyazaki poset. However its ideal lattice is a regular simple planar lattice and is level.
\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-5,2)(4,6)
\rput(-1.5,2){
\psline(0,0)(0,1.5)
\rput(0,0){$\bullet$}
\rput(0,1.5){$\bullet$}
\psline(2,0)(2,2)
\rput(2,0){$\bullet$}
\rput(2,1){$\bullet$}
\rput(2,2){$\bullet$}
\psline(0,0)(2,2)
\psline(0,1.5)(2,0)
}
\end{pspicture}
\end{center}
\caption{}\label{butterflyposet}
\end{figure}
The following result shows that for regular planar lattices the necessary condition for levelness formulated in Theorem~\ref{alsoviviana} is also sufficient.
\begin{Theorem}
\label{hibi}
Let $L$ be a regular planar lattice. Then the following conditions are equivalent:
\begin{enumerate}
\item[{\em (a)}] $L$ is level;
\item[{\em (b)}] $\height(x)+\depth(y) \leq \rank \hat{P}+1$ for all $x,y\in P$ with $x\gtrdot y$;
\item[{\em (c)}] for all $x,y\in P$ with $x\gtrdot y$, either $\depth(y) = \depth(x)+1$ or $\height(x) = \height(y)+1$.
\end{enumerate}
\end{Theorem}
\begin{Remark}
\label{check diagonal}
{\em (i) Observe that a Miyazaki poset satisfies condition (c). On the other hand, Figure~\ref{butterflyposet} shows a poset satisfying condition (c) which is not Miyazaki.
(ii) Let $C_1\cup C_2$ be a canonical chain decomposition of $P$. The inequality in (b) and the equations in (c) are always satisfied for those $x\gtrdot y$ for which $x$ and $y$ belong to the same chain in the decomposition. Hence it suffices to check the inequality in (b) and equations in (c) only for diagonals. Indeed, this fact follows directly from Lemma~\ref{regular}. For instance, if $x$ covers $y$ and both belong to the same chain, then, by Lemma~\ref{regular}, $\height (x)=\height(y)+1$. Thus $\height(x)+\depth(y)=\height(y)+1+\depth(y)\leq \rank \hat{P} +1$. For (c) we only need to observe that the condition $\height (x)=\height(y)+1$ for any $x$ that covers $y$ in the same chain, is always fulfilled, again by Lemma~\ref{regular}. }
\end{Remark}
\medskip
Before proving Theorem~\ref{hibi} we will need the following result.
\begin{Lemma}
\label{biggerone}
Let $L$ be a regular planar lattice. Let $C_1\union C_2$ be a canonical chain decomposition of $P$, and assume that $\ell(C_1)=\rank P$ (cf.\ Corollary~\ref{ell}). Suppose that $P$ satisfies condition (b) of Theorem~\ref{hibi}. Then for every $v\in \MT_0(\hat{P})$ we have $v(\max (C_1))=1$.
\end{Lemma}
\begin{proof}
Assume that $v(\max(C_1))>1$. Then $v(z)\geq \depth(z)+1$ for all $z\in C_1$.
Let
\[
v'(x)= \left\{ \begin{array}{ll}
v(x)-1, & \;\textnormal{if $v(x)\geq \depth(x)+1$\; (I),} \\ v(x), & \;\text{if $v(x)=\depth(x)$\hspace{0.8cm} (II),}
\end{array} \right.
\]
for all $x\in \hat{P}$.
We show that $v'\in \MT(\hat{P})$ and $v-v'\in \MS(\hat{P})$. Since $v'\neq v$, this will then show that $v\not\in \MT_0(\hat{P})$, a contradiction. Indeed, to see that $v'\in \MT(\hat{P})$ we have to show that $v'(x)<v'(y)$ for all $x\gtrdot y$. If both $x$ and $y$ satisfy (I) or (II), then the assertion is trivial. If $x$ satisfies (I) and $y$ satisfies (II), then $v'(x)=v(x)-1<v(y)=v'(y)$, and if $x$ satisfies (II) and $y$ satisfies (I), then $v(x)=\depth(x)\leq \depth(y)-1\leq v(y)-2$. Hence $v(x)<v(y)-1$, and this implies that $v'(x)<v'(y)$.
It remains to be shown that $v-v'\in \MS(\hat{P})$ which amounts to prove that $v(x)-v'(x)\leq v(y)-v'(y)$ for all $x\gtrdot y$. For this we only need to show that we cannot have $v'(x)=v(x)-1$ and $v(y)=v'(y)$, or, equivalently, that $v(x)\geq \depth(x)+1$ and $v(y)=\depth(y)$ is impossible.
Assume to the contrary that there exist $x\gtrdot y$ with $v(x)\geq \depth(x)+1$ and $v(y)=\depth(y)$ . Then $y\not\in C_1$ since $v(z)\geq \depth(z)+1$ for all $z\in C_1$. Thus, we may either have $x\in C_1$ and $y\in C_2$, or $x,y\in C_2$.
In the first case, since $\height(x)+\depth(y)\leq \rank \hat{P}+1$ by assumption, and since $\rank \hat{P}=\height(x)+\depth(x)$ due to the regularity of $L$ (see Corollary~\ref{ell}), we get $\depth(y)\leq \depth(x)+1$, and hence $\depth(y)=\depth(x)+1$. Therefore, $\depth(y)=\depth(x)+1\leq v(x)<v(y)$, a contradiction.
Finally, let $x,y\in C_2$. Since $v(x)<v(y)$, it follows that $\depth(y)>\depth(x)+1$. Therefore, the longest chain from $y$ to $\infty$ cannot pass through $x$. This implies that there exists $z\in C_1$ with $z\gtrdot y$. As in the first case we then deduce that $v(y)>\depth(y)$. So we get again a contradiction.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{hibi}]
(a) \implies (b) follows from Theorem~\ref{alsoviviana}.
(b) \implies (c): Let $C_1\cup C_2$ be a canonical chain decomposition of $P$ with $|C_1|\geq |C_2|$.
If $x,y \in C_1$ or $x,y \in C_2$, then by Lemma~\ref{regular}, it follows that $\height(x) = \height(y)+1$.
Next suppose that $x\in C_1$. Since $L$ is regular, we may apply Corollary~\ref{ell} and conclude that $\height(x)+\depth(x)=\rank \hat{P}$. Thus, by (b), we get $\depth (y)\leq \depth (x)+1$. On the other hand, it is clear that $\depth (y)\geq \depth (x)+1$. So that $\depth (y)=\depth (x)+1$.
Finally, if $y\in C_1$, then by Corollary~\ref{ell} we have $\height(y)+\depth(y)=\rank \hat{P}$. As in the previous case, we conclude that $\height(x) = \height(y)+1$.
(c)\implies (b): If $\depth(y)=\depth(x)+1$, then $\height(x)+\depth(y)=\height(x)+\depth(x)+1\leq \rank \hat{P}+1$, and if $\height (x)=\height(y)+1$, then $\height(x)+\depth(y)=\height (y)+\depth(y)+1\leq \rank \hat{P}+1$.
(b) \implies (a): As in Lemma~\ref{biggerone} we let $C_1\union C_2$ be a canonical chain decomposition of $P$, and may assume that $\ell(C_1)=\rank P\geq \ell(C_2)$. Let $v\in \MT_0(\hat{P})$. We will show that there exists $v'\in \MT(\hat{P})$ with $v'(-\infty)=\rank \hat{P}$ and such that $v-v'\in \MS(\hat{P})$. Since $v$ is a minimal generator it follows that $v=v'$, and we are done.
In order to construct $v'$ we consider the subposet $Q$ of $P$ which is obtained from $P$ by removing the maximal elements $\max(C_1)$ and $\max(C_2)$. We define on $\hat{Q}$ the strictly order reversing function $u$ by $u(\infty)=0$, and $u(z)=v(z)-1$ for all other $z\in \hat{Q}$. We notice that the ideal lattice of $Q$ is again a regular planar lattice satisfying (b). Indeed, assume that there exist $x\gtrdot y$ with $x,y\in Q$ such that $\height_{\hat{Q}}(x)+
\depth_{\hat{Q}}(y)>\rank \hat{Q}+1=\rank \hat{P}$. Since $\height_{\hat{Q}}(x)=\height(x)$ and $\depth(y)=\depth_{\hat{Q}}(y)+1$, it follows that
\[
\height(x)+\depth(y)=\height_{\hat{Q}}(x)+\depth_{\hat{Q}}(y)+1>\rank \hat{P}+1,
\]
a contradiction.
Therefore, by induction on the rank we may assume that the ideal lattice of $Q$ is level. Hence there exists $w\in \MT(\hat{Q})$ with $w(-\infty)=\rank \hat{Q}=\rank \hat{P}-1$ and such that $u-w\in \MS(\hat{Q})$. Set $v'(z) =1+w(z)$ for all $z\in A=Q\union\{-\infty\}$. Then $v'$ is a strictly order reversing function on $A$ with $v'(-\infty)=\rank \hat{P}$ and such that $v-v'$ is order reversing on $A$. It remains to define $v'(C_i)$ for $i=1,2$ in a way such that $v'\in \MT(\hat{P})$ and $v-v'\in \MS(\hat{P})$. We have to set $v'(\max(C_1))=1$ since $v(\max(C_1))=1$, and of course $v'(\infty)=0$. Let $x=\max(C_2)$ and let $z\in C_2$ be the unique element with $x\gtrdot z$. We set $v'(x)= v(x)-u(z)+w(z)= v(x)-v(z)+1+w(z)$, and claim that this $v'$ has the desired properties. Indeed, $v'(x)=v(x)-(v(z)-1-w(z))\leq v(x)$ and $v'(x)<1+w(z)=v'(z)$, since $v(x)<v(z)$. If $z$ is the only element covered by $x$, we are done. Otherwise, there exists $y\in C_1$ with $x\gtrdot y$ and it remains to be shown that $v'(y)>v'(x)=v(x)-v(z)+1+w(z)$. Suppose we know that $\depth_{\hat{Q}}(y)\geq w(z)$, then
\[
v'(y)=w(y)+1\geq \depth_{\hat{Q}}(y)+1>w(z)\geq v'(x),
\]
as desired, since $v(x)-v(z)+1\leq 0$. Thus in order to complete the proof we have to show that $\depth_{\hat{Q}}(y)\geq w(z)$. Since the ideal lattice of $\hat{Q}$ is regular, this is equivalent to showing that
\begin{eqnarray}
\label{last}
w(z)\leq \rank \hat{Q}-\height_{\hat{Q}}(y).
\end{eqnarray}
The assumption (b) and Corollary~\ref{ell}(c) imply that
\[
\height(x)+\depth(y)\leq \rank \hat{P}+1=\height(y)+\depth(y)+1,
\]
so that $\height(x)\leq \height(y)+1$. This yields
\begin{eqnarray}
\label{good}
\height(x)= \height(y)+1
\end{eqnarray}
since $\height(x)\geq \height(y)+1$ is always valid.
On the other hand, since $L$ is regular, Lemma~\ref{regular} implies that $\height_P(x)= \height_{C_2}(x) =\height_{C_2}(z)+1=\height_{P}(z)+1$. This implies that $\height (x)=\height (z)+1$. So together with (\ref{good}) we then conclude that $\height(y)=\height(z)$. Since $\height_{\hat{Q}}(y)= \height(y)$ and $\height(z)= \height_{\hat{Q}}(z)$, inequality (\ref{last}) becomes $w(z)\leq \rank \hat{Q}-\height_{\hat{Q}}(z)$, and since $w(-\infty) =\rank \hat{Q}$, this inequality indeed holds. This completes the proof of the theorem.
\end{proof}
In the following theorem we discuss an example of a poset for which the conditions of Theorem~\ref{alsoviviana} can be made more explicit. Let $P$ be a finite poset with a canonical chain decomposition $C_1\cup C_2$ with $2\leq |C_1|\leq |C_2|$. For $i=1,2$, let $x_i$ be the maximal and $y_i$ the minimal element of $C_i$. We call $P$ a {\em butterfly poset} (of type $(C_1,C_2)$), if $x_1\gtrdot y_2$ and $x_2\gtrdot y_1$ are the only diagonals of $P$. Figure~\ref{butterflyposet} displays a butterfly poset. Obviously, the ideal lattice of a butterfly poset is regular.
For the next result, we need some notation. The Hibi ring $K[L]$ can be presented as the quotient ring $T/I_L$, where $T$ is the polynomial ring over $K$ in the variables $x_{\alpha}$ with $\alpha\in L$ and where $I_L$ is generated by the binomials $x_{\alpha}x_{\beta}-x_{\alpha\vee \beta}x_{\alpha\wedge \beta}$. In the following theorem, we consider a monomial order $<$ given by a height reverse lexicographic monomial order, that is, the
reverse lexicographic monomial order induced by a total ordering of the variables satisfying $x_{\alpha} <
x_{\beta}$ if $\height_{L} (\alpha)> \height_{L} (\beta)$.
\begin{Theorem}
\label{butterfly}
Let $P$ be a butterfly poset of type $(C_1,C_2)$, and $L$ its ideal lattice. Then the following conditions are equivalent:
\begin{enumerate}
\item[{\em (a)}] $T/\mathrm{in}_<(I_L)$ is level;
\item[{\em (b)}] $L$ is level;
\item[{\em (c)}] $\height(x)+\depth(y) \leq \rank \hat{P}+1$ for all $x,y \in P$ with $x\gtrdot y$;
\item[{\em (d)}] for all $x,y \in P$ with $x\gtrdot y$, either $\depth(y) = \depth(x)+1$ or $\height(x) = \height(y)+1$;
\item[{\em (e)}] $|C_1|=2$.
\end{enumerate}
\end{Theorem}
\begin{proof}
(a)\implies (b) is well-known. Conditions (b), (c) and (d) are equivalent by Theorem~\ref{hibi}, since $L$ is regular.
(c)\implies (e): Since $2\leq |C_1|\leq |C_2|$, we have $\height (x_1)=|C_1|$, $\depth (y_2)=|C_2|$ and $\rank \hat{P}=|C_2|+1$.
On the other hand, by condition (b), we have $\height (x_1)+\depth (y_2)\leq \rank \hat{P}+1$. So, $|C_2|\leq 2$, and hence $|C_2|=2$.
(e)\implies (a): It is shown in \cite{H} that $\ini_<(I_L)$ is generated by the monomials $x_{\alpha}x_{\beta}$ where $\alpha$ and $\beta$ are incomparable elements of $L$. Thus, $\mathrm{in}_<(I_L)$ is the Stanley-Reisner ideal of the order complex $\Delta$ of $L$. It is known that $\Delta$ is pure shellable, and hence Cohen-Macaulay, see \cite[Theorem~6.1]{Bj}. Since $\dim (T/\ini_{<}(I_L))=\dim(T/I_L)=|P|+1$, it follows that $\pd (T/\ini_{<} (I_L))=\pd (T/I_L)=|L|-|P|-1$.
Moreover, since $T/\ini_{<}(I_L)$ and $T/I_L$ are Cohen-Macaulay, their regularity is given by the degree of the numerator polynomial of their Hilbert series. Hence, since both Hilbert series coincide, their regularity is the same and we obtain $\reg (T/\mathrm{in}_<(I_L))=\reg (T/I_L)=|P|-\rank P$, by \cite{EHS}. Since $|C_1|=2$, the lattice $L$ viewed as a ladder is of the form as shown in Figure~\ref{butterflylattice}. So, we see that $|L|=3|C_2|+1$. Therefore, since $|P|=|C_2|+2$, we see that $\pd (T/\mathrm{in}_<(I_L))=2|C_2|-2$ and $\reg (T/\mathrm{in}_<(I_L))=2$.
\texttt{}\begin{figure}[hbt]
\begin{center}
\psset{unit=1cm}
\begin{pspicture}(-5,0)(4,3)
\rput(-4,0){
\rput(-0.4,-0.4){$(0,0)$}
\psline(0,0)(2,0)
\psline(3,0)(5,0)
\psline(6,0)(7,0)
\psline(0,0)(0,1)
\psline(1,0)(1,2)
\psline(2,0)(2,2)
\psline(3,0)(3,2)
\psline(4,0)(4,2)
\psline(5,0)(5,2)
\psline(7,0)(7,2)
\psline(6,0)(6,2)
\psline(8,1)(8,2)
\psline(0,1)(2,1)
\psline(3,1)(5,1)
\psline(6,1)(8,1)
\psline(1,2)(2,2)
\psline(3,2)(5,2)
\psline(6,2)(8,2)
\psline[linestyle=dotted](2,0)(3,0)
\psline[linestyle=dotted](5,0)(6,0)
\psline[linestyle=dotted](2,1)(3,1)
\psline[linestyle=dotted](5,1)(6,1)
\psline[linestyle=dotted](2,2)(3,2)
\psline[linestyle=dotted](5,2)(6,2)
\rput(8.4,2.4){$(|C_2|,2)$}
}
\end{pspicture}
\end{center}
\caption{}\label{butterflylattice}
\end{figure}
So to prove (a), it is enough to show that $\beta_{2|C_2|-2,j}(T/\mathrm{in}_<(I_L))=0$, for all $j< 2|C_2|$. By Hochster's formula, we have
\[
\beta_{2|C_2|-2, j}(T/\mathrm{in}_<(I_L))=\sum_{W\subseteq V, |W|=j}\dim _{K}\widetilde{H}_{j-2|C_2|+1}({\Delta}_W;K),
\]
where $V$ is the vertex set of $\Delta$, and where ${\Delta}_W$ denotes as usual the subcomplex of $\Delta$ induced by $W$. For $W\subseteq V$ with
$|W|=j<2|C_2|-1$, $\dim _{K}\widetilde{H}_{j-2|C_2|+1}({\Delta}_W;K)=0$, since $j-2|C_2|+1<0$. Let $W\subseteq V$ with $|W|=2|C_2|-1$. We show that
$\dim _{K}\widetilde{H}_{0}({\Delta}_W;K)=0$ or equivalently ${\Delta}_W$ is a connected simplicial complex. As the connectedness of ${\Delta}_W$ is equivalent to the connectedness of the 1-skeleton ${\Delta}_W^{(1)}$ of ${\Delta}_{W}$, we show that the graph $G={\Delta}_W^{(1)}$ is connected. The vertices of $G$ correspond to the lattice points of the elements in $W$ which is a certain subset of the lattice points of the ladder displayed in Figure~\ref{butterflylattice}. Two vertices of $G$ are adjacent if they are contained in a chain in the lattice $L$. In other words, two vertices $(i,j)$ and $(i',j')$ with $j\leq j'$ are adjacent in $G$ if $j=j'$ or $i\leq i'$.
Let $v=(i,j)$ and $w=(i',j')$ be two nonadjacent vertices in $G$, so that we may assume $j<j'$ and $i'<i$. We show that there exists a path in $G$ connecting $v$ and $w$.
Case 1. There exists $(s,t)\in W$ such that either $s\leq i'$, $t\leq j$ or $s\geq i$, $t\geq j'$. Then $(s,t)$ is adjacent to both of $v$ and $w$, and hence $v$ and $w$ are connected.
Case 2. There exists $s$ with $i'<s<i$ such that $(s,j),(s,j')\in W$. Then we get a path in $G$ between $v$ and $w$ passing through $(s,j)$ and $(s,j')$.
Assume that $W$ does not satisfy the conditions given in Case 1 and Case 2.
Let $A=\{(s,t)\in W:i'<s<i, t=j~\mathrm{or}~t=j'\}$. Then $|A|\leq i-i'-1$.
First suppose that $j=0$ and $j'=1$. Also, set $A_1=\{(s,0):i\leq s\leq |C_2|-1\}$, $A_2=\{(s,1):0\leq s\leq i'\}$ and $A_3=\{(s,2):1\leq s<i\}$. Since we are not in Case 1 and Case 2, it follows that $W\subset A_1\cup A_2\cup A_3\cup A$. Therefore,
\begin{eqnarray*}
|W|&\leq & |A_1|+|A_2|+|A_3|+|A| \leq (|C_2|-i)+(i'+1)+(i-1)+(i-i'-1)\\
&=& |C_2|+i-1< 2|C_2|-1.
\end{eqnarray*}
The last inequality follows since $i<|C_2|$. This contradicts the fact that $|W|=2|C_2|-1$.
Now suppose that $j=0$ and $j'=2$ and set $A_1=\{(s,0):i\leq s\leq |C_2|-1\}$, $A_2=\{(s,1):0\leq s\leq |C_2|\}$, $A_3=\{(s,2):1\leq s\leq i'\}$.
We claim that $|A_2\cap W|\leq |A_2|-2$, unless there is a path in $G$ connecting $v$ and $w$. Indeed, we may assume that either $(i,1)$ or $(i',1)$ belongs to $W$. If both vertices belong to $W$, then there is a path in $G$ between $v$ and $w$. Without loss of generality, we may assume that $(i,1)\in W$ and $(i',1)\notin W$. In that case, non of the elements in the nonempty set $\{(s,1):s<i'\}$ belongs to $W$, because otherwise we get a path connecting $v$ and $w$ in $G$. This proves the claim.
Since we are not in Case 1 and Case 2, we have $W\subset A_1\cup (A_2\cap W)\cup A_3\cup A$. As we may assume that $|A_2\cap W|\leq |A_2|-2$, we get
\begin{eqnarray*}
|W|&\leq & |A_1|+|A_2\cap W|+|A_3|+|A| \leq (|C_2|-i)+(|C_2|-1)+(i')+(i-i'-1)\\
&=& 2|C_2|-2,
\end{eqnarray*}
a contradiction.
Finally, $j=1$ and $j'=2$ is similarly treated as the case $j=0$ and $j'=1$.
\end{proof}
As a straightforward consequence of the next result it can be seen that the implication (a)\implies (c) in Theorem~\ref{hibi} is in general not valid for non-planar lattices, not even for hyper-planar lattices.
\begin{Theorem}
\label{longchain}
Let $P=P_1\union P_2$ be a finite poset with the property that the elements of $P_1$ and $P_2$ are incomparable to each other, and suppose that $P_2$ is a chain of length $r$. Let $L$ be the ideal lattice of $P$. Then $L$ is level for all $r\gg 0$.
\end{Theorem}
\begin{proof}
Let $L_1$ by the ideal lattice of $P_1$, and $L_2$ that of $P_2$. It is observed in \cite{HHR} and easy to see that $K[L]\iso K[L_1]*K[L_2]$ (which is the Segre product of $K[L_1]$ and $K[L_2]$). By \cite[Theorem 4.3.1]{GW} we have $\omega_L\iso \omega_{L_1}*\omega_{L_2}$, where for graded $K[L_1]$-module $M$ and a graded $K[L_2]$-module $N$ the homogeneous components of the Segre product $M*N$ are given by $(M*N)_i=M_i\tensor_KN_i$ for all $i$. Now if $L_2$ is a chain of length $r$, then $S=K[L_2]$ is a polynomial ring of dimension $r+2$, and hence $\omega_{L_2}\iso S(-r-2)$, see for example \cite[Proposition 3.6.11 and Example 3.6.10]{BH}. Hence we obtain
\begin{eqnarray}
\label{truncation}
\omega_L\iso \omega_{L_1}*S(-r-2)\iso (\omega_{L_1})_{\geq r+2},
\end{eqnarray}
where for a graded module $M$ and any integer $s$ we set $M_{\geq s}=\Dirsum_{i\geq s}M_i$. Let $t$ be the highest degree of a generator in a minimal set of generators of $\omega_{L_1}$. Then (\ref{truncation}) implies that $\omega_{L_1}$ is generated in the single degree $r+2$ if $r+2\geq t$. Thus for any $r\geq t-2$ we see that $L$ is level.
\end{proof}
We would like to mention that the arguments given in the proof of Theorem~\ref{longchain} yield the following slightly more general result: for an arbitrary finite poset $P$ we set $\gamma(P)=\max\{v(-\infty)\: \; v\in \MT_{0}(\hat{P})\}$. Note that $\gamma(P)$ is the highest degree of a generator in a minimal set of generators of the canonical module of the ideal lattice of $P$. Now let $P=P_1\union P_2$ and suppose that the elements of $P_1$ and $P_2$ are incomparable. Furthermore, assume that the ideal lattice $L_2$ of $P_2$ is level. Then the ideal lattice $L$ of $P$ is level if $\gamma(P_2)\geq \gamma(P_1)$.
\medskip
Computational evidence leads us to conjecture that the equivalent conditions given in Theorem~\ref{hibi} do hold for any planar lattice (without any regularity assumption). In support of this conjecture we have the following result.
\begin{Theorem}
\label{diagonalposet}
Let $L$ be a simple planar lattice whose poset $P$ of join-irreducible elements has the single diagonal $x\gtrdot y$ with respect to a canonical chain decomposition. Then the following conditions are equivalent:
\begin{itemize}
\item[(a)] $L$ is level;
\item[(b)] $\height(x)+\depth(y) \leq \rank \hat{P}+1$;
\item[(c)] either $\depth(y) = \depth(x)+1$ or $\height(x) = \height(y)+1$.
\end{itemize}
\end{Theorem}
Observe that for all covering pairs $u,v\in P$ which are different from the diagonal, conditions (b) and (c) are always satisfied.
\begin{proof}[Proof of Theorem~\ref{diagonalposet}]
(a) $\Rightarrow$ (b) follows from Theorem~\ref{alsoviviana}.
(b) $\Rightarrow$ (c): Let $C_1\cup C_2$ be a canonical chain decomposition of $P$ and $x\gtrdot y$ its unique diagonal with $x\in C_1$ and $y\in C_2$. We define the integers $a=|\{z\in \hat{P}:z>x\}|$, $b=|\{z\in \hat{P}: z<x,z\notin C_2\}|$, $c=|\{z\in \hat{P}:z>y, z\notin C_1\}|$ and $d=|\{z\in \hat{P}:z<y\}|$, see Figure~\ref{diagonal}. Then $\height (y)=d$ and $\depth (x)=a$. On the other hand,
$\depth (y)=\max\{c, a+1\}$ and $\height (x)=\max\{b, d+1\}$.
\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-5,-2)(4,6)
\rput(-1.5,-1){
\rput(0,0){$\bullet$}
\rput(0,2.9){$\bullet$}
\rput(0,5){$\bullet$}
\psline(1,-1)(0,0)
\psline(1,-1)(2,0)
\psline(1,6)(0,5)
\psline(1,6)(2,5)
\rput(1,-1){$\bullet$}
\rput(1,6){$\bullet$}
\rput(2,0){$\bullet$}
\rput(2,1.9){$\bullet$}
\rput(2,5){$\bullet$}
\psline(0,2.9)(2,1.9)
\rput(-0.65,1.43){$b$}
\rput(-0.65,4.3){$a$}
\rput(2.65,4.3){$c$}
\rput(2.65,1.43){$d$}
\rput(0.3,3.05){$x$}
\rput(1.7,1.75){$y$}
\rput(1,-1.3){$-\infty$}
\rput(1,6.3){$\infty$}
\psline[linestyle=dotted](0,0)(0,5)
\psline[linestyle=dotted](2,0)(2,5)
}
\end{pspicture}
\end{center}
\caption{}\label{diagonal}
\end{figure}
Suppose that none of the equalities in condition (c) hold. Then $\depth (y)>\depth (x)+1$ and $\height (x)> \height (y)+1$. So, $\max \{c,a+1\}>a+1$, and hence $c>a+1$ and $\depth (y)=c$. Similarly, $\max \{b,d+1\}>d+1$, and hence $b>d+1$ and $\height (x)=b$. Note that $\hat{P}$ has just three maximal chains whose lengths are $a+b$, $c+d$ and $a+d+1$, so that $\rank \hat {P}=\max \{a+b, c+d\}$. Thus (b) implies $b+c\leq \max \{a+b+1 , c+d+1\}$, a contradiction. Therefore one of the desired equalities in (c) hold for $x,y$.
(c)\implies (a):
Let
\begin{eqnarray*}
C_{1} &:& - \infty = x'_{b} < x'_{b-1} < \cdots < x'_{1} < x
< x_{1} < \cdots < x_{a-1} < x_{a} = \infty, \\
C_{2} &:& - \infty = y'_{d} < y'_{d-1} < \cdots < y'_{1} < y
< y_{1} < \cdots < y_{c-1} < y_{c} = \infty.
\end{eqnarray*}
The condition (c) guarantees that either $a + 1 \geq c$ or $b \leq d + 1$.
We may assume that $a + 1 \geq c$. The case $b \leq d + 1$ is treated similarly by replacing $P$ by $P^{\vee}$.
Case 1. Suppose that a longest chain of $\hat{P}$ is
\[
- \infty = y'_{d} < y'_{d-1} < \cdots < y'_{1} < y < x <
x_{1} < \cdots < x_{a-1} < x_{a} = \infty.
\]
In other words, one has $b \leq d + 1$.
The rank of $\hat{P} = a + d + 1$.
Let $v \in\MT(\hat{P})$
with $v(-\infty) > a + d + 1$. We distinguish several cases:
\medskip
(i) Let $v(x) > a$. Then there are $i$ and $j$ with
$v(x_{i}) - v(x_{i+1}) \geq 2$ and $v(y_{j}) - v(y_{j+1}) \geq 2$,
where $x_{0} = x$ and $y_{0} = y$.
Let $u \in \MS(\hat{P})$
with $u(z) = 1$ if either $z \leq x_{i}$ or $z \leq y_{j}$
and $u(z) = 0$, otherwise. Then $v - u\in \MT(\hat{P})$.
\medskip
(ii) Let $v(x)=a$. In this case $v(y)\geq a+1$.
If $v(y) = a + 1$, then there are $i$ and $j$ with
$v(x'_{i+1}) - v(x'_{i}) \geq 2$ and $v(y'_{j+1}) - v(y'_{j}) \geq 2$,
where $x'_{0} = x$ and $y'_{0} = y$.
Let $u \in\MS(\hat{P})$ with
with $u(z) = 1$ if either $z \leq x'_{i+1}$ or $z \leq y'_{j+1}$
and $u(z) = 0$, otherwise. Then $v - u\in\MT(\hat{P})$.
If $v(y) > a + 1$, then there are $i$ and $j$ with
$v(y_{j}) - v(y_{j+1}) \geq 2$
and $v(x'_{i+1}) - v(x'_{i}) \geq 2$,
where $x'_{0} = x$ and $y_{0} = y$.
Let $u\in\MS(\hat{P})$
with $u(z) = 1$ if either $z \leq x'_{i+1}$ or $z \leq y_{j}$
and $u(z) = 0$, otherwise. Then $v - u\in\MT(\hat{P})$.
\bigskip
Case 2. Suppose that a longest chain of $\hat{P}$ is
\[
- \infty = x'_{b} < x'_{b-1} < \cdots < x'_{1} < x <
x_{1} < \cdots < x_{a-1} < x_{a} = \infty.
\]
In other words, one has $b \geq d + 1$.
The rank of $\hat{P} = a + b$.
Let $v \in \MT(\hat{P})$
with $v(-\infty) > a + b$. We distinguish several cases:
\medskip
(i) Let $v(x) > a$. Then there are $i$ and $j$ with
$v(x_{i}) - v(x_{i+1}) \geq 2$ and $v(y_{j}) - v(y_{j+1}) \geq 2$,
where $x_{0} = x$ and $y_{0} = y$.
Let $u \in\MS(\hat{P})$
with $u(z) = 1$ if either $z \leq x_{i}$ or $z \leq y_{j}$
and $u(z) = 0$, otherwise. Then $v - u\in \MT(\hat{P})$.
\medskip
(ii) Let $v(x) = a$.
Then there is $i$ with
$v(x'_{i+1}) - v(x'_{i}) \geq 2$,
where $x'_{0} = x$.
Since $b \geq d + 1$,
one has either
$v(y) - v(x) \geq 2$ or
there is $j$ with
$v(y'_{j+1}) - v(y'_{j}) \geq 2$,
where $y'_{0} = y$.
If $v(y) - v(x) \geq 2$, then there is $j$ with
$v(y_{j}) - v(y_{j+1}) \geq 2$, where $y_{0} = y$.
Let $u\in \MS(\hat{P})$
with $u(z) = 1$ if either $z \leq x'_{i+1}$ or $z \leq y_{j}$
and $u(z) = 0$ otherwise. Then $v - u\in \MT(\hat{P})$.
Suppose that there is $j$ with
$v(y'_{j+1}) - v(y'_{j}) \geq 2$,
where $y'_{0} = y$.
Let $u\in \MS(\hat{P})$
with $u(z) = 1$ if either $z \leq x'_{i+1}$ or $z \leq y'_{j+1}$
and $u(z) = 0$, otherwise. Then $v - u\in \MT(\hat{P})$.
\medskip
The discussions in both cases show that every
$v\in \MT(\hat{P})$
which belongs to ${\mathcal T}_{0}(\hat{P})$
satisfies $v(-\infty) = \rank(\hat{P})$.
Hence $L$ is level, as desired.
\end{proof}
\texttt{}\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-5,0)(4,6)
\rput(-4,0){
\rput(3.4,3.6){$s$}
\psline(0,0)(7,0)
\psline(7,0)(7,5)
\psline(0,0)(0,2.2)
\psline(0,2.2)(3.7,2.2)
\psline(3.7,2.2)(3.7,5)
\psline(3.7,5)(7,5)
\rput(3.5,-0.3){$n$}
\rput(7.3,2.5){$m$}
\rput(1.85,2.5){$t$}
}
\end{pspicture}
\end{center}
\caption{}\label{ladder with one corner}
\end{figure}
\begin{Corollary}
\label{onecorner}
Let $I$ be the ladder determinantal ideal of 2-minors of the ladder as displayed in Figure~\ref{ladder with one corner}. Then $S/I$ is level if and only if $\min\{m,n\}\leq s+t$.
\end{Corollary}
\begin{proof}
First note that the ideal lattice of the poset with one diagonal depicted in Figure~\ref{diagonal} can be identified with the one-sided ladder as shown in Figure~\ref{ladder with one corner} with $n=c+d-1$, $m=a+b-1$, $s=a$ and $t=d$. Thus, by Theorem~\ref{diagonalposet}, $S/I$ is level if and only if condition (c) of Theorem~\ref{diagonalposet} holds. According to Figure~\ref{diagonal}, this is equivalent to say that $b\leq d+1$ or $c\leq a+1$, or equivalently $m\leq s+t$ or $n\leq s+t$, which implies the assertion.
\end{proof}
\section{Generalized Hibi rings}
\label{generalized}
Let $P=\{x_1,\ldots,x_n\}$ be a finite poset and $r$ a positive integer. In Section~\ref{pseudo} we identified the Hibi ring $K[L]$ as a subring of the polynomial ring $K[s,t_x\: x\in P]$. There is a different natural embedding of $K[L]$ into a polynomial ring, namely into the polynomial ring $K[s_x,t_x\:x\in P]$ where $P$ is the set of join-irreducible elements of $L$, see \cite{HH}. For this embedding the generators of $K[L]$ are the monomials $u_\alpha=\prod_{x\in \alpha}t_x\prod_{x\in P\setminus\alpha}s_x$. This suggests an extension of the notion of Hibi rings as introduced in \cite{EHM}, see also \cite{EH}.
An {\em $r$-multichain} of $\MI(P)$ \index{multichain} is a chain of
poset ideals of length $r$,
\[
\MI: \emptyset=\alpha_0\subseteq \alpha_1\subseteq \alpha_2\subseteq \cdots \subseteq \alpha_r=P.
\]
We define a partial order on the set $\MI_r(P)$ of all $r$-multichains of $\MI(P)$ by setting $\MI\leq \MI'$ if $\alpha_k\subseteq \alpha_k^\prime$ for
$k=1,\ldots,r$.
Observe that the partially ordered set $\MI_r(P)$ is a distributive lattice, with meet and join defined as follows: for
$\MI\: \alpha_0\subseteq \alpha_1\subseteq \cdots \subseteq \alpha_r$ and $\MI^\prime\: \alpha^\prime_0\subseteq \alpha^\prime_1\subseteq \cdots \subseteq \alpha^\prime_r$ we let $\MI\wedge \MI^\prime$ be the multichain with $(\MI\wedge \MI^\prime)_k=\alpha_k\sect \alpha_k^\prime$ for $k=1,\ldots,r$, and
$\MI\vee \MI^\prime$ the multichain with $(\MI\vee \MI^\prime)_k=\alpha_k\union \alpha^\prime_k$ for $k=1,\ldots,r$.
With each $r$-multichain $\MI$ of $\MI_r(P)$ we associate a monomial $u_{\MI}$ in the polynomial ring $K[x_{ij} : 1\leq i\leq r,
1\leq j\leq n]$ in $rn$ indeterminates which is defined as
\[
u_\MI=x_{1\gamma_1}x_{2\gamma_2}\cdots x_{r\gamma_r},
\]
where $x_{k\gamma_k}=\prod_{x_\ell\in \gamma_k}x_{k\ell}$ and $\gamma_k=\alpha_k\setminus \alpha_{k-1}$\quad for $k=1,\ldots,r$.
The {\em generalized Hibi ring} $\MR_r(P)$ is the toric ring generated by the monomials $u_\MI$ with $\MI\in \MI_r(P)$. By what we said at the beginning of this section it is clear that $K[L]=\MR_2(P)$.
In \cite[Theorem 4.1]{EHM} is shown that $\MR_r(P)$ can be identified with the ordinary Hibi ring $K[L_r]$ where $L_r$ is the ideal lattice of $\MI_r(P)$, and further it is shown in \cite[Theorem 4.3]{EHM} that $\MI_r(P)$ is isomorphic to the poset $P_r=P\times Q_{r-1}$ where $Q_{r-1}=[r-1]$ with the natural order of its elements, and where $P\times Q_{r-1}$ denotes the direct product of the posets $P$ and $Q_{r-1}$. In general the direct product $P\times Q$ of two posets $P$ and $Q$ is defined to be the poset which as a set is just the cartesian product of two sets $P$ and $Q$ and with partial order given by $(x,y)\leq (x',y')$ if and only if $x\leq x'$ and $y\leq y'$.
This identification of $\MR_r(P)$ with $K[L_r]$ was used in \cite{EHM} to prove that for any given $r\geq 2$, the Hibi ring $K[L]$ is Gorenstein if and only if $K[L_r]$ is Gorenstein.
We denote by $\type(R)$ the {\em Cohen--Macaulay type} of a Cohen--Macaulay ring $R$. It is defined to be the number of generators of $\omega_R$. Here we show
\begin{Theorem}
\label{new}
Let $L$ be a finite distributive lattice, and $r\geq 2$ an integer. Then
\begin{enumerate}
\item[{\em (a)}] $\type(K[L])\leq \type(K[L_r])$;
\item[{\em (b)}] $L$ is pseudo-Gorenstein if and only if $L_r$ is pseudo-Gorenstein;
\item[{\em (c)}] If $L_r$ is level then $L$ is level.
\end{enumerate}
\end{Theorem}
\begin{proof}
(a) We need to show that $|\MT_0(\hat{P})|\leq |\MT_0(\hat{P}_r)|$. In order to prove this we define an injective map $\iota \: \MT_0(\hat{P})\to \MT_0(\hat{P}_r)$. Given $v\in \MT_0(\hat{P})$ we let $\iota(v)(x,i)=v(x)+(r-1-i)$ and $\iota(v)(\infty)=0$ and $\iota(v)(-\infty)=v(-\infty)+(r-2)$. Obviously, $\iota(v)\neq \iota(w)$ for $v, w\in \MT_0(\hat{P})$ with $v\neq w$ and $\iota(v)\in \MT(\hat{P}_r)$. Thus it remains to show that $\iota(v)$ actually belongs to $\MT_0(\hat{P_r})$. We set $v'=\iota(v)$, and show that if $v'-u\in \MS(\hat{P}_r)$ for some $u\in \MT(\hat{P}_r)$, then $v'=u$.
For any $w\in \MT(\hat{P}_r)$ and for $i\in [r-1]$ we define the function $w_i$ on $\hat{P}$ as follows:
\[
w_i(x)=w(x,i)-(r-1-i) \quad \text{for all} \quad x\in P,
\]
$w_i(\infty)=0$ and $w_i(-\infty)=\max\{w_i(x)\: x\in P\}+1$. Then $w_i\in \MT(\hat{P})$.
Since $v'-u\in \MS(\hat{P}_r)$ it follows that $v-u_i=v'_i-u_i\in \MS(\hat{P})$. Since $v\in \MT_0(\hat{P})$ we see that $v=u_i$ for all $i$. This shows that $v'=u$.
(b) Let $x\in P$ and $i\in [r-1]$. We claim that $\height_{\hat{P}_r}(x,i)=\height_{\hat{P}}(x)+(i-1)$ and $\depth_{\hat{P}_r}(x,i)=\depth_{\hat{P}}(x)+(r-i-1)$. If $\height_{\hat{P}_r} (x,i)=1$, then there is nothing to prove. Let $\height_{\hat{P}_r} (x,i)>1$ and let $x=x_0>x_1>\cdots>x_d>-\infty$ be a maximal chain of length $\height_{\hat{P}}(x)$ in $\hat{P}$. Then
$(x,i)=(x_0,i)>(x_1,i)>\cdots>(x_d,i)>(x_d,i-1)>\cdots>(x_d,1)>-\infty$ is a maximal chain of length $\height_{\hat{P}}(x)+(i-1)$ in ${\hat{P}_r}$. It follows that
$\height_{\hat{P}_r}(x,i)\geq \height_{\hat{P}}(x)+(i-1)$. To prove the other inequality we use induction on height. Let $(x,i)=z_0>z_1>\cdots >z_t>-\infty$ be a maximal chain of length $\height_{\hat{P}_r}(x,i)$ in $\hat{P}_r$. Then $z_1=(x,i-1)$ or $z_1=(x',i)$ where $x'$ is an element of $P$ covered by $x$. If $z_1=(x,i-1)$, then by induction hypothesis, we get $\height_{\hat{P}_r} (x,i-1)\leq \height_{\hat{P}} (x)+(i-2)$, and hence $\height_{\hat{P}_r} (x,i)=\height_{\hat{P}_r} (x,i-1)+1\leq \height_{\hat{P}} (x)+(i-1)$. If $z_1=(x',i)$, then our induction hypothesis implies that $\height_{\hat{P}_r} (x',i)\leq \height_{\hat{P}} (x')+(i-1)$, and hence $\height_{\hat{P}_r} (x,i)=\height_{\hat{P}_r} (x',i)+1\leq \height_{\hat{P}} (x)+(i-1)$, where the last inequality follows from the fact that
$\height_{\hat{P}} (x')\leq \height_{\hat{P}} (x)-1$. So, $\height_{\hat{P}_r}(x,i)=\height_{\hat{P}}(x)+(i-1)$. A similar argument can be applied to prove the claimed formula regarding the depth. As a side result, we obtain that $\rank \hat{P}_r=\rank \hat{P}+(r-2)$.
It follows that $\height_{\hat{P}}(x)+\depth_{\hat{P}}(x)=\rank \hat{P}$ if and only if $\height_{\hat{P}_r}(x,i)+\depth_{\hat{P}_r}(x,i)=\rank \hat{P}_r$. Thus Theorem~\ref{classification} yields the desired result.
(c) Suppose that $L$ is not level. Then there exists $v\in \MT_0(\hat{P})$ with $v(-\infty)>\rank \hat{P}$. Then $\iota(v)$, as defined in the proof of part (a), belongs to $\MT_0(\hat{P}_r)$ and
\[
\iota(v)(-\infty)=v(-\infty)+(r-2)>\rank \hat{P}+(r-2)=\rank \hat{P}_r.
\]
This shows that $L_r$ is not level.
\end{proof}
Note that if $\type(K[L])=1$, then $\type(K[L_r])=1$ for all $r\geq 2$, since by \cite[Corollary~4.5]{EHM}, $K[L]$ is Gorenstein if and only if $K[L_r]$ is Gorenstein. But the following example given in Figure~\ref{type} shows that the inequality in Theorem~\ref{new} (a) may be strict. Indeed, let $P$ be the poset depicted in the left side of Figure~\ref{type}. Then $P_3$ is the poset which is shown in right side of Figure~\ref{type}. It can be easily checked, by considering all possible strictly order reversing functions on $\hat{P}$ and $\hat{P}_3$ which correspond to the minimal generators of the canonical module, that $\type(K[L])=2$, but $\type(K[L_3])=3$.
\begin{figure}[hbt]
\begin{center}
\psset{unit=0.8cm}
\begin{pspicture}(-10.5,2)(4,5)
\rput(-1.5,2){
\psline(-6,0)(-6,1.5)
\rput(-6,0){$\bullet$}
\rput(-6,1.5){$\bullet$}
\rput(-4.5,0){$\bullet$}
\psline(1,0)(0,1)
\rput(1,0){$\bullet$}
\rput(0,1){$\bullet$}
\psline(1,0)(2,1)
\rput(1,0){$\bullet$}
\rput(2,1){$\bullet$}
\psline(0,1)(1,2)
\rput(0,1){$\bullet$}
\rput(1,2){$\bullet$}
\psline(2,1)(1,2)
\rput(2,1){$\bullet$}
\rput(1,2){$\bullet$}
\psline(3.5,0)(3.5,1.5)
\rput(3.5,0){$\bullet$}
\rput(3.5,1.5){$\bullet$}
}
\end{pspicture}
\end{center}
\caption{}\label{type}
\end{figure}
In Theorem~\ref{new}(c) we only could prove that if $L_r$ is level, then $L$ is level. However we expect
that the other implication also holds. An indication that this might be true is Miyazaki's theorem and the easy to prove fact that a poset $P$ is Miyazaki
if and only if $P_r$ is Miyazaki. Moreover, if the necessary condition for levelness given in Theorem~\ref{alsoviviana} would also be sufficient, which
indeed we expect, then one could also conclude that $L$ is level if and only if $L_r$ is level, since $\hat{P}$ satisfies the inequalities (\ref{inequality})
if and only if this is the case for $\hat{P}_r$.
|
\section{Introduction}
Solar \emph{prominences} are thread-like clouds consisting of relatively cool and dense magnetized plasma, suspended in the hot and tenuous corona. Known also as \emph{filaments} (used interchangeably with prominences hereafter), they appear as dark features along the polarity inversion line (PIL) when viewed on the disk, typically in H$\alpha$ filtergrams. Twisted/sheared magnetic field plays a crucial role in the equilibrium and dynamic evolution of prominences \citep[][and references therein]{Mackay2010a}. The remarkably stable equilibrium achieved by prominences and their sudden eruptions pose a great challenge for our understanding of the physics governing the destabilizing of the solar corona. Prominence eruptions have close association with flares and coronal mass ejections \citep[CMEs; see the reviews by][]{Low1996a, Chen2011a}. The three eruptive phenomena are hence suggested to be different manifestations of a single physical process, which involves the large-scale disruption and restructuring of the coronal magnetic field \citep{forbes00,pf02,lsb03}.
\par
It is widely accepted that the corona is energized by photospheric or sub-photospheric activities, such as shearing motions near PILs \citep[e.g.,][]{kusano02, moon02}, emerging magnetic flux \citep[e.g.,][]{Feynman1995a}, and flux cancellation \citep[e.g.,][]{livi89}. But it is still under debate how exactly coronal eruptions are triggered. A large number of mechanisms have been proposed, which include, but are not limited to, tether-cutting reconnection in a sheared arcade \citep{Moore2001a}, breakout reconnection at a magnetic null point in a quadrupolar configuration \citep{Antiochos1999a}, flux emergence \citep{Chen2000a}, catastrophic loss of equilibrium \citep{fi91} through either photospheric flux cancellation \citep[e.g.,][]{linker03} or an artificial increase in either the poloidal or the axial flux of a flux rope \citep[e.g.,][]{su11}, and ideal MHD instabilities such as the helical kink instability \citep[e.g.,][]{fan05} and the torus instability \citep{kt06}. Recently, \citet{liu12} studied a ``double-decker'' filament, which was composed of two branches separated in height. They found that prior to the eruption of the upper branch, multiple filament threads within the lower branch brightened up, rose upward, and merged into the upper branch. This transfer of magnetic flux and current to the upper branch is suggested to be the key mechanism responsible for its loss of equilibrium by reaching the limiting flux that can be stably held down by the overlying field \citep{su11} or by reaching the threshold of the torus instability \citep{kt06}.
\par
In this paper, we present the observation of a similar transfer of magnetic flux to a prominence through multiple rising `mini-prominences' originally located on the surface. In the sections that follows, we investigate the evolution and eruption of the prominence, which was embedded in the side-arcade of a quadrupolar field configuration (\sect{sec:Observations}). We argue that these mini-prominences are of the same nature as chromospheric fibrils (\S\ref{sec:nature}), and then discuss the relevant mechanisms for the prominence eruption (\S\ref{sec:instability} and \S\ref{sec:reconnection}), which was apparently coupled to the eruption of the other side-arcade in the quadrupolar field.
\section{Observations and Analysis}
\label{sec:Observations}
\begin{figure*}
\includegraphics[width=\textwidth]{euvi_original.jpg}
\caption{The prominence and the loop system from STA's perspective in 304 and 195~{\AA}}
\label{fig:orig}
\end{figure*}
\subsection{Instruments}
The prominence was observed on the west limb in EUV by the Atmospheric Imaging Assembly \citep[AIA;][]{lemen12} onboard the Solar Dynamics Observatory \citep[SDO;][]{pesnell12}, and in H$\alpha$ by the Kanzelh\"{o}he Solar Observatory (KSO). Images taken by the Extreme Ultraviolet Imager \citep[EUVI;][]{wuelser04}) of the Sun Earth Connection Coronal and Heliospheric Investigator \citep[SECCHI;][]{howard08} imaging package onboard the Solar TErrestrial RElations Observatory \citep[STEREO;][]{kaiser08}) were utilized to provide a different perspective of this prominence, which appeared as a filament in the field of view (FOV) of STEREO's `Ahead' spacecraft (hereafter STA). The CME resulting from this eruption was observed by the Large Angle and Spectrometric Coronagraph \citep[LASCO;][]{brueckner95} onboard the Solar and Heliospheric Observatory (SOHO). Magnetograms obtained by the Helioseismic and Magnetic Imager \citep[HMI;][]{schou12a,schou12b} onboard SDO provide the magnetic context of the eruption's source region.
\subsection{Eruptive Process}
\label{sec:erupt}
\begin{figure*}
\includegraphics[width=\textwidth]{erupt_total.jpg}
\caption{Eruption of the prominence observed by STA (top) and SDO (bottom). Top panels show running-difference images in EUVI 304~{\AA}, and bottom panels the corresponding original images in AIA 304~{\AA} taken at approximately the same time. The loop system is marked as `LS', and the prominence as `P'. In Panel (a), the curve denotes the solar limb as seen by SDO, and the inset plots the positions of the STEREO spacecrafts (blue dots) relative to the Sun (yellow dot) and Earth (green dot) in the plane of the Earth's orbit, with STA ahead of, and STB behind, the Earth. An animation of AIA and EUVI 304~{\AA} images is available online.}
\label{fig:total_erupt}
\end{figure*}
The prominence as observed by SDO erupted from the west limb on the early 2012 October 22. It was observed simultaneously by STA/EUVI as a filament located in the southeast quadrant of the disk, not far from the disk center (\fig{fig:orig}). \fig{fig:total_erupt} shows the eruptive process in 304~{\AA} as monitored by both satellites. From SDO's perspective, a loop system (hereafter LS) was overlying the prominence (labeled `P') in projection (\fig{fig:total_erupt}(e)); from STA's perspective LS was located to the west of the prominence (Fig.~\ref{fig:orig}). The eruption started as early as 23:49 UT on 2012 October 21 (see \S\ref{subsec:height}). The LS erupted southwestward and left the disk at about 03:16 UT (STA's perspective; \fig{fig:total_erupt}(d)), whereas the prominence erupted northwestward, and was still projected onto the disk center by the same time. This does not necessarily imply a difference in their propagation speeds, but due probably to LS's faster expansion, as can be seen from the bottom panels of \fig{fig:total_erupt}. It is noteworthy that the prominence was apparently writhed at the onset of the eruption (SDO's perspective; \fig{fig:total_erupt}(e)), taking on a projected forward S-shape on the disk (STA's perspective; \fig{fig:orig}). During the eruption, the prominence underwent a clockwise rotation of its axis, and consequently the S-shape was apparently straightened (\fig{fig:total_erupt}(b), see also the accompanying animation). This is opposite to the conversion of magnetic twist into writhe, in which case a counterclockwise rotation is expected if the flux rope assumes a forward S-shape \citep{green07,torok10}.
\begin{figure*}
\includegraphics[width=\textwidth]{euvi_195.jpg}
\caption{Eruptions and the atmospheric response as observed in STA/EUVI 195~{\AA}. All images are subtracted by the `base' image taken at 09:40:30 UT on 2012 October 21. Panel (a) shows an pre-eruption image, with the source regions of the LS and the prominence (P) marked by arrows. The eruptive process is featured in Panels (b)-(f) with transient brightenings (labeled B1-B3) and dimmings marked by arrows. Brightening loops evolving from B1 and B2 are labeled T1 and T2, respectively, in Panel (e).} An animation of both original and running-difference 195~{\AA} images is available online.
\label{fig:erupt_195}
\end{figure*}
The sequence of the eruption is best demonstrated by EUVI 195~{\AA} base-difference images in Fig.~\ref{fig:erupt_195}, though the LS was not quite visible on the disk in 195~{\AA}. A sequence of reconnection events is characterized by successive brightenings at the surface. The 1st episode of brightening (B1) occurred in the central region between the prominence and the LS at about 01:40 UT (Fig.~\ref{fig:erupt_195}(c)), during the initial phase of the prominence eruption. Then B1 separated into two ribbon-like structures moving away from each other (Fig.~\ref{fig:erupt_195}(d)). The 2nd episode of brightening (B2) appeared to be related to LS's eruption (Fig.~\ref{fig:erupt_195}(d)), and the 3rd episode (B3) took on the form of a two-ribbon flare (Fig.~\ref{fig:erupt_195}(f)), associated with the prominence eruption. Both B1 and B2 had an irregular, moss-like appearance initially and later became the footprints of some transient brightening loops labeled T1 and T2 in Fig.~\ref{fig:erupt_195}(e), see also the accompanying animation). Prior to the brightenings B1 and B2, two pairs of dimming regions were observed to be located at both sides of the prominence and the LS, respectively (marked by arrows in \fig{fig:erupt_195}(c) and (d)). Coronal dimmings are often interpreted as a mass deficit due to eruptions \citep[e.g.,][]{sh97,harrison03}. Dimmings in pair have only been occasionally observed and was suggested to represent the feet of an eruptive flux rope \citep[e.g.,][]{thompson98,thompson00,webb00,liuc07}.
\begin{figure}
\includegraphics[width=0.45\textwidth]{cme.jpg}
\caption{Height-time evolution of the CME. The inset shows a LASCO/C3 white-light observation of the CME with two fronts (see the text for details).}
\label{fig:cme}
\end{figure}
The eruption results in two CME fronts as observed by LASCO/C3 (see the inset of \fig{fig:cme}). We may identify the source of the leading front with the LS and that of the trailing front with the prominence. Linear fitting of their height-time profiles yields that the leading front propagated at (328$\pm$2) km s$^{-1}$ in the plane of sky, slightly slower than the trailing front, which propagated at (412$\pm$4) km s$^{-1}$. Hence, the trailing front might eventually catch up with the leading front and interact with it \citep[e.g.,][]{shen12}.
\subsection{Pre-Eruption Dynamics}
\label{sec:pre-eruption}
To explore the physical mechanism of the eruption as described in \S\ref{sec:erupt}, we investigate the pre-eruption processes that might help make the prominence `ready' to erupt. What stands out is that within 2 days prior to its eruption the prominence was `fed' for at least three times by mini-prominences originally resting on the surface. This process is referred to as `flux feeding' hereafter. In this paper, we emphasize on the role of magnetic flux as far as `feeding' is concerned, although this process involves both magnetic and mass flux (see \S\ref{sec:instability}).
\subsubsection{Flux Feeding}
\label{sec:perturbations}
The mini-prominences appear similar in emissivity as, but much smaller in spatial scales ($\sim 1/4$ in length and width) than, the target prominence in EUV images. In \fig{fig:total_slice} we study these feeding processes by placing a virtual slit along the rising direction of the mini-prominences, and present the resultant stack plots in a logarithm scale. During each feeding process, a miniature prominence rose upward apparently from the solar surface at a speed of tens of kilometers per second, interacted with, and eventually merged into, the target prominence. The interaction is characterized by an enhancement in brightness, and a decrease in speed, as the mini-prominences approached the target prominence. Each feeding process lasted for about half an hour.
\begin{figure*}
\includegraphics[width=\textwidth]{slice_total.jpg}
\caption{Flux-feeding episodes. On each row, an individual episode is presented by the space-time stack plot in the right panel, which is made through the virtual slit marked by the dashed line in the left panel.}
\label{fig:total_slice}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{disturb1.jpg}
\caption{First episode of flux feeding. Panels (a) and (b) show a 304~{\AA} and an 171~{\AA} image just before the feeding process, respectively, with two virtual slits A and B denoted by white dashed lines; Snapshots of the feeding process as observed in 304~{\AA} are shown in Panels (c), (e) and (g), and the corresponding base-difference images in Panels (d), (f) and (h); Panels (i) and (j) are the space-time stack plots obtained from the virtual slits in (a) and (b), respectively. Slit A is the same as that in \fig{fig:total_slice}(a). Fitting results on various features on the stack plots are also given (see the text for details). An animation of AIA 304 and 195~{\AA} images is available online.}
\label{fig:perturbation1}
\end{figure*}
The first feeding process, which took place at about 09:00 UT on 2012 October 20, is particularly interesting (see the animation accompanying Fig.~\ref{fig:perturbation1}). The upward-moving mini-prominence apparently drove the oscillation of two threads within the target prominence (\fig{fig:total_slice}(d)). Detailed analysis is shown in Figure \ref{fig:perturbation1}. The mini-prominence became visible as early as 08:20 UT, as a fibril-like structure. It started to rise at approximately 08:55 UT, with an acceleration of \accel{22$\pm$1} (\fig{fig:perturbation1}(i)). At about 09:05 UT, the upward moving turned into a deceleration of \accel{-5.9$\pm$0.6}, indicating an interaction with the prominence overhead. This compression process may result in the brightening of the mini-prominence from 09:05 UT onward, as well as the oscillation of the prominence threads at higher altitudes shortly after that time (see also \fig{fig:total_slice}(d)). The oscillation can be well fitted with a damped cosine function:
\begin{equation*}
h(t)=h_{0}+H\cos (\dfrac{2\pi}{T}t+\phi)\mathrm{e}^{-t/\tau},
\end{equation*}
where $H$, $T$, and $ \tau $ corresponds to the amplitude, period, and e-folding damping time, respectively. The two threads oscillated with essentially the same period, but the oscillation of the upper thread started slightly later, had a smaller amplitude, and decayed slower than the lower thread (Fig.~\ref{fig:perturbation1}(i)), suggesting that the oscillations were due to an upward-propagating wave which was excited by the interaction of the mini-prominence with the target prominence. The velocity amplitude was about 19 and 14 km~s$^{-1}$ for the lower and upper thread, respectively, significantly larger than that of small-amplitude oscillations (from 0.1 to several kilometers per second) that are apparently ever-present in prominences \citep{arregui12}. The present observation therefore provides an alternative cause for large-amplitude prominence oscillations (velocity amplitude $\gtrsim 20$ km~s$^{-1}$), which are relatively rare and have been suggested to be triggered by waves and disturbances produced by flares or jets \citep{tij09}.
The eventual merge of the mini-prominence with the target prominence is characterized by knots of filament material moving along the prominence axis bi-directionally at tens of kilometers per second, reminiscent of counter-streaming flows in prominences \citep[e.g.,][]{zem98,lew03,lin05,ahn10,alexander13}. The space-time plot (\fig{fig:perturbation1}(j)) obtained from the virtual slit parallel to the prominence axis (Slit B in Figure \ref{fig:perturbation1}(b)) clearly shows that such horizontal motions became appreciable in 171~{\AA} images only after the merge at about 09:20 UT (also see the animation accompanying Fig.~\ref{fig:perturbation1}). However, counter-streaming flows are believed to be ubiquitous in prominences, despite that its cause remains unclear \citep[see][for a discussion]{chen14}. The fact that the horizontal motion excited by the disturbance from the rising mini-prominence well resembles counter-streaming flows suggests that such flows are dictated by the magnetic nature of prominences. It is known that the prominence field is dominantly horizontal and directed along the prominence axis \citep[e.g.,][]{leroy89}, which may explain the observed horizontal motions as well as the absence of vertical motions within the perturbed prominence.
The hight-time profile of the mini-prominence is fitted with a piecewise parabolic function, with a uniform acceleration followed by deceleration. The fitting results are given in \fig{fig:perturbation1}(i). The second and third episodes of flux feeding are shown in the middle and bottom panels of \fig{fig:total_slice}, respectively. Unlike the first episode, there were no discernible oscillations resulting from the interaction, and the acceleration phase was less appreciable, so that both hight-time profiles can be well fitted with a uniform deceleration function with $v_0\approx 30$ km~s$^{-1}$ and $a\approx -8$ m s$^{-2}$.
\subsubsection{Height-Time Evolution}
\label{subsec:height}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{height.png}
\caption{Height-time evolution of the prominence. The leading edge of the prominence is measured with both SDO/AIA and STEREO-B/EUVI images until it left the AIA FOV. The profile after the third episode of flux feeding is fitted by an exponential function (shown as the solid curve). The inset shows the height-time profile starting on 2012 October 17 till October 22; three dotted lines mark the occurrences of flux feeding; the dashed line denotes the linear fitting result for the height-time evolution prior to any flux-feeding episodes.}
\label{fig:height}
\end{figure*}
That the flux-feeding processes affected the evolution of the prominence is evidenced by its height-time profile. With the \texttt{SCC\_MEASURE} procedure in SolarSoftWare, the `true' height, rather than projected height, of the prominence can be obtained. The inset of \fig{fig:height} shows the height-time profile of the prominence starting from 2012 October 17 till the prominence eruption on October 22. The vertical dashed lines mark the occurrences of the three episodes of flux feeding. The average speed till the occurrence of the 1st episode is $v_{0-1}\approx (0.071\pm0.002)$ km~s$^{-1}$. Similarly, $v_{1-2}\approx (0.150\pm0.070)$ km~s$^{-1}$ denotes the average speed between the 1st and 2nd episode, and $v_{2-3}\approx (0.227\pm0.028)$ km~s$^{-1}$ the average speed between the 2nd and 3rd episode. One can see that the average rising speed of the prominence was significantly enhanced with the flux-feeding processes, as compared with the long time interval before the occurrence of the 1st episode. This might be due to the magnetic-flux increase of the prominence, and therefore the strengthening of the outward magnetic pressure of the prominence field over the inward magnetic tension of the external field.
The eruptive process on 2012 October 22 can be well fitted by an exponential function with an initial height $h_0$ and velocity $v_0$:
\begin{equation*}
h(t)=h_0+v_0t+c\mathrm{e}^{t/\tau}.
\end{equation*}
The fitting yields that $v_0=(1.00\pm0.03)$ km~s$^{-1}$ and $\tau=(1636\pm14)$ s. Let the linear term equal to the exponential term, we are able to determine that the prominence eruption started at 0.154~$R_\odot$ at 23:49 UT on 2012 October 21.
\section{Discussion and Conclusion}
\label{sec:Discussion}
\subsection{Nature of Mini-Prominences} \label{sec:nature}
The flux-feeding process is reminiscent of the flux transfer within the double-decker filament as reported by \citet{liu12}. In the present study, the merge of the mini-prominences into the target prominence also feeds magnetic flux and current to the latter, resulting in an increasing speed of its quasi-static ascent, and eventually leading up to its unstableness. So, are the mini-prominences part of the lower branch of a double-decker filament? The main body of the lower branch could be lying beneath the photosphere so that only the upper branch was observed as the target prominence. If the lower branch emerges, then a double-deck configuration ensues. The Hinode observation that a flux rope emerges under a pre-existing filament \citep{okamoto08} might be such a case.
\begin{figure*}
\includegraphics[width=\textwidth]{fibril.png}
\caption{Mini-prominence in AIA 304~{\AA} and H$\alpha$ as observed on 2012 October 20. The white arrow in Panel (a) marks the mini-prominence in 304~{\AA}, which is hardly visible in the H$\alpha$ image obtained at the same time (Panel (b)). Black arrows in Panels (a) and (b) mark a fibril structure. Panel (c) shows an H$\alpha$ image with the disk overexposed, which highlights the rising mini-prominence, as marked by a red arrow, on its way to merge into the target prominence.}
\label{fig:fibril}
\end{figure*}
Back to the present study, the mini-prominences were observed in AIA 304~{\AA} as thin elongated structures lying on the surface before rising upward to interact with the target prominence. In morphology, they are very similar to chromospheric fibrils, which cover most of the disk in H$\alpha$ line core. However, taking the first mini-prominence as an example (marked by a white arrow in Fig.~\ref{fig:fibril}(a)), one cannot easily find its counterpart in H$\alpha$ (Fig.~\ref{fig:fibril}(b)). In contrast, a slightly thicker fibril (marked by black arrows) at about $(900'',-230'')$ to the northeast of the mini-prominence can be seen in both 304~{\AA} and H$\alpha$. Hence, this could be due to the relatively poor resolution of the H$\alpha$ images, whose contrast are further plagued by the seeing conditions. As it rose above the limb, the mini-prominence can also be observed in H$\alpha$ (Fig.~\ref{fig:fibril}(c)), suggesting it is indeed of the same nature as fibrils. This is not surprising as filaments and fibrils are closely related in the sense that a prerequisite for a filament is a channel of chromospheric fibrils aligned with the polarity boundary, known as ``filament channel'' \citep{gaizauskas98,martin98}.
Arguably the tracer of the chromospheric field \citep{cs11, jing11}, fibrils can be regarded as small flux tubes. Apparently attracted to the target prominence, these rising fibrils must carry currents in the same direction, or, helicity of the same sign, as the prominence, in light of the MHD simulations of interactions between parallel flux tubes done by \citet{linton01}. However, there still exists a possibility that these fibrils actually belong to a flux rope lying beneath the surface, serving as the lower branch of a double-deck configuration.
It is worth commenting that the rising fibrils are distinct from buoyant plumes detected in off-limb observations \citep{berger10}, which are dark, bubble-like features in visible-light spectral bands, rising and inflating through the bright prominence emission with approximately constant speeds. In contrast to the plumes, the fibrils appear in emission above the limb and in absorption on the disk, same as the prominence in terms of emissivity; they rise from the surface and merge into the prominence with obvious deceleration, but no significant inflation.
\subsection{Role of Instability} \label{sec:instability}
The low-lying forward S-shaped prominence rotated clockwise at the onset of the eruption, and it is not clear whether it was still kinked during the eruption. That the S-shape was apparently straightened due to the clockwise rotation implies a reduction of writhe, therefore excluding the helical kink instability as the trigger of the eruption \citep{torok10,liu12}. However, the exponential rise of the prominence and the lack of intense heating during the initial phase of the eruption suggests that a certain instability or loss of equilibrium may play an important role. Here we discuss three related mechanisms: a) flux imbalance, b) mass loading, and c) torus instability.
Numerical studies have suggested that a flux rope could become unstable due to an increase of the axial flux, whose amount may possess a threshold for the existence of stable equilibria \citep{bobra08, sv09, su11}. The threshold appears to be only 10\%--20\% of the total flux in the region. For the quiescent prominence in question, a rather modest amount of flux transfer to it through the rising fibrils may be significant enough to reach the critical point.
On the other hand, the rising fibrils also input mass into the target prominence. Mass loading could help hold down current-carrying flux, therefore raising the amount of free magnetic energy that can be stored in the pre-eruption configuration \citep{low03}. Thus, mass loading may also play a role in the present case, except that we do not see a significant increase in the darkness or thickness of the target prominence during the pre-eruption evolution, as reported in some cases \citep[e.g.,][]{kilper09, guo10, liu12}. However, both darkness and thickness could be modulated by the solar rotation: the apparent darkness of the prominence in EUV is expected to decrease as it rotated off the west limb, due to a `deeper' line-of-sight integration of EUV emission in the foreground; the thickness is affected by projection effect as the shape of the prominence is by no means symmetric along the line of sight. It is therefore difficult to determine quantitatively how much mass has been loaded as time progresses.
The torus instability \citep{kt06, Torok2007a} sets in if a flux rope rises to a critical height \citep{liuk12} at which the overlying field declines with height at a sufficiently steep rate \citep{liuy08,aulanier10,oz10,fan10}, i.e., the decay index $n\equiv-d\log(B_h)/d\log(h)$ exceeds a critical value of 1.5, where $B_h$ is the horizontal component of the potential field external to the flux rope. However, the equilibrium of the system becomes unstable already when $n$ approaches $n_\text{crit}$. For example, \citet{da10} found that $n_\text{crit}$ typically falls in the range [1.1--1.3] for both circular and straight current channels. Here, we calculate the average decay index along the filament at different altitudes using the potential-field source-surface (PFSS) approximation \citep[][see the right panel of Fig.~\ref{fig:hmi}]{sd03}. The critical height $h_\text{crit}$ at $n_\text{crit}=1.3-1.5$ is about 0.15--0.27~$R_\odot$, where the filament became nominally torus unstable. The prominence indeed takes off at about 0.15~$R_\odot$, according to the exponential fit (\S\ref{subsec:height}). Hence, we conclude that the torus instability is the major mechanism in triggering the prominence eruption. In light of flux imbalance, the role of flux feeding is to force the prominence to seek for equilibrium at higher and higher altitudes, as evidenced by the enhanced slow-rise speed after each flux feeding episode. Consequently, the prominence reached the unstable height much earlier: at the average slow-rising speed of 0.07 km~s$^{-1}$ prior to the flux-feeding episodes (\S\ref{subsec:height}), the filament would have reached the critical height of 0.15~$R_\odot$ by 23:56 UT on October 29. In other words, this quiescent prominence might have been quite stable without flux feeding.
\subsection{Role of Reconnection} \label{sec:reconnection}
One can see that the prominence was embedded in a quadrupolar configuration (Fig.~\ref{fig:hmi}; the four polarities are labeled P1--N1 and P2--N2) by superimposing the line-of-sight component of the photospheric magnetic field upon the H$\alpha$ image taken on 2012 October 14 when the prominence was crossing the central meridian. A small bipolar active region located to the west of the filament was composed of a leading sunspot of positive polarity (P1) followed by diffused flux of negative polarity (N1). With the filament and the sunspot serving as landmarks, one can see that the LS as identified in EUV observations (Figs.~\ref{fig:orig} and \ref{fig:total_erupt}) must be connecting N1 and P1. During the eruptive process (\S\ref{sec:erupt}), successive surface brightenings in EUV (Fig.~\ref{fig:erupt_195}) first took place between the filament and the LS (B1), then in the active region on the west, associated with the eruption of the LS (B2), and finally on the east, associated with the prominence eruption (B3). All three brightening episodes, especially B1 and B3, were similar to two-ribbon flares in terms of both morphology and dynamics. B2 appeared to have only one ribbon (\fig{fig:erupt_195}(d)), but similar to B1 and B3 it had the moss-like appearance initially and later became the footprint of transient brighteing loops T2 below the erupting LS (\fig{fig:erupt_195}(e)).
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{ha_hmi2.png}
\caption{Magnetic environment of the prominence in question as it crossed the central meridian on 2012 October 14. \emph{Left panel}: KSO H$\alpha$ image overlaid by the contours of the line-of-sight component of the photospheric magnetic field as obtained by SDO/HMI; contours levels are $\pm$50, $\pm$200, and $\pm$800 G, with red (blue) colors denoting positive (negative) polarities. Coronal field lines traced using the PFSS model are superimposed to demonstrate the large-scale magnetic connectivities.Four polarities of the quadrupolar configuration are labeled P1--N1 and P2--N2. The transient brightening loops T1 and T2 (Fig.~\ref{fig:erupt_195}(e)) that evolve from B1 and B2 are sketched with thick yellow curves. \emph{Right panel}: variation of the decay index $n$ with height, which is calculated using the PFSS model and averaged over the hand-picked points (green `+' symbols in the left panel) along the filament. The error bars reflect the standard deviation.}
\label{fig:hmi}
\end{figure*}
Based on these observations, we interpret the prominence eruption on 2012 October 22 in the framework of a schematic quadrupolar configuration (\fig{fig:cartoon}). Both the prominence and the LS are represented by a flux rope embedded in the two side-arcades. Within the two-day period prior to the eruption, multiple chromospheric fibrils rise upward and merge into the target prominence. The fibrils are apparently parallel to the target prominence and their interaction with the prominence results in horizontal flows along the prominence axis. We therefore speculate that primarily axial flux is ejected into the field of the prominence. With the accumulation of the axial flux, the prominence has to seek for equilibrium at higher heights. At certain point, it starts to interact with the flux rope embedded in the west side-arcade. The reconnection between the two flux ropes is evidenced by the first episode of two-ribbon brightening (B1) underneath the central arcade. The reconnection also cuts the `tethers' that hold down both flux ropes, leading to their rapid rise. Both rising flux ropes stretch their overlying fields and result in further reconnections underneath, which is evidenced by brightenings B2 and B3. It is remarkable that magnetic reconnection, as demonstrated by the surface brightenings, set in almost two hours after the eruption onset (\S\ref{subsec:height}). Thus, ideal instability must dominate the initial phase of the eruption, though it was later coupled to reconnection to drive the eruption.
We further conjecture that being held down by dense material causes the prominence eruption to progress initially on a slower pace than the LS in the west side-arcade, as evidenced by the fact that B2 precedes B3; but later on, the draining of the prominence material back to the surface (see the animation accompanying Fig.~\ref{fig:total_erupt}) may help the CME front resulting from the prominence eruption to catch up with the front caused by the erupting LS (see Fig.~\ref{fig:cme}). However, one must be aware of the limitation of this simplified scenario: despite deviating significantly from the potential field (Fig.~\ref{fig:hmi}), the observed brightening loops (T1) evolving from B1 do not connect B1's two ribbons, which cannot be explained by this 2D cartoon, but might be a reflection of the complex 3D nature of the reconnection between two flux ropes.
\begin{figure}
\includegraphics[width=0.45\textwidth]{cartoon.png}
\caption{Schematic of the quadrupolar magnetic field configuration in which the prominence is embedded. A gray slab indicates the body of the prominence. `X' symbols mark the locations of magnetic reconnections, which result in paired brightening ribbons on the surface. The observed brightenings, B1, B2, and B3, are marked by dashed arrows (see the text for details). }
\label{fig:cartoon}
\end{figure}
To conclude, we have described a new mechanism for a prominence to become torus unstable, i.e., chromospheric fibrils that carry the helicity of the same sign as the prominence could feed flux and current into the prominence, which results in the faster quasi-static ascent of the prominence, eventually leading up to its unstableness. We have also described a new paradigm of quadrupolar eruptions, i.e., two flux ropes embedded in the two side-arcades first interact to cut the constraining `tethers', and consequently erupt in close succession and proximity, effectively manifesting as a `twin' eruption \citep[e.g.][]{shen13}.
\acknowledgments The authors are grateful to the \sat{sdo}, \sat{stereo} and \sat{soho} consortium for the free access to the data. KSO H$\alpha$ data are provided through the Global H-alpha Network operated by New Jersey Institute of Technology. RL acknowledges the Thousand Young Talents Programme of China, NSFC 41222031 and NSF AGS-1153226. This work was also supported by NSFC 41131065 and 41121003, 973 key project 2011CB811403, CAS Key Research Program KZZD-EW-01-4, the fundamental research funds for the central universities WK2080000031.
|
\section{Introduction}
After the recent discovery of tensor modes at BICEP2 experiment \cite{Ade:2014xna}\footnote{Throughout this work, we will assume that although the exact numbers of BICEP2 may change,
sizeable tensor modes, {\it i.e.} $ r \gtrsim .1 $ are an actual feature that will stay}, the theory of cosmological inflation \cite{inflation} can claim to be the current (undisputed) paradigm of early universe cosmology. Inflation cannot only solve most of the problems of the Standard Big Bang Model, but it offers the only available explanation for the origin of the large-scale structure of the universe based on causal physics. Even more, cosmological inflation is a predictive theory. It calls for an almost scale invariant spectrum of curvature perturbations which anticipates the characteristic oscillations in the angular power spectrum of cosmic microwave anisotropy maps, observed with high accuracy by WMAP \cite{Bennett:2012zja} and Plank \cite{Ade:2013uln}.
Inflation is simply the assumption that there was a short epoch in the very early universe where the scale factor (space) grew at an accelerated pace, typically in an exponential way. Such an accelerated expansion flattens out and widens up a microscopic size of space, solving the longstanding ``size problem'' of standard cosmology. Not only that, the accelerated expansion decreases the contribution of any pre-existing curvature to the total energy budget of the universe and therefore turns the spatially flat universe into a local attractor in initial-condition space, solving this way yet another cosmological puzzle, the ``flatness problem''.
Unfortunately, inflation comes at a cost. Successful models of inflation, {\it i.e.} successful inflationary potentials require unusual features: the potentials have to be extremely flat so that enough inflation is produced to actually solve the above-mentioned issues, and observations seem to require the inflaton field to travel over transplanckian distances in field space. In fact, following an argument due to Lyth \cite{Lyth:1996im} \footnote{This bound has been generalized in the context of effective field theories of inflation in Ref.~\cite{Baumann:2011ws}. However, for the sake of this work the original bound still holds.} we have,
\beq
\frac{\Delta \theta}{M_{\rm Pl}} \gtrsim 5.8 \left(\frac{N_e}{50}\right) \left(\frac{r}{0.2}\right)^{1/2}
\eeq
with $\Delta \theta$ the variation of the field during inflation, $r \simeq 13.8 ~\epsilon$ the tensor-to-scalar ratio with $\epsilon$ the usual slow-roll parameter, $N_e$ the number of e-folds of inflation since the relevant scales left the horizon till the end of inflation and $M_{\rm Pl} = (16 \pi G)^{-1/2}$. Therefore, the value of $r = 0.2^{+0.07}_{-0.05}$ measured at BICEP2 implies transplanckian values for the inflaton field,
$\Delta \theta/M_{\rm Pl} \gtrsim 5.8$. Fortunately, large (transplanckian) field values do not necessarily involve large (transplanckian) energies, which is the reason why transplanckian field values are not total anathema. In fact, transplanckian field values have been the norm rather than the exception in the inflationary game \cite{Chialva:2014rla,Antusch:2014cpa,Kehagias:2014wza,Mazumdar:2010sa}. There is (almost) no single-field inflationary model which can be kept below Planck scale all the way.
Yet another problem which has not been devoted enough attention to is the fact that the energy scale of inflation and the Planck scale are not that far from each other and therefore it is easy to imagine that corrections to Planck scale physics are bounded to play a role.
Whether this role is significant or not is clearly a debatable issue.
Going back to the transplanckian field values, one of the reasons why it is safe to entertain transplanckian field values (once checked that the observables are well behaved) is that a field is, after all, a ``dummy'' variable, {\it i.e.} it is ``per-se'' meaningless. Just a field redefinition will turn its value into the desired domain at no expense, all the observables will remain invariant. Nevertheless, field redefinitions may be gratis observable-wise, but they are not innocent. They will surface somewhere else: in a change of the kinetic terms, the couplings in the potential, etc. In the same way that the mass matrix in the quark sector can be made real, but then the removed (physical) phase will show up in the charged and neutral current interactions, a field redefinition to turn the inflationary field subplanckian may end up shedding light on the shape of gravity close to the Planck scale.
In this work, we conjecture about the possibility that the trasplanckian field values arising in single-field inflationary models may be due to the fact that we are ``forcing'' our model to have Einstein gravity. We will show that well-behaved and subplanckian modified gravity, as non-minimally coupled scalar fields and/or scalar tensor theories, can become transplanckian once forced to behave as minimally coupled scalar field theories in Einstein gravity. Therefore, the observed tensor-to-scalar ratio can be obtained in
single-field inflation models, in the presence of non-minimal couplings
to gravity, working always in the subplanckian regime and in the slow-roll approximation.
This work is organised as follows. We begin with a basic review of
models with non-minimal coupling to gravity and recall the use of
conformal transformations to go from the Jordan frame (with
non-minimal coupling to gravity) to the Einstein frame in section
\ref{sec:jordan}. In section \ref{sec:models}, we present several
realistic examples showing the effect of conformal transformations in
the field values, making subplanckian field values in the Jordan
frame transplanckian in the Einstein frame. Finally, results and
conclusions are summarised in section \ref{sec:conclusions}.
\section{Inflation in theories with non-minimal couplings to gravity.}
\label{sec:jordan}
We start from a general theory with gravity coupled to a single scalar field that will play the role of the inflaton. The action in the Jordan frame, with non-minimal coupling to gravity and assuming canonical kinetic terms\footnote{The introduction of a non-canonical kinetic term will complicate
unnecessarily the theory and is not needed to make our point clear}, would be,
\begin{eqnarray}
S = - \int d^ 4 x ~\sqrt{-g} \left[\frac{k^ 2}{4} D(\theta)~ R - \frac{1}{2} ~g^ {\mu \nu}~ \partial_\mu \theta \partial_\nu \theta + V(\theta) \right]
\end{eqnarray}
where $R$ is the scalar curvature and $\theta $ our scalar field\footnote{This action becomes non-renormalizable, once the field is above the cutt-off scale in the Einstein frame, a fact that may {\it per-se} be signaling the need to introduce a non-minimal coupling to gravity as the true driver of inflation as higher correction are always kept under control in this frame} .
In the absence of any other sources of matter, and specialising for the case of a Friedmann-Robertson-Walker metric, $g_{\mu\nu} = \mbox{diag} \{ 1, -a(t)^2, -a(t)^2,-a(t)^2 \}$
the equations for the Hubble rate and the $\theta$ field become,
\begin{eqnarray}
\label{einstein}
&D(\theta)~ H^2 = \frac{\dot{\theta}^2}{3 k^2 } + \frac{2 V (\theta)}{ 3 k^2} - \dot{D}(\theta)~ H& \\
&\ddot{\theta} + 3 H \dot{\theta} + \frac{k^2}{4}~D'(\theta)~R +V'({\theta})=0 &\, ,\nonumber
\end{eqnarray}
with $k^2=M_{\rm Pl}^2/(4 \pi)$, from where it is straightforward to obtain,
\begin{eqnarray}
\label{dotH}
\dot{H} = -\frac{\dot{\theta}^2}{k^2 D} + \frac{\dot{D}(\theta)}{D(\theta)}~ \frac{H}{2} - \frac{\ddot{D}(\theta)}{2 D(\theta)}\, .
\end{eqnarray}
Due to the addition of the extra source for perturbations we have introduced, $D(\theta)$, we need to include two more slow-roll parameters as compared to the standard case \footnote{Throughout this work we assume the slow-roll regime. In principle the field could fast roll and in this case, quantum corrections can become sizeable and the slow-roll solution might stop being an attractor}. The scalar-type perturbations will be affected by both of them, although only one ($ \epsilon_3 $) will be relevant for the tensor perturbations \cite{Hwang:1996xh,Noh:2001ia},
\begin{eqnarray}
\label{slow-roll}
\epsilon_1 &=& \frac{\dot{H}}{H^2} =\frac{H' \dot{\theta}}{H^2} \\
\epsilon_2 &=& \frac{\ddot{\theta}}{H \dot{\theta}} \\
\epsilon_3 &=& \frac{1}{2} \frac{\dot{D}}{H D} = \frac{1}{2} \frac{D' \dot{\theta}}{H D} \\
\epsilon_4 &=& \frac{1}{2} \frac{\dot{E}}{H E} = \frac{1}{2} \frac{E' \dot{\theta}}{H E} \, ,
\end{eqnarray}
with $E= 3k^2 (D')^2/2 + D$. Assuming $\dot{\epsilon}_i =0$
and to linear order in the slow-roll parameters
\begin{eqnarray}
n_s &=& 1+ 2\left( 2\epsilon_1 - \epsilon_2 + \epsilon_3 - \epsilon_4 \right) \\
n_T &=& 2\left( \epsilon_1 - \epsilon_3 \right) \\
r &=& 13.8 \mid \epsilon_1 - \epsilon_3 \mid \label{rjordan}
\end{eqnarray}
As it is well-known, this model, as any non-standard theory of gravity, can be mapped into a standard theory of gravity at the expense of having a more complicated matter sector by a conformal transformation. Such a transformation is not just a coordinate redefinition (being general relativity a covariant theory, a coordinate redefinition would become trivial) rather, it is a transformation that mixes up the matter and gravitational degrees of freedom\footnote{Every single choice of field (and metric) definition among the family of transformations goes under the name of {\bf frame} and obviously the frame where gravity takes the form of Einstein's theory is called Einstein's frame.}.
The mapping we are alluding to, takes the original metric $g_{\mu \nu}$ into a new metric $ \tilde{g}_{\mu \nu}$ according to,
\begin{equation}
\tilde{g}_{\mu \nu} = e^{2 \omega} g_{\mu \nu}\, ,
\end{equation}
with $ e^{2 \omega} = D(\theta)$.
The Hubble rate transforms as,
\beq
\label{Htrans}
\tilde H = \frac{ H + \dot D(\theta)/ (2 D(\theta))}{\sqrt{D(\theta)}}\,,
\eeq
with $\dot D = \partial D /d t$, and the canonically normalised field replacing $\theta$ in the Einstein frame is,
\begin{equation}
\label{fieldtrans}
\phi(\theta) = \pm \int \sqrt{\frac{3}{2} \left(\frac{D'(\theta)}{D(\theta)}\right)^ 2 + \frac{2}{k^{2} D(\theta)}}~d \theta\,.
\end{equation}
In terms of this rescaled field the action takes the form,
\begin{eqnarray}
S = - \int d^ 4 x ~\sqrt{-\tilde g} \left[\frac{k^ 2}{4} \tilde R - \frac{1}{2} ~\tilde g^ {\mu \nu}~ \partial_\mu \phi \partial_\nu \phi + \tilde V(\phi) \right] \, ,
\end{eqnarray}
with $\tilde V(\phi) = V(\phi(\theta))/D(\theta)^2$.
It is trivial to show that the slow-roll parameters in both frames are related as,
\begin{eqnarray}
\label{Einst-roll}
\tilde \epsilon &=& \frac{k^2}{4}\left(\frac{\tilde H'}{\tilde H}\right)^2 = \epsilon_1 -\epsilon_3 \\
\tilde \eta &=& \frac{k^2}{4}\frac{\tilde H''}{\tilde H}= \epsilon_2 - 3 \epsilon_3 + \epsilon_4 \\
\tilde{n}_s &=&1+ 2\left( 2\tilde{\epsilon} - \tilde{\eta} \right) = 1+ 2\left( 2\epsilon_1 - \epsilon_2 + \epsilon_3 - \epsilon_4 \right) = n_s \\
\tilde{n}_T &=& 2\tilde{\epsilon} = 2\left( \epsilon_1 - \epsilon_3 \right)= n_T \\
\tilde{r} &=& 13.8 \mid \tilde{\epsilon} \mid = 13.8 \mid \epsilon_1 - \epsilon_3 \mid =r \label{rEinstein}
\end{eqnarray}
Leaving all the observables invariant, as it is obvious given the fact that changing from one frame to another one does not correspond to a change in the physics. However as the conformal transformation changes the space-time curvature (and also the scalar/matter field) phenomena that appear to be due to gravity in one frame may appear to be originated in the scalar sector in another.
Besides, it is easy to see that as a result of the fact that the inflaton field in Einstein and Jordan frames are related in a highly non-trivial way, it can be expected that subplanckian values in a given frame, may correspond to transplanckian values in the second frame.
The analysis in this paper is done in the framework of and effective
field theory neglecting terms suppressed by the cutoff scale, which in
the Einstein frame is $M_{Pl}$. However, as pointed out recently by
Hertzberg in Ref.~\cite{Hertzberg:2010dc}, in the presence of non-minimal couplings to
gravity the validy regime of the effective theory may change. As shown in
this work, in single field models, the relevant cutoff scale, in the
gravitational and kinetic sector, is still the Planck mass. Regarding
the potential interactions the situation is model dependent and we
will check it in a case by case basis.
\section{Single-field non-minimal models of inflation}
\label{sec:models}
As shown in the previous section, the field values in two different
frames are correlated by an non-trivial function in a rather complicated
way. Here, we will show that is possible and, in fact, quite natural and easy to
find realistic examples of theories with non-minimal coupling to
gravity, modified gravity or scalar tensor theories, that have
transplanckian field values if we insist on imposing a minimal
coupling to gravity, but are always subplanckian in their ``natural''
frame. This clearly does not imply that any conformal
transformation will turn transplanckian minimally coupled scalar
fields into the subplanckian regime once non-minimally coupled to
gravity or once allowed to live in a modified-gravity framework but, in our
scheme, observations would select a subclass of conformal transformations.
\subsection{Monomial Potentials}
\label{monomial}
As a first toy-model, we assume that the potential in the Einstein frame is exactly given by
the well-known potential $V(\phi) = \lambda \phi^4$ in the Einstein frame.
In the slow-roll regime $\dot \phi^2 \ll V(\phi)$, using the Eqs. of motion, we have,
\begin{eqnarray}
H& =& \sqrt{\frac{2 \lambda}{3}} \frac{\phi^2}{k} \\
\dot \phi &=& - 2 \sqrt{\frac{2 \lambda k^2}{3}} \phi
\end{eqnarray}
and therefore, we have,
\begin{eqnarray}
\epsilon &=& \frac{\dot H}{H^2} = \frac{H' \dot \phi}{H^2} = - \frac{4 k^2 }{\phi^2} \\
\eta &=& \frac{\ddot \phi}{H \dot \phi} = - \frac{2 k^2 }{\phi^2} \\
n_s &=& 1 + 4 \epsilon - 2 \eta = 1 - \frac{12 k^2 }{\phi^2}
\end{eqnarray}
Now, the number of e-foldings fixes the value of the field at which the scales of interest at present left the horizon,
\begin{equation}
N = \int H dt = - \left.\frac{\phi^2}{4 k^2}\right|^{\phi_f}_{\phi_i} \simeq \frac{\phi_i^2}{4 k^2}
\end{equation}
Using $N \simeq 62$ we need $\phi_i \simeq 11 \times k \simeq 3.1 \times M_{\rm Pl}$, and we obtain $n_s= 1 - 3 /N \simeq 0.95$ and $r \simeq 13 ~\epsilon = 13/N \simeq 0.21$. Therefore, we see that this potential would be able to reproduce approximately the observed values for the spectral index and the tensor-to-scalar ratio, but only if the field values during inflation are well above the Planck mass.
However, if our field makes excursions well-beyond the Planck scale, or gets very close to it, we can expect gravitational corrections
to come into play and to be very relevant. For example, higher-order curvature invariants could appear, and it is then
natural to consider also non-minimal couplings of the inflaton to gravity.
Now, let's assume that our inflaton field, $\theta$, has a
non-minimal coupling to gravity \footnote{In this case, although the
potential is exactly quartic in the Einstein frame, the effects of this
non-minimal coupling would still appear in this frame as higher order terms in the action that we have not considered} of the form $D(\theta)=(1 - \theta^2/( 3 k^2))$. The Einstein equations and the $\theta$ equations of motion in this frame, the Jordan frame, are,
\begin{eqnarray}
&D(\theta)~ \tilde H^2 = \frac{\dot{\theta}^2}{3 k^2} + \frac{2 \tilde V (\theta)}{ 3 k^2} - \dot{D}(\theta)~ \tilde H& \\
&\ddot{\theta} + 3 \tilde H \dot{\theta} + \frac{k^2}{4}~D'(\theta)~R + \tilde V'({\theta})=0 &\, ,\nonumber
\end{eqnarray}
with $\tilde H$ the Hubble rate in the Jordan frame related to the Hubble rate in the Einstein frame by Eq.~(\ref{Htrans}). Then, the potential in the Jordan frame is,
\beq
\label{thetapot}
\tilde V(\theta) = \frac{V\left(\phi(\theta)\right)}{(D(\theta))^2} = \lambda \frac{ \left(\phi(\theta)\right)^4}{\left(1 -\theta^2/(3 k^2) \right)^2}
\eeq
Using this potential we can obtain, folowing \cite{Hertzberg:2010dc},
the Jordan-frame cutoff scale wich signals the validity regime of the
effective theory after taking into account the quantum corrections
incorporated a la Coleman-Weimberg. In this case the cutoff is even larger
than the Einstein frame one, and it is given by $\Lambda = 6 M_{Pl}$.
The fields in the Jordan and Einstein frames, are related by Eq.~(\ref{fieldtrans}), that in this case can be integrated analytically,
\begin{equation}
\label{fieldredef}
\phi(\theta) = 2 \sqrt{6 \pi}~k \arctanh\left[\frac{\theta}{\sqrt{3}~ k}\right] \,,
\end{equation}
or,
\begin{equation}
\theta(\phi) = \sqrt{3}~k \tanh\left[\frac{\phi}{2\sqrt{6\pi}~ k}\right] \,.
\end{equation}
Therefore, we can see clearly that in the Jordan frame, the field $\theta$ is always subplanckian, $\theta \leq \sqrt{3/(4\pi)}~ M_{\rm Pl} \simeq 0.489~ M_{\rm Pl}$, and when $\phi \simeq 3.1~ M_{\rm Pl}$ the Jordan field $\theta \simeq 0.42~ M_{\rm Pl}$\footnote{Here, we consider only the leading term in the potential. It is clear that higher-order operators could play a role. However, it is always possible to deal with them through a symmetry. For example, in this case, we could impose a $Z_4$ symmetry.}. It is a trivial exercise to show that using Eqs~(\ref{slow-roll}--\ref{rjordan}) and (\ref{Einst-roll}--\ref{rEinstein}), and despite the fact the slow-roll parameters are different in both actions, two non-vanishing in one case corresponding to the usual slow-roll parameters, and four in the Jordan frame, all the observables are identical. Moreover, even in the Jordan frame, the potential is approximately quartic in $\theta$ at low field values, $\theta/(\sqrt{3} k) \ll 1$, as can be seen from
Eqs.~(\ref{thetapot}) and (\ref{fieldredef}). Therefore, already by the end of inflation,
both theories are nearly indistinguishable\footnote{An interesting possibility would be to have the same non-minimal coupling to gravity to play a role in the current accelerated expansion.}.
We can repeat the same exercise starting from a quadratic potential $V(\phi) = \mu^2 \phi^2$. In this case, we obtain $\epsilon = 1/(2 N_e)$, $n_s = 1 - 2/N_e$ and $r=6.9/N_e$, with $N_e = \phi_i^2/(2 k^2)$ the number of e-folds needed for inflation to solve the cosmological problems, which require $\phi_i\simeq 2.2 \times M_{\rm Pl}$. So, again we need transplanckian values for the field in the Einstein frame (less transplanckian than in the previous case due to the flatter potential), but clearly, assuming the same non-minimal coupling to gravity, the relation between the Jordan and Einstein-frame fields is the same as in Eq.~(\ref{fieldredef}), and therefore, as in the $V=\lambda \phi^4$ case, transplanckian field values become subplanckian, $\theta_i \simeq 0.35 \times M_{\rm Pl}$, once allowed to couple to the curvature.
\subsection{Generic scalar-tensor theories}
\label{generic}
In the previous model, we have specified the potential in the Einstein frame and the transformation to the Jordan frame, and we have seen that the transplanckian values of the field may be simply due to our attempt to write in Einstein form a theory that has a non-minimal coupling to gravity.
Here we will use a different strategy, we will start from an ansatz that guarantees inflation in the Jordan frame and obtain the potential and the non-minimal coupling to gravity from there.
We start from the requirement that space inflates exponentially with the inflaton field being
responsible for it, $ a = \exp{\left( -\theta/b \right)}$ and therefore $H = \dot a /a = - \dot \theta/ b$ \cite{Barenboim:2007bu}.
This ansatz establishes also the number of e-foldings in this scenario, which is given by
\beq
N_e = \int H dt = - \int \frac{\dot \theta}{b} dt = \frac{-1}{b} \int d\theta = \frac{1}{b} \left(\theta_i - \theta_f\right) \simeq \frac{\theta_i}{b}\,,
\eeq
where the fact that the scalar field is rolling down ($\theta$ is decreasing)
becomes transparent and we have chosen $\theta_f =0$ at the end of inflation for simplicity.
Using Eqs.~(\ref{einstein}) and (\ref{dotH}), with $H=-\dot \theta/b$ and $\dot f = f' \dot \theta$, we
can now obtain the relation between the Hubble rate and the coupling to gravity that will sustain the
exponential period of expansion we are longing to have,
\beq
\frac{H'}{H} = \frac{ 2 b/k^2 +b D'' + D'}{2 D - b D'}
\eeq
The following step is clear, we need to choose either a coupling to gravity (as we did in the previous section) or a Hubble rate,
and then obtain the other one via this second order differential equation.
In this section, we are going to choose the form of the Hubble rate and, from there, obtain the non-minimal coupling to gravity. For simplicity we want to obtain analytic expressions for this coupling, and then not many choices for $H'/H$ are possible.
Unfortunately the most natural and easiest choice,
$H'/H \simeq 0$, which gives,
\beq
D (\theta) = -\frac{2 b \theta}{k^2} - A~ b e^{- \theta /b} + B\,,
\eeq
where $A$ and $B$ are the two integration constants, with $B$ dimensionless and $[A] = 1/[b]= 1/[\theta]$, fails phenomenologically. It gives a red spectral index,
($n_s > 1 $) and therefore must be abandoned.
The next choice (in order of simplicity) would be $H'/H = -1/ M$. We can still solve exactly the equation
for the coupling to curvature although the solution is not as simple as before,
\bea
\label{Dtheta}
D (\theta) = -\alpha \frac{M^2 }{k^2} +e^{\frac{\left(\alpha -1 +\sqrt{\alpha ^2-10\alpha +1 }\right) \theta}{2 \alpha M}} (\alpha \frac{M^2 }{k^2}+ 1 -B) +e^{\frac{\left(\alpha -1-\sqrt{\alpha ^2-10 \alpha +1}\right) \theta}{2 \alpha M }} B\,,
\eea
with $B$ a dimensionless integration constant, which in the following we fix at $B=0$ for convenience, and $\alpha = b/M$. We have fixed the second integration constant requiring that $D(\theta_f=0)=1$, so that at the end of inflation we naturally land in an
Einstein gravity regime. Despite its rather complicated form, we will see that, for the set
of parameters needed to produce the correct inflation phenomenology, the behaviour of $D(\theta)$ is
not as sophisticated as it can appear by looking at its full expression.
\begin{figure}
\centerline{\epsfxsize 4. truein \epsfbox {dfun.eps} }
\caption
{Coupling of the scalar field to the Ricci scalar curvature $R$. The figure is produced for the slow-roll regime, $\alpha = 0.007$ and taking $B=0$. The value of $M$ is irrelevant as long as $M \lesssim 0.1 M_{\rm Pl}$. The field value is given in units of $M_{\rm Pl}$
\label{dfun}}
\end{figure}
The number of e-foldings in this scenario can be written in terms of the new mass scale $M$ and
the parameter $\alpha$ as $N_e \simeq \theta_i/b = \theta_i/(\alpha M)$.
For the slow-roll parameters defined in Eqs.~(\ref{slow-roll}), we have,
\begin{eqnarray}
\label{sroll2}
\epsilon_1 &=& \frac{H' \dot{\theta}}{H^2} = -b \frac{H'}{H} = -\alpha \\
\epsilon_2 &=& \frac{\ddot{\theta}}{H \dot{\theta}} = \frac{(\dot{\theta})'}{H } = - b \frac{H'}{H} = -\alpha\\
\epsilon_3 &=&\frac{1}{2} \frac{D' \dot{\theta}}{H D} = -\frac{b}{2} \frac{D'}{D} \simeq - \frac{\left( 1 + \frac{M^2}{k^2} \alpha\right) \cdot \omega(\alpha)e^{-\frac{N_e}{2}\cdot \omega(\alpha)} }{4\left(\frac{M^2}{k^2} \alpha - \left( 1 + \frac{M^2}{k^2} \alpha\right) e^{-\frac{N_e}{2}\cdot\omega(\alpha)} \right)}\\
\epsilon_4 &=& \frac{1}{2} \frac{E' \dot{\theta}}{H E} = -b \frac{2 D' + 3 k^2 D' D''}{4D + 3 k^2 (D')^2} \simeq -\left(1+ \frac{M^2}{k^2} \alpha\right)~ e^{-\frac{N_e}{2}\cdot \omega(\alpha)} \cdot \\
&& \frac{\frac{M^2}{k^2} \alpha^2 \cdot \omega(\alpha) + 3 \left(1 + \frac{M^2}{k^2} \alpha \right) \left((1-4\alpha +\alpha^ 2) \cdot \omega(\alpha) - 4 \alpha (1-\alpha)\right)~ e^{-\frac{N_e}{2}\cdot \omega(\alpha)}}{4 \frac{M^4}{k^4} \alpha^3 -4 \frac{M^2}{k^2} \alpha^2 \left(1 + \frac{M^2}{k^2} \alpha \right)~ e^{-\frac{N_e}{2}\cdot \omega(\alpha)}+ 3 \left(1 + \frac{M^2}{k^2} \alpha \right)^2((1-\alpha) \cdot \omega(\alpha) - 4 \alpha) ~ e^{-N_e\cdot \omega(\alpha)}} \, ,\nonumber
\end{eqnarray}
with $\omega(\alpha)=(1 - \alpha -\sqrt{1 - 10 \alpha +\alpha^2})$ and always taking $B=0$ in Eq.~(\ref{Dtheta}).
These expressions are simplified in the limit $\alpha \ll 1$,
corresponding to the slow-roll regime,
\begin{eqnarray}
\label{sroll3}
\epsilon_3 &\simeq& - \alpha~ e^{-2 N_e\alpha} \frac{(1 +3 \alpha) }{\left(\frac{M^2}{k^2} \alpha - e^{- 2 N_e\alpha} \right)} \\
\epsilon_4 &\simeq& -\alpha ~ e^{-2 N_e\alpha} \frac{\frac{M^2}{k^2} (1+ 3 \alpha) + 12 ~ e^{- 2 N_e\alpha}}{\frac{M^4}{k^4} \alpha - \frac{M^2}{k^2}~ e^{- 2 N_e \alpha}- 6~ e^{- 4 N_e \alpha} } \, .
\end{eqnarray}
The usual slow-roll parameters, in the interesting region $\frac{M^2}{k^2}< e^{- 2 N_e \alpha}$ ( {\it i.e.} basically $120~ \alpha \sim O(1)$), become,
\bea
\tilde \epsilon& =& \epsilon_1 - \epsilon_3 \simeq -2 \alpha \\
\tilde \eta &=& \epsilon_2 + \epsilon_4 - 3 \epsilon_3 \simeq - 2 \alpha \\
n_s &=& 1 + 4 \tilde \epsilon - 2 \tilde \eta \simeq 1 -4 \alpha\\
r &=& 13.8 \left| \tilde \epsilon \right| \simeq 27.6 \alpha\,.
\eea
And, for $\alpha = 0.007$, we obtain $r\simeq 0.19$ and $n_s\simeq 0.97$. As explained before, the value of $M$ is irrelevant for these observables as long as $M \lesssim 0.1 M_{\rm Pl}$.
As before, we can get the potential, which has a rather baroque expression in full form, although it is basically an exponential potential $ e^{-\frac{2 \theta }{M}}$,
\bea
V(\theta)&=&\frac{3 \; k^2}{4} e^{- \frac{2 \theta }{M}}\left[\left(1+\frac{M^2 }{k^2} \alpha\right)( 2 + \omega(\alpha))~ e^{-\frac{- \theta \cdot \omega (\alpha ) }{2 M \alpha }} -\frac{2}{3} \frac{M^2}{k^2}\alpha (3 + \alpha) \right] \,,
\eea
as can be seen in the left-side plot in Fig.~\ref{potential}. Here, we see that the potential in the $\theta$ field is decreasing and seems not able to produce inflation. However, in the Jordan frame, we must take also into account the effects of the non-minimal coupling to gravity, and then the corresponding potential in the Einstein frame becomes much more attractive. This is shown in the right-side plot in Fig.~\ref{potential}, where we plot the potential in the Einstein frame as a function of the $\theta$ field (the Einstein potential in terms of the $\phi$ field is obtained changing variables using Fig.~\ref{field}).
\begin{figure}
\centerline{\epsfxsize 3.3 truein \epsfbox {potential.eps} \epsfxsize 3.3 truein \epsfbox{potphi.eps} }
\caption
{Potential of the scalar field in the Jordan frame in terms of the non-minimally
coupled field (left). The potential in the Einstein frame in terms of the non-minimally coupled
variable is shown in the right-side panel. The figures are produced for the slow
roll regime, $\alpha = 0.007$ and taking $B=0$. The value of $M$ is irrelevant as long as $M \lesssim 0.1 M_{\rm Pl}$. The field value is given in units of $M_{\rm Pl}$.
\label{potential}}
\end{figure}
In this case, the exponential potential is well-behaved and the
quantum corrections at one-loop are again small in the region of
interest for inflation. The inflaton mass is always smaller than
$M_{Plank}$. Thus as shown in Ref.~\cite{Hertzberg:2010dc} the cutoff
is still $M_{Plank}$ and there are no UV issues below that scale.
Courtesy of the sophisticated form of $D(\theta )$, it is clear that the conformal
transformation, which will turn the $\theta $ field into a more familiar minimally
coupled scalar field enjoying Einstein gravity, cannot be carried out analytically.
In Fig.~\ref{field}, we show the relation between both fields, where, as before, our point becomes
transparent: the non-minimally coupled field is subplanckian at $\theta_i= \alpha M N_e \simeq 0.01~ M_{\rm Pl}$, while the minimally coupled one is not, $\phi_i \simeq 140~ M_{\rm Pl}$.
At this point, it is important to stress that, in this case, we have not engineered the
coupling to gravity in order to support our point. Instead, we have only asked our scale factor
to sustain an inflationary period and looked for the simplest possible choice allowing us to
solve analytically the second order differential equation which relates the Hubble rate to
the curvature coupling. In this context, the emergence of a subplanckian field value in the
Jordan frame cannot be considered the result of a fine-tuning.
\begin{figure}
\centerline{\epsfxsize 4. truein \epsfbox {field.eps} }
\caption
{Minimally coupled scalar field in terms of the non-minimally
coupled field. The figure is produced for the slow
roll regime, $\alpha = 0.007$, $M=0.02$ and taking $B=0$.
Both field values are given in units of $M_{\rm Pl}$. From the figure it becomes apparent
that while the field in the Einstein frame is transplanckian, the one in the Jordan frame is not.
\label{field}}
\end{figure}
\subsection{f(R) gravity models}
\label{fR}
Unlike the previous cases, where inflation was given by the interaction between the matter
sector and the modified gravity sector,
to finish we will consider the case where inflation is entirely nourished by gravity \cite{Sotiriou:2008rp},
\beq
L = \frac{k^2}{4} f(R) \,.
\eeq
In this case, equations for the background become,
\bea
H^2 &=& \frac{1}{3 F(R)} \left( \frac{R F(R) -f(R)}{2} - 3 H \dot{F}(R) \right)\,,
\nonumber \\
\dot{H}& =& - \frac{1}{2 F(R)} \left( \ddot{F}(R) - H\dot{F}(R) \right)\,,
\eea
where $F(R) = \partial f(R)/\partial R$ and $R = - 6 (2 H^2 + \dot{H}) $.
Face value, this case is quite distant from the previous ones, as now there is no conformal
transformation capable of driving us to the Einstein frame. However, once we depart
in a non-trivial way from the standard gravity, the field equations for $R$ become
higher-order effectively, signalling the presence of additional degrees of freedom.
This feature can be taken care by the introduction of an auxiliary scalar field and going, as an intermediate step, through a Brans-Dicke form of our model \cite{Sotiriou:2008rp}.
Then, in a similar way as in the previous cases, a conformal transformation will take us
away from our $f(R)$ gravity to the kingdom of Einstein gravity plus a minimally coupled scalar field with a specific potential. In the case we are studying,
under a conformal transformation, the metric is redefined as,
\beq
\hat{g}_{ab} = \Omega^2 g_{ab}\,,
\eeq
where $\Omega $ is a spacetime position-dependent factor and is defined to be,
\beq
\Omega^2 = F(R) = \exp{\left(\sqrt{\frac{2}{3 k^2}} \phi\right)}\,,
\eeq
and $\phi$ is the new dynamical variable we have obtained after conformal transformation of the Brans-Dicke auxiliary field,
\beq
\phi = \sqrt{\frac{3 k^2}{2}} \ln F(R)\,.
\eeq
The Lagrangian in terms of the $\phi$ field is given by,
\beq
L = - \left( \frac{k^2}{4}\hat{R} - \frac{1}{2} g^{\mu \nu} \partial_\mu \phi \partial_\nu \phi + V(\phi) \right)\,,
\eeq
where $\hat{R}$ is the conformally transformed Ricci scalar and the potential has the form,
\beq
V(\phi) = \frac{k^2}{4} \frac{f(R) - R F(R)}{ F^2(R)}\,.
\eeq
For the sake of concreteness and to make our point even more transparent, we will consider the following ad-hoc gravity during inflation \footnote{Although every single inflationary model is a toy model, we would like to stress that there is not a physics motivation for the proposed modification of gravity. At the same time, we should also bear in mind that physics without assumptions or caveats is unimaginable. We thus, left the reader judge by himself the degree of skepticism that is appropriate when considering modified gravity models of inflation like the one presented here.},
\beq
f(R) = R \left( 1 + (R/M^2) ^{5/4} \right) \,,
\eeq
where $M$ is an arbitrary mass scale and we assume that during inflation the second term dominates over the first one, implying that during inflation $H^2 \gg M^2$. Clearly, once the inflationary phase is over, we smoothly approach Einstein gravity
In this case,
\beq
F(R) = \exp{\left( \sqrt{\frac{2}{3}} \phi\right)}\,,
\eeq
and
\beq
V(\phi) = \frac{-5 k^2}{54} \left(\frac{2}{3}\right)^{3/5} M^2 e^{-2 \sqrt{\frac{2}{3}} \phi} \left(e^{\sqrt{\frac{2}{3}} \phi}-1\right)^{9/5}\,.
\eeq
As before, the form of the potential looks rather complicated but its shape is pretty simple,
as can be seen from figure~\ref{fr}. In fact, already by eye, we can guess that this kind of potential
should be able to accommodate a decent period of inflation.
\begin{figure}
\centerline{\epsfxsize 4. truein \epsfbox {fr.eps} }
\caption
{Potential for the scalar field introduced in the conformal transformation. From the figure it is obvious that the potential does have the right shape to inflate. However transplanckian masses are involved.
\label{fr}}
\end{figure}
As seen in Model B, the cutoff scale in this kind of potentials is again $M_{Pl}$ and the effective theory is UV safe.
But we can do way better that guessing; the analysis of our potential is straightforward\footnote{Notice that unlike the previous cases, now we are not using the Hamilton-Jacobi formalism any longer and therefore we
need to resort to another set of slow-roll parameters, the ones calculated directly from the potential. Their relation with the spectral index takes into account this difference (see for instance \cite{Liddle:1994dx}).},
\bea
\epsilon_V &=& \frac{k^2}{2} \left( \frac{V'}{V} \right)^2 =
\frac{1}{300} \left(11-9~ \coth \left(\frac{\phi}{\sqrt{6}}\right)\right)^2\,, \\
\eta_V&=& k^2 \frac{V''}{V} =
\frac{2}{75} \left(\frac{36}{\left(e^{\sqrt{\frac{2}{3}} \phi}-1\right)^2}-\frac{63}{e^{\sqrt{\frac{2}{3}} \phi}-1}+1\right)\,,\\
n_s&=& 1 - 6 \epsilon_V + 2 \eta_V =\frac{1}{150} \left(198 \coth \left(\frac{\phi}{\sqrt{6}}\right)-171 \csch^2\left(\frac{\phi}{\sqrt{6}}\right)-52\right)\,.
\eea
The modes we are interested in studying are those that left the horizon 60 efoldings before the end of inflation, where $\phi_{\rm end}$, the field value at the end of inflation, is calculated by asking $\epsilon_V =1$. Then, the field value at horizon exit is obtained from,
\beq
N = 62= \int_{\phi_{\mbox{\tiny end}}}^{\phi_{\mbox{\tiny hor}}} 5 \sqrt{\frac{3}{2}} \left(-\frac{9}{e^{\sqrt{\frac{2}{3}} \phi }-10}-1\right) d\phi \,,
\eeq
which is independent of the specific value of $M$ and, for our choice of parameters, is well beyond $M_{\rm Pl}$, $\phi_{\rm hor} \approx 15 \; M_{\rm Pl}$, giving,
\bea
n_s = 0.97, \qquad r= 0.16 \,.
\eea
Once again, we see that an innocent modification of gravity, when casted as a minimally coupled scalar field rolling down a potential, ends up giving transplanckian field values.
Once more, we would like to stress that, as in the previous example, we have not
designed a modification of gravity able to accommodate transplanckian field values. We
have just chosen an $f(R)$ capable of producing sizeable tensor modes and found that this corresponds
to transplanckian field values once analysed as a minimally coupled scalar field.
\section{Conclusions}
\label{sec:conclusions}
In this work, we have explored the possibility that the transplanckian field values needed to accommodate the experimental results in minimally coupled single-field inflation models are only due to our insistence of imposing a minimal coupling of the inflaton field to gravity. If the theory responsible for inflation
includes a non-minimal
coupling to gravity, the energies and field values can be subplanckian
during the full inflation era in the Jordan frame, while they may
appear transplankian in the Einstein frame.
We are perfectly aware that the field value by itself carries no information, it is
after all a ``dummy'' variable, but the fact that its vacuum expectation value turns out to be
well above the Planck mass may be telling us that it is gravity (or its couplings to gravity),
and not only the inflaton potential couplings, the true drivers of inflation.
We have shown (Section \ref{monomial}: Monomial potentials ) that not only it is possible to turn the most popular inflationary potentials
($\phi^4$, $\phi^2$) into the desired regime by choosing an appropriate coupling to gravity,
but also that scalar tensor theories, designed exclusively to sustain inflation by asking the
scale factor to grow exponentially (Section \ref{generic}: Generic scalar-tensor theories), also turn subplanckian even in the simplest cases
(let us remind the reader that the case $H'/H = 0 $ was discarded, not because it does not
satisfy our conjecture, but because it leads to a spectral index larger than 1).
We have also presented a case (Section \ref{fR}: f(R) gravity models) where gravity itself is solely responsible for inflation,
and again results in transplanckian field values once interpreted as single-field inflation.
In summary, we have seen that is possible, and in fact quite natural
and easy, to find realistic examples of theories with non-minimal
coupling to gravity that have transplanckian field values if we insist
on imposing a minimal coupling to gravity, but are always subplanckian
in their ``natural'' frame. Thus, we have proven that single-field inflation models can
still accommodate a large tensor-to-scalar ratio with subplankian
field values in the presence of non-minimal coupling to gravity.
\section*{Acknowledgements}
The authors are grateful to Scott Dodelson and Joe Lykken
for useful discussions. We acknowledge support from the MEC and
FEDER (EC) Grants
FPA2011-23596 and the Generalitat Valenciana under grant PROMETEOII/2013/017.
G.B. acknowledges partial support from the European Union FP7 ITN INVISIBLES (Marie
Curie Actions, PITN-GA-2011-289442).
|
\section{Introduction}
Hilbert space frames have traditionally been used in signal processing. But over the last few years, frame theory has become one of the most applied subjects in mathematics. Fundamental to the notion of a Bessel sequence, and more specifically a frame, is that it is a {\it possibly redundant} sequence of vectors $\Phi := \{\varphi_i\}_{i\in I}$ in a Hilbert space for which the frame expansions of a vector $x$,
\begin{equation*}
Sx = \sum_{i\in I} \innerp{x}{\varphi_i} \varphi_i,
\end{equation*}
are unconditionally convergent series. But until now, no work has been done on understanding the precise unconditional behavior of the frame expansions of vectors in the Hilbert space.
This constant is important for a number of reasons. First, it is quite useful that the Bessel bound for Bessel sequences is still a Bessel bound for all subsequences. But for frame expansions, subsequences can have larger norms than the original sequence. This issue occurs regularly in frame theory where one often partitions a frame into two subsets and then works with the frame operators for the subsets of the frame. So for quantitative estimates, we need to know the unconditional constants for frame expansions of subsets of the frame. Second, when using finite dimensional methods to approximate a frame (see \cite{finitemethod}), in general the approximation constants depend on the unconditional constant of the frame.
In this paper we will show that the unconditional constants (for all standard forms of unconditional convergence) for the frame expansions are bounded above by $\sqrt{\frac{B}{A}}$ where $A,B$ are the frame bounds of the frame. See Proposition \ref{usefulthm}. This means that {\it tight frames} have 1-unconditionally convergent series for their frame expansions. We will then expand this to a classification of Bessel sequences by showing that the frame expansions are 1-unconditional if and only if the Bessel sequence is an orthogonal sum of tight frames. See Theorem \ref{bigthm}. This is surprising at first since we have not assumed the family has any lower frame bound but conclude that locally it does have lower frame bounds. It follows that this Bessel sequence is a frame if and only if the tight frame bounds of the orthogonal parts are uniformly bounded away from zero.
We will also examine closely when the unconditional constants are actually equal to $\sqrt{\frac BA}$ and whether $\sqrt{\frac BA}$ can be attained with specific choices of $x$, or as a limit. See the results following Corollary \ref{tightcor} and the example section, Section \ref{examples}.
Finally, we will conclude the paper with a relationship to frame multipliers.
\section{Frame Theory Preliminaries}\label{ftp}
A brief introduction to frame theory is given in this section. For a thorough approach to the basics of frame theory, see \cite{petesbook, ole_book}. For the entirety of the paper, $\mathscr{H}$ will denote a separable, finite or infinite dimensional, real or complex Hilbert space while $\mathscr{H}^d$ will denote an $d$-dimensional, real or complex Hilbert space. Our index set, $I$, will either be $[N] := \{1,2,\dots,N\}$ or $\mathbb{N}$. Finally, $\mathbb{I}$ will represent the identity operator.
\begin{defn}
A family of vectors $\Phi := \{\varphi_i\}_{i \in I}$ in a Hilbert space $\mathscr{H}$ is said to be a \emph{frame} if there are constants $0 < A \leq B < \infty$ so that for all $x \in \mathscr{H}$,
\begin{align}\label{frameeq}
A \norm{x}^2 \leq \sum_{i\in I} \abs{ \innerp{x}{\varphi_i} }^2 \leq B \norm{x}^2.
\end{align}
The constant $A$ is called a \emph{lower frame bound} and $B$ an \emph{upper frame bound}. If only $B$ is assumed, then it is called a \emph{$B$-Bessel sequence} or simply \emph{Bessel} when reference to the bound is unnecessary. If $A = B$, it is said to be a \emph{tight frame} and if $A = B = 1$, it is a \emph{Parseval frame}.
If there is a constant $c$ so that $\norm{\varphi_i} = c$ for all $i \in I$, it is an \emph{equal norm frame}. If there is a constant $d >0$ so that $|\langle \phi_i, \phi_j\rangle| = d$ for all $i \neq j$, then it is called \emph{equiangular}.
\end{defn}
\begin{rmk}
We will always assume the stated frame bounds are optimal, that is, we always work with the largest value of $A$ and smallest value of $B$ for which the frame inequality (\ref{frameeq}) holds.
\end{rmk}
\begin{defn}
If $\Phi := \{\varphi_i\}_{i \in I}$ is a $B$-Bessel sequence of vectors in $\mathscr{H}$, then the \emph{synthesis operator} of $\Phi$ is the operator $T \colon \ell^2(I) \to \mathscr{H}$ that gives a linear combination of the vectors, and the associated \emph{analysis operator} is the adjoint operator ${T^* \colon\mathscr{H} \to \ell^2(I)}$ given by
\begin{align*}
T \{c_i\}_{i \in I} &:= \sum_{i \in I} c_i \varphi_i, \\
T^*x &:= \{ \innerp{x}{\varphi_i} \}_{i\in I} = \{ \varphi_i^* x \}_{i\in I}
\end{align*}
respectively, with norms $\norm{T} = \norm{T^*} = \sqrt{B}$.
\end{defn}
In the finite dimensional setting, it is often convenient to work with the matrix representation. Just as a basis can be viewed as a matrix whose columns are the basis vectors, we can view a frame as a matrix whose columns are the frame vectors, and this matrix is the matrix representation for the synthesis operator. In addition to the columns of the synthesis operator square-summing to the norm of the vectors, when represented against the eigenbasis of the frame operator (definition below), the rows are also orthogonal to each other and square-sum to the eigenvalues of $S$. See \cite{amspaper} for more on this representation.
Recall that for fixed $x,y \in \mathscr{H}$ the definition of an outer product, $(x y^*)(z) = \innerp{z}{y} x$ for all $z \in \mathscr{H}$. Some readers may be accustomed to $(x \otimes y) (z) = \innerp {z}{y} x$.
\begin{defn}
The \emph{frame operator} $S\colon \mathscr{H}\to\mathscr{H}$ is the self-adjoint operator defined by $S := TT^*$ satisfying
\begin{equation*}
Sx = TT^*x = \sum_{i \in I} \innerp{x}{\varphi_i} \varphi_i = \sum_{i \in I} \phi_i \phi_i^* x = \left(\sum_{i \in I} \phi_i \phi^* \right)x
\end{equation*}
for any $x \in \mathscr{H}$ with norm $\norm{S} = \norm{TT^*} = \norm{T}^2 = B$, thus the series
\begin{equation*}
S = \sum_{i \in I} \phi_i \phi_i^*
\end{equation*}
converges in the strong operator topology. The leftmost sum for $Sx$ is called the \emph{frame expansion} of the vector $x$.
The \emph{Gramian operator} $G:\ell^2(I) \to \ell^2(I)$ is defined by $G:= T^*T$ satisfying
\begin{align*}
G \{c_j\}_{j \in I} = T^*T \{c_j\}_{j \in I} = \sum_{j \in I} c_j \{ \innerp{\varphi_j}{\varphi_i} \}_{i\in I}.
\end{align*}
In the finite case, the \emph{Gramian matrix} $[G_{ij}]_{i,j\in[N]}$ is the matrix representation of $G$ with respect to the standard orthonormal basis of $\ell^2(I)$ and has entries $G_{ij} = \langle \phi_j, \phi_i\rangle$ for all $i, j \in [N]$. This is a convenient way of viewing the norm of the vectors as well as the angles between them.
\end{defn}
\begin{defn}
We say that an operator $F:\mathscr{H} \to \mathscr{H}$ is \emph{positive} if for any $x \in \mathscr{H}$, we have $\innerp{Fx}{x} \geq 0$. It is \emph{strictly positive} if the inequality is strict. The operator is (strictly) negative if $-F$ is (strictly) positive. Given operators $F, G: \mathscr{H}\to \mathscr{H}$, we write $G \geq F$ if $G - F \geq 0$, where $0$ denotes the zero operator on $\mathscr{H}$.
\end{defn}
If $\Phi$ is a frame with lower bound and upper bounds $A$ and $B$, respectively, then
\begin{equation*}
\innerp{Ax}{x} \le \innerp{Sx}{x} = \norm{T^* x}^2 = \sum_{i \in I} \abs{\innerp{x}{\varphi_i}}^2 \le \innerp{Bx}{x},
\end{equation*}
for any $x \in \mathscr{H}$. Hence, the operator inequality
\begin{equation*}
A \cdot \mathbb{I} \leq S \leq B \cdot \mathbb{I}
\end{equation*}
holds and the frame operator is strictly positive and invertible. If $\Phi$ is $B$-Bessel, then $0 \leq S \leq B \cdot \mathbb{I}$ and $S$ is positive.
\section{Unconditional Constants of the Frame Expansions}
This section is devoted to analyzing the precise unconditional behavior of the frame expansions. First we introduce the definition of unconditional convergence and the constants involved.
\begin{defn}
Given a sequence $\{v_i\}_{i \in I}$ of vectors in $\mathscr{H}$, the series
$ \sum_{i \in I} v_i $
is said to \emph{converge unconditionally} if for any $\sigma \subset I$,
$ \sum_{i \in \sigma} v_i $
converges. There are two other equivalent definitions in which we require $\sum_{i \in I} \epsilon_i v_i$ to converge for all $\epsilon_i \in \{-1,1\}$, or alternatively require $\sum_{i \in I} a_i v_i$ to converge for any scalars $|a_i| <1$. The smallest constants $E_\sigma$, $E_\varepsilon$, and $E_a$ so that
\begin{alignat*}{2}
\norm{\sum_{i \in \sigma} v_i} \le E_\sigma &\norm{\sum_{i \in I} v_i} & \quad &\mbox{for all } \sigma \subset I, \\
\norm{\sum_{i \in I} \epsilon_i v_i} \le E_\epsilon &\norm{\sum_{i \in I} v_i} & \quad &\mbox{for all } \epsilon_i \in \{-1,1\}, \\
\norm{\sum_{i \in I} a_i v_i} \le E_a &\norm{\sum_{i \in I} v_i} & \quad &\mbox{for all } \abs{a_i} \le 1,
\end{alignat*}
hold respectively, are called the \emph{unconditional constants}. Such constants exist by the Uniform Boundedness Principle. Note that these are always at least one since the full sum $\norm{\sum_{i \in I} v_i}$ is permitted on the left-hand-side of all three definitions. Futhermore, these constants are related by $E_\sigma \le E_\epsilon \le 2E_\sigma$ and $E_\epsilon \le E_a \le 2 E_\epsilon$. See \cite{Heil_book} for more on unconditional convergence and the constants involved. For the remainder of the paper, $C_\sigma$, $C_\varepsilon$, and $C_a$ will denote the respective unconditional constants of the frame expansions of a vector $x$ with respect to a Bessel sequence $\{\varphi_i\}_{i \in I}$ (that is, where $v_i = \innerp{x}{\varphi_i}\varphi_i$ for each $i$).
\end{defn}
We begin our quest by showing that $\sqrt{\frac{B}{A}}$ is an upper bound for all three unconditional constants of the frame expansions, where $A$ and $B$ are the lower and upper frame bounds, respectively.
\begin{prop}\label{usefulthm}
Let $\Phi := \{\varphi_i\}_{i \in I}$ be a frame for $\mathscr{H}$ with lower and upper frame bounds $A$ and $B$ respectively. Then all three of the unconditional constants of the frame expansions with respect to $\Phi$ are bounded above by $\sqrt{\frac{B}{A}}$. That is,
\begin{enumerate}[(i)]
\item \label{uc_ba1} For any $\sigma \subset I$ and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in \sigma} \innerp{x}{\varphi_i} \varphi_i } \leq
\sqrt{\dfrac{B}{A}} \ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}.
\end{align*}
\item \label{uc_ba2} For any sequence $\{\varepsilon_i\}_{i\in I}$, with $\varepsilon_i \in \{-1,1\}$ for all $i$, and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in I} \varepsilon_i \innerp{x}{\varphi_i} \varphi_i } \leq \sqrt{\dfrac{B}{A}}\ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}.
\end{align*}
\item \label{uc_ba3} For every sequence of real numbers $\{a_i\}_{i \in I}$ with $\abs{a_i} \leq 1$ for all $i$ and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in I} a_i \innerp{x}{\varphi_i} \varphi_i } \leq \sqrt{\dfrac{B}{A}}\ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}.
\end{align*}
\end{enumerate}
\end{prop}
\begin{proof}
Let $T$ and $S$ by the synthesis operator and frame operator of $\Phi$, respectively. To prove (\ref{uc_ba1}), (\ref{uc_ba2}), and (\ref{uc_ba3}) let $D: \ell^2(I) \to \ell^2(I)$ be the ``diagonal" operator defined by $D\{c_{i}\}_{i \in I} := \{d_i c_i\}_{i \in I}$ with
\begin{enumerate}[(i)]
\item $d_i = 1$ if $i \in \sigma$ and $d_i = 0$ otherwise,
\item $d_i = \varepsilon_i$, or
\item $d_i = a_i$.
\end{enumerate}
Observe that the left-hand-side of the inequality in (\ref{uc_ba1}), (\ref{uc_ba2}), and (\ref{uc_ba3}), respectively, is precisely $\norm{T D T^* x}$ and the right-hand-side is $\sqrt{\frac{B}{A}}\norm{S x}$. We have that $\norm{D} \leq 1$ in all cases which implies
\begin{align*}
\norm{T D T^* x}^2 &\leq \norm{T}^2\norm{D}^2\norm{T^* x}^2 \leq B \innerp{Sx}{x} = B \innerp{ S^{1/2} x }{ S^{1/2} x }
\end{align*}
and because $A \cdot \mathbb{I} \leq S$, this is bounded above by
\begin{align*}
B \innerp{ \left(\frac{1}{A} S\right) S^{1/2} x }{ S^{1/2} x } = \frac{B}{A} \norm{Sx}^2.
\end{align*}
Taking square roots gives the desired inequality.
\end{proof}
An immediately corollary of Proposition \ref{usefulthm} is that tight frames have $1$-unconditional frame expansions.
\begin{cor}\label{tightcor}
If $\Phi$ is a tight frame for $\mathscr{H}$, then the frame expansions with respect to $\Phi$ have unconditional constant $1$ for all forms of unconditional.
\end{cor}
One might wonder if the value $\sqrt{\frac{B}{A}}$ is optimal in the sense that there are frames for which $\sqrt{\frac BA}$ is actually equal to the unconditional constants, and not simply a bound. Corollary \ref{tightcor} clearly implies this is the case for all tight frames. However, this is not necessarily the case in general. We will show in Corollary \ref{Ex2Cor} that if a vector $x\in\mathscr{H}$ and $\sigma \subset I$ exists so that equality holds in Theorem \ref{usefulthm}(\ref{uc_ba1})
(thus, in this case, $C_\sigma = \sqrt{\frac BA}$ and so all unconditional constants are equal to this),
then $A = B$ and the frame is necessarily tight. As a consequence, a simple compactness argument implies that a non-tight frame for a finite-dimensional space has $C_\sigma < \sqrt{\frac{B}{A}}$. It is still an open question as to whether $\sqrt{\frac{B}{A}}$ can be arbitrarily approached in the infinite dimensional case. We first introduce some convenient notation.
\begin{defn}\label{subops}
Let $\{\varphi_i \}_{i \in I}$ be $B$-Bessel sequence of vectors with synthesis operator $T$ and frame operator $S$. For any $\sigma \subset I$ denote by $T_\sigma$, $T^{*}_{\sigma}$, and $S_\sigma$ the following related operators
\begin{alignat*}{2}
\quad T_\sigma\{c_i\}_{i \in \sigma} :=& \sum_{i \in \sigma} c_i \varphi_i, & \quad \{c_i\}_{i \in \sigma} &\in \ell^2(\sigma) \\
T_\sigma^* x :=& \left\{\innerp{x}{\varphi_i} \right\}_{i \in \sigma}, & x &\in \mathscr{H} \\
S_\sigma :=& T_\sigma T_\sigma^* = \sum_{i \in \sigma} \varphi_i \varphi_i^*,
\end{alignat*}
respectively.
\end{defn}
\begin{rmk}
It follows that $0 \le S_\sigma$ and $S = S_\sigma + S_{\sigma^c}$ for any choice of $\sigma \subset I$. As a consequence, $S_\sigma \le S$, that is, $\|S_\sigma^{1/2}x\| \le \norm{S^{1/2}x}$ for all $x \in \mathscr{H}$.
\end{rmk}
\begin{rmk}
Note that $S_\sigma \leq S$ does not imply that $\norm{S_\sigma x} \leq \norm{Sx}$ for all $x \in \mathscr{H}$. See Proposition \ref{squares} and the discussion leading up to it for more details on this distinction.
\end{rmk}
\begin{cor} \label{Ex2Cor}
If for some $x \in \mathscr{H}$ and $\sigma \subset I$ we have the equality $\norm{S_\sigma x} = \sqrt{\frac{B}{A}} \norm{Sx}$, then
\begin{enumerate}[(i)]
\item $\Phi$ is tight, that is $A = B$, and
\item every frame vector removed is orthogonal to $x$, that is $\innerp{x}{\varphi_i} = 0$ for all $i \in \sigma^c$.
\end{enumerate}
\end{cor}
\begin{proof}
The proof to Proposition \ref{usefulthm}(\ref{uc_ba1}) used the following inequality for all $x$
\begin{equation*}
\norm{T_\sigma^* x}^2 = \norm{D^{1/2}T^* x}^2 \le \norm{T^* x}^2 = \innerp{Sx}{x}.
\end{equation*}
If $\norm{S_\sigma x} = \sqrt{\frac{B}{A}} \norm{Sx}$ holds for some $x$, then every inequality in the proof to Proposition \ref{usefulthm}(\ref{uc_ba1}) is an equality for that $x$, and $\norm{T_\sigma^* x}^2 = \norm{T^* x}^2$ holds. This first implies (ii) because
\begin{align*}
0 &= \norm{T^* x}^2 - \norm{T_\sigma^* x}^2 \\
&= \sum_{i \in I} \abs{\innerp{x}{\varphi_i}}^2 - \sum_{i \in \sigma} \abs{\innerp{x}{\varphi_i}}^2 \\
&= \sum_{i \in \sigma^c} \abs{\innerp{x}{\varphi_i}}^2
\end{align*}
which means that $\innerp{x}{\varphi_i} = 0$ for all $i \in \sigma^c$. Secondly, this implies that for this same $x$, we have $S_{\sigma^c}x = \sum_{i \in \sigma^c} \innerp{x}{\varphi_i}\varphi_i = 0$, so that
\begin{align*}
Sx &= S_\sigma x + S_{\sigma^c}x \\
&= S_\sigma x.
\end{align*}
But this implies $\displaystyle \norm{S_\sigma x} = \norm{Sx}$, which with the original equality gives $\sqrt{\frac{B}{A}} = 1$, giving (i).
\end{proof}
\begin{rmk}
Finding a vector $x\in\mathscr{H}$ and $\sigma \subset I$ so that $\norm{S_\sigma x} = \norm{Sx}$, does not imply the frame is tight. These results follow precisely because equality was obtained with $\sqrt{\frac{B}{A}}$.
\end{rmk}
This leads to one wondering just how large $\norm{S_\sigma x}$ can be compared to $\norm{Sx}$? How close can $\norm{S_\sigma x}/\norm{Sx}$ get to $\sqrt{\frac{B}{A}}$? In Section \ref{examples}, Example \ref{eg2} we show that there are frames in which (i) $\norm{S_\sigma x} \approx \sqrt{\frac{B}{A}}\norm{Sx}$ for some $x \in \mathscr{H}$ and $\sigma \subsetneq I$, (ii) $C_\sigma > \frac{1}{2}\sqrt{\frac{B}{A}}$, and (iii) $\sqrt{\frac BA}$ is arbitrarily large. Hence, $\norm{S_\sigma x}$ can be as large as one would like when compared to $\norm{Sx}$. This warrants a discussion since it seems like a contradiction at a first glance.
The remark after Definition \ref{subops} noted that $S_\sigma \leq S$ for any $\sigma \subset I$.
At first, it looks like this should imply the inequality $\norm{S_\sigma x} \leq \norm{Sx}$. However, this is not true in general. In fact, $S_\sigma \leq S$ is equivalent to $\|S_\sigma^{1/2} x\| \leq \norm{S^{1/2} x}$.
To conclude that $\norm{S_\sigma x} \leq \norm{Sx}$, we
would need $S_\sigma^2 \leq S^2$ so that
\begin{align*}
\norm{S_\sigma x}^2 &= \innerp{ S_\sigma x }{ S_\sigma x } = \innerp{ S_\sigma^2 x }{x} \\
&\leq \innerp{S^2 x}{x} = \innerp{Sx}{Sx} = \norm{Sx}^2.
\end{align*}
But $S_\sigma \leq S$ does not imply that $S_\sigma^2 \leq S^2$ in general. To guarantee this, $S_\sigma$ and $S$ would need to have the same eigenvectors, which is not certain to hold. The following proposition slightly generalizes the preceding discussion.
\begin{prop}\label{squares}
Let $\Phi := \{\varphi_i\}_{i \in I}$ be a Bessel sequence for $\mathscr{H}$ with frame operator $S$ and let $C > 0$ be a fixed constant. The following are equivalent:
\begin{enumerate}[(i)]
\item For all $\sigma \subset I$,
$$
S_{\sigma}^2 \le C S^2.
$$
\item For every $x\in \mathscr{H}$ and any $\sigma \subset I$,
$$
\ensuremath{\norm{ \sum_{i \in \sigma} \innerp{x}{\p_i} \p_i }}^2 = \norm{S_\sigma x}^2 \leq C \norm{Sx}^2 = C \ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}^2.
$$
\end{enumerate}
\end{prop}
\begin{proof}
Note that $S_\sigma^2 \leq C S^2$ holds if and only if
\begin{align*}
\norm{S_\sigma x}^2 &= \innerp{S_\sigma x}{S_\sigma x} = \innerp{S_\sigma^2 x}{x} \\
&\leq C \innerp{S^2 x}{x} = C \innerp{Sx}{Sx} = C \norm{Sx}^2,
\end{align*}
giving the desired equivalence.
\end{proof}
\begin{rmk}
Corollary \ref{tightcor} combined with Proposition \ref{squares} implies that ${S^2_\sigma \leq S^2}$ when the frame is tight.
\end{rmk}
In Section \ref{examples}, Example \ref{eg4}, we give a frame for an infinite dimensional space so that
\begin{equation}
\dfrac{\norm{\sum\limits_{i = 1}^\infty \epsilon_i\langle y_n, \phi_i \rangle \phi_i}}{\norm{\sum\limits_{i = 1}^\infty\langle y_n, \phi_i \rangle \phi_i}} \to \sqrt{\dfrac{B}{A}}\label{blahhhh}
\end{equation}
for a sequence of vectors $\{y_n\}_{n =1}^\infty$ and a specific choice of $\epsilon_i \in \{-1,1\}$. That is, $\sqrt{\frac{B}{A}}$ can be arbitrarily approached and so $C_\epsilon = C_a = \sqrt{\frac{B}{A}}$. However, it is still an open question whether equality (rather than a limit) in (\ref{blahhhh}) can be achieved with a vector $x$ and scalars $\{\varepsilon_i\}_{i=1}^\infty$ satisfying $\varepsilon_i \in \{-1,1\}$, or with $\{a_i\}_{i =1}^\infty$ satisfying $|a_i| \leq 1$ in place of $\epsilon_i$.
\section{One-unconditional convergence and sums of orthogonal tight frames}\label{maintheorem}
The main result of the paper will be presented in this section. It will be shown in Theorem \ref{bigthm} that frame expansions have unconditional constants all equal to $1$ if and only if the sequence is an orthogonal sum of tight frames. First, we extend Proposition \ref{usefulthm} and Corollary \ref{tightcor} to a Bessel sequence made up of an orthogonal sum of frames.
\begin{thm}\label{converse_bigthm}
Let $J_i$ be finite or infinite index sets for all $i \in I$. Assume that $\Phi := \{\Phi_i\}_{i \in I} = \{\varphi_{ij}\}_{i \in I,j \in J_i}$ is a Bessel sequence
for the Hilbert space
\begin{equation*}
\mathscr{H} := \left( \sum_{i\in I}\oplus \mathscr{H}_i \right)_{\ell^2}
\end{equation*}
so that for each $i \in I$, the sequence $\Phi_i := \{\varphi_{ij}\}_{j \in J_i}$ is a frame with lower and upper frame bounds $A_i$ and $B_i$ for the Hilbert space $\mathscr{H}_i$, and assume that the frame bounds satisfy $\sup\{B_i/A_i: i \in I\} <\infty$. Then the unconditional constants of the frame expansions with respect to $\Phi$ are bounded above by $\sup\{\sqrt{\frac{B_i}{A_i}} : i \in I\}$.
\end{thm}
\begin{proof}
If $S_i$ is the frame operator for $\Phi_i$ for each $i$, then $\mathcal{S} := \sum_{i \in I} \oplus S_i$ is the
frame operator of $\Phi$. Now Proposition \ref{usefulthm} can now be applied to each frame $\Phi_i$ to obtain the desired result.
\end{proof}
\begin{rmk}
Note $\Phi$ may not itself be a frame unless $\inf\{A_i : i \in I\} > 0$.
\end{rmk}
A direct result of Theorem \ref{converse_bigthm} is that the unconditional constants are $1$ when each individual frame $\Phi_i$ is tight.
\begin{cor}\label{ortho_cor}
If $\Phi := \{\Phi_i\}_{i \in I}$ is an orthogonal sum of tight frames $\Phi_i$, then the frame expansions with respect to $\Phi$ have unconditional constant $1$ for all forms of unconditional.
\end{cor}
The main theorem is the converse of Corollary \ref{ortho_cor}. To prove it, we first need a lemma.
\begin{lem}\label{bigthmlem}
If $x,y \in \mathscr{H}$ are nonzero vectors satisfying
\begin{align*}
xy^* + yx^* \geq 0,
\end{align*}
then $y = \lambda x$ for some nonzero $\lambda \in \mathbb{C}$.
\end{lem}
\begin{proof}
First define $u_1 := x/\norm{x}$, let $u_2$ be a unit vector orthogonal to $u_1$ so that $y = su_1 + tu_2$ for some $s,t \in \mathbb{C}$, and define
$$M : = x y^* + y x^*,$$
in which we have assumed that $M \geq 0$. The matrix representation of $M$ with respect to $\{u_1,u_2\}$ is
\begin{align*}
\norm{x}
\begin{bmatrix}[cc]
s + \overline{s} & \overline{t} \\
t & 0
\end{bmatrix}
\end{align*}
and because $M \geq 0$ and $\norm{x} \neq 0$, this matrix must be positive semidefinite and hence $t = 0$. Therefore, $y = \lambda x$ with $\lambda := s/\norm{x}$.
\end{proof}
\begin{thm}\label{bigthm}
If $\Phi := \{\varphi_i\}_{i \in I}$ is a spanning Bessel sequence of vectors in $\mathscr{H}$, then the following are equivalent:
\begin{enumerate}[(i)]
\item \label{s21} For every $\sigma \subset I$ and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in \sigma} \innerp{x}{\varphi_i} \varphi_i } \leq \ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}
\end{align*}
\item \label{s24} For any sequence $\{\varepsilon_i\}_{i\in I}$ with $\varepsilon_i \in \{-1,1\}$ and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in I} \varepsilon_i \innerp{x}{\varphi_i} \varphi_i } \leq \ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}
\end{align*}
\item \label{s23} For every sequence of real numbers $\{a_i\}_{i \in I}$ with $\abs{a_i} \leq 1$ for all $i$ and for every $x \in \mathscr{H}$,
\begin{align*}
\norm{ \sum_{i \in I} a_i \innerp{x}{\varphi_i} \varphi_i } \leq \ensuremath{\norm{ \sum_{i \in I} \innerp{x}{\p_i} \p_i }}
\end{align*}
\item \label{s22} There is a partition $\{\mu_j\}_{j \in J}$ of $I$ satisfying:
\begin{enumerate}
\item \label{s221} For every $j \in J$, $\{\varphi_i\}_{i \in \mu_j}$ is a tight frame for its closed span,
\item \label{s222} For any $j_1,j_2 \in J$ with $j_1 \neq j_2$, it follows that $\innerp{\varphi_{k_1}}{\varphi_{k_2}} = 0$ for any $k_1 \in \mu_{j_1}$ and $k_2 \in \mu_{j_2}$.
\end{enumerate}
In other words,
\begin{align*}
\mathscr{H} = \left( \sum_{j \in J} \oplus \, \overline{\spn} \{\varphi_i : i \in \mu_j\} \right)_{\ell^2}
\end{align*}
where $\{\varphi_i\}_{i \in \mu_j}$ is a tight frame for $\overline{\spn} \{\varphi_i : i \in \mu_j\}$.
\end{enumerate}
Hence, when any of the above equivalences hold, $\Phi$ is a frame if and only if the infimum of the
tight frame bounds is not equal to zero.
\end{thm}
\begin{proof}
The implications (\ref{s22}) $\Rightarrow$ (\ref{s21}), (\ref{s24}), (\ref{s23}) is given by Corollary \ref{ortho_cor}. Also, (\ref{s23}) $\Rightarrow$ (\ref{s21}), (\ref{s24}) and (\ref{s24}) $\Rightarrow$ (\ref{s21}) are immediate from how the unconditional constants are related. All that is left to show is (\ref{s21}) $\Rightarrow$ (\ref{s22}).
Thus, assume (\ref{s21}) holds. Recall that (\ref{s21}) is equivalent to $S_\sigma^2 \leq S^2$ for all $\sigma \subset I$ (see Proposition \ref{squares}). Also, without loss of generality we can assume that $\varphi_i \neq 0$ for all $i \in I$. First, we show that all vectors in the Bessel sequence are eigenvectors of $S$. Let $i \in I$ and put $\sigma := I \backslash \{i\}$ so that $S_\sigma = S - \phi_i \phi_i^*$. Hence,
\begin{align*}
S^2 \ge S_\sigma^2 &= (S - \varphi_i \varphi_i^*)^2 \\
&= S^2 - S(\phi_i \phi_i^*) - (\phi_i \phi_i^*) S + (\phi_i \phi_i^*)^2
\end{align*}
which further implies that
\begin{equation*}
(S\phi_i) \phi_i^* + \phi_i (S\phi_i)^* \ge (\phi_i \phi_i^*)^2 > 0.
\end{equation*}
It therefore follows that $S\phi_i \neq 0$ and that $(S\phi_i) \phi_i^* + \phi_i (S\phi_i)^* \geq 0$ so that Lemma \ref{bigthmlem} can be applied to give $S\phi_i = \lambda_i \phi_i$ for some nonzero scalar $\lambda_i$. That is, $\phi_i$ is an eigenvector of $S$.
Finally, let $\{\lambda_j\}_{j \in J}$ be an enumeration of the distinct nonzero eigenvalues of $S$ and let $\mu_j := \{i \in I : \lambda_i = \lambda_j\}$. Then $\{\mu_j\}_{j \in J}$ is a partition of $I$ for which $\{\varphi_i\}_{i \in \mu_j}$ is a tight frame for its closed span with tight frame bound $\lambda_j$. This follows from using the facts that $S = \sum_{j \in J} S_{\mu_j}$ and $S_{\mu_k} S_{\mu_\ell} = 0$ for $k \neq \ell$ to show that $P_j := S_{\mu_j}/\lambda_j$ is a projection.
\end{proof}
\begin{rmk}
Note that the previous theorem classifies where signed frames, scaled frames (with scaling factors $\abs{a_i} \le 1$), and weighted frames can produce larger norms for the frame operator than the original frame, and Proposition \ref{usefulthm} gives the bounds.
\end{rmk}
\begin{rmk}
Theorem \ref{converse_bigthm} and Theorem \ref{bigthm} can both be reformulated in terms of fusion frames (see \cite{amspaper} for related definitions). If a fusion frame is made up of orthogonal subspaces $W_i$ for $i \in I$ and $\{\phi_{ij}\}_{j \in J_i}$ is a frame for $W_i$ with bounds $A_i, B_i$, then Theorem \ref{converse_bigthm} gives that the fusion frame expansions have unconditional constants bounded above by $\sup\{\sqrt{\frac{B_i}{A_i}} : i \in I\}$. Furthermore, Theorem \ref{bigthm} implies that a 1-unconditional fusion frame $(W_i,w_i)$ is actually an orthogonal sum of the fusion subspaces.
\end{rmk}
\section{Examples}\label{examples}
This section contains two fundamental examples. The first shows a family of frames for which $C_\sigma > \frac{1}{2}\sqrt{\frac{B}{A}}$ and that $\norm{S_\sigma x}$ can be arbitrarily large when compared to $\norm{Sx}$ for appropriate choices of $\sigma \subsetneq I$ and $x \in \mathscr{H}$. The second gives a non-tight infinite frame for which $C_\varepsilon = \sqrt{\frac BA}$.
\begin{eg}\label{eg2}
We will establish a family of frames indexed by a positive integer $N$, such that $\sqrt{\frac BA} = \sqrt{N}$, and show
\begin{align*}
\frac{\sqrt{N}}{2} \norm{S x} < \norm{S_\sigma x} \le \sqrt{N} \norm{Sx}.
\end{align*}
This establishes that $C_\sigma > \frac{1}{2}\sqrt{\frac{B}{A}}$ and the proportionality $\norm{S_\sigma x} \approx \sqrt{N} \norm{Sx}$, so that the ratio $\norm{S_\sigma x} \big/ \norm{Sx}$ grows arbitrarily large as $N$ is taken towards infinity.
Fix a positive integer $N \geq 3$. Let $\{e_i\}_{i \in [N]}$ denote the standard orthonormal basis for $\ell_2(N)$. Set $v$ to be the so called ``all ones" vector and let $P$ be the orthogonal projection onto this vector:
\begin{align*}
v := \sum_{i=1}^{N}e_i =
\begin{bmatrix}
1 \\ \vdots \\ 1
\end{bmatrix}, \qquad
P := \frac{vv^*}{\norm{v}^2} = \frac{1}{N}
\begin{bmatrix}
1 & & 1 \\
& \ddots & \\
1 & & 1
\end{bmatrix}.
\end{align*}
Denote by $\mathrm{X}$, the $N-1$ dimensional hyperplane orthogonal to $v$. Then $(\mathbb{I}-P)$ is the projection onto $\mathrm{X}$ and has the form
\begin{align}
\mathbb{I}-P =
\begin{bmatrix}
1 - \frac{1}{N} & & - \frac{1}{N} \\
& \ddots & \\
- \frac{1}{N} & & 1 - \frac{1}{N}
\end{bmatrix}
= \left[\delta_{ij} - \dfrac{1}{N}\right]_{i,j \in [N]}. \label{IP}
\end{align}
Define the following sequence of vectors $\Phi := \{\varphi_i\}_{i\in[N]}$ and view them as columns in the synthesis operator for $\Phi$:
\begin{gather*}
\varphi_i := (\mathbb{I} - P) e_i = e_i - \frac{1}{N} v, \quad \mbox{for all } i \in [N], \\
T
:= \begin{bmatrix}
\varphi_1 & \cdots & \varphi_N
\end{bmatrix}
= (\mathbb{I}-P)
\begin{bmatrix}
e_1 & \cdots & e_N
\end{bmatrix}
= (\mathbb{I}-P)\mathbb{I} = (\mathbb{I}-P).
\end{gather*}
As an orthogonal projection, $(\mathbb{I}-P)$ is self-adjoint and idempotent so we have $T = T^* = S = G = (\mathbb{I}-P)$. Thus $(\mathbb{I}-P)$ serves many purposes. It gives the vectors of $\Phi$ as well as the synthesis, analysis, frame, and Gramian operators. As a projection onto $\mathrm{X}$, $S$ fixes every $x \in \mathrm{X}$ and so $\Phi$ forms a Parseval frame for its range, $\mathrm{X}$. Furthermore, since $G$ is given by (\ref{IP}), the equality of the diagonal entries and the equality of the off-diagonal entries implies that the frame is equal norm and equiangular, respectively.
Now, for ${\sigma\subset [N]}$ notice that
\begin{align}
\norm{\sum_{i \in \sigma} \varphi_i }^2
& = \innerp{ \sum_{i \in \sigma} \varphi_i }{ \sum_{j \in \sigma} \varphi_j } \notag \\
& = \sum_{i,j \in \sigma} \left( \delta_{i,j}- \frac{1}{N} \right) \label{e:1} \\
& = \abs{\sigma} \left( 1 - \frac{\abs{\sigma}}{N} \right). \notag
\end{align}
Therefore, if we sum over all $N$ frame vectors so that $\abs{\sigma}=N$ and take its norm, the computation in (\ref{e:1}) gives
\begin{equation*}
\label{e:3} \sum_{i=1}^N\varphi_i = 0.
\end{equation*}
That is, $\phi_1 \in \spn{\{\varphi_i\}_{i = 2}^N}$ and the removal of $\varphi_1$ still leaves a frame $\Psi:= \{\varphi_i\}_{i = 2}^N$ for $\mathrm{X}$ with frame operator $S^\Psi = S - \varphi_1\varphi_1^* = \mathbb{I} - P - \varphi_1\varphi_1^*$. The upper frame bound of $\Psi$ remains 1, however, the lower frame bound is $1/N$ since $\varphi_1\varphi_1^*\le \norm{\varphi_1}^2 (\mathbb{I}-P)$ combined with the fact that $\|\varphi_1\|^2 = 1 - \frac{1}{N}$ implies
\begin{equation*}
S^{\Psi} = \mathbb{I} - P - \varphi_1\varphi_1^* \ge (1 - \norm{\varphi_1}^2) (\mathbb{I} - P) = \frac{1}{N}(\mathbb{I} - P).
\end{equation*}
This lower bound is achieved since $S^{\Psi}\varphi_1= (\mathbb{I} - P)\varphi_1 - \phi_1\phi_1^*\phi_1 = {\varphi_1 - \varphi_1(1 - \frac 1N)} = \frac{1}{N}\varphi_1$. Proposition \ref{usefulthm} therefore gives that
\begin{equation}\label{baex3:1}
\norm{S_\sigma^\Psi x} \leq \sqrt{N} \norm{S^\Psi x}
\end{equation}
for all $\sigma \subset \{2,\dots,N\}$ and all $x \in \mathrm{X}$. We will show that there is a ${\sigma \subsetneq \{2,\dots,N\}}$ for which
\begin{equation}\label{baex3:2}
\dfrac{\sqrt{N}}{2} \norm{S^\Psi\varphi_1} < \norm{S_\sigma^\Psi \phi_1}.
\end{equation}
This shows that $C_\sigma > \frac{\sqrt{N}}{2} = \frac{1}{2}\sqrt{\frac{B}{A}}$. Also, (\ref{baex3:2}) combined with (\ref{baex3:1}) will give that
\begin{align*}
\norm{S^\Psi \varphi_1} \approx \sqrt{N}\norm{S_\sigma^\Psi \phi_1}
\end{align*}
so that taking $N$ arbitrarily large will show that $\norm{S^\Psi_\sigma \varphi_1}$ can be arbitrarily large when compared to $\norm{S^\Psi \varphi_1}$.
To prove (\ref{baex3:2}), first notice that for every $\sigma \subset \{2,\dots,N \}$ we have
\begin{align*}
S_\sigma^\Psi \phi_1 = \sum_{j \in \sigma} \innerp{\varphi_1}{\varphi_j} \varphi_j &= -\frac{1}{N} \sum_{j \in \sigma} \phi_j
\end{align*}
so that by (\ref{e:1}),
\begin{align}\label{e:2}
\norm{S_\sigma^\Psi \phi_1}^2 &= \frac{1}{N^2}\abs{\sigma}\left( 1 - \frac{\abs{\sigma}}{N} \right) = \frac{N\abs{\sigma}-\abs{\sigma}^2}{N^3}.
\end{align}
Now, putting $N-1$ in for $\abs{\sigma}$ implies
\begin{equation*}\label{baex3:3}
\norm{S^\Psi \varphi_1}^2 = \dfrac{N-1}{N^3}.
\end{equation*}
As a parabola in $\abs{\sigma}$, equation (\ref{e:2}) obtains its maximum at $\abs{\sigma} = N/2$ and has zeros at $\abs{\sigma} \in \{0,N\}$. Any nonempty subset of $\Psi$ will have $\abs{\sigma} \in \{1,\dots,N-1\}$ so that $\norm{S_\sigma^\Psi \phi_1}^2 \ge \norm{S^\Psi \varphi_1}^2$. We want to maximize this left hand side, so choose any $\sigma \subset \{2,\dots,N\}$ with $\abs{\sigma}$ the largest integer less than or equal to $N/2$. Note this will be a proper subset since $N \ge 3$. Therefore, because $\abs{\sigma} \ge \frac{N-1}{2}$, (\ref{e:2}) gives
\begin{equation*}
\norm{S^\Psi_\sigma\phi_1}^2 = \frac{N\abs{\sigma}-\abs{\sigma}^2}{N^3} \ge \frac{N^2-1}{4N^3} = \left( \frac{N+1}{4} \right) \left( \frac{N-1}{N^3} \right) > \frac N4 \norm{S^\Psi \varphi_1}^2
\end{equation*}
which implies (\ref{baex3:2}) by taking square roots.
\end{eg}
We next give a non-trivial example of an infinite dimensional frame in which $C_{\varepsilon} = \sqrt{\frac{B}{A}}$.
\begin{eg}\label{eg4}
Let $\{e_i\}_{i=1}^{\infty}$ be the unit vector basis of
$\ell_2(\mathbb{N})$ and define $\Phi := \{\phi_i\}_{i=1}^{\infty}$ by
\begin{equation*}
\phi_i:= e_i+\frac{1}{2}e_{i+1}
\end{equation*}
for all $i \in \mathbb{N}$. First, note that $\Phi$ spans $\ell^2(\mathbb{N})$ because
\begin{equation*}
\sum_{i=0}^{\infty}\frac{(-1)^{i}}{2^i}\phi_{i+j} = e_j
\end{equation*}
for all $j \in \mathbb{N}$.
Next, we compute the frame bounds. For every $x \in \ell^2(\mathbb{N})$ we have that
\begin{align*}
\norm{T^*x}^2 = \sum_{i = 1}^\infty \abs{\innerp{x}{\varphi_i}}^2 &= \sum_{i = 1}^\infty \abs{\innerp{x}{e_i + \dfrac{1}{2} e_{i + 1}}}^2 \\
&= \sum_{i = 1}^\infty \abs{\innerp{x}{e_i}}^2 + \frac{1}{4} \sum_{i = 1}^\infty \abs{\innerp{x}{e_{i+1}}}^2 + \re{\sum_{i = 1}^\infty \innerp{x}{e_i} \overline{\innerp{x}{e_{i+1}}}} \\
&= \norm{x}^2 + \dfrac{1}{4}\left(\|x\|^2 - |x_1|^2\right) + \re{\sum_{i = 1}^\infty \innerp{x}{e_i} \overline{\innerp{x}{e_{i+1}}}}.
\end{align*}
By the Cauchy-Schwarz inequality,
\begin{align*}
\re{\sum_{i = 1}^\infty \innerp{x}{e_i} \overline{\innerp{x}{e_{i+1}}}} &\leq \left(\sum_{i = 1}^\infty \abs{\innerp{x}{e_i}}^2 \right)^{1/2} \left(\sum_{i = 1}^\infty \abs{\innerp{x}{e_{i+1}}}^2 \right)^{1/2} \\
&=\|x\|\left(\|x\|^2 - |x_1|^2 \right)^{1/2}
\end{align*}
with equality if and only if $c\innerp{x}{e_i} = \innerp{x}{e_{i+1}}$ for all $i \in \mathbb{N}$ for some constant $c$. Therefore,
\begin{align*}
\|T^*x\|^2 \geq \|x\|^2 + \dfrac{1}{4} \left(\|x\|^2 - |x_1|^2\right) - \|x\|\left(\|x\|^2 - |x_1|^2 \right)^{1/2} \geq \dfrac{1}{4}\|x\|^2
\end{align*}
and
\begin{align*}
\|T^*x\|^2 \leq \|x\|^2 + \dfrac{1}{4} \left(\|x\|^2 - |x_1|^2\right)+ \|x\|\left(\|x\|^2 - |x_1|^2 \right)^{1/2} \leq \dfrac{9}{4}\|x\|^2.
\end{align*}
The first inequality in each line is equality for vectors of the form $x =(a,ca,c^2a,c^3a,\dots)$ with $|c| < 1$, where this assumption on $c$ guarantees it is in $\ell^2(\mathbb{N})$. The second inequality in each line becomes tight with $c$ negative or positive, respectively, as $a \to 0$. Therefore, the lower and upper frame bounds are $A:=1/4$ and $B:=9/4$, respectively, and so $B/A = 9$.
Now, define the vector $y$ to be so that $\innerp{y}{\phi_i} = (-1)^{i+1}$ for all $i\in\mathbb{N}$. The existence of such a vector can be checked by a recursive computation. Note that this vector will not lie in $\ell^2(\mathbb{N})$ since its image under the analysis operator does not. However, the truncated vectors $y_n$ which equal $y$ on the first $n$ coordinates and zero elsewhere are in $\ell^2(\mathbb{N})$ . For every $n \in \mathbb{N}$,
\begin{align*}
\norm{\sum_{i=1}^\infty \innerp{y_n}{\phi_i}\phi_i}^2
&= \norm{\sum_{i = 1}^n \innerp{y}{\phi_i}\phi_i}^2 = \norm{\sum_{i = 1}^n (-1)^{i+1} \phi_i}^2 \\
&= \norm{e_1 - \dfrac{1}{2} e_2 + \dfrac{1}{2}e_3 - \dfrac{1}{2}e_4 + \cdots + \dfrac{(-1)^{n+1}}{2}e_{n} + \dfrac{(-1)^{n+1}}{2} e_{n+1}}^2\\
&= 1 + \dfrac{1}{4}n
\end{align*}
and choosing $\{\epsilon_i\}_{i = 1}^\infty$ to be $\epsilon_i = (-1)^{i+1}$ for all $i \in \mathbb{N}$ gives
\begin{align*}
\norm{\sum_{i=1}^\infty \epsilon_i\innerp{y_n}{\phi_i}\phi_i}^2 &= \norm{\sum_{i=1}^\infty (-1)^{i+1}\innerp{y_n}{\phi_i}\phi_i}^2
= \norm{\sum_{i = 1}^n (-1)^{i+1}\innerp{y}{\phi_i}\phi_i}^2
= \norm{\sum_{i = 1}^n \phi_i}^2\\
&= \norm{e_1 + \dfrac{3}{2} e_2 + \dfrac{3}{2}e_3 + \dfrac{3}{2}e_4 + \cdots + \dfrac{3}{2}e_{n} + \dfrac{1}{2} e_{n+1}}^2\\
&= \dfrac{5}{4} + \dfrac{9}{4}(n-1).
\end{align*}
Putting all of this together, we obtain for all $n \in \mathbb{N}$,
\begin{align*}
\norm{\sum_{i=1}^\infty \epsilon_i\innerp{y_n}{\phi_i}\phi_i}^2 = \dfrac{9n-4}{n+4} \norm{\sum_{i=1}^\infty \innerp{y_n}{\phi_i}\phi_i}^2
\end{align*}
and taking letting $n \to \infty$ gives $C_\varepsilon^2 = 9 = B/A$.
\end{eg}
\section{Frame Multipliers}\label{frmults}
The operators $S_\sigma$ and the sums $ \sum_{i\in I} a_i \innerp{\cdot}{\varphi_i} \varphi_i$ and $\sum_{i\in I} \epsilon_i \innerp{\cdot}{\varphi_i} \varphi_i $, can be considered as special cases of multipliers $M_{(m_i)_{i\in I}, (\phi_i)_{i\in I}, (\psi_i)_{i\in I}}$ defined by
\begin{equation*}
M_{(m_i)_{i\in I}, (\varphi_i)_{i\in I}, (\psi_i)_{i\in I}} x = \sum_{i\in I} m_i \innerp{x}{\psi_i} \varphi_i
\end{equation*}
for those $x$ for which the sum converges.
Gabor multipliers (see e.g. \cite{mult1}) are used in applications, in particular in signal processing, where they are used as a way to implement time-variant filters. Later on, multipliers for general Bessel sequences were introduced and investigated in \cite{mult2}; multipliers for general sequences, unconditional convergence, and invertibility of multipliers were investigated in \cite{mult4,mult3,mult6,mult5}.
In this language, Theorem \ref{bigthm} gives that for a spanning Bessel sequence $\Phi$,
\begin{equation*}
\norm{M_{(a_i)_{i \in I},(\phi_i)_{i \in I},(\phi_i)_{i \in I}}} \le 1 \quad \text{for all } \abs{a_i} \le 1
\end{equation*}
if and only if $\Phi$ is an orthogonal sum of tight frames.
\bigskip
\noindent {\bf Acknowledgment:} We wish to thank Jameson Cahill, Karlheinz Gr\"ochenig, Mark Lammers, and Adam Marcus for helpful discussions related to this paper. We are also deeply indebted to the referee for giving the paper considerable attention, resulting in a significant improvement.
|
\section{Introduction}
With the discovery of the Higgs boson at the CERN Large Hadron Collider (LHC) \cite{atlas:2012gk,cms:2012gu} particle physics has entered a new era. Both LHC collaborations, ATLAS and CMS, have confirmed the existence of a boson with a mass of about 126~GeV and properties consistent with those of the scalar CP-even particle predicted by the Standard Model \cite{Aad:2013xqa,Chatrchyan:2012jja}. In order to fully establish the nature of the Higgs boson, a precise determination of its couplings to fermions and gauge bosons is essential \cite{Zeppenfeld:2000td,Duhrssen:2004cv,LHCHiggsCrossSectionWorkingGroup:2012nn}.
A rather clean environment for such coupling measurements is provided by the vector-boson fusion (VBF) production mode \cite{Rainwater:1998kj,Plehn:1999xi,Rainwater:1999sd,Kauer:2000hi,Rainwater:1997dg,Eboli:2000ze}, where the Higgs boson is produced via quark-scattering mediated by weak gauge boson exchange in the $t$-channel, $qq'\to qq'H$. Because of the low virtuality of the exchanged weak bosons, the tagging jets emerging from the scattered quarks are typically located in the forward and backward regions of the detector, while the central-rapidity region exhibits little jet activity due to the color-singlet nature of the $t$-channel exchange. These features can be exploited to efficiently suppress QCD backgrounds with a priori large cross sections at the LHC.
In the context of central-jet veto (CJV) techniques, events are discarded if they exhibit one or more jets in between the two tagging jets. To quantitatively employ such selection strategies, a precise knowledge of the VBF cross section with an additional jet, i.e. the reaction $pp\to H jjj$, is crucial.
Next-to-leading order (NLO) QCD corrections to VBF-induced $Hjjj$ production have first been computed in \cite{Figy:2007kv}, yielding results with only small residual scale uncertainties of order 10\% or less. In particular, in that approach the survival probability for the Higgs signal has been estimated to exhibit a perturbative accuracy of about 1\%. The calculation of \cite{Figy:2007kv} is implemented in the {\tt VBFNLO} package~\cite{Arnold:2008rz,Arnold:2011wj,Arnold:2012xn} in the form of a flexible parton-level Monte-Carlo program.
More recently, an NLO-QCD calculation for electroweak $Hjjj$ production has been presented~\cite{Campanario:2013fsa}, where several approximations of Ref.~\cite{Figy:2007kv} have been dropped.
In this work, we merge the parton-level calculation of \cite{Figy:2007kv} with a parton-shower Monte-Carlo in the framework of the \POWHEG{} formalism~\cite{Nason:2004rx,Frixione:2007vw}, a method for the matching of an NLO-QCD calculation with a transverse-momentum ordered parton-shower program. For our implementation we are making use of version~2 of the \POWHEGBOX{}~\cite{Alioli:2010xd,Nason:2013ydw}, a repository that provides the process-independent ingredients of the \POWHEG{} method. The code we develop yields precise, yet realistic predictions for VBF-induced $Hjjj$ production at the LHC in a public framework that can easily be used by the reader for further phenomenological studies.
This article is organized as follows: In Sec.~\ref{sec:tech} we describe some technical details of our implementation. Phenomenological results are presented in Sec.~\ref{sec:pheno}. We conclude in Sec.~\ref{sec:concl}.
\section{Technical details of the implementation}
\label{sec:tech}
The implementation of $Hjjj$ production via VBF in the context of the \POWHEGBOX{} requires, as major building blocks, the matrix elements for all relevant partonic scattering processes at Born level and at next-to-leading order. These have first been calculated in \cite{Figy:2007kv} and are publicly available in the {\tt VBFNLO} package~\cite{Arnold:2008rz}. We extracted the matrix elements from {\tt VBFNLO} and adapted them to the format required by the \POWHEGBOX.
At leading order (LO), processes of the type $qq'\to qq'gH$ and all crossing-related channels are taken into account, if they include the exchange of a weak boson in the $t$-channel . Some representative Feynman diagrams are depicted in Fig.~\ref{fig:lo-graphs}.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{./figs/Higgs3j_LO-crop.ps}
\caption{Representative tree-level diagrams for VBF $Hjjj$ production. }
\label{fig:lo-graphs}
\end{figure}
The gauge-invariant class of diagrams involving weak-boson exchange in the $s$-channel is considered as part of the Higgs-strahlung process, and disregarded in the context of our work on VBF-induced Higgs production. The interference of $t$-channel with $u$-channel diagrams in flavor channels with quarks of the same type is neglected. Once VBF-specific selection cuts are imposed, these approximations are well justified~\cite{Ciccolini:2007ec}.
Throughout, we assume a diagonal CKM matrix.
We refer to the electroweak $Hjjj$ production process at order $\mathcal{O}(\alpha_s\alpha^3)$ within these approximations as ``VBF $Hjjj$ production''.
The virtual corrections to this reaction comprise the interference of the Born amplitudes with one-loop diagrams where a virtual gluon is attached to a single fermion line [c.f.~Fig.~\ref{fig:virt-graphs}~(a)--(g)], and diagrams where a virtual gluon is exchanged between the two different fermion lines, see~Fig.~\ref{fig:virt-graphs}~(h),(i).
\begin{figure}[tp]
\vspace*{8cm}
\includegraphics[width=\textwidth,bb=0 0 447 0]{./figs/Higgs3j_virt-crop.ps}
\\
\includegraphics[width=\textwidth]{./figs/Higgs3j_penta-crop.ps}
\caption{
Representative one-loop diagrams for $qq'\to qq'gH$, with the virtual gluon being attached to one single fermion line [graphs ($a$)-($g$)], or to the two different fermion lines [graphs ($h$) and ($i$)]. }
\label{fig:virt-graphs}
\end{figure}
As discussed in some detail in \cite{Figy:2007kv}, the
latter contributions are strongly suppressed by color factors and due to the VBF dynamics. They can be neglected, if the respective color structures of the real-emission contributions are disregarded as well, as these would serve to cancel the infrared singularities of the pentagon and hexagon contributions that we drop.
The real-emission contributions involve subprocesses with four external (anti-)quarks and two gluons such as $qq'\to qq'ggH $, as well as pure quark scattering processes of the type $qq'\to qq'Q\bar QH $, and all crossing-related channels with $t$-channel weak boson exchange.
Because of the approximations we have employed in the virtual contributions, where we dropped color-suppressed contributions giving rise to pentagon and hexagon integrals, the respective color structures have to be disregarded in the real-emission contributions as well. In practice this means to neglect interference terms between diagrams where a given gluon is emitted once from the upper and once from the lower quark line in the Feynman graphs. For example, interference terms like $2\,\mathcal{R}e\left( {\cal B}^3_4 {\cal B^*}^4_3\right)$, with ${\cal B}^3_4$ as depicted in Fig.~\ref{fig:real-gg-supp}, are dropped,
\begin{figure}[tp]
\includegraphics[width=1\textwidth]{./figs/Higgs3j_real3-crop.ps}
\caption{
Representative diagrams of the color structure $\mathcal{B}_4^3$ as introduced in Ref.~\cite{Figy:2007kv} for the subprocess $qq' \rightarrow qq'\, gg \, H$.
}
\label{fig:real-gg-supp}
\end{figure}
while their squares or the squares of the topologies sketched in Fig.~\ref{fig:real-gg} are fully considered.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=1\textwidth]{./figs/Higgs3j_real1-crop.ps} \\
\vspace{0.5cm}
\hspace{1cm} \includegraphics[width=0.66\textwidth]{./figs/Higgs3j_real2-crop.ps}
\caption{Representative diagrams of the color structure $\mathcal{A}_{43}^{1a}$ as introduced in Ref.~\cite{Figy:2007kv} for the subprocess $qq' \rightarrow qq'\, gg \, H$. }
\label{fig:real-gg}
\end{center}
\end{figure}
Representative diagrams for a pure quark subprocess are depicted in Fig.~\ref{fig:real-QQ}. For this class of subprocesses we require that the $Q\bar Q$ pair stems from a gluon. Contributions involving the hadronic decay of a weak boson, $V\to Q\bar Q$, such as graph~\ref{fig:real-QQ}~(c), are disregarded within our VBF~setup.
\begin{figure}[tp]
\centering
\includegraphics[width=1\textwidth]{./figs/Higgs3j_realQQ-crop.ps}
\caption{
Representative diagrams for the subprocess $qq' \to qq' \, Q\bar{Q}\, H$. }
\label{fig:real-QQ}
\end{figure}
While in \cite{Figy:2007kv} soft and collinear singularities have been taken care of by a dipole subtraction procedure, the \POWHEGBOX{} makes use of the so-called FKS subtraction scheme \cite{Frixione:1995ms}. From the color- and spin-correlated amplitudes provided by the user, the \POWHEGBOX{} internally constructs the counterterms that are needed to cancel soft and collinear singularities in the real-emission contributions. Because we are disregarding certain color-suppressed contributions in the virtual and real-emission amplitudes, we have to make sure that only the counterterms relevant for our setup are constructed. This is achieved by passing only those color- and spin-correlated Born amplitudes to the \POWHEGBOX{} that correspond to the color structures we consider within our approximations.
We have carefully tested that the counter terms constructed in this way approach the real-emission amplitudes in the soft and collinear limits. Additionally, we have compared the tree-level and real emission amplitudes for selected phase space points with code generated by {\tt MadGraph} \cite{Stelzer:1994ta} that has been adapted to match the approximations of our calculation. We found agreement at the level of more than ten digits. The virtual amplitudes have been compared to {\tt VBFNLO}, again showing full agreement for single phase space points. We note that some care has been necessary in this latter check, as finite parts of the subtraction terms are included in the virtual amplitudes in the default setup of {\tt VBFNLO}.
To verify the entire setup of our code, we have compared cross sections and distributions for various sets of selection cuts at LO and NLO-QCD accuracy as obtained with the \POWHEGBOX{} with respective results of {\tt VBFNLO}. We found full agreement for all considered scenarios.
We note that special care is needed when performing the phase-space integration of VBF $Hjjj$\;~production in the framework of the \POWHEGBOX{}. In contrast to the VBF-induced $Hjj$ production cross section that is entirely finite at leading order, the inclusive VBF $Hjjj$ cross section diverges already at leading order when a pair of partons becomes collinear or a soft gluon is encountered in the final state. While such divergent contributions disappear after phenomenologically sensible selection cuts are imposed, their presence considerably reduces the efficiency of the numerical phase space integration. This effect can be avoided by appropriate phase-space cuts at generation level, or by a so-called Born-suppression factor $F(\Phi_n)$ that dampens the integrand whenever singular configurations in phase-space are approached. In order to
ensure that our results are independent of technical cuts in the phase-space integration, we recommend the use of a Born-suppression factor. In our \POWHEGBOX{} implementation we provide two alternative versions of Born-suppression factors:
\begin{itemize}
\item
In our first, multiplicative, approach, the factor is of the form
\begin{equation}
\label{eq:bsupp1}
F(\Phi_n) =
\prod_{i=1}^3
\left(
\frac{p_{T,i}^2}{p_{T,i}^2+\Lambda_p^2}
\right)^2
\prod_{i,j=1;\\
j\neq i}^3
\left(
\frac{m_{ij}^2}{m_{ij}^2+\Lambda_m^2}
\right)^2\,,
\end{equation}
where the $p_{T,i}$ and $m_{ij}=\sqrt{(p_i+p_j)^2}$ respectively denote the transverse momenta and invariant masses of the three final-state partons of the underlying Born configuration. The $\Lambda_p$ and $\Lambda_m$ are cutoff parameters that are typically set to values of a few GeV. %
\item
Following the procedure suggested for the related case of trijet production in the framework of the \POWHEGBOX{}~\cite{Kardos:2014dua}, we use an exponential suppression factor of the form
\begin{equation}
S_1 = \exp \left[
- \Lambda_1^4 \cdot \left(\sum_{i=1}^3 \frac{1}{p^4_{T,i}} + \sum_{i,j=1;\\
j\neq i}^3 \frac{1}{q^2_{ij}}\right)
\right]\,,
\end{equation}
with
\begin{equation}
q_{ij}=p_i \cdot p_j \, \frac{E_i \, E_j}{E_i^2 + E_j^2}\,,
\end{equation}
for the suppression of infrared divergent configurations in the underlying Born kinematics,
accompanied by a factor
\begin{equation}
S_2 = \left(\frac{H_T^2}{H_T^2 + \Lambda_2^2}\right)^2,
\end{equation}
where
\begin{equation}
H_T = p_{T,1} + p_{T,2} + p_{T,3}.
\end{equation}
The factor $S_2$ serves to suppress configurations where all partons are having small transverse momenta, and at the same time increase the fraction of events generated with large transverse momenta.
Combining $S_1$ with $S_2$, we construct
\begin{equation}
\label{eq:bsupp2}
F(\Phi_n) = S_1 \cdot S_2\,.
\end{equation}
\end{itemize}
For the generation of the phenomenological results shown below we are using a Born suppression factor of the form given in Eq.~(\ref{eq:bsupp2}) with $\Lambda_1 = 10 $~GeV and $\Lambda_2 = 30$~GeV, supplemented by a small generation cut on the transverse momenta of the three outgoing partons of the underlying Born configuration, $p_{T,i}^\mathrm{gen} > 1$~GeV.
To make sure our results do not depend on these technical parameters, in addition to our default setup we ran our code using the Born suppression factor of Eq.~(\ref{eq:bsupp1}) with $\Lambda=20$~GeV and, again, $p_{T,i}^\mathrm{gen} > 1$~GeV. The results in the two setups are in full agreement with each other and, at fixed order, also with respective results obtained with {\tt VBFNLO} that is using an entirely different phase-space generator.
\section{Phenomenological results}
\label{sec:pheno}
Our implementation of VBF $Hjjj$\; production at the LHC is made publicly available in version~2 of the \POWHEGBOX{}, and can be obtained as explained at the project webpage, {\tt http://powhegbox.mib.infn.it/}.
Here, we are providing phenomenological results for a representative setup at the LHC with a center-of-mass energy of $\sqrt{s}=8$~TeV.
We are using the CT10 fixed-four-flavor set \cite{Lai:2010vv} for the parton distribution functions of the proton as implemented in the {\tt LHAPDF} library \cite{Whalley:2005nh} and the accompanying value of the strong coupling, $\alpha_s(m_Z)=0.1127$. Jets are reconstructed via the anti-$k_T$ algorithm with a resolution parameter of $R=0.5$, with the help of the {\tt FASTJET}~package~\cite{Cacciari:2005hq,Cacciari:2008gp,Cacciari:2011ma}. As electroweak input parameters we are using the masses of the weak gauge bosons, $m_W=80.398$~GeV and $m_Z=91.188$~GeV, and the Fermi constant, $G_F=1.16639\times 10^{-5}$~GeV$^{-1}$. Other electroweak parameters are computed thereof via tree-level relations.
The widths of the massive gauge bosons are set to $\Gamma_W = 2.095$~GeV and $\Gamma_Z=2.51$~GeV, respectively. For the Higgs boson, we are using $m_H = 126$~GeV and $\Gamma_H=4.095$~MeV.
The renormalization and factorization scales are identified as $\mu_\mr{R}=\mu_\mr{F}=m_H/2$.
In order to assess uncertainties that remain after matching the NLO calculation with a parton shower program, we consider three different tools: \PYTHIA{}~{\tt 6.4.25} with the Perugia~0 tune~\cite{Sjostrand:2006za}, \HERWIGPP{}~{\tt 2.7.0}~\cite{Bahr:2008pv,Bellm:2013lba} with its default angular-ordered shower, and with a transverse-momentum ordered dipole shower~\cite{Platzer:2011bc} which we tag as {\tt PYT}, {\tt HER}, and {\tt DS++}, respectively. We note that wide-angle, soft radiation that is in principle needed when matching an NLO calculation with a parton-shower program using the \POWHEG{} method, is missing in the default angular-ordered \HERWIGPP{} shower. The impact of this missing piece on observables can only be estimated by a comparison with predictions obtained with transverse momentum ordered showers, such as the \DSPP{} version of \HERWIGPP{}. We do not consider hadronization, QED radiation, multiple parton interactions, and underlying event effects in this work.
In order to define a $Hjjj$ event, we demand at least three well-observable jets with
\begin{equation}
p_{T,j}>20 \text{ GeV}\,,
\quad
|y_j| < 4.5\,.
\end{equation}
In addition, we impose VBF-specific selection cuts. The two hardest jets, referred to as ``tagging jets'', are required to fulfill
\begin{equation}
p_{T,j}^\mathrm{tag}>30~\mathrm{GeV}\,,\quad
|y_j^\mathrm{tag}|<4.5\,,
\end{equation}
and be well-separated from each other,
\begin{equation}
|y_{j_1}^\mathrm{tag}-y_{j_2}^\mathrm{tag}|>4.0\,,
\quad
y_{j_1}^\mathrm{tag}\times y_{j_2}^\mathrm{tag}<0\,,
\quad
m_{jj}^\mathrm{tag}>500~\mathrm{GeV}\,.
\end{equation}
The kinematics of the Higgs boson is not restricted.
With these cuts, we obtain a cross section of $\sigma^\mathrm{NLO}=71.5 \pm 0.4$~fb at fixed order, where the error is the statistical error of the Monte Carlo calculation. After matching the NLO result with a parton shower, some of the events fail to pass the cuts, resulting in slightly smaller cross sections of $\sigma^{\tt PYT}=65.8 \pm 0.3$~fb, $\sigma^{\tt HER}=68.3 \pm 0.3$~fb, and $\sigma^{\tt DS++}=69.8\pm 0.5$~fb, respectively.
Apart from this change in normalization the impact of the parton shower on observables related to the tagging jets is very mild, as illustrated in Fig.~\ref{fig:tag-jet}
\begin{figure}[t]
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/ptj1.eps}
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/mjj.eps}
\caption{
Transverse momentum distribution of the hardest tagging jet (left) and invariant mass distribution of the two tagging jets (right) at NLO (black), and at NLO+PS level: \PYT{}~(red), \HERPP{}~(blue), \DSPP{}~(cyan).
The lower panels show the NLO+PS results normalized to the pure NLO prediction together with its statistical uncertainty (yellow band).
}
\label{fig:tag-jet}
\end{figure}
for the transverse momentum distribution of the hardest tagging jet and the invariant mass of the tagging jet pair.
In contrast to NLO calculations for VBF $Hjj$ production, where the third jet can be described only with LO accuracy, our calculation is NLO accurate in distributions related to the third jet. In Fig.~\ref{fig:jet3}, NLO+PS results for the transverse momentum and the rapidity distribution of the third jet are shown for different parton shower programs together with the fixed-order NLO result.
\begin{figure}[t]
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/ptj3.eps}
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/etaj3.eps}
\caption{
Transverse momentum and rapidity distributions of the third jet at NLO, and at NLO+PS level (line styles as in Fig.~\ref{fig:tag-jet}).
}
\label{fig:jet3}
\end{figure}
For all considered parton showers, the difference between the NLO and the NLO+PS results is small. However, \PYTHIA{} tends to produce slightly more jets in the central-rapidity region of the detector, while \HERWIGPP{} preferentially radiates in the collinear region between the two tagging jets and the beam axis. We will see below that this effect is more pronounced in the case of sub-leading jets.
Larger differences between the fixed-order and the various matched predictions occur in distributions related to the fourth jet.
In the parton-level NLO calculation a fourth jet can only stem from the real-emission contributions, and can thus be described only at tree-level accuracy. Larger theoretical uncertainties are therefore expected for observables related to the fourth jet. Fig.~\ref{fig:jet4}
\begin{figure}[t]
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/ptj4.eps}
\includegraphics[angle=-90,width=0.5\textwidth]{./plots/y4star.eps}
\caption{
Transverse momentum distribution of a fourth jet for our default setup with an extra cut of $p_{T,j_4}>1$~GeV~(left) and rapidity distribution of a fourth hard jet with $p_{T,j_4}>20$~GeV relative to the two tagging jets (right) at NLO, and at NLO+PS level (line styles as in Fig.~\ref{fig:tag-jet}).
}
\label{fig:jet4}
\end{figure}
illustrates the effect of the \POWHEG{}-Sudakov factor on the transverse momentum of the fourth jet and clarifies how extra radiation in the VBF setup is distributed by the different parton shower programs via the $y_4^\star$ variable. This quantity is defined as
\begin{equation}
y_4^\star=y_{j_4} - \frac{y_{j_1} + y_{j_2}}{2}\,,
\label{eq:y4star}
\end{equation}
in order to parameterize the rapidity of the fourth jet relative to the two hard tagging jets. The respective distribution shows, more pronouncedly than in the case of the third jet, that \PYTHIA{} and \HERWIGPP{} tend to produce radiation in different regions of phase space. The differences between the various NLO+PS curves can thus be considered as inherent uncertainty of the matched prediction.
\section{Conclusions}
\label{sec:concl}
In this work, we have presented an implementation of VBF $Hjjj$\; production in version~2 of the \POWHEGBOX{} repository. We have performed the matching of an existing NLO-QCD calculation with parton-shower programs using the \POWHEG{} formalism and presented phenomenological results for a representative setup at the LHC. The code we developed is publicly available and can be adapted to the user's need in a straightforward manner.
We have shown that theoretical uncertainties associated with the description of the third jet by genuinely different parton-shower programs are mild at NLO+PS level, contrary to what is observed in studies based on matrix elements for VBF $Hjj$ production that are only LO accurate in the third jet. Our implementation thus provides an important improvement in the theoretical assessment of central-jet veto observables that are crucial for VBF analyses at the LHC.
\section*{Acknowledgments}
We are grateful to Carlo Oleari for help with implementing the code in the \POWHEGBOX{} repository.
The work of
B.~J.\ is supported by the Institutional Strategy of the University of T\"ubingen (DFG, ZUK~63).
F.~S.\ is supported by the ``Karlsruher Schule f\"ur Elementar\-teilchen- und Astroteilchenphysik: Wissenschaft und Technologie (KSETA)'' and D.~Z.\ by the BMBF under ``Verbundprojekt 05H2012 -- Theorie''.
|
\section{Introduction}
\label{sec:int}
The measurement of beauty and charm production in $ep$ collisions at
HERA is an important testing ground for perturbative Quantum
Chromodynamics (pQCD), since the heavy-quark masses provide a hard
scale that allows perturbative calculations to be made.
At leading
order, the dominant process for heavy-quark production at HERA is
boson-gluon fusion (BGF).
In this process, a virtual photon emitted by
the incoming electron interacts with a gluon from the proton
forming a heavy quark--antiquark pair.
When the negative squared four-momentum of the virtual photon, \ensuremath{Q^{2}}\xspace, is
large compared to the proton mass, the interaction is referred to as
deep inelastic scattering (DIS). For heavy-quark transverse momenta
comparable to the quark mass, next-to-leading-order (NLO) QCD
calculations based on the dynamical generation of the massive
quarks~\cite{np:b452:109, pl:b353:535, pl:b359:423, np:b392:162, np:b392:229} are
expected to provide reliable predictions.
Beauty and charm production in DIS has been measured using several
methods by the
H1~\cite{zfp:c72:593, np:b545:21, pl:b528:199, epj:c38:447, epj:c40:349, epj:c41:453, epj:c45:23, epj:c51:271, Aaron:2009jy, Aaron:2009ut, Aaron:2010ib, Aaron:2011gp, Aaron:2011gp2}
and
ZEUS~\cite{pl:b407:402, epj:c12:35, pr:d69:012004, pl:b599:173, epj:c50:1434, pl:b649:111, Chekanov:2007ch, Chekanov:2008zz, epj:c63:2009:2:171-188, Chekanov:2009kj, Abramowicz:2010zq, Abramowicz:2010aa, Abramowicz:2011rs, jhep05.2013.023, jhep05.2013.097}
collaborations.
All but the two most recent measurements of charm production~\cite{jhep05.2013.023, jhep05.2013.097} and older data~\cite{pl:b649:111} have been
combined~\cite{epj.c73.2311}.
Predictions from NLO QCD describe all results reasonably well.
Inclusive jet cross sections in beauty and charm events are used in the analysis presented here to extract
the heavy-quark contribution to the proton structure function $F_{2}$
with high precision, and to measure the $b$-quark mass.
For this purpose, the long lifetimes of the
weakly decaying $b$ and $c$ hadrons,
which make the reconstruction of their decay vertices possible,
as well as their large masses were exploited.
Two
discriminating variables, the significance of the reconstructed decay
length and the invariant mass of the charged tracks associated with
the decay vertex (secondary vertex), were used.
This inclusive tagging method leads to a substantial increase in statistics with
respect to previous ZEUS measurements.
Differential cross sections as a function of \ensuremath{Q^{2}}\xspace,
the Bjorken scaling variable, $x$, jet transverse energy, \ensuremath{E_{T}^{\text{jet}}}\xspace, and
pseudorapidity, \ensuremath{\eta^{\text{jet}}}\xspace, were measured. They are compared to a
leading-order (LO) plus parton-shower (PS) Monte Carlo prediction and
to NLO QCD calculations. The beauty and charm contributions to the
proton structure function $F_{2}$, denoted as \ensuremath{F_{2}^{b\bar{b}}}\xspace and \ensuremath{F_{2}^{c\bar{c}}}\xspace,
respectively, as well as beauty and charm reduced cross sections
(\ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace and \ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace, respectively) were extracted from the
double-differential cross section as a function of \ensuremath{Q^{2}}\xspace and $x$. The
results are compared to previous measurements and to predictions from
perturbative QCD.
The running \hbox{$\overline{\rm MS}$}\xspace beauty-quark mass, $m_{b}$ at the scale $m_{b}$, denoted $m_b(m_b)$,
is measured using \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace,
following a procedure similar to that used for a recent extraction
of the charm-quark mass~\cite{epj.c73.2311}.
This represents the
first measurement of the $b$-quark mass using HERA or any other hadron collider
data.
\section{Experimental set-up}
\label{sec:exp}
This analysis was performed with data taken with the ZEUS detector
from 2004 to 2007, when HERA collided electrons\footnote{%
In this paper ``electron'' is used to denote both electron and positron.}
with energy $E_e=\SI{27.5}{\GeV}$ with protons of energy
\SI{920}{\GeV}, corresponding to a centre-of-mass energy $\sqrt{s} =
\SI{318}{\GeV}$. This data-taking period is denoted as HERA~II. The
corresponding integrated luminosity is \SI{354 \pm 7}{\per\pb}.
\Zdetdesc
\Zctdmvddesc{\ZcoosysfnBEeta}
\Zcaldesc
\Zlumidesc
\section{Monte Carlo simulations}
\label{sec:montecarlo}
To evaluate the detector acceptance and to provide predictions of
the signal and background distributions, Monte Carlo (MC) samples of
beauty, charm and light-flavour events were generated, corresponding
to eighteen, three and one times the integrated luminosity of the
data, respectively. The \textsc{Rapgap}\xspace~3.00 MC
program~\cite{cpc:86:147} in the massive mode ($m_{b} = \SI{4.75}{\GeV},
m_{c} = \SI{1.5}{\GeV}$) was used to generate the beauty and charm samples,
where the CTEQ5L~\cite{epj:c12:375} parameterisation for the proton
parton density functions (PDFs) was used. In \textsc{Rapgap}\xspace, LO matrix
elements are combined with higher-order QCD radiation simulated in the
leading-logarithmic approximation. Higher-order QED effects are
included through \textsc{Heracles}\xspace~4.6~\cite{cpc:69:155}.
Light-flavour MC
events were extracted from an inclusive DIS sample generated with
\textsc{Djangoh}\xspace~1.6~\cite{proc:hera:1991:1419} interfaced to \textsc{Ariadne}\xspace{}
4.12~\cite{cpc:71:15}.
The CTEQ5D~\cite{epj:c12:375} PDFs were used and quarks were taken to
be massless.
Fragmentation and particle decays were simulated using the
\textsc{Jetset}\xspace{}/\textsc{Pythia}\xspace{} model~\cite{cpc:82:74, cpc:135:238}.
The Bowler parameterisation~\cite{zfp:c11:169} of the fragmentation
function, as implemented in \textsc{Pythia}\xspace~\cite{hep-ph-0108264} (with $r_{Q} = 1$), was used for the heavy-flavour samples.
The generated
events were passed through a full simulation of the ZEUS detector
based on \textsc{Geant 3.21}~\cite{tech:cern-dd-ee-84-1}. The final
MC events were then subjected to the same trigger requirements and
processed by the same reconstruction program as the data.
For the acceptance determination, the \ensuremath{E_{T}^{\text{jet}}}\xspace and \ensuremath{\eta^{\text{jet}}}\xspace distributions
in the charm MC, as well as the \ensuremath{Q^{2}}\xspace distributions in both the beauty and charm
MCs, were reweighted in order to give a good description of the data.
The charm branching fractions and fragmentation fractions were adjusted to the world-average
values~\cite{Nakamura:2010zzi,arXiv:1112.3757,thesis:viazlo:2012}.
\section{Theoretical predictions and uncertainties}
\label{sec:theory}
Next-to-leading-order QCD predictions for differential cross sections
were obtained from the \textsc{Hvqdis}\xspace program~\cite{pr:d57:2806}.
The calculations were used to extrapolate the visible cross
sections to extract \ensuremath{F_{2}^{b\bar{b}}}\xspace, \ensuremath{F_{2}^{c\bar{c}}}\xspace, \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace and \ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace (see
Section~\ref{sec:F2q}).
The calculations are based on the
fixed-flavour-number scheme (FFNS) in which only light flavours are
present in the proton and heavy quarks are produced in the
interaction~\cite{np:b374:36}.
Therefore, the 3-flavour (4-flavour) FFNS variant of the ZEUS-S NLO QCD
fit~\cite{pr:d67:012007} was used for the proton PDF for the predictions of the
charm (beauty) cross sections.
As in the PDF fit, the value of
$\ensuremath{\alpha_{s}}\xspace(M_{Z})$ was set to 0.118 and the heavy-quark masses (pole
masses) were set to $m_{b} = \SI{4.75}{\GeV}$ and $m_{c} = \SI{1.5}{\GeV}$. The
renormalisation and factorisation scales, $\mu_{R}$ and $\mu_{F}$,
were chosen to be equal and set to
$\mu_{R}=\mu_{F}=\sqrt{\ensuremath{Q^{2}}\xspace+4m_{b(c)}^{2}}$.
The systematic uncertainty on the theoretical predictions with the
ZEUS-S PDFs were estimated by varying the quark masses and the
renormalisation and factorisation scales. Quark masses of $m_{b} = 4.5$
and \SI{5.0}{\GeV}, $m_{c} = 1.3$ and \SI{1.7}{\GeV} were used. The
scales $\mu_{R}$, $\mu_{F}$ were varied independently by a factor of
two up and down. Additionally, the experimental uncertainties of the
data used in the PDF fit were propagated to the predicted cross
sections. The total uncertainties were obtained by adding positive
and negative changes to the cross sections in quadrature. This
results in total uncertainties of \SIrange{10}{20}{\%} for beauty and
\SIrange{10}{50}{\%} for charm.
Predictions were also obtained using the 3- and 4-flavour variants of the
ABKM NLO PDFs~\cite{PhysRevD.81.014032} for the proton.
The pole masses of heavy quarks were set to $m_{b} = \SI{4.5}{\GeV}$ and
$m_{c} = \SI{1.5}{\GeV}$, both in the PDF fit and in the \textsc{Hvqdis}\xspace calculation.
The values of $\ensuremath{\alpha_{s}}\xspace(\mu_{R})$ were provided by
LHAPDF~\cite{Whalley:2005nh,LHAPDF} to
ensure that the same function was used as in the PDF fit. The
renormalisation and factorisation scales were both set to
$\mu_{R}=\mu_{F}=\sqrt{\ensuremath{Q^{2}}\xspace+4m_{b(c)}^{2}}$.
The NLO QCD predictions are given for parton-level jets. These
were reconstructed using the $k_{T}$ clustering
algorithm~\cite{pr:d48:3160} with a radius parameter $R = 1.0$ in the longitudinally invariant
mode~\cite{np:b406:187}. The $E$-recombination scheme, which produces
massive jets whose four-momenta are the sum of the four-momenta of the
clustered objects, was used.
The parton-level cross sections
were corrected for jet hadronisation effects to allow a
direct comparison with the measured hadron-level cross sections:
\begin{equation}
\sigma^{\text{had, NLO}} = \ensuremath{C^{\text{had}}}\xspace \sigma^{\text{parton,NLO}}\,,
\label{eq:chad}
\end{equation}
where the correction factors, $\ensuremath{C^{\text{had}}}\xspace = 1 + \ensuremath{\Delta^{\text{had}}}\xspace$, were derived from the \textsc{Rapgap}\xspace MC
simulation. The factors \ensuremath{C^{\text{had}}}\xspace are defined as the ratio of the
hadron-level jet to the parton-level jet cross sections, and the parton
level is defined as the result of the parton-showering stage of the
simulation.
Since \ensuremath{C^{\text{had}}}\xspace were derived from an LO plus parton shower MC,
but are applied to an NLO prediction,
the uncertainty on \ensuremath{C^{\text{had}}}\xspace cannot be estimated in a straightforward
way.
Within the framework of parton showering, MC subsets with different numbers of
radiated partons were investigated using \textsc{Rapgap}\xspace and \textsc{Pythia}\xspace samples.
These studies indicated that different approaches yield variations of
\ensuremath{\Delta^{\text{had}}}\xspace of typically a factor of two.
Since it is not clear if the variations can be interpreted as uncertainties
on \ensuremath{C^{\text{had}}}\xspace, no such uncertainties were included in the cross-section ($F_{2}$) predictions.
However, for the extraction of the $b$-quark mass,
such a theoretical uncertainty needs to be included.
\section{Data selection}
\label{sec:data}
Events containing a scattered electron were selected online by means
of a three-level trigger
system~\cite{zeus:1993:bluebook, proc:chep:1992:222}.
The trigger~\cite{thesis:roloff:2011} did not require the presence of a secondary vertex nor of a jet.
Offline, the scattered electron was reconstructed using an electron finder based on a neural network~\cite{nim:a365:508}. The hadronic system
was reconstructed from energy-flow objects
(EFOs)~\cite{epj:c1:81, thesis:briskin:1998} which combine the
information from calorimetry and tracking, corrected for energy
loss in the detector material.
The kinematic variables used in the cross-section measurements, \ensuremath{Q^{2}}\xspace
and $x$, were reconstructed using the double-angle method~\cite{proc:hera:1991:23}.
The following cuts were applied to select a clean DIS sample:
\begin{itemize}
\item the reconstructed
scattered electron~\cite{nim:a365:508, nim:a391:360} was required to
have an energy $E'_{e} > \SI{10}{\GeV}$;
\item the impact position of the scattered electron on the face of the
RCAL had to be outside the region \SI[parse-numbers=false]{26 \times 26}{\cm^{2}} centred on
$X = Y = 0$;
\item the primary vertex had to be within
\SI{\pm 30}{\cm} in $Z$ of the nominal interaction point;
\item the photon virtuality, \ensuremath{Q^{2}}\xspace, had to be within $5 < \ensuremath{Q^{2}}\xspace <
\SI{1000}{\GeV^{2}}$;
\item $ y_{\text{JB}} > 0.02$, where $y_{\text{JB}}$ is the
inelasticity reconstructed using the Jacquet-Blondel
method~\cite{proc:epfacility:1979:391};
\item $ y_{e} < 0.7$, where $y_{e}$ is the inelasticity reconstructed
using the electron method~\cite{proc:hera:1991:23};
\item $44 < (E-p_{Z}) < \SI{65}{\GeV}$,
where $(E-p_{Z}) = \sum_{i} (E_{i} - p_{Z,i})$ and $i$ runs over all final-state particles
with energy $E_{i}$ and
$Z$-component of momentum $p_{{Z,i}}$;
this selects fully
contained neutral-current $ep$ events for which
$E-p_{Z}=2 E_{e}$.
\end{itemize}
Jets were reconstructed from EFOs using the $k_{T}$ clustering
algorithm~\cite{pr:d48:3160} as was described for parton-level jets
in Section~\ref{sec:theory}.
Jets containing the identified scattered electron were not considered further.
Events were selected if
they contained at least one jet within the pseudorapidity range $-1.6
< \ensuremath{\eta^{\text{jet}}}\xspace < 2.2$ and with transverse energy, \ensuremath{E_{T}^{\text{jet}}}\xspace,
of
\begin{equation*}
\ensuremath{E_{T}^{\text{jet}}}\xspace = \ensuremath{p_{T}^{\text{jet}}}\xspace\frac{E^{\text{jet}}}{p^{\text{jet}}} > \SI[parse-numbers=false]{5\,(4.2)}{\GeV}
\end{equation*}
for beauty (charm), where $E^{\text{jet}}$, $p^{\text{jet}}$ and
\ensuremath{p_{T}^{\text{jet}}}\xspace are the jet energy, momentum and transverse momentum. The cut
on \ensuremath{E_{T}^{\text{jet}}}\xspace was optimised separately for beauty and charm measurements.
For beauty, a cut of $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{5}{\GeV}$ ensures a good correlation between momentum and angle for
reconstructed and hadron-level jets~\cite{thesis:kahle:2005}; for charm this cut was
$\SI{4.2}{\GeV}$ to reduce the extrapolation uncertainties for the
\ensuremath{F_{2}^{c\bar{c}}}\xspace and \ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace measurements at low \ensuremath{Q^{2}}\xspace.
In order to reconstruct potential secondary vertices related to $b$- and
$c$-hadron decays, tracks were selected if:
\begin{itemize}
\item they had a transverse momentum $\ensuremath{p_{T}}\xspace > \SI{0.5}{\GeV}$;
\item the total number of hits\footnote{Each MVD
layer provided two coordinate measurements.} on the track in the MVD was $\ge 4$.
\item if the track was inside the CTD acceptance, track recognition in the CTD was required;
the percentage of the tracks outside the CTD acceptance, and hence reconstructed using MVD hits only, was \SI{2.5}{\percent}.
\end{itemize}
Tracks were associated with the closest jet if they fulfilled the
criterion $\Delta R < 1$ with $\Delta R =
\sqrt{(\eta^{\text{trk}}-\eta^{\text{jet}})^{2}+(\phi^{\text{trk}}-\phi^{\text{jet}})^{2}}
$. If two or more of such tracks were associated with the jet, a
candidate vertex was fitted from the selected tracks using a
deterministic annealing
filter~\cite{RosePhysRevLett.65.945, Rose726788, Didierjean2010188}.
This fit provided the vertex position and its error matrix as
well as the invariant mass, \ensuremath{m_{\text{vtx}}}\xspace, of the charged tracks associated
with the reconstructed vertex. The charged-pion mass was assumed for all tracks
when calculating the vertex mass.
Vertices with $\chi^{2}/\ensuremath{\text{ndf}}\xspace < 6$, a
distance from the interaction point within \SI{\pm 1}{\cm} in the
$X$--$Y$ plane, \SI{\pm 30}{\cm} in the $Z$ direction, and
$\num{1} < \ensuremath{m_{\text{vtx}}}\xspace < \SI{6}{\GeV}$ were kept for further
analysis.
The MC gives a good description of the track efficiencies, except for a
small fraction of tracks that are affected by hadronic interactions
in the detector material between the interaction point and the CTD.
Efficiency corrections for
this effect were determined from a study of exclusive $ep\to e\rho^0p$
events~\cite{thesis:libov:2012}, using a special track reconstruction.
The number of the pions
from the $\rho^0$ decay that were reconstructed in the MVD alone and had no
extension in the CTD was measured.
The resulting track efficiency correction in the MC was applied by randomly rejecting
selected vertex tracks before the vertex fit, with a probability
that depends on the track parameters (around \SI{3}{\%} at $\eta=0$
and $\ensuremath{p_{T}}\xspace = \SI{1}{\GeV}$).
\section{Extraction of the heavy-flavour cross sections}
\label{sec:bc-extr}
Using the secondary-vertex candidates associated with jets, the decay length,
$d$, was defined as the vector in $X$--$Y$ between the secondary
vertex and the interaction point\footnote{%
In the $X$--$Y$ plane, the interaction point was defined as the
centre of the beam ellipse, determined using the average primary
vertex position for groups of a few thousand events, taking into
account the difference in angle between the beam direction and the
$Z$ direction. The $Z$ coordinate was taken as the $Z$ position of
the primary vertex of the event.}
projected onto the jet axis in
the $X$--$Y$ plane. The sign of the decay length was assigned using
the axis of the jet to which the vertex was associated; if the decay-length vector
was in the same hemisphere as the jet axis, a
positive sign was assigned to it, otherwise the sign of the decay
length was negative. Negative decay lengths, which originate from
secondary vertices reconstructed on the wrong side of the interaction
point with respect to the direction of the associated jets, are
unphysical and caused by detector resolution effects. A small smearing
correction~\cite{thesis:libov:2012} to the MC decay-length
distribution was applied in order to reproduce the data with negative
values of decay length.
The beauty and charm content in the selected sample was determined
using the shape of the decay-length significance distribution together
with the secondary-vertex mass distribution, \ensuremath{m_{\text{vtx}}}\xspace. The
decay-length significance, $S$, is defined as $d/\delta{d}$, where
$\delta{d}$ is the uncertainty on $d$. The invariant mass of the
tracks fitted to the secondary vertex provides a distinguishing
variable for jets from $b$ and $c$ quarks, reflecting the different
masses of the $b$ and $c$ hadrons. Figure~\ref{fig:significance} shows
the decay-length significance, $S$, for $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{4.2}{\GeV}$
divided into four bins: $1 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{1.4}{\GeV}$, $1.4 <
\ensuremath{m_{\text{vtx}}}\xspace < \SI{2}{\GeV}$, $2 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{6}{\GeV}$ and
no restriction on \ensuremath{m_{\text{vtx}}}\xspace. The MC
simulation provides a good description of the data.
The separation into subsamples is described below.
The contents of the negative bins of the significance distribution,
$N(S^{-})$, were subtracted from the contents of the corresponding
positive bins, $N(S^{+})$, yielding a subtracted decay-length
significance distribution. In this way, the contribution from
light-flavour quarks is minimised.
An additional advantage of this subtraction is that symmetric systematic
effects, which might arise from discrepancies between the data and the MC, are
removed.
In order to reduce the
contamination of tracks originating from the primary vertex, a cut of
$|S| >4$ was applied.
To extract the contributions from beauty, charm and light flavours in
the data sample, a binned $\chi^{2}$ fit of
the subtracted significance distribution in the region $4 < |S| < 20$ was performed simultaneously
for three mass bins~\cite{thesis:roloff:2011}:
$1 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{1.4}{\GeV}$; $1.4 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{2}{\GeV}$;
$2 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{6}{\GeV}$.
All MC distributions were normalised to the
integrated luminosity of the data before the fit. The overall MC
normalisation was constrained by requiring it to be consistent with
the normalisation of the data in the significance distribution with
$|S| < 20$ and $1 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{6}{\GeV}$. The fit yielded scaling
factors $k_{b}$, $k_{c}$ and $k_{\text{lf}}$ for the beauty, charm and
light-flavour contributions, respectively, to obtain the best
description of the data.
The correlation coefficients were as follows: $\rho_{b,c}=-0.68(-0.67)$,
$\rho_{b,\text{lf}}=0.58(0.57)$ and $\rho_{c,\text{lf}}=-0.98(-0.98)$ for
$\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI[parse-numbers=false]{4.2(5.0)}{\GeV}$.
The subtracted and fitted distributions
for $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{4.2}{\GeV}$ are shown in
Fig.~\ref{fig:subtracted_sig}.
A good agreement between data and MC is observed.
The first
two mass bins corresponding to the region $1 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{2}{\GeV}$
are dominated by charm events. In the third mass bin,
beauty events are dominant at high values of significance.
The fit procedure was
repeated for every bin of a given observable to obtain differential
cross sections. For the beauty cross-section extraction, the fit
procedure was repeated with the higher cut on \ensuremath{E_{T}^{\text{jet}}}\xspace, $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{5}{\GeV}$.
Control distributions of \ensuremath{E_{T}^{\text{jet}}}\xspace, \ensuremath{\eta^{\text{jet}}}\xspace,
$\log_{10}\ensuremath{Q^{2}}\xspace$ and
$\log_{10}x$ are shown in Fig.~\ref{fig:beauty_enriched} after beauty
enrichment cuts ($2 < \ensuremath{m_{\text{vtx}}}\xspace < \SI{6}{\GeV}$ and $|S| > 8$) for
$\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{5.0}{\GeV}$ and in
Fig.~\ref{fig:charm_enriched} after charm enrichment cuts ($1 <
\ensuremath{m_{\text{vtx}}}\xspace < \SI{2}{\GeV}$ and $|S| > 4$) for $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI{4.2}{\GeV}$. All
data distributions are reasonably well
described by the MC.
The differential cross sections for jet production in beauty or charm events,
$q = b,c$, corrected to QED Born level, in a bin
$i$ of a given observable, $Y$, are given by:
\begin{equation}
\label{eq:xsec4}
\frac{\text{d} \sigma_{q}^{\text{jet}}}{\text{d} Y_{i}} =
k_{q}(Y_{i}) \frac{N_{q}^{\text{had,MC}}(Y_{i})}{%
\mathcal{L} \cdot \Delta Y_{i}} \frac{1}{\ensuremath{C^{\text{rad}}}\xspace},
\end{equation}
where $\Delta Y_{i}$ is the width of the bin, $k_{q}$ denotes the
scaling factor obtained from the fit,
$N_{q}^{\text{had,MC}}$ is the number of generated jets in beauty or charm events
at the MC hadron level, $\ensuremath{C^{\text{rad}}}\xspace$ is the QED radiative correction
and $\mathcal{L}$ is the corresponding integrated luminosity.
Hadron-level jets were obtained by running the
$k_{T}$ clustering algorithm
on all stable final-state particles,
in the same mode as for the data.
Weakly decaying $b$ and $c$ hadrons were treated as stable particles
and were decayed only after the application of the jet algorithm.
The predictions from the \textsc{Hvqdis}\xspace program are given at the QED Born
level with a running coupling, \ensuremath{\alpha_{\text{em}}}\xspace. Hence, a correction of the
measured cross sections for QED radiative effects is necessary in
order to be able to compare them directly to the \textsc{Hvqdis}\xspace predictions.
The corrections were obtained using the \textsc{Rapgap}\xspace Monte Carlo as
$\ensuremath{C^{\text{rad}}}\xspace=\sigma_{\text{rad}}/\sigma_{\text{Born}}$, where
$\sigma_{\text{rad}}$ is the cross section with full QED corrections,
as used in the standard MC samples, and $\sigma_{\text{Born}}$ was
obtained with the QED corrections turned off but keeping \ensuremath{\alpha_{\text{em}}}\xspace
running. Both cross sections, $\sigma_{\text{rad}}$ and
$\sigma_{\text{Born}}$, were obtained at the hadron level.
\section{Systematic uncertainties}
\label{sec:syst}
The systematic uncertainties were evaluated by varying the analysis
procedure or by changing the selection cuts
and repeating the extraction of the cross section.
The following sources of experimental systematic uncertainties were
identified~\cite{thesis:libov:2012, thesis:roloff:2011};
the uncertainties on the integrated cross sections determined for each
source are summarised in Table~\ref{tab:syst} to indicate the sizes of the
different effects:
\begin{enumerate}[$\delta_1$]
\item DIS selection -- the cuts for DIS event selection were varied in
both data and MC.
The cut on the scattered electron energy was varied
between $9 < E'_{e} < \SI{11}{\GeV}$ ($\delta_{1}^{E_{e}}$), the cut on the inelasticity was varied
between $0.01 < \ensuremath{y_{\text{JB}}}\xspace < 0.03$ ($\delta_{1}^{y}$),
and the lower cut on $E-p_{Z}$
was changed by $\pm \SI{2}{\GeV}$ ($\delta_{1}^{E-p_{Z}}$);
\item trigger efficiency -- the uncertainty on the trigger efficiency
was evaluated by comparing events taken with independent triggers;
\item tracking efficiency correction -- the size of the correction was varied
by its estimated uncertainty of $\pm \SI{50}{\%}$;
\item decay-length smearing -- the fraction of secondary vertices for
which the decay length was smeared was varied separately in the
core ($\delta_{4}^{\text{core}}$) and the tails ($\delta_{4}^{\text{tail}}$) of the distribution
such that the agreement between data and MC remained reasonable;
\item signal extraction procedure -- the systematic uncertainty on the
signal extraction procedure was estimated by changing the lower
$|S|$ cut from $|S| > 4$ to $|S| > 3$ and $|S| > 5$;
\item jet energy scale -- the calorimetric part of the transverse jet
energy in the MC was varied by its estimated uncertainty of $\pm \SI{3}{\%}$;
\item electron energy scale -- the reconstructed energy of the scattered
electron was varied in the MC by its estimated uncertainty of $\pm \SI{2}{\%}$;
\item \label{syst:MCc}MC model dependence -- the
\ensuremath{Q^{2}}\xspace ($\delta_{8}^{\ensuremath{Q^{2}}\xspace}$),
\ensuremath{\eta^{\text{jet}}}\xspace ($\delta_{8}^{\ensuremath{\eta^{\text{jet}}}\xspace}$) and
\ensuremath{E_{T}^{\text{jet}}}\xspace ($\delta_{8}^{\ensuremath{E_{T}^{\text{jet}}}\xspace}$) reweighting
corrections in the charm MC were varied in a range for which the description of data by MC remained reasonable.
The same relative variations were applied to the beauty MC;
\item light-flavour background -- the light-flavour contribution to
the subtracted significance distribution includes a contribution
from long-lifetime strange-hadron decays. To estimate the
uncertainty due to modelling of this effect, the MC light-flavour
distribution of $N(S^{+}) - N(S^{-})$ was scaled by $\pm
\SI{30}{\%}$~\cite{Aaron:2009ut} and the fit was repeated;
\item[$\delta_{10}$] charm fragmentation function -- to estimate the sensitivity to
the charm fragmentation function, it was changed in the MC from the
Bowler
to the Peterson~\cite{pr:d27:105} parameterisation with
$\epsilon=0.062$~\cite{1126-6708-2009-04-082};
\item[$\delta_{11}$] beauty fragmentation function -- to estimate the sensitivity to
the beauty fragmentation function, it was changed in the MC from the
Bowler
to the Peterson parameterisation with
$\epsilon=0.0041$~\cite{Abbiendi:2002vt};
\item[$\delta_{12}$] charm branching fractions ($\delta^{\text{BR}}_{12}$) and
fragmentation fractions ($\delta^{\text{frag}}_{12}$) -- these
were varied
within the uncertainties of the world-average
values~\cite{Nakamura:2010zzi,arXiv:1112.3757,thesis:viazlo:2012};
\item[$\delta_{13}$] luminosity measurement -- a $\SI{1.9}{\%}$ overall
normalisation uncertainty was associated with the luminosity
measurement.
\end{enumerate}
To evaluate the total
systematic uncertainty on the integrated cross sections, the contributions from the different
systematic uncertainties were added in quadrature, separately for the
negative and the positive variations.
The same procedure was applied
to each bin for the differential cross sections. However,
the luminosity measurement uncertainty was not included.
In the case of beauty, the dominant effects arise from the
uncertainties on the track-finding inefficiencies, the beauty fragmentation function
and MC modelling. For charm, the uncertainties on the branching
fractions, the light-flavour asymmetry as well as on the MC modelling
contribute most to the total systematic uncertainty.
\section{Cross sections}
\label{sec:results}
Cross sections for inclusive jet production in beauty (charm) events were measured
in the range $\ensuremath{E_{T}^{\text{jet}}}\xspace > \SI[parse-numbers=false]{5 (4.2)}{\GeV}$,
$-1.6 < \ensuremath{\eta^{\text{jet}}}\xspace < 2.2$
for DIS events with $5 < \ensuremath{Q^{2}}\xspace < \SI{1000}{\GeV^{2}}$ and
$0.02 < y < 0.7$, where the jets are defined as in Section~\ref{sec:bc-extr}.
The single-differential cross sections for jet production in beauty
and charm events were measured as a function of $\ensuremath{E_{T}^{\text{jet}}}\xspace, \ensuremath{\eta^{\text{jet}}}\xspace,
\ensuremath{Q^{2}}\xspace$ and $x$.
The results of the measured cross
sections are given in Tables~\ref{tab:diffet}--\ref{tab:diffx} and
shown in Figs.~\ref{fig:diff_et}--\ref{fig:diff_x}.
The measurements
are compared to the \textsc{Hvqdis}\xspace NLO QCD predictions obtained using ZEUS-S and ABKM
as proton PDFs, and to the \textsc{Rapgap}\xspace predictions scaled by a factor of
1.49 for beauty and 1.40 for charm. The scale factors correspond to
the ratio of the measured integrated visible cross section to the \textsc{Rapgap}\xspace
prediction. The shapes of all measured beauty cross sections are reasonably
well described by \textsc{Hvqdis}\xspace and the \textsc{Rapgap}\xspace MC. \textsc{Rapgap}\xspace provides a worse
description of the shape of the charm cross sections than \textsc{Hvqdis}\xspace.\footnote{For
the acceptance corrections, the Monte Carlo was reweighted as discussed in
Section~\ref{sec:montecarlo}.} For charm, the data are
typically \SIrange{20}{30}{\%} above the \textsc{Hvqdis}\xspace NLO prediction, but
in reasonable agreement within uncertainties. Differences between the
NLO predictions using the different proton PDFs are mostly very small.
Double-differential cross sections
as a function of $x$ for different
ranges of \ensuremath{Q^{2}}\xspace for inclusive jet production in beauty and charm events are
listed in Tables~\ref{tab:ddiffb} and~\ref{tab:ddiffc}, respectively.
\section{\boldmath Extraction of \ensuremath{F_{2}^{q\bar{q}}}\xspace and \ensuremath{\sigma_{\text{r}}^{q\bar{q}}}\xspace}
\label{sec:F2q}
The heavy-quark contribution to the proton structure function $F_{2}$,
\ensuremath{F_{2}^{q\bar{q}}}\xspace with $q = b, c$, can be defined in terms of
the inclusive heavy flavour double-differential cross section as a function of $x$ and \ensuremath{Q^{2}}\xspace,
\begin{equation*}
\frac{\text{d}^{2}\sigma_{q\bar{q}}}{\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace} =
\frac{2\pi \ensuremath{\alpha_{\text{em}}}\xspace^{2}}{xQ^{4}} {\Big\{[1 + (1 - y^{2})] \ensuremath{F_{2}^{q\bar{q}}}\xspace(x,\ensuremath{Q^{2}}\xspace)-y^{2} \ensuremath{F_{L}^{q\bar{q}}}\xspace(x
,\ensuremath{Q^{2}}\xspace)\Big\}} \,,
\end{equation*}
where \ensuremath{F_{L}^{q\bar{q}}}\xspace is the heavy-quark contribution to the
structure function $F_{L}$.
To extract \ensuremath{F_{2}^{q\bar{q}}}\xspace from the visible jet production cross sections in heavy-quark events,
measured in bins of $x$ and \ensuremath{Q^{2}}\xspace, an
extrapolation from the measured range in \ensuremath{E_{T}^{\text{jet}}}\xspace and \ensuremath{\eta^{\text{jet}}}\xspace to the
full kinematic phase space was performed.
This implicitly takes into account the jet multiplicity.
The measured values of \ensuremath{F_{2}^{q\bar{q}}}\xspace at a
reference point in the $x$--\ensuremath{Q^{2}}\xspace plane were calculated using
\begin{equation}
\ensuremath{F_{2}^{q\bar{q}}}\xspace(x,\ensuremath{Q^{2}}\xspace) =
\frac{\text{d}^{2}\sigma_{q}^{\text{jet}} /\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace}
{\text{d}^{2}\sigma_{q}^{\text{had,NLO}}/\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace}
F_{2}^{q\bar{q},\text{NLO}}(x,\ensuremath{Q^{2}}\xspace)\,,
\label{eq:F2q}
\end{equation}
where
$\text{d}^{2}\sigma_{q}^{\text{jet}}/\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace$ is determined in analogy to
Eq.~\eqref{eq:xsec4}, and $F_{2}^{q\bar{q},\text{NLO}}$ and
$\text{d}^{2}\sigma_{q}^{\text{had,NLO}}/\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace$ were
calculated at NLO in the FFNS using the \textsc{Hvqdis}\xspace program with the factor \ensuremath{C^{\text{had}}}\xspace applied as in Eq.~\eqref{eq:chad}. The proton
PDFs were obtained from the FFNS variant of the HERAPDF~1.0 NLO QCD
fit~\cite{epj.c73.2311}.
This PDF was used in order to be consistent with the HERA combined results~\cite{epj.c73.2311}.
The strong coupling constant $\ensuremath{\alpha_{s}}\xspace(M_Z)$ was set to 0.105 as in
the PDF fit. Other settings were as described in
Section~\ref{sec:theory} for the ZEUS-S variant.
As discussed in
Section~\ref{sec:bc-extr}, $\text{d}^{2}\sigma_{q}^{\text{jet}}/\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace$
was multiplied by $1/\ensuremath{C_{q}^{\text{rad}}}\xspace$, hence \ensuremath{F_{2}^{q\bar{q}}}\xspace is given at QED Born level,
consistent with the usual convention.
The procedure of Eq.~\eqref{eq:F2q} also
corrects for the \ensuremath{F_{L}^{q\bar{q}}}\xspace contribution to the cross section. This assumes that the
calculation correctly predicts the ratio $\ensuremath{F_{L}^{q\bar{q}}}\xspace/\ensuremath{F_{2}^{q\bar{q}}}\xspace$.
The
extrapolation factors for beauty due to cuts on \ensuremath{E_{T}^{\text{jet}}}\xspace and \ensuremath{\eta^{\text{jet}}}\xspace
typically range from 1.3 to 1.0, decreasing with increasing \ensuremath{Q^{2}}\xspace. The factor
is up to 1.7 at high values of $x$.
For charm, the extrapolation
factors are typically about 4 in the region $5 < \ensuremath{Q^{2}}\xspace < \SI{20}{\GeV^{2}}$
and about 2 in the region $20 < \ensuremath{Q^{2}}\xspace < \SI{60}{\GeV^{2}}$. The uncertainty on
the extrapolation from the measured range to the full kinematic phase
space was estimated by varying the parameters of the calculation for the
extrapolation factors and adding the resulting uncertainties in
quadrature. For charm, the same variations were performed as for
the HERA combined results~\cite{epj.c73.2311}:
the charm mass was varied by
$\pm \SI{0.15}{\GeV}$; the strong coupling constant $\ensuremath{\alpha_{s}}\xspace(M_Z)$ was changed by $\pm 0.002$;
renormalisation and factorisation scales were multiplied
simultaneously by 0.5 or 2.
Uncertainties resulting from the proton PDF uncertainty
are small~\cite{zeush1.web}
and were neglected.
For beauty, the same variations of \ensuremath{\alpha_{s}}\xspace and scales were made and
the beauty mass was varied by $\pm \SI{0.25}{\GeV}$.
For each bin, a reference
point in $x$ and \ensuremath{Q^{2}}\xspace was defined (see Table~\ref{tab:f2b}) to
calculate the structure function.
In addition, beauty and charm reduced cross sections were determined.
They are defined as
\begin{equation*}
\ensuremath{\sigma_{\text{r}}^{q\bar{q}}}\xspace = \frac{\text{d}^{2}\sigma_{q\bar{q}}}{\text{d} x\,\text{d} \ensuremath{Q^{2}}\xspace} \cdot
\frac{xQ^{4}}{2\pi \ensuremath{\alpha_{\text{em}}}\xspace^{2}[1 + (1 - y^{2})]} = F_{2}^{q\bar{q}}(x,\ensuremath{Q^{2}}\xspace)-\frac{y^{2}}{1 + (1 - y^{2})} F_{L}^{q\bar{q}}(x
,\ensuremath{Q^{2}}\xspace) \,,
\end{equation*}
and are extracted in analogy to \ensuremath{F_{2}^{q\bar{q}}}\xspace as described above except that
no assumption on \ensuremath{F_{L}^{q\bar{q}}}\xspace is required.
The extracted values of \ensuremath{F_{2}^{b\bar{b}}}\xspace and \ensuremath{F_{2}^{c\bar{c}}}\xspace are given in
Tables~\ref{tab:f2b} and~\ref{tab:f2c}, respectively, while \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace and
\ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace are shown in Tables~\ref{tab:reduced_beauty}
and~\ref{tab:reduced_charm}. The total uncertainties of the
measurements were calculated from the statistical and systematic
uncertainties of the measured cross sections
(Tables~\ref{tab:ddiffb},~\ref{tab:ddiffc},~\ref{tab:beauty_systematics_part1}--\ref{tab:charm_systematics_part2}) and of the extrapolation uncertainty
(Tables~\ref{tab:f2b_extrapolation}--\ref{tab:redc_extrapolation}),
added in quadrature.
The structure function \ensuremath{F_{2}^{c\bar{c}}}\xspace is shown in Fig.~\ref{fig:f2c_x} as a
function of $x$ for different values of \ensuremath{Q^{2}}\xspace. The measurements are
compared to the NLO QCD HERAPDF~1.5~\cite{herapdf.web} predictions,
the most recent official release of the HERAPDF, based on
the RT~\cite{springerlink:10.1140/epjc/s10052-009-1072-5} general-mass
variable-flavour-number scheme (GMVFNS).
The predictions are consistent with the measurements.
In Fig.~\ref{fig:sigma_red_c}, the measured \ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace
values are compared to the HERA combined results~\cite{epj.c73.2311} as well as to the two recent results from
ZEUS~\cite{jhep05.2013.097, jhep05.2013.023} which are not yet included
in the combination. For the comparison, some of the measured values of
this analysis were swum in \ensuremath{Q^{2}}\xspace and $x$ using \textsc{Hvqdis}\xspace. This measurement
is competitive, especially at high \ensuremath{Q^{2}}\xspace, where the extrapolation
uncertainty is low, and is in agreement with the HERA combined
measurements.
The structure function \ensuremath{F_{2}^{b\bar{b}}}\xspace is shown in Fig.~\ref{fig:f2b_x} as a
function of $x$ for different values of \ensuremath{Q^{2}}\xspace. The measurements
are compared to HERAPDF~1.5 GMVFNS predictions. The increase in the uncertainty on the prediction
around $\ensuremath{Q^{2}}\xspace = m_{b}^{2}$ is a feature of the GMVFNS scheme used.
The predictions
are consistent with the measurements.
The \ensuremath{F_{2}^{b\bar{b}}}\xspace measurement is also shown as a function of \ensuremath{Q^{2}}\xspace for fixed
$x$ in Fig.~\ref{fig:f2b_q2}, and is compared to previous ZEUS and H1
measurements. Again, \textsc{Hvqdis}\xspace was used to swim the measured values in
\ensuremath{Q^{2}}\xspace and $x$ to match the previous measurements.
In a wide range of \ensuremath{Q^{2}}\xspace, this measurement represents the most precise
determination of \ensuremath{F_{2}^{b\bar{b}}}\xspace at HERA. It is in good agreement with the
previous ZEUS analyses and H1 measurements. Several NLO and NNLO QCD
predictions based on the fixed- or variable-flavour-number
schemes~\cite{herapdf.web,PhysRevD.78.013004,
springerlink:10.1140/epjc/s10052-009-1072-5,
Thorne:2008xf, jimenez:2009,
Alekhin2009166, Alekhin:2009ni, Alekhin:2010iu}
are also compared to the measurements.
Predictions from different theoretical approaches agree
well with each other.
All predictions provide a reasonable description of the data.
\section{Measurement of the running beauty-quark mass}
\label{sec:mb}
The reduced beauty cross sections, \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace,
(Fig.~\ref{fig:sigma_red_b} and Table~\ref{tab:reduced_beauty})
together with inclusive DIS data were used
to determine
the beauty-quark mass, in a simultaneous fit of the
mass and the parton densities.
The measurement procedure follows closely the method presented in a
recent H1-ZEUS publication~\cite{epj.c73.2311}, where the running
charm-quark mass in the \hbox{$\overline{\rm MS}$}\xspace scheme was extracted using a
simultaneous QCD fit of the combined HERA~I inclusive DIS
data~\cite{Aaron:2009aa} and the HERA combined charm DIS
data~\cite{epj.c73.2311}.
This approach
was also used and extended
by a similar independent analysis~\cite{adlm2},
and was preceded by a
similar analysis of a partial charm data set~\cite{adlm}.
The fit for the running beauty-quark mass was performed within the
HERAFitter\xspace~\cite{herafitter} framework choosing the ABM implementation
of the fixed-flavour-number
scheme at next-to-leading order~\cite{ABKMMSbar, ABKMMSbar2,
Alekhin:2009ni, np:b392:162, np:b392:229}. The
OPENQCDRAD~\cite{openqcdrad} option in HERAFitter\xspace was used
in the \hbox{$\overline{\rm MS}$}\xspace running-mass mode.
The fit was applied to the beauty data listed in Table~\ref{tab:reduced_beauty}
and to the same inclusive DIS data~\cite{Aaron:2009aa} as in
the charm-quark mass fit~\cite{epj.c73.2311}.
A fit to the inclusive data only shows a very weak dependence on $m_{b}$.
In order to avoid technical complications,
no charm data were included in the simultaneous fit and only $m_b$ was extracted.
The PDFs resulting from the simultaneous fit changed only marginally with respect to
the nominal PDFs obtained from the fit to the inclusive DIS data only.
The $\chi^2$ of the QCD fit, including the beauty data, shows a clear dependence on the
beauty-quark mass, $m_b$, as can be seen in Fig.~\ref{fig:chi2}.
The total $\chi^2$ for the best fit
is 587 for 596 degrees of freedom, and the partial contribution from
the beauty data is 11.4 for 17 points. The beauty-quark mass and its
uncertainty are determined from a parabolic parameterisation. The best fit yields
\begin{equation*}
\mbMSbarmeas
\end{equation*}
for the \hbox{$\overline{\rm MS}$}\xspace running beauty-quark mass at NLO.
The fit uncertainty\footnote{For the charm-quark mass fit~\cite{epj.c73.2311} this uncertainty was denoted
``exp''.} (fit) is determined from
$\Delta\chi^2=1$. It contains the experimental uncertainties,
the extrapolation uncertainties, the uncertainties of the standard PDF
parameterisation, as well as an estimate of the uncertainty on the hadronisation
corrections, as detailed below.
In addition, the result has uncertainties attributed to the
choices of some extra model parameters (mod.), some additional
variations of the PDF parameterisation (param.) and uncertainties on the
perturbative QCD parameters
(theo.). Details of the uncertainty evaluation include:
\begin{description}
\item[Fit uncertainty:]
For the beauty data, all uncertainties from
Tables~\ref{tab:beauty_systematics_part1},
\ref{tab:beauty_systematics_part2} (experimental) and
\ref{tab:redb_extrapolation} (extrapolation), and the statistical uncertainty,
as summarised in
Table~\ref{tab:reduced_beauty}, were accounted for in the fit.
Following the discussion in Section~\ref{sec:theory}, an uncertainty of \SI{100}{\percent}
on $\ensuremath{\Delta^{\text{had}}}\xspace = \ensuremath{C^{\text{had}}}\xspace - 1$ (Table~\ref{tab:ddiffb}) was introduced
as an additional uncorrelated uncertainty.
The uncertainties arising from the default PDF parameterisation~\cite{epj.c73.2311},
including the so-called ``flexible'' gluon
parameterisation, are implicitly part of the fit uncertainty.
The statistical uncertainties and the uncertainties $\delta_1,
\delta_2, \delta_4^{\text{core}}, \delta_5$ and $\delta_{12}$ from
Tables~\ref{tab:beauty_systematics_part1} and~\ref{tab:beauty_systematics_part2} were treated as uncorrelated, while
all other uncertainties, including those from luminosity and
from Table~\ref{tab:redb_extrapolation}, were treated as
point-to-point correlated.
The ``multiplicative'' uncertainty option~\cite{herafitter} from HERAFitter\xspace was
used. In the case of asymmetric uncertainties, the larger was used
in both directions.
The uncertainties of the inclusive data were used as published.
Since the inclusive data were taken during the HERA~I phase
and the beauty data during the HERA~II phase, the two sets of
data were treated as uncorrelated.
\item[Model uncertainty:]
The model choices include an assumption on the strangeness fraction, $f_s$,
the minimum $Q^2$ used in the data selection, $Q^2_{\text{min}}$, and
$Q^2_0$, the starting value for the QCD evolution.
These were treated exactly as in the charm-quark mass fit~\cite{epj.c73.2311}.
Table~\ref{tab:result} lists the choices and variations and their
individual contributions to the model uncertainty attributed to the
model choices.
Another source of uncertainty is that the $b$-quark mass was
used earlier to extrapolate the measured visible cross sections to
the reduced cross sections. The corresponding uncertainty is
parameterised in Table~\ref{tab:redb_extrapolation} and used in the
fit, but the correlation of this uncertainty with the mass used in
the QCD fit is lost. Since the \textsc{Hvqdis}\xspace~\cite{pr:d57:2806} program
used for the extrapolation uses the pole-mass scheme, and no
differential calculations are available in the running-mass scheme,
no fully consistent treatment of this correlation is possible.
However, the equivalent uncertainty when using the pole-mass scheme
can be consistently estimated.
For this purpose, the fit was repeated with
the pole-mass option of OPENQCDRAD, which was checked to yield
results consistent with the \textsc{Hvqdis}\xspace predictions for \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace.
The result,
$m_b(\text{pole}) = \SI[parse-numbers=false]{4.35 \pm 0.14\,(\text{fit})}{\GeV}$,
has a fit uncertainty which is the same as
the fit uncertainty for the running-mass fit. However, since the pole-mass definition
includes an unavoidable additional theoretical uncertainty due to a
nonperturbative contribution,
no attempt to extract a pole-mass measurement with full systematic uncertainties was made.
To recover the correlation between the extrapolation and the mass fit,
the extrapolated cross sections were iteratively modified
using the predictions from the mass values obtained in each fit.
This removes the uncertainty on $m_{b}$ in the extrapolation
and takes the full correlations into account.
The resulting mass
$m_b(\text{pole}) = \SI[parse-numbers=false]{4.28 \pm
0.13\,(\text{fit})}{\GeV}$ is slightly lower.
The difference between the results from the two procedures
($\delta m_{\text{ext}} = \SI{-0.07}{\GeV}$) was treated as an additional
model uncertainty.
\item[PDF parameterisation uncertainty:]
The parameterisation of the PDFs is chosen as for the charm-quark mass
fit~\cite{epj.c73.2311}, including the ``flexible'' parameterisation
of the gluon distribution. An additional
uncertainty is estimated by freeing three extra PDF
parameters $D_{u_v}$, $D_{\bar D}$ and $D_{\bar U}$ in the fit which allow
for small shape variations in the $u_v$, $\bar{U}$ and $\bar{D}$ parton
distributions~\cite{epj.c73.2311}.
The change in $m_{b}$ due to leaving the parameters free is given as the systematic uncertainty in Table~\ref{tab:result}.
\item[Perturbative scheme and related theory uncertainty:]
The parameters used for the perturbative part of the QCD calculations
also introduce uncertainties;
the effects are listed in Table~\ref{tab:result}.
As in previous analyses~\cite{adlm, epj.c73.2311, adlm2}, the \hbox{$\overline{\rm MS}$}\xspace running-mass
scheme~\cite{mc:run, mc:run2, mc:run3}
was chosen for all calculations of the reduced cross sections and
the fit because it
shows better perturbative convergence behaviour than the pole-mass
scheme.
In order to allow the low-\ensuremath{Q^{2}}\xspace points of the inclusive DIS
measurement to be included without the need of additional charm-quark mass
corrections, the number of active flavours (NF) was set to three, i.e.\ the charm
contribution was also treated in the fixed-flavour-number scheme.
Accordingly, the strong coupling constant was set to
$\ensuremath{\alpha_{s}}\xspace(M_Z)^{\text{NF}=3} = 0.105 \pm 0.002$, corresponding to
$\ensuremath{\alpha_{s}}\xspace(M_Z)^{\text{NF}=5} = 0.116 \pm 0.002$.
The theoretical prediction
of the charm contribution to the inclusive DIS data is obtained using the
running charm-quark mass obtained from the fit to the combined HERA charm data~\cite{epj.c73.2311},
i.e.\ $m_c(m_c) = \SI{1.26(6)}{\GeV}$.
It was checked that, as expected, using this mass together with the central
PDF from the $m_{b}$ fit,
a good description of the combined HERA charm data~\cite{epj.c73.2311} was obtained.
Thus, the charm contribution to the inclusive data should be well described.
The renormalisation and factorisation scales were set to $\mu =
\mu_R = \mu_F = \sqrt{\ensuremath{Q^{2}}\xspace+4m^2}$ with $m=0, m_c, m_b$ for the light quark,
charm, and beauty contributions, respectively, and varied simultaneously by a factor two as in
previous analyses~\cite{adlm, adlm2}.
\end{description}
The measured beauty-quark mass is in
very good agreement with the world average $m_b(m_b)= \SI{4.18 \pm
0.03}{\GeV}$~\cite{PDG2012}.
The resulting predictions for the beauty cross
sections are shown together with the data in Fig.~\ref{fig:sigma_red_b}.
Figure~\ref{fig:sigma_red_b} also shows the change in the predictions
resulting from the fit when different $m_b$ values are assumed.
The largest sensitivity to
$m_b$ arises from the low-\ensuremath{Q^{2}}\xspace region, while at high \ensuremath{Q^{2}}\xspace the
impact of the beauty-quark mass is small.
\section{Conclusions}
\label{sec:conclusions}
Inclusive jet production cross sections in events
containing beauty or charm quarks have been
measured in DIS at HERA, exploiting the long lifetimes and large
masses of $b$ and $c$ hadrons. In contrast to previous analyses at ZEUS,
the measurement was not restricted to any particular final state. This
resulted in substantially increased statistics.
Differential cross sections as functions of \ensuremath{E_{T}^{\text{jet}}}\xspace, \ensuremath{\eta^{\text{jet}}}\xspace, \ensuremath{Q^{2}}\xspace and
$x$ were determined. Next-to-leading-order QCD predictions calculated using the \textsc{Hvqdis}\xspace
program using two different sets of proton PDFs are consistent with
the measurements.
The heavy-quark contributions to the proton structure function $F_{2}$
as well as beauty and charm reduced cross sections were extracted from
the double-differential cross sections as a function of $x$ and \ensuremath{Q^{2}}\xspace.
The \ensuremath{F_{2}^{b\bar{b}}}\xspace, \ensuremath{F_{2}^{c\bar{c}}}\xspace and \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace, \ensuremath{\sigma_{\text{r}}^{c\bar{c}}}\xspace values are in agreement with previous
measurements at HERA.
The results were also compared to several NLO and NNLO QCD
calculations, which provide a good description of the data. The
precision of the \ensuremath{F_{2}^{c\bar{c}}}\xspace measurement is competitive with other
analyses. For a wide range of \ensuremath{Q^{2}}\xspace, the \ensuremath{F_{2}^{b\bar{b}}}\xspace measurement represents
the most precise determination of \ensuremath{F_{2}^{b\bar{b}}}\xspace.
The running beauty-quark mass in the \hbox{$\overline{\rm MS}$}\xspace scheme was determined from
an NLO QCD fit in the fixed-flavour-number scheme to the \ensuremath{\sigma_{\text{r}}^{b\bar{b}}}\xspace
cross sections from
this analysis and to HERA~I inclusive DIS data:
\begin{equation*}
\mbMSbarmeas
\end{equation*}
This value agrees well with the world average.
\section*{Acknowledgements}
\label{sec:acknowledge}
We appreciate the contributions to the construction, maintenance and operation of
the ZEUS detector of many people who are not listed as authors. The
HERA machine group and the DESY computing staff are especially
acknowledged for their success in providing excellent operation of the
collider and the data-analysis environment. We thank the DESY
directorate for their strong support and encouragement.
It is a pleasure to thank the ABKM, CTEQ, JR and MSTW groups that
provided the predictions for \ensuremath{F_{2}^{b\bar{b}}}\xspace shown in
Fig.~\ref{fig:f2b_q2}.
We gratefully acknowledge the advice from S.~Alekhin
and R.~Pla\v{c}akyt\.{e} concerning the appropriate usage of OPENQCDRAD and
HERAFitter\xspace.
|
\section{Introduction}
The aim of this paper is to define the matching measure of an infinite lattice
$L$ and show how it can be used to analyze the behaviour of the monomer-dimer
model on $L$. The notion of matching measure has been recently introduced by
the first and second authors, Frenkel and Kun in \cite{ACFK}. There are
essentially two ways to define it: in this paper we take the path of giving a
direct, spectral definition for infinite vertex transitive lattices, using
self-avoiding walks and then connect it to the monomer-dimer model via graph
convergence. Recall that a graph $L$ is \emph{vertex transitive} if for any
two vertices of $L$ there exists an automorphism of $L$ that brings one vertex
to the other.
Let $v$ be a fixed vertex of the graph $L$. A walk in $L$ is
\emph{self-avoiding}, if it touches every vertex at most once. There is a
natural graph structure on the set of finite self-avoiding walks starting at
$v$: we connect two walks if one is a one step extension of the other. The
resulting graph is an infinite rooted tree, called the \emph{tree of
self-avoiding walks} \emph{of }$L$\emph{\ starting at }$v$.
\begin{Def}
\label{saw}Let $L$ be an infinite vertex transitive lattice. The
\emph{matching measure} $\rho_{L}$ is the spectral measure of the tree of
self-avoiding walks of $L$ starting at $v$, where $v$ is a vertex of $L$.
\end{Def}
By vertex transitivity, the definition is independent of $v$. For a more
general definition, also covering lattices that are not vertex transitive, see
Section~\ref{measure}.
To make sense of why we call this the matching measure, we need the notion of
Benjamini--Schramm convergence. Let $G_{n}$ be a sequence of finite graphs. We
say that $G_{n}$ Benjamini--Schramm converges to $L$, if for every $R>0$, the
probability that the $R$-ball centered at a uniform random vertex of $G_{n}$
is isomorphic to the $R$-ball in $L$ tends to $1$ as $n$ tends to infinity.
That is, if by randomly sampling $G_{n}$\ and looking at a bounded distance,
we can not distinguish it from $L$ in probability.
All Euclidean lattices $L$ can be approximated this way by taking sequences of
boxes with side lengths tending to infinity, by bigger and bigger balls in $L$
in its graph metric, or by suitable tori. When $L$ is a Bethe lattice (a
$d$-regular tree), finite subgraphs never converge to $L$ and the usual way is
to set $G_{n}$ to be $d$-regular finite graphs where the minimal cycle length
tends to infinity.
For a finite graph $G$ and $k>0$ let $m_{k}(G)$ be the number of monomer-dimer
arrangements with $k$ dimers (matchings of $G$ using $k$ edges). Let
$m_{0}(G)=1$. Let the \emph{matching polynomial}
\[
\mu(G,x)=\sum_{k}(-1)^{k}m_{k}(G)x^{\left\vert G\right\vert -2k
\]
and let $\rho_{G}$, the \emph{matching measure of }$G$ be the uniform
distribution on the roots of $\mu(G,x)$. Note that $\mu(G,x)$ is just a
reparametrization of the monomer-dimer partition function. The matching
polynomial has the advantage over the partition function that its roots are
bounded in terms of the maximal degree of $G$.
Using previous work of Godsil \cite{god3} we show that $\rho_{L}$ can be
obtained as the thermodynamical limit of the $\rho_{G_{n}}$.
\begin{Th}
\label{konv}Let $L$ be an infinite vertex transitive lattice and let $G_{n}$
Benjamini--Schramm converge to $L$. Then $\rho_{G_{n}}$ weakly converges to
$\rho_{L}$ and $\lim_{n\rightarrow\infty}\rho_{G_{n}}(\{x\})=\rho_{L}(\{x\})$
for all $x\in\mathbb{R}$.
\end{Th}
So in this sense, the matching measure can be thought of as the `root
distribution of the partition function for the infinite monomer-dimer model',
transformed by a fixed reparametrization.
It turns out that the matching measure can be effectively used as a substitute
for the Mayer series. An important advantage over it is that certain natural
functions can be integrated along this measure even in those cases when the
corresponding series do not converge. We demonstrate this advantage by giving
new, strong estimates on the free energies of monomer-dimer models for
Euclidean lattices, by expressing them directly from the matching measures.
The computation of monomer-dimer and dimer free energies has a long history.
The precise value is known only in very special cases. Such an exceptional
case is the Fisher-Kasteleyn-Temperley formula \cite{fis,kas,tem} for the
dimer model on $\mathbb{Z}^{2}$. There is no such exact result for
monomer-dimer models. The first approach for getting estimates was the use of the
transfer matrix method. Hammersley \cite{ham1,ham2}, Hammersley and Menon
\cite{ham3} and Baxter \cite{bax} obtained the first (non-rigorous) estimates
for the free energy. Then Friedland and Peled \cite{FP} proved the rigorous
estimates $0.6627989727\pm10^{-10}$ for $d=2$ and the range $[0.7653,0.7863]$
for $d=3$. Here the upper bounds were obtained by the transfer matrix method,
while the lower bounds relied on the Friedland-Tverberg inequality. The lower
bound in the Friedland-Peled paper was subsequently improved by newer and
newer results (see e.g. \cite{friegurv}) on Friedland's asymptotic matching
conjecture which was finally proved by L. Gurvits \cite{gur2}. Meanwhile, a
non-rigorous estimate $[0.7833,0.7861]$ was obtained via matrix permanents
\cite{huo}. The most significant improvement was obtained recently by D.
Gamarnik and D. Katz \cite{gam} via their new method which they called
sequential cavity method. They obtained the range $[0.78595,0.78599]$.
\bigskip
Here we only highlight one computational result. More data can be found in
Section~\ref{entropy-function}, in particular, in Table 1. Let $\tilde
{\lambda}(L)$ denote the monomer-dimer free energy of the lattice $L$, and let
$\mathbb{Z}^{d}$ denote the $d$-dimensional hyper-simple cubic lattice.
\begin{Th}
We have
\[
\tilde{\lambda}(\newline\mathbb{Z}^{3})=0.7859659243\pm9.88\cdot
10^{-7},
\
\[
\tilde{\lambda}(\newline\mathbb{Z}^{4})=0.8807178880\pm5.92\cdot
10^{-6}.
\]
\[
\tilde{\lambda}(\newline\mathbb{Z}^{5})=0.9581235802\pm4.02\cdot
10^{-5}.
\]
The bounds on the error terms are rigorous.
\end{Th}
Our method allows to get efficient estimates on arbitrary lattices. The
computational bottleneck is the tree of self-avoiding walks, which is famous
to withstand theoretical interrogation.
It is natural to ask what are the actual matching measures for the various
lattices. In the case of a Bethe lattice $\mathbb{T}_{d}$, the tree of
self-avoiding walks again equals $\mathbb{T}_{d}$, so the matching measure of
$\mathbb{T}_{d}$ coincides with its spectral measure. This explicit measure,
called Kesten-McKay measure has density
\[
\frac{d}{2\pi}\frac{\sqrt{4(d-1)-t^{2}}}{d^{2}-t^{2}}\chi_{\{|t|\leq
2\sqrt{d-1}\}}\text{.
\]
We were not able to find such explicit formulae for any of the Euclidean
lattices. However, using Theorem~\ref{konv} one can show that the matching
measures of hypersimple cubic lattices admit no atoms.
\begin{Th}
\label{atomless}The matching measures $\rho_{\mathbb{Z}^{d}}$ have no atoms.
\end{Th}
In Section~\ref{density} we prove a more general result which also shows that
for instance, the matching measure of the hexagonal lattice has no atoms. For
some images on the matching measures of $\mathbb{Z}^{2}$ and $\mathbb{Z}^{3}$
see Section~\ref{density}. We expect that the matching measures of all
hypersimple cubic lattices are absolutely continuous with respect to the
Lebesque measure. We also expect that the radius of support of the matching
measure (that is, the spectral radius of the tree of self-avoiding walks)
carries further interesting information about the lattice. Note that the
\emph{growth} of this tree for $\mathbb{Z}^{d}$ and other lattices has been under intense investigation
\cite{alm,dum,HSS}, under the name \emph{connective constant}. \bigskip
The paper is organized as follows. In Section~\ref{measure}, we define the
basic notions and prove Theorem~\ref{konv}. In Section~\ref{entropy-function} we
introduce the entropy function $\lambda_{G}(p)$ for finite graphs $G$ and
related functions, and we gather their most important properties. We also
extend this concept to lattices. In this section we provide the computational
data too. In Section~\ref{density}, we prove Theorem~\ref{atomless}.
\section{Matching measure}
\label{measure}
\subsection{Notations}
This section is about the basic notions and lemmas needed later. Since the
same objects have different names in graph theory and statistical mechanics,
for the convenience of the reader, we start with a short dictionary. \bigskip
\begin{center
\begin{tabular}
[c]{|c|c|}\hline
Graph theory & Statistical mechanics\\\hline
vertex & site\\\hline
edge & bond\\\hline
$k$-matching & monomer-dimer arrangement with $k$ dimers\\\hline
perfect matching & dimer arrangement\\\hline
degree & coordination number\\\hline
$d$-dimensional grid ($\mathbb{Z}^{d}$) & hyper-simple cubic lattice\\\hline
infinite $d$-regular tree ($\mathbb{T}_{d}$) & Bethe lattice\\\hline
path & self-avoiding walk\\\hline
\end{tabular}
\bigskip
\end{center}
Throughout the paper, $G$ denotes a finite graph with vertex set $V(G)$ and
edge set $E(G)$. The number of vertices is denoted by $|G|$. For an infinite
graph $L$, we will use the word \emph{lattice}. The \emph{degree} of a vertex
is the number of its neighbors. A graph is called $d$\emph{-regular} if every
vertex has degree exactly $d$. The graph $G-v$ denotes the graph obtained from
$G$ by erasing the vertex $v$ together with all edges incident to $v$.
For a finite or infinite graph $T$, let $l^{2}(T)$ denote the Hilbert space of
square summable real functions on $V(T)$. The \emph{adjacency operator}
$A_{T}:l^{2}(T)\rightarrow l^{2}(T)$ is defined by
\[
(A_{T}f)(x)=\sum\limits_{(x,y)\in E(T)}f(y)\text{ \ \ \ (}f\in l^{2
(T)\text{).
\]
When $T$ is finite, in the standard base of vertices, $A_{T}$ is a square
matrix, where $a_{u,v}=1$ if the vertices $u$ and $v$ are adjacent, otherwise
$a_{u,v}=0$. For a finite graph $T$, the characteristic polynomial of $A_{T}$
is denoted by $\phi(T,x)=\det(xI-A_{T})$.
A \emph{matching} is set of edges having pairwise distinct endpoints. A
$k$\emph{-matching} is a matching consisting of $k$ edges. A graph is called
\emph{vertex-transitive} if for every vertex pair $u$ and $v$, there exists an
automorphism $\varphi$ of the graph for which $\varphi(u)=v$.
\subsection{Matching measure and tree of self-avoiding walks}
The \emph{matching polynomial} of a finite graph $G$ is defined as
\[
\mu(G,x)=\sum_{k}(-1)^{k}m_{k}(G)x^{\left\vert G\right\vert -2k},
\]
where $m_{k}(G)$ denotes the number of $k$-matchings in $G$. Let $\rho_{G}$,
the \emph{matching measure} of $G$ be the uniform distribution on the zeros of
the matching polynomial of $G$.
The fundamental theorem for the matching polynomial is the following.
\begin{Th}
[Heilmann and Lieb \cite{hei}]\label{Hei} The roots of the matching polynomial
$\mu(G,x)$ are real, and if the largest degree $D$ is greater than $1$, then
all roots lie in the interval $[-2\sqrt{D-1},2\sqrt{D-1}]$.
\end{Th}
A walk in a graph is \emph{self-avoiding} if it touches every vertex at most
once. For a finite graph $G$ and a root vertex $v$, one can construct
$T_{v}(G)$, the \emph{tree of self-avoiding walks at }$v$ as follows: its
vertices correspond to the finite self-avoiding walks in $G$ starting at $v$,
and we connect two walks if one of them is a one-step extension of the other.
The following figure illustrates that in general, $T_{v}(G)$ very much depends
on the choice of $v$.
\begin{figure}[h!]
\centering
\tikzsetnextfilename{sawtree}
\begin{tikzpicture}
\begin{scope}[every node/.style={circle, draw, scale=.75, inner sep=.25mm}, xshift=-3.7cm]
\node (1) at (0,0) {1}; \node (2) at (0,.5) {2}; \node (3) at (.5,0) {3}; \node (4) at (0,-.5) {4}; \node (5) at (-.5,0) {5};
\draw (1) -- (2) (1) -- (3) (1) -- (4) (1) -- (5) (2) -- (3) (2) -- (5) (3) -- (4) (4) -- (5);
\end{scope}
\begin{scope}[grow cyclic, rotate=-45, level distance=5mm, inner sep=.25mm,
level 1/.style={sibling angle=-90}, level 2/.style={sibling angle=-75}, level 3/.style={sibling angle=-35},
level 4/.style={sibling angle=-20}, level 5/.style={sibling angle=-3}, level 6/.style={sibling angle=0},
every node/.style={circle, draw, thin, scale=.75}]
\node(1){1}child{node(12){2}child{node(125){5}child[missing]child{node(1254){4}child[missing]child{node(12543){3}
child[missing]child[missing]}}}child{node(123){3}child{node(1234){4}child{node(12345){5}child[missing]child[missing]}
child[missing]}child[missing]}}child{node(13){3}child{node(132){2}child[missing]child{node(1325){5}child[missing]
child{node(13254){4}child[missing]child[missing]}}}child{node(134){4}child{node(1345){5}child{node(13452){2}
child[missing]child[missing]}child[missing]}child[missing]}}child{node(14){4}child{node(143){3}child[missing]
child{node(1432){2}child[missing]child{node(14325){5}child[missing]child[missing]}}}child{node(145){5}
child{node(1452){2}child{node(14523){3}child[missing]child[missing]}child[missing]}child[missing]}}child{node(15){5}
child{node(154){4}child[missing]child{node(1543){3}child[missing]child{node(15432){2}child[missing]child[missing]}}}
child{node(152){2}child{node(1523){3}child{node(15234){4}child[missing]child[missing]}child[missing]}child[missing]}};
\end{scope}
\begin{scope}[grow cyclic, level distance=5mm, inner sep=.25mm, xshift=5cm,
level 1/.style={sibling angle=-120}, level 2/.style={sibling angle=-75}, level 3/.style={sibling angle=-55},
level 4/.style={sibling angle=-30}, level 5/.style={sibling angle=-15}, level 6/.style={sibling angle=0},
every node/.style={circle, draw, thin, scale=.75}]
\node(2){2}child{node(23){3}child{node(234){4}child{node(2345){5}child[missing]child{node(23451){1}child[missing]
child[missing]child[missing]}}child{node(2341){1}child{node(23415){5}child[missing]child[missing]}child[missing]
child[missing]}}child{node(231){1}child{node(2314){4}child[missing]child{node(23145){5}child[missing]child[missing]}}
child{node(2315){5}child{node(23154){4}child[missing]child[missing]}child[missing]}child[missing]}}child{node(21){1}
child{node(213){3}child[missing]child{node(2134){4}child{node(21345){5}child[missing]child[missing]}child[missing]}}
child{node(214){4}child{node(2143){3}child[missing]child[missing]}child{node(2145){5}child[missing]child[missing]}}
child{node(215){5}child{node(2154){4}child[missing]child{node(21543){3}child[missing]child[missing]}}child[missing]}}
child{node(25){5}child{node(251){1}child[missing]child{node(2513){3}child[missing]child{node(25134){4}child[missing]
child[missing]}}child{node(2514){4}child{node(25143){3}child[missing]child[missing]}child[missing]}}child{node(254){4}
child{node(2541){1}child[missing]child[missing]child{node(25413){3}child[missing]child[missing]}}child{node(2543){3}
child{node(25431){1}child[missing]child[missing]child[missing]}child[missing]}}};
\end{scope}
\end{tikzpicture}
\newcommand{\circled}[1]{\raisebox{-1.5pt}{\tikzsetnextfilename{circled#1}\tikz\node[circle, draw, scale=.75, inner sep=.25mm] {#1};}}
\caption{The pyramid graph and its trees of self-avoiding walks starting from \protect\circled1 and \protect\circled2 respectively.}
\label{fig:sawtree}
\end{figure}
Recall that the spectral measure of a (possibly infinite) rooted graph $(T,v)$
is defined as follows. Assume that $T$ has bounded degree. Then the adjacency
operator $A_{T}:l^{2}(T)\rightarrow l^{2}(T)$ is bounded and self-adjoint,
hence it admits a spectral measure $P_{T}(X)$ ($X\subseteq\mathbb{R}$ Borel).
This is a projection-valued measure on $\mathbb{R}$ such that for any
polynomial $F(x)$ we have
\begin{equation}
F(A)=\int F(x)dP_{x}\tag{Sp
\end{equation}
where $P_{x}=P((-\infty,x))$. We define $\delta_{(T,v)}$, the \emph{spectral
measure of }$T$\emph{ at }$v$ by
\[
\delta_{(T,v)}(X)=\left\langle P_{T}(X)\chi_{v},P_{T}(X)\chi_{v}\right\rangle
=\left\langle P_{T}(X)\chi_{v},\chi_{v}\right\rangle \text{ \ (
X\subseteq\mathbb{R}\text{ Borel)
\]
where $\chi_{v}$ is the characteristic vector of $v$. It is easy to check that
$\delta_{(T,v)}$ is a probability measure supported on the spectrum of the
operator $A_{T}$. Also, by (Sp), for all $k\geq0$, the $k$-th moment of
$\delta_{(T,v)}$ equals
\[
\int x^{k}d\delta_{(T,v)}=\left\langle A^{k}\chi_{v},\chi_{v}\right\rangle
=a_{k}(T,v)
\]
where $a_{k}(T,v)$ is the number of returning walks of length $k$ starting at
$v$.
It turns out that the matching measure of a finite graph equals the average
spectral measure over its trees of self-avoiding walks.
\begin{Th}
\label{expected} Let $G$ be a finite graph and let $v$ be a vertex of $G$
chosen uniformly at random. Then
\[
\rho_{G}=\mathbb{E}_{v}\delta_{(T_{v}(G),v)}\text{.
\]
Equivalently, for all $k\geq0$, the $k$-th moment of $\rho_{G}$ equals the
expected number of returning walks of length $k$ in $T_{v}(G)$ starting at $v$.
\end{Th}
In particular, Theorem \ref{expected} gives one of the several known proofs
for the Heilmann-Lieb theorem. Indeed, spectral measures are real and the
spectral radius of a tree with degree bound $D$ is at most $2\sqrt{D-1}$.
To prove Theorem~\ref{expected} we need the following result of Godsil
\cite{god3} which connects the matching polynomial of the original graph $G$
and the tree of self-avoiding walks:
\begin{Th}
\label{saw-tree} \cite{god3} Let $G$ be a finite graph and $v$ be an arbitrary
vertex of $G$. Then
\[
\frac{\mu(G-v,x)}{\mu(G,x)}=\frac{\mu(T_{v}(G)-v,x)}{\mu(T_{v}(G),x)}.
\]
\end{Th}
We will also use two well-known facts which we gather in the following proposition:
\begin{Prop}
\label{prop} \cite{god3} (a) For any tree or forest $T$, the matching
polynomial $\mu(T,x)$ coincides with the characteristic polynomial $\phi(T,x)$
of the adjacency matrix of the tree $T$:
\[
\mu(T,x)=\phi(T,x).
\]
\medskip
\noindent(b) For any graph $G$, we have
\[
\mu^{\prime}(G,x)=\sum_{v\in V}\mu(G-v,x).
\]
\end{Prop}
\begin{proof}
[Proof of Theorem~\ref{expected}]First, let us use part (a) of
Proposition~\ref{prop} for the tree $T_{v}(G)$ and the forest $T_{v}(G)-v$:
\[
\frac{\mu(T_{v}(G)-v,x)}{\mu(T_{v}(G),x)}=\frac{\phi(T_{v}(G)-v,x)}{\phi
(T_{v}(G),x)}.
\]
On the other hand, for any graph $H$ and vertex $u$, we have
\[
\frac{\phi(H-u,x)}{\phi(H,x)}=x^{-1}\sum_{k=0}^{\infty}c_{k}(u)x^{-k},
\]
where $c_{k}(u)$ counts the number of walks of length $k$ starting and ending
at $u$. So this is exactly the moment generating function of the spectral
measure with respect to the vertex $u$. Putting together these with
Theorem~\ref{saw-tree} we see that
\[
\frac{\mu(G-v,x)}{\mu(G,x)}=\frac{\mu(T_{v}(G)-v,x)}{\mu(T_{v}(G),x)
=x^{-1}\sum_{k=0}^{\infty}a_{k}(v)x^{-k
\]
is the moment generating function of the spectral measure of the tree of
self-avoiding walks with respect to the vertex $v$.
Now let us consider the left hand side of Theorem~\ref{saw-tree}. Let us use
part (b) of Proposition~\ref{prop}:
\[
\mu^{\prime}(G,x)=\sum_{u\in V}\mu(G-u,x).
\]
This implies that
\[
\mathbb{E}_{v}\frac{\mu(G-v,x)}{\mu(G,x)}=\frac{1}{|G|}\frac{\mu^{\prime
}(G,x)}{\mu(G,x)}=x^{-1}\sum_{k=0}^{\infty}\mu_{k}x^{-k},
\]
where
\[
\mu_{k}=\frac{1}{|G|}\sum\lambda^{k},
\]
where the summation goes through the zeros of the matching polynomial. In
other words, $\mu_{k}$ is $k$-th moment of the matching measure defined by the
uniform distribution on the zeros of the matching polynomial. Putting
everything together we see that
\[
\mu_{k}=\mathbb{E}_{v}a_{k}(v).
\]
Since both $\rho_{G}$ and $\mathbb{E}_{v}\rho(v)$ are supported on $\left\{
\left\vert x\right\vert \leq\left\Vert A_{G}\right\Vert \right\} $, we get
that the two measures are equal.
\end{proof}
Now we define Benjamini--Schramm convergence.
\begin{Def}
For a finite graph $G$, a finite rooted graph $\alpha$ and a positive integer
$r$, let $\mathbb{P}(G,\alpha,r)$ be the probability that the $r$-ball
centered at a uniform random vertex of $G$ is isomorphic to $\alpha$. We say
that a graph sequence $(G_{n})$ of bounded degree is \emph{Benjamini--Schramm
convergent} if for all finite rooted graphs $\alpha$ and $r>0$, the
probabilities $\mathbb{P}(G_{n},\alpha,r)$ converge. Let $L$ be a vertex
transitive lattice. We say that {\emph{$(G_{n})$ Benjamini-Schramm converges to $L$}},
if for all positive integers $r$, $\mathbb{P}(G_{n},\alpha_{r},r)\rightarrow1$
where $\alpha_{r}$ is the $r$-ball in $L$.
\end{Def}
\begin{Ex}
Let us consider a sequence of boxes in $\mathbb{Z}^{d}$ where all sides
converge to infinity. This will be Benjamini--Schramm convergent graph
sequence since for every fixed $r$, we will pick a vertex which at least
$r$-far from the boundary with probability converging to $1$. For all these
vertices we will see the same neighborhood. This also shows that we can impose
arbitrary boundary condition, for instance periodic boundary condition means
that we consider the sequence of toroidal boxes. Boxes and toroidal boxes will
be Benjamini--Schramm convergent even together.
\end{Ex}
We prove the following generalization of Theorem~\ref{konv}.
\begin{Th}
\label{wc} Let $(G_{n})$ be a Benjamini--Schramm convergent bounded degree
graph sequence. Then the sequence of matching measures $\rho_{G_{n}}$ is
weakly convergent. If $(G_{n})$ Benjamini--Schramm converges to the vertex
transitive lattice $L$, then $\rho_{G_{n}}$ weakly converges to $\rho_{L}$ and
$\lim_{n\rightarrow\infty}\rho_{G_{n}}(\{x\})=\rho_{L}(\{x\})$ for all
$x\in\mathbb{R}$.
\end{Th}
\begin{Rem}
The first part of the theorem was first proved in \cite{ACFK}. The proof given
there relied on a general result on graph polynomials given in \cite{csi}. For completeness, we give an
alternate self-contained proof here.
\end{Rem}
We will use the following theorem of Thom \cite{thom}. See also
\cite{luckapprox} where this is used for Benjamini--Schramm convergent graph
sequences.
\begin{Th}[Thom]\label{Luck} Let $(q_{n}(z))$ be a sequence of monic polynomials with
integer coefficients. Assume that all zeros of all $q_{n}(z)$ are at most $R$
in absolute value. Let $\rho_{n}$ be the probability measure of uniform
distribution on the roots of $q_{n}(z)$. Assume that $\rho_{n}$ weakly
converges to some measure $\rho$. Then for all $\theta\in\mathbb{C}$ we have
\[
\lim_{n\rightarrow\infty}\rho_{n}(\{\theta\})=\rho(\{\theta\}).
\]
\end{Th}
\begin{proof}[Proof of Theorem~\ref{konv} and \ref{wc}]
For $k\geq0$ let
\[
\mu_{k}(G)=\int z^{k}\,d\rho_{G}(z)
\]
be the $k$-th moment of $\rho_{G}$. By Theorem~\ref{expected} we have
\[
\mu_{k}(G)=\mathbb{E}_{v}a_{k}(G,v)
\]
where $a_{k}(G,v)$ denotes the number of closed walks of length $k$ of the
tree $T_{v}(G)$ starting and ending at the vertex $v$.
Clearly, the value of $a_{k}(G,v)$ only depends on the $k$-ball centered at
the vertex $v$. Let $TW(\alpha)=a_{k}(G,v)$ where the $k$-ball centered at $v$
is isomorphic to $\alpha$. Note that the value of $TW(\alpha)$ depends only on the rooted graph $\alpha$ and does not depend on $G$.
Let $\mathcal{N}_{k}$ denote the set of possible $k$-balls in $G$. The size of
$\mathcal{N}_{k}$ and $TW(\alpha)$ are bounded by a function of $k$ and the largest
degree of $G$. By the above, we have
\[
\mu_{k}(G)=\mathbb{E}_{v}a_{k}(G,v)=\sum_{\alpha\in\mathcal{N}_{k}
\mathbb{P}(G,\alpha,k)\cdot TW(\alpha).
\]
Since $(G_{n})$ is Benjamini--Schramm convergent, we get that for every fixed
$k$, the sequence of $k$-th moments $\mu_{k}(G_{n})$ converges. The same holds
for $\int q(z)\,d\rho_{G_{n}}(z)$ where $q$ is any polynomial. By the
Heilmann--Lieb theorem, $\rho_{G_{n}}$ is supported on $[-2\sqrt{D-1
,2\sqrt{D-1}]$ where $D$ is the absolute degree bound for $G_{n}$. Since every
continuous function can be uniformly approximated by a polynomial on
$[-2\sqrt{D-1},2\sqrt{D-1}]$, we get that the sequence $(\rho_{G_{n}})$ is
weakly convergent.
Assume that $(G_{n})$ Benjamini--Schramm converges to $L$. Then for all
$k\geq0$ we have $\mathbb{P}(G_{n},\alpha_{k},k)\rightarrow 1$ where
$\alpha_{k}$ is the $k$-ball in $L$, which implies
\[
\lim_{n\rightarrow\infty}\mu_{k}(G_{n})=\lim_{n\rightarrow\infty}\sum
_{\alpha\in\mathcal{N}_{k}}\mathbb{P}(G_n,\alpha,k)\cdot TW(\alpha
)=TW(\alpha_{k})=a_{k}(L,v)
\]
where $v$ is any vertex in $L$. This means that all the moments of $\rho_{L}$
and $\lim\rho_{G_{n}}$ are equal, so $\lim\rho_{G_{n}}=\rho_{L}$.
Since the matching polynomial is monic with integer coefficients, Theorem
\ref{Luck} gives $\lim_{n\rightarrow\infty}\rho_{G_{n}}(\{x\})=\rho
_{L}(\{x\})$ for all $x\in\mathbb{R}$.
\end{proof}
\section{The function $\lambda_{G}(p)$}
\label{entropy-function}
Let $G$ be a finite graph, and recall that $|G|$ denotes the number of vertices of
$G$, and $m_{k}(G)$ denotes the number of $k$-matchings ($m_{0
(G)=1$). Let $t$ be the activity, a non-negative real number, and
\[
M(G,t)=\sum_{k=0}^{\lfloor|G|/2\rfloor}m_{k}(G)t^{k},
\]
We call $M(G,t)$ the matching generating function or the partition function of
the monomer-dimer model. Clearly, it encodes the same information as the
matching polynomial. Let
\[
p(G,t)=\frac{2t\cdot M^{\prime}(G,t)}{|G|\cdot M(G,t)},
\]
and
\[
F(G,t)=\frac{\ln M(G,t)}{|G|}-\frac{1}{2}p(G,t) \ln(t).
\]
Note that
\[
\tilde{\lambda}(G)=F(G,1)
\]
is called the monomer-dimer free energy.
The function $p=p(G,t)$ is a strictly monotone increasing function which maps
$[0,\infty)$ to $[0,p^{*})$, where $p^{*}=\frac{2\nu(G)}{|G|}$, where $\nu(G)
$ denotes the number of edges in the largest matching. If $G$ contains a
perfect matching, then $p^{*}=1$. Therefore, its inverse function $t=t(G,p)$
maps $[0,p^{*})$ to $[0,\infty)$. (If $G$ is clear from the context, then we will simply write $t(p)$ instead of $t(G,p)$.) Let
\[
\lambda_{G}(p)=F(G,t(p))
\]
if $p<p^{*}$, and $\lambda_{G}(p)=0$ if $p>p^{*}$. Note that we have not
defined $\lambda_{G}(p^{*})$ yet. We simply define it as a limit:
\[
\lambda_{G}(p^{*})=\lim_{p\nearrow p^{*}}\lambda_{G}(p).
\]
We will show that this limit exists, see part (d) of Proposition~\ref{asymp}.
Later we will extend the definition of $p(G,t), F(G,t)$ and $\lambda_{G}(p) $
to infinite lattices $L$.
The intuitive meaning of $\lambda_{G}(p)$ is the following. Assume that we
want to count the number of matchings covering $p$ fraction of the vertices.
Let us assume that it makes sense: $p=\frac{2k}{|G|}$, and so we wish to count
$m_{k}(G)$. Then
\[
\lambda_{G}(p)\approx\frac{\ln m_{k}(G)}{|G|}.
\]
The more precise formulation of this statement will be given in
Proposition~\ref{asymp}. To prove this proposition we need some preparation.
\bigskip
We will use the following theorem of Darroch.
\begin{Lemma}
[Darroch's rule \cite{dar}]Let $P(x)=\sum_{k=0}^{n}a_{k}x^{k}$ be a
polynomial with only positive coefficients and real zeros. If
\[
k-\frac{1}{n-k+2}<\frac{P^{\prime}(1)}{P(1)}<k+\frac{1}{k+2},
\]
then $k$ is the unique number for which $a_{k}=\max(a_{1},a_{2},\dots, a_{n})
$. If, on the other hand,
\[
k+\frac{1}{k+2}<\frac{P^{\prime}(1)}{P(1)}<k+1-\frac{1}{n-k+1},
\]
then either $a_{k}$ or $a_{k+1}$ is the maximal element of $a_{1},a_{2},\dots,
a_{n}$.
\end{Lemma}
\begin{Prop}
\label{asymp} Let $G$ be a finite graph. \newline(a) Let $nG$ be $n$ disjoint
copies of $G$. Then
\[
\lambda_{G}(p)=\lambda_{nG}(p).
\]
(b) If $p<p^{*}$, then
\[
\frac{d}{dp}\lambda_{G}(p)=-\frac{1}{2}\ln t(p).
\]
(c) The limit
\[
\lambda_{G}(p^*)=\lim_{p\nearrow p^{*}}\lambda_{G}(p)
\]
exists.
\newline(d) Let $k\leq\nu(G)$ and $p=\frac{2k}{|G|}$. Then
\[
\left|\lambda_{G}(p)-\frac{\ln m_{k}(G)}{|G|}\right| \leq\frac{\ln|G|
{|G|}.
\]
(e) Let $k=\nu(G)$, then for $p^{*}=\frac{2k}{|G|}$ we have
\[
\lambda_{G}(p^{*})=\frac{\ln m_{k}(G)}{|G|}.
\]
(f) If for some function $f(p)$ we have
\[
\lambda_{G}(p)\geq f(p)+o_{|G|}(1)
\]
then
\[
\lambda_{G}(p)\geq f(p).
\]
\end{Prop}
\begin{proof}
\noindent(a) Let $nG$ be the disjoint union of $n$ copies of $G$. Note that
\[
M(nG,t)=M(G,t)^{n
\]
implying that $p(nG,t)=p(G,t)$ and $\lambda_{nG}(p)=\lambda_{G}(p)$. \medskip
\noindent(b) Since
\[
\lambda_{G}(p)=\frac{\ln M(G,t)}{|G|}-\frac{1}{2}p(G,t) \ln(t)
\]
we have
\[
\frac{d\lambda_{G}(p)}{dp}=\left( \frac{1}{|G|}\cdot\frac{M^{\prime
(G,t)}{M(G,t)}\cdot\frac{dt}{dp}-\frac{1}{2}\left( \ln(t)+p\cdot\frac{1
{t}\cdot\frac{dt}{dp}\right) \right) =-\frac{1}{2}\ln(t),
\]
since
\[
\frac{1}{|G|}\cdot\frac{M^{\prime}(G,t)}{M(G,t)}=\frac{p}{2t
\]
by definition. \medskip
\noindent(c) From $\frac{d}{dp}\lambda_{G}(p)=-\frac{1}{2}\ln t(p)$ we see that if $p>p(G,1)$, the function $\lambda_G(p)$ is
monotone decreasing. (Note that we also see
that $\lambda_{G}(p)$ is a concave-down function.) Hence
\[
\lim_{p\nearrow p^{*}}\lambda_{G}(p)=\inf_{p>p(G,1)}\lambda_{G}(p).
\]
\medskip
\noindent(d) First, let us assume that $k<\nu(G)$. In case of $k=\nu(G)$, we
will slightly modify our argument. Let $t=t(p)$ be the value for which
$p=p(G,t)$. The polynomial
\[
P(G,x)=M(G,tx)=\sum_{j=0}^{n}m_{j}(G)t^{j}x^{j
\]
considered as a polynomial in variable $x$, has only real zeros by
Theorem~\ref{Hei}. Note that
\[
k=\frac{p|G|}{2}=\frac{P^{\prime}(G,1)}{P(G,1)}.
\]
Darroch's rule says that in this case $m_{k}(G)t^{k}$ is the unique maximal
element of the coefficient sequence of $P(G,x)$. In particular
\[
\frac{M(G,t)}{|G|}\leq m_{k}(G)t^{k}\leq M(G,t).
\]
Hence
\[
\lambda_{G}(p)-\frac{\ln|G|}{|G|}\leq\frac{\ln m_{k}(G)}{|G|}\leq\lambda
_{G}(p).
\]
Hence in case of $k<\nu(G)$, we are done. \medskip
If $k=\nu(G)$, then let $p$ be arbitrary such that
\[
k-\frac{1}{2}<\frac{p|G|}{2}<k.
\]
Again we can argue by Darroch's rule as before that
\[
\lambda_{G}(p)-\frac{\ln|G|}{|G|}\leq\frac{\ln m_{k}(G)}{|G|}\leq\lambda
_{G}(p).
\]
Since this is true for all $p$ sufficiently close to $p^{*}=\frac{2\nu
(G)}{|G|}$ and
\[
\lambda_{G}(p^{*})=\lim_{p\nearrow p^{*}}\lambda_{G}(p),
\]
we have
\[
\left| \frac{\ln m_{k}(G)}{|G|}-\lambda_{G}(p^{*})\right| \leq\frac{\ln
|G|}{|G|
\]
in this case too. \medskip
\noindent(e) By part (a) we have $\lambda_{nG}(p)=\lambda_{G}(p)$. Note also
that if $k=\nu(G)$, then $m_{nk}(nG)=m_{k}(G)^{n}$. Applying the bound from
part (d) to the graph $nG$, we obtain that
\[
\left| \frac{\ln m_{k}(G)}{|G|}-\lambda_{G}(p^{*})\right| \leq\frac{\ln
|nG|}{|nG|}.
\]
Since
\[
\frac{\ln|nG|}{|nG|}\to0
\]
as $n\to\infty$, we get that
\[
\lambda_{G}(p^{*})=\frac{\ln m_{k}(G)}{|G|}.
\]
\medskip
\noindent(f) This is again a trivial consequence of $\lambda_{nG
(p)=\lambda_{G}(p)$.
\end{proof}
Our next aim is to extend the definition of the function $\lambda_{G}(p)$ for
infinite lattices $L$. We also show an efficient way of computing its values
if $p$ is sufficiently separated from $p^{*}$.
The following theorem was known in many cases for thermodynamic limit.
\begin{Th}
\label{entropy} Let $(G_{n})$ be a Benjamini--Schramm convergent
sequence of bounded degree graphs. Then the sequences of functions\newline(a)
\[
p(G_{n},t),
\]
(b)
\[
\frac{\ln M(G_{n},t)}{|G_{n}|
\]
converge to strictly monotone increasing continuous functions on the interval
$[0,\infty)$. \newline If, in addition, every $G_{n}$ has a perfect matching
then the sequences of functions \newline(c)
\[
t(G_{n},p),
\]
(d)
\[
\lambda_{G_{n}}(p)
\]
are convergent for all $0\leq p<1$.
\end{Th}
\begin{Rem}
In part (c), we used the extra condition to ensure that $p^{*}=1$ for all
these graphs. We mention that H. Nguyen and K. Onak \cite{ngu}, and
independently G. Elek and G. Lippner \cite{ele} proved that for a
Benjamini--Schramm convergent graph sequence $(G_{n})$, the following limit
exits:
\[
\lim_{n\to\infty}\frac{2\nu(G_{n})}{|G_{n}|}=\lim_{n\to\infty}p^{*}(G_{n}).
\]
In particular, one can extend part (c) to graph sequences without perfect
matchings. Since we are primarily interested in lattices with perfect matchings,
we leave it to the Reader.
\end{Rem}
To prove Theorem~\ref{entropy}, we essentially repeat an argument of the paper
\cite{ACFK}.
\begin{proof}
[Proof of Theorem~\ref{entropy}]First we prove part (a) and (b). For a graph
$G$ let $S(G)$ denote the set of zeros of the matching polynomial $\mu(G,x)$,
then
\[
M(G,t)=\prod_
\genfrac{}{}{0pt}{}{\lambda\in S(G) }{\lambda>0
}(1+\lambda^{2}t)=\prod_{\lambda\in S(G)}(1+\lambda^{2}t)^{1/2}.
\]
Then
\[
\ln M(G,t)=\sum_{\lambda\in S(G)}\frac{1}{2}\ln\left( 1+\lambda^{2}t\right)
.
\]
By differentiating both sides we get that
\[
\frac{M^{\prime}(G,t)}{M(G,t)}=\sum_{\lambda\in S(G)}\frac{1}{2}\frac
{\lambda^{2}}{1+\lambda^{2}t}.
\]
Hence
\[
p(G,t)=\frac{2t\cdot M^{\prime}(G,t)}{|G|\cdot M(G,t)}=\frac{1}{|G|
\sum_{\lambda\in S(G)}\frac{\lambda^{2} t}{1+\lambda^{2}t}=\int\frac{tz^{2
}{1+tz^{2}}\, d\rho_{G}(z).
\]
Similarly,
\[
\frac{\ln M(G,t)}{|G|}=\frac{1}{|G|}\sum_{\lambda\in S(G)}\frac{1}{2
\ln\left( 1+\lambda^{2}t\right) =\int\frac{1}{2}\ln\left( 1+tz^{2}\right)
\, d\rho_{G}(z).
\]
Since $(G_{n})$ is a Benjamini--Schramm convergent sequence of bounded degree graphs,
the sequence $(\rho_{G_{n}})$ weakly converges to some $\rho^{*}
$ by Theorem~\ref{wc}. Since both functions
\[
\frac{tz^{2}}{1+tz^{2}}\ \ \ \ \mbox{and}\ \ \ \ \frac{1}{2}\ln\left(
1+tz^{2}\right)
\]
are continuous, we immediately obtain that
\[
\lim_{n\to\infty}p(G_{n},t)=\int\frac{tz^{2}}{1+tz^{2}} \, d\rho^{*}(z),
\]
and
\[
\lim_{n\to\infty}\frac{\ln M(G_{n},t)}{|G_{n}|}=\int\frac{1}{2}\ln\left(
1+tz^{2}\right) \, d\rho^{*}(z).
\]
Note that both functions
\[
\frac{tz^{2}}{1+tz^{2}}\ \ \ \ \mbox{and}\ \ \ \ \frac{1}{2}\ln\left(
1+tz^{2}\right)
\]
are strictly monotone increasing continuous functions in the variable $t$.
Thus their integrals are also strictly monotone increasing continuous
functions. \bigskip
To prove part (c), let us introduce the function
\[
p(L,t)=\int\frac{tz^{2}}{1+tz^{2}} \, d\rho^{*}(z).
\]
We have seen that $p(L,t)$ is a strictly monotone increasing continuous
function, and equals $\lim_{n\to\infty}p(G_{n},t)$. Since for all $G_{n}$,
$p^{*}(G_{n})=1$, we have $\lim_{t \to\infty}p(G_{n},t)=1$ for all $n$. This
means that $\lim_{t \to\infty}p(L,t)=1$. Hence we can consider inverse
function $t(L,p)$ which maps $[0,1)$ to $[0,\infty)$. We show that
\[
\lim_{n\to\infty}t(G_{n},p)=t(L,p)
\]
pointwise. Assume by contradiction that this is not the case. This means that
for some $p_{1}$, there exists an $\varepsilon$ and an infinite sequence
$n_{i}$ for which
\[
\left| t(L,p_{1})-t(G_{n_{i}},p_{1})\right| \geq\varepsilon.
\]
We distinguish two cases according to \newline(i) there exists an infinite
sequence $(n_{i})$ for which
\[
t(G_{n_{i}},p_{1})\geq t(L,p_{1})+\varepsilon,
\]
or (ii) there exists an infinite sequence $(n_{i})$ for which
\[
t(G_{n_{i}},p_{1})\leq t(L,p_{1})-\varepsilon.
\]
In the first case, let $t_{1}=t(L,p_{1})$, $t_{2}=t_{1}+\varepsilon$ and
$p_{2}=p(L,t_{2})$. Clearly, $p_{2}>p_{1}$. Note that
\[
t(G_{n_{i}},p_{1})\geq t(L,p_{1})+\varepsilon=t_{2
\]
and $p(G_{n_{i}},t)$ are monotone increasing functions, thus
\[
p(G_{n_{i}},t_{2})\leq p(G_{n_{i}},t(G_{n_{i}},p_{1}))=p_{1}=p_{2
-(p_{2}-p_{1})=p(L,t_{2})-(p_{2}-p_{1}).
\]
This contradicts the fact that
\[
\lim_{n\to\infty}p(G_{n_{i}},t_{2})=p(L,t_{2}).
\]
In the second case, let $t_{1}=t(L,p_{1})$, $t_{2}=t_{1}-\varepsilon$ and
$p_{2}=p(L,t_{2})$. Clearly, $p_{2}<p_{1}$. Note that
\[
t(G_{n_{i}},p_{1})\leq t(L,p_{1})-\varepsilon=t_{2
\]
and $p(G_{n_{i}},t)$ are monotone increasing functions, thus
\[
p(G_{n_{i}},t_{2})\geq p(G_{n_{i}},t(G_{n_{i}},p_{1}))=p_{1}=p_{2
+(p_{1}-p_{2})=p(L,t_{2})+(p_{1}-p_{2}).
\]
This again contradicts the fact that
\[
\lim_{n\to\infty}p(G_{n_{i}},t_{2})=p(L,t_{2}).
\]
Hence $\lim_{n\to\infty}t(G_{n},p)=t(L,p)$. \medskip
Finally, we show that $\lambda_{G_{n}}(p)$ converges for all $p$. Let
$t=t(L,p) $, and
\[
\lambda_{L}(p)=\lim_{n\to\infty}\frac{\ln M(G_{n},t)}{|G_{n}|}-\frac{1}{2
p\ln(t).
\]
Note that
\[
\lambda_{G_{n}}(p)=\frac{\ln M(G_{n},t_{n})}{|G_{n}|}-\frac{1}{2}p\ln(t_{n}),
\]
where $t_{n}=t(G_{n},p)$. We have seen that $\lim_{n\to\infty}t_{n}=t$. Hence
it is enough to prove that the functions
\[
\frac{\ln M(G_{n},u)}{|G_{n}|
\]
are equicontinuous. Let us fix some $u_{0}$ and let
\[
H(u_{0},u)=\max_{z\in[-2\sqrt{D-1},2\sqrt{D-1}]}\left| \frac{1}{2}\ln\left(
1+u_{0}z^{2}\right) -\frac{1}{2}\ln\left( 1+uz^{2}\right) \right| .
\]
Clearly, if $|u-u_{0}|\leq\delta$ for some sufficiently small $\delta$, then
$H(u_{0},u)\leq\varepsilon$, and
\[
\left| \frac{\ln M(G_{n},u)}{|G_{n}|}-\frac{\ln M(G_{n},u_{0})}{v(G_{n
)}\right| =\left| \int\frac{1}{2}\ln\left( 1+u_{0}z^{2}\right) \,
d\rho_{G_{n}}(z)-\int\frac{1}{2}\ln\left( 1+uz^{2}\right) \, d\rho_{G_{n
}(z)\right| \leq
\]
\[
\leq\int\left| \frac{1}{2}\ln\left( 1+u_{0}z^{2}\right) -\frac{1}{2
\ln\left( 1+uz^{2}\right) \right| \, d\rho_{G_{n}}(z)\leq\int H(u,u_{0})\,
d\rho_{G_{n}}(z)\leq\varepsilon.
\]
This completes the proof of the convergence of $\lambda_{G_{n}}(p)$.
\end{proof}
\begin{Def}
Let $L$ be an infinite lattice and $(G_{n})$ be a sequence of finite graphs
which is Benjamini--Schramm convergent to $L$. For instance, $G_{n}$ can be
chosen to be an exhaustion of $L$. Then the sequence of measures $(\rho
_{G_{n}})$ weakly converges to some measure which we will call $\rho_{L}$, the
matching measure of the lattice $L$. For $t>0$, we can introduce
\[
p(L,t)=\int\frac{tz^{2}}{1+tz^{2}}\, d\rho_{L}(z)
\]
and
\[
F(L,t)=\int\frac{1}{2}\ln\left( 1+tz^{2}\right) \, d\rho_{L}(z)-\frac{1
{2}p(L,t) \ln(t).
\]
If the lattice $L$ contains a perfect matching, then we can choose $G_{n}$
such that all $G_{n}$ contain a perfect matching. Then $p(L,t)$ maps
$[0,\infty)$ to $[0,1)$ in a monotone increasing way, and we can consider its
inverse function $t(L,p)$. Finally, we can introduce
\[
\lambda_{L}(p)=F(L,t(L,p))
\]
for all $p\in[0,1)$. We will define $\lambda_{L}(1)$ as
\[
\lambda_{L}(1)=\lim_{p\nearrow1}\lambda_{L}(p).
\]
\end{Def}
\begin{Rem}
\label{Mayer} In the literature, the so-called Mayer series are computed for
various lattices $L$:
\[
p(L,t)=\sum_{n=1}^{\infty}a_{n}t^{n
\]
for small enough $t$. Let us compare it with
\[
p(L,t)=\int\frac{tz^{2}}{1+tz^{2}}\, d\rho_{L}(z)=\int\left( \sum
_{n=1}^{\infty}(-1)^{n+1}z^{2n}t^{n}\right) \, d\rho_{L}(z)=\sum
_{n=1}^{\infty}(-1)^{n+1}\left( \int z^{2n}d\rho_{L}(z)\right) t^{n}.
\]
Hence if we introduce the moment sequence
\[
\mu_{k}=\int z^{k}d\rho_{L}(z),
\]
we see that
\[
\mu_{2n}=\int z^{2n}d\rho_{L}(z)=(-1)^{n+1}a_{n}.
\]
Note that $\mu_{0}=1$ and $\mu_{2n-1}=0$ since the matching measures are
symmetric to $0$. Since the support of the measure $\rho_{L}$ lie in the
interval $[-2\sqrt{D-1},2\sqrt{D-1}]$, we see that the Mayer series converges
whenever $|t|<\frac{1}{4(D-1)}$. We also would like to point out that the
integral is valid for all $t>0$, while the Mayer series does not converge if
$t$ is 'large'.
\end{Rem}
\subsection{Computation of the monomer-dimer free energy}
The monomer-dimer free energy of a lattice $L$ is $\tilde{\lambda}(L)=F(L,1)$.
Its computation can be carried out exactly the same way as we proved its
existence: we use that
\[
\tilde{\lambda}(L)=F(L,1)=\int\frac{1}{2}\ln\left( 1+z^{2}\right) \,d\rho_{L}(z).
\]
Assume that we know the moment sequence $(\mu_{k})$ for $k\leq N$. Then let us
choose a polynomial of degree at most $N$, which uniformly approximates the
function
\[
\frac{1}{2}\ln\left( 1+z^{2}\right)
\]
on the interval $[-2\sqrt{D-1},2\sqrt{D-1}]$, where $D$ is the coordination
number of $L$. A good polynomial approximation can be found by Remez's
algorithm. Assume that we have a polynomial
\[
q(z)=\sum_{k=0}^{N}c_{k}z^{k
\]
for which
\[
\left\vert \frac{1}{2}\ln\left( 1+z^{2}\right) -q(z)\right\vert
\leq\varepsilon
\]
for all $z\in\lbrack-2\sqrt{D-1},2\sqrt{D-1}]$. Then
\[
\left\vert \tilde{\lambda}(L)-\int q(z)\,d\rho_{L}(z)\right\vert \leq\int\left\vert
\frac{1}{2}\ln\left( 1+z^{2}\right) -q(z)\right\vert d\rho_{L
(z)\leq\varepsilon,
\]
and
\[
\int q(z)\,d\rho_{L}(z)=\sum_{k=0}^{N}c_{k}\mu_{k}.
\]
Hence
\[
\left\vert \tilde{\lambda}(L)-\sum_{k=0}^{N}c_{k}\mu_{k}\right\vert \leq\varepsilon.
\]
How can we compute the moment sequence $(\mu_{k})$? One way is to use its
connection with the Mayer series (see Remark~\ref{Mayer}). A good source of Mayer
series coefficients is the paper of P. Butera and M. Pernici \cite{bupe},
where they computed $a_{n}$ for $1\leq n\leq24$ for various lattices. (More
precisely, they computed $d_{n}=a_{n}/2$ since they expanded the function
$\rho(t)=p(t)/2$.) This means that we know $\mu_{k}$ for $k\leq49$ for
these lattices. The other strategy to compute the moment sequence is to use
its connection with the number of closed walks in the self-avoiding walk tree.
Since the moment sequence is missing for the honeycomb lattice (hexagonal lattice), we
computed the first few elements of the moment sequence for this lattice:
\medski
\[
1, 0, 3, 0, 15, 0, 87, 0, 543, 0, 3543, 0, 23817, 0, 163551, 0, 1141119, 0,
8060343, 0,
\]
\[
57494385, 0, 413383875, 0, 2991896721, 0, 21774730539, 0, 159227948055, 0,
\]
\[
1169137211487, 0, 8615182401087, 0, 63683991513351, 0, 472072258519041, 0,
\]
\[
3508080146139867, 0, 26127841824131313, 0, 194991952493587371, 0,
\]
\[
1457901080870060919, 0, 10918612274039599755, 0, 81898043907874542705
\]
\bigskip
The following table contains some numerical results. The bound on the error terms are rigorous.
\begin{center
\begin{tabular}[c]{|c|c|c|c|c|}\hline
Lattice & $\tilde{\lambda}(L)$ & Bound on error & $p(L,1)$ & Bound on error \\\hline
2d & $0.6627989725$ & $3.72\cdot10^{-8}$ & $0.638123105$ &
$5.34\cdot10^{-7}$ \\\hline
3d & $0.7859659243$ & $9.89\cdot10^{-7}$ & $0.684380278$ &
$1.14\cdot10^{-5}$ \\\hline
4d & $0.8807178880$ & $5.92\cdot10^{-6}$ & $0.715846906$ &
$5.86\cdot10^{-5}$ \\\hline
5d & $0.9581235802$ & $4.02\cdot10^{-5}$ & $0.739160383$ &
$3.29\cdot10^{-4}$ \\\hline
6d & $1.0237319240$ & $1.24\cdot10^{-4}$ & $0.757362382$ &
$8.91\cdot10^{-4}$ \\\hline
7d & $1.0807591953$ & $3.04\cdot10^{-4}$ & $0.772099489$ &
$1.95\cdot10^{-3}$ \\\hline
hex &$0.58170036638$ & $1.56\cdot10^{-9}$ &$0.600508638$ &
$2.65\cdot10^{-8}$ \\\hline
\end{tabular}
\end{center}
\section{Density function of matching measures.}
\label{density}
It is natural problem to investigate the matching measure. One particular
question is whether it is atomless or not. In general, $\rho_{L}$ can contain
atoms. For instance, if $G$ is a finite graph then clearly $\rho_{G}$ consists
of atoms. On the other hand, it can be shown that for all lattices in Table 1,
the measure $\rho_{L}$ is atomless. We use the following lemmas.
\begin{figure}[h!]
\tikzsetnextfilename{2dgridmm} \begin{tikzpicture}
\begin{axis}[small, height=4cm, width=12cm, xmin=-5, xmax=5, ymin=0, ymax=0.25, axis lines=left, /pgf/number format/fixed]
\addplot[black!80, fill=black, fill opacity=0.15] table {2dgridmm.dat};
\end{axis}
\end{tikzpicture}
\caption{An approximation for the matching measure of $\mathbb{Z}^{2}$,
obtained by smoothing the matching measure of the finite grid $C_{10}\times
P_{100}$ by convolution with a triweight kernel.
\label{fig:2dgridmm
\end{figure}
We will only need part (a) of the following lemma, we only give part (b) for
the sake of completeness.
\begin{Lemma}
\cite{god3,hei} \label{pathcover} (a) The maximum multiplicity of a zero of
$\mu(G,x)$ is at most the number of vertex-disjoint paths required to
cover $G$. \medskip
\noindent(b) The number of distinct zeros of $\mu(G,x)$ is at least the length of
the longest path in $G$.
\end{Lemma}
The following lemma is a deep result of C. Y. Ku and W. Chen \cite{ku}.
\begin{Lemma}
\label{transitive} \cite{ku} If $G$ is a finite connected vertex transitive
graph, then all zeros of the matching polynomial are distinct.
\end{Lemma}
Now we are ready to give a generalization of Theorem~\ref{atomless}.
\begin{Th}
\label{atomless-general} Let $L$ be a lattice satisfying one of the following
conditions. \medskip
\noindent(a) The lattice $L$ can be obtained as a Benjamini--Schramm limit of
a finite graph sequence $G_{n}$ such that $G_{n}$ can be covered by
$o(|G_{n}|)$ disjoint paths. \medskip
\noindent(b) The lattice $L$ can be obtained as a Benjamini--Schramm limit of
connected vertex transitive finite graphs.
\medskip
Then the matching measure $\rho_{L}$ is atomless.
\end{Th}
\begin{proof}
We prove the two statements together. Let $\mbox{mult}(G_{n},\theta)$ denote
the multiplicity of $\theta$ as a zero of $\mu(G_{n},x)$. Then by
Theorem~\ref{Luck} we have
\[
\rho_{L}(\{\theta\})=\lim_{n\rightarrow\infty}\frac{\mbox{mult}(G_{n},\theta
)}{|G_{n}|}.
\]
Note that by Lemma~\ref{pathcover} we have $\mbox{mult}(G_{n},\theta)$ is at
most the number of paths required to cover the graph $G_{n}$. In case of
connected vertex transitive graphs $G_{n}$, we have $\mbox{mult}(G_{n
,\theta)=1$ by Lemma~\ref{transitive}. This means that in both cases $\rho
_{L}(\{\theta\})=0$.
\end{proof}
\begin{proof}
[Proof of Theorem~\ref{atomless}]Note that $\mathbb{Z}^{d}$ satisfies both
conditions of Theorem~\ref{atomless-general} by taking boxes or using part (b),
taking toroidal boxes.
\end{proof}
\begin{center}
\begin{figure}[h!]
\tikzsetnextfilename{3dgridmm} \begin{tikzpicture}
\begin{axis}[small, height=4cm, width=12cm, xmin=-5, xmax=5, ymin=0, ymax=0.25, axis lines=left, /pgf/number format/fixed]
\addplot[black!80, fill=black, fill opacity=0.15] table {3dgridmm.dat};
\end{axis}
\end{tikzpicture}
\caption{An approximation for the matching measure of $\mathbb{Z}^{3}$.
Working with reasonably sized finite grids would have been computationally too
expensive, so this time we took the $L_{2}$ projection of the infinite measure
to the space of degree 48 polynomials which can be calculated from the
sequence of moments.
\label{fig:3dgridmm
\end{figure}
\end{center}
|
\section{introduction}
Let $B$ denote an $n$-dimensional compact Riemannian manifold, let $\KK\in\{\R,\C,\HH\}$, and let $k\in\Z^+$. To classify the $\KK^k$-vector bundles over $B$, one considers the Grassmannian of $k$-dimensional $\KK$-linear subspaces in $\KK^N$, denoted $\GG$,
for sufficiently large $N$. The \emph{universal vector bundle} over $\GG$ has total space $\{(\sigma,V)\mid\sigma\in\GG, V\in\sigma\}$ and projection map $(\sigma,V)\mapsto\sigma$. It is well-known that every $\KK^k$-bundle over $B$ is the pull-back of this universal vector bundle via some \emph{classifying map} $\varphi:B\ra\GG$, and that the homotopy class of $\varphi$ determines the isomorphism class of the pulled-back bundle.
This universal vector bundle has a natural connection defined such that the covariant derivative of a section $t\mapsto(\sigma(t),V(t))$ equals the section $t\mapsto\left(\sigma(t),V'(t)^{\sigma(t)}\right)$, where the superscript denotes the orthogonal projection onto the subspace. Narasimhan and Ramanan proved in~\cite{NR} (see also~\cite{Schlafly}) that for sufficiently large $N$, this connection is \emph{universal} in the sense that every connection in the pulled-back bundle over $B$ can be obtained as a pullback of this universal connection by correctly choosing $\varphi$ within the homotopy class representing the bundle. This theorem has seen many abstract applications within mathematics and physics, but to the best of our knowledge, it has never been used to prove or disprove the existence of connections in vector bundles with specific desirable geometric properties.
Let $\varphi:B\ra\GG$ be an explict classifying map, let $\pi_E:E\ra B$ denote the pulled-back bundle, let $\nabla$ denote the pulled-back connection, and let $R^\nabla$ denote its curvature tensor. Let $\pi_1:E^1\ra B$ denote the sphere-bundle over $B$ formed from all unit-length vectors in $E$.
The first property we will study is \emph{fatness}, which was defined by Weinstein in~\cite{Weinstein} and studied by many authors (see~\cite{Ziller} for a survey). To define fatness, it is useful to select a metric on $B$ and a rotationally symmetric ``fiber metric'' on $\KK^k$. Together with $\nabla$, these choices induce a unique \emph{connection metric} on $E$ (and thus also on $E^1$). The connection $\nabla$ called \emph{fat} if the sectional curvatures of all $\pi_1$-vertizontal planes are positive. Notice that we're calling $\nabla$ \emph{fat} when the induced connection in the sphere bundle is fat as defined by Weinstein. This fatness condition turns out not to depend on the choice of base metric or fiber metric, so fatness is a property only of the connection.
In Section~\ref{S:calc}, we will describe the general condition on $\varphi$ under which $\nabla$ is fat. In the special case where $k=1$ and either $\KK=\C$ (so that $\varphi:B\ra\CP^{N-1}$) or $\KK=\HH$ (so that $\varphi:B\ra\HP^{N-1}$), fatness turns out to be equivalent to $\varphi(B)$ having bounded Wirtinger angles, defined as follows:
\begin{definition}\label{DD} Let $\KK\in\{\C,\HH\}$ and let $\mathcal{J}$ denote...
\begin{itemize}
\item[...]$\{J\}$ if $\KK=\C$, where $J$ denotes the standard almost complex structure on $\CP^N$,
\item[...]$\{I,J,K\}$ if $\KK=\HH$, where $\{I,J,K\}$ denotes a local coordinate expressions for the standard almost quaternionic structure on $\HP^N$.
\end{itemize}
Let $S\subset\KK\mathbb{P}^N$ be an immersed submanifold, $p\in S$ and $X\in T_pS$. The \emph{Wirtinger angle}, $\theta(X)\in[0,\pi/2]$, is the maximum angle that a vector in $\text{span}_\R\{\mathfrak{J}X\mid\mathfrak{J}\in\mathcal{J}\}$ makes with $T_pS$. If $\theta(X)<\pi/2$ for all $X\in TS$, then $S$ is said to have \emph{bounded Wirtinger angles}.
\end{definition}
\begin{theorem} \label{P:fat}If $k=1$ and $\KK\in\{\C,\HH\}$, then $\nabla$ is fat if and only if $\varphi$ is an immersion with bounded Wirtinger angles.
\end{theorem}
There are several well-studied conditions that imply bounded Wirtinger angles. For example, an immersion is called \emph{K\"ahler} (when $\KK=\C$) or \emph{quaternionic} (when $\KK=\HH$) if $\theta$ is constant at zero. More generally, a submanifold of $\CP^N$ or $\HP^N$ is called \emph{slant} if $\theta$ is constant, or \emph{semi-slant} if its tangent bundle decomposes into two sub-bundles on which $\theta$ is constant respectively at zero and at another value. In Section~\ref{S:lit}, we will survey some of the literature related to these conditions and it's implications to the search for fat connections.
Going the other direction, there are several known obstructions to fat connections. For example, a trivial bundle can not admit a fat connection~\cite{Weinstein}, which implies that:
\begin{cor} A homotopically trivial immersed submanifold of $\CP^N$ or $\HP^N$ could not have bounded Wirtinger angles.
\end{cor}
Among $\HH$-bundles over $S^4$, only the Hopf bundle admits a fat connection~\cite{DR}, which implies:
\begin{cor} For any $N>0$, there is at most one element of $\pi_4(\HP^N)$ that admits an immersed representative $\varphi:S^4\ra\HP^N$ with bounded Wirtinger angles.
\end{cor}
In fact, if $N$ is large enough to ensure universality, then there is \emph{exactly} one such element. Similarly, among circle bundles over the flag $F^6=SU(3)/T^2$, there is an infinite family which admit fat connections (corresponding to the positively curved Aloff Wallach spaces) and there are two which do not (the trivial bundle and the $W_{1,0}$ bundle); see~\cite{Ziller}. Thus:
\begin{cor} For sufficiently large $N$, there are infinitely many homotopy classes of maps $F^6\ra\CP^N$ which admit an immersed representatives with bounded Wirtinger angles, and two which do not.
\end{cor}
We are not aware of any work specifically addressing the possible homotopy classes of submanifolds of $\KP^N$ with bounded Wirtinger angles, but the above discussion suggests that this question is both natural and subtle.
Next, we will study conditions on $\varphi$ under which $\nabla$ is \emph{parallel} or \emph{radially symmetric}:
\begin{definition} A connection $\nabla$ is called \emph{parallel} if the covariant derivative of its curvature tensor vanishes; that is, $D_ZR^\nabla(X,Y)W=0$ for all $p\in B$, all $X,Y,Z\in T_p B$ and all $W\in E_p=\pi_E^{-1}(p)$. The connection is called \emph{radially symmetric} if this condition holds when $Z=X$; that is, if $D_XR^\nabla(X,Y)W=0$ for all $p\in B$, all $X,Y\in T_p B$ and all $W\in E_p$.
\end{definition}
Notice that these conditions depend on both the connection and on the choice of metric on the base space. We therefore assume for the remainder of this section that $\varphi$ is an immersion and that $B$ has the pull-back metric. This added assumption sacrifices universality, since there is no reason to expect that a given metric on $B$ and a given connection can be simultaneously achieved from a single classifying map $\varphi$.
Strake and Walschap proved that if $B$ has positive sectional curvature, then any radially symmetric connection in any vector bundle over $B$ will induce a connection metric with nonnegative sectional curvature on the vector bundle~\cite{SW}. However, all known examples of radially symmetric connections are parallel, and parallel connections appear to be rare. For example, Guijarro, Sadun and Walschap proved in~\cite{parallel} that an $\R^k$ vector bundle over a compact simply connected irreducible symmetric space with a parallel connection must be isomorphic to an \emph{associated bundle}. Since associated bundles trivially admit submersion metrics of nonnegative curvature, this result seems to limit one's ability to obtain topologically new examples of nonnegatively curved vector bundles using parallel connections (at least over symmetric base spaces). For $\R^2$-bundles, radially symmetric is equivalent to parallel~\cite{STT}, but for higher rank bundles, the gap between these hypotheses is not well understood.
Let $II$ denote the second fundamental form of $\varphi(B)$ and let $S$ denote the shape operator, so that for $p\in B$, $X,Y\in T_pB\cong\varphi_*(T_p B)$ and $\eta\perp \varphi_*(T_p B)$, we have $\lb S_\eta X,Y\rb = \lb II(X,Y),\eta\rb$.
\begin{theorem}\label{P:par}
If $k=1$ and $\KK\in\{\C,\HH\}$, then $\nabla$ is parallel if and only if
$$\lb S_{(\mathfrak{J}X)^\perp}Y - S_{(\mathfrak{J}Y)^\perp}X,Z\rb = 0$$
for all $p\in B$, $X,Y,Z\in T_p B\cong\varphi_*(T_p B)$ and all $\mathfrak{J}\in\text{span}\{\mathcal{J}\}$, where ``$\perp$'' denotes the component orthogonal to $\varphi_*(T_pB)$. Furthermore, $\nabla$ is radially symmetric if and only if the above condition is true in the special case $Z=X$.
\end{theorem}
In particular, if $\varphi(B)$ is totally geodesic, then $\nabla$ is parallel. But in this case, $\varphi(B)$ is a symmetric space and the bundle is an associated bundle, as we will explain in Section~\ref{S:tg}.
There are examples in which $\nabla$ is parallel even though $\varphi(B)$ is not totally geodesic. Specifically, if $k=1$ and $\varphi(B)$ is K\"ahler/quaternionic, then $\nabla$ is fat (as explained above) and also parallel (because $(\mathfrak{J}X)^\perp = (\mathfrak{J}Y)^\perp =0$). The case $\KK=\HH$ is less interesting here because quaternionic implies totally geodesic. But in the case $\KK=\C$, there are many examples of K\"ahler submanifolds of $\CP^{N-1}$ which are not totally geodesic.
The final property of the connection $\nabla$ which we wish to interpret in terms of the geometry of the classifying map $\varphi$ is the following inequality:
\begin{gather}
\text{For all }p\in B,\text{ all } X,Y\in T_pB, \text{ and all } W,V\in E_p, \notag\\
\lb (D_XR^\nabla)(X,Y)W,V\rb^2 \leq k_B(X,Y)\cdot |R^\nabla(W,V)X|^2,\label{ineq}
\end{gather}
where $k_B$ denotes the unnormalized sectional curvature of $B$. This inequality was proven in~\cite{SW} to be a necessary condition for $\nabla$ (together with the given metric on $B$) to induce a connection metric with nonnegative sectional curvature on $E$. Further, if this inequality is \emph{strictly} satisfied (for all orthonormal choices of $X,Y,V,W$), then it was proven in~\cite{T} that $\nabla$ induces a connection metric of nonnegative curvature in $E$ and of positive curvature in $E^1$.
This inequality relates the two previously-discussed properties of a connection: its left side vanishes if and only if $\nabla$ is radially symmetric, while on its right side, $\nabla$ is fat if and only if the expression $|R^\nabla(W,V)X|^2$ is strictly positive for orthonormal $X,W,V$~\cite[Equation 11]{SW}. Thus, a fat radially symmetric connection over a positively curved base space will satisfy the inequality strictly, and will therefore induce a connection metric of nonnegative curvature on $E$ and of positive curvature on $E^1$.
Our translation of Inequality~\ref{ineq} becomes particularly simple for $\C^1$ and $\HH^1$-bundles:
\begin{theorem}\label{USA}If $k=1$ and $\KK\in\{\C,\HH\}$, then Inequality~\ref{ineq} is satisfied if and only if the following inequality is satisfied for all $p\in B$, $X,Y\in T_p B\cong\varphi_*(T_p B)$ and all $\mathfrak{J}\in\text{span}\{\mathcal{J}\}$:
\begin{equation}\label{loki}\lb S_{(\mathfrak{J}X)^\perp}Y - S_{(\mathfrak{J}Y)^\perp}X,X\rb^2\leq k_B(X,Y)\cdot|\text{proj}(\mathfrak{J}X)|^2,
\end{equation}
where ``proj'' denotes the orthogonal projection onto $\varphi_*(T_pB)$. Further, Inequality~\ref{ineq} is \emph{strictly} satisfied (for orthonormal $X,Y,W,V$) if and only if Inequality~\ref{loki} is \emph{strictly} satisfied (for orthonormal $X,Y$ and unit-length $\mathfrak{J}$).
\end{theorem}
Inequality~\ref{loki} is clearly satisfied if the following quantities are \emph{both} sufficiently close to zero for all $p\in B$:
\begin{itemize}
\item $|S(p)|=\max\{|S_\eta X|\mid X\in\varphi_*(T_pB), \eta\perp\varphi_*(T_pB), |X|=|\eta|=1\},$
\item $\theta(p)=\max\{\theta(X)\mid X\in \varphi_*(T_pB)\}.$
\end{itemize}
In fact, the closer one of these quantities is to zero, the further the other one can move away from zero while still satisfying the inequality:
\begin{cor}\label{bable}Assume that the metric on $\KP^{N-1}$ is normalized to have maximal sectional curvature $1$. If $|S(p)|^2<\frac{1}{16\tan^2\theta(p) + 8}$ for all $p\in B$, then Inequality~\ref{loki} is strictly satisfied, so $\nabla$ induces a connection metric of nonnegative curvature on $E$ and of positive curvature on $E^1$.
\end{cor}
For example, if the immersion is K\"ahler/quaternionic ($\theta=0$), then the inequality becomes $|S(p)|^2< 1/8$, which is the bound that insures that $B$ has positive curvature. But as $\theta(p)\ra\pi/2$, the required upper bound on $|S(p)|^2$ goes to $0$.
The author is pleased to thank Luis Guijarro and Wolfgang Ziller for helpful comments and suggestions on this work.
\section{Calculations for the pull-back connection}\label{S:calc}
In this section, we assume that $\KK=\R$ and that $\varphi:B\ra\GR$ is an \emph{isometric immersion}. Our goal is to describe $R^\nabla$ and $DR^\nabla$ in terms of the classifying map $\varphi$.
We require some notation. Let $H\subset K\subset G$ equal the triple $O(N-k)\subset O(k)\times O(N-k)\subset O(N)$. Let $\mh\subset\mk\subset\mg$ denote their Lie algebras. Let $g_0$ be the biinvariant metric on $G=O(N)$ defined as $g_0(X,Y) = \frac 12\text{trace}(A\cdot B^T)$ for all $A,B\in\mg$. We will sometime write $\lb A,B\rb_0$ to mean $g_0(A,B)$ and write $|A|^2_0$ to mean $g_0(A,A)$. We endow $\GR=G/K$ with the normal homogeneous metric induced by $g_0$. Denote $\mm=\mk\ominus\mh$ and $\mp=\mg\ominus\mk$, where ``$\ominus$'' denotes the $g_0$-orthogonal complement, so $\mg=\mh\oplus\mm\oplus\mp$. Let $h:G\ra G/K=\GR$ denote the projection, which is a Riemannian submersion with respect to the above-mentioned metrics.
Let $p\in B$, and choose $g\in G$ such that $h(g)=\varphi(p)$. Let $\mathfrak{T}\subset\mp$ denote the subspace such that $h_*(dL_g(\mathfrak{T}))=\varphi_*(T_pB)$. For any $Z\in T_pB$, let $\tilde{Z}\in\mathfrak{T}$ denote the unique vector such that $h_*(dL_g(\tilde{Z})) = \varphi_* Z$. Let $\widetilde{II}:\mathfrak{T}\times\mathfrak{T}\ra\mp\ominus\mathfrak{T}$ denote the lift of the second fundamental form, $II$, of $\varphi(B)$. In other words, $h_*(dL_g(\widetilde{II}(\tilde{X},\tilde{Y}))) = II(\varphi_*X,\varphi_*Y)$ for all $X,Y\in T_p B$.
\begin{prop}\hspace{.5in}\label{all}
\begin{enumerate}
\item $\nabla$ is fat if and only if $\left[\tilde{X},\alpha\right]^{\mathfrak{T}}\neq 0,$ and
\item $\nabla$ is parallel if and only if
$\left\langle \left[\tilde{X},\widetilde{II}(\tilde{Z},\tilde{Y})\right]
-\left[\tilde{Y},\widetilde{II}(\tilde{Z},\tilde{X})\right], \alpha\right\rangle_0=0, and$
\item Inequality~\ref{ineq} is satisfied if and only if the following is satisfied:
\begin{equation}
\left\langle\left[\tilde{X},\widetilde{II}(\tilde{X},\tilde{Y})\right] -
\left[\tilde{Y},\widetilde{II}(\tilde{X},\tilde{X})\right],\alpha\right\rangle^2
\leq k_B(X,Y)\cdot \left|\left[\tilde{X},\alpha\right]^{\mathfrak{T}}\right|_0^2, \label{INEQ}
\end{equation}
\end{enumerate}
for all $p\in B$, all $X,Y,Z\in T_p B$ and all nonzero decomposable $\alpha\in\mm\cong so(k)\cong\Lambda^2(\R^k)$.
\end{prop}
Notice that the validity of these conditions do not depend on the choice of $g\in h^{-1}(\varphi(p))$, even though some of the individual terms do.
As we will explain later in this section, $\alpha$ represents the plane spanned by $W$ and $V$, and Inequalities~\ref{ineq} and~\ref{INEQ} match each other term-for-term in the obvious manner, from which parts (1) and (2) of the proposition follow. Thus, all that is really required to prove Proposition~\ref{all} is to describe $R^\nabla$ and $DR^\nabla$ in terms of $\varphi$. The remainder of this section is devoted to this task.
It is already clear that we intend to work in the setting of principle bundles, rather than vector bundles. The total space of the universal principle $O(k)$-bundle over $\GR$ is the collection of ``frames'', i.e., ordered sets of $k$ orthonormal vectors in $\R^N$:
$$\Gf = O(N)/O(N-k)$$
Let $\pi:\Gf\ra\GR$ denote the projection map, which sends each frame to its span.
This universal principal bundle, $O(k)\hookrightarrow\Gf\stackrel{\pi}{\ra}\GR$, can be re-described as the following homogenous bundle:
\begin{equation}\label{got} K/H\hookrightarrow G/H\stackrel{\pi}{\ra}G/K.
\end{equation}
Notice that $\pi$ becomes a Riemannian submersion when $G/H$ and $G/K$ are endowed with the normal homogeneous metrics induced by $g_0$. Let $\V$ and $\Hor$ denote the vertical and horizontal distributions of $\pi$. We will refer to $\Hor$ as the ``universal connection'' because:
\begin{lem} $\Hor$ is a principle connection in the principle bundle; in fact, it equals the universal connection constructed in~\cite{NR}.
\end{lem}
\begin{proof} Notice that $K=H\times O(k)$, so $H$ is normal in $K$, and $K/H=O(k)$. Since $H$ commutes with $O(k)$, the right-$O(k)$-action on $G/H$ is well-defined and isometric. Thus, the base space, $G/K$, of $\pi$ can be identified with $(G/H)/O(k)$. Under this identification, $\pi$ is simply the quotient map from $G/H$ to $(G/H)/O(k)$. In summary, the right $O(k)$-action on $G/H$ preserves each $\pi$-fiber (and therefore preserves $\V$) and is isometric (so it also preserves $\Hor$). Thus, $\Hor$ is invariant under this principal $O(k)$-action, which makes it a principle connection.
Notice that the left $G$-action on itself induces a transitive isometric $G$-action on $G/H$ which sends $\pi$-fibers to $\pi$-fibers and therefore preserves $\Hor$ and $\V$. Thus, to verify that $\Hor$ is the same as the universal connection constructed in~\cite{NR}, which has this same homogeneity property, it suffices to check a single point, which is straightforward.
\end{proof}
Consider the chain of Riemannian submersions $$G\stackrel{f}{\ra}G/H=\Gf\stackrel{\pi}{\ra}G/K=\GR,$$
and denote $h=\pi\circ f$. Let $\omega_\Hor$ denote the connection form of $\Hor$, which is an $o(k)$-valued $1$-form on $\Gf$. Let $\Omega_\Hor$ denote its curvature form, which is an $o(k)$-valued $2$-form on $\Gf$.
Let $g\in G$ be arbitrary, and denote $q=f(g)\in \Gf$. Recall that $L_g:G\ra G$ induces an isometry $\Gf\ra\Gf$ which preserves $\V$ and $\Hor$. Therefore,
$$\Hor_q = \{f_*(dL_gX)\mid X\in\mp\}\,\,\,\text{ and }\,\,\,\
\V_q = \{f_*(dL_gV)\mid V\in\mm\}.$$
Notice that for any $V\in\mm\cong o(k)$, we have
\begin{equation}\label{ironman}\omega_{\Hor}(f_*(dL_gV)) = V,\end{equation}
simply because the action fields for the principle right $O(k)$-action on $\Gf$ are projections via $f_*$ of left-invariant fields on $G$.
\begin{lem}\label{L:OmegaH} For all $X,Y\in\mp$, we have: $\Omega_\Hor(f_*(dL_gX),f_*(dL_gY)) = \frac 12 [X,Y]^\mm$.
\end{lem}
\begin{proof}
We first consider the special case $g=e$ (the identity of $G$). In this case, let $t\mapsto g(t)$ denote the geodesic in $G$ with $g(0)=e$ and $g'(0)=X$. Let $t\mapsto Y(t)=dL_{g(t)}Y$ denote the left-invariant extension of $Y$ along this geodesic. Notice that $t\mapsto (f\circ g)(t)$ is a $\pi$-horizontal geodesic in $\Gf$, and that $t\mapsto \hat{Y}(t) = f_*(Y(t))$ is a $\pi$-horizontal extension of $f_*Y$ along this horizontal geodesic. Therefore the O'Neill tensor, $A$, of $\pi$ at the point $q$ satisfies:
$$A(f_*X,f_*Y) = \left(\hat{Y}'(0)\right)^\V = \left(f_*(Y'(0))\right)^\V = \frac 12 \left(f_*[X,Y]\right)^\V= \frac 12 f_*\left([X,Y]^\mm\right).$$
Now we return to the case where $g\in G$ is arbitrary. The previous formula implies:
$$A(f_*(dL_gX),f_*(dL_gY)) = \frac 12 f_*\left(dL_g\left([X,Y]^\mm\right)\right)\in\V_{f(g)}.$$
Together with Equation~\ref{ironman}, this gives:
\begin{equation}\label{Omega_q}\Omega_\Hor(f_*(dL_g X),f_*(dL_g Y))
= \omega_{\Hor}(A(f_*(dL_gX),f_*(dL_gY)))
=\frac 12 [X,Y]^\mm.\end{equation}
\end{proof}
Let $\pi_P:P\ra B$ be the principle $O(k)$-bundle obtained as the pull-back via $\varphi$ of the universal frame bundle. Explicitly, $P=\{(p,a)\in B\times \Gf\mid \varphi(p)=\pi(a)\}$ with projection $\pi_P(p,a) = p$. Let $\overline{\varphi}:P\ra\Gf$ denote the corresponding bundle homomorphism, defined as $\overline{\varphi}(p,a)=a$.
Let $\omega$ denote the connection form of the principle connection in $P$ obtained as the pull-back via $\varphi$ of the universal connection in the universal frame bundle. Let $\Omega$ denote the curvature form of $\omega$. Thus, $\omega$ is an $o(k)$-valued 1-form on $P$, and $\Omega$ is an $o(k)$ valued $2$-form on $P$.
Let $\pi_E:E\ra B$ denote the vector bundle associated to $P$. Explicitly, $E=P\times_{O(k)}\R^k$. Let $p\in B$. For any $x\in\pi_P^{-1}(p)$ and $U\in\R^k$, we let $x\diamond U\in E_p=\pi_E^{-1}(p)$ denote the image of $(x,U)$ under the projection $P\times\R^k\ra E$. Let $\nabla$ the connection in $E$ associated to $\omega$, and let $R^\nabla$ denote its curvature tensor.
Let $X,Y\in T_p B$ be orthonormal, and let $W,V\in E_p$ be orthonormal.
For simplicity, we initially assume that $\varphi(p) = h(e)\in \GR$. Let $a=f(e)\in \Gf$, and let $a'=(a,p)\in P$. Notice that $\R^k$ can be identified with the fiber $E_p$ via the map which sends $\hat{U}\in\R^k$ to $U=a'\diamond\hat{U}\in E_p$. Let $\hat{W},\hat{V}\in\R^k$ be associated with $W,V\in E_p$ in this way. There exists a unique $\alpha\in o(k)$ such that $\alpha\cdot \hat{W}=\hat{V}$, $\alpha\cdot\hat{V}=-\hat{W}$, and $\alpha\cdot\hat{U}=0$ for all $\hat{U}\in\R^k$ orthogonal to $\text{span}\{\hat{W},\hat{V}\}$, where the dots denote matrix multiplication. We claim that:
\begin{equation}\label{notnot}\lb \beta\cdot\hat{W},\hat{V}\rb = \lb\beta,\alpha\rb_0
\,\,\,\,\,\,\,\,\,\text{ for all }\beta\in o(k).\end{equation}
To see this, just choose an ordered orthonormal basis of $\R^k$ beginning with $\hat{W},\hat{V}$, and then express $\alpha,\beta$ in terms of the corresponding standard basis for $o(k)$.
For any $Z\in T_p B$, we let $\overline{Z}\in T_{a'}P$ denote its $\pi_P$-horizontal lift, and we let $\tilde{Z}\in\mp$ denote the unique vector such that $h_*\tilde{Z} = \varphi_* Z$. In the following calculation, the maximum is taken over all unit-length $Z\in T_pB$:
\begin{eqnarray*}
|R^\nabla(W,V)X| & = & \max\lb R^\nabla(W,V)X,Z\rb = \max\lb R^\nabla(X,Z)W,V\rb \label{T1}\\
& = & \max\lb \Omega(\overline{X},\overline{Z})\cdot \hat{W},\hat{V}\rb
=
\text{max}\lb\Omega(\overline{X},\overline{Z}),\alpha\rb_0 \\
& = & \text{max}\lb\Omega_\Hor
(\overline{\varphi}_*\overline{X},\overline{\varphi}_*\overline{Z}),\alpha\rb_0
= \frac 12\text{max}\lb[\tilde{X},\tilde{Z}]^\mm,\alpha\rb_0 \\
& = & \frac 12\text{max}\lb[\tilde{X},\tilde{Z}],\alpha\rb_0
= \frac 12\text{max}\lb[\tilde{X},\alpha],\tilde{Z}\rb_0= \frac 12|[\tilde{X},\alpha]^\mathfrak{T}|_0.
\end{eqnarray*}
The last equality uses that $\varphi$ is an isometric immersion, so maximizing over all unit-length $Z\in T_p B$ is the same as maximizing over all unit-length $\tilde{Z}\in\mathfrak{T}$. Recall that
where $\mathfrak{T}\subset\mp$ is the subspace such that $h_*(\mathfrak{T})=\varphi_*(T_pB)$. In summary:
\begin{equation}\label{m1}
|R^\nabla(W,V)X| = \frac 12|[\tilde{X},\alpha]^\mathfrak{T}|_0.
\end{equation}
It remains to express the expression $\lb(D_ZR^\nabla)(X,Y)W,V\rb$ in terms of the geometry of $\varphi$. For this, let $t\mapsto c(t)$ denote the geodesic in $B$ with $c(0)=p$ and $c'(0)=Z$. Let $t\mapsto\overline{c}(t)$ denote its $\pi_P$-horizontal lift beginning at $ a' \in P$. We can write $\overline{c}(t) = (c(t),\beta(t))$ where $t\mapsto\beta(t)$ is the $\pi$-horizontal lift of $t\mapsto \varphi(c(t))$ to $\Gf$ beginning at $\beta(0)=a$. Let $t\mapsto g(t)$ be the $h$-horizontal lift to $G$ of $t\mapsto\varphi(c(t))$ beginning at $g(0)=e$. Define $W(t)=\overline{c}(t)\diamond\hat{W}\in E_{c(t)}$ and $V(t)=\overline{c}(t)\diamond\hat{V}\in E_{c(t)}$. Notice that $V(t)$ and $W(t)$ are parallel because $\overline{c}(t)$ is horizontal.
Let $X(t)$ and $Y(t)$ denote the parallel extensions of $X,Y$ along $t\mapsto c(t)$. Let $\overline{X}(t)$ and $\overline{Y}(t)$ denote the $\pi_P$-horizontal lifts of these fields along $t\mapsto\overline{c}(t)$. Let $\tilde{X}(t)$ and $\tilde{Y}(t)$ denote the $h$-horizontal lifts along $t\mapsto g(t)$ of the fields $t\mapsto \varphi_*X(t)$ and $t\mapsto \varphi_*Y(t)$. Notice that:
$$R^\nabla(X(t),Y(t))W(t) = \overline{c}(t)\diamond\left(\Omega(\overline{X}(t),\overline{Y}(t))\cdot\hat{W}\right)
\in E_{c(t)}.$$
In the following calculation, we will use $\frac{D}{dt}$ to denote covariant differentiation with respect to $\nabla$, and $\frac{d}{dt}$ to denote the usual differentiation of a path of vectors in the Euclidean spaces $\R^k$ and $o(k)$, and prime $'$ to denote covariant differentiation with respect to the Levi Civita connection in $(G,g_0)$. With this notation, we have:
\begin{eqnarray*}
\lb (D_Z R^\nabla)(X,Y)W,V\rb
& = & \left\langle \frac{D}{dt}\Big|_{t=0}
\overline{c}(t)\diamond\left(\Omega(\overline{X}(t),\overline{Y}(t))
\cdot\hat{W}\right),V\right\rangle \\
& = & \left\langle \overline{c}(0)\diamond\left(\frac{d}{dt}\Big|_{t=0}
\Omega(\overline{X}(t),\overline{Y}(t))\cdot\hat{W}\right),V\right\rangle \\
& = & \left\langle \left(\frac{d}{dt}\Big|_{t=0}
\Omega(\overline{X}(t),\overline{Y}(t))\right)\cdot\hat{W},\hat{V}\right\rangle \\
& = & \left\langle \frac{d}{dt}\Big|_{t=0}
\Omega(\overline{X}(t),\overline{Y}(t)),\alpha\right\rangle_0
= \left\langle \frac{d}{dt}\Big|_{t=0}
\Omega_{\Hor}(\overline{\varphi}_*\overline{X}(t),
\overline{\varphi}_*\overline{Y}(t)),\alpha\right\rangle_0 \\
& = & \frac 12 \left\langle \frac{d}{dt}\Big|_{t=0}
\left[dL_{g(t)}^{-1}\tilde{X}(t) ,dL_{g(t)}^{-1}\tilde{Y}(t)\right]^\mm,\alpha\right\rangle_0 \\
& = & \frac 12 \left\langle \left[\frac{d}{dt}\Big|_{t=0}
dL_{g(t)}^{-1}\tilde{X}(t),\tilde{Y}\right]^\mm +\left[
\tilde{X},\frac{d}{dt}\Big|_{t=0}dL_{g(t)}^{-1}\tilde{Y}(t)\right]^\mm,\alpha\right\rangle_0.
\end{eqnarray*}
To interpret these terms, let $\{E_i\}$ denote an orthonormal basis of $\mp$, and let $\{E_i(t)\}$ denote their left-invariant extensions along $g(t)$; that is, $E_i(t)=dL_{g(t)}E_i$. Then,
\begin{eqnarray*}
\frac{d}{dt}\Big|_{t=0}dL_{g(t)}^{-1}\tilde{Y}(t)
& = & \frac{d}{dt}\Big|_{t=0}\sum\lb dL_{g(t)}^{-1}\tilde{Y}(t),E_i\rb_0 E_i
= \frac{d}{dt}\Big|_{t=0}\sum\lb\tilde{Y}(t),E_i(t)\rb_0 E_i \\
& = & \sum\lb\tilde{Y}'(0),E_i\rb_0 E_i + \sum\lb\tilde{Y},E_i'(0)\rb_0 E_i\\
& = & \tilde{Y}'(0) + \frac 12 \sum\lb\tilde{Y},[\tilde{Z},E_i]\rb_0 E_i \\
& = & \tilde{Y}'(0) - \frac 12 \sum\lb E_i,[\tilde{Z},\tilde{Y}]\rb_0 E_i\\
& = & \tilde{Y}'(0) - \frac 12[\tilde{Z},\tilde{Y}].
\end{eqnarray*}
Since $t\mapsto Y(t)$ is a parallel vector field along the geodesic $t\mapsto c(t)$ in $B$, we have $\tilde{Y}'(0) = \widetilde{II}(\tilde{Z},\tilde{Y}) + A_h(\tilde{Z},\tilde{Y})$, where $A_h$ denotes the O'Neill tensor of $h$. In summary,
\begin{eqnarray*}
\left[\tilde{X},\frac{d}{dt}\Big|_{t=0}dL_{g(t)}^{-1}\tilde{Y}(t)\right]^\mm
& = & \left[\tilde{X},\tilde{Y}'(0)-\frac 12[\tilde{Z},\tilde{Y}]\right]^\mm\\
& = & \left[\tilde{X},\widetilde{II}(\tilde{Z},\tilde{Y})+A_h(\tilde{Z},\tilde{Y}) - \frac 12[\tilde{Z},\tilde{Y}]\right]^\mm\\
& = & \left[\tilde{X},\widetilde{II}(\tilde{Z},\tilde{Y})\right]^\mm.
\end{eqnarray*}
The last equality follows from the fact that $\tilde{X}\in\mp$ while $A_h(\tilde{Z},\tilde{Y})\in\mk$ and $[\tilde{Z},\tilde{Y}]\in\mk$ (because $\mk\subset\mg$ is a symmetric pair). We similarly have that:
$$ \left[\tilde{Y},\frac{d}{dt}\Big|_{t=0}dL_{g(t)}^{-1}\tilde{X}(t)\right]^\mm
= \left[\tilde{Y},\widetilde{II}(\tilde{Z},\tilde{X})\right]^\mm.$$
Therefore, since $\alpha\in\mm$, we have:
\begin{equation}\lb (D_Z R^\nabla)(X,Y)W,V\rb = \frac 12\left\langle \left[\tilde{X},\widetilde{II}(\tilde{Z},\tilde{Y})\right]
-\left[\tilde{Y},\widetilde{II}(\tilde{Z},\tilde{X})\right], \alpha\right\rangle_0.\label{parallelcurv}\end{equation}
\begin{proof}[Proof of Proposition~\ref{all}]
By~\cite[Equation 11]{SW}, $\nabla$ is fat if and only if $|R^\nabla(W,V)X|>0$ for all $p\in B$, nonzero $X\in T_pB$ and all \emph{orthonormal} $W,V\in E_p$ (this condition depends only on $\nabla$, even though a metric on $B$ must be chosen in order for the expression $R^\nabla(W,V)X$ and its norm to be defined). Also notice that $\nabla$ is parallel if and only if $\lb D_ZR^\nabla(X,Y) W,V\rb=0 $ for all $p\in B$, $X,Y\in T_p B$ and \emph{orthonormal} $W,V\in E_p$. Also notice that Inequality~\ref{ineq} is satisfied if and only if it is satisfied for all \emph{orthonormal} choices of $X,Y,W,V$.
Proposition~\ref{all} now follows from Equations~\ref{m1} and~\ref{parallelcurv}. Recall that in these equations, $W,V$ were assumed to be orthonormal, and $\alpha\in\mm$ was selected to represent $\text{span}\{W,V\}$ in the sense of Equation~\ref{notnot}. In fact, the nonzero \emph{decomposable} elements of $o(k)\cong\Lambda^2(\R^k)$ are exactly the elements which represent planes in this manner.
\end{proof}
\section{Totally geodesic classifying maps}\label{S:tg}
In this section, we assume that $\KK=\R$ and that the classifying map $\varphi:B\ra\GR$ is a totally geodesic isometric imbedding. By Proposition~\ref{all}, this implies that $\nabla$ is parallel. If additionally $B$ has positive sectional curvature, then $\nabla$ induces a connection metric of nonnegative curvature on $E$~\cite{SW}.
Since $\GR$ contains many totally geodesic submanifolds (which have not yet been fully classified) including many with positive curvature, this might appear to be a hopeful source for topologically new examples of vector bundles which admit metrics of nonnegative curvature. But we will explain in this section why no new examples can be obtained in this way.
First notice that $B$ is a symmetric space because it is a totally geodesic submanifold of the symmetric space $\GR=G/K$ (here, as before, $H\subset K\subset G$ denotes the triple $O(N-k)\subset O(k)\times O(N-k)\subset O(N)$). More precisely, $B=G'/K'$ where $G'\subset G$ and $K'=G'\cap K$. Let $\rho':K'\ra O(k)$ denote the composition of the inclusion map into $K$ with the projection onto the first factor. The following was observed by Rigas in~\cite{Rigas}.
\begin{prop}[Rigas]\label{rig}The bundle $\pi_E:E\ra B$ is isomorphic to the associated bundle $G'\times_{\rho'}\R^k\ra G'/K'$.
\end{prop}
\begin{proof} Let $\rho:K\ra O(k)$ denote the projection onto the first factor.
The universal principal bundle $O(k)\hookrightarrow\Gf\ra\GR$ was re-described in Equation~\ref{got} as the homogenous bundle $K/H\hookrightarrow G/H\ra G/K.$ This homogeneous bundle can again be re-described as the associated bundle $K/H\hookrightarrow G\times_{\rho}(K/H)\ra G/K.$ This implies that the universal vector bundle over $\GR$ can be re-described as $\R^k\hookrightarrow G\times_\rho\R^k\ra G/K$. Consider the following commutative diagram, in which the right arrows denote the natural inclusion maps:
$$\begin{CD} G'\times_{\rho'}\R^k @> >> G\times_{\rho}\R^k \\
@V VV @VV V \\
G'/K' @>\varphi>> G/K
\end{CD}$$
the bundle on the right is the universal vector bundle over $\GR=G/K$, so the bundle on the left is isomorphic to the pull-back bundle $\pi_E:E\ra B$, as desired.
\end{proof}
In conclusion, if the classifying map $\varphi:B\ra\GR$ is a totally geodesic isometric imbedding, then $\nabla$ is parallel, but in this case the bundle is isomorphic to an associated bundle, which trivially admits a submersion metric of nonnegative sectional curvature.
\section{The cases of $\C^1$ and $\HH^1$ bundles}\label{S:k1}
The primary technical difficulty in applying Proposition~\ref{all} is that ``bracketing with $\alpha$'' is difficult to interpret geometrically in general. However, in this section, we will provide a very natural geometric interpretation in the special case where $k=1$ and $\KK\in\{\C,\HH\}$ (so that $\varphi:B\ra\CP^{N-1}$ or $\varphi:B\ra\HP^{N-1}$). Theorems~\ref{P:fat},~\ref{P:par} and~\ref{USA} will follow from this interpretation.
Even though we assumed in the Section~\ref{S:calc} that $\KK=\R$, almost all of the calculations generalize in the obvious way to the cases $\KK\in\{\C,\HH\}$. The only exception involves the manner in which $\alpha$ was chosen to represent a particular plane in the fiber. Recall that we were given arbitrary orthonormal vectors $W,V\in\R^k$, and we were able to choose $\alpha\in o(k)$ to represent the plane $\text{span}\{W,V\}$ in the sense that:
$$\lb \beta\cdot W,V\rb = \lb\beta,\alpha\rb_0
\,\,\,\,\,\,\,\,\,\text{ for all }\beta\in o(k).$$
To generalize this proof to the case $\KK=\C$ (respectively $\KK=\HH$), we would be given arbitrary $\R$-orthonormal vectors $W,V\in\KK^k$, and we would need to choose $\alpha\in u(k)$ (respectively $\alpha\in sp(k)$) so $\lb \beta\cdot W,V\rb_\R = \lb\beta,\alpha\rb_0$ for all $\beta\in u(k)$ (respectively $\beta\in sp(k)$). Unfortunately, this is not generally possible unless we additionally assume that $W\perp\text{span}\{\mathfrak{J}V\mid \mathfrak{J}\in\mathcal{J}\}$.
However, when $k=1$, there is no trouble with choosing $\alpha$ as desired. In fact, the choice $\alpha=V\cdot\overline{W}$ works, and all of the calculations in the previous section go through.
More specifically, when $k=1$ and $\KK=\C$ (so that $\varphi:B\ra\CP^{N-1}$), the chain $H\subset K\subset G$ from Section~\ref{S:calc} becomes $SU(N-1)\subset S(U(1)\times U(N-1))\subset SU(N)$, and $\mm=u(1)$ is spanned by a unique (up to sign) unit-length vector $\alpha\in\mm$. It is not hard to see that for any $q\in G/K=\CP^{N-1}$, the map $\text{ad}_\alpha:\mp\ra\mp$ (which sends $\tilde{X}$ to $[\alpha,\tilde{X}]$) induces an involution of $T_q(\CP^{N-1})$ that is well-defined in the sense that it is independent of the choice of $g\in h^{-1}(q)$ through which the lift $\tilde{X}$ is defined as in Proposition~\ref{all}. In fact, this is one natural way in which to define the standard almost complex structure on $\CP^{N-1}$. Therefore in Proposition~\ref{all}, bracketing with $\alpha$ can be interpreted as applying the almost-complex structure $J$.
When $k=1$ and $\KK=\HH$ (so that $\varphi:B\ra\HP^{N-1}$), the chain $H\subset K\subset G$ from the previous section becomes $Sp(N-1)\subset Sp(1)\times Sp(N-1)\subset Sp(N)$. Identify $\mathcal{J}=\{I,J,K\}$ with an oriented orthonormal basis of $\mm=sp(1)=\text{Im}(\HH)$. For any $q\in G/K=\HP^{N-1}$, the triple of maps $\text{ad}_I,\text{ad}_J,\text{ad}_K:\mp\ra\mp$ induces a triple of involutions of $T_q(\HP^{N-1})$ which satisfy the familiar properties of an almost quaternionic structure. Changing to a different $g\in h^{-1}(q)$ has the effect of conjugating to a different oriented orthonormal basis of $\mm$, so the family of triples:
$$\{\{\text{ad}_I,\text{ad}_J,\text{ad}_K\}\mid\{I,J,K\} \text{ is an oriented orthonormal basis of }\mm \}$$
determines a well-defined family of triples of involutions of $T_q\HP^{N-1}$. This is one way to define the natural almost quaternionic structure on $\HP^{N-1}$. Recall that on an almost quaternionic manifold, a choice of basis $\{I,J,K\}$ for the almost quaternionic structure generally only exists locally, which is reflected in the dependence on $g\in h^{-1}(q)$ described above. In any case, bracketing with all possible $\alpha\in\mm$ can be interpreted in Proposition~\ref{all} as applying all possible elements of $\text{span}\{\mathcal{J}\}$.
\begin{proof}[Proof of Theorem~\ref{P:fat}] Recall that $\nabla$ is fat if and only if $|R^\nabla(W,V)X|>0$ for all $p\in B$, nonzero $X\in T_pB$ and all orthonormal $W,V\in E_p$.
If $\varphi$ is not an immersion, then there exits $p\in B$ and $X\in T_pB$ such that $\varphi_*X=0$, which implies that $R^\nabla(W,V)X=0$ for any choice of $W,V$. Thus, $\nabla$ is not fat.
Next assume that $\varphi$ is an immersion, and choose the pull-back metric for $B$. Let $\alpha\in\mm$ represent the plane $\text{span}\{W,V\}$ in the sense of Equation~\ref{notnot}. Let $\mathfrak{J}\in\text{span}\{\mathcal{J}\}$ represent $\alpha$ as described previously in this section. By Equation~\ref{m1},
$$2\left|R^\nabla(W,V)X\right| = \left|[\tilde{X},\alpha]^\mathfrak{T}\right|_0 = \left|(\mathfrak{J}X)^{\varphi_*(T_pB)}\right|\geq|X|\cdot\cos(\theta(X)),$$
so $\nabla$ is fat if and only if $\theta(X)<\pi/2$ for all nonzero $X\in TM$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{P:par}]
The connection $\nabla$ is parallel if and only if the following equals zero for all $p\in B$, $X,Y,Z\in T_pB$ and $W,V\in E_p$:
\begin{eqnarray*}2 \lb (D_Z R^\nabla)(X,Y)W,V\rb
& = & \left\langle \left[\tilde{X},\widetilde{II}(\tilde{Z},\tilde{Y})\right]
-\left[\tilde{Y},\widetilde{II}(\tilde{Z},\tilde{X})\right], \alpha\right\rangle_0\\
& = & \left\langle \left[\alpha,\tilde{X}\right],\widetilde{II}(\tilde{Z},\tilde{Y})\right\rangle_0
-\left\langle \left[\alpha,\tilde{Y}\right],\widetilde{II}(\tilde{Z},\tilde{X})\right\rangle_0\\
& = & \left\langle \mathfrak{J}X,II(Z,Y)\right\rangle
-\left\langle \mathfrak{J}Y,II(Z,X)\right\rangle\\
& = & \left\langle S_{(\mathfrak{J}X)^\perp}Y,Z\right\rangle
- \left\langle S_{(\mathfrak{J}Y)^\perp}X,Z\right\rangle\\
& = & \left\langle S_{(\mathfrak{J}X)^\perp}Y - S_{(\mathfrak{J}Y)^\perp}X,Z\right\rangle,
\end{eqnarray*}
where $\alpha\in\mm$ represent the plane $\text{span}\{W,V\}$ in the sense of Equation~\ref{notnot}, and $\mathfrak{J}\in\text{span}\{\mathcal{J}\}$ represents $\alpha$ as described previously in this section. Furthermore, $\nabla$ is radially symmetric if and only if the above is true in the special case $Z=X$.
\end{proof}
Theorem~\ref{USA} follows immediately from the calculations of the previous two proofs. It remains only to prove Corollary~\ref{bable}.
\begin{proof}[Proof of Corollary~\ref{bable}]
The inequality of Theorem~\ref{USA} is:
$$\lb S_{(\mathfrak{J}X)^\perp}Y - S_{(\mathfrak{J}Y)^\perp}X,X\rb^2\leq k_B(X,Y)\cdot|\text{proj}(\mathfrak{J}X)|^2.$$
The terms of this inequality at a point $p\in B$ are bounded as follows:
\begin{eqnarray*}
\lb S_{(\mathfrak{J}X)^\perp}Y - S_{(\mathfrak{J}Y)^\perp}X,X\rb^2
& \leq & 4|S(p)|^2\sin^2\theta(p).\\
k_B(X,Y) & = & k_{\KP^N}(X,Y) +\lb II(X,X),II(Y,Y)\rb - |II(X,Y)|^2\\
& \geq & k_{\KP^N}(X,Y) - 2|S(p)|^2\\
& \geq & 1/4 - 2|S(p)|^2.\\
|\text{proj}(\mathfrak{J}X)|^2 & \geq & \cos^2\theta(p).
\end{eqnarray*}
Thus, the inequality is satisfied if:
$$4|S(p)|^2\sin^2\theta(p) \leq (1/4 - 2|S(p)|^2)\cos^2\theta(p),$$
which can be re-expressed as:
$$|S(p)|^2\leq\frac{1}{16\tan^2\theta(p) + 8}.$$
\end{proof}
\section{Well studied classes of immersions into $\CP^N$ and $\HP^N$}\label{S:lit}
Theorems~\ref{P:fat},~\ref{P:par} and~\ref{USA} empower one to construct connections with certain natural geometric properties in $\KK^1$-vector bundles over $B$ (with $\KK\in\{\C,\HH\}$) by constructing immersions of $B$ into $\KP^{N-1}$ that satisfy certain hypotheses. There is a large body of literature on immersed submanifolds of projective spaces with natural properties. Some of these properties overlap the hypotheses required by our theorems. In this section, we review some of the literature and and discuss it's relevance to the search for nice connections in $\KK^1$ bundles.
As mentioned previously, if $\varphi$ is an isometric K\"ahler/quaternionic immersion, then $\nabla$ is fat and parallel. If additionally $B$ has positive curvature, then the inequality is strictly satisfied, so $\nabla$ induces a connection metric of nonnegative curvature in $E$ and of positive curvature in $E^1$. Since these are strong conclusions, it is worthwhile to begin by surveying some of the relevant literature on isometric K\"ahler/quaternionic immersions. We start with the case $\KK=\C$.
Calabi's rigidity theorem from~\cite{Calabi} says that if $f_1,f_2:B\ra\CP^N$ are both isometric K\"{a}hler immersions, then there exists a unitary transformation $U$ of $\CP^N$ such that $f_2=U\circ f_1$. Calabi further classified all isometric K\"{a}hler imbeddings of $\CP^n(c_1)$ into $\CP^N(c_2)$, where $c_1,c_2$ denote the constant holomorphic curvature. For any fixed $n$, he proved there exists a countably infinite family of imbeddings, $f_i:\CP^n(c/i)\ra\CP^{N_i}(c)$, where $i\in\Z^+$ and $N_i=\frac{(n+i)!}{n!i!}-1$. The map $f_i$ is sometimes called the $i^{\text{th}}$ Veronese imbedding. It is not totally geodesic if $i>1$. Each $f_i$ induces a parallel fat connection in the pulled-back $\C^1$-bundle over $\CP^n$, and the total space, $E_i^1$, of the corresponding circle-bundle inherits a connection metric of positive sectional curvature.
These examples are not new. The main result of~\cite{parallel} implies that any vector bundle over the symmetric space $\CP^n=SU(n+1)/S(U(1)\times U(n))$ with a parallel connection must be isomorphic to an associated bundle, which means it has the form $SU(n+1)\times_{\rho}\C\ra \CP^n$ for some representation $\rho:S(U(1)\times U(n))\ra U(1)$. There is a one parameter family of such representations coming from powers of the determinant of $A\in U(n)\cong S(U(1)\times U(n))$: $\rho_j(A)=\det(A)^j$. The total space of each circle bundle therefore has the following form for some $j\in\Z^+$:
\begin{equation}\label{franklin}E^1_j=SU(n+1)\times_{\rho_j}U(1),\end{equation}
which can be shown to be diffeomorphic to the lens space $S^{2n+1}/\Z_j$.
Nakagawa and Takagi in~\cite{NT} classified the isometric K\"{a}hler imbeddings of all other compact simply connected irreducible Hermitian symmetric space into $\CP^N$. For each such space, they obtained a countably infinite families of imbeddings analogous to the Veronese imbeddings. Notice that $\CP^n$ is the only such space with positive curvature, so no new examples of connection metrics with positive curvature in circle bundles could be obtained by pulling back the universal connection via these imbeddings. As above, these pulled-back connections are parallel and fat, and the bundles are associated bundles.
More recently, Di Scala, Ishi, and Loi studied isometric K\"ahler immersions of the form $f:B\ra\CP^N(1)$, where $B$ is a homogeneous K\"ahler manifold~\cite{DIL}. They proved that $f$ must be an imbedding and that $B$ must be simply connected. Moreover, they conjectured that (some rescaling of) any simply connected homogeneous K\"ahler manifold, $B$, whose associated K\"ahler form is integral must admit an isometric K\"ahler imbedding into $\CP^N(1)$ for some $N$. This conjecture would imply that over each such space there exists a $\C^1$-bundle that admits a parallel fat connection.
There is no classification of the isometric K\"ahler immersions $f:B\ra\CP^N(1)$ for which $B$ has positive or nonnegative sectional curvature, except under added hypotheses. For example, If $B$ complex dimension $\geq 2$ and sectional curvature $>1/8$, then Ros and Verstraelen proved that $f(B)$ must be totally geodesic~\cite{RV}. If $B$ has positive sectional curvature and has holomorphic curvature $\geq 1/2$, then Ros proved that $f$ must be a one of a list of standard imbeddings~\cite{Ros}. If $B$ has nonnegative sectional curvature and has complex codimension less than its complex dimension, then Shen obtained a similar conclusion~\cite{Shen}. But without any added hypotheses, no classification is known. Any new example would be interesting, especially if $B$ had positive sectional curvature, for then the pulled back circle bundle over $B$ would inherit positive sectional curvature as well.
There are other conditions on immersions (more general than the K\"ahler/quaternionic condition) that imply bounded Wirtinger angles. For example, an immersion $\varphi:B\ra\KP^N$ is called called \emph{slant} if $\theta$ is constant, or \emph{semi-slant} if its tangent bundle decomposes into two sub-bundles on which $\theta$ is constant respectively at zero and at another value. Slant submanifolds were defined by Chen, who summarized the early results in his book~\cite{chen}. The pullback of any proper slant (or semi-slant) immersion would yield a $\KK^1$ bundle with a fat connection. The literature on slant and semi-slant submanifolds consists primarily of rigidity results. However, Maeda, Ohnita and Udagawa in~\cite{MOU} constructed examples of slant submanifolds of $\CP^N$, including families of proper slant imbeddings of $\CP^n$ into $\CP^N$ which generalize the Veronese imbeddings.
\bibliographystyle{amsplain}
|
\section{Introduction}
\label{intro}
Presently, the standard cosmological model ($\Lambda$CDM) includes a constant term in Einstein equations and assumes that the universe is constituted, besides baryons, photons and neutrinos, by a dominant weakly interacting component of unknown nature dubbed dark matter. This model gives the best representation of different independent data sets as distances to type Ia supernovae, the angular power spectrum of the cosmic microwave background (CMB) or the angular distance scale of the baryon acoustic oscillations (BAO) (see, for instance, \cite{kowal08,tsuji04} and references therein).
Despite these successes, arguments against the inclusion of a cosmological term are often found in the literature. Among others, the so-called ``coincidence" problem and the dramatic difference between predicted and observed values of the cosmological constant, when interpreted as the vacuum energy density (however, see \cite{bianchi} for counter-arguments). Face to these hypothetical difficulties, alternative models able to explain the present accelerated phase of the cosmic expansion have been proposed in the past years (see \cite{tsuji04} for a short review). Among these models, we mention those in which a negative pressure, responsible for the present accelerated phase of the cosmic expansion, is associated to a process of particle creation \cite{zeldovich,calvao,zimdahl94,lima99,pavon,cardenas,lima09}.
In \cite{roany} cosmological models based on creation of cold dark matter particles (hereafter simply CCDM) and on open thermodynamic systems \cite{prigogine} were reviewed and, in particular, the case in which the entropy density production rate is equal to the particle density variation rate. Models satisfying this condition include practically all CCDM scenarios found in the literature. The derived background dynamics mimics that of the $\Lambda$CDM cosmology only if the (negative) pressure related to the creation process is constant. However, in the linear regime the growth of density perturbations differs from the standard model because, after attaining a maximum amplitude that depends on the adopted creation rate, the density contrast decreases \cite{roany}. This is certainly a difficulty for this class of cosmology but these results were questioned by \cite{lima10}, who criticised the Newtonian approach adopted by \cite{roany} in their analysis of linear perturbations.
Since CCDM models are still being investigated as a viable alternative to the $\Lambda$CDM cosmology (see, for instance, \cite{ioav,pan,berezin,Chakraborty:2014oya}), we report here a new investigation that differs from previous analyses in two main aspects. Firstly, we consider the evolution of both dark matter and baryons either in the background or in the linear regime. Secondly, a fully relativistic analysis of the linear perturbations of both components is performed. As found by previous authors, we will show that the background dynamics of the CCDM cosmology is formally identical to the $\Lambda$CDM model and, consequently, satisfies the same tests mentioned above. However, in the CCDM model the density ratio between dark matter particles and baryons varies and presently, as we shall see, the theory predicts a value of about 19.5, which is considerably higher than that derived from the analysis of massive X-ray clusters. We will show that in the linear regime the growth of baryons is not affected by the particle creation process but that of dark matter is, as it was pointed out by \cite{roany}. These aspects represent potential problems to be surpassed by the CCDM cosmology. This paper is organised as follows: in Section~\ref{Sec2} the main properties of the CCDM model are reviewed, in Section~\ref{Sec3} the relativistic linear equations are derived and finally, in Section~\ref{Sec4} the main conclusions are given.
\section{The CCDM model}\label{Sec2}
In the CCDM cosmology, non-relativistic dark matter particles are assumed to be produced continuously and are supposed to act like a fluid. According to \cite{prigogine}, consider that the universe is an open thermodynamic system in which particles are created at the expense of the gravitational field. Then, let $\mu$ be the chemical potential associated to the variation of the number of particles. In this case, the Euler relations can be written as
\begin{equation}
\label{eum}
\mu n = \mathfrak{h} -Ts\;,
\end{equation}
and
\begin{equation}
Tds = d\rho - \mu dn\;.
\label{edois}
\end{equation}
In the equations above, $n$ is the dark matter particle density, $s$ and $\rho$ are respectively the entropy and the energy densities, $\mathfrak{h}=P+\rho$ is the enthalpy density and $T$ is the temperature. From these equations one obtains trivially
\begin{equation}
\label{etres}
\left(\frac{d\rho}{dt}-\frac{\mathfrak{h}}{n}\frac{dn}{dt}\right)=sT\left(\frac{1}{s}\frac{ds}{dt}-
\frac{1}{n}\frac{dn}{dt}\right)\;,
\end{equation}
which is the Gibbs relation for the dark matter fluid. As already mentioned, a common assumption in most CCDM models is that the right side of eq.~\eqref{etres} be zero. This is equivalent to say that the entropy production and the particle variation rates are equal. In other words, the entropy per particle $s/n$ remains constant during the expansion process. It should be emphasized that although this last condition be also verified in the $\Lambda$CDM model, there is an important difference with respect to the CCDM cosmology: in the former the expansion is adiabatic while in the later entropy is produced as a consequence of the particle creation process. In this case, under the assumption of a constant $s/n$ ratio eq.~\eqref{etres} becomes simply
\begin{equation}
\label{equatro}
\frac{d\rho}{dt}=\frac{\mathfrak{h}}{n}\frac{dn}{dt}\;.
\end{equation}
Following \cite{roany}, let us introduce the stress-energy tensor for the dark matter fluid as
\begin{equation}
\label{ecinco}
T_{ik}=(\rho + \Pi)u_iu_k - \Pi g_{ik}\;,
\end{equation}
where $\Pi = P(\rho)+P_{\rm c}$ is the effective pressure acting on the fluid, with the first term representing the pressure due to kinetic motions and interactions between particles while the second represents the effective pressure associated to the process of particle creation. Assuming a flat Friedmann-Robertson-Walker (FRW) metric, i.e.,
\begin{equation}
\label{eseis}
ds^2 = -dt^2 + a^2(t)\left(dr^2 + r^2d\Omega^2\right)\;,
\end{equation}
and using the conservation equation $T^k{}_{0;k}=0$, one obtains
\begin{equation}
\label{esete}
\frac{d\rho}{dt}+3H(\mathfrak{h}+P_{\rm c})=0\;,
\end{equation}
where, as usual, $H={\dot a}/a$ is the Hubble parameter. Combining now eqs.~\eqref{equatro} and \eqref{esete}, an explicit expression for the pressure associated to the particle creation process can be obtained, i.e.,
\begin{equation}
\label{eoito}
P_{\rm c} = -\frac{\mathfrak{h}}{3H}\left(3H+\frac{1}{n}\frac{dn}{dt}\right)\;.
\end{equation}
Note that if ${\dot n}>0$ and the background is expanding ($H>0$), the pressure $P_{\rm c}$ is negative. Moreover, if the created particles are ``cold" (non-relativistic) or, in other words, their kinetic energy and pressure can be neglected, their enthalpy is simply
$\mathfrak{h}=\rho=mnc^2$, where $m$ is the mass of the dark matter particles. In this case, eq.~\eqref{eoito} can be rewritten as
\begin{equation}
3HP_{\rm c} = -3Hnmc^2-mc^2\frac{dn}{dt}\;.
\end{equation}
Multiplying both sides of this equation by $a^3$, one obtains easily
\begin{equation}
\label{enove}
P_{\rm c}\frac{da^3}{dt}+\frac{d(nmc^2a^3)}{dt}=0\;.
\end{equation}
The equation above permits a simple interpretation of the creation pressure $P_{\rm c}$. The work done by this stress to expand a unit comoving volume of the universe is equal to the rate of energy in the form of new particles appearing in the same comoving volume. A past study by \cite{mccrea} on matter creation in the context of general relativity led to the same interpretation.
\subsection{The Boltzmann-Einstein approach}
In the previous Section, the basic equation defining $P_{\rm c}$ was derived using a thermodynamic approach and assuming an open system. Here, for the sake of completeness, the Boltzmann equation in a curved space is used to re-derive the equations of the model.
With the metric given by eq.~\eqref{eseis}, the Boltzmann equation for the dark matter fluid is \cite{oliver}
\begin{equation}
\label{boltzmann}
\frac{\partial f}{\partial t}-Hp\frac{\partial f}{\partial p}=C[f]\;.
\end{equation}
In the equation above, using the same notation as in \cite{oliver}, $p$ is the modulus of the 3-momentum and $f$ is the one-particle distribution function satisfying the condition
\begin{equation}
\label{density}
n = \int f(p)d^3p\;.
\end{equation}
The right side of eq.~\eqref{boltzmann} is not a truly collisional term but a source term intending to represent phenomenologically the creation process. A simple expression for this term can be derived in the following way. First, we integrate over momenta eq.~\eqref{boltzmann} to obtain
\begin{equation}
\label{firstmomentum}
{\dot n}+3Hn = \int C[f]d^3p\;.
\end{equation}
Comparing the equation above with eq.~\eqref{esete} and recalling that $h=\rho=nmc^2$, one obtains
\begin{equation}
\label{collisionterm}
\int C[f]d^3p = -\frac{3HP_{\rm c}}{mc^2}\;.
\end{equation}
In a further step, assume the following ansatz for the distribution function, $f(p,a)=f_0(pa)g(a)$, where $f_0(pa)$ is a solution of eq.~\eqref{boltzmann} without the source or ``collisional" term (see \cite{oliver}). Replacing this into eq.~\eqref{boltzmann}, one obtains
\begin{equation}
\label{collisionb}
f_0\frac{dg(a)}{da}Ha = C[f]\;.
\end{equation}
Substitute the equation above into eq.~\eqref{collisionterm} to get, after some straightforward calculations,
\begin{equation}
\label{enoveb}
\frac{dg(a)}{da}=-\frac{3P_{\rm c}}{\rho(a)}\frac{g(a)}{a}\;.
\end{equation}
The solution of eq.~\eqref{esete} gives the evolution of the dark matter energy density and is given by eq.~\eqref{equatorze} for the case in which the creation pressure is constant, an assumption also made here. Replacing the expression for $\rho(a)$ into eq.~\eqref{enoveb} gives
\begin{equation}
\label{enovec}
\frac{dg(a)}{da}=\frac{3a^2}{a^3+\beta}g(a)\;,
\end{equation}
where the parameter $\beta$ was defined by
\begin{equation}
\label{beta}
\beta + 1=-\frac{\rho_{\rm dm0}}{P_{\rm c}}\;,
\end{equation}
with $\rho_{\rm dm0}$ being the present dark matter energy density. Integration of eq.~\eqref{enovec} is trivial and one gets $g(a)=K(\beta+a^3)$, where $K$ is an integration constant. Now, substitute this result into eq.~\eqref{collisionb} to obtain
\begin{equation}
\label{source1}
C[f]=3KHa^3f_0\;.
\end{equation}
Using these results, the distribution function that satisfies eq.~\eqref{boltzmann} with the collisional term above is $f=Kf_0(\beta+a^3)$. Combining this with eq.~\eqref{source1} permits to recast the collisional term as
\begin{equation}
\label{source}
C[f]=\frac{3Ha^3}{\beta + a^3}f\;.
\end{equation}
Note that when $P_{\rm c}$ goes to zero from the negative side, $\beta$ goes to infinity and the collisional term goes to zero as wished, since no particles are being created. Using the same reasoning, the integration constant may be put equal to $K=1/\beta$ since in the limit of zero creation pressure, the distribution function must be essentially $f_0$, the solution of the Vlasov-Einstein equation.
In a further step, replace the derived expression for the source term into eq.~\eqref{boltzmann}. Then multiply both sides of the equation by the particle energy $E$ and integrate over the 3-momentum to obtain
\begin{equation}
\label{energy1}
{\dot\rho}+3H(\rho + P)=\frac{3Ha^3}{\beta + a^3}\rho\;,
\end{equation}
where the energy density and the pressure were defined as usually, i.e.,
\begin{equation}
\rho = \int Efd^3p\;,
\end{equation}
and
\begin{equation}
\label{pressure}
P = \int\frac{p^2}{3E}fd^3p\;.
\end{equation}
Moving the right side of eq.~\eqref{energy1} to the left side, it is easy to verify that this is equivalent to add to the kinetic pressure a term equal to
\begin{equation}
P_{\rm c} = -\frac{\rho a^3}{\beta + a^3}\;.
\end{equation}
Since a constant creation pressure was assumed, the equation above is, as expected, the same as that resulting from the integration of eq.~\eqref{esete}, indicating consistency between both approaches.
It is interesting to mention that the creation process affects the kinetic pressure of dark matter particles. This can be seen integrating eq.~\eqref{pressure} with a new variable $x = pa$. In this case, one obtains
\begin{equation}
P=\frac{4\pi}{3}\frac{\beta + a^3}{\beta a^4}\int dxf_0(x)\frac{x^4}{\sqrt{x^2+m^2a^2}}\;.
\end{equation}
Since the created particles are non-relativistic, they satisfy the condition $m^2a^2 \gg x^2$. In this case, the above integral can be simplified and one obtains for the pressure
\begin{equation}
P=\frac{4\pi}{3m}\frac{\beta + a^3}{\beta a^5}\int x^4f_0(x)dx\;.
\end{equation}
When particles are not being created ($\beta \rightarrow \infty$), the kinetic pressure decays as $P\propto a^{-5}$, a well known result. However, when $\beta$ is finite, at high redshifts the pressure decays as expected but at redshift of the order of $1+z\sim 1/\beta$, the logarithmic slope $(d\log~P/d\log~a)$ is higher than -5, indicating effects of the creation process.
\subsection{Dynamics of the CCDM model}
In the present study the contribution of both baryons and dark matter to the dynamics of the universe is taken into account. However, the continuous particle creation process affects only the dark matter component, since baryons are supposed to be conserved. As it was shown in \cite{roany}, the CCDM model can mimic the $\Lambda$CDM cosmology only if a {\it constant} creation pressure is assumed, which here will be denoted as $P_{\rm c} = -\lambda$, where $\lambda$ is a positive constant with the dimension of an energy density. Assuming further that the thermodynamic pressure of baryons and dark matter could be neglected, the stress-energy tensor for dark matter and baryons can be recast as
\begin{equation}
\label{edez}
T_{ik}({\rm dm})=\left(\rho_{\rm dm}-\lambda\right)u_iu_k + \lambda g_{ik}\;,
\end{equation}
and
\begin{equation}
\label{eonze}
T_{ik}({\rm b})=\rho_{\rm b}u_iu_k\;,
\end{equation}
The subscripts ``dm" and ``b" refer respectively to dark matter and baryons. The conservation equations $T^k{}_{0;k}(\rm dm)=0$ and $T^k{}_{0;k}(\rm b)=0$ give respectively
\begin{equation}
\label{edoze}
\frac{d\rho_{\rm dm}}{dt}+3H\rho_{\rm dm} = 3H\lambda\;,
\end{equation}
and
\begin{equation}
\label{etreze}
\frac{d\rho_{\rm b}}{dt}+3H\rho_{\rm b} = 0\;.
\end{equation}
These equations are easily integrated in terms of the scale factor $a$ and one obtains
\begin{equation}
\label{equatorze}
\rho_{\rm dm} = \lambda + \frac{(\rho_{\rm dm0}-\lambda)}{a^3}\;,
\end{equation}
and
\begin{equation}
\label{equinze}
\rho_{\rm b} = \frac{\rho_{b0}}{a^3}\;.
\end{equation}
The constants $\rho_{\rm dm0}$ and $\rho_{b0}$ denote the present ($a=1$) values of the dark matter and the baryon energy densities.
Using eqs.~\eqref{equatorze} and \eqref{equinze}, the Hubble equation can be written as
\begin{equation}
\label{edezesseis}
H^2 = \frac{8\pi G}{3c^2}\left[\lambda+\frac{(\rho_{\rm dm0}+\rho_{\rm b0}-\lambda)}{a^3}\right]\;.
\end{equation}
The first point to be noticed in this equation is that the present value of the Hubble parameter is fixed only by the present total matter content $\rho_{\rm t0} = \rho_{\rm dm0} + \rho_{\rm b0}$ of the universe as in the Einstein-de Sitter model, namely
\begin{equation}
\label{hubble}
H^2_0 = \frac{8\pi G}{3c^3}\rho_{\rm t0}\;.
\end{equation}
The second point is that eq.~\eqref{edezesseis} can be written in a form equivalent to the $\Lambda$CDM model, i.e., $H^2=H^2_0(\Omega_{\rm v}+\Omega_{\rm m0}a^{-3})$, if the following identifications are made
\begin{equation}
\label{omega}
\Omega_{\rm v} = \frac{\lambda}{\rho_{\rm t0}}\;,
\end{equation}
and
\begin{equation}
\label{omegam}
\Omega_{\rm m0} = \frac{(\rho_{\rm dm0}+\rho_{\rm b0}-\lambda)}{\rho_{\rm t0}}\;.
\end{equation}
Since the Hubble equation in the CCDM model is formally identical to that of the $\Lambda$CDM cosmology, one should expect that data on supernova distances, BAO angular distances, variation of the Hubble parameter with $z$ will be equally explained by both cosmological models. This identification led to different authors to conclude that CCDM and $\Lambda$CDM models are indistinguishable at the background level. However there are some subtleties in this analysis. From the equations above, the ratio between dark matter and baryons energy densities is
\begin{equation}
\label{edezesete}
\frac{\rho_{\rm dm}}{\rho_{\rm b}}=\frac{\left[(1-\Omega_{\rm m0})a^3+(\Omega_{\rm m0}-\Omega_{\rm b0}\right]}{\Omega_{\rm b0}}\;.
\end{equation}
The cosmological data from Planck \cite{planck} gives for the total matter density parameter $\Omega_{\rm m0}=0.315$ whereas that of the baryons is $\Omega_{\rm b0}=0.04872$. Using these values, eq.~\eqref{edezesete} at the baryon-radiation decoupling or at still higher $z$ gives a ratio $\rho_{\rm dm}/\rho_{\rm b}$ = 5.46. This is essentially the Planck value for the $\Lambda$CDM model, implying that the CCDM cosmology predicts the same amplitude for the peaks of the angular power spectrum of the CMB and again both models agree. However, since dark matter particles are being produced continuously, the present value of the ratio between dark matter and baryon energy densities predicted by the CCDM model is $\rho_{\rm dm}/\rho_{\rm b}$ = 19.5. Observations of clusters of galaxies in the mass range $6\times 10^{13} - 1\times 10^{15} M_{\odot}$ reported by \cite{gonzalez} indicate a nearly constant value for this ratio and equal to $\rho_{\rm dm}/\rho_{\rm b}$ = 6.52. The $\rho_{\rm dm}/\rho_{\rm b}$ ratio derived from massive clusters is about 20\% higher than the ``cosmic" value derived from CMB data and the authors of this study suggest that a small baryon deficiency could be present, consequence of different processes operating in the formation of these large structures as, for instance, the feedback of AGNs. Turning the argument, this value also imposes some constraints on the possible amount of dark matter created after decoupling, which is considerably smaller than the growth by a factor of about 3.6 predicted by the CCDM model. This discrepancy represents a potential problem for the CCDM model already at the background level and may represent a test able to distinguish both cosmologies.
However, the authors of reference \cite{ioav} have a different interpretation for the evolution of the baryon-to-dark matter ratio. They assume that the dark matter energy density is the sum of two distinct terms: the first, denoted by $\rho_{\rm cre}$ corresponds to particles that are being created continuously, forming a uniform background. According to them, these particles do not participate in the formation process of structures; the second term, denoted by $\rho_{\rm con}$, corresponds to particles formed in a short timescale at very high redshifts, which are ``conserved" and are able to cluster. In eq.~\eqref{equatorze}, they made the identifications $\rho_{\rm cre}=\lambda$ and $\rho_{\rm con}= (\rho_{\rm dm0}-\lambda)/a^3$. Clearly, if only ``conserved" particles participate in the formation of structures, the aforementioned problem with the baryon-to-dark matter ratio disappear. Such an interpretation can be criticised by the following reasons. The first concerns the identification of $\rho_{\rm cre}$ with $\lambda$. Since the density of this component is constant, this implies that newly created particles are permanently in a steady state or, in other words, the rate at which particles are being created compensates exactly the losses due to the expansion of the universe, requiring fine tuning that represents a very special situation. The second aspect concerns the clustering of the ``conserved" term only. This is based on the fact that the perturbation of the dark matter energy density is
\begin{equation}
\delta\rho_{\rm dm}=\delta\left(\rho_{\rm cre}+\rho_{\rm con}\right)=\delta\rho_{\rm con}\;,
\end{equation}
since $\delta\rho_{\rm cre}=0$, because $\rho_{\rm cre}$ is a constant. In this case, only the term $\delta\rho_{\rm con}$ contributes to the Poisson equation or to the gravitational potential, leading to the idea that particles created continuously do not give any contribution to the potential and thus, they do not cluster. In other words, the newly created particles would not feel the gravitational forces, representing a unlikely situation since all dark matter particles are expected to be indistinguishable. In the general case, using the same components as defined in \cite{ioav}, insert the total dark matter energy density $\rho_{\rm dm} = \rho_{\rm cre}+\rho_{\rm con}$ into eq.~\eqref{edoze} that can be split in two equations: one including the source term, which is obeyed by the component $\rho_{\rm cre}$ and another, without the source term, which is satisfied by the component $\rho_{\rm con}$. Under these conditions, the general solution for the created component is
\begin{equation}
\rho_{\rm cre}=\lambda + \frac{K_1}{a^3}\;.
\end{equation}
The integration constant $K_1$ can be estimated by requiring that when $\lambda \rightarrow 0$ the created component must disappear. Then, $K_1=\lambda$. On the other side, the general solution for the conserved component is
\begin{equation}
\rho_{\rm con}=\frac{K_2}{a^3}\;.
\end{equation}
The integration constant $K_2$ can be derived by imposing that at $a=1$ (present time) the dark matter energy density be equal to $\rho_{\rm dm0}$. Consequently, one obtains $K_2=(\rho_{\rm dm0}-2\lambda$) and hence the total dark matter energy density is recovered by adding both components, i.e.,
\begin{equation}
\rho_{\rm dm}=\left[\lambda + \frac{\lambda}{a^3}\right]+\left[\frac{(\rho_{\rm dm0}-2\lambda)}
{a^3}\right]=\lambda+\frac{(\rho_{\rm dm0}-\lambda)}{a^3}\;.
\end{equation}
By such a procedure eq.~\eqref{equatorze} is recovered and it becomes clear that the ``conserved" term includes both components in contradiction with the hypothesis raised by \cite{ioav}.
\section{Linear perturbations}\label{Sec3}
In the present investigation, the evolution of density perturbations either of the baryonic or of the dark matter fluids will be considered using a fully relativistic approach. We follow here the treatment and the notation as in \cite{ma}, adopting a synchronous gauge. In this case the perturbed metric can be written as (the convention $c=1$ is adopted, excepting cases when recovering all constants is necessary)
\begin{equation}
ds^2=-dt^2+a^2(t)\left(\delta_{ik}+h_{ik}\right)dx^idx^k\;.
\end{equation}
Combining eqs. 20a and 20c given in reference \cite{ma} but including both dark matter and baryons and taking the derivatives with respect to the cosmic time and not with respect to the conformal time one obtains
\begin{equation}
\label{pum}
{\ddot h} + 2H{\dot h} = -8\pi G\left(\rho_{\rm dm}\delta_{\rm dm}+\rho_{\rm b}\delta_{\rm b}\right)\;,
\end{equation}
where $h$ is the trace of the metric perturbation tensor $h_{ik}$ and, as usually, the density contrast for baryons $\delta_{\rm b} = \delta\rho_{\rm b}/\rho_{\rm b}$ and for dark matter $\delta_{\rm dm}=\delta\rho_{\rm dm}/\rho_{\rm dm}$ was introduced.
Perturbations of the conservation equations are given by eqs. 28 and 29 of reference \cite{ma}. Here they are explicitly given for the two fluid components under the following assumptions: zero sound velocity for both fluids and the parameter $w=P/\rho$ being zero for baryons and equal to $w=-\lambda/\rho_{\rm dm}$ for dark matter. These approximations imply that the perturbations are adiabatic.
However, since there is entropy production in the matter creation process,
non-adiabatic perturbations are also possible and they will be analyzed in a
future work. Taking again the derivatives with respect to the cosmic time one obtains
\begin{equation}
\label{pdois}
{\dot\delta_{\rm b}}=-\frac{1}{2}{\dot h}-\frac{\theta_{\rm b}}{a}\;,
\end{equation}
and
\begin{equation}
\label{ptres}
{\dot\delta_{\rm dm}}=-(1+w)\left(\frac{{\dot h}}{2}+\frac{\theta_{\rm dm}}{a}\right)+3Hw\delta_{\rm dm}\;,
\end{equation}
where $\theta = \delta u^i{}_{;i}$ is the divergence of the peculiar velocity, obeying the equations below, derived also from eqs. 28 and 29 given in reference \cite{ma} and under the assumptions mentioned above,
\begin{equation}
\label{pquatro}
{\dot\theta_{\rm b}}=-H\theta_{\rm b}\;,
\end{equation}
and
\begin{equation}
\label{pcincoa}
{\dot\theta_{\rm dm}}=-(1-3w)H\theta_{\rm dm}-\frac{{\dot w}}{1+w}\theta_{\rm dm}\;.
\end{equation}
Performing the time derivative of the relation $w=-\lambda/\rho_{\rm dm}$ and using eq.~\eqref{edoze} one obtains
\begin{equation}
\label{evinte}
{\dot w}=3Hw(1+w)\;.
\end{equation}
Replacing this result into eq.~\eqref{pcincoa}, one obtains simply
\begin{equation}
\label{pcincob}
{\dot\theta_{\rm dm}}=-H\theta_{\rm dm}\;.
\end{equation}
The integration of eqs.~\eqref{pquatro} and \eqref{pcincob} is trivial and one obtains $\theta_{\rm b} \propto 1/a$ and $\theta_{\rm dm} \propto 1/a$. These are decaying solutions that justify to neglect the $\theta$ terms in eqs.~\eqref{pdois} and \eqref{ptres}. Under
this condition, eq.~\eqref{pdois} simplifies to
\begin{equation}
\label{pdoisb}
{\dot\delta_{\rm b}}=-\frac{1}{2}{\dot h}\;.
\end{equation}
Replacing this equation and its time derivative into eq.~\eqref{pum} results
\begin{equation}
\label{pumb}
{\ddot\delta_{\rm b}}+2H{\dot\delta_{\rm b}}-4\pi G(\rho_{\rm b}\delta_{\rm b}+\rho_{\rm dm}\delta_{\rm dm})=0\;.
\end{equation}
Neglecting the $\theta$ term also in eq.~\eqref{ptres} and using eq.~\eqref{pdoisb} results
\begin{equation}
\label{ptresb}
{\dot\delta_{\rm dm}}=(1+w){\dot\delta_{\rm b}}+3Hw\delta_{\rm dm}\;.
\end{equation}
A further step in the solution of this equation is to introduce the reduced dark matter density contrast by the relation
\begin{equation}
\label{pseis}
{\tilde\delta_{\rm dm}}=\frac{\delta\rho_{\rm dm}}{(\rho_{\rm dm}-\lambda)}=
\frac{\delta_{\rm dm}}{1+w}\;.
\end{equation}
Taking the time derivative of this equation and using eq.~\eqref{evinte} one obtains
\begin{equation}
\label{psete}
\frac{d{\delta_{\rm dm}}}{dt}=(1+w)\frac{d{\tilde\delta_{\rm dm}}}{dt}+3Hw(1+w)
{\tilde\delta_{\rm dm}}\;.
\end{equation}
Substituting eq.~\eqref{psete} into eq.~\eqref{ptresb} gives simply $d{\tilde\delta_{\rm dm}}/dt=d\delta_{\rm b}/dt$ or ${\tilde\delta_{\rm dm}}=\delta_{\rm b}+ \rm constant$. Taking the integration constant equal to zero and making use of the definition of the reduced density contrast, one obtains finally
\begin{equation}
\label{dmcontrast}
\delta_{\rm dm}=(1+w)\delta_{\rm b}\;.
\end{equation}
Replacing the equation above into eq.~\eqref{pumb} one gets for the evolution of the baryon density contrast
\begin{equation}
\label{pumc}
{\ddot\delta_{\rm b}}+2H{\dot\delta_{\rm b}}-4\pi G\left[\rho_{\rm b}+(1+w)\rho_{\rm dm}\right]\delta_{\rm b}=0\;.
\end{equation}
Making use of eqs.~\eqref{hubble}, \eqref{edezesete} and after some algebra, the equation above can be recast as
\begin{equation}
\label{pumd}
{\ddot\delta_{\rm b}}+2H{\dot\delta_{\rm b}}-\frac{3}{2}H_0^2\Omega_{\rm m0}a^{-3}\delta_{\rm b}=0\;.
\end{equation}
This equation for the evolution of the baryon density contrast is formally the same as that for the $\Lambda$CDM model, indicating that baryons are not directly affected by the matter creation process. However, in the CCDM model, the evolution of the dark matter density contrast and that of the baryons are not the same as it occurs in the $\Lambda$CDM cosmology. They are related by eq.~\eqref{dmcontrast} and since the parameter $w$ is not a constant, the evolution of both components is not proportional to each other. Figure \ref{fig:1} shows the evolution of the density contrast for both fluids, using the cosmological parameters given in reference \cite{planck}. The linear growth of baryons coincides with that of the $\Lambda$CDM model but the dark matter fluid displays a different behaviour. After reaching a maximum around $z\sim$ 1, the amplitude of the density contrast decreases, in agreement with the conclusions by \cite{roany}, who have performed a Newtonian analysis and with those of \cite{ioav}, who have performed an one-fluid investigation based either on the Neo-Newtonian approximation or on a relativistic approach.
\begin{figure}
\center
\includegraphics[width=0.5\columnwidth]{Figure1.eps}
\caption{Evolution of the density contrast for baryons (upper curve) and dark matter (lower curve as a function of the scale factor $a$.}
\label{fig:1}
\end{figure}
How to explain the behaviour of the dark matter density contrast? The evolution of the background dark matter density is given by eq.~\eqref{equatorze}. This equation indicates that at high $z$ (or $a \ll 1$) the dark matter background evolves as the baryon background, i.e., inversely proportional to the cube of the scale parameter $a$. Using the spherical model as a guide, in the early phases of the linear regime the background ``expands" faster than the density perturbation, leading to an increase of density contrast for both fluids as it can be seen in fig.~\ref{fig:1}. However, as a consequence of the particle creation process, in late phases the dark matter background varies slower than the baryon background (not affected in the creation process), tending to a constant value fixed by the parameter $\lambda$. Such a modification in the background expansion rate produces a decrease in the amplitude
of the dark matter density contrast. A further argument in favour of this explanation is provided by eq.~\eqref{dmcontrast}. Using eqs.~\eqref{omega} and \eqref{omegam}, the eq.~\eqref{dmcontrast} can be recast as
\begin{equation}
\delta_{\rm dm}=\left[1+\frac{\Omega_{\rm v}a^3}{(\Omega_{\rm m0}-\Omega_{\rm b0})}\right]^{-1}\delta_{\rm b}\;.
\end{equation}
This equation shows clearly that when $a\ll 1$, the contrast of both components is the same but as soon as the particle creation term becomes relevant, the dark matter density contrast decreases with respect to that of the baryon component, since now the background of both fluids evolves differently as explained previously.
\begin{figure}
\center
\includegraphics[width=0.5\columnwidth]{Figure2.eps}
\caption{Evolution of the total matter density contrast as a function of the scale factor $a$.}
\label{fig:2}
\end{figure}
In figure \ref{fig:2} is shown the evolution of the total matter density contrast defined by
\begin{equation}
\delta_{\rm m} =\frac{\rho_{\rm b}}{(\rho_{\rm b}+\rho_{\rm dm})}\delta_{\rm b}+\frac{\rho_{\rm dm}}{(\rho_{\rm b}+\rho_{\rm dm})}\delta_{\rm dm}\;.
\end{equation}
Since the dark matter component is dominant, the behaviour of the total matter density contrast is similar to that of dark matter.
\begin{figure}
\center
\includegraphics[width=0.5\columnwidth]{Figure3.eps}
\caption{Evolution of the growth factor for the $\Lambda$CDM model (upper curve) and for the CCDM model (lower curve) as a function of the scale parameter $a$.}
\label{fig:3}
\end{figure}
In figure \ref{fig:3}, the growth factor $f(z)=d\log\delta_{\rm m}/d\log a$ is shown for both the standard model and the CCDM model. Note that the decreasing amplitude branch of the density contrast in the CCDM model implies in {\it negative} values for growth factor, in disagreement with observational data given by \cite{hudson}, but that are
consistent with predictions of the $\Lambda$CDM model.
Different conclusions were reached by \cite{ioav} since they have defined a density contrast only for the component able to cluster. In this case, the evolution of the density contrast has a weak dependence on the parameter $\lambda$ and the growth factor is not inconsistent with data.
\section{Conclusions}\label{Sec4}
The main physical properties of the creation cold dark matter (CCDM) cosmology were revisited. When the thermodynamical approach by \cite{prigogine} is considered, the CCDM model is able to mimic the standard model if two main assumptions are made: the first is to consider that the entropy density production rate is equal to the particle variation rate and the second, is to hypothesise that the pressure associated to the creation process is constant. Generally, even if the creation pressure is not assumed to be constant, the first hypothesis and the conservation equation $T^k{}_{0;k}(\rm dm)=0$ lead to a simple interpretation of such a pressure: it corresponds to the work required to expand a unit comoving volume, which is equal to the energy of the (``cold") particles created in the same volume.
The solution of the Boltzmann-Einstein equation, including a pseudo-collisional term intending to represent the matter creation process, is consistent with the thermodynamical approach. The cosmic evolution of the kinetic or thermodynamic pressure of the newly created particles is affected by the creation process itself. At high redshifts, the kinetic pressure varies as $P\propto a^{-5}$, as expected for the standard model but at late phases the pressure decreases with a slower rate, consequence of the creation process.
The resulting Hubble equation for the CCDM cosmology is formally identical to that of a flat $\Lambda$CDM model. Thus, both models are indistinguishable with respect to tests like the supernova distances, the angular distances of the baryon acoustic oscillations and the variation of the Hubble parameter with redshift. However, fixing the parameters of the Hubble equation in order to identify both models, an important difference between both cosmologies become evident. Since dark matter particles are being created continuously, the ratio between dark matter and baryon energy densities varies. At high redshift it agrees with Planck's data and no conflict with predicted and observed amplitudes of the angular power spectrum of the CMB exists. However, the dark matter-to-baryon ratio increases with decreasing redshift and presently is expected to be about 19.5. Such a high value disagrees with data on massive clusters of galaxies, which indicates values about 3 times smaller. This is certainly a difficulty for the CCDM model.
Previous analysis of the density contrast evolution in the linear regime using a Newtonian approximation has shown that there is a maximum amplitude that depends on the creation rate \cite{roany}. The present investigation adopts a fully relativistic approach and takes into account both baryons and dark matter particles. The present study indicates that baryons are not affected by the creation process and that the growth of perturbations in the linear regime coincides with that in the $\Lambda$CDM model. However, this is
not the case for dark matter particles. After reaching a maximum amplitude around $z\sim$1, the density contrast decreases as a consequence of the matter creation process that modifies the evolution of background, as explained in the previous section. The decreasing branch of the matter density contrast produces a negative growth rate of cosmic structures in contradiction with observational data. The difficulties pointed here raise doubts about the CCDM theory as a viable alternative to the standard model.
\acknowledgments
JAFP thanks the program ``Science without Borders" of the Brazilian Government for the financial support to this research and the Federal University of the Espírito Santo State (Brazil) for his hospitality. The authors thank CNPq and Fapes for support and are also grateful to Ioav Waga for discussions.
\bibliographystyle{JHEP}
|
\section{Introduction}
Let $X$ be a connected smooth projective variety of dimension $n$.
We will denote in this paper ${\rm CH}_i(X)$ the Chow groups of $X$
with rational coefficients and ${\rm CH}_i(X)/{\rm alg}$ the groups
of $i$-cycles of $X$ with $\mathbb{Q}$-coefficients modulo algebraic
equivalence.
Let
$a\in{\rm CH}_0(X)$ be a $0$-cycle of degree $1$ on $X$. Following
Gross-Schoen \cite{grossschoen} and O'Grady \cite{ogrady}, let us
consider for $m\geq2$ the following $n$-cycle $\Gamma^m(X,a)$ in
$X^m$, which is a modification of the $m$-th small diagonal of $X$:
\begin{eqnarray} \label{eqformgammam} \Gamma^m(X,a)=\sum_{I\subset
\{1,\ldots,m\},|I|=i<m}(-1)^ip_I^*(a^{*i})\cdot p_J^*\Delta_{m-i}\in {\rm
CH}_n(X^m)_\mathbb{Q}, \end{eqnarray}
where
\begin{itemize}
\item $\{1,\ldots,m\}$ is the disjoint union of $I$ and $J$,
\item $p_I:X^m\rightarrow X^i$, resp. $p_J:X^m\rightarrow X^{m-i}$ are the
projections onto the products of factors indexed by $I$, resp. $J$,
\item $\Delta_{m-i}$ is the small diagonal of $X^{m-i}$, $\Delta_1=X$,
\item $a^{*i}\in {\rm
CH}_0(X^i)$ is defined by
\begin{eqnarray} \label{eqstar} a^{*i}=p_1^*a\cdot\ldots \cdot p_i^*a.
\end{eqnarray}
\end{itemize}
For example, for $m=2$, we have $\Gamma^2(X,a)=\Delta_X-a\times
X-X\times a$ and $\Gamma^2(X,a)=0$ if and only if $X=\mathbb{P}^1$
or a point. The modified small diagonal $\Gamma^3(X,a)$ appears in
several recent works. Gross and Schoen prove that $\Gamma^3(X,a)=0$
if $X$ is a hyperelliptic curve and $a$ is a Weierstrass point. This
result was greatly extended in \cite{colombovgeemen}
by Colombo and
van Geemen, who worked with
$1$-cycles modulo algebraic equivalence and proved that, for a
$d$-gonal curve $X$, the cycle $\Gamma^{d+1}
(X,a)$
is
algebraically equivalent to $0$. Although they do not state their
result in this form, but as the vanishing modulo algebraic
equivalence of the components $Z_s,\,s\geq d-1$ of the Beauville
decomposition (see \cite{beau}) of $X$ in its Jacobian, one can show
that this is equivalent to the vanishing of $\Gamma^{d+1}(X,a)$
modulo algebraic equivalence. For completeness, we will prove this
fact in subsection \ref{subsec}.
Concerning higher dimensional varieties,
Beauville and the author proved in \cite{beauvoi} the following theorem:
\begin{theo}\label{BV} Let $X$ be a $K3$ surface. Then there exists a
canonical degree $1$
zero-cycle $o_X$ of $X$ such that
\begin{eqnarray}
\Gamma^3(X,o_X)=0 \,\,{\rm in}\,\,{\rm CH}_2(X^3).
\label{eqformk3}
\end{eqnarray} In fact,
$o_X$ can be defined as the class in ${\rm CH}_0(X)$ of any point of $X$ lying on a (singular)
rational curve in $X$.
\end{theo}
In the paper \cite{ogrady}, O'Grady investigates $\Gamma^m(X,a)$ for
higher $m$. He proves the following results (for $X$ smooth
projective connected):
\begin{theo}\label{theoog} (O'Grady \cite{ogrady}) (i) The cycle $\Gamma^{n+1}(X,a)$ is
cohomologous to $0$, for $n={\rm dim}\,X$ and $q(X)=0$. More
generally $\Gamma^{m+1}(X,a)$ is cohomologous to $0$ if and only if
$m\geq {\rm dim}\,X+d$, where $d$ is the dimension of the image of
$X$ in its Albanese variety.
(ii) If $\Gamma^m(X,a)=0$ then $\Gamma^{m'}(X,a)=0$ for $m'\geq m$.
(iii) If $p:X\rightarrow Y$ is a ramified double cover and $a$ is a
branch point such that $\Gamma^m(Y, a)=0$, then for $m=2$ or
$m=3$, $\Gamma^{2m-1}(X,b)=0$, where $p(b)=a$.
\end{theo}
He conjectures that (iii) holds for any $m$ (see \cite[Conjecture
5.1]{ogrady}). One of our results is the proof of O'Grady's
conjecture, see (i) below, and a generalization to any degree, see
(ii) and (iii) below.
\begin{theo}\label{theointro2} Let $p:X\rightarrow Y$ be a degree $d$
finite morphism, where $X,\,Y$ are smooth projective and connected.
(i) Assume $d=2$, $a\in{\rm CH}_0(Y)$ is a $0$-cycle of degree $1$
supported on the branch locus of $p$, and $b:=\frac{1}{2}p^*a\in
{\rm CH}_0(X)$; if
$\Gamma^m(Y,a)=0$, then
$\Gamma^{2m-1}(X,b)=0$.
(ii) For any $d$, assume $a\in Y$ is a point such that the subscheme $p^{-1}(a)$
is supported on a point
$b\in X$. If
$\Gamma^m(Y,a)=0$, then
$\Gamma^{d(m-1)+1}(X,b)=0$.
(iii) For any $d$, let $b:=\frac{1}{d}p^*a$ for some
$0$-cycle $a\in{\rm CH}_0(Y)$ of degree $1$. If
$\Gamma^m(Y,a)=0$ in ${\rm CH}_n(Y^m)/{\rm alg}$, then
$\Gamma^{d(m-1)+1}(X,b)=0$ in ${\rm CH}_n(X^{d(m-1)+1})/{\rm alg}$.
\end{theo}
Statement (i) of Theorem \ref{theointro2} has been obtained
independently by Moonen and Yin \cite{moonenyin}.
\begin{rema}\label{remaref}{\rm When $Y=\mathbb{P}^n$, and $d\leq n+1$,
there always exists a point $a\in \mathbb{P}^n$ as in (ii) (cf.
\cite{fula}). In this case, we have $\Gamma^m(Y,a)=0$, with $m=n+1$,
hence we conclude that for $d$-th covers $X$ of $\mathbb{P}^n$ with
$d\leq n+1$, $\Gamma^{dn+1}(X,b)=0$, with $b=\frac{1}{d}p^*(pt)$.
Note also that any curve $X$ of genus $g$ admits a morphism of
degree $d\leq g+1$ to $\mathbb{P}^1$, which is totally ramified at
one given point $x$. Hence we get $\Gamma^{g+2}(X,x)=0$ for any
$x\in X$. This last result is also proved by Moonen and Yin
\cite{moonenyin} using the Colombo-van Geemen vanishing result.}
\end{rema}
\begin{rema} \label{remanew25aout} {\rm In the case where $Y$ is $\mathbb{P}^1$, so $X$ is a $d$-gonal
curve, Theorem \ref{theointro2}, (iii) gives the vanishing
$\Gamma^{d+1}(X, b)=0$ in ${\rm CH}_1(X^{d+1})/{\rm alg}$. As explained
in Subsection \ref{subsec}, this is equivalent to the
Colombo-van Geemen theorem \cite{colombovgeemen} mentioned above.}
\end{rema}
Another application of Theorem \ref{theointro2} is the following
result, which will be deduced from it in Section \ref{sec1} using
the smash nilpotence result
of \cite{voe} for cycles algebraically equivalent to $0$:
\begin{coro}\label{corosec1} Let $X$ be a smooth projective(connected) variety of dimension $n$. Then for any
$a\in {\rm CH}_0( X)$ of degree
$1$, there exists an integer $m$ such that $\Gamma^m(X,a)=0$ in ${\rm CH}^{(m-1)n}(X^m)$.
\end{coro}
Our second result is
the following more precise statement:
\begin{theo}\label{theointro1preciseintro}
Let $X$ be smooth projective connected of dimension $n$ and let
$a\in {\rm CH}_0( X)$ be of degree $1$. Then, if
$X $ is swept-out by irreducible curves of
genus $g$ supporting a zero-cycle rationally equivalent to $a$, and
$m\geq (n+1)(g+1)$,
one has
$\Gamma^m(X,a)=0 $ in ${\rm CH}^{(m-1)n}(X^m)$.
\end{theo}
Note that such a $g$ always exists: Indeed, consider curves in $X$
which are complete intersections of ample hypersurfaces containing
the support of the cycle $a$. For sufficiently high degree such
hypersurfaces, these curves will sweeep-out $X$, and thus we can
take for $g$ the genus of the generic such curves.
In the case where $X$ is a $K3$ surface, we know that $X$ is swept-out
by elliptic curves supporting the canonical $0$-cycle. Hence we get from Theorem
\ref{theointro1preciseintro}
the vanishing $\Gamma^6(X,o_X)=0$, which is not optimal in view
of
the relation (\ref{eqformk3}) in Theorem \ref{BV}.
We finally turn to the case of hyper-K\"ahler manifolds.
For $K3$ surfaces, one can get as a consequence of (\ref{eqformk3})
the following properties of $o_X$ (note however that property
\ref{itemi} below is used to prove (\ref{eqformk3}) so that we do
not actually recover it from (\ref{eqformk3}). Nevertheless, the
consequences \ref{itemi} and \ref{itemii} indicate that surfaces
satisfying (\ref{eqformk3}) are quite special):
\begin{enumerate}
\item \label{itemi} The intersection of two divisors $D,D'$ on $X$ is proportional to $o_X$ in ${\rm CH}_0(X)$.
\item \label{itemii} The second Chern class $c_2(X)$ is equal to $24\, o_X$.
\end{enumerate}
In the paper \cite{ogrady}, O'Grady formulates the following generalization
of (\ref{eqformk3}):
\begin{conj}\label{conjogrady} (O'Grady, \cite[Conjecture 0.1]{ogrady}) Let $X$ be a hyper-K\"ahler $n$-fold.
Then there exists a canonical $0$-cycle $o_X\in {\rm CH}_0(X)$ of degree $1$ such that
$\Gamma^{n+1}(X,o_X)=0$ in ${\rm CH}_n(X^{n+1})$.
\end{conj}
Note that by Theorem \ref{theoog}, (i), we have
$[\Gamma^{n+1}(X,o_X)]=0$ in $H^*(X^{n+1},\mathbb{Q})$ and that this
is optimal. Conjecture \ref{conjogrady} thus states that the cycles
$\Gamma^k(X,o_X)$ vanish in ${\rm CH}(X^k)$ once they vanish in
$H^*(X^k,\mathbb{Q})$, which is very different from the situation
encountered in the case of curves (except for the hyperelliptic
ones).
O'Grady establishes this conjecture for the punctual
Hilbert schemes $S^{[2]}$ and $S^{[3]}$ of a $K3$ surface. The
canonical $0$-cycle $o_X$, for $X=S^{[n]}$, is naturally defined as
the class in ${\rm CH}_0(X)$ of any point of $X$ lying over $no_S\in
S^{(n)}$, for some representative $o_S\in S$ of the canonical
$0$-cycle of $S$. We prove in section \ref{sec3} Conjecture
\ref{conjogrady} for punctual Hilbert schemes $X=S^{[n]}$ of $K3$
surfaces, and for any $n$,
using methods from \cite{voisinpamq} and recent results of Yin \cite{yin}:
\begin{theo} \label{theoHKintro} Let $S$ be a $K3$ surface, and let $X=S^{[m]}$. Then
\begin{eqnarray}
\Gamma^{n+1}(X,o_X)=0 \,\,{\rm in \,\, CH}_n(X^{n+1}).
\label{eqformk3prouve}
\end{eqnarray}
where $o_X$ is the canonical $0$-cycle on $X$ coming from the
canonical $0$-cycle of $S$, and $n={\rm dim}\,X=2m$.
\end{theo}
Note that one can recover from (\ref{eqformk3prouve}) the following
result, which had been in fact already proved in \cite[Theorem 1.5]{voisinpamq}.
\begin{coro} The intersection of $n$ divisors on $X$ is proportional to
$o_X$ in ${\rm CH}_0(X)$.
\end{coro}
For the proof of Theorem \ref{theoHKintro}, we will need three
tools. The first ingredient is similar to what we did in
\cite{voisinpamq}, namely we will use the de Cataldo-Migliorini
theorem \cite{decami} and will prove Proposition
\ref{proppourtheofin} in order to reduce to computations in the Chow
rings of the self-products $S^{k}$. The second ingredient is very
new and it is provided by Yin's recent result \cite{yin} saying that
the cohomological relations between the big diagonals of a regular
surface and the pull-back of the class of a point are generated
(modulo trivial relations) by the pull-backs of the Kimura relation
and the cohomological counterpart $[\Gamma^3(S,o_S)]=0$ in
$H^8(S^3,\mathbb{Q})$ of the relation (\ref{eqformk3}) (see also
\cite[Proposition 1.3]{ogrady}). We then argue that the Kimura
relation is not needed in our context, while the relation
$\Gamma^3(S,o_S)=0$ is satisfied in the Chow ring by Theorem
\ref{BV}.
To conclude, let us remark that the following conjecture in the same
spirit as Conjecture \ref{conjogrady} was stated first in
\cite{voisinpamq} for $K3$ surfaces, and then in \cite{shenvial} for
general hyper-K\"ahler manifolds:
\begin{conj} \label{conjabove} Let $X$ be a projective
hyper-K\"ahler manifold and $n> 0$ be an integer. Then there exists
a canonical $0$-cycle $o_X\in{\rm CH}_0(X)$ such that any polynomial
relation between the cohomology classes $pr_i^*[o_X],\,i\leq
n,\,pr_{ij}^*[\Delta_X],\,i\not=j\leq n$, already holds in ${\rm
CH}(X^n)$.
\end{conj}
O'Grady's conjecture \ref{conjogrady} is the particular case of
Conjecture \ref{conjabove} which concerns the class
$\Gamma^{n+1}(X,o_X)$, $n={\rm dim}\,X$. As explained in
\cite{voisinbook} in the case of $K3$ surfaces, Conjecture
\ref{conjabove} is extremely strong since it implies finite
dimensionality in the Kimura sense, with very important consequences
established by Kimura \cite{kimura}, in particular on the nilpotency
of self-correspondences homologous to $0$. O'Grady's conjecture
\ref{conjogrady} does not seem to have such implications, so it is
possibly of a nature different from Conjecture \ref{conjabove}.
The paper is organized as follows: in
section \ref{sec1} we introduce variants $\Gamma^{1,m}(X,a)$ of the
cycles $\Gamma^m(X,a)$ which lie in ${\rm CH}_n(X^{m+1})$, $n={\rm
dim}\,X$ and relate them to $\Gamma^m(X,a)$. In section \ref{sec2},
we will prove Theorem \ref{theointro1preciseintro}. Theorem
\ref{theointro2} will be proved in Section \ref{sec2bis} and Theorem
\ref{theoHKintro} will be proved in section \ref{sec3}. The last subsection \ref{secuniv} is devoted to the sketch of the proof of a general
theorem (Theorem
\ref{theopourtheofin}) concerning universally defined cycles on quasiprojective surfaces, which is used in the proof of Theorem \ref{theoHKintro}. This result
is of independent interest and its complete proof will
be given together with further applications in \cite{voisinuniversalcycles}.
\vspace{0.5cm}
{\bf Thanks.} This paper has been completed at the Institute for Advanced Study, which I thank for ideal working conditions. I also thank Lie Fu, Ben Moonen, Kieran O'Grady, Burt Totaro
and Qizheng Yin
for interesting discussions related to this work, and the referee for his/her
useful suggestions and comments.
\section{Cycles $\Gamma^{1,m}(X,a)$ \label{sec1}} We first introduce the
following notation: $X$ being smooth projective, and $a\in{\rm
CH}_0(X)_\mathbb{Q}$ being a zero-cycle of degree $1$, we define
$\Gamma^{1,m}(X,a)\in {\rm CH}_n(X^{m+1})_\mathbb{Q}$ by
\begin{eqnarray}\label{formulaforgamma1m}
\Gamma^{1,m}(X,a):=\prod_{1\leq i\leq
m}({p_{0i}}^*\Delta_X-{p_i}^*a),
\end{eqnarray}
where \begin{enumerate} \item $\Delta_X\subset X\times X$ is the
diagonal of $X$, \item $p_{0i}:X^{m+1}\rightarrow X\times X$ is the
projection on the product of the first and $i+1$-th factors,
\item
$p_i:X^{m+1}\rightarrow X$ is the projection on the $i+1$-th factor
(our factors are indexed by $\{0,\ldots,m\}$).
\end{enumerate}
The cycles $\Gamma^m(X,a)$ and $\Gamma^{1,m}(X,a)$ are related as
follows:
\begin{lemm}\label{legammamgamma1} We have the following formula:
\begin{eqnarray}\label{eqformproj}
\Gamma^m(X,a)=p'_{1,\ldots,m*}\Gamma^{1,m}(X,a),
\end{eqnarray}
where we index the factors of $X^{m+1}$ by $\{0,\ldots,m\}$ and
$p'_{1,\ldots,m}$ is the projection from $X^{m+1}$ to the products
$X^m$ of its last $m$ factors. We also have:
\begin{eqnarray}\label{eqformrec}
\Gamma^{m+1}(X,a)=\Gamma^{1,m}(X,a)-p_0^*a\cdot
{p'_{1,\ldots,m}}^*(\Gamma^m(X,a)).
\end{eqnarray}
\end{lemm}
\begin{proof} This is almost immediate. Developing the product in
(\ref{formulaforgamma1m}), we get
\begin{eqnarray}\label{eqformrec111}\Gamma^{1,m}(X,a)=\sum_{I\subset
\{1,\ldots,m\},|I|=i}(-1)^i{p'_I}^*(a^{*i})\cdot
p_{0,J}^*\Delta_{m+1-i}, \end{eqnarray}
where $I\bigsqcup J=\{1,\dots,
m\}$, $p_{0,J}$ is the projection from $X^{m+1}$ to the product
$X^{m+1-i}$ of factors indexed by $\{0\}\cup J$ and $p'_I$ is the
projection from $X^{m+1}$ to the product $X^i$ of the factors
indexed by $I\subset \{1,\ldots,m\}$. Applying
$p'_{1,\ldots,m*}:{\rm CH}_n(X^{m+1})_\mathbb{Q}\rightarrow {\rm
CH}_n(X^{m})_\mathbb{Q}$, we get by the projection formula, using
the fact that $p'_I=p_I\circ p'_{1,\ldots,m}$:
$$p'_{1,\ldots,m*}\Gamma^{1,m}(X,a)=\sum_{I\subset \{1,\ldots,m\},|I|=i}(-1)^ip_I^*(a^{*i})\cdot
p'_{1,\ldots,m*}(p_{0,J}^*\Delta_{m+1-i}).$$ Formula
(\ref{eqformproj}) then follows from the fact that
$p_J^*\Delta_{m-i}=p'_{1,\ldots,m*}(p_{0,J}^*\Delta_{m+1-i})$ in
${\rm CH}_n(X^{m})$.
As for (\ref{eqformrec}), we first write formula
(\ref{eqformgammam}) for $X^{m+1}$, where as above we index the
factors of $X^{m+1}$ by $\{0,\ldots,m\}$. This gives us
\begin{eqnarray} \label{eqformgammam+1}
\Gamma^{m+1}(X,a)=\sum_{I\subset \{0,\ldots,m\},i=|I|\leq
m}(-1)^i{p'_I}^*(a^{*i})\cdot {p'_J}^*\Delta_{m+1-i}\in {\rm
CH}_n(X^{m+1})_\mathbb{Q}. \end{eqnarray} We now separate the terms
where $0\not\in I$, which by (\ref{eqformrec111}) exactly give
$\Gamma^{1,m}(X,a)$, and the terms where $0\in I$, which exactly
give $-p_0^*a\cdot {p'_{1,\ldots,m}}^*(\Gamma^m(X,a))$.
\end{proof}
We deduce the following
\begin{prop}\label{propeq} The vanishing of $\Gamma^m(X,a)$ in ${\rm CH}_n(X^m)$ is equivalent to the
vanishing of $\Gamma^{1,m}(X,a)$ in ${\rm CH}_n(X^{m+1})$.
\end{prop}
\begin{proof} Indeed, if $\Gamma^{1,m}(X,a)=0$ then
$\Gamma^m(X,a)=0$ by (\ref{eqformproj}). Conversely, if
$\Gamma^m(X,a)=0$, then \cite[Proposition 2.4]{ogrady} shows that
also $\Gamma^{m+1}(X,a)=0$. Formula (\ref{eqformrec}) then implies
that $\Gamma^{1,m}(X,a)=0$.
\end{proof}
A consequence of this result is the following statement comparing
$\Gamma^m(X,a)$ and $\Gamma^m(X,b)$, for two $0$-cycles
$a,\,b\in{\rm CH}_0(X)$ of degree $1$.
\begin{coro} If $\Gamma^m(X,a)=0$ and the cycle $b-a$ satisfies $(b-a)^{*k}=0$ in
${\rm CH}_0(X^k)$, then $\Gamma^{m+k}(X,b)=0$.
\end{coro}
Here we refer to (\ref{eqstar}) for the definition of the $*$-product (or external product) of cycles.
\begin{proof} Indeed, by Proposition \ref{propeq}, the assumption is equivalent to
the vanishing conditions:
$$\Gamma^{1,m}(X,a)=\prod_{i=1}^{i=m}(p_{0i}^*(\Delta_X)-p_i^*a)=0\,\,{\rm in}\,\, {\rm CH}_n(X^{m+1}),
\,\,n={\rm dim}\,X$$
$$\prod_{i=1}^{i=k}p_{i}^*(b-a)=0\,\,{\rm in}\,\,{\rm CH}_0(X^k).$$
On the other hand, the conclusion is equivalent to the vanishing
$$\Gamma^{1,m+k}(X,a)=\prod_{i=1}^{m+k}(p_{0i}^*(\Delta_X)-p_i^*b)=0\,\,{\rm in}\,\,
{\rm CH}_n(X^{m+k+1}).$$
We now write $b=a+(b-a)$, getting
$$\Gamma^{1,m+k}(X,b)=\prod_{i=1}^{m+k}((p_{0i}^*(\Delta_X)-p_i^*a)-p_i^*(b-a))$$
and develop the product.
In the developed expression, the product of
$\geq m$ terms of the form $p_{0i}^*(\Delta_X)-p_i^*a$ is $0$ and the product of
$\geq k$ terms of the form $p_i^*(b-a)$ is $0$. Hence we conclude that
each monomial in the development is $0$.
\end{proof}
Here is another corollary of Proposition \ref{propeq}. It shows how to deduce Corollary \ref{corosec1} from Theorem
\ref{theointro2}, and thus gives another proof of the nilpotency statement
of Theorem \ref{theointro1preciseintro}, with no estimate on the nilpotency index.
\begin{coro} Let $X$ be a smooth projective connected variety and let $a$ be a $0$-cycle
of degree $1$ on $X$ such that $\Gamma^m(X,a)=0$ in ${\rm
CH}(X^m)/{\rm alg}$. Then for any $0$-cycle $b$ of degree $1$ on
$X$, there is an integer $M$ such that $\Gamma^M(X,b)=0$ in ${\rm
CH}(X^M)$.
\end{coro}
\begin{proof} As $a$ and $b$ are algebraically equivalent, we also have
$\Gamma^m(X,b)=0$ in ${\rm CH}(X^m)/{\rm alg}$. By Proposition
\ref{propeq}, which is true and proved in the same way for cycles
modulo algebraic equivalence (observing that \cite[Proposition
2.4]{ogrady} is true as well for cycles modulo algebraic
equivalence), this is equivalent to the fact that
$\Gamma^{1,m}(X,b)$ is algebraically equivalent to $0$ in $X^{m+1}$.
By the smash-nilpotence result of Voevodsky \cite{voe}, there is an
integer $N$ such that the cycle $\Gamma^{1,m}(X,b)^{*N}$ vanishes
identically in ${\rm CH}(X^{N(m+1)})$. Thus its restriction to
$X^{Nm+1}$ embedded in $X^{N(m+1)}$ by the small diagonal on the
factors of index $0,m+1,2m+1,\ldots,(N-1)m+1$ also vanishes in ${\rm
CH}(X^{Nm+1})$. But this restricted cycle is nothing but
$\Gamma^{1,Nm}(X,b)$.
\end{proof}
The following criterion for the vanishing of $\Gamma^m(X,a)$ will be
used in Section \ref{sec2bis}. Here we consider more generally the
vanishing of $\Gamma^m(X,a)$ modulo an adequate equivalence relation
$R$ which in applications will be rational or algebraic equivalence.
We need an assumption on the $0$-cycle $a$ of degree $1$, namely
\begin{eqnarray}\label{eqcyclea}
p_1^*a\cdot p_2^*a=\Delta_*a \,\,{\rm in}\,\,{\rm CH}_0(X\times X)/R,
\end{eqnarray}
where $\Delta$ is the diagonal inclusion map of $X$ in $X\times X$.
This assumption is satisfied for any $R$
if $a$ is a point, or for any $0$-cycle if $R$ is algebraic equivalence, and $X$ is connected.
\begin{prop}\label{propourdm1}
Assume $a$ satisfies (\ref{eqcyclea}). Then
$\Gamma^m(X,a)=0$ in ${\rm CH}(X^m)/R$ if and only if
$$\Gamma^{1,m-1}(X,a)=p_0^*a\cdot \Gamma\,\,{\rm in}\,\,{\rm CH}_n(X^m)/R,\,n={\rm dim}\,X$$
for some cycle $\Gamma\in {\rm CH}_{2n}(X^m)/R$.
\end{prop}
The proof of Proposition \ref{propourdm1} will use the following
\begin{lemm}\label{lepourprop22avril} Assume the degree $1$ zero-cycle $a$ of $X$
satisfies (\ref{eqcyclea}). Then for any
$Y$ and any cycle $\Gamma\in {\rm CH}(X\times Y)/R$, the following formula holds:
$$p_X^*a\cdot \Gamma=p_X^*a\cdot p_Y^*\Gamma_a \,\,{\rm in}\,\,{\rm
CH}(X\times Y)/R,$$ where
$$\Gamma_a:=p_{Y*}(p_X^*a\cdot \Gamma)\in {\rm CH}(Y)/R.$$
\end{lemm}
\begin{proof} Let $a=\sum_in_ia_i$, where $a_i\in X$. Then
$$p_X^*a\cdot \Gamma=\sum_in_ip_X^*a_i\cdot \Gamma=\sum_in_i a_i\times \Gamma_{a_i}=\sum_in_ip_X^*a_i\cdot
p_Y^*\Gamma_{a_i},$$ where $\Gamma_{a_i}\in {\rm CH}(Y)/R$ is the
restriction of $\Gamma$ to $a_i\times Y$. So we need to prove that,
assuming (\ref{eqcyclea}),
\begin{eqnarray}\label{eqpouraa22avril} \sum_in_ip_X^*a_i\cdot
p_Y^*\Gamma_{a_i}=p_X^*a\cdot p_Y^*\Gamma_a\,\,{\rm in}\,\,{\rm
CH}(X\times Y)/R,
\end{eqnarray}
where $\Gamma_a=\sum_in_i\Gamma_{a_i}\in {\rm CH}(Y)/R$. Note that
(\ref{eqcyclea}) is exactly the case of (\ref{eqpouraa22avril})
where $X=Y$ and $\Gamma$ is the diagonal of $X$. The general case
is then deduced from this one by introducing $X\times X\times Y$
with its various projections to $X, \,Y$ and $X\times Y$. Namely,
let \begin{itemize} \item $p_{i,X,Y}:X\times X\times Y\rightarrow
X\times Y$, $i=1,\,2$, be the projections onto the product of the
first and the third (resp. the second and the third) factor,
\item
$p_{i,X}:X\times X\times Y\rightarrow X,\,i=1,\,2$ be the projection
on the
first, resp. second, factor
and \item $p_{X,X}:X\times X\times Y\rightarrow X\times Y\rightarrow
X\times X$ be the projection onto the product of the first two
factors. \end{itemize}
The left hand side of (\ref{eqpouraa22avril})
is clearly equal to
$$p_{2,X,Y*}(p_{1,X}^*a\cdot p_{X,X}^*\Delta_X\cdot p_{2,X,Y}^*\Gamma).$$
Formula (\ref{eqcyclea}) tells us that
on $X\times X$, $p_1^*a\cdot\Delta_X=p_1^*a\cdot p_2^*a$ modulo $R$,
so that
\begin{eqnarray}\label{autreeqpouraa22avril}p_{2,X,Y*}(p_{1,X}^*a\cdot p_{X,X}^*\Delta_X\cdot
p_{2,X,Y}^*\Gamma)=
p_{2,X,Y*}(p_{1,X}^*a\cdot p_{2,X}^*a\cdot p_{2,X,Y}^*\Gamma)
\,\,{\rm in}\,\,{\rm CH}(X\times Y)/R.
\end{eqnarray}
As $a$ has degree $1$, the right hand side of
(\ref{autreeqpouraa22avril}) is equal by the projection formula to
$$p_X^*a\cdot p_Y^*\Gamma_a,$$
proving (\ref{eqpouraa22avril}).
\end{proof}
\begin{proof}[Proof of Proposition \ref{propourdm1}] We have by (\ref{eqformrec})
$$\Gamma^m(X,a)=\Gamma^{1,m-1}(X,a)-p_0^*a\cdot p_{1,\ldots,m-1}^*\Gamma^{m-1}(X,a)$$
so if $\Gamma^m(X,a)=0$ in ${\rm CH}(X^m)/R$, we get
$$\Gamma^{1,m-1}(X,a)=p_0^*a\cdot p_{1,\ldots,m-1}^*\Gamma^{m-1}(X,a)\,\,{\rm in }\,\,{\rm CH}(X^m)/R.$$
This proves one direction (for which we do not need
(\ref{eqcyclea})). In the other direction, we assume
(\ref{eqcyclea}) and
\begin{eqnarray}\label{autreeqpouraa22avril111}
\Gamma^{1,m-1}(X,a)=p_0^*a\cdot \Gamma\,\,{\rm in}\,\,{\rm
CH}_n(X^m)/R
\end{eqnarray}
for some cycle $\Gamma\in {\rm CH}_{2n}(X^m)/R$.
We now use Lemma \ref{lepourprop22avril} which gives
$$p_0^*a\cdot \Gamma=p_0^*a\cdot p_{1,\ldots,m-1}^*(p_{1,\ldots,m-1*}(p_0^*a\cdot \Gamma)).$$
By (\ref{autreeqpouraa22avril111}), this gives
$$p_0^*a\cdot \Gamma=p_0^*a\cdot p_{1,\ldots,m-1}^*(p_{1,\ldots,m-1*}(\Gamma^{1,m-1}(X,a)))
\,\,{\rm in}\,\,{\rm CH}_n(X^m)/R.$$
As $p_{1,\ldots,m-1*}(\Gamma^{1,m-1}(X,a))=\Gamma^{m-1}(X,a)$ by
(\ref{eqformproj}), we get
$$\Gamma^{1,m-1}(X,a)=p_0^*a\cdot p_{1,\ldots,m-1}^*(\Gamma^{m-1}(X,a))\,\,{\rm in}\,\,{\rm CH}_n(X^m)/R.$$
Using (\ref{eqformrec}), we conclude that
$$\Gamma^m(X,a)=\Gamma^{1,m-1}(X,a)-p_0^*a\cdot
p_{1,\ldots,m-1}^*(\Gamma^{m-1}(X,a))=0\,\,{\rm in}\,\,{\rm
CH}_n(X^m)/R.$$
\end{proof}
\section{Proof of Theorem \ref{theointro1preciseintro} \label{sec2}}
We prove in this section Theorem \ref{theointro1preciseintro}, that
is the following statement :
\begin{theo}\label{theo1precise} Let $X$ be a variety of dimension $n$ and let $a\in {\rm CH}_0(X)$
be of degree $1$.
If $X $ is swept-out by irreducible curves of genus $\leq g$
supporting a $0$-cycle rationally equivalent to $a$, and $m\geq
(n+1)(g+1)$, then $\Gamma^m(X,a)=0 $.
\end{theo}
Note that for $g=0$, we get the following corollary:
\begin{coro} \label{coroRC} Let $X$ be a rationally connected manifold of dimension $n$.
Then $\Gamma^{n+1}(X,o)=0$ for any point $o\in X$.
\end{coro}
This corollary will be improved at the end of this section in
Theorem \ref{theovariantRC}.
\begin{proof}[Proof of Theorem \ref{theo1precise}] By Proposition
\ref{propeq}, it suffices to prove the vanishing of
$\Gamma^{1,m}(X,a)$. Let us see $\Gamma^{1,m}(X,a)$ as a
correspondence between $X$ and $X^m$. Then for any $x\in X$, we have
$$ \Gamma^{1,m}(X,a)_{\mid x\times X^m}=(x-a)^{*m}\,\,{\rm
in}\,\,{\rm CH}_0(X^{m})_\mathbb{Q}.$$
Recall now the following result proved in \cite{voe}, \cite{voivoe}:
\begin{lemm} \label{levoe} Let $C$ be a smooth connected curve of genus $g$, and let
$z\in {\rm CH}_0(C)_\mathbb{Q}$ be a $0$-cycle of degree $0$ on $C$.
Then for $k>g$, $z^{*k}=0$ in ${\rm CH}_0(C^k)_\mathbb{Q}$.
\end{lemm}
Our assumption is now that $X$ is swept out by irreducible curves
of genus $\leq g$ supporting a $0$-cycle rationally equivalent to
$a$. This means that for any $x\in X$, there is a smooth connected
curve $C_x$ of genus $\leq g$ mapping to $X$ via a morphism $f_x$, a
point $x'\in C_x$ such that $f_x(x')=x$ and a $0$-cycle $a'\in {\rm
CH}_0(C_x)_\mathbb{Q}$ of degree $1$, such that $f_{x*}(a')=a$ in
${\rm CH}_0(X)_\mathbb{Q}$. It is then clear that
$${f_x^k}_*((x'-a')^{*k})=(x-a)^{*k}\,\,{\rm in}\,\,{\rm
CH}_0(X^k)_\mathbb{Q}.$$ We thus conclude by Lemma \ref{levoe} that
for $k>g$, and for any $x\in X$
\begin{eqnarray} \label{eq0k}
\Gamma^{1,k}(X,a)_{\mid x\times X^k}=(x-a)^{*k}=0\,\,{\rm
in}\,\,{\rm CH}_0(X^k)_\mathbb{Q}.
\end{eqnarray}
We use now the following
general principle which is behind
the Bloch-Srinivas decomposition of the diagonal
\cite{blochsrinivas}, see \cite[3.1]{voisinbook}:
\begin{theo} Let $\phi:W\rightarrow Y$ be a morphism, where $W$
is smooth of dimension $m$. Let $Z$ be a codimension $k$ cycle on
$W$. Assume that, for general $y\in Y$, the restriction $Z_{\mid
W_y}$ vanishes in ${\rm CH}^k(W_y)$. Then there is a dense Zariski
open set $U\subset Y$, such that $Z_U=0$ in ${\rm CH}^k(W_U)$.
Equivalently, there exist a nowhere dense closed algebraic
subset $D\subsetneqq Y$ and a cycle $Z'\in {\rm
CH}_{m-k}(W_D)_\mathbb{Q}$ such that
$$Z=Z'\,\,{\rm in\,\,CH}^k(W)_\mathbb{Q}.$$
\end{theo}
(Here we use the notation $W_D:=\phi^{-1}(D),\,W_U:=\phi^{-1}(U)$.)
Applying this statement to $Y=X,\,W=X^{k+1},\,\phi$ the projection
to the first factor and $Z=\Gamma^{1,k}(X,a)$, we conclude from
(\ref{eq0k}) that under the assumptions of Theorem
\ref{theo1precise}, there exists for $k>g$ a proper closed algebraic
subset $D\subsetneqq X$, such that $\Gamma^{1,k}(X,a)$ is rationally
equivalent to a cycle supported on $D\times X^k$.
Recall now the formula (\ref{formulaforgamma1m}) defining
$\Gamma^{1,k}$:
$$\Gamma^{1,k}(X,a):=\prod_{1\leq i\leq
k}(p_{0i}^*\Delta_X-p_i^*a).
$$
It follows immediately that
\begin{eqnarray}
\label{eqkk'} \Gamma^{1,k+k'}(X,a)=p_{0,1\leq i\leq
k}^*\Gamma^{1,k}(X,a)\cdot p_{0,k+1\leq i\leq
k+k'}^*\Gamma^{1,k'}(X,a), \end{eqnarray}
where $$p_{0,1\leq i\leq
k}:X^{k+k'+1}\rightarrow X^{k+1}$$ is the projection on the product
of the $k+1$ first factors and
$$p_{0,k+1\leq i\leq
k+k'}:X^{k+k'+1}\rightarrow X^{k'+1}$$ is the projection on the
product of the first factor (indexed by $0$) and the last $k'$
factors.
For $m\geq (n+1)(g+1)$, we write $m=(n+1)(g+1)+r$, for some $r\geq
0$ and we get from (\ref{eqkk'}):
$$\Gamma^{1,m}(X,a)=
p_{0,1\leq i\leq g+1}^*(\Gamma^{1,g+1})\cdot p_{0, g+2\leq i\leq
2(g+1)}^*(\Gamma^{1,g+1})$$ $$\ldots p_{0,n(g+1)+1\leq i\leq
(n+1)(g+1)}^*(\Gamma^{1,g+1})\cdot p_{0,(n+1)(g+1)+1\leq i\leq
(n+1)(g+1)+r}^*(\Gamma^{1,r}) .$$
Now we proved that the cycle $\Gamma^{1,g+1}$ is supported (via
the first projection $X^{g+2}\rightarrow X$) over a proper algebraic
subset $D\varsubsetneqq X$, and by the easy moving Lemma
\ref{lemoving} below, we can choose closed algebraic subsets
$D_1,\ldots, D_{n+1}$ such that $\cap_iD_i=\emptyset$ and
$\Gamma^{1,g+1}$ is supported (via the first projection
$X^{g+2}\rightarrow X$) over the proper algebraic subset
$D_i\varsubsetneqq X$ for each $i$.
Then we conclude that for $m\geq (n+1)(g+1)$, $\Gamma^{1,m}(X,a)$ is
supported (via the first projection $X^{(n+1)(g+1)+r+1}\rightarrow
X$) over the proper algebraic subset $\cap_iD_i=\emptyset$, and thus
is equal to $0$.
\end{proof}
\begin{lemm}\label{lemoving} Let $Y$ be irreducible and let $Z$ be a cycle on a product
$Y\times W$. Assume there exists a proper closed algebraic subset
$D\varsubsetneqq Y$ such that $Z$ is rationally equivalent to a
cycle $Z'$ supported on $D\times W$. Then for any finite set of
points $y_1,\ldots,y_l\in Y$, there is a $D'\varsubsetneqq Y$ such
that none of the $y_j$'s belongs to $D'$ and $Z$ is rationally
equivalent to a cycle $Z''$ supported on $D'\times W$.
\end{lemm}
\begin{proof} Let $\tau:\widetilde{D}\rightarrow D$
be a desingularization of $D\stackrel{i}{\hookrightarrow}Y$.
The cycle $Z'$ of $D\times W$ with rational coefficients lifts to a cycle
$\widetilde{Z}'$ of $\widetilde{D}\times W$. Let $\tilde{i}=i\circ
\tau:\widetilde{D}\rightarrow Y$ be the natural map and let
$\Gamma_{\tilde{i}}\subset \widetilde{D}\times Y$ be its graph.
Since $\Gamma_{\tilde{i}}$ has codimension $n={\rm dim}\,Y$, and
dimension $\leq n-1$, there is a cycle $\Gamma'\subset
\widetilde{D}\times Y$ rationally equivalent to $\Gamma_{\tilde{i}}$
and not intersecting $\widetilde{D}\times \{y_1,\ldots,y_l\}$. In
other words, $pr_2({\rm Supp}\,\Gamma')$ does not contain any of the
points $y_i$. We have by assumption
$$Z=({i},Id_W)_*Z'=(\tilde{i},Id_W)_*\widetilde{Z}'=(\Gamma_{\tilde{i}},Id_W)_*(\widetilde{Z}')$$
$$=(\Gamma',Id_W)_*(\widetilde{Z}')$$
in ${\rm CH}(Y\times W)$. Now, the cycle
$(\Gamma',Id_W)_*(\widetilde{Z}')$ is supported on $pr_2({\rm
Supp}\,\Gamma')\times W$, so the result is proved with $D'=pr_2({\rm
Supp}\,\Gamma')$, and $Z''=(\Gamma',Id_W)_*(\widetilde{Z}')$.
\end{proof}
To conclude this section, let us observe that the same scheme of proof proof applies to give the following
result, which is a generalization of Corollary \ref{coroRC}:
\begin{theo} \label{theovariantRC} Let $X$ be a connected smooth projective variety with ${\rm CH}_0(X)=\mathbb{Z}$.
Then for the canonical degree $1$
$0$-cycle $o$ on $X$, $\Gamma^{n+1}(X,o)=0$ in ${\rm CH}_n(X^{n+1})$, where $n={\rm dim}\,X$.
\end{theo}
\begin{proof} Indeed, the Bloch-Srinivas decomposition of the diagonal \cite{blochsrinivas}
gives an equality
$$\Delta_X-X\times o=Z\,\,{\rm in}\,\,{\rm CH}_n(X\times X),$$
where $Z$ is supported over $D\times X$, for some divisor $D\subset X$.
By Lemma \ref{lemoving}, we can write such a decomposition with
$n+1$ divisors $D_1,\ldots,D_{n+1}$ such that $\cap_iD_i=\emptyset$.
We then conclude that $\Gamma^{1,n+1}(X,o)=\prod_{i=1}^{n+1}p_{0i}^*(\Delta_X-X\times o)$ is equal to $0$ in
${\rm CH}_n(X^{n+2})$, and it follows from
Proposition \ref{propeq} that $\Gamma^{n+1}(X,o)=0$ in
${\rm CH}_n(X^{n+1})$.
\end{proof}
\section{Proof of Theorem \ref{theointro2}\label{sec2bis}}
We will first give the proof of Theorem \ref{theointro2}, (i). Let us recall the statement:
\begin{theo}\label{theogradyconj} Let $Y$ be smooth projective, and let $\pi:X\rightarrow Y$ be a
degree $2$ finite morphism, where $X$ is smooth projective. Let
$a\in {\rm CH}_0(Y)$ be a $0$-cycle of degree $1$ supported on the
branch locus of $\pi$. Then if $\Gamma^m(Y,a)=0$, we have
$\Gamma^{2m-1}(X,b)=0$, where $b=\frac{1}{2}\pi^*a\in {\rm
CH}_0(X)$.
\end{theo}\begin{rema}{\rm The assumption made on $a$ and $b$ is maybe not
optimal, but in any case the condition $b=\frac{1}{2}\pi^*a$ is not
sufficient. Indeed, consider the case where $Y$ is connected with
$\Gamma^m(Y,a)=0$, and $X$ consists of two copies of $Y$ with
$b=\frac{1}{2}\pi^*a\in {\rm CH}_0(X)$. Then $\Gamma^{k}(X,b)$ is
different from $0$ for any $k$ (in fact it is not even cohomologous
to $0$).}
\end{rema}
We will
denote by $\pi_2=(\pi,\pi):X\times X\rightarrow Y\times Y$. Let
$i:X\rightarrow X$ be the involution of $X$ over $Y$ and
$\Gamma_i\subset X\times X$ be its graph. We then have
$$\pi_2^*(\Delta_Y)=\Delta_X+\Gamma_i.$$
Let
$$\Delta_X^+=\pi_2^*(\Delta_Y)=\Delta_X+\Gamma_i,\,\,\Delta_X^-=\Delta_X-\Gamma_i.$$
We thus have \begin{eqnarray} \label{eqpetiteform}
2\Delta_X=\Delta_X^++\Delta_X^-.
\end{eqnarray}
\begin{lemm} \label{leutilepour23mars} Under the assumptions of Theorem \ref{theogradyconj}, we have
the following
equalities in ${\rm CH}_n(X\times X\times X)$, $n:={\rm dim}\,X$.
\begin{eqnarray}\label{eqegacruc}
p_{12}^*\Delta_X^-\cdot p_{13}^*\Delta_X^-=p_{12}^*\Delta_X^+\cdot
p_{23}^*\Delta_X^-,
\\
p_2^*b\cdot p_{23}^*\Delta_X^-=0, \label{eqegacrucab}
\end{eqnarray}
hence
\begin{eqnarray}\label{uneeqpourd2} p_{12}^*\Delta_X^-\cdot p_{13}^*\Delta_X^-=
p_{12}^*(\Delta_X^+-2p_2^*b)\cdot
p_{23}^*\Delta_X^-.
\end{eqnarray}
\end{lemm}
\begin{proof} We compute the left hand side of (\ref{eqegacruc}); we have:
$$p_{12}^*\Delta_X^-\cdot
p_{13}^*\Delta_X^-=p_{12}^*(\Delta_X-\Gamma_i)\cdot
p_{13}^*(\Delta_X-\Gamma_i)$$
$$=p_{12}^*\Delta_X\cdot p_{13}^*\Delta_X-p_{12}^*\Delta_X\cdot
p_{13}^*\Gamma_i-p_{12}^*\Gamma_i\cdot
p_{13}^*\Delta_X+p_{12}^*\Gamma_i\cdot p_{13}^*\Gamma_i.$$ We
observe now that
$$p_{12}^*\Delta_X\cdot p_{13}^*\Delta_X=p_{12}^*\Delta_X\cdot
p_{23}^*\Delta_X,\,\,\, p_{12}^*\Delta_X\cdot
p_{13}^*\Gamma_i=p_{12}^*\Delta_X\cdot p_{23}^*\Gamma_i,$$
$$p_{12}^*\Gamma_i\cdot
p_{13}^*\Delta_X=p_{12}^*\Gamma_i\cdot
p_{23}^*\Gamma_i,\,\,\,p_{12}^*\Gamma_i\cdot
p_{13}^*\Gamma_i=p_{12}^*\Gamma_i\cdot p_{23}^*\Delta_X.$$ It thus
follows that
\begin{eqnarray}\label{eqn1} p_{12}^*\Delta^-\cdot p_{13}^*\Delta^-=p_{12}^*\Delta_X\cdot
p_{23}^*\Delta_X-p_{12}^*\Delta_X\cdot
p_{23}^*\Gamma_i-p_{12}^*\Gamma_i\cdot
p_{23}^*\Gamma_i+p_{12}^*\Gamma_i\cdot p_{23}^*\Delta_X.
\end{eqnarray}
The right hand side of (\ref{eqn1}) is clearly equal to
$$(p_{12}^*\Delta_X +p_{12}^*\Gamma_i )\cdot(p_{23}^*\Delta_X-
p_{23}^*\Gamma_i),
$$
which is by definition $p_{12}^*\Delta_X^+\cdot p_{23}^*\Delta_X^-$,
thus proving formula (\ref{eqegacruc}).
In order to prove formula (\ref{eqegacrucab}), we use the fact that
the $0$-cycle $b$ can be written as $\sum_jn_j x_j$, where the
$x_j$'s are $i$-invariant. By linearity, it thus suffices to prove
(\ref{eqegacrucab}) when $b$ is an $i$-invariant point of $X$. Now
we have
$$p_2^*b\cdot p_{23}^*\Delta^-=p_{23}^*((b,b)-(b,ib))=0.$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{theogradyconj}]
By (\ref{eqformproj}), we have to prove that
$$p_{1,\ldots,2m-1*}(\Gamma^{1,2m-1}(X,b))=0\,\,{\rm in}\,\,{\rm CH}_n(X^{2m-1})_\mathbb{Q}.$$ Now, by
(\ref{formulaforgamma1m}) and (\ref{eqpetiteform}), using
$$2b=\pi^*a,\,\,\Delta_X^+=\pi_2^*\Delta_Y,$$
we get:
\begin{eqnarray}\label{eqnpourtheo1}
2^{2m-1}\Gamma^{1,2m-1}(X,b)=p_{01}^*(\pi_2^*\Delta_Y^a+\Delta_X^-)\cdot
\ldots \cdot p_{0,2m-1}^*(\pi_2^*\Delta_Y^a+\Delta_X^-).
\end{eqnarray}
Here we use the notation
$$\Delta_Y^a=\Delta_Y-p_2^*a\in {\rm CH}_n(Y\times Y)_\mathbb{Q},\,\,$$
so that we have $\Delta_X^+-2p_2^*b=\pi_2^*\Delta_Y^a$ and
(\ref{uneeqpourd2}) can be written as
\begin{eqnarray}\label{uneeqpourd2bis} p_{12}^*\Delta_X^-\cdot p_{13}^*\Delta_X^-=
p_{12}^*(\pi_2^*\Delta_Y^a)\cdot
p_{23}^*\Delta_X^-.
\end{eqnarray}
our assumption $\Gamma^m(Y,a)=0$ on $Y$ can be written using
Proposition \ref{propeq} as
\begin{eqnarray}\label{eqhyp}q_{01}^*\Delta_Y^a\cdot q_{02}^*\Delta_Y^a\cdot\ldots\cdot
q_{0m}^*\Delta_Y^a=0\,\,{\rm in}\,\,{\rm CH}_n(Y^{m+1})_\mathbb{Q},
\end{eqnarray}
where the $q_{0i}:Y^{m+1}\rightarrow Y\times Y$ are the projectors
onto the product of the first and $i+1$-th factors.
Denote by $\pi_r:X^r\rightarrow Y^r$. We then clearly have for any
$r$
\begin{eqnarray}
\label{eqntire} \pi_{r+1}^*(q_{01}^*\Delta_Y^a \cdot\ldots\cdot
q_{0r}^*\Delta_Y^a)=p_{01}^*(\pi_2^*\Delta_Y^a)\cdot \ldots \cdot
p_{0r}^*(\pi_2^*\Delta_Y^a)\,\,{\rm in}\,\,{\rm CH}(X^r),
\end{eqnarray}
and similarly for any
choice of indices $i_1,\ldots,i_r$ in $\{1,\ldots,2m-1\}$.
Developing now the product in (\ref{eqnpourtheo1}), we get a sum of
monomials which up to reordering the factors, take the form
\begin{eqnarray}
\label{eqntire1} p_{01}^*(\pi_2^*\Delta_Y^a)\cdot \ldots \cdot
p_{0r}^*(\pi_2^*\Delta_Y^a)\cdot
p_{0,r+1}^*\Delta_X^-\cdot\ldots\cdot p_{0,2m-1}^*\Delta_X^-
\end{eqnarray}
for some $r$.
These
terms vanish for $r\geq m$ by (\ref{eqntire}) and (\ref{eqhyp}).
We now conclude the proof as follows: The terms
$p_{0i}^*\Delta_{X}^-$ for $i\geq r+1$ can be grouped by pairs, and
there are at least $\llcorner\frac{2m-1-r}{2}\lrcorner$ such pairs.
By (\ref{uneeqpourd2bis}), for each such pair, we have
$$p_{0i}^*\Delta_{X}^-\cdot p_{0,i+1}^*\Delta_{X}^-=p_{0i}^*(\pi_2^*\Delta_Y^a)\cdot
p_{i,i+1}^*\Delta^-.$$ Hence each such pair produces a summand
$p_{0i}^*(\pi_2^*\Delta_Y^a)$. In total we get in (\ref{eqntire1})
at least $r+\llcorner\frac{2m-1-r}{2}\lrcorner$ factors of the form
$p_{0j}^*(\pi_2^*\Delta_Y^a)$. Now we have
$r+\llcorner\frac{2m-1-r}{2}\lrcorner\geq m$ unless $r=0$, and it
follows that (\ref{eqntire1}) vanishes for $r\geq1$. Hence we
proved that the only possibly nonzero monomial of the form
(\ref{eqntire1}) in the developed product (\ref{eqnpourtheo1}) is
$p_{01}^*(\Delta_X^-)\cdot \ldots \cdot p_{0,2m-1}^*\Delta_X^-$.
Thus we proved that
\begin{eqnarray}\label{eqnpourtheo2}
2^{2m-1}\Gamma^{1,2m-1}(X,b)=p_{01}^*(\Delta_X^-)\cdot \ldots \cdot
p_{0,2m-1}^*\Delta_X^-\,\,{\rm in}\,\,{\rm CH}(X^{2m}).
\end{eqnarray}
Let $i'$ be the involution $(i,Id,\ldots,Id)$ acting on $X^{2m}$.
Observe that each cycle $p_{0j}^*\Delta_X^-$ is skew-invariant under
${i'}^*$. It follows from (\ref{eqnpourtheo2}) that
$p_{01}^*(\Delta_X^-)\cdot \ldots \cdot p_{0,2m-1}^*\Delta_X^-$ is
skew-invariant under ${i'}^*$, hence also under $i'_*={i'}^*$. But
as we have $p_{1,\ldots,2m-1}\circ i'=p_{1,\ldots,2m-1}$, we have
$$\Gamma^{2m-1}(X,b)=p_{1,\ldots,2m-1*}(\Gamma^{1,2m-1}(X,b))=p_{1,\ldots,2m-1*}\circ i'_*(\Gamma^{1,2m-1}(X,b))$$
$$= -p_{1,\ldots,2m-1*}(\Gamma^{1,2m-1}(X,b))=-\Gamma^{2m-1}(X,b),$$ so that
$\Gamma^{2m-1}(X,b)=0$ in ${\rm CH}_n(X^{2m-1})$.
\end{proof}
We now turn to the proof of Theorem \ref{theointro2}, (ii) and (iii) : in fact, the
result will take the following
more precise form:
\begin{theo}\label{theodm} Let $\pi:X\rightarrow Y$ be a finite morphism of
degree $d$. If $\Gamma^m(Y,a)=0$ in ${\rm CH}(Y^m)/R$ for some
adequate equivalence relation $R$, and $b=\frac{1}{d}p^*a$ satisfies
\begin{eqnarray} \label{eqimportant} b*b=\Delta_*(b)\,\,{\rm in}\,\,{\rm CH}_0(X\times X)/R,
\end{eqnarray}
where $\Delta:X\rightarrow X\times X$ is the diagonal inclusion map,
then $\Gamma^{d(m-1)+1}(X,b)=0$ in ${\rm CH}(X^{dm})/R$.
\end{theo}
Statement (ii) of Theorem \ref{theointro2} is the case where $R$ is rational equivalence
(that is $R=0$) and $b$ is the class of
a point of $X$, as all points satisfy (\ref{eqimportant}) modulo rational equivalence.
Statement (iii) of Theorem \ref{theointro2} is the case where $R$ is algebraic equivalence.
Indeed, Theorem \ref{theodm} applies since the equality
$b*b=\Delta_*(b)\,\,{\rm in}\,\,{\rm CH}_0(X\times X)$ modulo algebraic equivalence
is satisfied by $0$-cycles of degree $1$ on a connected variety.
We first introduce some notation. Let as above
$\Delta_Y^a:=\Delta_Y-p_2^*a\in {\rm CH}_n(Y\times Y)$ and similarly
$\Delta_X^b:=\Delta_X-p_2^*b\in {\rm CH}_n(X\times X)$. In both
expressions, $p_2$ is the projection from $Y\times Y$, resp.
$X\times X$ onto its second factor. The proof of Theorem
\ref{theodm} will use the following result (which will replace
formula (\ref{eqegacruc}) used previously when $d=2$):
\begin{prop} \label{legroupd}
The morphism $\pi:X\rightarrow Y$ and the $0$-cycle $b$ being as in
Theorem \ref{theodm}, there exist cycles $\Gamma_i\in {\rm
CH}^{(d-1)n}(X^{d+1})$ such that
\begin{eqnarray}\label{eqformcrucpourdn}
\prod_{i=1}^dp_{0i}^*\Delta_X^{b}=\sum_ip_{0i}^*(\pi_2^*\Delta_Y^a)\cdot
p_{0,D\setminus\{i\}}^*\Gamma_i \,\,{\rm in}\,\, {\rm
CH}^{nd}(X^{d+1})/R,
\end{eqnarray}
where $D$ is the set $\{1,\ldots,d\}$ and as usual
$p_{0,D\setminus\{i\}}$ is the projection onto the product of the factors indexed
by the set $\{0\}\cup D\setminus\{i\}$.
\end{prop}
Before giving the proof, we will first prove a similar statement of independent interest
for $\Delta_X$ and $\Delta_Y$, instead of $\Delta_X^b$ and $\Delta_Y^a$, as the proof is much
simpler to write and we will use similar but slightly more involved arguments to
prove Proposition \ref{legroupd}. Namely we have the following result:
\begin{prop} \label{peutservir} Let $\pi:X\rightarrow Y$ be a finite morphism of degree
$d$. There exist cycles $\Gamma'_i\in {\rm CH}^{n(d-1)}(X^{d+1})$
such that
\begin{eqnarray}\label{eqformcrucpourdn2}
\prod_{i=1}^dp_{0i}^*\Delta_X=\sum_ip_{0i}^*(\pi_2^*\Delta_Y)\cdot
p_{0,D\setminus\{i\}}^*\Gamma'_i \,\,{\rm in}\,\, {\rm
CH}^{nd}(X^{d+1}).
\end{eqnarray}
\end{prop}
\begin{proof}
Indeed, let us denote by $E_{k}\subset {\rm CH}(X^{k+1})$ the ideal
generated by the elements $p_{0i}^*(\pi_2^*\Delta_Y)$, $i=1,\ldots,
k$. Next let \begin{eqnarray}\label{sigma118avril}
\Sigma_1:=\pi_2^{-1}(\Delta_Y)- \Delta_X\in {\rm CH}(X\times X).
\end{eqnarray} Note that, because
$\pi$ is finite of degree $d$, $\Sigma_1$ is the class of the
Zariski closure in $X\times X$ of the subvariety $\{(x,x_1)\in
X^0\times X^0,\,\pi(x_1)=\pi(x),\,x_1\not=x\}$ where
$X^0:=\pi^{-1}(Y^0)$ and $Y^0$ is the open set of $Y$ over which
$\pi$ is \'etale of degree $d$. The first projection
$pr_1:\Sigma_1\rightarrow X$ has degree $d-1$. Let us denote more
generally by $\Sigma_k\subset X^{k+1}$ the Zariski closure in
$X^{k+1}$ of the subvariety
\begin{eqnarray}
\label{defsigmak}\{(x,x_1,\ldots,x_k)\in
(X^0)^{k+1},\pi(x_i)=\pi(x),\,\,x_i\not=x_j\,\,{\rm
for}\,\,i\not=j,\,x_i\not=x\,\,{\rm for \,\,all}\,\,i\}.
\end{eqnarray}
The contents of formula (\ref{eqformcrucpourdn2}) is that
$\prod_{i=1}^dp_{0i}^*\Delta_X$ belongs to $E_{d}$. It is therefore a consequence of the following
statement:
\begin{claim} For any integer $k\geq 1$, one has
\begin{eqnarray}\label{eqsigma18avril}
\alpha_k\prod_{i=1}^kp_{0i}^*\Delta_X=\Sigma_k \,\,{\rm in
\,\,CH}(X^{k+1})/E_{k+1},
\end{eqnarray}
with $\alpha_k=(-1)^kk!$. In particular,
$\prod_{i=1}^dp_{0i}^*\Delta_X=0$ in ${\rm CH}(X^{d+1})/E_{d}$.
\end{claim}
The second statement follows from the first since $\Sigma_d$ is empty.
The first statement is proved by induction on $k$. For $k=1$, the
result is (\ref{sigma118avril}). The induction step is immediate:
we have the following equalities in ${\rm CH}(X^{k+1})$:
\begin{eqnarray}\label{eqsigma18avril111}\prod_{i=1}^{k+1}p_{0i}^*\Delta_X=p_{0,\ldots,k}^*
(\prod_{i=1}^{k}p_{0i}^*\Delta_X)\cdot p_{0,k+1}^*\Delta_X\\
\nonumber =-(\prod_{i=1}^{k}p_{0i}^*\Delta_X)\cdot
p_{0,k+1}^*\Sigma_1\,\,{\rm mod}\,\,E_{k+1}\\
\nonumber =-\frac{1}{\alpha_k}p_{0,\ldots,k}^*(\Sigma_k)\cdot
p_{0,k+1}^*\Sigma_1\,\,{\rm mod}\,\,E_{k+1}. \end{eqnarray} On the
other hand, we observe that $\Sigma_{k+1}$ is obtained from
$p_{0,\ldots,k}^*(\Sigma_k)\cdot p_{0,k+1}^*\Sigma_1$ by removing in
the fibered product the components where $x_{k+1}$ equals one of the
$x_i$'s for $i=1,\ldots,k$. This gives rise to the following
identity:
\begin{eqnarray}\label{numero118avril}
p_{0,\ldots,k}^*(\Sigma_k)
\cdot p_{0,k+1}^*\Sigma_1=\Sigma_{k+1}+\sum_{i=1}^kp_{0,\ldots,k}^*(\Sigma_k)\cdot
p_{i,k+1}^*\Delta_X.
\end{eqnarray}
In the right hand side of (\ref{numero118avril}), we can replace
(using again the induction hypothesis) $\Sigma_k$ by
${\alpha_k}\prod_{j=1}^{k}p_{0j}^*\Delta_X$ mod $E_k$ and we also
observe that
\begin{eqnarray}\label{numero118avri1111l}\prod_{j=1}^{k}p_{0j}^*\Delta_X\cdot
p_{i,k+1}^*\Delta_X=\prod_{i=1}^{k+1}p_{0i}^*\Delta_X
\end{eqnarray}
for any
$i=1,\ldots,k$. Hence we get, using (\ref{eqsigma18avril111}),
(\ref{numero118avril}) and (\ref{numero118avri1111l}),
$$\prod_{i=1}^{k+1}p_{0i}^*\Delta_X
=-\frac{1}{\alpha_k}\Sigma_{k+1}-k\prod_{i=1}^{k+1}p_{0i}^*\Delta_X.$$
This finally provides
$$ \alpha_{k+1}\prod_{i=1}^{k+1}p_{0i}^*\Delta_X=\Sigma_{k+1}$$
with $\alpha_{k+1}=-(k+1)\alpha_k$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{legroupd}]
We follow the above argument with $\Delta_X$, $\Delta_Y$ replaced
by $\Delta_X^b$ and $\Delta_Y^a$, in order to prove Lemma
\ref{le20avril} below. We use the following notation: we will work
with the $n$-cycle $\Sigma_k^b$ of $X^{k+1}$ obtained by replacing
formally in the definition (\ref{defsigmak}) of $\Sigma_k$ each
$x_i$ by $x_i-b$ and developing multilinearly. More rigorously,
$\Sigma_k$ admits morphisms $p,\,p_i:\Sigma_k\rightarrow X$,
obtained by restricting the projections $X^{k+1}\rightarrow X$
(where the factors are indexed by $\{0,\ldots,k\}$ and $p=p_0$). Let
$\Gamma_i\subset \Sigma_{k}\times X$ be the graphs of these
projections. Then we can obviously define $\Sigma_k\subset X^{k+1}$
as $(p,pr_{X^k})_*(\prod_{i=1}^kpr_{\Sigma_k,i}^*\Gamma_i)$, where
\begin{itemize}
\item $pr_{X^k}:\Sigma_k\times X^k\rightarrow X^k$ is the second
projection and $(p,pr_{X^k}):\Sigma_k\times X^k\rightarrow X^{k+1}$
is the obvious morphism.
\item $pr_{\Sigma_k,i}: \Sigma_k\times X^k\rightarrow \Sigma_k\times
X$ is the projection on the product of the first factor and the
$i$-th factor of $X^k$.
\end{itemize}
On the other hand, we also have in $\Sigma_{k}\times X$ the graph
$\Sigma_k\times\{b\}$ of the constant morphism mapping to $b$ if $b$
is a point, or more generally the $n$-cycle $pr_X^*b$ if $b$ is any
$0$-cycle of degree $1$. We then define analogously $\Sigma_k^b$ as
follows:
\begin{eqnarray} \Sigma_k^b=(p,pr_{X^k})_*(\prod_{i=1}^kpr_{\Sigma_k,i}^*(\Gamma_i-pr_X^*
b))\,\,{\rm in}\,\,{\rm CH}(X^{k+1}).
\end{eqnarray}
Developing the product above, we see that the formula for $\Sigma_k^b$ is
of the form
\begin{eqnarray}\label{eqpourSigmakb}\Sigma_k^b=\sum_{I\subset
\{1,\ldots,k\}}(-1)^{k-i}\lambda_{k,i,d}p_{0,I}^*\Sigma_i\cdot
p_{J}^*b^{*j}\in {\rm CH}_n(X^{k+1}),
\end{eqnarray}
where in the formula above, $I\sqcup J=\{1,\ldots,k\}$, $i=|I|$, and
the $\lambda_{k,j,d}$ are combinatorial coefficients given by the
formula
\begin{eqnarray}\label{eqpourlambdaij}
\lambda_{k,i,d}=(d-i-1)(d-i-2)\ldots(d-k).
\end{eqnarray}
Indeed, the reason for (\ref{eqpourlambdaij}) is the fact that the
projection map
$$p_{0,I}:\Sigma_k\rightarrow \Sigma_i\subset X^{i+1}$$
has degree $(d-i-1)(d-i-2)\ldots(d-k)$.
Note in particular, that $\Sigma_k^b=0$ for $k\geq d$.
Next we define
$E_{k,a,R}\subset {\rm CH}(X^{k+1})/R$ as the ideal generated by the
$p_{0,i}^*\Delta_Y^a$ for $i=1,\ldots,k$. Recall that
$\Gamma^{1,k}(X,b)=\prod_{i=1}^kp_{0i}^*\Delta_X^b$.
\begin{lemm} \label{le20avril}
The morphism $\pi:X\rightarrow Y$ and the $0$-cycle $b$ being as in Theorem \ref{theodm},
for any integer $k\geq 1$, one has
\begin{eqnarray}\label{eqsigma20avril}
\alpha_k\Gamma_{1,k}(X,b)=\Sigma_{k}^b \,\,{\rm in\,\,
CH}(X^{k+1})/E_{k,a,R},
\end{eqnarray}
\end{lemm}
\begin{proof}
We have by (\ref{eqpourSigmakb}), (\ref{eqpourlambdaij})
$$\Delta_X^b=\pi_2^*\Delta_Y^a-\Sigma_1^b,$$ which can be written as
$\Delta_X^b=-\Sigma_1^b$ mod $E_{1,a,R}$, proving the case $k=1$.
Assume the formula is proved for $k$. Then we have
\begin{eqnarray}\label{eqsigma20avril1}p_{0,1,\ldots,k}^*\Sigma_{k}^b\cdot p_{0,k+1}^*\Sigma_1^b=
-\alpha_k p_{0,1,\ldots,k}^*\Gamma^{1,k}(X,b)\cdot
p_{0,k+1}^*\Delta_X^b
\\ \nonumber
=-\alpha_k \Gamma^{1,k+1}(X,b)\,\,{\rm in}\,\,{\rm
CH}(X^{k+1})/E_{k,b,R}.
\end{eqnarray}
Next we claim that we have the following relation in ${\rm
CH}(X^{k+2})/R$:
\begin{eqnarray}\label{eqsigma21avril1}p_{0,1,\ldots,k}^*\Sigma_{k}^b\cdot p_{0,k+1}^*\Sigma_1^b=
\Sigma_{k+1}^b+\sum_{i=1}^{k}p_{0,\ldots,k}^*\Sigma_k^b\cdot p_{i,k+1}^*\Delta_X^b\\
\nonumber -\sum_{i=1}^{k}p_{0,\ldots,\hat{i},k+1}^*\Sigma_k^b\cdot
p_i^*b.
\end{eqnarray}
This relation uses in a crucial way the identity
\begin{eqnarray}\label{eqiddiagb26aout}
\Delta_*b=p_1^*b\cdot p_2^*b\,\,{\rm in}\,\,{\rm CH}_0(X\times X)/R.
\end{eqnarray}
The beginning
$$p_{0,1,\ldots,k}^*\Sigma_{k}^b\cdot p_{0,k+1}^*\Sigma_1^b=
\Sigma_{k+1}^b+\sum_{i=1}^{k}p_{0,\ldots,k}^*\Sigma_k^b\cdot
p_{i,k+1}^*\Delta_X^b+...$$ of the formula (\ref{eqsigma21avril1})
is easily understood: it expresses the fact that in the left hand
side, we include all possible $x_{k+1}\not=x$, while in
$\Sigma_{k+1}^b$, we have to take into account the restriction
$x_{k+1}\not=x_i$ for $i=1,\ldots,k$. The last term in
(\ref{eqsigma21avril1}) is explained as follows. The intersection
with $p_{i,k+1}^*\Delta_X^b=p_{i,k+1}^*\Delta_X-p_{k+1}^*b$ produces
a term
$$\Delta_*(x_i-b)-(x_i-b,b)=(x_i,x_i)-\Delta_*b-(x_i,b)+p_i^*bp_{k+1}^*b$$
$$= (x_i,x_i)-(x_i,b)$$
on the product of the $i$th and $k+1$th factors. On the other hand,
we had on the left in (\ref{eqsigma21avril1}) the term
$$(x_i-b,x_i-b)=(x_i,x_i)-(x_i,b)-(b,x_i)+(b,b)$$ on the product of the $i$th and
$k+1$th factors, which is unwanted in the development of
$\Sigma_{k+1}^b$. Hence we also have to add on the right the extra
term $-(b,x_i-b)$ on the product of the $i$th and $k+1$th factors,
which is exactly the meaning of the term
$-p_{0,\ldots,\hat{i},k+1}^*\Sigma_k^b\cdot p_i^*b$.
Thus the
claim is proved.
Combined with
(\ref{eqsigma20avril1}) and the inductive assumption,
(\ref{eqsigma21avril1}) gives
\begin{eqnarray}\label{eqsigma20avril2}
-\alpha_k \Gamma^{1,k+1}(X,b)= \Sigma_{k+1}^b
\\ \nonumber
+\alpha_k(\sum_{i=1}^{k}p_{0,\ldots,k}^* \Gamma^{1,k}(X,b)\cdot
p_{i,k+1}^*\Delta_X^b-\sum_{i=1}^{k}p_{0,\ldots,\hat{i},k+1}^*
\Gamma^{1,k}(X,b)\cdot p_i^*b).
\end{eqnarray}
The equality above holds in ${\rm CH}(X^{k+2})/E_{k+1,a,R}$. Let us
now prove that for any $i$,
\begin{eqnarray}\label{eq27mai}p_{0,\ldots,k}^*\Gamma^{1,k}(X,b)\cdot
p_{i,k+1}^*\Delta_X^b-p_{0,\ldots,\hat{i},k+1}^*
\Gamma^{1,k}(X,b)\cdot p_i^*b=\Gamma^{1,k+1}(X,b)
\end{eqnarray}
in ${\rm CH}(X^{k+2})/R$.
As
$$\Gamma^{1,k}(X,b)=\prod_{i=1}^kp_{0i}^*\Delta_X^b,\,\,\Gamma^{1,k+1}(X,b)=
\prod_{i=1}^{k+1}p_{0i}^*\Delta_X^b,$$ it clearly suffices to show
that the cycles $p_{01}^*\Delta_X^b\cdot
p_{12}^*\Delta_X^b-p_1^*b\cdot p_{02}^*\Delta_X^b$
and $p_{01}^*\Delta_X^b\cdot p_{02}^*\Delta_X^b$
of $X^3$ are equal in ${\rm CH}(X^{3})/R$. We have
$$p_{01}^*\Delta_X^b\cdot p_{12}^*\Delta_X^b-p_1^*b\cdot p_{02}^*\Delta_X^b= (p_{01}^*\Delta_X-p_1^*b)\cdot
(p_{12}^*\Delta_X-p_2^*b)-p_1^*b\cdot p_{02}^*\Delta_X+p_1^*b\cdot p_2^*b$$
$$=p_{01}^*\Delta_X\cdot p_{12}^*\Delta_X-p_{01}^*\Delta_X\cdot p_2^*b-
p_1^*b\cdot p_{12}^*\Delta_X+p_1^*b\cdot p_2^*b-p_1^*b\cdot p_{02}^*\Delta_X+p_1^*b\cdot p_2^*b$$
$$=p_{01}^*\Delta_X\cdot p_{02}^*\Delta_X-p_{01}^*\Delta_X
\cdot p_2^*b-p_1^*b\cdot p_{02}^*\Delta_X+p_1^*b\cdot p_2^*b$$ in
${\rm CH}(X^3)/R$ because we assumed $p_1^*b\cdot
p_{12}^*\Delta_X=p_1^*b\cdot p_2^*b $ in ${\rm CH}(X^3)/R$ (cf.
(\ref{eqimportant})). On the other hand,
$$p_{01}^*\Delta_X^b\cdot p_{02}^*\Delta_X^b=(p_{01}^*\Delta_X-p_1^*b)\cdot (p_{02}^*\Delta_X-p_2^*b)$$
$$=p_{01}^*\Delta_X\cdot p_{02}^*\Delta_X-p_{01}^*\Delta_X\cdot p_2^*b-
p_1^*b\cdot p_{02}^*\Delta_X+p_1^*b\cdot p_2^*b.$$ Hence we proved
that both terms in (\ref{eq27mai}) are equal; using
(\ref{eqsigma20avril2}), we then get:
$$-\alpha_k \Gamma^{1,k+1}(X,b)=
\Sigma_{k+1,b}+\alpha_k(\sum_{i=1}^{k}\Gamma^{1,k+1}(X,b))
,$$
hence
$$-(k+1)\alpha_k \Gamma^{1,k+1}(X,b)=\Sigma_{k+1,b}\,\,{\rm in}\,\,{\rm CH}(X^{k+2})/R$$
and Lemma \ref{le20avril} is proved.
\end{proof}
Finally, Lemma \ref{le20avril} for $k=d$ implies Proposition
\ref{legroupd} since $\Sigma_d^b=0$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{theodm}]
By Lemma \ref{legroupd} applied to each set of $d$ indices
$\{1,\ldots,d\},\,\{d+1,\ldots,2d\}$, $\{(m-2)d+1,\ldots,(m-1)d\}$,
we can write $\prod_{i=1}^{d(m-1)}p_{0i}^*\Delta_X^{b}$ as a sum of
products of $m-1$ cycles, each of them being of the form
$p_{0i}^*(\pi_2^*\Delta_Y^a)\cdot \Gamma''$ for
an adequate index
$i$ (one in each of the sets above). We now apply Proposition
\ref{propourdm1} to both $Y$ and $X$. Thus the assumption
$\Gamma^m(Y,a)=0$ implies that for some cycle $\Gamma_Y$ on $Y^m$,
$$\prod_{i=1}^{m-1}p_{0i}^*\Delta_Y^a=p_0^*a\cdot \Gamma_Y\,\,{\rm in}\,\,{\rm CH}(Y^m)/R.$$
Applying this relation to each product of $m-1$ factors
$\prod_{k=1}^{m-1}p_{0i_k}^*(\pi_2^*\Delta_Y^a)$ for adequate
indices $i_k$ appearing above, we conclude that
$$\Gamma^{1,d(m-1)}(X,b)=\prod_{i=1}^{d(m-1)}p_{0i}^*\Delta_X^{b}=p_{0}^*b\cdot\Gamma_X\,\,{\rm in}\,\,{\rm CH}(X^{d(m-1)+1})/R$$
for some cycle $\Gamma_X$ on $X^{d(m-1)+1}$. By Proposition
\ref{propourdm1}, and using the fact that $b$ satisfies property
(\ref{eqimportant}), (that is, condition (\ref{eqcyclea}) in
Proposition \ref{propourdm1}), we conclude that
$\Gamma^{d(m-1)+1}(X,b)=0$ in ${\rm CH}(X^{d(m-1)+1})/R$.
\end{proof}
\subsection{Case of curves\label{subsec}}
A special case of Theorem \ref{theointro2}, (iii) is the case where
$Y=\mathbb{P}^1$, so $X$ is a $d$-gonal curve. We then get the
vanishing $\Gamma^{d+1}(X,b)=0$ in ${\rm CH}_1(X^{d+1})/{\rm alg}$,
where $b$ is any point of $X$. Recall now the Beauville
decomposition of cycles on an abelian variety $A$ modulo rational or
algebraic equivalence:
$${\rm CH}_i(A)=\oplus_s{\rm CH}_i(A)_s,$$
with
$${\rm CH}_i(A)_s:=\{z\in {\rm CH}_i(A),\,\,\mu_{k*}z=k^{2i+s}z\,\,{\rm for\,\,all}\,\,k\in
\mathbb{Z}^*\}$$ and similarly for Chow groups modulo algebraic
equivalence. Here $\mu_k:A\rightarrow A$ is the morphism $a\mapsto
ka$. Let now $X$ be a smooth genus $g$ projective curve and
$A:=J(X)$. $X$ has an embedding in $J(X)$ which is canonical up to
translation, hence determines a $1$-cycle $Z$ in $J(X)$, well
defined modulo algebraic equivalence. Thus we have a Beauville
decomposition $$Z=\sum_s Z_s\,\,{\rm in}\,\, {\rm CH}_1(A)/{\rm
alg}.$$ For nonvanishing results concerning the cycles $Z_s$ (when
$X$ is very general) and its decomposition, let us mention
\cite{fakh}, \cite{voiinfjac} (in the later paper, it is proved that
if $g\geq s^2/2$, then $Z_{s}\not=0$ modulo algebraic equivalence
for a very general curve $X$ of genus $g$).
Let us show the
following:
\begin{prop} \label{propsansnom12mai} The vanishing of $\Gamma^{d+1}(X,b)$ in ${\rm
CH}_1(X^{d+1})/{\rm alg}$ is equivalent to the vanishing of
$Z_s,\,\forall s\geq{d-1}$, in ${\rm CH}_1( J(X))/{\rm alg}$.
\end{prop}
\begin{proof} It suffices to prove the result for $d\leq g-1$,
because we know by Theorem \ref{theointro2}, (iii) (see Remark
\ref{remanew25aout}) that $\Gamma^{d+1}(X,b)=0$ in ${\rm
CH}_1(X^{d+1})/{\rm alg}$ for some $d\leq g-1$. Assuming the
proposition proved for $d\leq g-1$, this implies that $Z_s=0$ in
${\rm CH}_1( J(X))/{\rm alg}$ for all $ s\geq{g-1}$, and thus for
$d\geq g$, both vanishing statements are true.
We thus assume $d\leq g-1$; note that the
cycle $\Gamma^{d+1}(X,b)$ is a $1$-cycle of $X^{d+1}$ which is
invariant under the action of the symmetric group $
\mathfrak{S}_{d+1}$, so that its vanishing in ${\rm
CH}_1(X^{d+1})/{\rm alg}$ is equivalent to the vanishing of its
image $\overline{\Gamma}^{d+1}(X,b)$ in ${\rm CH}_1(X^{(d+1)})/{\rm
alg}$. We now consider the inclusion
$$b_{g-d-1}:X^{(d+1)}\rightarrow X^{(g)}$$
$$z\mapsto z+(g-d-1)b$$
and claim that $\overline{\Gamma}^{d+1}(X,b)=0$ in ${\rm
CH}_1(X^{(d+1)})/{\rm alg}$ if and only if $b_{g-d-1
\,*}(\overline{\Gamma}^{d+1}(X,b))=0$ in ${\rm CH}_1(X^{(g)})/{\rm
alg}$. Indeed, there is an incidence correspondence
$$\Sigma\subset X^{(d+1)}\times
X^{(g)},\,\Sigma=\{(z,z'),\,z'=z+z''\,\,{\rm for\,\,some}\,\,z''\in
X^{(g-d-1)}\}.$$ It is not hard to see that, due to its special
form, the cycle $\overline{\Gamma}^{d+1}(X,b)$ satisfies
$$\Sigma^*(b_{g-d-1\,*}(\overline{\Gamma}^{d+1}(X,b)))=\overline{\Gamma}^{d+1}(X,b),$$
which proves the claim.
The next step is to observe that the Griffiths group
of $1$-cycles homologous to $0$ modulo algebraic equivalence is a
birational invariant. This is elementary to show using resolution of
indeterminacies of birational maps, as it is invariant under blow-up
and is functorial under pushforward and pullbacks under generically
finite morphisms. As $X^{(g)}$ is birational to $J(X)$ via the Abel
map, we conclude that $\overline{\Gamma}^{d+1}(X,b)=0$ in ${\rm
CH}_1(X^{(d+1)})/{\rm alg}$ if and only if its image $W$ in $J(X)$
under the Abel map vanishes in ${\rm CH}_1(J(X))/{\rm alg}$.
Finally, we observe that a cycle appearing in the formula
(\ref{eqformgammam}) for $\Gamma^{d+1}(X,b)$, which is up to
permutation of the form
$$\{(x,\ldots,x,b,\ldots,b),\,x\in X\},$$
where $x$ appears $k$ times and $b$ appears $d+1-k$ times, maps
under the Abel map to a $1$-cycle of $J(X)$ algebraically equivalent
to $\mu_{k*}(Z)$. The vanishing of $W$ in ${\rm CH}_1(J(X))/{\rm
alg}$ thus gives
\begin{eqnarray}
\label{eqnvaniW1}\sum_{k=1}^{d+1}(-1)^{d+1-k}\binom{d+1}{k}\mu_{k*}Z=0\,\,{\rm
in }\,\,{\rm CH}_1(J(X))/{\rm alg}.
\end{eqnarray}
Writing the Beauville decomposition
$$Z=\sum_s Z_s,$$
the vanishing of $W$ in ${\rm CH}_1(J(X))/{\rm alg}$ is equivalent
to
\begin{eqnarray}
\label{eqnvaniW2}
\sum_{k=1}^{d+1}(-1)^{d+1-k}\binom{d+1}{k}k^{2+s}Z_s=0,\,{\rm
in}\,\,{\rm CH}_1(J(X))/{\rm alg}
\end{eqnarray}
for any $s$.
We now have the following easy lemma:
\begin{lemm} We have
$\sum_{k=1}^{d+1}(-1)^{d+1-k}\binom{d+1}{k}k^{2+s}=0$ for $s\leq
d-2$, and
$$\sum_{k=1}^{d+1}(-1)^{d+1-k}\binom{d+1}{k}k^{2+s}\not=0$$ for
$s\geq d-1$.
\end{lemm}
This shows that the vanishing (\ref{eqnvaniW2}) is equivalent to the
vanishing of $Z_s$ for $s\geq d-1$.
\end{proof}
\begin{rema}{\rm Proposition \ref{propsansnom12mai} is also proved
in \cite{moonenyin}, where it is used to deduce the vanishing
$\Gamma^{g+2}(X,a)=0$ of Remark \ref{remaref}, for any point $a\in
X$, from the main result of Colombo and van Geemen
\cite{colombovgeemen}.}
\end{rema}
\section{Hyper-K\"ahler manifolds\label{sec3}}
\subsection{Proof of Theorem \ref{theoHKintro}}
We prove in this section the following theorem (cf. Theorem
\ref{theoHKintro} of the introduction):
\begin{theo} \label{theoHK} Let $S$ be a $K3$ surface, and let $X=S^{[n]}$. Then
\begin{eqnarray}
\Gamma^{2n+1}(X,o_X)=0 \,\,{\rm in \,\, CH}_{2n}(X^{2n+1}),
\label{eqformk3prouvefin}
\end{eqnarray}
where $o_X$ is the canonical $0$-cycle on $X$ constructed from the canonical $0$-cycle of
$S$.
\end{theo}
Here the cycle $o_S$ appears in the
following theorem from \cite{beauvoi} providing a list of relations
which hold in the Chow ring of a self-product
of a $K3$ surface.
\begin{theo}\label{theorappelrelbeauvoi} Let $S$ be a smooth projective $K3$ surface. Then
there is a degree $1$ zero-cycle $o_S\in {\rm CH}_0(S)$ satisfying
the following equalities (which are all polynomial relations in
${\rm CH}(S^k)$ for adequate $k$, between the cycles
$p_i^*o_S,\,p_j^*L,\,p_{st}^*\Delta_S$):
\begin{enumerate}
\item $L^2-{\rm deg}\,(L^2)\,o_S=0\,\,{\rm in}\,\,{\rm CH}_0(S),\,{\rm for\,\,any\,\,}L\in Pic\,S$.
\item $\Delta_S.p_1^*L-L\times o_S-o_S\times L=0$ in ${\rm CH}_1(S\times S)$
for any $L\in {\rm Pic}\,S$,
where $p_1$ is the first projection from $S\times S$ to $S$, and
$L\times o_S=p_1^*L\cdot p_2^*o_S$.
\item \label{5} $\Gamma^3(S,o_S)=0$ in ${\rm
CH}_2(S\times S\times S)$. (Using formula (\ref{eqformgammam}) and
the identity
$\Delta_3=p_{12}^*\Delta_S\cdot p_{13}^*\Delta_S$,
we can also see Property \ref{5} as a polynomial relation in ${\rm
CH}(S^3)$ involving the classes $p_{ij}^*\Delta_S$ and
$p_k^*(o_S)$.)
\item \label{itemsansnom} $\Delta_S^2=24\,p_1^*o_S\cdot p_2^*o_S$ in ${\rm CH}_0(S\times S)$.
\item \label{item3} $\Delta_S.p_1^*o_S-p_1^*o_S\cdot p_2^*o_S=0$ in ${\rm CH}_0(S\times S)$.
\end{enumerate}
\end{theo}
Note that property \ref{item3} is (\ref{eqimportant}) and is easily
satisfied because $o_S$ is the class of a point in $S$. Property
\ref{itemsansnom} is a consequence of Property \ref{5} which implies
$c_2(S)=24 \, o_S$ in ${\rm CH}_0(S)$, and Property \ref{item3}.
\begin{rema}\label{rematrivialrel} {\rm The above relations are the nontrivial relations
involving $p_i^*(o_S),\,p_j^*L,\,L\in{\rm CH}^1(S) $ and the $p_{kl}^*\Delta_S$
and with the property that in at least one monomial, an index is repeated. To make a complete
list of such relations, one should add the ``trivial relations'', which hold on any surface, namely
\begin{enumerate}
\item $o_X\cdot L=0$ in ${\rm CH}(S)$, $L\in {\rm CH}^1(S)$,
\item $o_X\cdot o_X=0$ in ${\rm CH}(S)$,
\item $p_{12}^*\Delta_S\cdot p_{23}^*\Delta_S=p_{13}^*\Delta_S\cdot p_{23}^*\Delta_S$ in ${\rm CH}(S\times S\times S)$.
\end{enumerate}}
\end{rema}
As in
\cite{voisinpamq}, the ingredients of the proof of Theorem \ref{theoHK} are
1) the results of de Cataldo-Migliorini
\cite{decami}, which will allow, thanks to Proposition
\ref{proppourtheofin}, to translate the problem into
computations
in ordinary
self-products $S^N,\, N\leq (2n+1)n$, of a $K3$ surface; 2) the
relations listed in Theorem \ref{theorappelrelbeauvoi}; 3) the
recent result of Yin \cite{yin}. The latter says basically that for
a regular surface $S$, the {\it cohomological} polynomial relations on
$S^N$ between the diagonal classes and the pull-back under the
various projections of the class of a point are generated by the
relations listed above (or rather, their cohomological counterpart)
and the Kimura relation (cf. \cite{kimura},
\cite[3.2.3]{voisinbook}) which holds when the motive of $S$ is
finite dimensional. A key point of the proof will be thus
the fact that the Kimura relation is not needed to express the pull-back to
$S^N$ of the vanishing relation
$[\Gamma^{2n+1}(X,o_X)]=0$.
We first recall some notation related to $S^n$ and $S^{[n]}$, for
any smooth surface $S$. Let $\mu=\{A_1,\ldots,A_l\},\,l=:l(\mu)$ be
a partition of $\{1,\ldots,n\}$, where all the $A_i$'s are nonempty.
Let $S^\mu\cong S^{l(\mu)}\subset S^n$ be the set $$\{(s_1,\ldots,
s_n),\,s_i=s_j\,\,{\rm if}\,\,i,\,j\in A_k\,\,{\rm
for\,\,some}\,\,k\}.$$ The image $\overline{S^{(\mu)}}$ of $S^\mu$
in $S^{(n)}$ is a stratum of $S^{(n)}$. It is not normal in general, but its
normalization $S^{(\mu)}$ is the quotient of $S^\mu$ by the subgroup
$\mathfrak{S}_\mu$ of $\mathfrak{S}_n$ preserving $S^\mu$, that is
acting on $\{1,\ldots,n\}$ by permuting the $A_i$'s with the same
cardinality.
Let $c: S^{[n]}\rightarrow S^{(n)}$ be
the Hilbert-Chow morphism and let $E_\mu:=S^\mu\times_{S^{(n)}}
S^{[n]}\subset S^\mu\times S^{[n]}$. It is known that $E_\mu$ is
irreducible of dimension $n+l(\mu)$. We see $E_\mu$ as a
correspondence between $S^\mu$ and $S^{[n]}$.
\begin{theo} (de Cataldo-Migliorini \cite{decami}) \label{dedecami} The collection $(E_\mu)_\mu$ of
correspondences identifies the motive of $S^{[n]}$ to a submotive of
the disjoint union $\sqcup_\mu S^\mu$. More precisely, for some
combinatorial coefficients $\lambda_\mu$,
$$\Delta_X=\sum_\mu\lambda_\mu(E_{\mu},E_\mu)_*(\Delta_{S^\mu})\,\,{\rm
in}\,\,{\rm CH}_{2n}(X\times X).$$
\end{theo}
The result above implies in particular: \begin{coro}\label{coroinj}
Let $X:=S^{[n]}$. For any integer $k$, the map
$$\oplus_{(\mu_1,\ldots,\mu_k)}(E_{\mu_1},\ldots,E_{\mu_k})^*:{\rm CH}^*(X^k)\rightarrow
\oplus_{(\mu_1,\ldots,\mu_k)}{\rm CH}^*(S^{\mu_1}\times\ldots\times
S^{\mu_k})$$ is injective.
\end{coro}
We now have the following result: Let $n$ and $k$ be fixed.
Let us denote
by $\Delta_k\subset X^k$ the small diagonal of $X^k$, where
$X:=S^{[n]}$, for
a smooth projective surface $S$.
\begin{prop} \label{proppourtheofin} For any $k$-uple $(\mu_1,\ldots,\mu_k)$ of partitions of
$\{1,\ldots,n\}$, there exists a universal (i.e. independent of $S$) polynomial $P_{\mu_\cdot}$ (in many
variables) with the following property: For any smooth quasi-projective surface
$S$,
$$(E_{\mu_1},\dots,E_{\mu_k})^*(\Delta_k)=
P_{\mu_\cdot}(pr_i^*c_2(S),\,pr_j^*(K_S),pr_{st}^*(\Delta_S))\,\,{\rm
in}\,\, {\rm CH}(S^{\mu_1}\times\ldots\times S^{\mu_k}) ,$$ where
the $pr_i$'s are the projections from $\prod_iS^{\mu_i}\cong S^N$ to
its factors (isomorphic to $S$), and the $pr_{st}$'s are the
projections from $\prod_iS^{\mu_i}$ to the products of two of its
factors (isomorphic to $S\times S$).
\end{prop}
\begin{proof}
Proposition \ref{proppourtheofin} is a particular case of Theorem
\ref{theopourtheofin} whose proof will be sketched in Subsection
\ref{secuniv} and will be completed in \cite{voisinuniversalcycles}, because
the cycles $(E_{\mu_1},\dots,E_{\mu_k})^*(\Delta_k) \in {\rm CH}(S^N)
$ are clearly universally defined cycles in the sense of Definition \ref{defiunivdef}. Indeed, for any family $\mathcal{S}\rightarrow B$ of smooth quasi-projective surfaces,
we can construct the smooth family of relative Hilbert schemes
$\mathcal{X}:=\mathcal{S}^{[n/B]}$ and its relative small diagonals
$$\Delta_{k/B}(\mathcal{X})\subset \mathcal{X}^{k/B}.$$
Then we have the relative correspondences $E_{\mu_i}\subset \mathcal{S}^{\mu/B}\times_B\mathcal{X}$,
which are proper over the first summand,
and we have thus the relative cycle
$$E_{\mu_\cdot}^*(\Delta_{k/B}(\mathcal{X}))\in{\rm CH}(\mathcal{S}^{[N/B]}),\,\,N=l(\mu_1)+\ldots+l(\mu_k),$$
satisfying the functoriality properties stated in Definition
\ref{defiunivdef}, because the morphisms $E_{\mu_i}\rightarrow B$ are flat.
\end{proof}
\begin{rema} {\rm One may have the feeling that the canonical class
is not necessary in Proposition \ref{proppourtheofin}, as set theoretically one wants the set of
$(s_1,\ldots,s_k)\in S^{\mu_1}\times \ldots\times S^{\mu_k}$ such
that there is a subscheme $x\in S^{[n]}$ whose associated cycle is
$s_i$ (or rather its image in $S^{(n)}$) and this does not seem to
involve the intrinsic geometry of $S$, except for the
self-intersection of the diagonal, thus only $c_2$. In fact, due to
excess formulas, the canonical class actually appears, as the
simplest example shows: Let $X$ be $S^{[2]}$, $k=3$, and
$\mu_1=\mu_2=\mu_3$ be the partition of $\{1,2\}$ consisting of a
single set with $2$ elements. Then $E_{\mu_1}=E_{\mu_2}=E_{\mu_3}=E$
is the exceptional divisor of $S^{[2]}$ and we have
$$(E_{\mu_1},E_{\mu_2},
E_{\mu_3})^*(\Delta_3)=\Delta_*(p_*(E^2_{\mid E})),$$ where
$\Delta:S\rightarrow S^3$ is the diagonal inclusion, and
$p:E\rightarrow S$ is the natural map. But $p_*(E^2_{\mid E})\in
{\rm CH}^1(S)$ is a nonzero multiple of the canonical class of $S$.}
\end{rema}
\begin{rema}{\rm We proved in \cite{voisinpamq}
similar statement where instead of the small
diagonal, arbitrary polynomials in the Chern classes of the
tautological sheaf on $X^{[n]}$ and the Chern classes of the ideal
sheaf of the incidence correspondence in $S^{[n-1]}\times S$) are
considered; the same kind of arguments used in {\it loc. cit.}, which are in fact borrowed from
\cite{EGL}, can be applied to prove Proposition \ref{proppourtheofin}, but the proofs
are very intricated and lengthy and in fact
all these results can also be obtained as Proposition \ref{proppourtheofin},
as a consequence
of Theorem
\ref{theopourtheofin}.
}
\end{rema}
We now show how Theorem \ref{theoHK} follows from Proposition
\ref{proppourtheofin}.
\begin{proof}[Proof of Theorem \ref{theoHK}] We have to prove the vanishing of
$\Gamma^{2n+1}(X,o_X)$, where
$S$ is a smooth projective $K3$
surface and $X=S^{[n]}$. By Corollary \ref{coroinj}, it suffices to show that for
any $2n+1$-uple $(\mu_1,\ldots,\mu_{2n+1})$ of partitions of
$\{1,\ldots,n\}$, we have
\begin{eqnarray}\label{vanemustar}
(E_{\mu_1},\dots,E_{\mu_{2n+1}})^*(\Gamma^{2n+1}(X,o_X))=0\end{eqnarray}
in ${\rm CH}(S^{\mu_1}\times\ldots\times S^{\mu_{2n+1}})$. As
$\Gamma^{2n+1}(X,o_X)$ is a combination of cycles which up to
permutation of factors are of the form $\Delta_k\times o_X^{2n+1-k}$
and $E_\mu^*o_X=0$ if $\mu\not=\{\{1\},\ldots,\{n\}\}$, and is equal
to $n!(o_S,\ldots,o_S)$ if $\mu=\{\{1\},\ldots,\{n\}\}$, it follows
from Proposition \ref{proppourtheofin} that there exists a
polynomial $Q_{\mu_\cdot}$ (in many variables) with the following
property: For any smooth projective surface $S$, and any point
$o_S\in S$,
\begin{eqnarray}
\label{eqEmustarQ}(E_{\mu_1},\dots,E_{\mu_{2n+1}})^*(\Gamma^{2n+1}(X,o_X))=
Q_{\mu_\cdot}(pr_i^*c_2(S),\,pr_j^*(K_S),\,pr_l^*o_S,\,pr_{st}^*(\Delta_S))
\end{eqnarray}
in ${\rm CH}(S^{\mu_1}\times \ldots\times S^{\mu_{2n+1}})$.
We know by \cite[Proposition 1.3]{ogrady} (see also Theorem \ref{theoog}, (i))
that for any regular surface
$S$, and any point $o_S\in S$,
$\Gamma^{2n+1}(X,o_X)$ is cohomologous to $0$, where $o_X$ is any
point of $X=S^{[n]}$ over $no_S\in S^{(n)}$. It follows that for
each $2n+1$-uple $(\mu_1,\ldots,\mu_{2n+1})$ as above, the cycle
$$(E_{\mu_1},\dots,E_{\mu_{2n+1}})^*(\Gamma^{2n+1}(X,o_X))$$ is
cohomologous to $0$ in $S^{\mu_1}\times\ldots\times S^{\mu_{2n+1}}$.
Hence the polynomial $Q_{\mu_\cdot}$ has the property that for a
regular surface $S$,
\begin{eqnarray}\label{eqvancohpol}
Q_{\mu_\cdot}(pr_i^*[c_2(S)],\,pr_j^*([K_S]),\,pr_l^*[o_S],\,pr_{st}^*([\Delta_S]))=0\,\,{\rm
in}\,\, H^*(S^{\mu_1}\times\ldots\times S^{\mu_{2n+1}},\mathbb{Q}).
\end{eqnarray}
Here the brackets denote the cohomology class of the corresponding
cycles. In this equation, we can of course replace $[c_2(S)]$
by $\chi_{top}(S)[o_S]$, with $\chi_{top}(S)$
determined by the polynomial relation (this is relation \ref{itemsansnom} in Theorem \ref{theorappelrelbeauvoi})
$[\Delta_S]^2=\chi_{top}(S)pr_1^*[o_S]\cup pr_2^*[o_S]$ in $H^4(S\times S,\mathbb{Q})$. We now follow \cite{voisinpamq} (see also \cite{yin}): The
cohomological version of the equations given in Theorem
\ref{theorappelrelbeauvoi} with $L=K_S$ holds on any smooth
projective surface with $b_1=0$, and if the canonical class
satisfies $[K_S]=0$ or $[K_S]^2\not=0$, one can reduce modulo
these relations any polynomial expression in the variables
$$pr_i^*[pt],\,pr_j^*[K_S],\,pr_{st}^*[\Delta_S]$$
to a linear combination of monomials in the variables
$pr_i^*[pt],\,pr_j^*[K_S],\,pr_{st}^*[\Delta_S]^0$,
with the property that no index appears twice in the monomial. Here
the class $[\Delta_S]^0$ is the class
$$[\Delta_S]-pr_1^*[pt]-pr_2^*[pt]-\lambda pr_1^*[K_S]\cup pr_2^*[K_S],$$
where the coefficient $\lambda$, when $K_S\not=0$, is determined by
the relation $\lambda [K_S]^2=1$ (the class $[\Delta_S]^0\in
H^4(S\times S,\mathbb{Q})$) is the projector onto
$H^2(S,\mathbb{Q})^{{\perp}[K_S]}$).
Now, it is clear by K\"unneth decomposition that if
a linear combination of such monomials vanishes in $H^*(S^N,\mathbb{Q})=H^*(S,\mathbb{Q})^{\otimes N}$, then for fixed distinct indices $i_1,\ldots,i_m,\,j_1,\ldots,j_p,\,k_1,\ldots,k_q$, the
sum of such monomials of the form
$$pr_{i_1}^*[pt]\cdot\ldots \cdot pr_{i_m}^* [pt]\cdot pr_{j_1}^*[K_S]\cdot\ldots\cdot
pr_{j_p}^*[K_S]\cdot pr_{k_1}^*1_S\cdot\ldots\cdot pr_{k_q}^*1_S\cdot \prod_{s_1,t_1,\ldots,s_l,t_l}pr_{s_i,t_i}^*[\Delta_S]^0$$
where the indices $s_i,\,t_j$ exhaust the remaining indices, are all distinct and
are different from
the $i_s,\,j_s,\,k_s$, has to be $0$.
This way, we reduced the problem to linear combinations of monomials of the form
\begin{eqnarray}
\label{eqmonomial}
pr_{s_1t_1}^*[\Delta_S]^0\cdot\ldots\cdot
pr_{s_lt_l}^*[\Delta_S]^0 \end{eqnarray}
on $S^{2l}$,
where no index is repeated.
We now have the following result due to Yin \cite{yin}: The ``Kimura
relation'' is a relation between monomials of the above type. It
says that, for $M={\rm dim}\,H^2(S,\mathbb{Q})^{{\perp}[K_S]}$, the
cohomology class of the projector onto
$\bigwedge^{M+1}H^2(S,\mathbb{Q})^{{\perp}[K_S]}\subset
H^{2M+2}(S^{M+1},\mathbb{Q})$ is $0$, which is obvious since
$\bigwedge^{M+1}H^2(S,\mathbb{Q})^{{\perp}[K_S]}=0$.
The
class of this projector is the class \begin{eqnarray}
\label{eqkimura} \sum_{\sigma\in
\mathfrak{S}_{M+1}}\epsilon(\sigma)\prod_{i=1}^{M+1}pr_{i,M+1+\sigma(i)}^*[\Delta_S]^0\in
H^{4M+4}(S^{2M+2},\mathbb{Q}).
\end{eqnarray}
and the Kimura relation is thus the vanishing of (\ref{eqkimura}).
\begin{theo} (Yin, \cite{yin})\label{theoyin} For any integer $m$,
the relations in $H^*(S^m,\mathbb{Q})$ between the monomials
(\ref{eqmonomial}) with no repeated indices are generated by the
pull-back to $S^m$ of the Kimura relation via a projection (and a
permutation) $S^m\rightarrow S^{2N+2}$.
\end{theo}
We deduce the following
\begin{coro}\label{corodelafinfin} The polynomial $Q_{\mu_\cdot}$ belongs to the ideal
generated by the trivial relations (see Remark \ref{rematrivialrel}), the relation $c_2(S)=\chi_{top}(S) o_S$ (where we
recover $\chi_{top}(S)$ as the self-intersection of $\Delta_S$) and
the relations listed in Theorem \ref{theorappelrelbeauvoi} with
$L=K_S$.
\end{coro}
\begin{proof} Indeed, choose for $S$ a smooth projective surface
with $b_1(S)=0$ and $b_2(S)> \frac{n(2n+1)}{2}$. Then by Theorem
\ref{theoyin}, there are no linear relations between the monomials
(\ref{eqmonomial}) with no repeated index if $s\leq (2n+1)n$. On
the other hand, we have the vanishing of the cohomology class $$
Q_{\mu_\cdot}(pr_i^*[c_2(S)],\,pr_j^*[K_S],\,pr_l^*[o_S],\,pr_{st}^*[\Delta_S])
\in H^*(S^N,\mathbb{Q}),$$ where $N=\sum_il(\mu_i)\leq (2n+1)n$.
It
then follows from the above reduction that the polynomial
$Q_{\mu_\cdot}$, where one substitutes $\chi_{top}(S)[o_S]$ to
$[c_2(S)]$, belongs to the ideal generated by the cohomological
version of the relations given in Theorem
\ref{theorappelrelbeauvoi}, with $L=K_S$.
\end{proof}
The proof of Theorem \ref{theoHK} is now finished. Indeed, $S$ being
now a $K3$ surface, we know by Theorem \ref{theorappelrelbeauvoi}
that the relation $\chi_{top}(S)o_S=c_2(S)$ holds in $CH_0(S)$ and
that the relations listed in Theorem \ref{theorappelrelbeauvoi} hold
in ${\rm CH}(S^k)$ for adequate $k$. As the polynomial
$Q_{\mu_\cdot}$, where one substitutes $\chi_{top}(S)o_S$ to
$c_2(S)$, belongs to the ideal generated by the relations given in
Theorem \ref{theorappelrelbeauvoi} and the trivial relations, we conclude that
$Q_{\mu_\cdot}=0$ in ${\rm CH}(S^N)$. By (\ref{eqEmustarQ}), we proved the vanishing
(\ref{vanemustar})
$$(E_{\mu_1},\dots,E_{\mu_{2n+1}})^*(\Gamma^{2n+1}(X,o_X))=0\,\,{\rm
in}\,\,{\rm CH}(S^N),$$ which
concludes the proof.
\end{proof}
\subsection{Universally defined cycles \label{secuniv} }
This subsection is devoted to introducing the notion of ``universally defined
cycles'' and to sketching the proof of a quite general statement which will be fully proved in \cite{voisinuniversalcycles}. It concerns ``universally defined" cycles on self-products of surfaces.
We first explain the meaning of this expression. In the following, we work with Chow groups
with integral coefficients, and we will write
${\rm CH}(X)_\mathbb{Q}$ for cycles with $\mathbb{Q}$-coefficients.
\begin{Defi}\label{defiunivdef} Let $n,\,N$ be integers. A universally defined cycle on the $N$-th power of
smooth complex algebraic varieties $X$ of a given dimension $n$ consists in the following data: for
each smooth family of $n$-dimensional algebraic varieties $\mathcal{X}\rightarrow B$, where
$B$ is smooth quasiprojective, a cycle $z_\mathcal{X}\in {\rm CH}(\mathcal{X}^{N/B})$ is given, satisfying
the following conditions:
(i)
If $r:B'\rightarrow B$ is a morphism, with induced morphism
$$R_N:(\mathcal{X}')^{N/B'}\rightarrow
\mathcal{X}^{N/B},\,\mathcal{X}':=\mathcal{X}\times_BB',$$ then
$$ z_{\mathcal{X}'}=R_N^*z_{\mathcal{X}}\,\,{\rm in}\,\,{\rm CH}((\mathcal{X}')^{N/B'}).$$
(ii) If $\mathcal{X}\rightarrow B$ is a family as above and
$\mathcal{Y}\subset \mathcal{X}$ is a Zariski open set, then
$$z_\mathcal{Y}=z_{\mathcal{X}\mid \mathcal{Y}^{N/B}}\,\,{\rm in}\,\,{\rm CH}(\mathcal{Y}^{N/B}).$$
\end{Defi}
\begin{theo}\label{theopourtheofin} For any universally defined cycle $z$ on $N$-th powers of surfaces,
there exists a polynomial $P$ with rational coefficients, depending only on $z$, such that
for any smooth algebraic surface
$S$ defined over $\mathbb{C}$, $$z_S=P(pr_i^*c_1(S),pr_j^*c_2(S),pr_{rs}^*\Delta_S)\,\,{\rm in}\,\,{\rm CH}(S^N)_\mathbb{Q}.$$
\end{theo}
\begin{rema}{\rm One could introduce as well
universally defined cycles with $\mathbb{Q}$-coefficients, by replacing everywhere in the definition
above ${\rm CH}$ by ${\rm CH}_\mathbb{Q}$.
It is possible that the conclusion holds as well for universally defined cycles
with $\mathbb{Q}$-coefficients, but our present proof uses the integral structure.
}
\end{rema}
We will give some hints on the proof, with a complete proof only in the case $N=1$
(Proposition \ref{projusteS}) and the construction of the desired polynomials
(Corollary \ref{corpourthefinsd} and Proposition \ref{coroonnesaitjamais}). We refer
to \cite{voisinuniversalcycles} for a full treatment.
Let us
first show how to produce such polynomials.
Let $G:=G(2,5)$ be the Grassmannian of $2$-dimensional vector subspaces
in $\mathbb{C}^5$. Any smooth complex projective surface can be embedded in $G$,
for example by choosing $5$ general sections of a very ample vector bundle $E$ on $S$.
Let $\mathcal{O}_G(1)$ be the Pl\"ucker line bundle on $G$, and let $c\in {\rm CH}^2(G)$ be the second
Chern class of the tautological rank $2$ vector bundle on $G$.
We choose an integer $d$, and consider the universal family
$\mathcal{S}_d\rightarrow B$ of smooth surfaces in $G$ which are
complete intersections of $4$ members of $|\mathcal{O}_G(d)|$.
The smooth variety $B$ is thus
the vector space $H^0(G,\mathcal{O}_G(d))^4$ and
$$\mathcal{S}_d\subset \mathcal{S}_{d,univ}$$
is the Zariski open set consisting of points where
$\mathcal{S}_{d,univ}\rightarrow B$ is smooth.
Here
$$\mathcal{S}_{d,univ}:=\{(b,x)\in B\times G,\,b=(f_{1,b},\ldots,f_{4,b}),\,f_{i,b}(x)=0\,\,\forall i\}.$$
There is an obvious morphism
$$f:\mathcal{S}_d\rightarrow G$$
given by the restriction to $\mathcal{S}_d$ of the second projection
$\mathcal{S}_{d,univ}\rightarrow G$,
which induces for any $N\geq 1$ the morphism
$$f_N: \mathcal{S}_d^{N/U}\rightarrow G^N$$
with induced pull-back morphism
$f_N^*:{\rm CH}(G^N)\rightarrow
{\rm CH}(\mathcal{S}_d^{N/U})$.
We now use the following result, which is one of the main ingredients in the proof of Theorem
\ref{theopourtheofin}:
\begin{prop} \label{leunivsurfd} For any integer $N>0$ and
sufficiently large $d$,
${\rm CH}(\mathcal{S}_d^{N/U})$ is generated as a ${\rm CH}(G^N)$-module by
the relative partial diagonals
$\Delta_{I/U}(\mathcal{S}_d)$.
\end{prop}
Here $I$ denotes as usual a partition of $ \{1,\ldots,N\}$, determining a partial diagonal.
\begin{proof}[Proof of Proposition \ref{leunivsurfd}] By the localization exact sequence, it suffices
to prove the result with
$\mathcal{S}_d$ replaced by $\mathcal{S}_{d,univ}$.
Next consider the natural morphism
$$
f_N:\mathcal{S}_{d,univ}^{N/B}\rightarrow G^d.$$
The fiber of $f_N$ over a $N$-uple $(x_1,\ldots x_N)$ consists of the set of
$4$-uples $(\sigma_1,\ldots,\sigma_4)\in H^0(G,\mathcal{O}_G(d))^4$ having the property
that the $\sigma_i$'s vanish at all points $x_i$.
As $d $ is large compared to $N$, any $k$ distinct points of $G$ with $k\leq N$ impose independent
conditions to $H^0(G,\mathcal{O}_G(d))$, and thus, denoting by $G^N_k$ the locally closed
subvariety of $G^N$ consisting of $N$-uples with exactly $k$-distinct points,
which is the disjoint union of the diagonals $\Delta_I(G)$ with $l(I)=k$ (or rather of
the $\Delta_I^0(G):=\Delta_I(G)\setminus \cup_{J,l(J)<k}\Delta_J(G)$), we find that
$f_N^{-1}(G^N_k)$ is a Zariski open set in a vector bundle over $G^N_k$. It follows from the
localization exact sequence and $A^1$-invariance that ${\rm CH}( G^N_k)\stackrel{f_N^*}
{\rightarrow }
{\rm CH}(f_N^{-1}(G^N_k))$ is surjective.
Writing $G^N$ as the disjoint union of the $\Delta_I^0(G)$,
we conclude from the above and the localization exact sequence that
$$\oplus_I{\rm CH}(\Delta_I
(G))\stackrel{(j_{I*}\circ f_I^{*})}
{\rightarrow}
{\rm CH}(\mathcal{S}_{d,univ}^{N/B})$$
is surjective, where $f_I$ is the restriction of
$f_N$ to $f_N^{-1}(\Delta_I(G))\subset G^N$ and $j_I$ is the inclusion
of $f_N^{-1}(\Delta_I(G))$ in $\mathcal{S}_d^{N/U}$. Note that
$f_I^{-1}(\Delta_I(G))=\Delta_{I/B}(\mathcal{S}_{d,univ})$.
Finally, we observe that
the restriction map
$${\rm CH}(G^N)\rightarrow {\rm CH}(\Delta_I(G))$$
is surjective, and that for any $\alpha\in {\rm CH}(G^N)$,
$$j_{I*}
\circ f_I^*(\alpha_{\mid \Delta_I(G)})=f_N^*\alpha\cdot (j_{I*}
\circ f_I^*)(1)=f_N^*\alpha\cdot
\Delta_{I/B}(\mathcal{S}_{d,univ}),$$
and this finishes the proof.
\end{proof}
\begin{coro} \label{corpourthefinsd} For any universally defined cycle $z$ on $N$-th powers of surfaces and for sufficiently large $d$,
there exists a polynomial $P_d$ with rational coefficients, depending only on $z$ and $d$ such that
for any smooth complete intersection surface
$S_d\subset G$ as above,
\begin{eqnarray}
\label{eqpourPd}
z_{S_d}=P_d(pr_i^*c_1(S_d),pr_j^*c_2(S_d),pr_{rs}^*\Delta_{S_d})\,\,{\rm in}\,\,{\rm CH}(S_d^N)_\mathbb{Q}.
\end{eqnarray}
Furthermore, $(4d-5)^{2N}P_d$ has integral coefficients.
\end{coro}
\begin{proof} As $z$ is universal, there exists a cycle $z_{\mathcal{S}_d}\in {\rm CH}(\mathcal{S}_d^{N/B})$ such that for any surface $S_d$ as above,
$$z_{S_d}=(z_{\mathcal{S}_d})_{\mid S_d^N},$$
where we see $S_d$ as a fiber of the universal family
$\mathcal{S}_d\rightarrow B$.
We next use Proposition \ref{leunivsurfd} to write, for $d>>0$, $\mathcal{Z}_d$ as a combination
$\sum_I f_N^*\alpha_I\cdot \Delta_{I/B}(\mathcal{S}_d)$, where $\alpha_I\in {\rm CH}(G(2,5))$.
Furthermore, is it immediate to prove that ${\rm CH}(G^N)={\rm CH}(G)^{\otimes N}$, so that
we can write each $\alpha_I$ as a polynomial with integral coefficients in
$pr_i^*m,\,m:=c_1(\mathcal{O}_G(1))=c_1(E)$ and $pr_j^*c,\,c:=c_2(E)$ where $E$ is the
dual of the tautological subbundle on $G$. Of course, under restriction to $S_d^N$, only polynomials of weighted degree $\leq 2$ in each set of
variables $pr_i^*m,\,pr_i^*c$ will survive.
We now observe that the restriction of $m$ to $S_d$ is a rational multiple of $c_1(S)$ (more precisely,
$K_{S_d}=
\mathcal{O}_G(4d-5)_{\mid S}$ by the adjunction formula), and the restriction of $c$ to $S_d$ is an adequate
linear combination of $\frac{1}{(4d-5)^2}c_1(S_d)^2,\,c_2(S_d)$.
Putting everything together and using the fact
that the relative diagonals $\Delta_{I/B}(\mathcal{S}_d)$ restrict to
$\Delta_I(S_d)$, we get the result.
\end{proof}
\begin{rema}\label{remaaddendum}{\rm Note that Corollary \ref{corpourthefinsd} is true
more generally
for the regular and complete intersection
locus $S_{reg}$ of any set of $4$ degree $d$ equations on $G$.
The proof uses Proposition \ref{leunivsurfd} (which works for the family
$\mathcal{S}_d\rightarrow B$ of smooth complete intersection quasi-projective surfaces), and
both conditions (i) and (ii) in Definition \ref{defiunivdef}.}
\end{rema}
The corollary above proves Theorem \ref{theopourtheofin} for smooth complete intersection surfaces
of degree $d$, and more generally
for the regular and complete intersection
locus of any set of $4$ degree $d$ equations on $G$. What remains to be done is to prove that the polynomial above works for all
surfaces.
Note that the polynomial $P_d$ is in fact not uniquely defined
as only its value on the set of variables
$pr_i^*c_1(S_d),\,pr_j^*c_2(S_d),\,pr_{st}^*\Delta_{S_d}$ is well defined in
${\rm CH}(S_d^N)_\mathbb{Q}$. Hence {\it a priori} $P_d$ is only defined modulo the relations
in ${\rm CH}(S_d^N)_\mathbb{Q}$
between these variables. However, the following result shows that
a part of $P_d$ is in fact independent of $d$ for large $d$.
\begin{prop}\label{coroonnesaitjamais} For any universally
defined cycle $z$ on $N$-th powers of surfaces, there exists a polynomial $Q$
in the variables $pr_{st}^*\Delta_S$, depending only on $z$, with the following property:
For any smooth surface $S$, there is a Zariski dense open set
$V\subset S$ such that
$z_V=Q(pr_{st}^*\Delta_S)$ in ${\rm CH}(V^N)_\mathbb{Q}$.
\end{prop}
\begin{proof} Let $Q_d$ be the part of the polynomial $P_d$ which involves only the diagonals.
Then let $U_d\subset S_d$ be the complement of a hyperplane section
defined by the choice of a codimension $2$ vector subspace $W\subset \mathbb{C}^5$ in general position. As
$c_1(\mathcal{O}_{S_d}(1))$ and $c_2(E)$ vanish in ${\rm CH}(U_d)$, we deduce
from (\ref{eqpourPd}) that
\begin{eqnarray}\label{eqncyrestUd}
z_{U_d}=Q_d(pr_{st}^*\Delta_{U_d})\,\,{\rm in}\,\,{\rm CH}(U_d^N).
\end{eqnarray}
We observe now that
for $d'\leq d$, a surface $S_{d'}$ which is the complete intersection in $G$
of hypersurfaces of degree $d'$ is an irreducible component
of a (singular) surface $\Sigma_d=S_{d'}\cup T$ defined as the complete
intersection in $G$ of four degree $d$ hypersurfaces
containing $S_{d'}$ and that, denoting
$C:=S_{d'}\cap T$, the open set $U'_{d'}:=S_{d'}\setminus C$ is contained in the smooth locus
of $\Sigma_d$. From Remark \ref{remaaddendum},
we thus get that
\begin{eqnarray}\label{eqncyrestUdsurU'}
z_{U'_{d'}}=P_d(pr_i^*c_1(U'_{d'}),pr_j^*c_2(U'_{d'}),pr_{st}^*\Delta_{U'_{d'}})\,\,{\rm in}\,\,{\rm CH}({U'_{d'}}^N)_\mathbb{Q},
\end{eqnarray}
and after restriction
to $V_{d'}:=U_{d'}\cap U'_{d'}$, this becomes
\begin{eqnarray}\label{eqncyrestUdsurU'surV}
z_{V_{d'}}=Q_d(pr_{st}^*\Delta_{V_{d'}})\,\,{\rm in}\,\,{\rm CH}({V}_{d'}^N)_\mathbb{Q},
\end{eqnarray}
On the other hand, we also have (\ref{eqncyrestUd}) for $d'$, which provides after restriction
to $V_{d'}$
\begin{eqnarray}\label{eqncyrestUdsurd'}
z_{V_{d'}}=Q_{d'}(pr_{st}^*\Delta_{V_{d'}})\,\,{\rm in}\,\,{\rm CH}({V}_{d'}^N)_\mathbb{Q}.
\end{eqnarray}
Hence $Q_d-Q_{d'}$ belongs to the kernel
of the map
$${\rm ev}_{d'}:\mathbb{Q}[X_{rs}]_{1\leq r\not=s\leq N}\rightarrow {\rm CH}(V^N),$$
$$f\mapsto f(p_{rs}^*\Delta_V),$$
where $V$ is a sufficiently small Zariski open set of a general
complete intersection of four hypersurfaces of degree $d'$ in $G$.
On the other hand, it follows from the above construction
that ${\rm Ker}\,{\rm ev}_d\subset {\rm Ker}\,{\rm ev}_{d'}$ for $d'\leq d$. As the polynomials we consider are homogeneous of given degree (equal to half the codimension of $z$),
they live in a finite dimensional vector space and
we conclude that these kernels are in fact stationary, equal to $K$ for $d\geq d_0$. So we
finally conclude that there exists a $d_0$ such that
for $d\geq d'\geq d_0$,
$Q_d-Q_{d'}$ belongs to $K$. It follows that for any $d$, for any reduced
complete intersection
of four degree $d$ hypersurfaces in $G$, and for a dense Zariski open set $V\subset S$
$$z_{V}=Q_{d_0}(pr_{st}^*\Delta_{V}
)\,\,{\rm in}\,\,{\rm CH}(V^N)_\mathbb{Q}.$$
As any smooth quasi-projective surface has a dense Zariski open set which
is contained in the smooth locus of such a complete intersection for $d$ large enough,
the proposition is proved, with $Q=Q_{d_0}$.
\end{proof}
We finish this section with the proof of Theorem \ref{theopourtheofin} in the case $N=1$.
\begin{prop}\label{projusteS} Let $z$ be
a universally defined cycle on surfaces.
Then there is a polynomial $P$ independent of $S$ and with integral coefficients, such that
for any smooth quasi-projective surface $S$,
$$z_S=P(c_2(S),\,c_1(S))\in {\rm CH}(S)_\mathbb{Q}.$$
\end{prop}
\begin{proof}
Let us first treat the case of $z\in {\rm CH}^1(S)$ universally defined.
For complete intersections $S_d$ of four hypersurfaces of degree
$d$ in $G$, we must have by Corollary \ref{corpourthefinsd}
$$z=\alpha_d K_{S_d},$$
for some rational number $\alpha_d$ such that $(4d-5)\alpha_d\in\mathbb{Z}$,
and for any surface $S$, choosing a very ample vector bundle
$E$ of rank $2$ on $S$ to embed $S$ in $G$, and choosing $d$ large enough, we get
\begin{eqnarray}
\label{eqfin4nov}z_{S\mid U}=\alpha_dK_U\,\,{\rm in}\,\,{\rm CH}^1(U)_\mathbb{Q},
\end{eqnarray}
where $U=S\setminus C$, the surface
$S\cup_CT=\Sigma_d$ being the complete intersection of four degree $d$ hypersurfaces
containing $S$ in $G$.
The curve $C$ belongs to the linear system
$|(4d-5)L-K_S|$, where $L={\rm det}\,E=\mathcal{O}_G(1)_{\mid S}$. For a general choice of
equations and $d$ large enough,
the curve $C$ will be irreducible, so by the localization exact sequence,
(\ref{eqfin4nov}) rewrites as
\begin{eqnarray}
\label{eqfin4nov}z_{S}=\alpha_dK_S+\beta_d C\,\,{\rm in}\,\,{\rm CH}^1(S)_\mathbb{Q},
\end{eqnarray}
If $K_S$ and $L$ are linearly independent in ${\rm CH}^1(S)_\mathbb{Q}$, this implies, because
the left hand side is independent of $L$, that $\beta_d=0$ and thus
$z_S=\alpha_d K_S$, with $\alpha_d=:\alpha$ necessarily independent of $d$.
If not, we simply blow up $S$ at one point and choose $L$ on
$\widetilde{S}$ linearly independent of $K_{\widetilde{S}}$ in ${\rm CH}^1(\widetilde{S})_\mathbb{Q}$.
Then the above conclusion applies to $\widetilde{S}$, hence we get
\begin{eqnarray}
\label{eqfin4novtructruc}z_{\widetilde{S}}=\alpha K_{\widetilde{S}}\,\,{\rm in}\,\,{\rm CH}^1(\widetilde{S}).
\end{eqnarray}
As
$$S\setminus\{p\}\cong \widetilde{S}\setminus E_p,\,\,{\rm CH}^1(S)_\mathbb{Q}\cong
{\rm CH}^1(S\setminus\{p\})_\mathbb{Q},$$
(\ref{eqfin4novtructruc}) is also true for $S$ by condition (ii) in Definition
\ref{defiunivdef}. Finally $\alpha$ has to be an integer since $(4d-5)\alpha\in \mathbb{Z}$
for any $d$.
This proves the result for $z_S\in {\rm CH}^1(S)$ universally defined.
Let us now prove the result for a universally defined cycle $z_S\in {\rm CH}^2(S)$.
We start the proof exactly as before, and we have by Corollary \ref{corpourthefinsd} and Remark
\ref{remaaddendum} that
for $S$ the regular locus of a complete intersection of four hypersurfaces of degree $d$ in $G$,
$z_S=\mu_d c_2(S)+\nu_d c_1(S)^2$ in ${\rm CH}_0(S)_\mathbb{Q}$.
With the notation above, we conclude as before that for any pair $(S,E) $ consisting of a
smooth surface $S$ and a very ample rank $2$ vector bundle in a bounded family
(depending on $d$), we have
\begin{eqnarray}
\label{equtilepourtheofin66}
z_{S\mid U}=\mu_d c_2(S)_{\mid U}+\nu_dc_1(S)^2_{\mid U}\,\,{\rm in}\,\,{\rm CH}_0(U)_\mathbb{Q},
\end{eqnarray}
where $U=S\setminus C$ and the curve $C$ is defined as above.
By the localization exact sequence, this is equivalent
to
\begin{eqnarray}
\label{equtilepourtheofin77}
z_{S}=\mu_d c_2(S)+\nu_dc_1(S)^2+z'_{S}\,\,{\rm in}\,\,{\rm CH}_0(S)_\mathbb{Q},
\end{eqnarray}
where $z'_S$ is a $0$-cycle supported on $C$.
If $d$ is large enough (with respect to the given
bounded family), the surface $S$ is schematically
cut out by hypersurfaces of degree $d$ in $G$ and this implies
that the morphism of sheaves
$$I_S(d)\otimes\mathcal{O}_S\rightarrow N_{S/G}^*(d),\,\,N_{S/G}^*=\mathcal{I}_S/\mathcal{I}_S^2,$$
is surjective. The curve $C$ is the curve obtained as the degeneracy locus of
the morphism
$$f:\mathcal{O}_S^4\rightarrow N_{S/G}^*(d),$$
deduced from the choice of $4$ degree $d$ equations defining $S$ in $G$.
It is in fact better to see such a curve as
embedded in
$\mathbb{P}(N_{S/G}(-d))$, as the complete intersection of $4$ hypersurfaces
in the linear system $|\mathcal{O}_{\mathbb{P}(N_{S/G}(-d))
}(1)|$.
Let $B_S$ be the Zariski open set of $I_S(d)^4$ consisting of the
$(f_1,\ldots, f_4)\in I_S(d)^4$ such that
the $f_i$'s form a regular sequence cutting $S$ at its generic point.
For each element $b\in B_S$, let $C_b\subset \mathbb{P}(N_{S/G}(-d))$ be the corresponding curve.
We have
a family of complete intersection surfaces
$$\mathcal{S}'\rightarrow B_S $$
which is a subfamily of the family
$\mathcal{S}'_d\rightarrow B$ of all complete intersection surfaces of four degree $d$
hypersurfaces in $G$.
The family
$\mathcal{S}'\rightarrow B_S $ contains $B_S\times S$, which itself
contains the family of curves
$\mathcal{C}\rightarrow B_S$ with fiber $C_b$ over $b\in B_S$. The surfaces
$b\times (S\setminus C_b)$ are contained the smooth locus of
the surfaces $S'_b$.
In conclusion, the inclusion
$$\mathcal{S}'\subset \mathcal{S}'_d$$
of families of complete intersection surfaces restricts, by considering the Zariski
open sets where the surfaces are smooth, to
an open inclusion of smooth families of surfaces
$$(B_S\times S)\setminus \mathcal{C}\hookrightarrow \mathcal{S}_d\times_BB_S.$$
Applying the conditions (i) and (ii)
in Definition \ref{defiunivdef} and Proposition \ref{leunivsurfd}, we get
that
$$(pr_S^*(\mu_d c_2(S)+\nu_d c_1(S)^2))_{\mid (B_S\times S)\setminus \mathcal{C}}=(z_{B_S\times S})_{\mid (B_S\times S)\setminus \mathcal{C}}\,\,{\rm in}\,\,{\rm CH}((B_S\times S)\setminus \mathcal{C})_\mathbb{Q},$$
which rewrites equivalently, by the localization exact sequence,
as
\begin{eqnarray} \label{equtiledu24oct} (pr_S^*(\mu_d c_2(S)+\nu_d c_1(S)^2))-z_{B_S\times S}=i_*(w)\,\,{\rm in}\,\,{\rm CH}(B_S\times S)_\mathbb{Q},
\end{eqnarray}
where $w\in {\rm CH}^1(\mathcal{C})_\mathbb{Q}$ and $i:\mathcal{C}\rightarrow B_S\times S$ is the natural morphism. Here, as already mentioned, it is much better to consider the variant of
$\mathcal{C}$ which is contained in $B_S\times \mathbb{P}(N_{S/G}(-d))$, (and the
morphism $i$ is only at the general point of
$B_S$ an embedding) because it is then the universal complete intersection
of four hypersurfaces and this allows to conclude with exactly the same proof as in Proposition
\ref{leunivsurfd}:
\begin{lemm}\label{propCHCuniv} The restriction map ${\rm CH}^1(\mathbb{P}(N_{S/G}(-d)))
\rightarrow {\rm CH}^1(\mathcal{C})$ is surjective.
\end{lemm}
Note that $${\rm CH}^1(\mathbb{P}(N_{S/G}(-d)))={\rm CH}^1(S)\oplus \mathbb{Z}c_1(\mathcal{O}_{\mathbb{P}(N_{S/G}(-d))}(1)).$$
The cycle $w$ of (\ref{equtiledu24oct}) thus decomposes as
$w=(w_S)_{\mid C}+(w_P)_{\mid C}$, with
$$w_P=\alpha c_1(\mathcal{O}_{\mathbb{P}(N_{S/G}(-d))}(1))\in {\rm CH}^1(\mathbb{P}(N_{S/G}(-d)))_\mathbb{Q}.$$
As
$C_b\subset \mathbb{P}(N_{S/G}(-d))$ is the intersection of four
members of the linear system $|\mathcal{O}_{\mathbb{P}(N_{S/G}(-d))}(1)) |$ for any given $b\in B_S$, we get
\begin{eqnarray}
\label{eqnou24oct}
i_*((w_P)_{\mid C_b})=\alpha s_2(N_{S/G}(-d)),
\end{eqnarray}
where $s_2$ is the second Segre class.
Note that by (\ref{eqs2}), $s_2(N_{S/G}(-d))
$ can be explicitly computed as
a linear combination with integral coefficients of $L^2,\,c_2(E), c_2(S),\,c_1(S)^2$, involving non trivially
$c_2(E)$.
Thus we get from (\ref{equtiledu24oct}) and (\ref{eqnou24oct}), by restricting to any point
$b\in B_S$
\begin{eqnarray}\label{eqder5no} z_S=\mu_d c_2(S)+\nu_d c_1(S)^2)-\alpha s_2(N_{S/G}(-d))-((4d-5)L-K_S)\cdot w_S\,\, {\rm in}\,\,{\rm CH}_0(S)_\mathbb{Q}.
\end{eqnarray}
Let us analyze in general the cycle $w_S\in {\rm CH}^1(S)$.
We observe first that this cycle is universally defined for triples
$(S,E,d)$, where $d$ has to be large and we conclude as in the previous
proof that there are rational numbers $\gamma,\,\delta$
such that
$$w_S=\gamma c_1(S)+\delta L,$$
where $L=c_1(E)$.
Combining this with
(\ref{eqder5no}), we get, using the fact that the
curve $i(C_b)\subset S$ belongs to the linear system $|(4d-5)L-K_S|$:
\begin{eqnarray} \label{equtiledu24oct11} z_{S}=\mu_d c_2(S)+\nu_dc_1(S)^2
-\alpha s_2(N_{S/G}(-d))
\\
\nonumber
-((4d-5)L-K_S)\cdot (\gamma K_S+\delta L)\,\,{\rm in}\,\,{\rm CH}_0( S)_\mathbb{Q},
\end{eqnarray}
where the coefficients $\alpha,\,\beta,\,\gamma$ a priori depend on $S,\,E$ and $d$.
Finally we observe that the left hand side depends only on $S$, while in the right hand side,
for fixed large $d$, we can freely change $c_2(E),\,L=c_1(E)$ (staying in a bounded family).
Assume first that no nonzero linear combination of the cycles $s_2(N_{S/G}(-d)),\,L^2, \, K_S\cdot L$ belongs to
the linear span of the cycles
$z_{S},\,c_2(S),\,c_1(S)^2$. Then (\ref{equtiledu24oct11}) easily implies that
$\alpha=\delta=\gamma=0$, so that we get in this case
\begin{eqnarray} \label{equtiledu24oct222} z_{S}=\mu_d c_2(S)+\nu_dc_1(S)^2
\,\,{\rm in}\,\,{\rm CH}_0( S)_\mathbb{Q}.
\end{eqnarray}
To treat the general case, we will first prove the result assuming
${\rm CH}_0(S)$ is nontrivial, that is, if $S$ is projective
and connected, ${\rm CH}_0(S)\not=\mathbb{Z}$. It then easily follows that ${\rm CH}_0(S)$ is uncountable,
and one deduces immediately that for the blow-up $S'$ of $S$ at $6$ general points, and for adequate
choice of $E'$, (and thus $L'={\rm det}\,E'$)
no combination of the cycles $s_2(N_{S'/G}(-d)),\,L'^2, \, K_{S'}\cdot L'$ belongs to
the linear span of the cycles
$z_{S'},\,c_2(S'),\,c_1(S')^2$. Thus we conclude from (\ref{equtiledu24oct222}) that
for a general blow-up $S'$ of $S$ along $6$ points,
\begin{eqnarray} \label{equtiledu24oct333}z_{S'}=\mu_d c_2(S')+\nu_dc_1(S')^2
\,\,{\rm in}\,\,{\rm CH}_0( S')_\mathbb{Q}.
\end{eqnarray}
But this immediately implies
that for the original surface $S$, (\ref{equtiledu24oct222}) also holds, because
we can restrict (\ref{equtiledu24oct333}) to
the Zariski open set $S'\setminus \cup_{i=1}^6E_i=S\setminus \{p_1,\ldots,p_6\}$,
where the $p_i$'s are the points blown-up in $S$ and the $E_i$'s are
the exceptional curves over them.
Using property (ii) in Definition \ref{defiunivdef}, we find that
(\ref{equtiledu24oct333}) provides:
\begin{eqnarray} \label{equtiledu24oct333isb}(z_{S})_{\mid S\setminus \{p_1,\ldots,p_6\}}=\mu_d c_2(S)+\nu_dc_1(S)^2
\,\,{\rm in}\,\,{\rm CH}_0( S\setminus \{p_1,\ldots,p_6\})_\mathbb{Q}.
\end{eqnarray}
As the $p_i$'s are general points, this immediately implies that
$$z_{S}=\mu_d c_2(S)+\nu_dc_1(S)^2
\,\,{\rm in}\,\,{\rm CH}_0( S)_\mathbb{Q}.$$
Finally, the above discussion was concerning polarized surfaces in a bounded family
(depending on $d$). Taking $d$ larger, we conclude that the coefficients
$\mu_d$ and $\nu_d$ do not depend on $d$, which concludes the proof, under the assumption made
on ${\rm CH}_0(S)$.
It now only remains to prove the result for surfaces with ${\rm CH}_0(S)=\mathbb{Z}$.
In this case, we can use the following trick, using the fact that we already know that
$$\mu_d=\mu,\,\,\nu_d=\nu$$
are independent of $d$ (supposed to be large), so that
$$z_{S}=\mu c_2(S)+\nu c_1(S)^2$$
when $S$ is the regular locus of the complete intersection
of four hypersurfaces of degree $d$ in $G$. Here $\mu,\,\nu$ are {\it a priori} rational numbers
but in fact, they are integers because $(4d-5)^5\mu_d,\,(4d-5)^2\nu_d$ are
integers by Corollary \ref{corpourthefinsd}.
We observe now that with the notation introduced for Proposition
\ref{leunivsurfd},
the equality
$$z_{\mathcal{S}_d}=\mu c_2(K_{\mathcal{S}_d/B})+\nu c_1(K_{\mathcal{S}_d/B})^2$$
must hold in ${\rm CH}^2(\mathcal{S}_d)$, because both sides are integral cycles,
the equality is satisfied in ${\rm CH}^2(\mathcal{S}_d)_\mathbb{Q}$, and
${\rm CH}^2(\mathcal{S}_d)$ has no torsion.
We choose as before a very ample rank $2$ vector bundle $E$ on $S$ with $c_2(E)=c\in {\rm CH}_0(S)=\mathbb{Z}$, and embed
$S$ in $G$ using $5$ general sections of $E$. For large $d$, we then
have as before
the family $(B_S\times S)\setminus \mathcal{C}\subset \mathcal{S}_d$, where
$B_S$ is a Zariski open set of $I_S(d)^4$, and
we conclude by restriction, using (i) and (ii) of Definition
\ref{defiunivdef} that
$$(pr_S^*z_S)_{\mid (B_S\times S)\setminus \mathcal{C}}=(\mu pr_S^*c_2(K_{S})+\nu pr_S^*c_1(K_S)^2)_{\mid (B_S\times S)\setminus \mathcal{C}}$$
holds in ${\rm CH}^2((B_S\times S)\setminus \mathcal{C})$.
In other words, we proved the same equalities as before, but in ${\rm CH}$ instead of
${\rm CH}_\mathbb{Q}$.
We now apply the localization exact sequence and Lemma \ref{propCHCuniv} to conclude
that (\ref{eqder5no})
holds in fact with integral cycles, namely
\begin{eqnarray}\label{eqder5nopasder} z_S=\mu c_2(S)+\nu c_1(S)^2-\alpha s_2(N_{S/G}(-d))-((4d-5)L-K_S)\cdot w_S\,\, {\rm in}\,\,{\rm CH}_0(S),
\end{eqnarray}
where $\alpha$ is an integer and $w_S\in {\rm CH}^1(S)$.
We now choose $d$, $c$ and $L={\rm det}\,E$
in such a way that an arbitrarily large given integer $M$ divides
both $s_2(N_{S/G}(-d))$ and $(4d-5)L-K_S\in {\rm NS}(S)$. This is possible because
\begin{eqnarray}
\label{eqs2}s_2(N_{S/G}(-d))=c+Q(L,K_S),
\end{eqnarray}
where $Q$ is a degree $2$ polynomial with integer coefficients.
Then we apply
formula (\ref{eqder5nopasder}), and we
conclude that
$M$ divides $z_S-\mu c_2(S)-\nu c_1(S)^2 $ in ${\rm CH}_0(S)=\mathbb{Z}$. As
$M$ is arbitrarily large, it follows that
$z_S=\mu c_2(S)+\nu c_1(S)^2$.
\end{proof}
|
\section{Introduction}
Rapid development in technology makes it possible to
collect measurements intensively over an entire time domain.
This forms the so-called \textit{sample curve}. In functional data analysis,
one may regress the response variable on the sample curve using
(generalized) functional linear models, as in, for example, \cite
{MS05,DPZ12}. Functional principle component analysis (FPCA) is
commonly used for analyzing such models; see, for instance, \cite
{BIW11,YMW05b,HMW06,CH06,HH07,HMV13,L13}. For example, M\"{u}ller and
Stadtm\"{u}ller \cite{MS05} proposed a set of FPCA-based inference
procedures, while Dou et al. \cite{DPZ12} established minimax estimation
rates in a similar framework. The success of these FPCA-based
approaches hinges on the availability of a good estimate of the
functional principal components for the slope function; see \cite
{CY12}. On the other hand, the truncation parameter in the FPCA changes
in a discrete manner, which may yield an imprecise control on the model
complexity, as pointed out in \cite{RS05}. Recently,
Crambes et al. \cite{CKS09}, Yuan and Cai \cite{YC10} and Cai and Yuan \cite{CY12}, among others, have
proposed roughness regularization methods that circumvent the
aforementioned perfect alignment requirement and allow one to
regularize the model complexity in a continuous manner. As far as we
are aware, these works focus mostly on the \textit{estimation or
prediction} in the functional linear models. An exception is the
prediction intervals obtained in \cite{CKS09} under the restrictive
Gaussian errors; see (5.3) therein. However, it is yet unknown how to
handle a broader range of inference problems such as (adaptive)
hypothesis testing for generalized functional linear models in the
above roughness regularization framework.
The major goal of this paper is to systematically conduct asymptotic
inference in the class of generalized functional linear models, which
cover $\ell_2$ regression, logistic regression and exponential family
models. Specifically, we construct confidence intervals for regression
mean, prediction intervals for future response and various statistical
procedures for hypothesis testing. As far as we are aware, all these
inference results are new. In particular, these inference procedures
maintain the modeling and computation flexibility by taking advantage
of the roughness regularization. However, this practical superiority
comes at the price of a much harder theoretical investigation. A key
technical tool we develop in this paper is the \textit{Bahadur
representation for functional data}, which provides a unified treatment
for various inference problems. Due to the involvement of a covariance
operator, we note that this new Bahadur representation is dramatically
different from that recently established in the nonparametric
regression framework \cite{SC13}. In addition, we employ the
integro-differential equation techniques \cite{T1927,TL1928,T1930} to
explicitly characterize the underlying eigen-system that leads to more
transparent inference procedures; see Proposition~\ref{validA3}. As a
side remark, our general theory does not require the Sacks--Ylvisaker
(SY) conditions as in \cite{YC10}, although assuming a pseudo version
of SY conditions (given in Section~S.2) can
facilitate the implementation.
To be more specific, we show that the proposed confidence/prediction
intervals asymptotically achieve
the desirable coverage probability. We also propose a procedure for
testing functional contrast
and show the null limit distribution as a standard normal distribution.
As for testing global behaviors of the slope function, we propose a
penalized likelihood ratio test (PLRT) that achieves the minimax rate
of testing established in \cite{HMV13}. In the particular case of
functional linear models, we observe a new version of the Wilks
phenomenon \cite{W38,FZZ01} arising from PLRT, by which we mean that
the null limit distribution, which is derived as a Chi-square
distribution with diverging degrees of freedom, is free of the true
model parameters. A major advantage of the Wilks type of results is
that we can directly simulate the null limit distribution (without
resorting to bootstrap) in practice. In PLRT, we also point out that
the class of functions in the alternative hypothesis is allowed to be
infinite-dimensional in contrast to the parametric class considered in
\cite{HMV13}. Besides, the rejection region of PLRT is based on the
asymptotic distribution, which makes the procedure more applicable in
general modeling setup, that is, in \textit{generalized} functional linear models.
In reality, the smoothness of the slope function and the structure of
the predictors are typically unknown.
To address this issue, we modify the above PLRT in an ``adaptive''
fashion. Explicitly, we conduct a sequence of standardized PLRT
procedures over multiple smoothness levels, and then use the maximal
one as the new test (after a second standardization). This new testing
method does not rely on prior knowledge of the above two crucial
quantities, and is shown to achieve the minimax rate of testing (up to
logarithm term) established in \cite{HMV13}.
In fact, our adaptive procedures can be viewed as a generalization of
the adaptive Neyman test studied in \cite{F96,FL98} to functional
data. Due to the distinct model structure and test construction, the
Darling--Erd\H{o}s theorem used in \cite{F96,FL98} is no longer
applicable. Instead, we adapt the impressive and powerful Gaussian
approximation tool recently proposed in \cite{CCK13} to show that in
both Gaussian and sub-Gaussian settings, the null limit is a type of
extreme value distribution. Our adaptive testing procedures differ from
the FPCA-based tests such as those considered in \cite{HMV13,L13} in
two ways: (i) our tests work for non-Gaussian models; (ii) our tests
provide an asymptotic null limit distribution, from which the correct
test size can be achieved. Besides, our tests do not require the
``eigen-gap'' condition in the FPCA literature, as in, for example,
\cite{L13}. Simulation results demonstrate the advantages of our
methods in terms of desirable sizes and powers. In particular, we
observe that PLRT is more powerful than the adaptive testing
procedures. This is reasonable since PLRT incorporates prior knowledge
on smoothness of the covariance and reproducing kernels. However, their
difference quickly vanishes when the sample size is large or the signal
strength is strong.
The rest of this paper is organized in the following way. In
Section~\ref{secpreliminary}, basic assumptions on model and parameter space
are given. Section~\ref{secbrfd} presents the key technical device of
this paper:
Bahadur representation for functional data. In Section~\ref
{secasympCI}, asymptotically valid confidence intervals
for regression mean and prediction intervals for future response are
constructed.
In Section~\ref{sechypothesistesting}, a procedure for testing
functional contrast and a global testing for the slope function, that
is, PLRT, are established. Theoretical properties are also
demonstrated. Section~\ref{secadaptivePLRT} contains two adaptive
testing procedures for either Gaussian or sub-Gaussian errors. Their
null limit distributions and minimax properties are carefully examined.
A simulation study is provided in Section~\ref{secsimulation}. The
generalized cross validation (GCV) is used to select the roughness
penalty parameter in the simulations. Section~\ref{secdisc} discusses
the technical connection between our work and \cite{SC13}. All
technical proofs are deferred to the Supplementary Material \cite{SCFDA}.
\section{Preliminaries}\label{secpreliminary}
\subsection{Model assumptions}
Suppose the data $(Y_i,X_i(t))$, $i=1,\ldots,n$, are i.i.d. copies of
$(Y,X(t))$,
where $Y$ is a univariate response variable taking values in $\mathcal{Y}$,
a subset of real numbers, and $X(t)$ is a real-valued random predictor
process over $\mathbb{I}=[0,1]$.
Consider the following generalized functional linear model:
\begin{equation}
\label{gflm} \mu_0(X)\equiv E\{Y\mid X\}=F \biggl(
\alpha_0+\int_0^1 X(t)\beta
_0(t)\,dt \biggr),
\end{equation}
where $F$ is a known link function,
$\alpha_0$ is a scalar and $\beta_0(\cdot)$ is a real-valued
function. The conditional mean w.r.t. $X=X(\cdot)$ can be understood
as a function of a collection of random variables
$\{X(t)\dvtx 0\le t\le1\}$ throughout the paper. Let $\beta\in
H^m(\mathbb{I})$, the $m$-order Sobolev space
defined by
\begin{eqnarray*}
H^m(\mathbb{I})&=&\bigl\{\beta\dvtx \mathbb I\mapsto\mathbb R\mid
\beta^{(j)}, j=0,\ldots,m-1,
\\
&&\hspace*{4pt}\mbox{are absolutely continuous, and }\beta^{(m)}\in L^2(
\mathbb{I}) \bigr\}.
\end{eqnarray*}
Therefore, the unknown parameter $\theta\equiv(\alpha,\beta)$
belongs to $\mathcal{H}\equiv\bbR^1\times H^m(\mathbb{I})$. We
further assume $m>1/2$ such that $H^m(\mathbb{I})$ is a reproducing
kernel Hilbert space.
In this paper, we consider a general loss function $\ell(y;a)$ defined
over $y\in\mathcal{Y}$ and $a\in\bbR$, which covers two important
classes of statistical models: (i)~$\ell(y; a)=\log p(y; F(a))$, where
$y\mid x\sim p(y;\mu_0(x))$ for a conditional distribution $p$;
(ii)~$\ell
(y; a)=Q(y; F(a))$, where $Q(y; \mu)\equiv\int_{y}^\mu
(y-s)/\mathcal V(s)\,ds$ is a quasi-likelihood with some known
positive-valued function $\mathcal{V}$ satisfying $\mathcal V(\mu
_0(X))=\operatorname{Var}(Y\mid X)$; see \cite{W74}. Note that these two criterion
functions coincide under some choices of~$\mathcal V$. The regularized
estimator is given by
\begin{eqnarray}
\label{estimation} \qquad &&(\widehat{\alpha}_{n,\lambda},\widehat{\beta
}_{n,\lambda
})\nonumber
\\
&&\qquad = \arg\sup_{(\alpha,\beta)\in\mathcal{H}}\ell_{n,\lambda
}(\theta)
\\
&&\qquad \equiv \arg\sup_{(\alpha,\beta)\in\mathcal{H}} \Biggl\{ \frac
{1}{n}\sum
_{i=1}^n\ell\biggl(Y_i;
\alpha+\int_0^1 X_i(t)\beta(t)\,dt
\biggr)-(\lambda/2)J(\beta,\beta) \Biggr\},\nonumber
\end{eqnarray}
where $J(\beta,\widetilde{\beta})=\int_0^1\beta^{(m)}(t)\widetilde
{\beta}^{(m)}(t)\,dt$ is a roughness penalty. Here, we use $\lambda/2$
to simplify future expressions. In the special $\ell_2$-regression,
Yuan and Cai \cite{YC10} study the minimax optimal estimation and
prediction by assuming the same roughness penalty.
We next assume the following smoothness and tail conditions on $\ell$.
Denote the first-, second- and third-order derivatives of $\ell(y;a)$
w.r.t. $a$ by $\dot\ell_a(y;a)$, $\ddot{\ell}_a(y;a)$ and $\ell
'''_a(y;a)$, respectively.
\begin{Assumption}\label{A1}
\textup{(a)} $\ell(y;a)$ is three times continuously
differentiable and strictly concave w.r.t $a$.
There exist positive
constants $C_0$ and $C_1$ s.t.,
\begin{eqnarray}
\label{A1aeq1} E \Bigl\{\exp\Bigl(\sup_{a\in\bbR}\bigl\llvert
\ddot{\ell}_a(Y;a)\bigr\rrvert/C_0 \Bigr)\big| X \Bigr\}
&\le& C_1,
\nonumber\\[-8pt]\\[-8pt]\nonumber
E \Bigl\{\exp\Bigl(\sup_{a\in\bbR} \bigl\llvert
\ell'''_a(Y;a)\bigr\rrvert
/C_0 \Bigr) \big| X \Bigr\}&\le& C_1,\qquad\mbox{a.s.}
\end{eqnarray}
(b) There exists a positive constant $C_2$ s.t.,
\[
C_2^{-1}\le B(X)\equiv-E \biggl\{\ddot{
\ell}_a \biggl(Y;\alpha_0+\int_0^1X(t)
\beta_0(t)\,dt \biggr)\big| X \biggr\}\le C_2\qquad\mbox{a.s.}
\]
In addition, $X$ is weighted-centered in the sense that $E\{B(X)X(t)\}
=0$ for any $t\in\mathbb{I}$.
(c)
$\epsilon\equiv
\dot{\ell}_a (Y;\alpha_0+\int_0^1X(t)\beta_0(t)\,dt )$ satisfies
$E\{\epsilon\mid X\}=0$ and $E\{\epsilon^2\mid X\}= B(X)$, a.s.
\end{Assumption}
The weighted center condition in Assumption~\ref{A1}(b) is only used to simplify our technical arguments. Actually, it
always holds after a simple data transformation; see the Supplementary
Material~\cite{SCFDA}, Section~S.1. Next, we give three
examples to illustrate the validity of Assumption~\ref{A1}.
\begin{Example}[(Gaussian model)]\label{exa1pss}
In the functional linear models under Gaussian errors, that is,
$Y=\alpha_0+\int_0^1 X(t)\beta_0(t)\,dt+v$ and $v\mid X\sim N(0,\sigma
^2)$, we can easily verify Assumption~\ref{A1} with $B(X)=\sigma
^{-2}$ and $\epsilon=v/\sigma^2$ given that $E\{X(t)\}=0$.
\end{Example}
\begin{Example}[(Logistic model)]\label{exa2logit}
In the logistic regression, we assume
$P(Y=1\mid X)=1-P(Y=0\mid X)=\exp(\alpha_0+\int_0^1X(t)\beta_0(t)\,dt)/
(1+\exp(\alpha_0+\int_0^1X(t)\beta_0(t)\,dt))$. It is easy to see
that $\ell(y;a)=ay-\log(1+\exp(a))$ and
$B(X)=\exp(\alpha_0+\int_0^1 X(t)\beta_0(t)\,dt)/(1+\exp(\alpha
_0+\int_0^1X(t)\beta_0(t)))^2\leq1$. Assumption~\ref{A1}(a) follows from simple algebra. Assumption~\ref{A1}(b)
follows from data transformation and the following $L^2$ bounded
condition: $\int_0^1X^2(t)\,dt\leq c$ a.s. The latter condition implies
that the range\vspace*{1pt} of $\mu_0(X)$ is finite, and thus $B(X)$ is bounded
away from zero. Since $\epsilon=Y-\exp(X^T\theta_0+g_0(Z))/(1+\exp
(X^T\theta_0+g_0(Z)))$, Assumption~\ref{A1}(c) can be
verified by direct calculations.
\end{Example}
\begin{Example}[(Exponential family)]\label{exa2exponentialfamily}
Let $(Y,X)$ follow the one-parameter exponential family
\[
Y\mid X\sim\exp\biggl\{Y\biggl(\alpha_0+\int_0^1
X(t)\beta_0(t)\,dt\biggr)+A(Y)-G\biggl(\alpha_0+\int
_0^1 X(t)\beta_0(t)\,dt\biggr)
\biggr\},
\]
where $A(\cdot)$ and $G(\cdot)$ are known, and $\dot{G}=F$ [recall that
$F$ is the link function satisfying (\ref{gflm})]. We assume that $G$
has bounded second- and third-order derivatives,
and $\ddot{G}\ge\delta$ for some constant $\delta>0$; see similar
conditions on page~738\vspace*{1pt} of \cite{MS05}. It is easy to see that $\ell
(y;a)=ya+A(y)-G(a)$, and hence, $\dot{\ell}_a(y;a)=y-\dot{G}(a)$,
$\ddot{\ell}_a(y;a)=-\ddot{G}(a)$ and $\ell'''_a(y;a)=-\dddot
{G}(a)$. Clearly, $\ddot{\ell}_a$ and $\ell'''_a$ are both bounded,
and hence
Assumption~\ref{A1}(a) holds.
Furthermore, $B(X)=\ddot{G}(\alpha_0+\int_0^1 X(t)\beta_0(t)\,dt)$
satisfies Assumption~\ref{A1}(b).
Since $\epsilon=Y-\dot{G}(\alpha_0+\int_0^1 X(t)\beta
_0(t)\,dt)=Y-\mu_0(X)$,
it is easy to see that $E\{\epsilon\mid X\}=E\{Y\mid X\}-\mu_0(X)=0$,
and $E\{\epsilon^2\mid X\}=\operatorname{Var}(Y\mid X)=\ddot{G}(\alpha_0+\int_0^1
X(t)\beta_0(t)\,dt)$
(see \cite{M82}), and therefore, Assumption~\ref{A1}(c) holds.
\end{Example}
\subsection{Reproducing kernel Hilbert space}
We introduce an inner product in $H^m(\mathbb{I})$, that is,
\begin{equation}
\label{innerprod1} \langle\beta,\widetilde{\beta}\rangle_1=V(\beta,
\widetilde{\beta})+\lambda J(\beta,\widetilde{\beta}),
\end{equation}
where $V(\beta,\widetilde{\beta})\equiv\int_0^1\int_0^1
C(s,t)\beta(t)\widetilde{\beta}(s)\,ds\,dt$ and $C(s,t)\equiv E\{
B(X)X(t)X(s)\}$ is a weighted covariance function. Denote the
corresponding norm as $\llVert\cdot\rrVert_1$. Define a
linear bounded operator $C(\cdot)$ from $L^2(\mathbb{I})$ to
$L^2(\mathbb{I})$: $(C\beta)(t)=\int_0^1 C(s,t)\beta(s)\,ds$. Below
we assume a regularity condition on $C\beta$, which implies the
positive definiteness of $V$, such that the above inner product (\ref
{innerprod1}) is well defined.
\begin{Assumption}\label{A2}
$C(s,t)$ is continuous on $\mathbb{I}\times\mathbb{I}$.
Furthermore, for any $\beta\in L^2(\mathbb{I})$ satisfying $C\beta
=0$, we have $\beta=0$.
\end{Assumption}
Suppose that $C$ is continuous over $\mathbb{I}\times
\mathbb{I}$. By Mercer's theorem, $C$ admits the spectral
decomposition\vspace*{1pt} $C(s,t)=\sum_{\nu=1}^\infty\zeta_\nu\psi_\nu
(s)\psi_\nu(t)$, where $\{\psi_\nu(\cdot), \zeta_\nu\ge0\}_{\nu
\ge1}$ forms an orthonormal basis in $L^2(\mathbb{I})$ under the
usual $L^2$-norm. Therefore, for any $\beta\in L^2(\mathbb I)$, we
have $\beta(\cdot)=\sum_{\nu=1}^\infty b_\nu\psi_\nu(\cdot)$
and $V(\beta,\beta)=\sum_{\nu=1}^\infty\zeta_\nu b_\nu^2$ for a
sequence of square summable $b_\nu$'s. Assumption~\ref{A2} directly
implies that all the eigenvalues of $C$ are positive,
that is, $\zeta_\nu>0$ for all $\nu\ge1$. Therefore, if \mbox{$V(\beta,\beta
)=0$}, that is, $\sum_{\nu=1}^\infty\zeta_\nu b_\nu^2=0$, we
can easily show that $\beta=\sum_{\nu=1}^\infty b_\nu\psi_\nu=0$.
Hence $\langle\cdot,\cdot\rangle_1$ is well defined. Moreover,
together with Proposition 2 of \cite{YC10}, Assumption~\ref{A2}
implies that $H^m(\mathbb{I})$ is indeed a reproducing kernel Hilbert
space (RKHS) under $\langle\cdot,\cdot\rangle_1$. We denote its
reproducing kernel function as $K(s,t)$.
As for the joint parameter space $\mathcal H$, we also need to assume a
proper inner product under which\vspace*{1pt} it is a well-defined Hilbert space.
Define, for any $\theta=(\alpha,\beta)$, $\widetilde{\theta
}=(\widetilde{\alpha},\widetilde{\beta})\in\mathcal{H}$,
\begin{eqnarray}\label{innerprod}
\langle\theta, \widetilde{\theta}\rangle&\equiv& E
\biggl\{B(X) \biggl(\alpha+\int_0^1X(t)\beta(t)\,dt \biggr)
\biggl(\widetilde{\alpha}+\int_0^1X(t)
\widetilde{\beta}(t)\,dt \biggr) \biggr\}
\nonumber\\[-8pt]\\[-8pt]\nonumber
&&{} +\lambda J(\widetilde\beta, \beta).\nonumber
\end{eqnarray}
By rewriting $\langle\theta, \widetilde{\theta}\rangle=E\{B(X)\}
\alpha\widetilde\alpha+\langle\beta,\widetilde\beta\rangle_1$,
we note that $\langle\cdot,\cdot\rangle$ is a well-defined inner
product under Assumptions~\ref{A1}(b) and~\ref{A2}. The
corresponding norm is denoted as $\llVert\cdot\rrVert$.
Given the above relation between $\langle\cdot,\cdot\rangle$ and
$\langle\cdot,\cdot\rangle_1$, it is easy to show that $\mathcal H$
inherits the completeness of $H^m(\mathbb I)$. This means $\mathcal
{H}$ is indeed a Hilbert space as described in Proposition~\ref
{Hhilbert} below.
\begin{Proposition}\label{Hhilbert}
Under $\langle\cdot,\cdot\rangle$, $\mathcal{H}$ is a Hilbert space.
\end{Proposition}
In the literature, the estimation/prediction rate results in the
(generalized) functional linear models are mostly expressed in terms of
$L^2$-norm; see \cite{MS05,HMW06,HH07,YC10}. We remark that our norm
$\llVert\cdot\rrVert$ is stronger than the $L^2$-norm used
in the above literature under Assumption~\ref{A1}(b).
We next assume a sequence of basis functions in $H^m(\mathbb{I})$
which can simultaneously diagonalize $V$ and $J$. Let $\llVert
\cdot\rrVert_{L^2}$ and $\llVert\cdot\rrVert_{\sup}$
denote the $L^2$ and supremum norms, respectively. Denote $a_n\asymp
b_n$ if and only if there exist positive
constants $c_1,c_2$ such that $c_1\le a_\nu/b_\nu\le c_2$ for all
$\nu$.
\begin{Assumption}\label{A3}
There exists a sequence of functions $\{\varphi_\nu\}_{\nu\ge
1}\subset H^m(\mathbb{I})$ such that
$\llVert\varphi_\nu\rrVert_{L^2}\le C_\varphi\nu^{a}$
for each $\nu\ge1$, some constants $a\ge0, C_\varphi>0$ and
\begin{equation}
\label{simuldiag} V(\varphi_\nu,\varphi_\mu)=
\delta_{\nu\mu},\qquad J(\varphi_\nu,\varphi_\mu)=
\rho_\nu\delta_{\nu\mu}\qquad\mbox{for any }\nu,\mu\ge1,
\end{equation}
where $\delta_{\nu\mu}$ is Kronecker's notation, and $\rho_\nu$ is
a nondecreasing nonnegative
sequence satisfying $\rho_\nu\asymp\nu^{2k}$ for some constant $k>a+1/2$.
Furthermore, any $\beta\in H^m(\mathbb{I})$ admits the
Fourier expansion $\beta=\sum_{\nu=1}^\infty V(\beta,\varphi_\nu
)\varphi_\nu$ with
convergence in $H^m(\mathbb{I})$ under $\langle\cdot,\cdot\rangle_1$.
\end{Assumption}
We remark that Assumption~\ref{A3} is the price we need to
pay for
making valid statistical inference in addition to those required for
minimax estimation, as in, for example, \cite{YC10,CY12}.
Assumption~\ref{A3} can be directly implied by the pseudo
Sacks--Ylvisaker (SY) conditions, which are slightly different from the
conventional SY conditions proposed in \cite
{SY1966,SY1968,SY1970,RWW95}; see Section~S.2 in
the Supplementary Material \cite{SCFDA} for more details. Proposition
\ref{validA3} below discusses the construction of an eigen-system
satisfying Assumption~\ref{A3} under this condition.
\begin{Proposition}[(Eigen-system construction)]\label{validA3}
Suppose the covariance function $C$
satisfies Assumption~\ref{A2} and the pseudo SY conditions of order
$r\ge0$ specified in Section~\textup{S.2}.
Furthermore, the boundary value problem \textup{(S.2)} in Section~\textup{S.3} is regular in the sense of \cite{Birk1908}.
Consider the following integro-differential equations:
\begin{equation}
\label{bvp1} \cases{\displaystyle (-1)^m y^{(2m)}_\nu(t)=
\rho_\nu\int_0^1C(s,t)y_\nu(s)\,ds,
\vspace*{5pt}\cr
\displaystyle y_\nu^{(j)}(0)=y_\nu^{(j)}(1)=0, \qquad\mbox{$j=m,\ldots,2m-1$.}}
\end{equation}
Let $(\rho_\nu,y_\nu)$ be the corresponding eigenvalues and
eigenfunctions of problem (\ref{bvp1}),
and let $\varphi_\nu=y_\nu/\sqrt{V(y_\nu,y_\nu)}$. Then $(\rho
_\nu,\varphi_\nu)$ satisfy Assumption~\ref{A3}
with $k=m+r+1$ and $a=r+1$ if \textit{one} of the following additional
assumptions is satisfied:
\begin{longlist}[(ii)]
\item[(i)]$r=0$;
\item[(ii)]$r\geq1$, and for $j=0,1,\ldots,r-1$, $C^{(j,0)}(0,t)=0$ for\vspace*{1pt}
any $0\le t\le1$,
where $C^{(j,0)}(s,t)$ is the $j$th-order partial derivative with
respect to $s$.
\end{longlist}
\end{Proposition}
The proof of Proposition~\ref{validA3} relies on a
nontrivial application of
the general integro-differential equation theory developed in \cite
{T1927,TL1928,T1930}.
In particular, the order of $\rho_\nu$ in problem (\ref{bvp1}) is,
in general, equivalent
to the order of eigenvalues in an ordinary differential problem; see
\cite{T1927}, Theorem 7.
More explicitly, $\rho_\nu\approx(c\pi\nu)^{2k}$ as $\nu
\rightarrow\infty$ for some constant $c>0$; see
\cite{TL1928}, equation (20).
In the Gaussian model with unit variance (see Example~\ref{exa1pss}),
it can be shown with arguments similar to those
in \cite{SY1966,SY1968,SY1970} that
the covariance function $C$ satisfies the pseudo SY conditions of order $r=0$
when $X(t)$ is Brownian motion with $C(s,t)=\min\{s,t\}$.
We also note that the boundary condition in Proposition~\ref{validA3}(ii) was also assumed in \cite{RWW95}
when $X$ is Gaussian of order $r>0$. The integro-differential equations
(\ref{bvp1}) can be
translated into easily computable differential equations. More
specifically, we rewrite $y_\nu(t)$ in
(\ref{bvp1}) as $\ddot{g}_\nu(t)$, and thus obtain that
\begin{equation}
\label{bvpexample} \cases{ (-1)^{m+1}g_\nu^{(2m+2)}(t)=
\rho_\nu g_\nu(t),
\vspace*{3pt}\cr
g_\nu^{(j)}(0)=g_\nu^{(j)}(1)=0,\qquad\mbox{$j=m+2,\ldots,2m+1$,}
\vspace*{3pt}\cr
g_\nu(0)=\dot{g}_\nu(1)=0.}
\end{equation}
Note that (\ref{bvp1}) and (\ref{bvpexample}) share the same eigenvalues.
Numerical examinations show that
$\rho_\nu\approx(\pi\nu)^{2(m+1)}$.
The function $g_\nu$'s have closed forms
\[
g_\nu(t)=\operatorname{Re} \Biggl(\sum_{j=1}^{2(m+1)}a_{\nu, j}
\exp\bigl(\rho_\nu^{1/(2(m+1))}z_jt\bigr) \Biggr),\qquad
\nu=1,2,\ldots,
\]
where $\operatorname{Re}(\cdot)$ means the real part of a complex number,
$z_1,\ldots,z_{2(m+1)}$ are the complex (distinct) roots of
$z^{2(m+1)}=(-1)^{m+1}$
and $a_{\nu,1},\ldots,a_{\nu,2(m+1)}$ are complex constant coefficients
determined by the boundary value conditions in (\ref{bvpexample}).
It follows by Proposition~\ref{validA3} that the resultant
$\rho_\nu$ and the corresponding scaled functions $\varphi_\nu
=y_\nu/\sqrt{V(y_\nu,y_\nu)}$
satisfy Assumption~\ref{A3},
where recall that $y_\nu$ is the second-order derivative of $g_\nu$.
In the logistic regression (Example~\ref{exa2logit}) or exponential
family models (Example~\ref{exa2exponentialfamily}),
the approach given in Section~S.5 (Supplementary Material \cite{SCFDA})
can be used to find $(\rho_\nu,\varphi_\nu)$
without verifying the pseudo SY conditions. To do so, we need to
replace the kernel function $C$
by its sample version $C_n(s,t)\equiv n^{-1}\sum_{i=1}^n \widehat
{B}(X_i)X_i(s)X_i(t)$,
where $\widehat{B}(X)$ is the plug-in estimate of $B(X)$.
Recall that $K$ is the reproducing kernel function for $H^m(\mathbb
{I})$ under $\langle\cdot,\cdot\rangle_1$.
For any $t\in\mathbb{I}$, define $K_t(\cdot)=K(t,\cdot)\in
H^m(\mathbb I)$. Under Assumption~\ref{A3}, we may write
$K_t=\sum_{\nu\ge1}a_\nu\varphi_\nu$ for a real sequence $a_\nu$.
Clearly, $\varphi_\nu(t)=\langle K_t,\varphi_\nu\rangle_1=a_\nu
(1+\lambda\rho_\nu)$, for all $\nu\ge1$.
So $K_t=\sum_{\nu\ge1}\frac{\varphi_\nu(t)}{1+\lambda\rho_\nu
}\varphi_\nu$. Define $W_\lambda$ as an operator from $H^m(\mathbb
{I})$ to $H^m(\mathbb{I})$ satisfying $\langle W_\lambda\beta,\widetilde
{\beta}\rangle_1=\lambda J(\beta,\widetilde{\beta})$,
for all $\beta,\widetilde{\beta}\in H^m(\mathbb{I})$. Hence
$W_\lambda$ is linear, nonnegative definite and self-adjoint.
For any $\nu\ge1$, write $W_\lambda\varphi_\nu=\sum_\mu b_\mu
\varphi_\mu$.
Then by Assumption~\ref{A3}, for any $\mu\ge1$, $\lambda\rho_\nu
\delta_{\nu\mu}=
\lambda J(\varphi_\nu,\varphi_\mu)=\langle W_\lambda\varphi_\nu,\varphi
_\mu\rangle_1=b_\mu(1+\lambda\rho_\mu)$.
Therefore, $b_\nu=\lambda\rho_\nu/(1+\lambda\rho_\nu)$ and
$b_\mu=0$ if $\mu\neq\nu$,
which implies $W_\lambda\varphi_\nu=\frac{\lambda\rho_\nu
}{1+\lambda\rho_\nu}\varphi_\nu$.
Thus we have shown the following result.
\begin{Proposition}\label{PropRPK} Suppose Assumption~\ref{A3} holds.
For any $t\in\mathbb{I}$,
\[
K_t(\cdot)=\sum_\nu
\frac{\varphi_\nu(t)}{1+\lambda\rho_\nu} \varphi_\nu(\cdot),\vspace*{-2pt}
\]
and for any $\nu\ge1$,\vspace*{-2pt}
\[
(W_\lambda\varphi_\nu) (\cdot)= \frac{\lambda\rho_\nu}{1+\lambda\rho_\nu}
\varphi_\nu(\cdot).
\]
\end{Proposition}
Propositions~\ref{propR} and~\ref{propP} below define two
operators, $R_x$ and $P_\lambda$, that will be used in the Fr\'{e}chet
derivatives of the criterion function $\ell_{n, \lambda}$. We first
define $\tau(x)$ as follows. For any $L^2$ integrable function
$x=x(t)$ and $\beta\in H^m(\mathbb{I})$,
$\mathcal{L}_x(\beta)\equiv\int_0^1 x(t)\beta(t)\,dt$ defines a
linear bounded functional. Then by the Riesz representation theorem,
there exists an element in $H^m(\mathbb{I})$, denoted as $\tau(x)$,
such that $\mathcal{L}_x(\beta)=\langle\tau(x),\beta\rangle_1$
for all $\beta\in H^m(\mathbb{I})$.
If we denote $\tau(x)=\sum_{\nu=1}^\infty x^*_\nu\varphi_\nu$,
then $x^*_\nu(1+\lambda\rho_\nu)=\langle\tau(x),\varphi_\nu
\rangle_1=\int_0^1x(t)\varphi_\nu(t)\,dt\equiv x_\nu$ for\vspace*{1pt} any $\nu
\ge1$.
Thus $\tau(x)=\sum_{\nu=1}^\infty\frac{x_\nu}{1+\lambda\rho_\nu
}\varphi_\nu$.
\begin{Proposition}\label{propR}
For any $x\in L^2(\mathbb{I})$,
define $R_x=(E\{B(X)\}^{-1},\tau(x))$.
Then $R_x\in\mathcal{H}$ and $\langle R_x,\theta\rangle=\alpha
+\int_0^1 x(t)\beta(t)\,dt$ for any $\theta=(\alpha,\beta)\in
\mathcal{H}$.
\end{Proposition}
It should be noted that $R_x$ depends on $h$ according to the
definition of $\tau(x)$.
\begin{Proposition}\label{propP}
For\vspace*{2pt} any $\theta=(\alpha,\beta)\in\mathcal{H}$, define $P_\lambda
\theta=(0,W_\lambda\beta)$.
Then $P_\lambda\theta\in\mathcal{H}$ and
$\langle P_\lambda\theta,\widetilde{\theta}\rangle=\langle
W_\lambda\beta,\widetilde{\beta}\rangle_1$ for any $\widetilde
{\theta}=(\widetilde{\alpha},\widetilde{\beta})\in\mathcal{H}$.
\end{Proposition}
For notational convenience, denote $\Delta\theta=(\Delta\alpha,\Delta
\beta)$
and $\Delta\theta_j=(\Delta\alpha_j,\Delta\beta_j)$ for
$j=1,2,3$. The Fr\'{e}chet derivative of $\ell_{n,\lambda}(\theta)$
w.r.t. $\theta$ is given by
\[
S_{n,\lambda}(\theta)\Delta\theta\equiv D\ell_{n,\lambda}(\theta)\Delta
\theta=\frac{1}{n}\sum_{i=1}^n\dot{
\ell}_a\bigl(Y_i;\langle R_{X_i},\theta\rangle
\bigr)\langle R_{X_i},\Delta\theta\rangle-\langle P_\lambda
\theta,\Delta\theta\rangle.
\]
The second- and third-order Fr\'{e}chet derivatives of $\ell
_{n,\lambda}(\theta)$ can be shown to be, respectively,
\begin{eqnarray*}
&&DS_{n,\lambda}(\theta)\Delta\theta_1\Delta\theta_2
\\[-2pt]
&&\qquad \equiv D^2\ell_{n,\lambda}(\theta)\Delta\theta_1
\Delta\theta_2
\\[-2pt]
&&\qquad = \frac{1}{n}\sum_{i=1}^n \ddot{
\ell}_a\bigl(Y_i;\langle R_{X_i},\theta\rangle
\bigr)\langle R_{X_i},\Delta\theta_1\rangle\langle
R_{X_i},\Delta\theta_2\rangle-\langle P_\lambda
\Delta\theta_1,\Delta\theta_2\rangle
\end{eqnarray*}
and
\begin{eqnarray*}
&&D^2S_{n,\lambda}(\theta)\Delta\theta_1\Delta
\theta_2\Delta\theta_3
\\[-2pt]
&&\qquad \equiv D^3\ell_{n,\lambda}(\theta)\Delta\theta_1
\Delta\theta_2\Delta\theta_3
\\[-2pt]
&&\qquad =\frac{1}{n}\sum_{i=1}^n
\ell'''_a
\bigl(Y_i;\langle R_{X_i},\theta\rangle\bigr) \langle
R_{X_i},\Delta\theta_1\rangle\langle R_{X_i},
\Delta\theta_2\rangle\langle R_{X_i},\Delta
\theta_3\rangle.
\end{eqnarray*}
Define $S_n(\theta)=\frac{1}{n}\sum_{i=1}^n\dot{\ell
}_a(Y_i;\langle R_{X_i},\theta\rangle)R_{X_i}$,
$S(\theta)=E\{S_n(\theta)\}$ and $S_\lambda(\theta)=E\{S_{n,\lambda
}(\theta)\}$
with expectations taken under the true model.
\section{Bahadur representation for functional data}\label{secbrfd}
In this section, we extend the functional Bahadur representation
originally established in the nonparametric regression framework \cite
{SC13} to the generalized functional linear models. This new technical
tool is fundamentally important in the sense that it provides a unified
treatment for various inference problems.
Denote $h=\lambda^{1/(2k)}$, where $k$ is specified in Assumption~\ref
{A3}. An auxiliary norm is introduced for technical purpose: $\llVert
\theta\rrVert_2=\llvert \alpha\rrvert +\llVert\beta\rrVert
_{L^2}$ for any $\theta=(\alpha,\beta)\in\mathcal{H}$. The
following result gives a useful relationship between the two norms
$\llVert\cdot\rrVert_2$ and $\llVert\cdot\rrVert
$. Recall that $a$ is defined in Assumption~\ref{A3}.
\begin{Lemma}\label{crucialrelation}
There exists a constant $\kappa>0$ such that for any $\theta\in
\mathcal{H}$,
$\llVert\theta\rrVert_2\le\kappa h^{-(2a+1)/2}\llVert
\theta\rrVert$.
\end{Lemma}
To obtain an appropriate Bahadur representation for the functional
data, we need the following regularity conditions on $X$. Recall that
$\llVert X\rrVert_{L^2}^2=\int_0^1X^2(t)\,dt$.
\begin{Assumption}\label{A4}
There exists a constant $s\in(0,1)$ such that
\begin{equation}
\label{A41} E\bigl\{\exp\bigl(s\llVert X\rrVert_{L^2}\bigr)\bigr\}<
\infty.
\end{equation}
Moreover, suppose that there exists a constant $M_0>0$ such that for
any $\beta\in H^m(\mathbb{I})$,
\begin{equation}
\label{amomconv} E \biggl\{\biggl\llvert\int_0^1
X(t)\beta(t)\,dt\biggr\rrvert^4 \biggr\}\le M_0 \biggl[E
\biggl\{\biggl\llvert\int_0^1 X(t)\beta(t)\,dt
\biggr\rrvert^2 \biggr\} \biggr]^2.
\end{equation}
\end{Assumption}
It is easy to see that (\ref{A41}) holds for any bounded stochastic
process $X$, that is, $\llVert X\rrVert_{L^2}\le c$ a.s.
for some constant $c>0$.
This applies to Example~\ref{exa2logit} which usually requires $X$
to be almost surely bounded in terms of $L^2$-norm.
Equation~(\ref{A41}) also holds for the Gaussian process
as described in Proposition~\ref{validA4} below.
The result applies to Examples~\ref{exa1pss} and~\ref
{exa2exponentialfamily}
where $X$ can be Gaussian.
\begin{Proposition}\label{validA4}
If $X$ is a Gaussian process with square-integrable mean function,
then (\ref{A41}) holds for any $s\in(0,1/4)$.
\end{Proposition}
The fourth moment condition (\ref{amomconv})
is valid for $M_0=3$ when $X$ is a Gaussian process; see \cite{YC10}
for more discussions.
The following result shows that (\ref{amomconv}) actually holds in
more general settings.
\begin{Proposition}\label{propvalidfourthmomcond}
Suppose $X(t)=u(t)+\sum_{\nu=1}^\infty\xi_\nu\omega_\nu\psi_\nu(t)$,
where $u(\cdot)\in L^2(\mathbb{I})$ is nonrandom,
$\psi_\nu$ is orthonormal $L^2(\mathbb{I})$-basis,
$\omega_\nu$ is a real square-summable sequence
and $\xi_\nu$ are independent random variables
drawn from some symmetric distribution with finite fourth-order moment.
Then for any $\beta(t)=\sum_{\nu=1}^\infty b_\nu\psi_\nu(t)$
with $b_\nu$ being real square-summable,
(\ref{amomconv}) holds with $M_0=\max\{E\{\xi_\nu^4\}/E\{\xi_\nu
^2\}^2,3\}$.
\end{Proposition}
Lemma~\ref{eplemma} below proves a concentration inequality as a
preliminary step in obtaining the Bahadur representation. Denote
$T=(Y,X)\in\mathcal{T}$ as the data variable. Let $\psi_n(T;\theta
)$ be a function over
$\mathcal{T}\times\mathcal{H}$, which might depend on $n$. Define
\[
H_n(\theta)=\frac{1}{\sqrt{n}}\sum_{i=1}^n
\bigl[\psi_n(T_i;\theta)R_{X_i}-E_T
\bigl\{\psi_n(T;\theta)R_X\bigr\} \bigr],
\]
where $E_T\{\cdot\}$
means the expectation w.r.t. $T$. Define $\mathcal{F}_{p_n}=
\{\theta=(\alpha,\beta)\in\mathcal{H}\dvtx \llvert \alpha\rrvert \le1,
\llVert
\beta\rrVert_{L^2}\le1, J(\beta,\beta)\le p_n\}$,
where $p_n\ge1$.
\begin{Lemma}\label{eplemma}
Suppose Assumptions~\ref{A1} to~\ref{A4} hold.
In addition, $\psi_n(T_i;0)=0$ a.s., there exists a constant $C_\psi
>0$ s.t., and
the following Lipschitz continuity holds:
\begin{equation}
\label{lipcond} \bigl\llvert\psi_n(T;\theta)-\psi_n(T;
\widetilde{\theta})\bigr\rrvert\le C_\psi\llVert\theta-\widetilde{
\theta}\rrVert_2\qquad\mbox{for any }\theta,\widetilde{\theta}\in
\mathcal{F}_{p_n}.
\end{equation}
Then as $n\rightarrow\infty$,
\[
\sup_{\theta\in\mathcal{F}_{p_n}} \frac{\llVert H_n(\theta)\rrVert
}{p_n^{1/(4m)}\llVert
\theta\rrVert_2^{\gamma}+n^{-1/2}}=O_P\bigl(
\bigl(h^{-1}\log{\log{n}}\bigr)^{1/2}\bigr),
\]
where $\gamma=1-1/(2m)$.
\end{Lemma}
Our last assumption is concerned with the convergence rate of $\widehat
{\theta}_{n,\lambda}$.
Define $r_n=(nh)^{-1/2}+h^k$. Recall that $k$ is specified in
Assumption~\ref{A3}.
\begin{Assumption}\label{A5}
$\llVert\widehat{\theta}_{n,\lambda}-\theta_0\rrVert=O_P(r_n)$.
\end{Assumption}
Proposition~\ref{rateconv} states that Assumptions~\ref{A1} to~\ref
{A4} are actually sufficient to imply the above rate of convergence if
the smoothing parameter is properly chosen. Note that no estimation
consistency is required in Proposition~\ref{rateconv}.
\begin{Proposition}\label{rateconv}
Suppose that Assumptions~\ref{A1} to~\ref{A4} hold, and that the
following rate conditions on $h$ (or equivalently, $\lambda$) are satisfied:
\begin{eqnarray}\label{ratehrateconv}
h&=&o(1),\nonumber
\\
n^{-1/2}h^{-1}&=&o(1),
\\
n^{-1/2}h^{-(a+1)-\sklvafrac{2k-2a-1}{4m}}(
\log{n})^2(\log\log{n})^{1/2}&=&o(1).\nonumber
\end{eqnarray}
Then Assumption~\ref{A5} is satisfied.
\end{Proposition}
It follows by Proposition~\ref{propvalidfourthmomcond} that
if $X(t)=u(t)+\sum_{\nu=1}^\infty\xi_\nu\omega_\nu\psi_\nu(t)$
with $\xi_\nu$ being independent random variables
following symmetric distribution with bounded support, say, $[-N,N]$,
then (\ref{amomconv}) holds. In this case, $X$ is almost surely
$L^2$ bounded since
$\llVert X\rrVert_{L^2}\le\llVert u\rrVert
_{L^2}+\sqrt{\sum_\nu\xi_\nu^2\omega_\nu^2}
\le\llVert u\rrVert_{L^2}+N\sqrt{\sum_\nu\omega_\nu
^2}$, a.s.
Then Proposition~\ref{rateconv} states that if Assumptions~\ref{A1}
to~\ref{A4} hold
and the smoothing parameter is tuned to satisfy (\ref{ratehrateconv}),
then Assumption~\ref{A5} holds for the above $L^2$ bounded~$X$.
Condition (\ref{ratehrateconv}) is satisfied for a suitable range
of $h$. To illustrate this point, we consider the following simple but
representative case.
Under the setup of Proposition~\ref{validA3}, we have $a=r+1$ and $k=m+r+1$.
Suppose $r=0$, that is, $X$ corresponds to zero-order covariance.\vspace*{1pt} Thus
$a=1$ and $k=m+1$.
Denote $h^\ast\asymp n^{-1/(2k+1)}$, $h^{\ast\ast}\asymp
n^{-2/(4k+1)}$ and $h^{\ast\ast\ast}\asymp n^{-1/(2k)}$.
It can be shown that
when $m>(3+\sqrt{5})/4$, $h^\ast$, $h^{\ast\ast}$ and $h^{\ast\ast
\ast}$
all satisfy the conditions of (\ref{ratehrateconv}).
It should be mentioned that $h^\ast$ yields the optimal estimation\vspace*{1pt}
rate $n^{-k/(2k+1)}$ \cite{YC10},
$h^{\ast\ast}$ yields the optimal testing rate $n^{-2k/(4k+1)}$ as
will be shown in later sections,
and $h^{\ast\ast\ast}$ yields the optimal prediction rate \cite{CH06}.
Now we are ready to present the Bahadur representation based on the
functional data.
\begin{Theorem}[(Bahadur representation for functional data)] \label{UBR}
Suppose that Assumptions
\ref{A1}--\ref{A5} hold, and as $n\rightarrow\infty$,
$h=o(1)$ and $\log(h^{-1})=O(\log{n})$.
Furthermore, (\ref{amomconv}) holds.
Then, as $n\rightarrow\infty$,
$\llVert\widehat{\theta}_{n,\lambda}-\theta_0-S_{n,\lambda
}(\theta_0)\rrVert=O_P(a_n)$,
where
\[
a_n=n^{-1/2}h^{-\vafrac{4ma+6m-1}{4m}}r_n(
\log{n})^2(\log\log{n})^{1/2}+C_\ell
h^{-1/2}r_n^2,
\]
and
\[
C_\ell\equiv\sup_{x\in L^2(\mathbb{I})}E\Bigl\{\sup
_{a\in\bbR}\bigl\llvert\ell'''_a(Y;a)
\bigr\rrvert \big| X=x\Bigr\}.
\]
\end{Theorem}
We next give an example rate of $a_n$ in Theorem~\ref{UBR} when $\ell
$ is quadratic. In this case, we have $C_\ell=0$. Suppose $a=1$ and
$k=m+1$; see the discussions below Proposition~\ref{rateconv}.
Direct examinations show that $a_n$ is of the order $o(n^{-1/2})$ when
$m>1+\sqrt{3}/2\approx1.866$
and $h=h^\ast$, $h^{\ast\ast}$ and $h^{\ast\ast\ast}$.
An immediate consequence of Bahadur representation is the following
point-wise limit distribution of the slope function estimate. This
local result is new and of independent interest, for example,
point-wise CI.
\begin{Corollary}\label{localasymp}
Suppose that the conditions of Theorem~\ref{UBR} are satisfied, $\sup
_{\nu\ge1}\llVert\varphi_\nu\rrVert_{\sup}
\le C_{\varphi}\nu^a$ for $\nu\ge1$ and that
$E\{\exp(s\llvert \epsilon\rrvert )\}<\infty$, for some constant $s>0$.
Furthermore, as $n\rightarrow\infty$, $nh^{2a+1}(\log
(1/h))^{-4}\rightarrow\infty$,
$n^{1/2}a_n=o(1)$ and $\sum_{\nu=1}^\infty
\frac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}\asymp
h^{-(2a+1)}$.
Then we have for any $z\in\mathbb{I}$,
\[
\frac{\sqrt{n}(\widehat{\beta}_{n,\lambda}(z)-\beta
_0(z)+(W_\lambda\beta_0)(z))}{
\sqrt{\sum_{\nu=1}^\infty
\sklfrac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}}}
\stackrel{d} {\longrightarrow}N(0,1).
\]
In addition, if $\sqrt{n}(W_\lambda\beta
_0)(z) /\sqrt{\sum_{\nu
=1}^\infty
\frac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}}=o(1)$,
then
\[
\frac{\sqrt{n}(\widehat{\beta}_{n,\lambda}(z)-\beta_0(z))}{
\sqrt{\sum_{\nu=1}^\infty
\sfrac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}}}
\stackrel{d} {\longrightarrow}N(0,1).
\]
\end{Corollary}
Corollary~\ref{localasymp} applies\vspace*{1pt} to any point $z\in\mathbb{I}$ satisfying
$\sum_{\nu=1}^\infty
\frac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}\asymp
h^{-(2a+1)}$.
Validity of this condition is discussed in Section~S.14
of the Supplementary Material~\cite{SCFDA}.
The condition $\sqrt{n}(W_\lambda\beta
_0)(z)/\sqrt{\sum_{\nu
=1}^\infty
\frac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho_\nu)^2}}=o(1)$
holds if
$nh^{4k}=o(1)$ and the true slope function $\beta_0=\sum_\nu b_\nu
\varphi_\nu$ satisfies the condition (\textbf{U}):\break $\sum_\nu b_\nu
^2\rho_\nu^2<\infty$. To see this,
observe that
\begin{eqnarray*}
\bigl\llvert(W_\lambda\beta_0) (z)\bigr\rrvert
&=&\biggl\llvert\sum_\nu b_\nu
\frac{\lambda\rho_\nu}{1+\lambda\rho_\nu
}\varphi_\nu(z)\biggr\rrvert
\\
&\le& C_\varphi\lambda\sum_{\nu}\llvert
b_\nu\rrvert\frac{\rho_\nu\nu
^a}{1+\lambda\rho_\nu}
\\
&\le& C_\varphi\lambda\sqrt{\sum_\nu
b_\nu^2\rho_\nu^2} \sqrt{\sum
_\nu\frac{\nu^{2a}}{(1+\lambda\rho_\nu)^2}},
\end{eqnarray*}
where the last term is of the order $O(\lambda h^{-(2a+1)/2})$. Hence,
it leads to
\[
\sqrt{n}(W_\lambda\beta_0) (z)\bigg/\sqrt{\sum
_{\nu=1}^\infty\frac{\llvert \varphi_\nu(z)\rrvert ^2}{(1+\lambda\rho
_\nu)^2}}\asymp\sqrt
{nh^{2a+1}}(W_\lambda\beta_0) (z)=o(1).
\]
\begin{Remark}[(Convergence rate)]
Corollary~\ref{localasymp} derives the convergence
rate of the local estimate $\widehat{\beta}_{n,\lambda}(z)$ as
$\sqrt{nh^{2a+1}}$.
The factor $a$ (defined in Assumption~\ref{A3}) generically reflects
the impact of the covariance operator on the convergence rate. For
example, Proposition~\ref{validA3} shows that $a=r+1$ with $r$ being
the order of the covariance function under the pseudo SY condition. The
above observation coincides with the arguments in \cite{HH07}, that
the covariance effect in general influences the (global) rate
expressions. When the eigenfunctions are uniformly bounded, that is,
$a=0$, the above rate becomes $\sqrt{nh}$, which is exactly the rate
derived in the general nonparametric regression setup; see Theorem 3.5
in \cite{SC13}.
\end{Remark}
\begin{Remark}
(Undersmoothing) Assumption~\ref{A3} implies that $\beta_0\in
H^m(\mathbb{I})$ has the property that $\sum_\nu b_\nu^2\rho_\nu
<\infty$. However, the condition (\textbf{U}) imposes a faster decay
rate on the generalized Fourier coefficients $b_\nu$, which in turn
requires more smoothness of $\beta_0$. Since we still employ the $m$th
order penalty in~(\ref{estimation}), condition~(\textbf{U}) can be
treated as a type of undersmoothing condition. More generally, similar
conditions will be implicitly imposed in the inference procedures to be
presented later.
\end{Remark}
\section{Confidence/prediction interval}\label{secasympCI}
In this section, we consider two inter-connected inference procedures:
(i) confidence interval for the conditional mean and (ii) prediction
interval for a future response.
\subsection{Confidence interval for conditional mean}\label{secCI}
For any (nonrandom) $x_0\in L^2(\mathbb{I})$,
we construct a confidence interval $\mu_0(x_0)=E\{Y\mid X=x_0\}$ by
centering around the plug-in estimate $\widehat{Y}_0\equiv F(\widehat
{\alpha}_{n,\lambda}+\int_0^1 x_0(t)\widehat{\beta}_{n,\lambda}(t)\,dt)$.
Define $\mu_0'(x_0)=\dot{F}(\alpha_0+\int_0^1x_0(t)\beta_0(t)\,dt)$,
$\sigma_n^2=E\{B(X)\}^{-1}+\sum_{\nu=1}^\infty\frac{\llvert x_\nu
^0\rrvert ^2}{(1+\lambda\rho_\nu)^2}$,
where $x_\nu^0=\int_0^1 x_0(t)\varphi_\nu(t)\,dt$.
\begin{Theorem}[(Confidence interval construction)]\label{CImean}
Let Assumptions~\ref{A1} through
\ref{A5} be satisfied for the true parameter
$\theta_0=(\alpha_0,\beta_0)$,
and $\mu_0'(x_0)\neq0$. Furthermore,\vspace*{1pt} assume (\ref{amomconv}) and
$E\{\exp(s\llvert \epsilon\rrvert )\}<\infty$ for some $s>0$.
If $h=o(1)$, $\log(h^{-1})=O(\log{n})$, $nh^{2a+1}(\log{n})^{-4}
\rightarrow\infty$, $na_n^2=o(1)$ and $\llVert R_{x_0}\rrVert\asymp
\sigma_n$,
then as $n\rightarrow\infty$,
\begin{eqnarray*}
&&\frac{\sqrt{n}}{\sigma_n} \biggl(\widehat{\alpha}_{n,\lambda
}+\int
_0^1 x_0(t)\widehat{
\beta}_{n,\lambda}(t)\,dt -\alpha_0-\int_0^1x_0(t)
\beta_0(t)\,dt
\\
&&\hspace*{137pt}{} -\int_0^1
x_0(t) (W_\lambda\beta_0) (t)\,dt \biggr)
\\[-1pt]
&&\qquad\stackrel{d} {\longrightarrow}N(0,1).
\end{eqnarray*}
Furthermore, if $\beta_0=\sum_\nu b_\nu\varphi_\nu$
with $\sum_\nu b_\nu^2\rho_\nu^2<\infty$ and $nh^{4k}=o(1)$, then
$\frac{\sqrt{n}}{\sigma_n}\int_0^1 x_0(t)(W_\lambda\beta
_0)(t)\,dt=o(1)$ so that
we have
\begin{equation}
\label{CIlimitdistr} \frac{\sqrt{n}}{\sigma_n\mu_0'(x_0)} \bigl
(\widehat Y_0-\mu
_0(x_0) \bigr)\stackrel{d} {\longrightarrow}N(0,1).
\end{equation}
Hence the $100(1-\widetilde{\alpha})\%$ confidence interval for $\mu
_0(x_0)$ is
\begin{equation}
\label{CIconditionalmean} \bigl[\widehat Y_0 \pm n^{-1/2}z_{\widetilde
{\alpha}/2}
\sigma_n \widehat\mu_0'(x_0)
\bigr],
\end{equation}
where $z_{\widetilde{\alpha}/2}$ is the $(1-\widetilde{\alpha
}/2)$-quantile of $N(0,1)$ and $\widehat\mu_0'(x_0)\equiv\dot
{F}(\widehat\alpha_{n, \lambda}+\break \int_0^1x_0(t)\widehat\beta
_{n,\lambda}(t)\,dt)$.
\end{Theorem}
In\vspace*{1pt} the Gaussian model (Example~\ref{exa1pss}) with $B(X)\equiv1$,
if $X$ is Brownian motion with $C(s,t)=\min\{s,t\}$,
then $\sigma_n^2$ has an explicit form with $\rho_\nu\approx(2\pi
\nu)^{2(m+1)}$
and $\varphi_\nu$ solved by (\ref{bvpexample}). As for Examples
\ref{exa2logit} and~\ref{exa2exponentialfamily}, one can obtain
$\sigma_n^2$ by following the approach outlined in Section~S.5
(Supplementary Material \cite{SCFDA}).
A direct byproduct of Theorem~\ref{CImean} is the prediction rate
$\sigma_n/\sqrt{n}$. Proposition~\ref{proppredrate} further
characterizes this rate in various situations. Suppose that $\llvert
x_\nu
^0\rrvert \asymp\nu^{a-d}$ for some constant $d$.
A larger $d$ usually yields a smoother function $x_0$.\vadjust{\goodbreak}
\begin{Proposition}\label{proppredrate}
The prediction rate in Theorem~\ref{CImean} satisfies
\[
\sigma_n/\sqrt{n}= \cases{ n^{-1/2}, &\quad if $d-a>1/2$,
\vspace*{3pt}\cr
n^{-1/2}\bigl(\log(1/h)\bigr)^{1/2}, &\quad if $d-a=1/2$,
\vspace*{3pt}\cr
n^{-1/2}h^{d-a-1/2}, &\quad if $d-a<1/2$.}
\]
In\vspace*{1pt} particular, if $d-a<1/2$ and $h=h^{\ast\ast\ast}\asymp n^{-\afrac
{1}{2(m+a)}}$, then
$\sigma_n/\sqrt{n}=n^{-\vafrac{d+m-1/2}{2(m+a)}}$. Furthermore, if $k=m+a$
as in the setting of Proposition~\ref{validA3}, then $\sigma_n/\sqrt
{n}$ is minimax optimal when $h=h^{\ast\ast\ast}$.
\end{Proposition}
Proposition~\ref{proppredrate} states that when $d-a>1/2$, that is,
the process $x_0$ is sufficiently smooth, then the prediction can be
conducted in
terms of root-$n$ rate regardless of the choice of $h$. This result
coincides with \cite{CH06}
in the special FPCA setting. Moreover, when $d-a<1/2$ and $h=h^{\ast
\ast\ast}$, the rate becomes optimal. Again, this is consistent with
\cite{CH06} in the setting
that the true slope function belongs to a Sobolev rectangle.
Interestingly, it can be checked that $h=h^{\ast\ast\ast}$
satisfies the rate conditions in Theorem~\ref{CImean}
if $a=1$, $k=m+1$ and $m>1+\sqrt{3}/2$; see the discussions below
Theorem~\ref{UBR}.
\subsection{Prediction interval for future response}
Following Theorem~\ref{CImean}, we can establish the prediction
interval for the future response $Y_0$ conditional on \mbox{$X=x_0$}. Write
$Y_0-\widehat Y_0=\xi_n+\epsilon_0$, where $\xi_n=\mu
_0(x_0)-\widehat Y_0$ and $\epsilon_0=\dot\ell_a(Y_0;\alpha_0+\int
_0^1x_0(t)\beta_0(t)\,dt)$. Since $\epsilon_0$ is independent of $\xi
_n$ depending on all the past data $\{Y_i,X_i\}_{i=1}^n$, we can easily
incorporate the additional randomness from $\epsilon_0$ into the
construction of the prediction interval. This leads to a nonvanishing
interval length as sample size increases. This is crucially different
from that of confidence interval.
Let $F_{\xi_n}$ and $F_{\epsilon_0}$ be the distribution functions of
$\xi_n$ and $\epsilon_0$, respectively.
Denote the distribution function of $\xi_n+\epsilon_0$ as $G\equiv
F_{\xi_n}*F_{\epsilon_0}$, and $(l_{\widetilde{\alpha}},
u_{\widetilde{\alpha}})$ as its $(\widetilde{\alpha}/2)$th and
$(1-\widetilde{\alpha}/2)$th quantiles, respectively. Then the
$100(1-\widetilde{\alpha})\%$ prediction interval for $Y_0$ is given as
\[
[\widehat Y_0+l_{\widetilde{\alpha}}, \widehat Y_0+u_{\widetilde{\alpha}}
].
\]
Theorem~\ref{CImean} directly implies that $\xi_n\stackrel{a}{\sim
}N(0, (n^{-1/2}\sigma_n\widehat\mu_0'(x_0))^2)$,
where $\stackrel{a}{\sim}$ means \textit{approximately distributed}. If
we further assume that $\epsilon_0\sim N(0, B^{-1}(x_0))$ [see
Assumption~\ref{A1}(c)], that is, $B(x_0)$ is the reciprocal
error variance for the $L^2$ loss, the above general formula reduces to
\begin{equation}
\label{PIGaussian} \Bigl[\widehat Y_0\pm z_{\widetilde{\alpha}/2}\sqrt
{B(x_0)+\bigl(n^{-1/2}\sigma_n\widehat
\mu_0'(x_0)\bigr)^2} \Bigr].
\end{equation}
The unknown quantities in (\ref{CIconditionalmean}) and (\ref
{PIGaussian}) can be estimated by plug-in approach.
\section{Hypothesis testing}\label{sechypothesistesting}
We consider two types of testing for the generalized functional linear
models: (i) testing the \textit{functional contrast} defined as $\int_0^1
w(t)\beta(t)\,dt$ for some given weight function $w(\cdot)$, for
example, $w=X$ and (ii) testing the intercept value and the global
behavior of the slope function, for example, $\alpha=0$ and $\beta$
is a linear function.
\subsection{Testing functional contrast}
In practice, it is often of interest to test the \emph{functional
contrast}. For example, we may test single frequency or frequency
contrast of the slope function; see Examples~\ref{egCT1} and~\ref
{egCT2}. In general, we test $H_0^{CT}\dvtx \int_0^1 w(t)\beta(t)\,dt=c$
for some known $w(\cdot)$ and $c$.
Consider the following test statistic:
\begin{equation}
\label{testcontr} CT_{n,\lambda}=\frac{\sqrt{n}(\int_0^1w(t)\widehat
{\beta
}_{n,\lambda}(t)\,dt-c)}{
\sqrt{\sum_{\nu=1}^\infty\sklfrac{w_\nu^2}{(1+\lambda\rho_\nu)^2}}},
\end{equation}
where $w_\nu=\int_0^1 w(t)\varphi_\nu(t)\,dt$. Recall that $(\varphi
_\nu,\rho_\nu)$ is the eigensystem satisfying Assumption~\ref{A3}.
Let $w\in L^2(\mathbb{I})$ and $\tau(w)\in H^m(\mathbb{I})$ be such
that $\langle\tau(w),\beta\rangle_1=\int_0^1w(t)\beta(t)\,dt$,
for any $\beta\in H^m(\mathbb{I})$. We can verify that $\tau(w)=\sum
_{\nu=1}^\infty\frac{w_\nu}{1+\lambda\rho_\nu}\varphi_\nu$.
Then, under $H_0^{CT}$, $CT_{n,\lambda}$ can be rewritten as
\begin{equation}
\frac{\sqrt{n}\langle\tau(w), (\widehat\beta_{n,\lambda}-\beta
)\rangle_1}{\llVert\tau(w)\rrVert_1}.\label{cttest}
\end{equation}
It follows from Theorem~\ref{CImean} that
(\ref{cttest}) converges weakly to a standard normal distribution.
This is summarized in the following theorem.
Define $M_a=\sum_{\nu=1}^\infty\frac{w_\nu^2}{(1+\lambda\rho_\nu
)^a}$ for $a=1,2$.
\begin{Theorem}[(Functional contrast testing)]\label{limitCT}
Suppose that Assumptions~\ref{A1}
through~\ref{A5} hold. Furthermore,
let $\beta_0=\sum_\nu b_\nu\varphi_\nu$ with
$\sum_\nu b_\nu^2\rho_\nu^2<\infty$, and
assume~(\ref{amomconv}), $E\{\exp(s\llvert \epsilon\rrvert )\}<\infty
$ for some
$s>0$, and as $n\rightarrow\infty$,
$h=o(1)$, $\log(h^{-1})=O(\log{n})$,
$nh^{2a+1}(\log(1/h))^{-4}\rightarrow\infty$,
$M_1\asymp M_2$, $na_n^2=o(1)$, $nh^{4k}=o(1)$.
Then, under $H_0^{CT}$, we have
$CT_{n,\lambda}\stackrel{d}{\longrightarrow}N(0,1)$ as $n\rightarrow
\infty$.
\end{Theorem}
\begin{Example}[(Testing single frequency)]\label{egCT1}
Suppose that the slope function has an expansion $\beta=\sum_{\nu
=1}^\infty b_\nu\varphi_\nu$,
and we want to test whether $b_{\nu^\ast}=0$ for some \mbox{$\nu^\ast\ge
1$}. In other words,
we are interested in knowing whether the \mbox{$\nu^\ast$-}level frequency
of $\beta$ vanishes.
Let $w(t)=(C\varphi_{\nu^\ast})(t)$. Then it is easy to see that
$\int_0^1w(t)\*\beta(t)\,dt=b_{\nu^\ast}$. That is, the problem reduces
to testing
$H_0^{CT}\dvtx \int_0^1w(t)\beta(t)\,dt=0$.
It can be shown directly that $M_a=(1+\lambda\rho_{\nu^\ast
})^{-a}\asymp1$ for $a=1,2$.
If $r=0$ (see Proposition~\ref{validA4} for validity),
then it can be shown that when $m>(3+\sqrt{5})/4\approx1.309$, the
rate conditions in Theorem~\ref{limitCT}
are satisfied for $h=h^\ast$. This means that $H_0^{CT}$ is rejected
at level $0.05$
if $\llvert CT_{n,\lambda}\rrvert >1.96$.
\end{Example}
\begin{Example}[(Testing frequency contrast)]\label{egCT2}
Following Example~\ref{egCT1}, we now test whether $\sum_{\nu
=1}^\infty\mathfrak{c}_\nu b_\nu=0$,
for some real sequence $\mathfrak{c}_\nu$ satisfying $0<\inf_{\nu
\ge1}\llvert \mathfrak{c}_\nu\rrvert \le
\sup_{\nu\ge1}\llvert \mathfrak{c}_\nu\rrvert <\infty$.
Suppose that the covariance function $C(\cdot,\cdot)$ satisfies the
conditions in Proposition
\ref{validA3} with order $r>0$.
It follows from\vspace*{1pt} Proposition~\ref{validA3} and its proof that the
eigenfunction $\varphi_\nu$
can be managed so that $\llVert\varphi^{(2m)}_\nu\rrVert
_{\sup}\le C_\varphi\nu^{2m+r+1}$
and $\llVert C\varphi_\nu\rrVert_{\sup}\le\rho_\nu
^{-1}\llVert\varphi^{(2m)}_\nu\rrVert_{\sup}\asymp\nu
^{-(r+1)}$, for all $\nu\ge1$.
So the function $w(t)=\sum_{\nu=1}^\infty\mathfrak{c}_\nu(C\varphi
_\nu)(t)$ is well defined since
the series is absolutely convergent on $[0,1]$. It is easy to see that
$\int_0^1w(t)\beta(t)\,dt=\sum_\nu\mathfrak{c}_\nu b_\nu$
and $\mathfrak{c}_\nu=\int_0^1w(t)\varphi_\nu(t)\,dt=w_\nu$.
So the problem reduces to testing $H_0^{CT}\dvtx \int_0^1w(t)\beta(t)\,dt=0$.
For $a=1,2$
\[
M_a=\sum_{\nu=1}^\infty
\frac{w_\nu^2}{(1+\lambda\rho_\nu
)^a}=\sum_{\nu=1}^\infty
\frac{\mathfrak{c}_\nu^2}{(1+\lambda\rho
_\nu)^a} \asymp\sum_{\nu=1}^\infty
\frac{1}{(1+\lambda\rho_\nu)^a}\asymp h^{-1}.
\]
It can be shown that when $m>(3+\sqrt{5})$, the rate conditions in
Theorem~\ref{limitCT}
are satisfied for $h=h^\ast$. We reject $H_0^{CT}$ at level 0.05 if
$\llvert CT_{n,\lambda}\rrvert >1.96$.
\end{Example}
\subsection{Likelihood ratio testing}\label{secPLRT}
Consider the following simple hypothesis:
\begin{equation}
\label{simplehypothesis} H_0\dvtx \theta=\theta_0\quad\mbox{versus}
\quad H_1\dvtx \theta\in\mathcal H-\{\theta_0\},
\end{equation}
where $\theta_0\in\mathcal{H}$. The penalized likelihood ratio test
statistic is defined as
\begin{equation}
\mathrm{PLRT}=\ell_{n,\lambda}(\theta_0)-\ell_{n,\lambda}(\widehat{
\theta}_{n,\lambda}).
\end{equation}
Recall that $\widehat{\theta}_{n,\lambda}$ is the maximizer of $\ell
_{n,\lambda}(\theta)$ over $\mathcal{H}$. The proposed likelihood
ratio testing also applies to the composite hypothesis; that is,
$\theta$ belongs to a certain class. See Remark~\ref{remcomp} for
more details.
Theorem~\ref{FDAPLRT} below derives the null limiting distribution of
$\mathrm{PLRT}_{n,\lambda}$.
\begin{Theorem}[(Likelihood ratio testing)]\label{FDAPLRT}
Suppose that $H_0$ holds, and Assumptions
\ref{A1} through~\ref{A5} are satisfied for
the hypothesized value $\theta_0$. Let $h$
satisfy the following rate conditions: as $n\rightarrow\infty$,
$nh^{2k+1}=O(1)$, $nh\rightarrow\infty$, $n^{1/2}a_n=o(1)$,
$nr_n^3=o(1)$, $n^{1/2}h^{-(a+\sfrac{1}{2}+\vafrac
{2k-2a-1}{4m})}r_n^2(\log{n})^2\times\break (\log\log{n})^{1/2}=o(1)$
and $n^{1/2}h^{-(2a+1+\vafrac{2k-2a-1}{4m})}\times
r_n^3(\log{n})^3\times\break (\log\log{n})^{1/2}=o(1)$.
Furthermore, there exists a constant $M_4>0$ s.t.\break
$E\{\epsilon^4\mid X\}\le M_4$, a.s.
Then as $n\rightarrow\infty$,
\begin{equation}
\label{PLRTlim} -(2u_n)^{-1/2} \bigl(2n\sigma^2
\cdot \mathrm{PLRT}+u_n+n\sigma^2\llVert W_\lambda
\beta_0\rrVert_1^2 \bigr)\stackrel{d} {
\longrightarrow}N(0,1),
\end{equation}
where $u_n=h^{-1}\sigma_1^4/\sigma_2^2$, $\sigma^2=\sigma
_1^2/\sigma_2^2$ and $\sigma_l^2=h\sum_\nu(1+\lambda\rho_\nu
)^{-l}$ for $l=1,2$.
\end{Theorem}
By carefully examining the proof of Theorem~\ref{FDAPLRT},
it can be shown that $n\llVert W_\lambda\beta_0\rrVert
_1^2=o(n\lambda)=o(u_n)$. Therefore, $-2n\sigma^2\cdot \mathrm{PLRT}$ is
asymptotically\break $N(u_n, 2u_n)$ which is nearly $\chi_{u_n}^2$ as
$n\rightarrow\infty$. Hence we claim the null limit distribution as
being approximately $\chi_{u_n}^2$, denoted as
\begin{equation}
\label{wilks} -2n\sigma^2\cdot \mathrm{PLRT}\stackrel{a} {\sim}
\chi_{u_n}^2,
\end{equation}
where $\stackrel{a}{\sim}$ means \textit{approximately distributed}; see
\cite{FZZ01}.
If $C$ satisfies the conditions of Proposition~\ref{validA3} with
order $r\ge0$,
then $\rho_\nu\approx(c\nu)^{2k}$
(see the comments below Proposition~\ref{validA3}),
where $k=m+r+1$ and $c>0$ is constant.
It is easy to see that $\sigma_l^2\approx c^{-1}\int_0^\infty
(1+x^{2k})^{-l}\,dx$ for $l=1,2$.
In Example~\ref{exa1pss}, since the covariance function is free of
the model parameters $\alpha_0,\beta_0$, we can see that $c$ is also
free of the model parameters. In particular, when $B(X)\equiv1$, $m=2$
and $X(t)$ is Brownian motion with $C(s,t)=\min\{s,t\}$ and $r=0$, we
have $k=3$ and $c=\pi$ [by solving~(\ref{bvpexample})]. This yields
$\sigma_l^2\approx0.2876697$, $0.2662496$ for $l=1,2$, respectively.
In the end, we obtain $\sigma^2\approx1.080451$ and $u_n\approx
0.3108129/h$ in (\ref{wilks}). As seen above, the null limiting
distribution has the nice property that it is free of the unknown model
parameters, that is, so-called Wilks phenomenon \cite{W38,FZZ01}.
Hence we unveil a new version of Wilks phenomenon that applies to the
functional data. This Wilks type of result enables us to simulate the
null limit distribution directly without resorting to bootstrap or
other resampling methods.
The quantities $\sigma_1^2,\sigma_2^2$
in Theorem~\ref{FDAPLRT} depend on the population eigenpairs.
However, it is possible to replace these quantities by suitable estimators
so that the results become more applicable. In
Section~S.18 (Supplementary Material \cite{SCFDA}), we discuss the validity of this ``plug-in''
approach for Theorems~\ref{CImean},~\ref{limitCT} and~\ref{FDAPLRT}.
\begin{Remark}[(Composite hypothesis)]\label{remcomp}
By examining the proof of Theorem~\ref{FDAPLRT}, we find that the
null limiting distribution derived therein
remains the same even when the hypothesized value $\theta_0$ is
unknown. An important consequence is that the
proposed likelihood ratio approach can also be used to test a composite
hypothesis such as $H_0\dvtx \alpha=\alpha_0$ and $\beta\in
\mathcal{P}_j$, where $\mathcal{P}_j$ represents the class of the
$j$th-order polynomials. Under $H_0$, $\beta
$ is of the form $\beta(t)=\sum_{l=0}^j b_l t^l$
for some unknown vector $\mathbf{b}=(b_0,b_1,\ldots,b_j)^T$. In this
case, the slope function and intercept
can be estimated through the following ``parametric'' optimization:
\begin{eqnarray}\label{H0mle}
\bigl(\widehat{\alpha}^0,\widehat{
\mathbf{b}}^0\bigr) &=&\arg\max_{\alpha,b_0,\ldots,b_j\in\bbR}
n^{-1}\sum_{i=1}^n\ell
\Biggl(Y_i;\alpha+\sum_{l=0}^j
b_l\int_0^1 X_i(t)t^l\,dt
\Biggr)
\nonumber\\[-8pt]\\[-8pt]\nonumber
&&{} -(\lambda/2)\mathbf{b}^T D\mathbf{b},
\end{eqnarray}
where $D=[D_{l_1L^2}]_{l_1,l_2=0,\ldots,j}$ is a $(j+1)\times(j+1)$
matrix with $D_{l_1l_2}
=J(t^{l_1},t^{l_2})$.
The corresponding slope function estimate is $\widehat{\beta
}^0(t)=\sum_{l=0}^j\widehat{b}^0_l t^l$.
The test\vspace*{1pt} statistic for this composite hypothesis is defined as
$\mathrm{PLRT}=\ell_{n,\lambda}(\widehat{\alpha}^0,
\widehat{\beta}^0)
-\ell_{n,\lambda}(\widehat{\alpha}_{n,\lambda},\widehat{\beta
}_{n,\lambda})$. Let $\theta_0=(\alpha_0,\beta_0)
$ be the unknown true model parameter under $H_0$, where $\beta_0$ can
be represented as $\sum_{l=0}^j b_l^0
t^l$. Hence, we can further decompose the above $\mathrm{PLRT}$ as
$\mathrm{PLRT}_1-\mathrm{PLRT}_2$, where $\mathrm{PLRT}_1=\ell_{n,\lambda}
(\theta_0)-\ell_{n,\lambda}(\widehat{\alpha}_{n,\lambda},\widehat
{\beta}_{n,\lambda})$,
$\mathrm{PLRT}_2=\ell_{n,\lambda}(\theta_0)-\ell_{n,\lambda}(\widehat
{\alpha}^0,\widehat{\beta}^0)$.
Note that $\mathrm{PLRT}_1$ is the test statistic for the simple hypothesis
$\theta=\theta_0$ versus $\theta\neq
\theta_0$, and $\mathrm{PLRT}_2$ for the parametric hypothesis $(\alpha,\mathbf{b})=(\alpha_0,
\mathbf{b}^0)$ versus $
(\alpha,\mathbf{b})\neq(\alpha_0,\mathbf{b}^0)$, where $\mathbf{b}^0=(b_0^0,\ldots,b_j^0)^T$.
Conventional\vspace*{1pt} theory on parametric likelihood ratio testing leads to
$-2n\cdot \mathrm{PLRT}_2=O_P(1)$. On the other hand, Theorem~\ref{FDAPLRT}
shows that $-2n\sigma^2\cdot
\mathrm{PLRT}_1\stackrel{a}{\sim}\chi_{u_n}^2$. Therefore, we conclude that
the null limit distribution for testing the
composite hypothesis also follows $\chi^2_{u_n}$.
\end{Remark}
In the end of this section, we show that the proposed PLRT is optimal
in the minimax sense \cite{I93} when $h=h^{**}$. To derive the
minimax rate of testing (also called as minimum separation rate), we consider
a local alternative written as $H_{1n}\dvtx \theta=\theta_{n0}$, where
the alternative value is assumed to deviate
from the null value by an amount of $\theta_n$, that is, $\theta
_{n0}=\theta_0+\theta_n$. For simplicity, we
assume $\theta_0=0$, and thus $\theta_{n0}=\theta_n$. Define the
alternative value set $\theta_n\in\Theta_b
\equiv\{(\alpha,\beta)\in\mathcal{H}\dvtx \llvert \alpha\rrvert \le b,
\llVert
\beta\rrVert_{L^2}\le b, J(\beta,\beta)\le b\}$ for some
fixed constant $b>0$.
\begin{Theorem}\label{mrt}
Let Assumptions~\ref{A1}--\ref{A5} be satisfied uniformly under
$\theta=\theta_{n0}\in\Theta_b$.
Let $h$ satisfy $nh^{3/2}\rightarrow\infty$, as $n\rightarrow\infty$,
and also the rate conditions specified in Theorem~\ref{FDAPLRT}.
Furthermore, $\inf_{y\in\mathcal{Y},a\in\bbR}(-\ddot{\ell}_a(y;a))>0$,
and there is a constant $M_4>0$ s.t. for $\theta_n=(\alpha_n,\beta
_n)\in\Theta_b$,
$\epsilon_n\equiv\dot{\ell}_a(Y;\alpha_{n}+\int_0^1 X(t)\beta_{n}(t)\,dt)$
satisfies $E\{\epsilon^4_n\mid X\}\le M_4$, a.s.
Then for any $\varepsilon>0$,
there exist positive constants $N_\varepsilon$ and $C_\varepsilon$
s.t. when $n\ge N_\varepsilon$,
\[
\inf_{\theta_n\in\Theta_b\dvtx \llVert\theta_n\rrVert\ge
C_\varepsilon\eta_n} P_{\theta_n}(\mbox{reject }H_0)
\ge1-\varepsilon,
\]
where $\eta_n\asymp\sqrt{(nh^{1/2})^{-1}+\lambda}$.
\end{Theorem}
The model assumption $\inf_{y\in\mathcal{Y},a\in\bbR
}(-\ddot{\ell}_a(y;a))>0$ trivially holds for Gaussian regression and
exponential family considered in Examples~\ref{exa1pss}
and~\ref{exa2exponentialfamily}. As for the logistic model with
$L^2$ bounded $X$,
this condition can be replaced by $\inf_{y\in\mathcal{Y},a\in
\mathcal{I}}(-\ddot{\ell}_a(y;a))
>0$ under which the same conclusion as in Theorem~\ref{mrt} holds,
where $\mathcal{I}$ is some bounded open interval including the range
of $\langle R_X,\theta_{n0}\rangle$
for every $\theta_{n0}\in\Theta_b$.
Theorem~\ref{mrt} states that the PLRT is able to detect any local
alternative with separation rate no faster than $\eta_n$. In
particular, the minimum separation rate, that is, $n^{-2k/(4k+1)}$, is
achieved when $h=h^{\ast\ast}$. Note that $h^{\ast\ast}$ satisfies
the rate conditions required by Theorem~\ref{mrt}. For example, when
$k=m+r+1$, $a=r+1$, $r=0$
(see the discussions below Proposition~\ref{rateconv}), we can verify
this fact for $m>(7+\sqrt{33})/8\approx
1.593$ by direct calculations. In the specific $\ell_2$ regression,
Corollary~4.6 of \cite{HMV13} proves that the above minimax rate, that
is, $n^{-2k/(4k+1)}$, is optimal but under the perfect alignment
condition. Therefore, we prove that the proposed PLRT can achieve the
minimax optimality under more general settings.
The likelihood ratio testing procedure developed in this section
requires prior knowledge on the smoothness of the true slope function
and covariance kernel function, which might not be available in
practice. This motivates us to propose two adaptive testing procedures
in the next section.
\section{Adaptive testing construction}\label{secadaptivePLRT}
In this section, we develop two adaptive testing procedures for
$H_0\dvtx \beta=\beta_0$ without knowing
$m$ and $r$, that is, the smoothness of the true slope function and
covariance kernel function. One works for Gaussian errors, and another
works for sub-Gaussian errors. The test statistics for both cases are
maximizers over a sequence of standardized PLRTs. We derive the null
limit distribution as an extreme value distribution using Stein's
method \cite{CCK13,Stein86}. Their minimax properties will also be
carefully studied. To the best of our knowledge, our adaptive testing
procedures are the first ones developed in the roughness regularization
framework, which forms an interesting complement to those based on FPCA
techniques \cite{HMV13,L13}.
In this section, we focus on the $\ell_2$ regression with two types of
error: Gaussian error (Section~\ref{testgau}) and sub-Gaussian error
(Section~\ref{testsubg}). For simplicity, we assume $\beta_0=0$,
$\alpha=0$, and the errors to be of unit standard deviations. In
addition, we assume that the covariate process $X(t)$ has zero mean and
is independent of the error term. We remark that it is possible to
extend our results in this section to the general loss functions, but
with extremely tedious technical arguments.
Our test statistic is built upon a modified estimator $\widetilde
{\beta}_{n,\lambda}$ that is constructed in the following three
steps. The first step is to find a sequence\vspace*{1pt} of empirical eigenfunctions
$\widehat{\varphi}_\nu$ that satisfy $\widehat{V}(\widehat{\varphi
}_\nu,\widehat{\varphi}_\mu)=\delta_{\nu\mu}$ for all $\nu,\mu
\ge1$, where $\widehat{V}(\beta,\widetilde{\beta})=\int_0^1\int
_0^1\widehat{C}(s,t)\beta(s)\widetilde{\beta}(t)\,ds\,dt$ and $\widehat
{C}(s,t)=n^{-1}\sum_{i=1}^nX_i(s)X_i(t)$. We offer two methods for
finding $\widehat{\varphi}_\nu$. The\vspace*{1pt} first method conducts a
spectral decomposition, $\widehat{C}(s,t)=\sum_{\nu=1}^\infty
\widehat{\zeta}_\nu\widehat{\psi}_\nu(s)\widehat{\psi}_\nu(t)$,
with some nonincreasing positive sequence $\widehat{\zeta}_\nu$
and orthonormal functions $\widehat{\psi}_\nu$ in the usual
$L^2$-norm. Construct $\widehat{\varphi}_\nu=\widehat{\psi}_\nu
/\sqrt{\widehat{\zeta}_\nu}$. This method is easy to implement, but
implicitly assumes the perfect alignment condition. Our second method
is more generally applicable, but requires more tedious implementation.
Specifically, we apply similar construction techniques as in
Section~S.5 (Supplementary Material \cite{SCFDA}) by using the sample versions of
$\widetilde
{K}$, $C$ and $T$ therein. In particular, we choose $m=1, 2$ such that
the true slope function is more possible to be covered.
The\vspace*{1pt} second step is to define a data-dependent parameter space. Note
that $H^m(\mathbb I)$ can be alternatively defined as $\{\sum_{\nu
=1}^\infty b_\nu\varphi_\nu\dvtx \sum_{\nu=1}^\infty b_\nu^2\nu
^{2k}<\infty\}$, where $k$ depends on $m$ in an implicit manner. An
approximate parameter space is $\mathcal{B}_k= \{\sum_{\nu
=1}^\infty b_\nu\widehat{\varphi}_\nu\dvtx \sum_{\nu=1}^\infty b_\nu
^2\nu^{2k}<\infty\}$.
The consistency of the sample eigenfunctions implies that $\mathcal{B}_k$
is a reasonable approximation of $H^m(\mathbb I)$; see \cite{HMW06}.
The data-dependent parameter space is thus defined as
\[
\mathcal{B}_{kn}\equiv\Biggl\{\sum_{\nu=1}^n
b_\nu\widehat{\varphi}_\nu\bigg| b_1,
\ldots,b_n\in\bbR\Biggr\}.
\]
In $\mathcal{B}_{kn}$, we can actually use the first $K_n\rightarrow
\infty$ ($K_n\ll n$) eigenfunctions as the basis. However, such a
general choice would give rise to unnecessary tuning of $K_n$ in practice.
In the last step, we obtain the desirable estimator as $\widetilde
{\beta}_{n,\lambda}=\break \arg\sup_{\beta\in\mathcal{B}_{kn}}\ell
_{n,\lambda}(\beta)$, where
\begin{equation}
\label{gauplik1} \ell_{n,\lambda}(\beta)=-\frac{1}{n}\sum
_{i=1}^n \Biggl(Y_i-\sum
_{\nu=1}^{n} b_\nu\omega_{i\nu}
\Biggr)^2\bigg/2-(\lambda/2)\sum_{\nu=1}^{n}
b_\nu^2\nu^{2k},
\end{equation}
and $\omega_{i\nu}=\int_0^1X_i(t)\widehat{\varphi}_\nu(t)\,dt$ for
$i=1,\ldots,n$ and $\nu\ge1$. The smoothing parameter $\lambda$
depends on both $n$ and $k$, denoted as $\lambda_k$. In particular, we
choose $\lambda_k$ as $c_0^{2k}n^{-4k/(4k+1)}(\log\log
{n})^{2k/(4k+1)}$ for some constant $c_0>0$ irrelevant to $k$. As will
be seen in later theorems, this choice yields the minimax optimality of
the adaptive testing. Define $Y=(Y_1,\ldots,Y_n)^T$,
$b=(b_1,\ldots,b_n)^T$, $\Lambda_k=\operatorname{diag}(1^{2k},
2^{2k}, \ldots, n^{2k})$,
$\Omega_i=(\omega_{i1},\ldots,\omega_{in})$
and $\Omega=(\Omega_1^T,\ldots,\Omega_n^T)^T$.
Hence we can rewrite $-2n\ell_{n,\lambda}(\beta)$ as
\[
(Y-\Omega b)^T(Y-\Omega b)+n\lambda_k b^T
\Lambda_k b,
\]
whose minimizer [equivalently, the maximizer of $\ell_{n,\lambda
}(\beta)$] is $\widehat{b}_{n,k}=(\Omega^T\Omega+n\lambda_k\Lambda
_k)^{-1}\Omega^TY$.
Note that $\Omega^T\Omega=nI_n$ by the following analysis: for any
$\nu,\mu\ge1$,
\begin{eqnarray*}
\sum_{i=1}^n\omega_{i\nu}
\omega_{i\mu}&=&\sum_{i=1}^n\int
_0^1X_i(t)\widehat{
\varphi}_\nu(t)\,dt\int_0^1X_i(s)
\widehat{\varphi}_\mu(s)\,ds
\\
&=&\int_0^1\int_0^1
\sum_{i=1}^nX_i(s)X_i(t)
\widehat{\varphi}_\nu(s)\widehat{\varphi}_\mu(t)\,ds\,dt
\\
&=&n\int_0^1\int_0^1
\widehat{C}(s,t)\widehat{\varphi}_\nu(s)\widehat{
\varphi}_\mu(t)\,ds\,dt=n\delta_{\nu\mu}.
\end{eqnarray*}
Therefore, $\widehat{b}_{n,k}=(nI_n+n\lambda_k\Lambda_k)^{-1}\Omega
^TY$ and $\widetilde\beta_{n,\lambda}=(\widehat\varphi_1,\ldots,\widehat
\varphi_n)\widehat b_{n,k}$.
In the above analysis, we implicitly assume $k$ to be known. However,
the value of $k$ is usually unavailable in practice. To resolve this
issue, we will conduct our testing procedure over a sequence of integer
$k$, that is, $\{1, 2,\ldots, k_n\}$, as will be seen in the next two
subsections. The full adaptivity of testing procedure is achieved when
we allow $k_n\rightarrow\infty$ so that the unknown $k$ can
eventually be captured by this sequence.
\subsection{Gaussian error}\label{testgau}
In this subsection, we denote the PLRT as $\mathrm{PLRT}_k\equiv\ell
_{n,\lambda}(0)-\ell_{n,\lambda}(\widetilde{\beta}_{n,\lambda})$
due to its\vspace*{1pt} dependence on $k$. By plugging in the above form of
$\widetilde\beta_{n,\lambda}$, we obtain
\begin{equation}
\mathrm{PLRT}_k=-\frac{1}{2n}Y^T\Omega(nI_n+n
\lambda_k\Lambda_k)^{-1}\Omega
^TY.\label{gauplrt1}
\end{equation}
We next derive a standardized version of $\mathrm{PLRT}_k$ under $H_0$. Define
$d_\nu(k)=1/(1+\lambda_k\rho_\nu(k))$, where $\rho_\nu(k)=\nu
^{2k}$, for any $\nu,k\ge1$. Under $H_0$, we have $Y=\epsilon
=(\epsilon_1,\ldots,\epsilon_n)^T$, and thus $-2n \mathrm{PLRT}_k=\sum_{\nu
=1}^{n}d_{\nu}(k)\eta_\nu^2$ for $\eta_1,\ldots,\eta_n\stackrel
{\mathrm{i.i.d.}}{\sim}N(0,1)$ by straightforward calculation. Hence, we have $E\{
-2n \mathrm{PLRT}_k\}=\sum_{\nu=1}^{n} d_\nu(k)$
and $\operatorname{Var}(-2n \mathrm{PLRT}_{k})=2\sum_{\nu=1}^{n} d_\nu^2(k)$.
The standardized version of $\mathrm{PLRT}_k$ can be written as
\begin{equation}
\label{deftauk} \tau_k=\frac{-2n\cdot \mathrm{PLRT}_k-\sum_{\nu=1}^{n}d_\nu(k)}{
(2\sum_{\nu=1}^{n}d_\nu(k)^2)^{1/2}}.
\end{equation}
Inspired by Theorem~\ref{limitCT}, $\tau_k$ is presumably of
standard normal distribution for any particular $k$. However, $k$ is
often unavailable in practice. As discussed previously, we shall
construct the adaptive testing based on a sequence of $\tau_k$ as
follows: (i)~define $AT_n^*=\max_{1\leq k\le k_n}\tau_k$, and (ii)
standardize $AT_n^\ast$ as
\[
AT_n=B_n\bigl(AT_n^*-B_n\bigr),
\]
where $B_n$ satisfies $2\pi B_n^2\exp(B_n^2)=k_n^2$; see \cite
{Hall79}. By Cram\'{e}r \cite{C46}, $B_n=\break \sqrt{2\log{k_n}}-\frac
{1}{2}(\log\log{k_n}+\log{4\pi})/\sqrt{2\log{k_n}}+O(1/\log
{k_n})\asymp\sqrt{2\log{k_n}}$
as $n$ becomes sufficiently large.
\begin{Theorem}\label{adapasym}
Suppose $k_n\asymp(\log{n})^{d_0}$, for some constant $d_0\in(0,1/2)$.
Then for any $\bar{\alpha}\in(0,1)$, we have under $H_0\dvtx \beta=0$,
\[
P(AT_n\le c_{\bar{\alpha}})\rightarrow1-\bar{\alpha}\qquad \mbox{as $n
\rightarrow\infty$},
\]
where $c_{\bar{\alpha}}=-\log(-\log(1-\bar{\alpha}))$.
\end{Theorem}
The proof of Theorem~\ref{adapasym} is mainly based on
Stein's leave-one-out method
\cite{Stein86} since under $H_0$, $\tau_k$ can be written as a sum of
independent random variables, that is, $\tau_k=\sum_{\nu=1}^n[d_\nu
(k)/s_{n, k}](\eta_\nu^2-1)$, where $s_{n,k}^2=2\sum_{\nu=1}^n
d_\nu(k)^2$.\vspace*{1pt}
In the end, we investigate the optimality of the proposed adaptive
testing procedure.
Consider the local alternative $H_{1n}\dvtx \beta\in\mathcal{B}_{k,1}$, where
\[
\mathcal{B}_{k,1}\equiv\Biggl\{\sum_{\nu=1}^\infty
b_\nu\widehat{\varphi}_\nu\dvtx \sum
_{\nu=1}^\infty b_\nu^2
\nu^{2k}\leq1 \Biggr\},
\]
for some fixed but \emph{unknown} integer $k\ge1$. For any real
sequence $\mathfrak{b}=\{b_\nu\}$ satisfying $\sum_{\nu=1}^\infty
b_\nu^2\nu^{2k}\leq1$, let $\beta_{\mathfrak{b}}=\sum_{\nu
=1}^\infty b_\nu\widehat{\varphi}_\nu$ be the alternative function
value, and let $P_{\mathfrak{b}}$ be the corresponding probability
measure. The following result shows that the adaptive test $AT_n$
achieves the optimal minimax rate (up to an logarithmic order), that
is, $\delta(n,k)\equiv n^{-2k/(4k+1)}(\log\log{n})^{k/(4k+1)}$, for
testing the hypothesis $H_0\dvtx \beta=0$,
with the alternative set being certain Sobolev ellipsoid $\mathcal
{B}_{k,1}$; see \cite{HMV13}.
Define $\llVert\mathfrak{b}\rrVert_{\ell^2}^2=\sum_{\nu
=1}^\infty b_\nu^2$ and $\llVert\mathfrak{b}\rrVert
_{k,\ell^2}^2=\sum_{\nu=1}^\infty b_\nu^2\rho_\nu(k)$.
\begin{Theorem}\label{minimaxadapt}
Suppose $k_n\asymp(\log{n})^{d_0}$, for some constant $d_0\in(0,1/2)$.
Then, for any $\varepsilon\in(0,1)$,
there exist positive constants $N_\varepsilon$ and $C_\varepsilon$
such that for any $n\ge N_\varepsilon$,
\[
\mathop{\inf_{\llVert\mathfrak{b}\rrVert_{\ell^2}\ge
C_\varepsilon\delta(n,k)}}_{\llVert\mathfrak{b}\rrVert
_{k,\ell^2}\le1}P_{\mathfrak{b}}(
\mbox{reject }H_0)\ge1-\varepsilon.
\]
\end{Theorem}
In Gaussian white noise models, Fan \cite{F96} and Fan and Lin \cite{FL98}
proposed an adaptive Neyman test based on multiple standardized test,
and derived the null limit distribution using the Darling--Erd\H{o}s
theorem. Theorems~\ref{adapasym}
and~\ref{minimaxadapt} can be viewed as extensions of such results to
functional data under Gaussian errors.\vadjust{\goodbreak} However, the
Darling--Erd\H{o}s theorem is no longer applicable in our setup due to
the difference in modeling and test construction. Instead, we employ
the Stein leave-one-out method. More interestingly, Stein's method can
even be applied to handle sub-Gaussian errors, as will be seen in
Section~\ref{testsubg}.
\subsection{Sub-Gaussian error}\label{testsubg}
In this subsection, we consider models with sub-Gaussian errors; that
is, there exists a positive constant $C_\epsilon$
such that $E\{\exp(t\epsilon)\}\le\exp(C_\epsilon t^2)$ for all
$t\in\bbR$. Further relaxation to the error term with finite fourth
moment is straightforward, but requires more stringent conditions on
the design.
For simplicity, we assume deterministic design, and suppose that
$X_i$'s satisfy the following moment condition:
\begin{equation}
\label{4thmom} \max_{1\le\nu\le n}\sum_{i=1}^n
\omega_{i\nu}^4=o\bigl(n^{8/5}(\log
\log{n})^{-14/5}\bigr).
\end{equation}
Recall that $\omega_{i\nu}=\int_{0}^1X_i(t)\widehat\varphi_{\nu
}(t)\,dt$ and is nonrandom under deterministic design.
Condition (\ref{4thmom}) implies that for any $\nu=1,\ldots,n$, the
magnitudes of $\omega_{1\nu},\ldots,\omega_{n\nu}$ should be
comparable given the restriction that $\sum_{i=1}^n\omega_{i\nu
}^2=n$. It\vspace*{1pt} rules out the situation that the sequence $\omega_{i\nu}$
is spiked at $i=\nu$, that is, $\omega_{\nu\nu}^2=n$ and $\omega
_{i\nu}=0$, for any $i\neq\nu$. This special situation essentially
gives rise to $\Omega=\sqrt{n}I$ such that $\mathrm{PLRT}_k$ defined in (\ref
{gauplrt1}) can be written as a scaled sum of independent centered
squares of the errors. The leave-one-out method employed in Theorem
\ref{adapasym} can handle this special case.
We first standardize $\mathrm{PLRT}_k$. The non-Gaussian assumption yields a
substantially different design matrix. Hence,
the scale factor is chosen to be different from the one used in
Section~\ref{testgau},
as described below. The standardized version is defined as
\[
\tilde{\tau}_k=\frac{-2n\cdot \mathrm{PLRT}_k-\sum_{\nu=1}^{n}d_\nu(k)}{
(2\sum_{i\neq j}a_{ij}^2(k))^{1/2}},
\]
where
$a_{ij}(k)$ is the $(i,j)$th entry of $A_k\equiv n^{-1}\Omega
(I_n+\lambda_k\Lambda_k)^{-1}\Omega^T$ for $1\le i,j\le n$. Note
that the scale factor in $\tilde{\tau}_k$, that is,
the term $(2\sum_{i\neq j}a_{ij}(k)^2)^{1/2}$, differs from the one in
$\tau_k$.
Technically, this new scale factor will facilitate the
asymptotic theory developed later in this section.
Let $AT_n^*=\max_{1\le k\le k_n}\tilde{\tau}_k$,
and $AT_n=B_n(AT_n^*-B_n)$, where $B_n$ satisfies $2\pi B_n^2\exp
(B_n^2)=k_n^2$.
\begin{Theorem}\label{adapasymnongauss}
Suppose $k_n\asymp(\log{n})^{d_0}$, for some constant $d_0\in(0,1/2)$.
Furthermore, $\epsilon$ is sub-Gaussian, and (\ref{4thmom}) holds.
Then for any $\bar{\alpha}\in(0,1)$, we have under $H_0\dvtx \beta=0$,
\[
P(AT_n\le c_{\bar{\alpha}})\rightarrow1-\bar{\alpha}\qquad\mbox{as $n
\rightarrow\infty$},
\]
where $c_{\bar{\alpha}}=-\log(-\log(1-\bar{\alpha}))$.
\end{Theorem}
The proof of Theorem~\ref{adapasymnongauss} is mainly
based on Stein's exchangeable pair method; see \cite{Stein86}.\vadjust{\goodbreak}
We conclude this subsection by showing that the proposed adaptive test
can still achieve the optimal minimax rate (up to a logarithmic order)
specified in \cite{HMV13}, that is, $\delta(n,k)$, even under
non-Gaussian errors. Recall that $\delta(n,k)$, $\llVert\mathfrak
{b}\rrVert_{\ell^2}$, $\llVert\mathfrak{b}\rrVert
_{k,\ell^2}$ and
$P_{\mathfrak{b}}$ are defined in Section~\ref{testgau}.
\begin{Theorem}\label{minimaxadaptnongauss}
Suppose $k_n\asymp(\log{n})^{d_0}$, for some constant $d_0\in(0,1/2)$.
Furthermore, $\epsilon$ is sub-Gaussian, and (\ref{4thmom}) holds.
Then, for any $\varepsilon\in(0,1)$,
there exist positive constants $N_\varepsilon$ and $C_\varepsilon$
such that for any $n\ge N_\varepsilon$,
\[
\mathop{\inf_{\llVert\mathfrak{b}\rrVert_{\ell^2}\ge
C_\varepsilon\delta(n,k)}}_{
\llVert\mathfrak{b}\rrVert_{k,\ell^2}\le1} P_{\mathfrak
{b}}(
\mbox{reject }H_0)\ge1-\varepsilon.
\]
\end{Theorem}
\section{Simulation study}\label{secsimulation}
In this section, we investigate the numerical performance of the
proposed procedures for inference.
We consider four different simulation settings. The settings in
Sections~\ref{setting1}--\ref{setting3}
are exactly the same as those in Hilgert et al. \cite{HMV13} and Lei \cite{L13}
so that we can fairly compare our testing results with theirs. We focus
on models with Gaussian error and choose $m=2$, that is, cubic spline.
Confidence interval in Section~\ref{secasympCI}, penalized
likelihood ratio test in Section~\ref{secPLRT} and adaptive testing
procedure in Section~\ref{testgau} are examined.
The setting in Section~\ref{setting4} is about functional linear
logistic regression.
Size and power of the PLRT test are examined.
\subsection{Setting 1}\label{setting1}
Data were generated in the same way as in Hilgert et al. \cite{HMV13}.
Consider the functional linear model $Y_i=\int_0^1 X_i(t)\beta_0(t)\,dt
+ \epsilon_i$, with $\epsilon_i$ being independent standard normal
for $i=1,\ldots,n$. Let $\lambda_j=(j-0.5)^{-2}\pi^{-2}$ and
$V_j(t)=\sqrt{2}\sin((j-0.5)\pi t)$, $t \in[0,1], j=1, 2, \ldots,
100$. The covariate curve $X_i(t)$ was Brownian motion simulated as
$X_i(t)=\sum_{j=1}^{100}\sqrt{\lambda_j}\eta_{ij}V_j(t)$, where
$\eta_{ij}$'s are independent standard normal for $i=1,\ldots,n$ and
$j=1,\ldots,100$. Each $X_i(t)$ was observed at $1000$ evenly spaced
points over $[0,1]$. The true slope function was chosen as
\[
\beta_0^{B,\xi}(t)= \frac{B}{\sqrt{\sum_{k=1}^{\infty}k^{-2\xi-1}}}\sum
_{j=1}^{100}j^{-\xi-0.5}V_j(t).
\]
Figure \ref{fig1} displays $\beta_0$. Four different signal strengths $B=(0,0.1,0.5,1)$ and three smoothness
parameters $\xi=(0.1,0.5,1)$
were considered. Note that $B=0$ implies $\beta_0=0$.
\begin{figure}
\includegraphics{1322f01.eps}
\caption{Plots of $\beta_0(t)$ in settings 1 and 2.}\label{fig1}
\end{figure}
For each case study, we considered sample sizes $n=100$ and $n=500$
respectively,
and ran 10,000 trials to investigate the Monte Carlo performance of our methods.
\begin{longlist}
\item[\textit{Case study} 1: 95\% \textit{confidence interval for conditional mean}.]
In this study, we set $\mu_0(x_0) = E\{Y \mid X_0 = x_0\}=\int_0^1
x_0(t)\beta_0(t)\,dt$ with $B=1,\xi=1$,
where $x_0$ is independent of $X_1,\ldots,X_n$
and randomly generated from the same distribution as~$X_1$.
From (\ref{CIconditionalmean}),
the $95\%$ confidence interval for $\mu_0(x_0)$ is
\[
\bigl[\widehat Y_0 - n^{-1/2}z_{0.025}
\sigma_n,\widehat Y_0 + n^{-1/2}z_{0.025}
\sigma_n \bigr],
\]
where $\sigma_n^2=1+\sum_{\nu=1}^\infty x_\nu^2/(1+\lambda\rho
_\nu)$,
$x_\nu=\int_0^1 x_0(t)\varphi_\nu(t)\,dt$. Here $\varphi_\nu$ and
$\rho_\nu$ are both obtained through
(\ref{bvpexample}).
With 10,000 replications, percentages of the conditional mean $\mu_0(x_0)$
beyond the scope of CI and the average lengths of the confidence intervals
are summarized in Table~\ref{ci1}.
\begin{table
\tabcolsep=0pt
\tablewidth=180pt
\caption{Case study 1: Percentage of $\mu_0(x_0)$ outside the $95 \%$
confidence intervals $\pm$ standard deviation
(average length of the $95 \%$ confidence intervals)}\label{ci1}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lc@{}}
\hline
$\bolds{n=100}$ & $\bolds{n=500}$ \\
\hline
$4.89\pm0.42~(0.56)$ & $5.01\pm0.19~(0.39)$\\
\hline
\end{tabular*}
\end{table}
\item[\textit{Case study} 2: \textit{Size of the tests}.]
Denote\vspace*{1pt} the testing methods proposed by Hilgert et al. \cite
{HMV13} as $\mathrm{HMV13}^{(1)}$ and $\mathrm{HMV13}^{(2)}$. Under
$H_0\dvtx \beta=0$, we calculated the sizes of PLRT and AT (adaptive
testing), that is,\vspace*{1pt} the percentages of rejecting $H_0$, and then
compared them with $\mathrm{HMV13}^{(1)}$ and $\mathrm{HMV13}^{(2)}$
(directly cited from \cite{HMV13}) in Table~\ref{size}. Numerically,
we found that AT converges to the Gumbel distribution very slowly,
which is a common phenomenon in the extreme value literature; see \cite
{F96,FL98}. Following an idea similar to \cite{FL98}, finite sample
distributions of AT based on one million replications were instead
used. Obviously, from Table~\ref{size}, the proposed PLRT and AT are
both valid test statistics achieving desirable sizes.
\begin{table}[b]
\tabcolsep=0pt
\tablewidth=180pt
\caption{Case study 2: Sizes of the tests}\label{size}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lcc@{}@{}}
\hline
& $\bolds{n=100}$ & $\bolds{n=500}$ \\
\hline
$\mathrm{HMV13}^{(1)}$ & $ 3.47~(\pm0.36)$ & $2.61~(\pm0.14)$\\
$\mathrm{HMV13}^{(2)}$ & $4.97~(\pm0.43)$ & $5.26~(\pm0.20)$\\
AT& $5.13~(\pm0.43)$ & $5.04~(\pm0.19)$ \\
PLRT & $5.45~(\pm0.45)$ & $5.19~(\pm0.20)$\\
\hline
\end{tabular*}
\end{table}
\item[\textit{Case study} 3: \textit{Power comparison}.]
In this study, we generated $\beta_0$ under different signal strengths
$B=(0.1,0.5,1)$ and smoothing parameters $\xi=(0.1,0.5,1)$.
Tables~\ref{power1} and~\ref{power2} summarize the powers of four
different testing methods, that is, the percentages of
rejecting $H_0\dvtx \beta=0$ at $95\%$ significance level, under $n=100$
and $n=500$. From $n=100$ to $n=500$, the powers of all tests increase.
In particular, PLRT generally performs better than AT since PLRT
incorporates known information from the model,
that is, $r=0$ (smoothness of the covariance kernel) and $m=2$
(smoothness of the functional parameter), while AT is adaptive on these
quantities. The power loss is the price paid for adaptiveness. We also
note that for weaker signals $B=0.1$, the\vspace*{1pt} powers of PLRT and AT improve
those of $\mathrm{HMV13}^{(1)}$, $\mathrm{HMV13}^{(2)}$, while for
stronger signals $B=0.5,1$, the powers of all tests are comparable.
\end{longlist}
\begin{table
\tabcolsep=0pt
\caption{Case study 3: $n=100$. Powers}\label{power1}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lca{5.7}a{5.7}a{5.7}@{\hspace*{-2pt}}}
\hline
& \textbf{Test} & \multicolumn{1}{c}{$\bolds{B=0.1}$} & \multicolumn{1}{c}{$\bolds{B=0.5}$} & \multicolumn{1}{c@{}}{$\bolds{B=1}$}\\
\hline
$\xi=0.1$ & $\mathrm{HMV13}^{(1)}$ & 3.88;(\pm0.38) & 21.41;(\pm0.8) & 77.24;(\pm0.82) \\
& $\mathrm{HMV13}^{(2)}$ & 5.80;(\pm0.46) & 26.38;(\pm0.86) & 81.78;(\pm0.76) \\
& AT & 6.12;(\pm0.47 ) & 30.77;(\pm0.90 ) & 81.56;(\pm0.76 ) \\
& PLRT & 21.27;(\pm0.80) & 42.34;(\pm0.97) & 84.20;(\pm0.71)
\\[3pt]
$\xi=0.5$ & $\mathrm{HMV13}^{(1)}$ & 4.74;(\pm0.42) & 46.47;(\pm0.98) & 98.68;(\pm0.22) \\
& $\mathrm{HMV13}^{(2)}$ & 6.65;(\pm0.49) & 52.79;(\pm0.98) & 99.06;(\pm0.19) \\
& AT & 8.28;(\pm0.54 ) & 71.08;(\pm0.89 ) & 99.86;(\pm0.07) \\
& PLRT & 23.13;(\pm0.83) & 74.74;(\pm0.85) & 99.70;(\pm0.11)
\\[3pt]
$\xi=1$ & $\mathrm{HMV13}^{(1)}$ & 4.8;(\pm0.42) & 62.67;(\pm0.95) & 99.75;(\pm0.10) \\
& $\mathrm{HMV13}^{(2)}$ & 7.07;(\pm0.5) & 68.30;(\pm0.91) & 99.84;(\pm0.08) \\
& AT & 9.47;(\pm0.57) & 83.20;(\pm0.73 ) & 99.98;(\pm0.03) \\
& PLRT & 23.95;(\pm0.84) & 84.03;(\pm0.72) & 99.98;(\pm0.03) \\
\hline
\end{tabular*}
\end{table}
\begin{table}[b]
\caption{Case study 3: $n=500$. Powers}\label{power2}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lca{5.7}a{5.7}a{3.4}@{}}
\hline
& \textbf{Test} & \multicolumn{1}{c}{$\bolds{B=0.1}$} & \multicolumn{1}{c}{$\bolds{B=0.5}$} & \multicolumn{1}{c@{}}{$\bolds{B=1}$}\\
\hline
$\xi=0.1$ & $\mathrm{HMV13}^{(1)}$ & 5.17;(\pm0.19) & 86.98;(\pm0.29) & 100;(\pm0) \\
& $\mathrm{HMV13}^{(2)}$ & 8.48;(\pm0.24) & 90.89;(\pm0.25) &100;(\pm0) \\
& AT & 9.57;(\pm0.26) &89.14;(\pm0.27 ) &100;(\pm0)\\
& PLRT & 20.00;(\pm0.35) & 88.19;(\pm0.28) & 100;(\pm0)
\\[3pt]
$\xi=0.5$ & $\mathrm{HMV13}^{(1)}$ & 8.81;(\pm0.25) & 99.85;(\pm0.03) & 100;(\pm0) \\
& $\mathrm{HMV13}^{(2)}$ & 13.07;(\pm0.30) & 99.88;(\pm0.03) &100;(\pm0) \\
& AT &20.20;(\pm0.35 ) &100;(\pm0) &100;(\pm0)\\
& PLRT & 29.47;(\pm0.40) & 99.90;(\pm0.03) & 100;(\pm0)
\\[3pt]
$\xi=1$ & $\mathrm{HMV13}^{(1)}$ & 11.38;(\pm0.28) & 99.99;(\pm0.01) & 100;(\pm0) \\
& $\mathrm{HMV13}^{(2)}$ & 16.13;(\pm0.32) & 100;(\pm0) & 100;(\pm0) \\
& AT &26.51;(\pm0.39) &100;(\pm0 ) &100;(\pm0)\\
& PLRT & 34.08;(\pm0.42) & 100;(\pm0) & 100;(\pm0) \\
\hline
\end{tabular*}
\end{table}
\subsection{Setting 2}\label{setting2}
Let the true slope function be
\[
\beta_0^{B, \tau}(t)=B \exp\biggl\{-\frac{(t-0.5)^2}{2\tau
^2} \biggr
\} \biggl\{\int_0^1\exp\biggl\{-
\frac{(x-0.5)^2}{\tau
^2} \biggr\}\,dx \biggr\}^{-1/2},
\]
where $B=(0.5,1,2)$ and $\tau=(0.01,0.02,0.05)$. The processes
$X_i(t)$ and the samples were generated in the same way
as in Setting~1.\vadjust{\goodbreak}
The powers in Setting~2 are summarized in Tables~\ref{power3} and~\ref
{power4}. We observe similar phenomena as in Setting~1, that under
weaker signals, say $\tau=0.01, B=0.5$, PLRT and AT demonstrate larger
powers, while the powers of all procedures become comparable under
stronger signals. Again, PLRT generally has larger powers than the
adaptive procedure AT. All the powers increase as sample size becomes larger.
\begin{table
\tabcolsep=0pt
\caption{Setting~2: $n=100$. Powers}\label{power3}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lca{5.7}a{5.7}a{5.7}@{\hspace*{-2pt}}}
\hline
& \textbf{Test} & \multicolumn{1}{c}{$\bolds{B=0.5}$} & \multicolumn{1}{c}{$\bolds{B=1}$} & \multicolumn{1}{c@{}}{$\bolds{B=2}$}\\
\hline
$\tau=0.01$ & $\mathrm{HMV13}^{(1)}$ & 4.94;(\pm0.42) & 11.85;(\pm0.63) & 46.69;(\pm0.98) \\
& $\mathrm{HMV13}^{(2)}$ & 7.25;(\pm0.51) & 15.49;(\pm0.71) &53.56;(\pm0.98) \\
& AT & 9.88;(\pm0.58) & 23.86;(\pm0.84 ) & 69.46;(\pm0.90 )\\
& PLRT & 17.9;(\pm0.75 )& 33.25;(\pm0.92 ) & 81.04;(\pm0.77)
\\[3pt]
$\tau=0.02$& $\mathrm{HMV13}^{(1)}$ & 7.33;(\pm0.51) & 23.09;(\pm0.83) & 80.26;(\pm0.78) \\
& $\mathrm{HMV13}^{(2)}$ & 10;(\pm0.59) & 28.54;(\pm0.89) &84.04;(\pm0.72) \\
& AT & 14.58;(\pm0.69) & 42.21;(\pm0.97) & 93.54;(\pm0.48) \\
& PLRT & 22.87;(\pm0.82 )& 53.21;(\pm0.98 ) & 97.83;(\pm0.29)
\\[3pt]
$\tau=0.05$ & $\mathrm{HMV13}^{(1)}$ & 13.85;(\pm0.68) & 56.51;(\pm0.97) & 99.48;(\pm0.14) \\
& $\mathrm{HMV13}^{(2)}$ & 18.13;(\pm0.50) & 63.09;(\pm0.95) &99.65;(\pm0.12) \\
& AT & 28.31;(\pm0.88) & 78.52;(\pm0.80 ) & 99.96;(\pm0.04)\\
& PLRT & 37.54;(\pm0.95) & 87.63;(\pm0.65) & 100;(\pm0)\\
\hline
\end{tabular*}\vspace*{-6pt}
\end{table}
\begin{table}[b]
\tabcolsep=0pt
\caption{Setting~2: $n=500$. Powers}\label{power4}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lca{5.7}a{5.7}a{5.7}@{\hspace*{-2pt}}}
\hline
& \textbf{Test} & \multicolumn{1}{c}{$\bolds{B=0.5}$} & \multicolumn{1}{c}{$\bolds{B=1}$} & \multicolumn{1}{c@{}}{$\bolds{B=2}$}\\
\hline
$\tau=0.01$ & $\mathrm{HMV13}^{(1)}$ & 12.41;(\pm0.42) & 54.6;(\pm0.63) & 99.75;(\pm0.98) \\
& $\mathrm{HMV13}^{(2)}$ & 17.99;(\pm0.51) & 63.16;(\pm0.71) &99.98;(\pm0.98) \\
& AT & 28.93;(\pm0.40 ) & 79.75;(\pm0.35 ) & 100;(\pm0 )\\
& PLRT & 34.77;(\pm0.42 )& 86.08;(\pm0.30 ) & 100;(\pm0 )
\\[3pt]
$\tau=0.02$& $\mathrm{HMV13}^{(1)}$ & 26.11;(\pm0.51) &88.91;(\pm0.83) & 100;(\pm0) \\
& $\mathrm{HMV13}^{(2)}$ & 33.95;(\pm0.59) & 92.62;(\pm0.89) &100;(\pm0) \\
& AT & 50.25;(\pm0.44 ) & 97.03;(\pm0.15) & 100;(\pm0) \\
& PLRT & 56.57;(\pm0.43)& 99.20;(\pm0.08) & 100;(\pm0)
\\[3pt]
$\tau=0.05$ & $\mathrm{HMV13}^{(1)}$ & 65.38;(\pm0.68) & 99.95;(\pm0.97) & 100;(\pm0) \\
& $\mathrm{HMV13}^{(2)}$ & 72.74;(\pm0.50) & 99.99;(\pm0.95) &100;(\pm0) \\
& AT & 86.92;(\pm0.30) & 100;(\pm0 ) & 100;(\pm0)\\
& PLRT & 92.07;(\pm0.24) & 100;(\pm0 ) & 100;(\pm0 ) \\
\hline
\end{tabular*}\vspace*{-7pt}
\end{table}
\subsection{Setting~3}\label{setting3}
In this setting, data were generated in the same way as in Section~4.2
of \cite{L13}. Hence we will compare our PLRT and AT with the testing
procedure in \cite{L13}, denoted as L13. Specifically, the covariance
operator has eigenvalues $\kappa_j = j^{-1.7}$ and eigenfunctions
$\phi_1(t)=1, \phi_j(t)=\sqrt{2}\cos((j-1)\pi t)$ for $j \geq2$.
The covariate processes are $X_i(t)=\sum_{j=1}^{100}\sqrt{\kappa
_j}\eta_j \phi_j(t)$, where $\eta_j$'s are independent standard
normal. Each $X_i(t)$ was observed on $1000$ evenly spaced points over~$[0,1]$.
In the first case denoted as $\operatorname{Model}(2,1)$, let $\theta_j = \bar
{\theta}_j/\llVert\bar{\theta}\rrVert_2$,
where $\bar{\theta}_j=0$ for $j>2$, $\bar{\theta}_j=b_j\cdot I_j$
for $j=1,2$, $b_1$ and $b_2$
are independent $\operatorname{Unif}(0,1)$, and\vadjust{\goodbreak} $(I_1,I_2)$ follows a multinomial
distribution $\operatorname{Mult}(1;0.5,0.5)$.
Let the true function be
$\beta_0(t)=r\sum_{j=1}^{100}\theta_j\phi_j(t)$, where
$r^2=(0,1,0.2,0.5,1.5)$.
\begin{table
\tabcolsep=0pt
\caption{Setting~3: Powers}\label{power5}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lcca{5.7}a{5.7}a{5.7}a{5.7}@{\hspace*{-2pt}}}
\hline
& \textbf{Sample size} & \textbf{Test} &
\multicolumn{1}{c}{$\bolds{r^2=0.1}$} &
\multicolumn{1}{c}{$\bolds{r^2=0.2}$} &
\multicolumn{1}{c}{$\bolds{r^2=0.5}$} &
\multicolumn{1}{c@{}}{$\bolds{r^2=1.5}$}\\
\hline
$\operatorname{Model} (2,1)$ & $n=50$ & L13 & 16.20;(\pm1.02) & 26.40;(\pm1.22) & 54.20;(\pm1.38) & 80.80;(\pm1.09) \\
& & AT & 47.81;(\pm1.38) & 64.94;(\pm1.32 )& 84.75;(\pm1.00) & 99.13;(\pm0.26 )\\
& & PLRT & 57.76;(\pm1.37) & 72.70;(\pm1.23) & 90.18;(\pm0.82) & 99.52;(\pm0.19) \\[3pt]
& $n=100$ & L13 & 25.80;(\pm0.86 ) & 42.20;(\pm0.97) & 68.20;(\pm0.91) & 90.40;(\pm0.58) \\
& & AT & 65.53;(\pm0.93 ) & 79.75;(\pm0.79 )& 96.97;(\pm0.34) & 99.99;(\pm0.02 ) \\
& & PLRT & 74.04;(\pm0.86 ) & 87.98;(\pm0.64) & 98.22;(\pm0.26) & 100;(\pm0) \\[3pt]
& $n=500$ & L13 & 67.20;(\pm0.41) & 84.60;(\pm0.32) & 94.40;(\pm0.20) & 97.20;(\pm0.14) \\
& & AT & 97.81;(\pm0.13 ) & 100;(\pm0)& 100;(\pm0 ) & 100;(\pm0 ) \\
& & PLRT & 98.5;(\pm0.11) & 99.94;(\pm0.02) & 100;(\pm0) &100;(\pm0)
\\[6pt]
$\operatorname{Model} (9,2)$ & $n=50$ & L13 & 9.00;(\pm0.79) & 14.00;(\pm0.96)& 29.60;(\pm1.27) & 43.40;(\pm1.37) \\
& & AT & 21.72;(\pm1.14) & 27.57;(\pm1.24 )& 37.67;(\pm1.34) & 53.33;(\pm1.38 )\\
& & PLRT & 39.54;(\pm1.36 ) & 46.22;(\pm1.38) & 56.92;(\pm1.37) & 73.42;(\pm1.22) \\[3pt]
& $n=100$ & L13 & 13.4;(\pm0.67) & 27.8;(\pm0.88) & 39.8;(\pm0.96) & 65.8;(\pm0.93) \\
& & AT & 27.86;(\pm0.88 ) & 21.63;(\pm0.81)& 47.61;(\pm0.98) & 65.69;(\pm0.93) \\
& & PLRT & 45.80;(\pm0.98) & 53.72;(\pm0.98) & 67.12;(\pm0.92) & 83.88;(\pm0.72) \\[3pt]
& $n=500$ & L13 & 42.40;(\pm0.43) & 47.8;(\pm0.44) & 72.4;(\pm0.39) & 93.4;(\pm0.22) \\
& & AT & 49.40;(\pm0.44 ) & 58.29;(\pm0.43)& 80.21;(\pm0.35) & 91.23;(\pm0.25 ) \\
& & PLRT & 69.44;(\pm0.40) & 80.00;(\pm0.35) & 91.70;(\pm0.24) & 99.22;(\pm0.08) \\
\hline
\end{tabular*}
\end{table}
In the second case denoted as $\operatorname{Model}(9,2)$, a different choice of
$\theta_j$ was considered.
Specifically, $\theta_j = \bar{\theta}_j/\llVert\bar{\theta
}\rrVert_2$, where $\bar{\theta}_j=0$ for $j>9$,
$\bar{\theta}_j=b_j\cdot I_j$ for $j=1,\ldots,9$, $b_1,\ldots,b_9$
are independent $\operatorname{Unif}(0,1)$, and $(I_1,\ldots,I_9)$ follows a
multinomial distribution $\operatorname{Mult}(2;1/9,\ldots,1/9)$.
In both cases, the samples were drawn from $Y_i=\int_{0}^1 X_i(t)\beta
(t)\,dt + \epsilon_i$, $i=1,\ldots,n$,
where $\epsilon_i$ are independent standard Gaussian. $5000$ Monte
Carlo trials
were conducted in each case under different sample sizes $n=50, 100$
and $500$.
Results are summarized in Table~\ref{power5},
from which we can see that the powers of AT and PLRT improve those
of L13, especially when $r^2=0.1,0.2$ (weaker signals). As $n$ increases,
the power of L13 becomes more comparable to those of PLRT and AT
especially when $r^2=1.5$ (stronger signal).
Again, PLRT generally has larger powers than adaptive methods.
\subsection{Setting~4}\label{setting4}
Let $Y\in\{0,1\}$ be a binary variable
generated from the following functional logistic regression model:
\[
P(Y=1\mid X)=\frac{\exp(\int_0^1X(t)\beta_0(t)\,dt)}{1+\exp(\int
_0^1X(t)\beta_0(t)\,dt)}.
\]
The predictor process $X_i$ was simulated as $X_i(t)=\sum
_{j=1}^{100}\sqrt{\lambda_j}\eta_{ij}V_j(t)$,
where $\lambda_j$ and $V_j(t)$ are exactly the same as in Setting~1,
$\eta_{ij}$'s are independent truncated normals,
that is, $\eta_{ij}=\xi_{ij}I_{\{\llvert \xi_{ij}\rrvert \leq0.5\}} +
0.5I_{\{\xi
_{ij}> 0.5\}} -0.5I_{\{\xi_{ij}< -0.5\}}$,
with $\xi_{ij}$ being a standard normal random variable.
Each $X_i(t)$ was observed at $1000$ evenly spaced points over $[0,1]$.
We intend to test $H_0\dvtx \beta=0$.
To examine the power, data were generated under $\beta
_0(t)=3*10^5(t^{11}(1-t)^6)$
for $t\in[0,1]$.
We examined two sample sizes: $n=100$ and $n=500$.
Results (summarized in Table~\ref{setting4sizepower})
were based on 10,000 independent trials.
It can be seen that when $n=100$ and $500$,
the test achieves the desired sizes.
The power at $n=100$ is small,
but the power at $n=500$ approaches one, demonstrating the asymptotic
property of the test.
\begin{table
\tabcolsep=0pt
\tablewidth=180pt
\caption{Setting~4: Size and power}\label{setting4sizepower}
\begin{tabular*}{\tablewidth}{@{\extracolsep{\fill}}@{}lcc@{}}
\hline
& $\bolds{n=100}$ & $\bolds{n=500}$ \\
\hline
Size & 0.054 & 0.046 \\
Power& 0.387 & 0.985 \\
\hline
\end{tabular*}
\end{table}
\section{Discussion}\label{secdisc}
The current paper and our previous work on nonparametric regression
models \cite{SC13} are both built upon the RKHS framework and theory.
Hence it seems necessary for us to comment their technical connections
and differences to facilitate the reading. Compared to \cite{SC13},
the RKHS considered in the current paper has a substantially different
structure that involves a covariance function of the predictor process.
This immediately causes a difference in building the eigensystems:
\cite{SC13} relies on an ODE system, but the current paper relies on
an integro-differential system. Hence the methods of analyzing both
systems are crucially different. Meanwhile, the asymptotic analysis on
the statistical inference such as the penalized likelihood ratio test
are also different. For example, \cite{SC13} only considers the
reproducing kernel, while the current work requires a delicate
interaction between
the reproducing kernel and the covariance kernel. More importantly, the
relaxation of perfect alignment between both kernels poses more
technical challenges.
Besides, Assumption~\ref{A3} requires $\llVert\varphi_\nu\rrVert
_{L^2}\le C_\varphi\nu^a$ for $\nu\ge1$ and a constant $a\ge
0$. The introduction of factor $a$ in Assumption~\ref{A3} is helpful
in simplifying our proofs. However,
it is interesting to investigate how to avoid imposing this seemingly
``redundant'' $a$.
As indicated by Proposition~\ref{validA3}, that $a$ relates to $C$
(and hence $V$), one possible strategy is to avoid the use of $V$.
Instead, one may use its empirical version, namely $V_n$, as suggested
by one referee. This would require a delicate analysis of the
convergence of $V_n$,
which may be handled by techniques in \cite{M10}. We leave this as a
future exploration.
\section*{Acknowledgments}
We thank Pang Du for providing us his R
code for functional logistic regression,
and thank Ph.D. student Meimei Liu at Purdue for help with the simulation
study. We also thank Co-editor
Runze Li, an Associate Editor, and two referees for helpful comments
that lead to
important improvements on the paper.
Guang Cheng was on sabbatical at Princeton while the revision of this
work was carried out; he would like
to thank the Princeton ORFE department for its hospitality and support.
\begin{supplement}[id=suppA]
\stitle{Supplement to ``Nonparametric inference in generalized functional~linear models''}
\slink[doi]{10.1214/15-AOS1322SUPP}
\sdatatype{.pdf}
\sfilename{aos1322\_supp.pdf}
\sdescription{Proofs are provided.}
\end{supplement}
|
\section{Introduction}
Junctions with confined electrons, like atomic, molecular or quantum-dot (QD) junctions, are among the most studied nanoscopic devices \cite{Andergassen}. In these structures electron-electron correlations within the junction together with the electronic properties of the contacts lead to a number of different phenomena concerning fundamental aspects of quantum charge and spin transport.
Modern technologies allowed building junctions that are close to the ultimate limit of miniaturization with normal, superconducting \cite{Yu, Morpurgo, Buitelaar, Doh, Jorgensen, Xiang, Jarillo} or ferromagnetic \cite{Ferro1, Ferro2} leads creating new opportunities for novel nanodevices with predefined functional properties. One-electron transistors, spin valves or superconducting spin qubits are some examples of such devices.
When a molecule or a QD bridges the gap between two metallic leads, the Coulomb energy tends to quantize the charge confined in the junctions, i.e. on the molecule or the dot, leading to the possibility of fabricating a junction with a confined spin. In the case of superconducting leads, the characteristic of the so-obtained Josephson junction depends on the total spin in the junctions.
Superconducting circuits with QD (S-QD-S junctions) have been extensively studied during the last decade \cite{Levy_Yeyati_Review,De_Franceschi_Review}. These junctions are fabricated by contacting superconducting leads to a normal nanostructure, typically a single walled carbon nanotube or a semiconducting nanowire. The structure may include gate electrodes that can be used to control the number of electrons in the dot.
S-QD-S junctions having a localized spin in the dot may have a global minimum of the free energy for a $\pi$-difference between the phases of the two superconducting contacts . The current-phase characteristic of these junctions is described by a Josephson equation with a {\it negative} critical current \cite{pi_junction1, pi_junction2}. These junctions, referred to as $\pi$-junctions, in contrast with standard "$0$-junctions", have interesting properties and potential applications in superconducting electronics, including phase or flux qubits \cite{Makhlin_Review}.
A superconducting ring containing a $\pi$-junction could generate a spontaneous current with (nearly) half a superconducting quantum flux threading the ring, a very convenient situation for experimental detection. Notice that this current structure is stable only if the ring self-inductance exceeds a critical value\cite{Sigrist_Rice,Jagla}.
Here we present results for a $\pi$-bijunction consisting of a QD connected to three superconducting leads, see Fig. \ref{fig:Triterminal_dot}. Graphene dots (GQD) are good candidates to build such a device\cite{graph1, graph2}. In fact graphene offers new opportunities for superconducting electronics as a new class of material that can be tailored and contacted to normal or superconducting leads. With gate electrodes controlling the number of electrons confined in the GQD, a non-zero spin can be localized at the dot which tends to generate a $\pi$-shift between each of the pairs of superconductors. This situation, in a way similar to Heisenberg magnets, generates frustration\cite{Sigrist_Rice}.The frustration can be resolved by canting the phases of the three superconductors, in a way that depends on the asymmetry of the device. This asymmetry, due to different couplings between the dot and the contacts, can be controlled by gates. Frustration leads to canting only for moderate asymmetries. In the canted phase, the equilibrium phase difference between two given superconductors can be controlled at will to a value $\varphi_m$, between $0$ and $\pi$. Such a tunable $\varphi$-junction can be probed by various geometries incorporating one or several superconducting loops. An important feature of the phase canted state is that it contains a spontaneous vorticity. Due to time-reversal symmetry, two equivalent solutions with opposite vorticities are found, corresponding to phases $\varphi_m$ and $-\varphi_m$. This symmetry can be broken with the help of a single loop and an applied magnetic field. In addition, the structure of the energy-phase profile of the bijunction makes the barrier between the two degenerate minima tunable, either through the bijunction parameters, or using the external flux. This might be an useful property for building a superconducting qubit.
The paper is organized as follows. In Section 2 a phenomenological phase model is first considered, with a stability analysis of the canted (frustrated) solution against the non-frustrated solution. Then a microscopic model for a dot with a single level allows a nonperturbative solution, which essentially confirms the existence of a canted phase below a critical asymmetry of the dot-contact couplings. Section 3 considers a single loop set-up, with an applied orbital magnetic field, then two-loop or three-loop set-ups.
\section{bijunction at equilibrium}
\subsection{A phenomenological model}
The quantum dot connects all three superconductors (Fig. \ref{fig:Triterminal_dot}). Each pair $(i,j)$ of the three superconductors ($i=1,2,3$) forms a Josephson junction. As a first approximation, one can write the total energy of the bijunction as the sum of those of separated junctions. Such an expression could be obtained in perturbation theory from a microscopic Hamiltonian, at fourth order in the tunnelling element between the superconductor states and the dot states. The bijunction is then equivalent to a triangular array of separated junctions. We assume that the presence of a $1/2$ spin on the dot creates $\pi$-junctions, and that this holds for all of them. The nonperturbative calculation presented in the next subsection shows that it is essentially the case, unless the couplings to the dot are very asymmetric. Denoting the superconducting phases as $\varphi_i$ ($i=1-3$), the bijunction energy thus reads :
\begin{equation}
\label{eq:energy_BJ_red}
E_{BJ}=E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos\varphi_{23}],
\end{equation}
with $\varphi_{ij}=\varphi_i-\varphi_j$, $E_0>0$ and where $g_0\geq0$, $g\geq0$ are parameters quantifying the bijunction asymmetry (Fig. \ref{fig:Triterminal_dot}).
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE1.pdf}
\caption{(Color online) Bijunction made of three superconductors and one quantum dot carrying a spin $S=1/2$. Inset show the equivalent triangular model, valid in the perturbative regime only. The asymmetry ratios between the junctions critical currents are indicated.
\label{fig:Triterminal_dot}
}
\end{figure}
Let us look for the equilibrium state. Setting to zero the partial derivatives of $E_{BJ}$ with respect to the $\varphi_i$'s is equivalent to imposing zero current $J_i$ in each lead $S_i$. One obtains from $J_1=0$ :
\begin{equation}
\label{zerocurrent}
g\sin\varphi_{12}+\sin\varphi_{13}=0,
\end{equation}
which, together with similar equations expressing that $J_2$ or $J_3=0$, yields :
\begin{equation}
\label{canting}
\cos\varphi_{23}=\frac{(g_0g)^2-1-g^2}{2g}.
\end{equation}
Such a nontrivial solution thus exists only if $|1-\frac{1}{g}|\leq g_0 \leq (1+\frac{1}{g})$. This is a canted (e.g. frustrated) phase solution (Fig. \ref{fig:phase_vortices}a,b), with two degenerate states obtained from each other by changing $\varphi_{2,3}$ into $-\varphi_{2,3}$. In such states, the current across any junction $S_i-S_j$ is nonzero. Yet, the total current in each lead is zero. Those two degenerate solutions therefore feature a phase vortex, with two opposite vorticities. While in a real triangular network this vorticity is associated to a circulating current, with a zero-dimensional quantum dot, it is difficult to define a path with a nonzero current circulating around the dot. Nevertheless, we show in the last Section that a true vortex can be induced on an adjacent loop.
In the opposite case $|(g_0g)^2-1-g^2|\geq 2g$, the energy minimum is obtained for $\varphi_2=0$ or $\pi$, $\varphi_3=0$ or $\pi$, replacing Eq. (\ref{canting}). This results in two of the three junctions being $\pi$-junctions and the other one a $0$-junction (Fig. \ref{fig:phase_vortices}c).
Later on we consider situations where the lead $1$ is disconnected (e.g. the phase $\varphi_1$ is floating), while leads $2$ and $3$ are connected to an external circuit. Then the convenient phase variable is $\varphi=\varphi_2-\varphi_3$. One can use gauge invariance and choose $\varphi_2=\frac{\varphi}{2}$, $\varphi_3=-\frac{\varphi}{2}$. Then Equation (\ref{zerocurrent}) yields
\begin{equation}
\tan\varphi_1=G\tan(\frac{\varphi}{2}).
\end{equation}
with $G=\frac{g-1}{g+1}$. The total energy reads :
\begin{equation}
\label{eq:energy_TJ}
E_{TJ}=E_{0}\Big [\cos\varphi - g_0(g+1)\frac{|\cos(\frac{\varphi}{2})|}{\sqrt{1+G^2\tan^2(\frac{\varphi}{2})}}\Big ].
\end{equation}
The variation of $\varphi_1$ with $\varphi$ (see Fig. \ref{fig:phenomenol_sans_flux}) shows that for a partially symmetric bijunction ($g=1$),
$\varphi_1$ jumps by $\pi$ each time $\varphi$ is an odd multiple of $\pi$. The energy profile of the bijunction is pictured on Fig.
\ref{fig:phenomenol_sans_flux}. Although it is $2\pi$-periodic, the plot between $-2\pi$ and $2\pi$ shows that, depending on the choice
of the minima modulo $2\pi$, the barrier between the equivalent minima can be different. Notice that if $g=1$, the $E_{TJ}(\varphi)$ curve
possesses a cusp at $\varphi=\pi$, but this cusp is rounded by any small asymmetry between leads $2$ and $3$ ($g\neq1$).
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE2.pdf}
\caption{(Color online) a-b)Two symmetric canted solutions in the frustrated situation, corresponding to a spontaneous phase vortex. The straight arrows represent the superconducting phase. c) non frustrated solution with two $\pi$-junctions ($S_1-S_2$ and $S_1-S_3$) and a $0$-junction ($S_2-S_3$).
\label{fig:phase_vortices}
}
\end{figure}
In the case $g=1$, the energy minimum corresponding to the canted solution satisfies $\cos\frac{\varphi}{2}=\pm\frac{g_0}{2}$. It spans from
$\varphi=\pi$ for $g_0=0$, corresponding to a single $\pi$-junction $S_2-S_3$ through the dot, to $\varphi=\frac{2\pi}{3}$ or $\frac{4\pi}{3}$
for $g_0=1$ (fully symmetric bijunction, Fig. \ref{fig:phase_vortices}a,b) and $\varphi=0$ for $g_0=2$ (Fig. \ref{fig:phase_vortices}c). If instead $g_0>2$, there is no canting
and the bijunction displays two $\pi$-junctions $S_2-S_1$ and $S_1-S_3$ in series, and a $0$-junction $S_2-S_3$ (Fig. \ref{fig:phase_vortices}c).
As an essential fact, in the canted case there are two equivalent solutions, obtained by changing $\varphi$ in $-\varphi$ or in $2\pi-\varphi$ (Fig. \ref{fig:phase_vortices}a,b). As shown in Section III, the choice of the minima and of the corresponding barrier can be monitored by an external flux.
\begin{figure}[htb]
\includegraphics[width=.8\columnwidth]{FIGURE3.pdf}
\caption{(Color online). (Top left) Variation of the "floating" phase $\varphi_1$ with the phase $\varphi$ across $S_2-S_3$, for $g=1$ (staircase) and $0.9$. Total energy of the bijunction inserted in a single loop, as a function of the phase $\varphi_{23}=\varphi$ for (top right) $g_0=1$, $g=1$; (middle left) $g_0=1$, $g=0.9$; (middle right) $g_0=1$, $g=0.7$; (bottom left) $g_0=0.5$, $g=0.9$; (bottom right) $g_0=1.5$, $g=0.9$. Despite the $2\pi$-periodicity, the plot between $\varphi=\pm 2\pi$ shows the different barriers depending on the couple of degenerate phase vortex states.
\label{fig:phenomenol_sans_flux}
}
\end{figure}
This simple calculation shows that i) frustration manifests itself in canting the phases from $0$ or $\pi$; ii) a doubly degenerate state is formed, with opposite phase vorticities; iii) a too asymmetric
bijunction does not sustain frustration, and yields two $\pi$-junctions and one $0$-junction. Phase vorticity appears as a spontaneous symmetry breaking, induced by the frustration brought by the existence of a localized spin creating $\pi$-junctions. The presence of the localized spin therefore induces a chirality in the bijunction.
\subsection{A microscopic model}
Let us now provide a nonperturbative calculation, describing the localized spin with the help of a local Zeeman (or exchange field $J$), as in Ref. \onlinecite{Benjamin}. This can be related to a model including the Coulomb interaction through a mean-field approximation. This excludes the possible formation of a Kondo state in the $0$-junction regime when the dot-lead couplings are large enough. The Hamiltonian of the system is $H=H_{S}+H_{D}+H_{T}$ where $H_{S}$, $H_{D}$ and $H_{T}$ respectively denote the lead, dot and lead-dot tunneling contributions. The dot part is written as:
\begin{equation}
H_D=E_0\sum_{s=\uparrow,\downarrow} d^{\dagger}_s d_s - J(d^{\dagger}_{\uparrow} d_{\uparrow}-d^{\dagger}_{\downarrow} d_{\downarrow}).
\end{equation}
where $E_0$ is the bare energy level. We assume that $E_0-J<0$ and $E_0+J>0$, such that for weak coupling to the leads, the dot level carries one electron with spin up. Writing $H$ in the Nambu notation $H=H_{S}+H_{D}+H_{T}$, and performing a gauge transformation to incorporate the superconducting phases $\varphi_j$ in the tunneling term $H_T$, one gets, up to an additive constant, the following expressions:
\begin{equation}\label{Hamilt}
H_S=\sum_{j=1,2,3}\sum_{k}
\Psi_{jk}^{\dagger}(\xi_{k}\sigma_{z}+\Delta_j \sigma_{x})
\Psi_{jk}, \Psi_{jk}=\left(\begin{array}{c} \psi_{jk,
\uparrow}\\
\psi_{j(-k),\downarrow}^{\dagger}
\end{array} \right)
\end{equation}
\begin{equation}
H_{D}=d^{\dagger}(E_{0} \sigma_{z}-J \sigma_0)d
\label{eq:dspinor}
\end{equation}
\begin{equation}
H_{T}=\sum_{jk}\Psi_{jk}^{\dagger} T_{j} d + h.c.,
\quad \quad d=\left(\begin{array}{c} d_{\uparrow}\\
d_{\downarrow}^{\dagger}
\end{array} \right),
\label{eq:dspinor}
\end{equation}
\noindent
with $ T_{j} =t_{j} \sigma_{z}e^{i\sigma_{z}\varphi_j/2}$ and $t_{j} $ is
the tunnelling amplitude between the lead $j$ and the dot. $\sigma_{0}$ is the identity matrix and $\sigma_{x,y,z}$ denote the Pauli matrices in the basis formed by electrons with spin $\uparrow$ and holes with spin $\downarrow$.
The procedure to obtain the Andreev bound states and the current-phase relationships by writing an effective action for the two dots can be found in Ref. [\onlinecite{Benjamin}]. One writes the partition function as
\begin{align}
Z = \int \mathcal{D} \left[ \bar{\psi},\psi , \bar{d}, d\right] e^{- S \left[ \bar{\psi},\psi , \bar{d}, d \right]} ,
\end{align}
e.g. as a functional integral over Grassmann fields for
the electronic degrees of freedom ($\Psi_{jk}, \bar\Psi_{jk}, d, \bar d$). The
Euclidean action reads:
\begin{equation}
S_A=S_{D}+\int_{0}^{\beta} \! d \tau
[\sum_{jk}{\bar\Psi_{jk}(\tau)}(\partial_{\tau} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})\Psi_{jk}(\tau)+{
H_{T}(\tau)}].
\end{equation}
\noindent $\beta$ is the inverse temperature, and
${H_{T}(\tau)}=\sum_{jk}{\bar\Psi_{jk}(\tau)}T_{j}d_{}(\tau)+h.c.$
while $S_{D}=\int_{0}^{\beta}\! d\tau [{\bar d(\tau)}(\partial_{\tau} \sigma_0+\epsilon_{}
\sigma_{z})d(\tau)]$.
After integrating out the leads we get $Z=\int\! \mathcal{D} \left[\bar{d}, d\right] \; e^{-S_{eff}}$
with
\begin{equation}
S_{eff}=S_{D}-\int_{0}^{\beta}\! d\tau \; d\tau' \;{\bar d(\tau)}{\check
\Sigma(\tau - \tau ')} d(\tau '),
\end{equation}
where $\check
\Sigma(\tau)=\sum_{j=1,2,3} T_{j}^{\dagger} G_{j}(\tau) T_{j}$ and
$G_{j}(\tau)=\sum_{k}
(\partial_{\tau} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})^{-1}
\delta(\tau)$.
We perform a Fourier transform on the Matsubara frequencies (with $\omega_{n}=(2n+1)\pi/\beta$):
$\delta(\tau)=\frac{1}{\beta} \sum_{\omega_n} e^{-i \omega_{n} \tau}$ and
$G(\tau)=\frac{1}{\beta} \sum_{\omega_n} e^{-i \omega_{n} \tau} G(i\omega_n)$, which gives for the Green function $G_j$:
\begin{eqnarray}
G_j(i\omega_n)=\int\! d \xi\; \nu(\xi)(-i
\omega_{n} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})^{-1}
\\
\nonumber
\simeq \frac{\pi
\nu(0)}{\sqrt{\Delta_j^2-(i\omega_n)^2}}(i\omega_{n} \sigma_0+\Delta_j \sigma_{x})
\end{eqnarray}
\noindent
Here $\nu(\xi)=\sum_{k}\delta(\xi-\xi_{k})$ is
approximated by a constant $\nu(0)$, the density of states at the
Fermi level in the normal leads. Let us assume for sake of simplicity the three gaps equal, $\Delta_j=\Delta$. One finally obtains the effective action
(introducing $d_{\alpha}(\tau)=\frac{1}{\sqrt{\beta}}\sum_{\omega_{n}} e^{-i\omega_{n} \tau} d_{\alpha}(i\omega_{n})$)
\begin{eqnarray} \nonumber
S_{eff}=&\sum_{\omega_{n}} \bar{\bf d}(\omega_{n}){\bf {\cal M}}(i\omega_n){\bf d}(i\omega_{n}) \\
{\bf {\cal M}}(i\omega_n) = &(-i \omega_{n}+J) \sigma_{0} +E_0
\sigma_{z}-{\check{ \bf \Sigma}_{i\omega_{n}}},
\end{eqnarray}
${\bf {\cal M}}(i\omega_n)$ is described by a $2$ x $2$ matrix, whose coefficients are given by
\begin{widetext}
\begin{equation}\label{Matrix}
\begin{aligned}
{\cal M}_{11}&=i\omega_n(1+\frac{\Gamma}{2\sqrt{\Delta^2-(i\omega_n)^2}})-E_0+J,\,\,\,\,\,\,
{\cal M}_{12}=-\frac{\Gamma \Delta}{2\sqrt{\Delta^2-(i\omega_n)^2}}(\sum_i\gamma_ie^{-i\varphi_i}),\\
{\cal M}_{21}&=-\frac{\Gamma \Delta}{2\sqrt{\Delta^2-(i\omega_n)^2}}(\sum_i\gamma_ie^{i\varphi_i}),\,\,\,\,\,\,
{\cal M}_{22}=i\omega_n(1+\frac{\Gamma}{2\sqrt{\Delta^2-(i\omega_n)^2}})+E_0+J,\\
\end{aligned}
\end{equation}
\end{widetext}
\noindent
with $\Gamma=2\pi \nu(0)\sum_i|t_i|^2$ and $\gamma_i=|t_i|^2/\sum_i|t_i|^2$. ${\bf {\cal M}}$ is an hermitian matrix once $i\omega_n$ is replaced by the real number $z$. The dispersion relation for the Andreev bound states is given by the eigenvalues of the effective action, replacing $i\omega_n$ by $z$. After integrating out the
$\{d,\bar d\}$ variables, the partition function is given by
\begin{equation}
Z=\int \! \mathcal{D} \left[ \bar{d}, d\right] \; e^{-S_{eff}}=\prod_{i\omega_n} \det
{\bf {\cal M}}(\omega_{n}).
\end{equation}
The free energy is given by:
\begin{equation}
F=-\frac{1}{\beta}\sum_{\omega_n}\ln(\det {\bf {\cal M}}(i\omega_n)).
\end{equation}
The Josephson current in $S_i$ is expressed as:
\begin{align}
\nonumber
I_{Ji}=-\frac{2e}{\hbar\beta}\frac{\partial}{\partial \varphi_{i}} \ln Z\\
=-\frac{2}{\beta}\frac{\partial}{\partial \varphi_{i}} \sum_{\omega_n} \ln
(\det {\bf {\cal M}}(i\omega_n))
\end{align}
Consider for simplicity the case of a bijunction symmetric by exchange of leads $2$ and $3$, e.g. $\gamma_2=\gamma_3$, to be compared
with the $g=1$ case of Section I. If the exchange field is sufficient to stabilize a local moment, one also finds a critical value of the asymmetry
$\frac{\gamma_1}{\gamma_{2,3}}$ above which frustration disappears and the bijunction is dominated by two $\pi$-junctions in series. In the
perturbative limit where $\Gamma$ is smaller than the single spin level $|E_0-J|$, one finds energy profiles $E_{TJ}(\varphi)$ similar to those of the
phenomenological model, with couplings $g_{ij}$ respectively proportionnal to $\gamma_i\gamma_j$. An example of an exact nonperturbative
solution is given in Fig. \ref{fig:exact}. More generally, the critical value of $J$ above which the $\pi$-junctions are stabilized is about
$\frac{J}{\Gamma}=0.5$. In this regime, because the $\pi$-junction is weaker than a $0$-junction, the perturbative calculation turns out to
be qualitatively correct, and the physics is well described by the phenomenological model.
\begin{figure}[htb]
\includegraphics[width=.6\columnwidth]{FIGURE4.pdf}
\caption{(Color online). Total energy of the bijunction, from the microscopic model, as a function of the phase $\varphi_{23}=\varphi$. Parameters are $\Delta=1$, $\Gamma=2$, $J=5$, $\varepsilon=0$, temperature $T=0.02$, and (a) $\gamma_{1,2,3}=(1/3,1/3,1/3)$, (b) $\gamma_{1,2,3}=(0.4,0.4,0.2)$.
\label{fig:exact}
}
\end{figure}
\section{bijunction in a circuit with loops}
Superconducting interference devices with embedded junctions can be used to measure their phase-current relation \cite{Dam, Cleuziou, Rocca}. These techniques imply inserting the junction in a multiple connected circuit. The above analysis shows that the presence of three $\pi$-junctions can create frustration and phase canting at the junctions. Having three superconducting reservoirs, the bijunction can be connected in various ways to an external circuit.
\subsection{Single loop}
Let us first consider here the simplest geometry obtained by connecting superconductors $S_2$ and $S_3$ by a loop, leaving superconductor $S_1$ disconnected. This implies that the phase $\varphi=\varphi_2-\varphi_3$ is accessible and controllable experimentally, while the "floating" phase $\varphi_1$ is determined by the condition $J_1=0$ (Equation \ref{zerocurrent}). Let us denote by $L$ the loop inductance, $\Phi_{ext}$ the external flux, and $LI$ the flux embedded in the loop, induced by the current $I$. Expressing flux quantification along the circuit enclosing the loop and passing through the dot yields :
\begin{equation}
\varphi=\frac{2\pi}{\Phi_0}\Phi [2\pi],
\end{equation}
where $\Phi=\Phi_{ext}+LI$ is the total embedded flux and $\Phi_0=hc/2e$ is the elementary flux quantum. The total energy becomes:
\begin{equation}
\label{eq:one loop}
E_{TJ1L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}(\varphi-\varphi_{ext})^2+E_{TJ}(\varphi)
\end{equation}
where $E_{TJ}$ is given by Equation \ref{eq:energy_TJ}, and with $\varphi_{ext}=\frac{2\pi}{\Phi_0}\Phi_{ext}$.
Consider first $\Phi_{ext}=0$. Then, if $L<L_c$, the only stable solution is $\Phi=0$, and there is no equilibrium phase difference at the
junction. Conversely, if $L>L_c$, a spontaneous flux appears in the loop, together with a phase difference $\varphi'_m$ at the junction
(Fig. \ref{fig:phenomenol_avec_flux}). When $LI_c>>\phi_0$ ($I_c$ is the critical current of the bijunction),
$\Phi\simeq\pm\frac{\Phi_0}{2\pi}\varphi_m$ thus a large loop stabilizes the two vortex solutions found in Section I. If on the contrary
$\Phi_{ext}=\Phi_0/2$, the loop stabilizes the solutions $\Phi\simeq\frac{\Phi_0}{2\pi}\varphi_m, \frac{\Phi_0}{2\pi}(2\pi-\varphi_m)$. These
two sets of solutions are equivalent, but the barrier between the two degenerate mimima are different. For a given set of the parameters
$g_0,g$, the highest barrier is encountered for one or the other of the applied fluxes. On the other hand, if $\Phi_{ext}$ is not a multiple of
$\Phi_0/2$, the minima are not equivalent (Fig. \ref{fig:phenomenol_avec_flux}). Fig. \ref{fig:phenomenol_avec_flux} shows that
one may keep the two minima at fixed values, say ($\pm \varphi_m$), and vary the asymmetry parameter $g_0$
(Fig. \ref{fig:phenomenol_avec_flux}, top panels), thus changing the barrier between the two minima. More interestingly, one can keep the same
bijunction parameters and change the flux from $\Phi_{ext}=0$ to $\Phi_{ext}=\Phi_0/2$ (Fig. \ref{fig:phenomenol_avec_flux}, left panels,
or right panels). This switches the pair of minima from ($\pm \varphi_m$) to ($\varphi_m$, $2\pi-\varphi_m$), with a strong change of the
barrier between them. Depending on whether $g_0$ is smaller or larger than $1$, the barrier may be decreased or increased.
This might offer a way of manipulating the pair of vortex solutions as a phase qubit, by tuning from three to two energy minima. Actually, tuning the flux between $\Phi_{ext}=0$ and $\Phi_{ext}=\Phi_0/2$ keeps two of the three states equally probable but allow to switch on or off the tunneling between them. On the other hand, fixing the flux to a value such as $\Phi_{ext}=\pm\Phi_0/4$ favours one or the other minima.
The above discussion shows that for a moderate asymmetry and a large inductance, this set-up allows a spontaneous current/flux to appear in the loop. Contrarily to the simple $\pi-$ junction where only a flux $\Phi_0/2$ can be stabilized, here the induced flux can take any value between $0$ and $\Phi_0$.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE5.pdf}
\caption{(Color online).Connecting the bijunction with one loop can stabilize an arbitrary flux. $J$ (blue arrow)
denotes the current flowing in junction $S_2-S_3$.
\label{fig:one_loop}
}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=0.8\columnwidth]{FIGURE6.pdf}
\caption{(Color online). Total energy of the bijunction inserted in a single loop, as a function of the phase $\varphi_{23}=\varphi$. The asymmetry betwen leads $2$ and $3$ is weak. Parameters are (left panels) $g_0=0.5$ and (right panels) $g_0=1.5$, and from top to bottom $\Phi_{ext}=0, \Phi_0/4, \Phi_0/2$.
\label{fig:phenomenol_avec_flux}
}
\end{figure}
\subsection{Two loops}
Let us now connect the bijunction by two loops, by closing for instance the junctions $S_1-S_2$ and $S_1-S_3$ (Fig. \ref{fig:two_loop}). Those loops respectively enclose fluxes $\Phi_{ext,3}$ and $\Phi_{ext,2}$. The quantification condition for each of the loops are:
\begin{equation}
\varphi_{12}=\frac{2\pi}{\Phi_0}\Phi_{3} [2\pi], \;\;\varphi_{13}=-\frac{2\pi}{\Phi_0}\Phi_{2} [2\pi],
\end{equation}
where $\Phi_{2}=\Phi_{ext,2}+LI_{13}$ and $\Phi_{3}=\Phi_{ext,3}-LI_{12}$ are the total embedded flux. Defining $\varphi_{ext,2}=\frac{2\pi}{\Phi_0}\Phi_{ext,2}$, $\varphi_{ext,3}=\frac{2\pi}{\Phi_0}\Phi_{ext,3}$, the total energy then reads :
\begin{equation}
\begin{aligned}
E_{TJ2L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}\Big[(\varphi_{12}-\varphi_{ext,3})^2
+(\varphi_{13}+\varphi_{ext,2})^2\Big]\\
+E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos(\varphi_{13}-\varphi_{12})].
\end{aligned}
\end{equation}
If $L$ is large, minimizing with respect to $\varphi_{12}$, $\varphi_{13}$ gives in the frustrated regime the two symmetric vortex
solutions of Section I, which induce nonzero but equal fluxes in the loops. The fluxes can take the values
$\Phi_2=\Phi_3\simeq\frac{\Phi_0}{2\pi}\varphi_m$, or $\Phi_2=\Phi_3\simeq-\frac{\Phi_0}{2\pi}\varphi_m$ (Fig. \ref{fig:two_loop}).
In the symmetric junction case, fluxes $(\Phi_0/3,\Phi_0/3)$ or $(-\Phi_0/3,-\Phi_0/3)$ can be stabilized. Those flux can be made
dissymetric either by acting on the junction parameters (with gates) of with en external flux.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE7.pdf}
\caption{(Color online).Connecting the bijunction with two loops stabilizes two symmetric spontaneous fluxes. The Figure corresponds to zero external flux, large inductance and symmetric bijunction. The blue arrows denote the currents circulating in the junctions.
\label{fig:two_loop}
}
\end{figure}
\subsection{Three loops}
Finally, the bijunction can be more symmetrically closed by three loops, each embedding an external flux $\Phi_{ext,i}$ (Fig. \ref{fig:three_loops}). Thus ($i=1,2,3)$:
\begin{equation}
\varphi_{ij}=\frac{2\pi}{\Phi_0}\Phi_{k}\delta_{ijk} [2\pi], \;\;\Phi_{k}=\Phi_{ext,k}+L_{ij}\delta_{ijk},
\end{equation}
where $\delta_{ijk}$ is $1$ if all i,j,k are different, and zero otherwise.
The total energy reads :
\begin{equation}
\begin{aligned}
E_{TJ3L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}\sum_{ij}(\varphi_{ij}-\varphi_{ext,k}\delta_{ijk})^2\\
+E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos(\varphi_{13}-\varphi_{12})].
\end{aligned}
\end{equation}
Large $L$ either yields $\Phi_1=\Phi_2=\Phi_3\simeq\frac{\Phi_0}{2\pi}\varphi_m$, or $\Phi_1=\Phi_2=\Phi_3\simeq-\frac{\Phi_0}{2\pi}\varphi_m$. For instance, in the fully symmetric case, each loop carries one third of the flux quantum $\Phi_0$. A property of the three-loop configuration is that it globally embeds one flux quantum, $\Phi_1+\Phi_2+\Phi_3=\pm\Phi_0$. This is a direct manifestation of the phase vorticity induced by frustration.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE8.pdf}
\caption{(Color online).Connecting the bijunction with three loops globally traps one flux quantum.
\label{fig:three_loops}
}
\end{figure}
\section{conclusion}
We have shown that a quantum dot carrying a $1/2$ spin, thus able to generate a Josephson $\pi$ junction, may induce frustration if inserted in a Josephson bijunction. If the coupling between two superconductors - say $S_2$ and $S_3$ - dominates, this results in the $S_2-S_3$ junction being a $\pi$-junction, while low asymmetry leads to frustration and canting of the equilibrium phases. On the opposite, a too small coupling between $S_2$ and $S_3$ results in both $S_1-S_2$ and $S_1-S_3$ being $\pi$-junctions, making $S_2-S_3$ an effective $0$-junction. The possibility of continuously tuning the junction $S_2-S_3$ between a $0$- and a $\pi$-junction is a first result of this work.
This phenomenon displays an interesting link between two kins of magnetism : spin magnetism in the dot, and orbital magnetism manifesting in spontaneous flux (vortex). It is remarkable that this is a topological property, related to the existence of a localized spin, and not to its direction. In fact the direction of the localized spin and the sign of the stabilized vortex are unrelated. This could be different in more complicated situations involving an additional spin-orbit coupling.
The second result is that frustration generates two equivalent states possessing opposite phase vorticities, each of them breaking time-reversal symmetry. These states can be revealed by inserting the bijunction in a set-up containing one, two or three loops. In the case of a single loop, the two phase vortex states result in a spontanous flux crossing the loop, which is different from $0$ or $\pi$. While a zero external flux, or a multiple of $\Phi_0/2$, preserve the symmetry of the two vortex states, any other value lifts the degeneracy and can be used to stabilize one or the other of these two states.
This might have some consequences in terms of using the above device for generating flux qubits ot flux qutrits \cite{Clarke,Langford,Buisson}. Indeed, in the two-loop scheme one may control the two distinct phases by the external fluxes and the bijunction parameters. Tunneling through the barrier separating the two states can be strongly varied if acting on, say, the coupling between lead $1$ and the dot, by split gates for instance. Quantum fluctuations of the trapped fluxes occur if the lead-dot junctions have finite capacitances, for instance if the reservoirs are Cooper pair boxes. Control of the "longitudinal" and "transversal" components of this flux qubit is thus possible, as a basic ingredient for applications. Further investigations must be carried out to derive an effective qubit model and check its feasibility.
The authors acknowledge the support of PICS CNRS-CONICET 5755, PICT 2010-1060 from ANPCyT, PIP 11220080101821 from CONICET, Argentina, grant 06-C400 from UNCuyo and the Laboratoire Franco-Argentin en Nanosciences (LIFAN).
\section{Introduction}
Junctions with confined electrons, like atomic, molecular or quantum-dot (QD) junctions, are among the most studied nanoscopic devices \cite{Andergassen}. In these structures electron-electron correlations within the junction together with the electronic properties of the contacts lead to a number of different phenomena concerning fundamental aspects of quantum charge and spin transport.
Modern technologies allowed building junctions that are close to the ultimate limit of miniaturization with normal, superconducting \cite{Yu, Morpurgo, Buitelaar, Doh, Jorgensen, Xiang, Jarillo} or ferromagnetic \cite{Ferro1, Ferro2} leads creating new opportunities for novel nanodevices with predefined functional properties. One-electron transistors, spin valves or superconducting spin qubits are some examples of such devices.
When a molecule or a QD bridges the gap between two metallic leads, the Coulomb energy tends to quantize the charge confined in the junctions, i.e. on the molecule or the dot, leading to the possibility of fabricating a junction with a confined spin. In the case of superconducting leads, the characteristic of the so-obtained Josephson junction depends on the total spin in the junctions.
Superconducting circuits with QD (S-QD-S junctions) have been extensively studied during the last decade \cite{Levy_Yeyati_Review,De_Franceschi_Review}. These junctions are fabricated by contacting superconducting leads to a normal nanostructure, typically a single walled carbon nanotube or a semiconducting nanowire. The structure may include gate electrodes that can be used to control the number of electrons in the dot.
S-QD-S junctions having a localized spin in the dot may have a global minimum of the free energy for a $\pi$-difference between the phases of the two superconducting contacts . The current-phase characteristic of these junctions is described by a Josephson equation with a {\it negative} critical current \cite{pi_junction1, pi_junction2}. These junctions, referred to as $\pi$-junctions, in contrast with standard "$0$-junctions", have interesting properties and potential applications in superconducting electronics, including phase or flux qubits \cite{Makhlin_Review}.
A superconducting ring containing a $\pi$-junction could generate a spontaneous current with (nearly) half a superconducting quantum flux threading the ring, a very convenient situation for experimental detection. Notice that this current structure is stable only if the ring self-inductance exceeds a critical value\cite{Sigrist_Rice,Jagla}.
Here we present results for a $\pi$-bijunction consisting of a QD connected to three superconducting leads, see Fig. \ref{fig:Triterminal_dot}. Graphene dots (GQD) are good candidates to build such a device\cite{graph1, graph2}. In fact graphene offers new opportunities for superconducting electronics as a new class of material that can be tailored and contacted to normal or superconducting leads. With gate electrodes controlling the number of electrons confined in the GQD, a non-zero spin can be localized at the dot which tends to generate a $\pi$-shift between each of the pairs of superconductors. This situation, in a way similar to Heisenberg magnets, generates frustration\cite{Sigrist_Rice}.The frustration can be resolved by canting the phases of the three superconductors, in a way that depends on the asymmetry of the device. This asymmetry, due to different couplings between the dot and the contacts, can be controlled by gates. Frustration leads to canting only for moderate asymmetries. In the canted phase, the equilibrium phase difference between two given superconductors can be controlled at will to a value $\varphi_m$, between $0$ and $\pi$. Such a tunable $\varphi$-junction can be probed by various geometries incorporating one or several superconducting loops. An important feature of the phase canted state is that it contains a spontaneous vorticity. Due to time-reversal symmetry, two equivalent solutions with opposite vorticities are found, corresponding to phases $\varphi_m$ and $-\varphi_m$. This symmetry can be broken with the help of a single loop and an applied magnetic field. In addition, the structure of the energy-phase profile of the bijunction makes the barrier between the two degenerate minima tunable, either through the bijunction parameters, or using the external flux. This might be an useful property for building a superconducting qubit.
The paper is organized as follows. In Section 2 a phenomenological phase model is first considered, with a stability analysis of the canted (frustrated) solution against the non-frustrated solution. Then a microscopic model for a dot with a single level allows a nonperturbative solution, which essentially confirms the existence of a canted phase below a critical asymmetry of the dot-contact couplings. Section 3 considers a single loop set-up, with an applied orbital magnetic field, then two-loop or three-loop set-ups.
\section{bijunction at equilibrium}
\subsection{A phenomenological model}
The quantum dot connects all three superconductors (Fig. \ref{fig:Triterminal_dot}). Each pair $(i,j)$ of the three superconductors ($i=1,2,3$) forms a Josephson junction. As a first approximation, one can write the total energy of the bijunction as the sum of those of separated junctions. Such an expression could be obtained in perturbation theory from a microscopic Hamiltonian, at fourth order in the tunnelling element between the superconductor states and the dot states. The bijunction is then equivalent to a triangular array of separated junctions. We assume that the presence of a $1/2$ spin on the dot creates $\pi$-junctions, and that this holds for all of them. The nonperturbative calculation presented in the next subsection shows that it is essentially the case, unless the couplings to the dot are very asymmetric. Denoting the superconducting phases as $\varphi_i$ ($i=1-3$), the bijunction energy thus reads :
\begin{equation}
\label{eq:energy_BJ_red}
E_{BJ}=E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos\varphi_{23}],
\end{equation}
with $\varphi_{ij}=\varphi_i-\varphi_j$, $E_0>0$ and where $g_0\geq0$, $g\geq0$ are parameters quantifying the bijunction asymmetry (Fig. \ref{fig:Triterminal_dot}).
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE1.pdf}
\caption{(Color online) Bijunction made of three superconductors and one quantum dot carrying a spin $S=1/2$. Inset show the equivalent triangular model, valid in the perturbative regime only. The asymmetry ratios between the junctions critical currents are indicated.
\label{fig:Triterminal_dot}
}
\end{figure}
Let us look for the equilibrium state. Setting to zero the partial derivatives of $E_{BJ}$ with respect to the $\varphi_i$'s is equivalent to imposing zero current $J_i$ in each lead $S_i$. One obtains from $J_1=0$ :
\begin{equation}
\label{zerocurrent}
g\sin\varphi_{12}+\sin\varphi_{13}=0,
\end{equation}
which, together with similar equations expressing that $J_2$ or $J_3=0$, yields :
\begin{equation}
\label{canting}
\cos\varphi_{23}=\frac{(g_0g)^2-1-g^2}{2g}.
\end{equation}
Such a nontrivial solution thus exists only if $|1-\frac{1}{g}|\leq g_0 \leq (1+\frac{1}{g})$. This is a canted (e.g. frustrated) phase solution (Fig. \ref{fig:phase_vortices}a,b), with two degenerate states obtained from each other by changing $\varphi_{2,3}$ into $-\varphi_{2,3}$. In such states, the current across any junction $S_i-S_j$ is nonzero. Yet, the total current in each lead is zero. Those two degenerate solutions therefore feature a phase vortex, with two opposite vorticities. While in a real triangular network this vorticity is associated to a circulating current, with a zero-dimensional quantum dot, it is difficult to define a path with a nonzero current circulating around the dot. Nevertheless, we show in the last Section that a true vortex can be induced on an adjacent loop.
In the opposite case $|(g_0g)^2-1-g^2|\geq 2g$, the energy minimum is obtained for $\varphi_2=0$ or $\pi$, $\varphi_3=0$ or $\pi$, replacing Eq. (\ref{canting}). This results in two of the three junctions being $\pi$-junctions and the other one a $0$-junction (Fig. \ref{fig:phase_vortices}c).
Later on we consider situations where the lead $1$ is disconnected (e.g. the phase $\varphi_1$ is floating), while leads $2$ and $3$ are connected to an external circuit. Then the convenient phase variable is $\varphi=\varphi_2-\varphi_3$. One can use gauge invariance and choose $\varphi_2=\frac{\varphi}{2}$, $\varphi_3=-\frac{\varphi}{2}$. Then Equation (\ref{zerocurrent}) yields
\begin{equation}
\tan\varphi_1=G\tan(\frac{\varphi}{2}).
\end{equation}
with $G=\frac{g-1}{g+1}$. The total energy reads :
\begin{equation}
\label{eq:energy_TJ}
E_{TJ}=E_{0}\Big [\cos\varphi - g_0(g+1)\frac{|\cos(\frac{\varphi}{2})|}{\sqrt{1+G^2\tan^2(\frac{\varphi}{2})}}\Big ].
\end{equation}
The variation of $\varphi_1$ with $\varphi$ (see Fig. \ref{fig:phenomenol_sans_flux}) shows that for a partially symmetric bijunction ($g=1$),
$\varphi_1$ jumps by $\pi$ each time $\varphi$ is an odd multiple of $\pi$. The energy profile of the bijunction is pictured on Fig.
\ref{fig:phenomenol_sans_flux}. Although it is $2\pi$-periodic, the plot between $-2\pi$ and $2\pi$ shows that, depending on the choice
of the minima modulo $2\pi$, the barrier between the equivalent minima can be different. Notice that if $g=1$, the $E_{TJ}(\varphi)$ curve
possesses a cusp at $\varphi=\pi$, but this cusp is rounded by any small asymmetry between leads $2$ and $3$ ($g\neq1$).
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE2.pdf}
\caption{(Color online) a-b)Two symmetric canted solutions in the frustrated situation, corresponding to a spontaneous phase vortex. The straight arrows represent the superconducting phase. c) non frustrated solution with two $\pi$-junctions ($S_1-S_2$ and $S_1-S_3$) and a $0$-junction ($S_2-S_3$).
\label{fig:phase_vortices}
}
\end{figure}
In the case $g=1$, the energy minimum corresponding to the canted solution satisfies $\cos\frac{\varphi}{2}=\pm\frac{g_0}{2}$. It spans from
$\varphi=\pi$ for $g_0=0$, corresponding to a single $\pi$-junction $S_2-S_3$ through the dot, to $\varphi=\frac{2\pi}{3}$ or $\frac{4\pi}{3}$
for $g_0=1$ (fully symmetric bijunction, Fig. \ref{fig:phase_vortices}a,b) and $\varphi=0$ for $g_0=2$ (Fig. \ref{fig:phase_vortices}c). If instead $g_0>2$, there is no canting
and the bijunction displays two $\pi$-junctions $S_2-S_1$ and $S_1-S_3$ in series, and a $0$-junction $S_2-S_3$ (Fig. \ref{fig:phase_vortices}c).
As an essential fact, in the canted case there are two equivalent solutions, obtained by changing $\varphi$ in $-\varphi$ or in $2\pi-\varphi$ (Fig. \ref{fig:phase_vortices}a,b). As shown in Section III, the choice of the minima and of the corresponding barrier can be monitored by an external flux.
\begin{figure}[htb]
\includegraphics[width=.8\columnwidth]{FIGURE3.pdf}
\caption{(Color online). (Top left) Variation of the "floating" phase $\varphi_1$ with the phase $\varphi$ across $S_2-S_3$, for $g=1$ (staircase) and $0.9$. Total energy of the bijunction inserted in a single loop, as a function of the phase $\varphi_{23}=\varphi$ for (top right) $g_0=1$, $g=1$; (middle left) $g_0=1$, $g=0.9$; (middle right) $g_0=1$, $g=0.7$; (bottom left) $g_0=0.5$, $g=0.9$; (bottom right) $g_0=1.5$, $g=0.9$. Despite the $2\pi$-periodicity, the plot between $\varphi=\pm 2\pi$ shows the different barriers depending on the couple of degenerate phase vortex states.
\label{fig:phenomenol_sans_flux}
}
\end{figure}
This simple calculation shows that i) frustration manifests itself in canting the phases from $0$ or $\pi$; ii) a doubly degenerate state is formed, with opposite phase vorticities; iii) a too asymmetric
bijunction does not sustain frustration, and yields two $\pi$-junctions and one $0$-junction. Phase vorticity appears as a spontaneous symmetry breaking, induced by the frustration brought by the existence of a localized spin creating $\pi$-junctions. The presence of the localized spin therefore induces a chirality in the bijunction.
\subsection{A microscopic model}
Let us now provide a nonperturbative calculation, describing the localized spin with the help of a local Zeeman (or exchange field $J$), as in Ref. \onlinecite{Benjamin}. This can be related to a model including the Coulomb interaction through a mean-field approximation. This excludes the possible formation of a Kondo state in the $0$-junction regime when the dot-lead couplings are large enough. The Hamiltonian of the system is $H=H_{S}+H_{D}+H_{T}$ where $H_{S}$, $H_{D}$ and $H_{T}$ respectively denote the lead, dot and lead-dot tunneling contributions. The dot part is written as:
\begin{equation}
H_D=E_0\sum_{s=\uparrow,\downarrow} d^{\dagger}_s d_s - J(d^{\dagger}_{\uparrow} d_{\uparrow}-d^{\dagger}_{\downarrow} d_{\downarrow}).
\end{equation}
where $E_0$ is the bare energy level. We assume that $E_0-J<0$ and $E_0+J>0$, such that for weak coupling to the leads, the dot level carries one electron with spin up. Writing $H$ in the Nambu notation $H=H_{S}+H_{D}+H_{T}$, and performing a gauge transformation to incorporate the superconducting phases $\varphi_j$ in the tunneling term $H_T$, one gets, up to an additive constant, the following expressions:
\begin{equation}\label{Hamilt}
H_S=\sum_{j=1,2,3}\sum_{k}
\Psi_{jk}^{\dagger}(\xi_{k}\sigma_{z}+\Delta_j \sigma_{x})
\Psi_{jk}, \Psi_{jk}=\left(\begin{array}{c} \psi_{jk,
\uparrow}\\
\psi_{j(-k),\downarrow}^{\dagger}
\end{array} \right)
\end{equation}
\begin{equation}
H_{D}=d^{\dagger}(E_{0} \sigma_{z}-J \sigma_0)d
\label{eq:dspinor}
\end{equation}
\begin{equation}
H_{T}=\sum_{jk}\Psi_{jk}^{\dagger} T_{j} d + h.c.,
\quad \quad d=\left(\begin{array}{c} d_{\uparrow}\\
d_{\downarrow}^{\dagger}
\end{array} \right),
\label{eq:dspinor}
\end{equation}
\noindent
with $ T_{j} =t_{j} \sigma_{z}e^{i\sigma_{z}\varphi_j/2}$ and $t_{j} $ is
the tunnelling amplitude between the lead $j$ and the dot. $\sigma_{0}$ is the identity matrix and $\sigma_{x,y,z}$ denote the Pauli matrices in the basis formed by electrons with spin $\uparrow$ and holes with spin $\downarrow$.
The procedure to obtain the Andreev bound states and the current-phase relationships by writing an effective action for the two dots can be found in Ref. [\onlinecite{Benjamin}]. One writes the partition function as
\begin{align}
Z = \int \mathcal{D} \left[ \bar{\psi},\psi , \bar{d}, d\right] e^{- S \left[ \bar{\psi},\psi , \bar{d}, d \right]} ,
\end{align}
e.g. as a functional integral over Grassmann fields for
the electronic degrees of freedom ($\Psi_{jk}, \bar\Psi_{jk}, d, \bar d$). The
Euclidean action reads:
\begin{equation}
S_A=S_{D}+\int_{0}^{\beta} \! d \tau
[\sum_{jk}{\bar\Psi_{jk}(\tau)}(\partial_{\tau} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})\Psi_{jk}(\tau)+{
H_{T}(\tau)}].
\end{equation}
\noindent $\beta$ is the inverse temperature, and
${H_{T}(\tau)}=\sum_{jk}{\bar\Psi_{jk}(\tau)}T_{j}d_{}(\tau)+h.c.$
while $S_{D}=\int_{0}^{\beta}\! d\tau [{\bar d(\tau)}(\partial_{\tau} \sigma_0+\epsilon_{}
\sigma_{z})d(\tau)]$.
After integrating out the leads we get $Z=\int\! \mathcal{D} \left[\bar{d}, d\right] \; e^{-S_{eff}}$
with
\begin{equation}
S_{eff}=S_{D}-\int_{0}^{\beta}\! d\tau \; d\tau' \;{\bar d(\tau)}{\check
\Sigma(\tau - \tau ')} d(\tau '),
\end{equation}
where $\check
\Sigma(\tau)=\sum_{j=1,2,3} T_{j}^{\dagger} G_{j}(\tau) T_{j}$ and
$G_{j}(\tau)=\sum_{k}
(\partial_{\tau} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})^{-1}
\delta(\tau)$.
We perform a Fourier transform on the Matsubara frequencies (with $\omega_{n}=(2n+1)\pi/\beta$):
$\delta(\tau)=\frac{1}{\beta} \sum_{\omega_n} e^{-i \omega_{n} \tau}$ and
$G(\tau)=\frac{1}{\beta} \sum_{\omega_n} e^{-i \omega_{n} \tau} G(i\omega_n)$, which gives for the Green function $G_j$:
\begin{eqnarray}
G_j(i\omega_n)=\int\! d \xi\; \nu(\xi)(-i
\omega_{n} \sigma_0+\xi_{k}\sigma_{z}+\Delta_j\sigma_{x})^{-1}
\\
\nonumber
\simeq \frac{\pi
\nu(0)}{\sqrt{\Delta_j^2-(i\omega_n)^2}}(i\omega_{n} \sigma_0+\Delta_j \sigma_{x})
\end{eqnarray}
\noindent
Here $\nu(\xi)=\sum_{k}\delta(\xi-\xi_{k})$ is
approximated by a constant $\nu(0)$, the density of states at the
Fermi level in the normal leads. Let us assume for sake of simplicity the three gaps equal, $\Delta_j=\Delta$. One finally obtains the effective action
(introducing $d_{\alpha}(\tau)=\frac{1}{\sqrt{\beta}}\sum_{\omega_{n}} e^{-i\omega_{n} \tau} d_{\alpha}(i\omega_{n})$)
\begin{eqnarray} \nonumber
S_{eff}=&\sum_{\omega_{n}} \bar{\bf d}(\omega_{n}){\bf {\cal M}}(i\omega_n){\bf d}(i\omega_{n}) \\
{\bf {\cal M}}(i\omega_n) = &(-i \omega_{n}+J) \sigma_{0} +E_0
\sigma_{z}-{\check{ \bf \Sigma}_{i\omega_{n}}},
\end{eqnarray}
${\bf {\cal M}}(i\omega_n)$ is described by a $2$ x $2$ matrix, whose coefficients are given by
\begin{widetext}
\begin{equation}\label{Matrix}
\begin{aligned}
{\cal M}_{11}&=i\omega_n(1+\frac{\Gamma}{2\sqrt{\Delta^2-(i\omega_n)^2}})-E_0+J,\,\,\,\,\,\,
{\cal M}_{12}=-\frac{\Gamma \Delta}{2\sqrt{\Delta^2-(i\omega_n)^2}}(\sum_i\gamma_ie^{-i\varphi_i}),\\
{\cal M}_{21}&=-\frac{\Gamma \Delta}{2\sqrt{\Delta^2-(i\omega_n)^2}}(\sum_i\gamma_ie^{i\varphi_i}),\,\,\,\,\,\,
{\cal M}_{22}=i\omega_n(1+\frac{\Gamma}{2\sqrt{\Delta^2-(i\omega_n)^2}})+E_0+J,\\
\end{aligned}
\end{equation}
\end{widetext}
\noindent
with $\Gamma=2\pi \nu(0)\sum_i|t_i|^2$ and $\gamma_i=|t_i|^2/\sum_i|t_i|^2$. ${\bf {\cal M}}$ is an hermitian matrix once $i\omega_n$ is replaced by the real number $z$. The dispersion relation for the Andreev bound states is given by the eigenvalues of the effective action, replacing $i\omega_n$ by $z$. After integrating out the
$\{d,\bar d\}$ variables, the partition function is given by
\begin{equation}
Z=\int \! \mathcal{D} \left[ \bar{d}, d\right] \; e^{-S_{eff}}=\prod_{i\omega_n} \det
{\bf {\cal M}}(\omega_{n}).
\end{equation}
The free energy is given by:
\begin{equation}
F=-\frac{1}{\beta}\sum_{\omega_n}\ln(\det {\bf {\cal M}}(i\omega_n)).
\end{equation}
The Josephson current in $S_i$ is expressed as:
\begin{align}
\nonumber
I_{Ji}=-\frac{2e}{\hbar\beta}\frac{\partial}{\partial \varphi_{i}} \ln Z\\
=-\frac{2}{\beta}\frac{\partial}{\partial \varphi_{i}} \sum_{\omega_n} \ln
(\det {\bf {\cal M}}(i\omega_n))
\end{align}
Consider for simplicity the case of a bijunction symmetric by exchange of leads $2$ and $3$, e.g. $\gamma_2=\gamma_3$, to be compared
with the $g=1$ case of Section I. If the exchange field is sufficient to stabilize a local moment, one also finds a critical value of the asymmetry
$\frac{\gamma_1}{\gamma_{2,3}}$ above which frustration disappears and the bijunction is dominated by two $\pi$-junctions in series. In the
perturbative limit where $\Gamma$ is smaller than the single spin level $|E_0-J|$, one finds energy profiles $E_{TJ}(\varphi)$ similar to those of the
phenomenological model, with couplings $g_{ij}$ respectively proportionnal to $\gamma_i\gamma_j$. An example of an exact nonperturbative
solution is given in Fig. \ref{fig:exact}. More generally, the critical value of $J$ above which the $\pi$-junctions are stabilized is about
$\frac{J}{\Gamma}=0.5$. In this regime, because the $\pi$-junction is weaker than a $0$-junction, the perturbative calculation turns out to
be qualitatively correct, and the physics is well described by the phenomenological model.
\begin{figure}[htb]
\includegraphics[width=.6\columnwidth]{FIGURE4.pdf}
\caption{(Color online). Total energy of the bijunction, from the microscopic model, as a function of the phase $\varphi_{23}=\varphi$. Parameters are $\Delta=1$, $\Gamma=2$, $J=5$, $\varepsilon=0$, temperature $T=0.02$, and (a) $\gamma_{1,2,3}=(1/3,1/3,1/3)$, (b) $\gamma_{1,2,3}=(0.4,0.4,0.2)$.
\label{fig:exact}
}
\end{figure}
\section{bijunction in a circuit with loops}
Superconducting interference devices with embedded junctions can be used to measure their phase-current relation \cite{Dam, Cleuziou, Rocca}. These techniques imply inserting the junction in a multiple connected circuit. The above analysis shows that the presence of three $\pi$-junctions can create frustration and phase canting at the junctions. Having three superconducting reservoirs, the bijunction can be connected in various ways to an external circuit.
\subsection{Single loop}
Let us first consider here the simplest geometry obtained by connecting superconductors $S_2$ and $S_3$ by a loop, leaving superconductor $S_1$ disconnected. This implies that the phase $\varphi=\varphi_2-\varphi_3$ is accessible and controllable experimentally, while the "floating" phase $\varphi_1$ is determined by the condition $J_1=0$ (Equation \ref{zerocurrent}). Let us denote by $L$ the loop inductance, $\Phi_{ext}$ the external flux, and $LI$ the flux embedded in the loop, induced by the current $I$. Expressing flux quantification along the circuit enclosing the loop and passing through the dot yields :
\begin{equation}
\varphi=\frac{2\pi}{\Phi_0}\Phi [2\pi],
\end{equation}
where $\Phi=\Phi_{ext}+LI$ is the total embedded flux and $\Phi_0=hc/2e$ is the elementary flux quantum. The total energy becomes:
\begin{equation}
\label{eq:one loop}
E_{TJ1L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}(\varphi-\varphi_{ext})^2+E_{TJ}(\varphi)
\end{equation}
where $E_{TJ}$ is given by Equation \ref{eq:energy_TJ}, and with $\varphi_{ext}=\frac{2\pi}{\Phi_0}\Phi_{ext}$.
Consider first $\Phi_{ext}=0$. Then, if $L<L_c$, the only stable solution is $\Phi=0$, and there is no equilibrium phase difference at the
junction. Conversely, if $L>L_c$, a spontaneous flux appears in the loop, together with a phase difference $\varphi'_m$ at the junction
(Fig. \ref{fig:phenomenol_avec_flux}). When $LI_c>>\phi_0$ ($I_c$ is the critical current of the bijunction),
$\Phi\simeq\pm\frac{\Phi_0}{2\pi}\varphi_m$ thus a large loop stabilizes the two vortex solutions found in Section I. If on the contrary
$\Phi_{ext}=\Phi_0/2$, the loop stabilizes the solutions $\Phi\simeq\frac{\Phi_0}{2\pi}\varphi_m, \frac{\Phi_0}{2\pi}(2\pi-\varphi_m)$. These
two sets of solutions are equivalent, but the barrier between the two degenerate mimima are different. For a given set of the parameters
$g_0,g$, the highest barrier is encountered for one or the other of the applied fluxes. On the other hand, if $\Phi_{ext}$ is not a multiple of
$\Phi_0/2$, the minima are not equivalent (Fig. \ref{fig:phenomenol_avec_flux}). Fig. \ref{fig:phenomenol_avec_flux} shows that
one may keep the two minima at fixed values, say ($\pm \varphi_m$), and vary the asymmetry parameter $g_0$
(Fig. \ref{fig:phenomenol_avec_flux}, top panels), thus changing the barrier between the two minima. More interestingly, one can keep the same
bijunction parameters and change the flux from $\Phi_{ext}=0$ to $\Phi_{ext}=\Phi_0/2$ (Fig. \ref{fig:phenomenol_avec_flux}, left panels,
or right panels). This switches the pair of minima from ($\pm \varphi_m$) to ($\varphi_m$, $2\pi-\varphi_m$), with a strong change of the
barrier between them. Depending on whether $g_0$ is smaller or larger than $1$, the barrier may be decreased or increased.
This might offer a way of manipulating the pair of vortex solutions as a phase qubit, by tuning from three to two energy minima. Actually, tuning the flux between $\Phi_{ext}=0$ and $\Phi_{ext}=\Phi_0/2$ keeps two of the three states equally probable but allow to switch on or off the tunneling between them. On the other hand, fixing the flux to a value such as $\Phi_{ext}=\pm\Phi_0/4$ favours one or the other minima.
The above discussion shows that for a moderate asymmetry and a large inductance, this set-up allows a spontaneous current/flux to appear in the loop. Contrarily to the simple $\pi-$ junction where only a flux $\Phi_0/2$ can be stabilized, here the induced flux can take any value between $0$ and $\Phi_0$.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE5.pdf}
\caption{(Color online).Connecting the bijunction with one loop can stabilize an arbitrary flux. $J$ (blue arrow)
denotes the current flowing in junction $S_2-S_3$.
\label{fig:one_loop}
}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=0.8\columnwidth]{FIGURE6.pdf}
\caption{(Color online). Total energy of the bijunction inserted in a single loop, as a function of the phase $\varphi_{23}=\varphi$. The asymmetry betwen leads $2$ and $3$ is weak. Parameters are (left panels) $g_0=0.5$ and (right panels) $g_0=1.5$, and from top to bottom $\Phi_{ext}=0, \Phi_0/4, \Phi_0/2$.
\label{fig:phenomenol_avec_flux}
}
\end{figure}
\subsection{Two loops}
Let us now connect the bijunction by two loops, by closing for instance the junctions $S_1-S_2$ and $S_1-S_3$ (Fig. \ref{fig:two_loop}). Those loops respectively enclose fluxes $\Phi_{ext,3}$ and $\Phi_{ext,2}$. The quantification condition for each of the loops are:
\begin{equation}
\varphi_{12}=\frac{2\pi}{\Phi_0}\Phi_{3} [2\pi], \;\;\varphi_{13}=-\frac{2\pi}{\Phi_0}\Phi_{2} [2\pi],
\end{equation}
where $\Phi_{2}=\Phi_{ext,2}+LI_{13}$ and $\Phi_{3}=\Phi_{ext,3}-LI_{12}$ are the total embedded flux. Defining $\varphi_{ext,2}=\frac{2\pi}{\Phi_0}\Phi_{ext,2}$, $\varphi_{ext,3}=\frac{2\pi}{\Phi_0}\Phi_{ext,3}$, the total energy then reads :
\begin{equation}
\begin{aligned}
E_{TJ2L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}\Big[(\varphi_{12}-\varphi_{ext,3})^2
+(\varphi_{13}+\varphi_{ext,2})^2\Big]\\
+E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos(\varphi_{13}-\varphi_{12})].
\end{aligned}
\end{equation}
If $L$ is large, minimizing with respect to $\varphi_{12}$, $\varphi_{13}$ gives in the frustrated regime the two symmetric vortex
solutions of Section I, which induce nonzero but equal fluxes in the loops. The fluxes can take the values
$\Phi_2=\Phi_3\simeq\frac{\Phi_0}{2\pi}\varphi_m$, or $\Phi_2=\Phi_3\simeq-\frac{\Phi_0}{2\pi}\varphi_m$ (Fig. \ref{fig:two_loop}).
In the symmetric junction case, fluxes $(\Phi_0/3,\Phi_0/3)$ or $(-\Phi_0/3,-\Phi_0/3)$ can be stabilized. Those flux can be made
dissymetric either by acting on the junction parameters (with gates) of with en external flux.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE7.pdf}
\caption{(Color online).Connecting the bijunction with two loops stabilizes two symmetric spontaneous fluxes. The Figure corresponds to zero external flux, large inductance and symmetric bijunction. The blue arrows denote the currents circulating in the junctions.
\label{fig:two_loop}
}
\end{figure}
\subsection{Three loops}
Finally, the bijunction can be more symmetrically closed by three loops, each embedding an external flux $\Phi_{ext,i}$ (Fig. \ref{fig:three_loops}). Thus ($i=1,2,3)$:
\begin{equation}
\varphi_{ij}=\frac{2\pi}{\Phi_0}\Phi_{k}\delta_{ijk} [2\pi], \;\;\Phi_{k}=\Phi_{ext,k}+L_{ij}\delta_{ijk},
\end{equation}
where $\delta_{ijk}$ is $1$ if all i,j,k are different, and zero otherwise.
The total energy reads :
\begin{equation}
\begin{aligned}
E_{TJ3L}(\varphi)=\frac{\Phi_0^2}{8\pi^2L}\sum_{ij}(\varphi_{ij}-\varphi_{ext,k}\delta_{ijk})^2\\
+E_{0}[g_0g\cos\varphi_{12}+g_0\cos\varphi_{13}+\cos(\varphi_{13}-\varphi_{12})].
\end{aligned}
\end{equation}
Large $L$ either yields $\Phi_1=\Phi_2=\Phi_3\simeq\frac{\Phi_0}{2\pi}\varphi_m$, or $\Phi_1=\Phi_2=\Phi_3\simeq-\frac{\Phi_0}{2\pi}\varphi_m$. For instance, in the fully symmetric case, each loop carries one third of the flux quantum $\Phi_0$. A property of the three-loop configuration is that it globally embeds one flux quantum, $\Phi_1+\Phi_2+\Phi_3=\pm\Phi_0$. This is a direct manifestation of the phase vorticity induced by frustration.
\begin{figure}[htb]
\includegraphics[width=1.0\columnwidth]{FIGURE8.pdf}
\caption{(Color online).Connecting the bijunction with three loops globally traps one flux quantum.
\label{fig:three_loops}
}
\end{figure}
\section{conclusion}
We have shown that a quantum dot carrying a $1/2$ spin, thus able to generate a Josephson $\pi$ junction, may induce frustration if inserted in a Josephson bijunction. If the coupling between two superconductors - say $S_2$ and $S_3$ - dominates, this results in the $S_2-S_3$ junction being a $\pi$-junction, while low asymmetry leads to frustration and canting of the equilibrium phases. On the opposite, a too small coupling between $S_2$ and $S_3$ results in both $S_1-S_2$ and $S_1-S_3$ being $\pi$-junctions, making $S_2-S_3$ an effective $0$-junction. The possibility of continuously tuning the junction $S_2-S_3$ between a $0$- and a $\pi$-junction is a first result of this work.
This phenomenon displays an interesting link between two kins of magnetism : spin magnetism in the dot, and orbital magnetism manifesting in spontaneous flux (vortex). It is remarkable that this is a topological property, related to the existence of a localized spin, and not to its direction. In fact the direction of the localized spin and the sign of the stabilized vortex are unrelated. This could be different in more complicated situations involving an additional spin-orbit coupling.
The second result is that frustration generates two equivalent states possessing opposite phase vorticities, each of them breaking time-reversal symmetry. These states can be revealed by inserting the bijunction in a set-up containing one, two or three loops. In the case of a single loop, the two phase vortex states result in a spontanous flux crossing the loop, which is different from $0$ or $\pi$. While a zero external flux, or a multiple of $\Phi_0/2$, preserve the symmetry of the two vortex states, any other value lifts the degeneracy and can be used to stabilize one or the other of these two states.
This might have some consequences in terms of using the above device for generating flux qubits ot flux qutrits \cite{Clarke,Langford,Buisson}. Indeed, in the two-loop scheme one may control the two distinct phases by the external fluxes and the bijunction parameters. Tunneling through the barrier separating the two states can be strongly varied if acting on, say, the coupling between lead $1$ and the dot, by split gates for instance. Quantum fluctuations of the trapped fluxes occur if the lead-dot junctions have finite capacitances, for instance if the reservoirs are Cooper pair boxes. Control of the "longitudinal" and "transversal" components of this flux qubit is thus possible, as a basic ingredient for applications. Further investigations must be carried out to derive an effective qubit model and check its feasibility.
The authors acknowledge the support of PICS CNRS-CONICET 5755, PICT 2010-1060 from ANPCyT, PIP 11220080101821 from CONICET, Argentina, grant 06-C400 from UNCuyo and the Laboratoire Franco-Argentin en Nanosciences (LIFAN).
|
\section{\label{sec:intro}Introduction}
\FloatBarrier Self-accelerating wave packets freely accelerate even without any external potential present. This intriguing phenomenon is of rapidly
growing interest since its advent in optics in 2007 \cite{Siv07a, Siv07, Ban08, Min11, Zha13, Kam13}. The most prominent example of self-accelerating
waves has been introduced by Siviloglou and coworkers \cite{Siv07, Siv07a}: They demonstrated, that an Airy-type wave packet exhibits a linear
transversal acceleration and is therefore following a parabolic trajectory.
Enlarging the scope and the versatility of Airy beams, nonparaxial generalizations in terms of full vectorial solutions of Maxwell's equations \cite{Kam12} as well as by the method of caustics \cite{Cou12} were investigated.
Their curved trajectories render classical Airy beams a powerful tool in many areas of application. For instance, in the field of particle manipulation, micro beads have been
guided and sorted in a new fashion \cite{Bau08} beyond the scope of classical optical tweezers. Moreover, it was shown that curved plasma channels
have many advantages over their straight counterparts, e.g., when it comes to the spatially resolved detection of secondary signals \cite{Pol09}.
Additionally, Airy wave packets inspired excessive fundamental research in the field of nonlinear optics \cite{Kam11,Lot11,Dol12,Bek11}, and boosted
the study of waves with intensity maxima that propagate along almost arbitrary trajectories \cite{Gre11}. Broadening the range of influence beyond
the scope of optics, Airy beams have been utilized in electron beam shaping \cite{Vol13} as well.
\begin{figure}[b]
\includegraphics[width=\linewidth, page=1]{figures}
\caption{Illustrative presentation showing the accelerative behavior of Airy beams (a) and radially self-accelerating field\,/\,intensity distributions (b).}
\label{fig:fig1}
\end{figure}
One common feature of all of the aforementioned waves - even in two-dimensional settings - is that they accelerate linearly, namely along a
specific Cartesian coordinate axis (see Fig.~\ref{fig:fig1}~(a)). This obvious limitation brings about a number of fundamental questions: Is it
possible to generate optical beams that show a self-accelerating behavior along different types of trajectories? Could such wave packets be
shape-invariant or even non-diffractive? Based on the numerous already existing applications of Airy-type beams discussed in the previous paragraph,
it should be obvious that beams for which the aforementioned questions
can be answered affirmatively would enrich the optical toolbox in many areas of
application and research. Moreover, from a fundamental point of view, it is essential to determine under which kind of approximations analytical
solutions for such beams can be found. In other words, will those solutions be restricted to the paraxial case or do they obey the scalar Helmholtz equation
or even Maxwell's equations?
In the present article, we report on a new class of self-accelerating diffraction-free waves that move along three-dimensional spiraling
trajectories. As such, they behave as if they were influenced by a radially symmetric external potential even though the propagation takes place in
free space. Observed from a rotating, co-moving frame of reference like the one depicted in Fig.~\ref{fig:fig1}~(b), the beams we present are
propagation-invariant. Within a thorough theoretical discourse we will derive a general, explicit, and analytical expression for this new class of
beams. We will show the great versatility of these beams, for which - while outperforming Airy-type beams - the transverse cross-section is highly
tunable after fixing predefined spiraling properties. In addition, we are going to verify our theoretical findings on an experimental basis utilizing
an intuitive understanding borrowed from Fourier-optics. In this regard, we present a simple yet powerful setup that enables one to address the
entire parameter range of the presented beams.
Our theoretical section is comprised of two parts. The first one consists of an extensive theoretical derivation regarding the most general expression of a beam which exhibits a field pattern that is invariant in a rotating frame of reference. As it will be discussed later, of more practical interest might be the implementation of beams with a rotating intensity distribution. For this reason, in the second part of our theory section, we pose conditions on the beam intensity only (i.e., no condition is posed on the phases) finding a more general class of beams.
First, we want to model a beam that is propagation-invariant in a co-moving, rotating frame. This wave is supposed to be a solution to the scalar Helmholtz
equation $\Delta E +k^2 E=0$, where $E$ is a scalar electric field and $k=\nicefrac{2\pi}{\lambda}$ the corresponding wave number. Since we are
dealing with rotating solutions, it is a natural choice to work with cylindrical coordinates. Then, the most general solution of the scalar Helmholtz
equation can be written as \begin{align} E(r,\varphi,z)=\sum_{n=-\infty}^{\infty}\int_0^{\infty}d\alpha C_n(\alpha)J_n(\alpha r)e^{i(n\varphi+\beta
z)}, \label{eq:generalSuperpos}
\end{align}
which is essentially a superposition of fundamental eigenmodes given in terms of diffraction-free Bessel waves. The spatial structure of each eigenmode is determined by $J_n(\alpha r)e^{i(n\varphi+\beta z)}$, where $J_n(\alpha r)$ represents the Bessel function of order $n$ and
$\beta=\sqrt{k^2-\alpha^2}$ is the longitudinal component of the wave vector, or propagation constant. For an arbitrary beam, the expansion coefficients $C_n(\alpha)$ are arbitrary as well. In the following, we will derive conditions for $C_n(\alpha)$ in order to obtain rotating self-similar solutions to the scalar Helmholtz equation. Note, that we restrict our analysis to beams that are propagating in the positive $z$-direction.
For a beam that is self-accelerating three major properties need to be fulfilled. First, no external potential or non-linear optical effect should be present. Second, the beam is diffraction-free in a certain frame of reference. Finally, an observer resting in the aforementioned reference frame
would experience a fictitious force. The first condition is fulfilled immediately as we start our analysis from the linear and time-independent scalar Helmholtz equation. For the second requirement, a coordinate transformation needs to exist with which the field distribution is no longer dependent
on the propagation direction. It can easily be shown that an electric field of the general form
\begin{align}
E(r,\varphi,z)\overset{!}{=} E(r,\varphi+\omega z) \label{eq:condRot}
\end{align}
fulfils this condition. Obviously, with the substitution $\varphi^\prime=\varphi+\omega z$ the field $E(r,\varphi^\prime)$ in Eq.~\eqref{eq:condRot}
is no longer dependent on the longitudinal position $z$ and thus remains unchanged for every $z$. Moreover, the aforementioned coordinate
transformation describes a reference frame which is rotating with an angular velocity $\omega$. As a consequence, the last requirement is satisfied,
as an observer resting in this rotating frame of reference experiences a centrifugal force.
\newcommand*{\mathrm{sign}}{\mathrm{sign}}
Since Eq.~\eqref{eq:condRot} has to hold for every $\varphi$ and $z$, from Eq.~\eqref{eq:generalSuperpos} it immediately follows that
\begin{align}
\nicefrac{\beta}{n}\overset{!}{=}\omega. \label{eq:cond1}
\end{align}
As $\beta$ was restricted to be positive, the first conclusion from condition~\eqref{eq:cond1} is that the signs of $n$ and $\omega$ have to be
equal, i.e., $\mathrm{sign}(n)=\mathrm{sign}(\omega)$. Moreover, since the propagation constant $\beta$ is a function of the transverse component of the wave vector
$\alpha$, condition~\eqref{eq:cond1} can be rewritten as \begin{align} \alpha \overset{!}{=} \alpha_n = \sqrt{k^2-\omega^2n^2} \label{eq:cond2}
\end{align}
Obviously, this restriction can only be fulfilled for the specific choice of coefficients
\begin{align}
C_n(\alpha)=\widetilde{C}_n\,\delta(\alpha-\alpha_n).\label{eq:Cndelta}
\end{align}
Applying the restrictions on $\mathrm{sign}(n)$ as well as on $C_n(\alpha$), Eq.~\eqref{eq:generalSuperpos} becomes
\begin{align}
E(r,\varphi,z)=\sum_{n=1}^{n_{\textrm{max}}} \widetilde{C}_n J_n(\alpha_n r)e^{i(\mathrm{sign}(\omega)n(\varphi+\omega z))}. \label{eq:Superpos}
\end{align}
This is the most general expression of a beam that rotates in a shape-invariant fashion with an angular velocity $\omega$. Note, that $n_{\textrm{max}}=\max\{n\in \mathbb{N}:\, k^2>\omega^2n^2 \}$ in order to ensure that evanescent waves are excluded from the sum. To give an intuitive description of this finding, it is helpful to consider the Fourier-transform of this specific field. In essence, the Fourier-transform is a discrete superposition of concentric rings with radius $\alpha_n$ whereas the amplitude of these rings is given by the coefficients $\widetilde{C}_n$. Note that for a given $\omega$ Eq.~\eqref{eq:Cndelta} states that for each order $n$ there is exactly one ring radius $\alpha_n$. Moreover, each ring of order $n$ carries a helical phase pitch of $2\pi n$.
The field pattern of a beam described by Eq.~\eqref{eq:Superpos} gives rise to screw-shaped trajectories, that show a non-degenerate periodicity in the azimuthal and the propagation direction. \autoref{fig:example_field} shows an appropriate example using four Bessel waves ranging from order $1$ to $4$, i.e., $\tilde{C}_{n}=1$ for $1\leq n \leq 4$ and $\tilde{C}_{n}=0$ for $n>4$. The depicted insets emphasize the fact that amplitude and phase are rotating synchronously as predicted. The angular frequency spectrum consists of four concentric rings with radii determined by Eq.~\eqref{eq:cond2}.
\begin{figure}
\centering
\includegraphics[width=\linewidth,page=5]{figures}
\caption{Exemplary illustration of a radially self-accelerating field distribution with $\tilde{C}_{n}=1$ for $1\leq n\leq4$ and $\tilde{C}_{n}=0$
for $n>4$. The figure consists of 1-dimensional representation of the superimposed Bessel functions (main plot), resulting intensity distribution
(upper inset row) and resulting phase pattern (lower inset row).} \label{fig:example_field}
\end{figure}
At the beginning of the theory section we indicated that for many practical applications, such as optical tweezing or micro-fabrication, only the intensity distribution of a beam might be of interest. For this reason, we want to state how the requirements for the beam profile change if condition \eqref{eq:condRot} is posed on the intensity, i.e., $I(r,\varphi,z)=\vert E(r,\varphi,z)\vert^2=I(r,\varphi+\omega z)$. It will be shown that this scenario is more general and offers a larger degree of freedom. Consequently, in our subsequent discussion (including our experimental section) we will exclusively concentrate on this more general case. As the derivation of the following results is somewhat more lengthy and does not convey much additional physical insight, it is contained in the Supplementary Material. However, from this derivation, one arrives at a constrain similar to \eqref{eq:cond2}, which reads
\begin{align}
\alpha_n=\sqrt{k^2-(\omega\,n+\beta_0)^2}. \label{eq:cond_I}
\end{align}
Moreover, under these conditions the field is given by
\begin{align}
E(r,\varphi,z)=e^{i\beta_0z}\sum_{n\in \mathcal{N}} \widetilde{C}_n J_n(\alpha_n r)e^{i(n(\varphi+\omega z))}. \label{eq:Superpos_I}
\end{align}
Note that there are two main differences between Eq.~\eqref{eq:Superpos_I} and Eq.~\eqref{eq:Superpos}. The first difference is that
Eq.~\eqref{eq:Superpos_I} contains the global phase factor $e^{i\beta_0z}$ with the propagation constant $\beta_0$, which can be regarded as a free
parameter for these beams. Consequently, one can already see that the field given by Eq.~\eqref{eq:Superpos_I} does not fulfill the requirement of a
rotation invariant field anymore (only the intensity is rotation invariant). It is important to be aware of the fact that $\beta_0$ does not only
determine the global phase factor but poses an important degree of freedom for scaling the transverse beam properties after fixing the rotation parameter $\omega$. This becomes apparent when comparing Eq.~\eqref{eq:cond2} and Eq.~\eqref{eq:cond_I}. The second difference is that the sum in Eq.~\eqref{eq:Superpos_I} contains also negative $n$, whereas Eq.~\eqref{eq:Superpos} covers only positive $n$. To be specific, the set $\mathcal{N}=\{n\in\mathbb{Z}:k^2>(\omega\,n+\beta_0)^2\}$ contains all integer numbers $n$ for which Eq.~\eqref{eq:cond_I} yields real values for
$\alpha_n$.
Please note that rotating intensity distributions based on pairs of superimposed Bessel-functions have been investigated theoretically and experimentally \cite{Pat96,Ter01, Vas09, Rop12, Rop12a, Ruf12,McG03,Paa98,Cha98}. Our theory, however, describes rotating field\,/\,intensity patterns that are composed of an arbitrary number of Bessel waves. The arising enlarged degree of freedom renders useful when modeling specific beam properties as will be discussed in the upcoming paragraph. \autoref{fig:example_intensity} shows an exemplary beam with $\tilde{C}_{n}=1$ for $-1\leq
n\leq +1$ and $\tilde{C}_{n}=0$ for $\left|n\right|>1$. The depicted insets demonstrate that the intensity distribution is indeed rotating during propagation while the corresponding phase is no longer synchronized. The angular frequency spectrum of the shown beam consists of three concentric
rings with radii that are determined by Eq.~\eqref{eq:cond_I}.
\begin{figure}
\centering
\includegraphics[width=\linewidth, page=4]{figures}
\caption{Exemplary illustration of a radially self-accelerating intensity distribution with $\tilde{C}_{n}=1$ for $-1\leq n\leq +1$ and
$\tilde{C}_{n}=0$ for $\left|n\right|>1$. The figure consists of 1-dimensional representation of the superimposed Bessel functions (main plot), resulting
intensity distribution (upper inset row) and resulting phase pattern (lower inset row).} \label{fig:example_intensity}
\end{figure}
An important, yet open question is how versatile the transverse cross-section of beams described by Eq.~\eqref{eq:Superpos_I} can be tailored. In
order to answer this question consider Eq.~\eqref{eq:Superpos_I} in the initial plane ($z=0$). For a given distance $R$ from the origin of the
coordinate system, Eq.~\eqref{eq:Superpos_I} can be written as
\begin{align}
E(R,\varphi,0)=\sum_{n\in\mathcal{N}} D_n e^{in\varphi}, \label{eq:Superpos_inital}
\end{align}
where $D_n=\widetilde{C}_n J_n(\alpha_n R)$. If the distance $R$ is chosen such that $J_n(\alpha_n R)\neq 0$, then Eq.~\eqref{eq:Superpos_inital}
represents a general Fourier-series. As an important consequence, the transverse beam profile can be tailored in a way that the field distribution
on a circle with radius $R$ can be chosen arbitrarily, in other words $E(R,\varphi,z=0)=f(\varphi)$, where the complex function $f(\varphi)$ can be
set without any condition. From this it follows that a cooking recipe to tailor these kind of rotating beams could be to fix the parameter $\beta_0$
as well as the function $f(\varphi)$. Then Eq.~\eqref{eq:cond_I} determines $\alpha_n$ and the expansion coefficients are given by
\begin{align}
\widetilde{C}_n =\frac{1}{2\pi J_n(\alpha_n R)}\int_{0}^{2\pi}f(\varphi)e^{-in\varphi}\mathrm{d}\varphi.
\end{align}
Note that as soon as the field is specified for one radius $R$, the entire field in the transverse plane is determined. This is simply due to the
fixed radial dependence of the Bessel functions.
In order to experimentally implement our findings, we make use of the fact that the presented beams show a multi-ring pattern with distinct helical
phase pitch in the angular frequency domain. Fourier transforming this pattern by means of a conventional lens will match the previously discussed
theory. For the experimental setup, different approaches are conceivable ranging from the use of axicons (conical lenses), ring slit apertures and
phase plates to the exclusive use of spatial-light-modulators (SLMs). In this work, we followed this last approach as it provides the highest amount
of flexibility. Our setup is presented in Fig.~\ref{fig:setup} and makes use of a technique introduced in Ref. \cite{Dav99}. This technique enables
simultaneous amplitude and phase modulation with a single phase-only SLM by multiplying the desired amplitude distribution with a blazed grating. In
our case this desired amplitude distribution is given by concentric rings - in other words, as previously mentioned, we implement the
Fourier-transform of the desired beam in the SLM-plane. After the Fourier-transforming lens, undesired grating orders are filtered by a pin-hole and
the primary signal is imaged by an additional 4f-setup. Finally, a movable CCD-camera allows to measure the change of the intensity profile in
propagation direction.
\begin{figure}
\includegraphics[width=\linewidth, page=2]{figures}
\caption{Experimental setup containing a telescope for beam expansion, SLM (Holoeye Pluto VIS) for amplitude and phase modulation, lens for Fourier
transformation ($f=300\,\mathrm{mm}$), aperture and 4f-arrangement ($f_1=f_2=200\,\mathrm{mm}$) for signal cleaning as well as a movable CCD Unit (Basler
Ace1600-20gm with Olympus Plan N $20\times$) for data acquisition.} \label{fig:setup}
\end{figure}
With the proposed setup, we are able to cover almost the entire parameter range provided by our theory. This of course would go beyond the scope of this letter. We will therefore present an exemplary set of parameters upon which we show the practicability of radially self-accelerating beams and discuss the experimental limitations. One of these limitations is already given by the fact that Bessel beams cannot be created to their full extend as they would carry an infinite amount of energy and require a non-finite aperture. This is also true for the presented radially self-accelerating beams, since they are a specific discrete superposition of Bessel beams. Another limitation arises from the fact that the Fourier-transform of a Bessel wave is a ring with infinitesimal thickness. In an experimental setting it is clear that only rings with finite width
can be generated. As an immediate consequence, the range over which the experimentally generated beam resembles the theoretical prediction will be limited.
For experimental realization the set of parameters presented in Fig. 3 was used.
For this purpose, a superposition of three rings was implemented in the SLM-plane. The subsequent Fourier-lens with focal length $f$ connects the ring radii $R_n$ on the SLM with the transverse components of the wave vector $\alpha_n$ via
\begin{align}
\alpha_n=\frac{k R_n}{f}.\label{equ:radii}
\end{align}
Fig.~\ref{fig:measurement} shows an experimental scan along the propagation direction together with a simulation based on the analytical solution
described by Eq.~\eqref{eq:Superpos_I}. In total, we were able to observe about two rotations over a length of $101.5\,\mathrm{mm}$. See Supplemental
Material at [URL will be inserted by publisher] for an animated representation of the full scan. The rotation rate was found to be
$\omega_{\mathrm{exp}}=\left(123.2\pm 2.4\right)\,\mathrm{\frac{rad}{m}}$ and is therefore in very good agreement with the intended value of
$\omega_{\mathrm{th}}=125\,\mathrm{\frac{rad}{m}}$.
\begin{figure}
\centering
\includegraphics[width=\linewidth, page=3]{figures}
\caption{Simulation (top) and experimental results (bottom) showing the propagation dynamics of a radially self-accelerating beam identical to the
one in Fig. \ref{fig:example_intensity}. Radii on the SLM have been $R_1=2.328\,\mathrm{mm}$, $R_2=1.958\,\mathrm{mm}$ and $R_3=1.501\,\mathrm{mm}$.}
\label{fig:measurement}
\end{figure}
In conclusion, we have demonstrated the existence of a new class of self-accelerating waves. Theoretically, those waves, which accelerate freely on
spiraling trajectories, were derived as solutions to the scalar Helmholtz equation. It was pointed out, that radially self-accelerating beams can be
generated as a discrete superposition of Bessel waves with well defined properties. As such, they are quasi-non-diffractive, meaning that they are
diffraction-free in a rotating, co-moving frame of reference. With the proposed experimental setup the study of beam properties under realistic
conditions was shown to be possible with great flexibility. In a first proof of principle experiment, it was verified that the beam shows indeed the
desired rotating behavior - yielding excellent agreement with the theoretical predictions - and, moreover, the transverse beam profile remains
unbroadened for a substantially long propagation distance.
We foresee a broad range of applications for this new class of radially accelerating beams ranging from particle
manipulation, e.g. as tractor beams, to material processing, e.g. photo lithography. In order to widen the range of possible applications even
further, it is also of interest to study the properties of these kind of beams in more detail on a fundamental basis. For instance, with particle
manipulation in mind, it is of great interest to investigate the self-healing behavior or the dynamics in random media. Regarding material
processing, for example, the dynamics of such field configurations in non-linear environments is of great importance.
\begin{acknowledgments}
The authors wish to thank the German Ministry of Education and Research (Center for Innovation Competence program, grant 03Z1HN31), the Thuringian Ministry for Education, Science and Culture (Research group Spacetime, grant no. 11027-514), the Deutsche Forschungsgemeinschaft (grant NO462/6- 1), and the German-Israeli Foundation for Scientific Research and Development (grant 1157-127.14/2011).
\end{acknowledgments}
\FloatBarrier
|
\section{Introduction}
\label{intro}
In literature, Darcy's equation is considered as law of hydrodynamics in porous media. This linear relation between the fluid velocity and pressure gradient was derived by Darcy in \cite{Darcybook}.
However, even in their early works, Darcy and Dupuit already observed deviations of fluid flows from this linear equation \cite{Darcybook, Dupuit1857}.
In early 1900s, Forchheimer proposed three models for nonlinear flows, the so-called two-term, three-term and power laws (cf. \cite{Forchh1901,ForchheimerBook}) . More experiments gave rise to more nonlinear models during 1940-60s (cf. \cite{BearBook,Nieldbook}). Despite that fact, most mathematical papers on fluids in porous media deal only with Darcy's law starting from 1960s. The mathematical investigations of Forchheimer equations and related Brinkman-Forchheimer equation started much later in 1990s and have been growing ever since (cf. \cite{ChadamQin,Qin1998,Payne1999b,Payne1999a,Franchi2003,CKU2005,CKU2006,Straughan2013}, see also \cite{Straughanbook}). Even so, most of these papers consider incompressible fluids only. In our previous works \cite{ABHI1,HI1,HI2,HIKS1},
we proposed and studied generalized Forchheimer equations for slightly compressible fluids in porous media. Such mathematical generalization is appropriate and useful. This is due to the nature of Forchheimer equations which are derived from experiments and have their physical parameters found by fitting real life data.
From mathematical point of view, it introduces a new class of degenerate parabolic equations into studies of porous medium equations \cite{VazquezPorousBook}.
In our mentioned papers, we study the properties of pressure in space $L^\alpha$ ($1\le \alpha<\infty$), of pressure gradient in space $L^{2-a}$ and of pressure's time derivative in $L^2$. Here $a$ is a number between $0$ and $1$ defined in terms the degree of the polynomial in the generalized Forchheimer equations (see \eqref{ab}).
In this paper, we focus on properties of fluid flows in higher regularity spaces.
Specifically, we will study the pressure and its time derivative in space $L^\infty$, the pressure gradient in $L^s$ for any $1\le s<\infty$ and the pressure Hessian in $L^{2-\delta}$ for $\delta>0$.
Moreover, our high priority is the long-time dynamical properties, including uniform estimates in time, asymptotic bounds and asymptotic stability.
Such topics of long-time dynamics of degenerate parabolic equation, particularly in $L^\infty$, is important, and the specific results are usually hard to obtain.
(See, for e.g., \cite{VazquezSmoothBook}, chapters 18--20 of \cite{VazquezPorousBook} for porous medium equations, \cite{RVV2009,RVV2010} for degenerate equations with Dirichlet boundary condition, \cite{LD2000} for systems.)
In order to work in these much higher regularity spaces and obtain estimates for large time, more sophisticated techniques are called for.
We combine iteration techniques by De Giorgi \cite{DeGiorgi57} and Ladyzhenskaya-Uraltseva \cite{LadyParaBook68}, which were primarily used for studying local properties of solutions to elliptic and parabolic problems, with those from long-time dynamics studies for nonlinear partial differential equations such as Navier-Stokes equations \cite{FMRTbook}. For our degenerate equations, we also use and refine relevant techniques in DiBenedetto's book \cite{DiDegenerateBook}.
Such a combination gives fruitful results on the estimates of solutions for large time as well as detailed continuous dependence of the solutions on time-dependent boundary data and coefficients of the Forchheimer polynomials. (The latter results are called structural stability.)
We also emphasize that the mentioned general techniques from parabolic equations must be used in accordance with the structure of our equation, in this case, the important monotonicity and perturbed monotonicity in Lemma \ref{quasimono-lem} below.
In the current work, we focus on the case of Degree Condition, see \eqref{deg} in the next section, for the following reasons.
First, it already covers most commonly used Forchheimer equations in practice, namely, the two-term, three-term and power laws.
Second, to take advantage of available estimates in our previous work \cite{HI2}.
Third, to make clear our ideas and techniques without involving much more complicated technical details in case that the Degree Condition is not met (see \cite{HIKS1}); such case will be investigated in our future work.
The paper is organized as follows.
\par
In section \ref{prelim}, we present the formulation of generalized Forchheimer equations and consequently obtain a degenerate parabolic equation for pressure $p$. Basic properties of this equation are reviewed and suitable parabolic Sobolev embeddings are presented.
\par
In section \ref{supestimate}, we study the initial boundary value problem (IBVP) for pressure with the time-dependent flux boundary data $\psi(x,t)$.
We derive various estimates for the shifted solution $\bar p$ (see \eqref{pbargam}).
This section is divided into four subsections.
\par
Subsection \ref{pressuresubsec} deals with $L^\infty$-estimates up to the boundary for $\bar p$.
In Theorem \ref{thm-1}, bounds are established for all time, large time (independent of initial data) and time infinity.
The precise estimates for both small and large data are achieved by using Lemma \ref{multiseq} in the Appendix -- a generalization of the classical fast geometric convergence of sequences of numbers.
\par
Subsection \ref{Lsgrad-sec} deals with interior $L^s$-estimates for $\nabla p$ for any $s\ge 1$, see Theorems \ref{cor311} and \ref{cor312}.
The key Sobolev embedding is Lemma \ref{LUK} with specific weight $K(|\nabla w|)$ related to our degenerate structure.
The main iteration step is \eqref{Kgrad1}.
\par
Subsection \ref{Lpt-sec} deals with interior $L^\infty$-estimates for $\bar p_t$, where De Giorgi's technique is applied to the time derivative with weighted parabolic Sobolev embedding.
The main estimates are in Theorem \ref{ptbar}, other particular large time and asymptotic estimates are in Theorems \ref{ptbar2} and \ref{smallqbar}.
\par
Subsection \ref{seconderiv} deals with interior $L^{2-\delta}$-estimates with $\delta\in(0,a]$ for Hessian $\nabla^2 p$.
The estimates in Theorems \ref{hessest}, \ref{HessThm3} and \ref{HessLim} for second derivatives were not considered in our previous works.
\par
\par
Sections \ref{datacont} and \ref{polycont} are devoted to the structural stability issues.
\par
In section \ref{datacont}, we prove the continuous dependence of the solutions on the boundary data. Specifically, it is established for $\bar p$
in interior $L^\infty$-norms (see Theorem \ref{thm46}) and for $\nabla p$ in interior $L^{2-\delta}$-norms (see Theorems \ref{GradThm1} and \ref{lem412} for finite time intervals, and Theorems \ref{newP}, \ref{newthmAbar2} and \ref{GradThm2} for $t\to\infty$).
The results show that even when each individual flux $\psi_1,\psi_2$ grows unbounded as time $t\to\infty$, the difference between two corresponding solutions $\bar p_1,\bar p_2$ can be small provided the difference $\Psi=\psi_1-\psi_2$ is small.
In order to obtain this, De Giorgi's iteration is combined, in Proposition \ref{cont:strictcond}, with the monotonicity available for our degenerate equation.
\par
In section \ref{polycont}, we prove the continuous dependence of the solutions on the Forchheimer polynomials (see Theorems \ref{thm54}, \ref{thmsupP}, \ref{GradThm3} and \ref{lastthm}).
Here, the perturbed monotonicity is combined with De Giorgi's technique as presented in Proposition \ref{theo49}.
It is proved in Theorems \ref{thm55} and \ref{GradThm4} that the smallness of the difference $\bar P(x,t)$ between two solutions corresponding to two Forchheimer equations, when $t\to\infty$, can be controlled by the difference between coefficient vectors of the two Forchheimer polynomials.
\clearpage
\section{Preliminaries}
\label{prelim}
We consider a fluid in a porous medium having velocity $u(x,t)\in \ensuremath{\mathbb R}^n$, pressure $p(x,t)\in \ensuremath{\mathbb R}$ and density $\rho (x,t)\in \ensuremath{\mathbb R}^+=[0,\infty)$, with the spatial variable $x\in \ensuremath{\mathbb R}^n$ and time variable $t\in\ensuremath{\mathbb R}$.
The space dimension $n=3$ in physics problems, but here we consider any $n\ge 2$.
Generalized Forchheimer equations, studied in \cite{ABHI1,HI1}, are of the form:
\begin{equation}\label{gForch} g(|u|)u=-\nabla p,\end{equation}
where $g(s)\ge 0$ is a function defined on $[0,\infty)$.
When
\begin{equation}
g(s)=\alpha,\ \alpha +\beta s,\ \alpha +\beta s+\gamma s^2,\ \alpha +\gamma_m s^{m-1},
\end{equation*}
where $\alpha,\beta,\gamma,m,\gamma_m$ are empirical constants, we have Darcy's law, Forchheimer's two-term, three-term and power laws, respectively.
In this paper, the function $g$ in \eqref{gForch} is a generalized polynomial with non-negative coefficients, that is,
\begin{equation}\label{gsa} g(s) =a_0s^{\alpha_0}+a_1 s^{\alpha_1}+\ldots +a_N s^{\alpha_N},\ s\ge 0,\end{equation}
where
$N\ge 1$, $\alpha_0=0<\alpha_1<\ldots<\alpha_N$ are real (not necessarily integral) numbers, the coefficients satisfy $a_0,a_N>0$ and $a_1,\ldots,a_{N-1}\ge 0$.
The number $\alpha_N$ is the degree of $g$ and is denoted by $\deg(g)$.
Denote the vectors of powers and coefficients by $\vec \alpha=(\alpha_0,\ldots,\alpha_N)$ and
$\vec a=(a_0,\ldots,a_N)$.
The class of functions $g(s)$ as in \eqref{gsa} is denoted by FP($N,\vec \alpha$), which is the abbreviation of ``Forchheimer polynomials''. When the function $g$ in \eqref{gForch} belongs to FP($N,\vec \alpha$), it is referred to as the Forchheimer polynomial.
From relation \eqref{gForch} one can solve velocity $u$ in terms of pressure gradient $\nabla p$ and
derives a nonlinear version of Darcy's equation:
\begin{equation}\label{u-forma} u= -K(|\nabla p|)\nabla p.\end{equation}
The function $K:\ensuremath{\mathbb R}_+\to\ensuremath{\mathbb R}_+$ is defined by
\begin{equation}\label{Kdef} K(\xi)=\frac1{g(s(\xi))},
\text{ where } s=s(\xi)\ge 0 \text{ satisfies } sg(s)=\xi, \ \text{ for }\xi\ge 0. \end{equation}
We will use notation $g(s,\vec a)$, $K(\xi,\vec a)$, $s(\xi,\vec a)$ to denote the corresponding functions in \eqref{gsa} and \eqref{Kdef} when the dependence on $\vec a$ needs be specified.
Equation \eqref{gForch} replaces the momentum equation in fluid mechanics. In addition to this, we have the equation of continuity
\begin{equation}\label{conti-eq} \frac{\partial \rho}{\partial t}+\nabla\cdot(\rho u)=0,\end{equation}
and the equation of state which, for slightly compressible fluids, is
\begin{equation}\label{slight-compress}
\frac{d\rho}{dp}=\frac{\rho}{\kappa},\quad \kappa>0.
\end{equation}
From equations \eqref{u-forma}, \eqref{conti-eq} and \eqref{slight-compress} one derives an equation for the pressure:
\begin{equation}\label{dafo-nonlin} \frac{\partial p}{\partial t}=\kappa\nabla \cdot (K(|\nabla p|)\nabla p) + K(|\nabla p|)|\nabla p|^2.\end{equation}
Since the constant $\kappa$ is very large for most slightly compressible fluids in porous media \cite{Muskatbook}, we neglect the second term on the right-hand side of \eqref{dafo-nonlin}. This results in the following reduced equation
\begin{equation}\label{lin-p} \frac{\partial p}{\partial t} = \kappa \nabla\cdot (K(|\nabla p|)\nabla p).\end{equation}
Note that this reduction is commonly used in engineering.
By rescaling the time variable, hereafter we assume that $\kappa=1$.
Let $g=g(s,\vec a)$ in FP($N,\vec \alpha$).
The following numbers are in our calculations:
\begin{equation}\label{ab} a=\frac{\alpha_N}{1+\alpha_N}\in(0,1),\ b= \frac{a}{2-a}= \frac{\alpha_N}{2+\alpha_N}\in(0,1),\end{equation}
\begin{equation}\label{Ag} \chi(\vec a)=\max\Big \{a_0,a_1,\ldots,a_N,\frac1{a_0},\frac1{a_N} \Big \}\in [1,\infty).\end{equation}
The following properties for $K(\xi,a)$ are proved Lemmas III.5 and III.9 of \cite{ABHI1}.
For $\xi\ge 0$, we have
\begin{equation}\label{K-est-2}
- a\, K(\xi,\vec a)\le \xi\, \frac{\partial K(\xi,\vec a)}{\partial\xi} \le 0.
\end{equation}
Consequently, $K(\xi,\vec a)$ is decreasing in the variable $\xi$ and hence
\begin{equation}\label{Kestzero}
K(\xi, \vec a)\le K(0,\vec a)=a_0^{-1}\le \chi(\vec a),\end{equation}
Moreover, $K(\xi,\vec a)\xi^m$ is increasing in $\xi$ for all $m\ge 1$.
For the type of degeneracy of $K(\xi,\vec a)$ in $\xi$, we recall the following facts.
\begin{lemma}[cf. \cite{HI1}, Lemma 2.1]\label{lem21}
Let $g(s,\vec a)$ be in class FP($N,\vec \alpha$). One has for any $\xi\ge 0$ that
\begin{equation}\label{Kesta}
\frac{ C_0^{-1} \chi(\vec a)^{-1-a} }{(1+\xi)^a}\le K(\xi, \vec a)\le \frac{ C_0 \chi(\vec a)^{1+a} }{(1+\xi)^a},
\end{equation}
and for any $m\ge 1$, $\delta>0$ that
\begin{equation}\label{Km}
c_0^{-1} \chi(\vec a)^{-1-a} \frac{\delta^a}{(1+\delta)^a} (\xi^{m-a}-\delta^{m-a}) \le K(\xi,\vec a)\xi^m\le c_0 \chi(\vec a)^{1+a}\xi^{m-a},
\end{equation}
where $c_0=c_0(N,\alpha_N)>0$ depends on $N$ and $\alpha_N$ only.
In particular, when $m=2$, $\delta=1$, one has
\begin{equation}\label{Kestn}
2^{-a}c_0^{-1}\chi(\vec a)^{-1-a}(\xi^{2-a}-1)\le K(\xi,\vec a)\xi^2\le c_0 \chi(\vec a)^{1+a}\xi^{2-a}.
\end{equation}
\end{lemma}
Same as in \cite{ABHI1,HI1}, we define
\begin{equation}\label{Hxi} H(\xi,\vec a)=\int_0^{\xi^2}K(\sqrt s,\vec a)ds \quad \hbox{for } \xi\ge 0.\end{equation}
When vector $\vec a$ is fixed, we denote $K(\cdot,\vec a)$ and $H(\cdot,\vec a)$ by $K(\cdot)$ and $H(\cdot)$, respectively.
The function $H(\xi)$ can be compared with $\xi$ and $K(\xi)$ by
\begin{equation}\label{Hcompare}
K(\xi)\xi^2\leq H(\xi)\le 2K(\xi)\xi^2,\quad c_1(\xi^{2-a}-1)\leq H(\xi)\le c_2\xi^{2-a},
\end{equation}
where $c_1,c_2>0$ depend on $\chi(\vec a)$.
Next, we recall important monotonicity properties.
For convenience, we use the following notation: let $\vec{x}=(x_1,x_2,\ldots)$ and $\vec{x}'=(x'_1,x'_2,\ldots)$ be two arbitrary vectors of the same length, including possible length $1$. We denote by $\vec{x}\vee \vec{x}'$ and $\vec{x}\wedge \vec{x}'$ their maximum and minimum vectors, respectively, with components
$(\vec{x}\vee \vec{x}')_j=\max\{x_j, x'_j\}$ and $(\vec{x}\wedge \vec{x}')_j=\min\{ x_j, x'_j\}$.
\begin{lemma}[cf. \cite{HI1}, Lemma 5.2]\label{quasimono-lem}
Let $g(s,\vec a)$ and $g(s,\vec a')$ belong to class FP($N,\vec \alpha$). Then for any $y,y'$ in $\ensuremath{\mathbb R}^n$, one has
\begin{multline}\label{quasimonotone}
(K(|y|,\vec a) y-K(|y'|,\vec a')y')\cdot (y-y')
\ge (1-a)\cdot K(|y|\vee|y'|,\vec a\vee \vec a')\cdot |y-y'|^2 \\
- N\cdot \max\{ \chi(\vec a),\chi(\vec a')\}\cdot |\vec a-\vec a'|\cdot K(|y|\vee |y'|,\vec a\wedge \vec a')\cdot (|y|\vee |y'|)\cdot |y-y'|,
\end{multline}
where $a\in(0,1)$ is defined in \eqref{ab}.
Particularly, when $\vec a=\vec a'$, we have
\begin{equation}\label{mono1}
(K(|y|,\vec a) y-K(|y'|,\vec a)y')\cdot (y-y') \ge (1-a)\cdot K(|y|\vee|y'|,\vec a)\cdot |y-y'|^2.
\end{equation}
\end{lemma}
When the fluid is confined in an open, bounded domain $U$ of $\ensuremath{\mathbb R}^n$, the Sobolev embeddings play an important role in our study. For $0<r<n$, we denote by $r^*$ the critical Sobolev exponent, i.e. $r^* = \frac{nr}{n-r}$. The following
fact is used frequently
\begin{equation}\label{rfact}
r^*>2\Leftrightarrow \frac{nr}{n-r}>2 \Leftrightarrow r>\frac{2n}{n+2}\Leftrightarrow (2-r)n<2r.
\end{equation}
With this notation, we define the \textbf{Degree Condition} as one of the following equivalent conditions:
\begin{equation}\label{deg} \deg(g)\le \frac{4}{n-2},\quad a\le \frac 4{n+2},\quad 2\le (2-a)^*,\quad 2-a\ge \frac{2n}{n+2}.\tag{\textbf{DC}}
\end{equation}
Similarly, we define the \textbf{Strict Degree Condition} as one of the following equivalent conditions:
\begin{equation}\label{strictdeg}
\deg(g)< \frac{4}{n-2},\quad a < \frac 4{n+2},\quad 2 < (2-a)^*,\quad 2-a> \frac{2n}{n+2}.\tag{\textbf{SDC}}
\end{equation}
We will assume the Strict Degree Condition very often in this paper, but not always. Whenever this condition is met, the Sobolev space $W^{1,2-a}(U)$ is continuously embedded into $L^2(U)$.
We now consider parabolic Sobolev embeddings.
Let us denote throughout $Q_T=U\times(0,T)$.
\begin{lemma}\label{ParSob-1}
If $2n/(n+2)\le r \le 2$, $r<n$ and $p=r(n+2)/n$ then
\begin{equation}\label{Sineq1}
\| u\|_{L^p(Q_T)} \le C(1+T^{1/p})[[u]]_{2,r;T},
\end{equation}
where $C=C(U,n,r)>0$ is independent of $T$ and
\begin{equation}\label{normtriple}
[[u]]_{2,r;T}=\mathop{\mathrm{ess\,sup}}_{t\in [0,T]} \| u(t)\|_{L^2(U)}+ \|\nabla u\|_{L^r(Q_T)}.
\end{equation}
In \eqref{Sineq1} above, we can remove $T^{1/p}$ whenever $u$ vanishes on the boundary of $U$.
\end{lemma}
\begin{proof
The proof is standard, cf. \cite{LadyParaBook68,DiDegenerateBook}.
The short proof is presented here for the sake of completeness, and also serves the next lemma.
For convenience we denote $[[\cdot]] \equiv [[\cdot]]_{2,r,T}$.
Note that $2\le p \le r^*$. We write
\begin{equation}\label{powers}
\frac 1 p=\frac \alpha 2+\frac\beta {r^*},\quad \alpha,\beta\ge 0,\quad \alpha+\beta=1.
\end{equation}
By interpolation inequality and Sobolev embedding, we have
\begin{equation}
\|u(t)\|_{L^p(U)} \le \| u(t)\|_{L^2(U)}^\alpha \|u(t)\|_{L^{r^*}(U)}^\beta
\le C \| u(t)\|_{L^2(U)}^\alpha (\delta \|u(t)\|_{L^r(U)}^\beta +\norm{\nabla u(t)}_{L^r(U)}^\beta),
\end{equation*}
where $\delta=1$ in general, and $\delta=0$ in case $u$ vanishes on the boundary $\partial U$.
Note that $r\le 2$ then $\|u(t)\|_{L^r(U)}\le C\|u(t)\|_{L^2(U)}$. Thus,
\begin{equation}\label{e1}
\|u(t)\|_{L^p(U)}
\le C \delta \| u(t)\|_{L^2(U)} + C \| u(t)\|_{L^2(U)}^\alpha \norm{\nabla u(t)}_{L^r(U)}^\beta.
\end{equation}
Raising \eqref{e1} to power $p$, integrating it in $t$ from $0$ to $T$ and using the fact that
$\| u(t)\|_{L^2(U)} \le [[u]]$ a.e. in $[0,T]$, we obtain
\begin{equation}\label{e2}
\norm{u}_{L^p(Q_T)}^p
\le C \delta T [[u]]^{p} + C [[u]]^{\alpha p} \int_0^T \|\nabla u(t)\|_{L^r(U)}^{\beta p} dt.
\end{equation}
From \eqref{powers}, we find $\beta=\frac n{n+2}$. Thus, $\beta p= r$ and we have from \eqref{e2} that
\begin{equation}\label{e3}
\begin{aligned}
\norm{u}_{L^p(Q_T)}^p
&\le C \delta T [[u]]^{p} + C [[u]]^{\alpha p} \Big(\int_0^T\int_U |\nabla u(x,t)|^{r} dx dt\Big )^\frac{\beta p}{r} \\
&\le C \delta T [[u]]^{p}
+ C [[u]]^{\alpha p} [[u]]^{\beta p}
= C [[u]]^{p} (\delta T+1).
\end{aligned}
\end{equation}
Therefore, we obtain \eqref{Sineq1}.
\end{proof}
The following parabolic Sobolev embedding with spatial weights will be useful in this paper.
\begin{lemma}\label{ParaSob-3}
Given $W(x,t)>0$ on $Q_T$. Suppose two numbers $m$ and $r $ satisfy $2n/(n+2)\le r\le 2$, $r<n$ and $r<m < r^*$.
Let
\begin{equation}\label{pm}
p= 2+m -\frac{2m}{r^*}.
\end{equation}
Then
\begin{equation}\label{weight1}
\|u\|_{L^p(Q_T)}
\le C [[u]]_{2,m,W;T} \Big\{ T^{1/p}+ \mathop{\mathrm{ess\,sup}}_{t\in[0,T]} \Big (\int_U W(x,t)^{-\frac r{m-r}}dx\Big )^\frac{m-r}{pr}\Big \},
\end{equation}
where
\begin{equation}\label{normquad}
[[u]]_{2,m,W;T}=\mathop{\mathrm{ess\,sup}}_{t\in [0,T]} \|u(t)\|_{L^2(U)}+\Big(\int_0^T \int_U W(x,t)|\nabla u|^m dx dt\Big)^\frac 1 m.
\end{equation}
Consequently, under the Strict Degree Condition, when $m=2$ and $r=2-a$ we have
\begin{equation}\label{Wemb}
\|u\|_{L^p(Q_T)}
\le C [[u]]_{2,2,W;T} \Big\{ T^{1/p}+ \mathop{\mathrm{ess\,sup}}_{t\in[0,T]} \Big (\int_U W(x,t)^{-\frac{2-a}{a}}dx\Big )^\frac{2}{p(2-a)}\Big \},
\end{equation}
where $2<p<(2-a)^*$ given explicitly by
\begin{equation}\label{pkey}
p=4\Big(1-\frac 1{(2-a)^*}\Big).
\end{equation}
In \eqref{weight1} and \eqref{Wemb} above, we can remove $T^{1/p}$ whenever $u$ vanishes on the boundary of $U$.
\end{lemma}
\begin{proof
In this proof we denote $[[\cdot]] = [[\cdot]]_{2,m,W;T}$.
By H\"older's inequality,
\begin{equation}\label{stemp}
\|\nabla u\|_{L^r(U)}
= \Big( \int_U W^{\frac rm} |\nabla u|^r \cdot W^{-\frac rm} dx \Big)^\frac1r
\le \Big (\int_U W|\nabla u|^m dx\Big )^\frac 1m \Big (\int_U W^{-\frac r{m-r}}dx\Big )^{\frac1r-\frac1m}.
\end{equation}
By \eqref{rfact} and the condition on $r$ we have $r^*\ge 2$. With exponent $p$ in \eqref{pm}, we write
\begin{equation}\label{manyp}
p-2= \frac{m(r^*-2)}{r^*},
\quad p-r^*= \frac{(r^*-2)(m-r^*)}{r^*},
\quad p-m=\frac{2(r^*-m)}{r^*},
\end{equation}
then we have $2\le p\le r^*$ and $p>m$.
Let $\alpha$ and $\beta$ be defined as in \eqref{powers}.
Then applying \eqref{e1} and \eqref{stemp} yields
\begin{multline}\label{e4}
\norm{u}_{L^p(Q_T) }^p \le C \delta T [[u]]^p+C [[u]]^{\alpha p}
\Big[ \int_0^T \Big (\int_U W|\nabla u|^m dx\Big )^\frac {\beta p} m d\tau \Big]\\
\cdot \mathop{\mathrm{ess\,sup}}_{[0,T]} \Big (\int_U W(x,t)^{-\frac r{m-r}}dx\Big )^{\beta p(\frac1r-\frac1m)} .
\end{multline}
From \eqref{powers} and \eqref{manyp}, we have
$\beta =\frac{r^*(p-2)}{p(r^*-2)}=\frac m p.$
Therefore, we rewrite \eqref{e4} as
\begin{align*}
\norm{u}_{L^p(Q_T) }^p
&\le C \delta T [[u]]^p+C [[u]]^{\alpha p} \Big (\int_0^T \int_U W|\nabla u|^m dx d\tau \Big )^\frac {\beta p} m\mathop{\mathrm{ess\,sup}}_{[0,T]} \Big (\int_U W(x,t)^{-\frac r{m-r}}dx\Big )^\frac{m-r}r\\
&\le C \delta T [[u]]^p+C [[u]]^{\alpha p} [[u]]^{\beta p} \mathop{\mathrm{ess\,sup}}_{[0,T]} \Big (\int_U W(x,t)^{-\frac r{m-r}}dx\Big )^\frac{m-r}r\\
&\le C [[u]]^p\Big\{ \delta T + \mathop{\mathrm{ess\,sup}}_{[0,T]}\Big (\int_U W(x,t)^{-\frac r{m-r}}dx\Big )^\frac{m-r}r\Big\}.
\end{align*}
Thus, we obtain \eqref{weight1}.
In particular, when $m=2$ and $r=2-a$ then \eqref{pm} becomes \eqref{pkey}.
Under the Strict Degree Condition, requirements on $r$ and $m$ are met.
Then \eqref{Wemb} follows \eqref{weight1}.
\end{proof}
Another type of embedding will be proved in Lemma \ref{LUK}.
\clearpage
\section{Estimates of solutions}
\label{supestimate}
Let $U$ be a bounded, open, connected subset of $\ensuremath{\mathbb R}^n$ with boundary $\Gamma$ of class $C^2$.
We consider a fluid flow in $U$ that satisfies the generalized Forchheimer
equation \eqref{gForch} with a fixed $g(s)=g(s,\vec a)\in FP(N,\vec \alpha)$.
We study the resulting parabolic equation for the pressure $p=p(x,t)$:
\begin{equation}\label{eqorig} \frac{\partial p}{\partial t}=\nabla \cdot (K(|\nabla p|)\nabla p), \quad x\in U,\ t>0.\end{equation}
In addition, we assume the flux condition on the boundary:
\begin{equation*} u\cdot \vec\nu =\psi(x,t),\quad x\in \Gamma,\ t>0,\end{equation*}
where $\vec\nu$ is the outward normal vector on $\Gamma$ and the flux $\psi(x,t)$ is known. Hence, by \eqref{u-forma} we have
\begin{equation}\label{BC} -K(|\nabla p|)\nabla p \cdot \vec\nu=\psi \quad \hbox{on}\quad \Gamma\times (0,\infty) .\end{equation}
The initial data
\begin{equation} \label{In-Cond} p(x,0) = p_0(x) \text{ is given.}\end{equation}
We will focus on the IBVP \eqref{eqorig}, \eqref{BC} and \eqref{In-Cond}. By integrating \eqref{eqorig} over $U$, we easily find
\begin{equation*} \frac{d}{dt} \int_U p(x,t)dx = \int_\Gamma K(|\nabla p|)\nabla p\cdot \vec\nu d\sigma=-\int_\Gamma \psi(x,t)d\sigma,\quad t>0. \end{equation*}
(Here $d\sigma$ is the surface area element.)
By the continuity of $\int_U p(x,t)dx$ and $\int_\Gamma \psi(x,t)d\sigma$ on $[0,\infty)$, we assert
\begin{equation}\label{pavg-rel} \int_U p(x,t)dx =\int_U p(x,0)dx - \int_0^t\int_\Gamma \psi(x,\tau)d\sigma d\tau,\quad t\ge 0.\end{equation}
Let $\bar p(x,t)=p(x,t)-\frac1{|U|}\int_U p(x,t)dx$, for $x\in U$ and $t\ge 0$, where $|U|$ denotes the volume of $U$.
Then $\bar p$ satisfies the zero average condition
\begin{equation}\label{pbar0}
\int_U \bar{p}(x,t) dx=0 \quad \hbox{for all }t\ge 0.
\end{equation}
It follows from \eqref{pavg-rel} that for $t\ge 0$,
\begin{equation}
\label{pbargam}
\bar{p}(x,t)=p(x,t)-\frac{1}{|U|}\int_U p_0(x) dx+\frac{1}{|U|}\int_0^t\int_\Gamma \psi(x,\tau) d\sigma d\tau.
\end{equation}
We call $\bar p(x,t)$ the shifted solution.
Then $\bar{p}$ satisfies the following IBVP
\begin{equation}\label{eqgamma}
\left \{
\begin{array}{lll}
& \dfrac{\partial \bar{p}}{\partial t}=\nabla \cdot (K(|\nabla \bar{p}|)\nabla \bar{p}) + \frac{1}{|U|}\int_\Gamma \psi(x,t) d\sigma,& \quad x\in U, t>0,\\
& -K(|\nabla\bar{p}|)\nabla\bar{p} \cdot \vec\nu=\psi(x,t), &\quad x\in \Gamma, t>0,\\
& \bar p(x,0)=p_0(x)- \frac{1}{|U|}\int_U p_0(x) dx, &\quad x\in U.
\end{array} \right.
\end{equation}
Note that even in the linear case, i.e., $K(\xi)\equiv const.$ and $\psi(x,t)$ is uniformly bounded on $\Gamma\times[0,\infty)$, the solution $p(x,t)$ can still be unbounded as $t\to\infty$.
Therefore instead of estimating $p(x,t)$ directly, we will estimate $\bar p(x,t)$.
Thanks to the explicit relation \eqref{pbargam} between $p$ and $\bar p$, the obtained results will have clear physical interpretations and the estimates for $p(x,t)$ can be easily retrieved from those for $\bar p(x,t)$.
The existence of weak solutions can be treated by the theory of nonlinear monotone operators (cf. \cite{BrezisMonotone,s97,z90}, see also section 3 of \cite{HIKS1} for our proof for the Dirichlet boundary condition). The regularity of weak solutions is treated in \cite{DiDegenerateBook}.
For simplicity, we always assume that the initial data $p_0$ and boundary data $\psi$ are sufficiently smooth and the solution of \eqref{eqgamma} exists \textit{for all} $t\ge 0$. Also, sufficient regularity of the solution is assumed.
For example, when we estimate $L^\infty$-norm of $\bar p$ we require $\bar p\in C(\bar U\times [0,\infty))$;
when we estimate $L^\infty$-norm of $\bar p_t$ we require $\bar p_t\in C( U\times (0,\infty))$;
when we estimate $L^s$-norm of $\nabla p$ we require $\nabla p\in C( [0,\infty),L_{loc}^s(U))$.
In the following subsections, we derive various estimates for pressure and its derivatives.
These estimates are important by themselves and for the next sections when we study the stability of the solutions.
\begin{notation}
In estimates below, constants $C$'s always depend on the dimension $n$, domain $U$ and the Forchheimer polynomial $g(s,\vec a)$. Additional dependence will be specified as needed.
We use short-hand notation $\| f \|_{L^2}$, $\| f \|_{L^\infty}$ for the norms $\| f \|_{L^2(U)}$, $\|f\|_{L^\infty(U)}$ whenever $f$ is defined on $U$. Similarly, if a function $\phi$ is defined on $\Gamma$, we use short-hand notation
$\| \phi \|_{L^2}$, $\|\phi \|_{L^\infty}$ for the norms $\| \phi \|_{L^2(\Gamma)}$, $\|\phi\|_{L^\infty(\Gamma)}$.
If $f(x,t)$ is a function of two variables, we denote by $f(t)$ the function $t\to f(\cdot,t)$, therefore $\|f(t)\|_{L^2}$ means $\|f(\cdot,t)\|_{L^2}$ and $\sup_{[0,T]}\| f\|_{L^2}$ means $\sup\{ \| f(\cdot,t)\|_{L^2}:t\in[0,T]\}$.
Throughout, we denote $Q_T=U\times (0,T)$.
Also, we use the following notation for partial derivatives: $\partial p/\partial t= p_t$ and $\partial /\partial x_i=\partial_i$.
\end{notation}
\subsection{Estimates for pressure}\label{pressuresubsec}
We begin our subsection by introducing the following result on the $L^\infty$-estimates of for the $\bar p$ of the IBVP \eqref{eqgamma}.
Throughout the paper, we denote
\begin{equation} \label{newmu}
{\mu_1}=(2-a)(n+2)/n.
\end{equation}
\begin{proposition} \label{p-T.est} Assume \eqref{strictdeg}. Then,
there is a constant $C>0$ such that any $T>0$, the following inequality holds
\begin{equation}\label{pbaronly}
\sup_{[0,T]} \|\bar p\|_{L^\infty}\le C \Big\{ \|\bar p_0\|_{L^\infty} + (1+T)^\frac2{2-a} (\sup_{[0,T]} \|\psi\|_{L^\infty}+1)^\frac{1}{1-a} \Big\}.
\end{equation}
\end{proposition}
\begin{proof}
We follow the celebrated De Giorgi's technique. Although, the method is standard (see, for example, \cite{LadyParaBook68}), calculations are tedious and new. Thus, for completeness, we provide its details here. For any $k\ge0$, let us denote
\begin{align*}
& \bar p^{(k)} =\max\{\bar p-k,0\},
\quad S_{k}(t)=\{ x\in U: \bar p^{(k)}(x,t)\ge 0\}, \quad \sigma_k =\int_0^T |S_k(t)|dt, \\
& F_k = \sup_{t\in [0,T]} \int_U |\bar p^{(k)}(x,t)|^2 dx + \int_0^T\int_U |\nabla\bar p^{(k)}(x,t)|^{2-a} dxdt.
\end{align*}
Assume $k\ge \|\bar p_0\|_{L^\infty}$, then $\bar p_0^{(k)}(x) =0$ a.e..
Multiplying the first equation of \eqref{eqgamma} by $\bar p^{(k)}$ and integrating
the resultant over the domain $U$, we obtain
\begin{align*}
& \frac 12 \frac{d}{dt} \int_U |\bar p^{(k)}|^2 dx + \int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dx
=\int_\Gamma \psi \bar p^{(k)} d\sigma +\frac{1}{|U|}\int_\Gamma \psi(x,t) d\sigma\int_U \bar{p}^{(k)} dx.
\end{align*}
We bound $|\psi(x,t)|$ and $|\int_\Gamma \psi(x,t) d\sigma|$ by $C\|\psi(t)\|_{L^\infty}$, and apply the trace theorem to the boundary integral, and then H\"older's inequality to have
\begin{align*}
& \frac 12 \frac{d}{dt} \int_U |\bar p^{(k)}|^2 dx + \int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dx
\le C \|\psi(t)\|_{L^\infty} \int_U \Big (|\bar p^{(k)}|+|\nabla \bar p^{(k)}| \Big ) dx \\
&\le C \|\psi(t)\|_{L^\infty} \left \{ \Big (\int_U |\bar p^{(k)}|^2dx\Big )^{1/2} |S_k(t)|^{1/2}+ \Big (\int_{S_k(t)} |\nabla \bar p^{(k)}|^{2-a}dx\Big )^\frac 1{2-a}|S_k(t)|^\frac{1-a}{2-a}\right \}.
\end{align*}
Using \eqref{Kestn} to compare $|\nabla \bar p^{(k)}|^{2-a}$ with $K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 +1$, we obtain
\begin{align*}
& \frac 12 \frac{d}{dt} \int_U |\bar p^{(k)}|^2 dx + \int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dx
\le C \|\psi(t)\|_{L^\infty} \Big \{ \Big (\int_U |\bar p^{(k)}|^2dx\Big )^{1/2} |S_k(t)|^{1/2} \\
& \quad \quad + \Big (\int_{S_k(t)}K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2dx\Big )^\frac 1{2-a}|S_k(t)|^\frac{1-a}{2-a} + |S_k(t)|\Big \}.
\end{align*}
Let $\varepsilon>0$. By Young's inequality, we have
\begin{align*}
& \frac 12 \frac{d}{dt} \int_U |\bar p^{(k)}|^2 dx + \frac 12 \int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dx
\le \varepsilon \int_U |\bar p^{(k)}|^2 dx
\\
&\quad \quad+ C \Big\{ \varepsilon^{-1} \|\psi(t)\|_{L^\infty}^2 +\|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a}
+\|\psi(t)\|_{L^\infty} \Big\}|S_k(t)|.
\end{align*}
Note that $p^{(k)}(0)=0$. For each $t \in [0,T)$, integrating the previous estimate on $(0, t)$, and the taking the supremum in $t$ yield
\begin{align*}
&\sup_{[0,T]} \int_U |\bar p^{(k)}|^2 dx + \int_0^T\int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dxdt\\
&\le 4\varepsilon \int_0^T \int_U |\bar p^{(k)}|^2 dx +C\int_0^T \Big( \varepsilon^{-1} \|\psi(t)\|_{L^\infty}^2
+ \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a} +\|\psi(t)\|_{L^\infty}\Big)|S_k(t)| dt\\
&\le 4\varepsilon T \sup_{[0, T]} \int_U |\bar p^{(k)}|^2 dx +C\int_0^T \Big( \varepsilon^{-1} \|\psi(t)\|_{L^\infty}^2 + \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a}+\|\psi(t)\|_{L^\infty}\Big)|S_k(t)| dt.
\end{align*}
By taking $\varepsilon=1/(8T)$, we obtain
\begin{align*}
&\sup_{[0,T]} \int_U |\bar p^{(k)}|^2 dx + \int_0^T\int_U K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2 dxdt\\
&\le C \int_0^T\left[ T \|\psi(t)\|_{L^\infty}^2 + \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a} + \|\psi(t)\|_{L^\infty} \right]|S_k(t)| dt\\
&\le C \int_0^T\left[ T \|\psi(t)\|_{L^\infty}^2 + \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a} + 1 \right]|S_k(t)| dt.
\end{align*}
Again, using \eqref{Kestn} to compare $K(|\nabla \bar p^{(k)}|)|\nabla \bar p^{(k)}|^2$ with $|\nabla \bar p^{(k)}|^{2-a}$ and using Young's inequality,
one gets
\begin{equation}\label{p-DG.cond}
F_k \mathbin{\buildrel \rm def \over \longequals} \sup_{[0,T]} \int_U |\bar p^{(k)}|^2 dx + \int_0^T \int_U |\nabla\bar p^{(k)}|^{2-a} dxdt \leq C_0 \alpha_T \sigma_k,
\end{equation}
where $C_0$ is a constant depending on $a$, $n$ and $U$, and $\alpha_T = T\|\psi\|_{L^\infty}^2 + \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a} + 1$.
We now iterate \eqref{p-DG.cond} to derive our desired estimate. Under the Strict Degree Condition \eqref{strictdeg}, ${\mu_1}>2$, hence, the parabolic Sobolev embedding in Lemma~\ref{ParSob-1} with $r=2-a$ implies
\begin{equation} \label{Fk.co}
\| \bar p^{(k)}\|_{L^{\mu_1}(Q_T)}\le C(1+ T)^{1/{\mu_1}} (F_k^{1/2}+F_k^{1/(2-a)}).
\end{equation}
Next, let $M_0 \geq \|\bar p(\cdot, 0)\|_{L^\infty}$ be a fixed number which will be determined.
For each $i \in \mathbb{N} \cup \{0\}$, let $k_i=M_0(2-2^{-i})$.
Then $k_i$ is increasing in $i$, and therefore $S_{k_i}$, $\sigma_{k_i}$ are decreasing.
Also, note that $k_i \geq M_0\ge \|\bar p_0\|_{L^\infty}$ for all $i\ge 0$.
We now define $\mathcal Q_k=\{ (x,t)\in U\times(0,T): \bar p(x,t)>k\}$.
From \eqref{Fk.co} and the fact that
\begin{equation} \label{pk.co}
\|\bar p^{(k_i)}\|_{L^{\mu_1}(Q_T)} \ge \|\bar p^{(k_i)}\|_{L^{\mu_1}(\mathcal Q_{k_{i+1}})}\ge \norm{k_{i+1} -k_i }_{L^{\mu_1}(\mathcal Q_{k_{i+1}})} \ge (k_{i+1}-k_{i})\sigma_{k_{i+1}}^{1/{\mu_1}},
\end{equation}
it follows
\[
\sigma_{k_{i+1}}^{1/{\mu_1}}\le \frac{C(1+T)^{1/{\mu_1}}}{ k_{i+1}-k_{i}} \Big[F_{k_i}^{1/2}+F_{k_i}^{1/(2-a)}\Big].
\]
This together with \eqref{p-DG.cond} yield
\[
\sigma_{k_{i+1}}^{1/{\mu_1}}\le C (1+T)^{1/{\mu_1}}\frac{ 2^i }{M_0}
\Big[ (\alpha_T \sigma_{k_i})^{1/2}+ (\alpha_T \sigma_{k_i})^{1/(2-a)}\Big].
\]
Equivalently,
\[
\sigma_{k_{i+1}}\le \frac{ C (1+T) \alpha_T^{{\mu_1}/(2-a)} }{M_0^{\mu_1}}2^{{\mu_1}i}
( \sigma_{k_i}^{{\mu_1}/2}+ \sigma_{k_i}^{{\mu_1}/(2-a)}).
\]
Let us denote $Y_i = \sigma_{k_i}$, $D = \frac{C (1+T) \alpha_T^{{\mu_1}/(2-a)} }{M_0^{\mu_1}}$, ${\theta_1} = {\mu_1}/(2-a) -1 >0$ and ${\theta_2} = {\mu_1}/2 - 1 >0 $. We now obtain
\[ Y_{i+1} \leq D 2^{{\mu_1}i} [Y_i^{1+{\theta_1}} + Y_i^{1+{\theta_2}}].
\]
Note that $k_0=M_0\ge \|\bar p_0\|_{L^\infty}$ and
\[ Y_0=\sigma_{M_0}\le |Q_T|= |U|T\le |U|(T+1). \]
To apply \eqref{Y0mcond} with $m=2$ in Lemma \ref{multiseq}, we choose $M_0$ sufficiently large such that
$$Y_0 \leq C \min\{D^{-1/{\theta_2}}, D^{-1/{\theta_1}}\},\quad \text{or sufficiently,} $$
\begin{align*}
|U|(T+1) \le C \min\Big\{ ((1+T) \alpha_T^{{\mu_1}/(2-a)} M_0^{-{\mu_1}})^{-1/{\theta_2}}, ((1+T) \alpha_T^{{\mu_1}/(2-a)} M_0^{-{\mu_1}})^{-1/{\theta_1}} \Big \}.
\end{align*}
It suffices to have
\begin{equation*} \label{M-p.cond}
\begin{split}
& M_0\ge C (1+T)^\frac{1+{\theta_2}}{{\mu_1}} \alpha_T^{1/(2-a)} = C (1+T)^\frac{1}{2} \alpha_T^\frac1{2-a}\quad
\text{and} \\
& M_0\ge C (1+T)^\frac{1+{\theta_1}}{{\mu_1}} \alpha_T^{1/(2-a)} = C (1+T)^\frac{1}{2-a} \alpha_T^\frac1{2-a}.
\end{split}
\end{equation*}
These and the initial requirement $M_0 \geq \|\bar{p}_0\|_{L^\infty}$ are satisfied if
we choose
\[
M_0\ge C \Big\{ \|\bar p_0\|_{L^\infty} + (1+T)^\frac1{2-a} \alpha_T^\frac{1}{2-a} \Big\}.
\]
Note that
$ (1+T)\alpha_T
\le C (T+1)^2 (\sup_{[0,T]}\|\psi\|_{L^\infty}+1)^\frac{2-a}{1-a}$.
Therefore we finally select
\[
M_0=C_1 \Big\{ \|\bar p_0\|_{L^\infty} + (1+T)^\frac2{2-a} (\sup_{[0,T]}\|\psi\|_{L^\infty}+1)^\frac{1}{1-a} \Big\}.
\]
Now, applying \eqref{Y0mcond} in Lemma \ref{multiseq}, we have $\sigma_{2M_0}=\displaystyle{\lim_{i \rightarrow \infty}Y_i} = 0$.
Thus, $\bar p(x,t)\le 2M_0$ in $Q_T$. Note that we used the continuity of $\bar p(x,t)$ on $\bar Q_T$.
By using the same argument with $p$ replaced by $-p$ and $\psi$ replaced by $-\psi$, one can show that $-\bar p \le 2M_0$ in $Q_T$.
Therefore, \eqref{pbaronly} follows and the proof is complete.
\end{proof}
We emphasize that the $L^\infty$-estimate of $\bar{p}$ in Proposition \ref{p-T.est} depends on $L^\infty$-norm of the initial data and on $T$. In the next proposition, we give a $L^\infty$-estimate which removes the former dependence. The result will be used flexibly to derive the time-uniform $L^\infty$-estimate of $\bar{p}$.
\begin{proposition} \label{local-L-infty} Assume \eqref{strictdeg}.
Then, there is a constant $C>0 $ such that for any $T_0\ge 0$, $T>0$, $\delta\in (0,1]$ and $\theta\in (0,1)$, the following inequality holds
\begin{multline}\label{mainest-1}
\sup_{[T_0+\theta T,T_0+T]} \|\bar p\|_{L^\infty}\le C\Big\{ \sqrt{\mathcal E}
+ (T+1)^\frac{n}{4-(n+2)a} \Big(1+\frac 1 { \delta^a \theta T}\Big)^\frac{n+2}{4-(n+2)a} \\
\cdot \Big(\|\bar p\|_{L^2(U\times (T_0,T_0+T))} + \|\bar p\|_{L^2(U\times (T_0,T_0+T))}^\frac{4}{4-(n+2)a}\Big) \Big\},
\end{multline}
where
\begin{equation*}
\mathcal E=\mathcal E_{T_0,T} \mathbin{\buildrel \rm def \over \longequals} \delta^{2-a}+ T\delta^{-a} \sup_{[T_0,T_0+T]}\|\psi\|_{L^\infty}^2 + \delta^{-\frac {a(2-a)}{1-a}} \sup_{[T_0,T_0+T]} \|\psi\|_{L^\infty}^\frac{2-a}{1-a}.
\end{equation*}
\end{proposition}
\begin{proof}
Without loss of generality, we assume $T_0=0$. Again, we use De Giorgi's iteration technique. But, unlike in the previous one, we will iterate using the $L^2$-norm of $\bar{p}$. The proof is therefore significantly different from that of Proposition \ref{p-T.est} and we give it here in full details.
For each $k\ge 0$, let $\bar p^{(k)}$ be as
in the proof of the previous theorem. Also, let $\zeta=\zeta(t)$ be a non-negative cut-off function on $[0,T]$
with $\zeta(0)=0$. Multiplying \eqref{eqgamma} by $\bar p^{(k)} \zeta$ and integrating over $U$, we have
\begin{equation}\label{prediff}
\frac12 \int_U \frac{\partial |\bar p^{(k)}|^2}{\partial t} \zeta dx +\int_U K(|\nabla\bar p^{(k)}|) |\nabla\bar p^{(k)}|^2 \zeta dx
=\int_\Gamma \psi \bar p^{(k)} \zeta d\sigma + \frac{1}{|U|}\int_\Gamma \psi(x,t) d\sigma\int_U \bar p^{(k)} \zeta dx.
\end{equation}
Denote by $\chi_k(x,t)$ the characteristic function of the set $\{ (x,t)\in \in U\times (0,T): \bar p^{(k)}(x,t) > 0\}$.
Then, using \eqref{Km} in Lemma \ref{lem21} to estimate $K(|\nabla\bar p^{(k)}|) |\nabla\bar p^{(k)}|^2\ge C\delta^a |\nabla\bar p^{(k)}|^{2-a} - C\delta^2$, and estimating the right-hand side of \eqref{prediff} the same way as in the proof of Proposition \ref{p-T.est}, we have
\begin{align*}
& \frac12 \frac{d}{dt} \int_U |\bar p^{(k)}|^2 \zeta dx - \frac12 \int_U |\bar p^{(k)}|^2 \zeta_t dx +C\delta^a \int_U |\nabla \bar p^{(k)}|^{2-a} \zeta dx -C\delta^2 \int_U \chi_k \zeta dx\\
& \le C \|\psi(t)\|_{L^\infty} \int_U (|\bar p^{(k)}| \zeta +|\nabla \bar p^{(k)}|\zeta) dx.
\end{align*}
Let $\varepsilon >0$. Then it follows from Young's inequality that
\begin{align*}
& \frac12 \frac{d}{dt} \int_U |\bar p^{(k)}|^{2} \zeta dx +C_1\delta^a \int_U |\nabla \bar p^{(k)}|^{2-a} \zeta dx \\
& \le \frac12 \int_U |\bar p^{(k)}|^2 \zeta_t dx + C\delta^2\int_U \chi_k(t)\zeta dx
+ \varepsilon \int_U |\bar p^{(k)}|^2 \zeta dx + C \varepsilon^{-1} \|\psi\|_{L^\infty}^2 \int_U \chi_k \zeta dx\\
&\quad + \frac{ C_1 \delta^{a} }2 \int_U |\nabla \bar p^{(k)}|^{2-a} \zeta dx
+ C \delta^{-\frac{a}{1-a}} \|\psi(t)\|_{L^\infty}^\frac{2-a}{1-a} \int_U \chi_k \zeta dx.
\end{align*}
Now, integrating this inequality on $(0, t)$ and taking supremum of the resultant on $[0,T]$, we obtain
\begin{align*}
& \frac{1}{2} \sup_{[0,T]} \int_U |\bar p^{(k)}|^2 \zeta dx +\frac{C_1\delta^a}2 \int_0^T\int_U |\nabla \bar p^{(k)}|^{2-a} \zeta dxdt \\
& \le \int_0^T\int_U |\bar p^{(k)}|^2 \zeta_t dxdt
+ 2 \varepsilon T \sup_{[0,T]} \int_U |\bar p^{(k)}|^2 \zeta dx \\
&\quad \quad +C \Big[ \delta^2 + \varepsilon^{-1} \|\psi\|_{L_t^\infty(0,T; L_x^\infty)}^2 +\delta^{-\frac a{1-a}} \|\psi\|_{L_t^\infty(0,T; L_x^\infty)}^\frac{2-a}{1-a}\Big] \int_0^T\int_U \chi_k\zeta dxdt.
\end{align*}
By choosing $\varepsilon=1/(8T)$, we obtain
\begin{equation}\label{prena}
\begin{aligned}
& \sup_{[0,T]} \int_U |\bar p^{(k)}|^2 \zeta dx + \int_0^T\int_U |\nabla \bar p^{(k)}|^{2-a} \zeta dx dt\\
&\leq C\delta^{-a}\int_0^T\int_U |\bar p^{(k)}|^2 \zeta_t dx dt +C\mathcal E \int_0^T\int_U \chi_k\zeta dx dt ,
\end{aligned}
\end{equation}
where
$\mathcal E=\delta^{2-a}+ T\delta^{-a} \sup_{[0,T]} \|\psi\|_{L^\infty}^2 + \delta^{-\frac{a(2-a)}{1-a}} \sup_{[0,T]} \|\psi\|_{L^\infty}^\frac{2-a}{1-a}.$
We will iterate this relation with $k=k_{i+1}$ and $\zeta=\zeta_i$. The choice of $k_i$ and $\zeta_i(t)$ is as follows.
Let $t_i =\theta T( 1- 2^{-i})$ for all $i\ge 0$, and let $\zeta_i$ be a piecewise linear function with
$\zeta_i(t)=0$ for $t\le t_i$, $\zeta_i(t)=1$ for $t\ge t_{i+1}$ and
\[|\zeta_{it}(t)|\le \frac 1{t_{i+1}-t_i}= \frac{ 2^{i+1} }{\theta T} \quad\forall\ t\in [0,T].\]
Next, let $M_0$ be a fixed positive number which will be determined, and let $k_i= M_0(1-2^{-i})$, for $i\ge 0$.
Note that, because of our choices, $k_0=0$ and $k_{i+1}-k_i = 2^{-i-1}M_0$ and $t_0=0<t_1<\ldots<\theta T$.
We also denote
$A_{i,j} =\{(x,t):u(x,t)>k_i,\ t\in(t_j,T)\}$ and $A_i=A_{i,i}$.
Then, by applying \eqref{prena} with $k=k_{i+1}$ and $\zeta=\zeta_i$, we obtain
\begin{equation}\label{pren1}
\begin{aligned}
& \sup_{[0,T]} \int_U |\bar p^{(k_{i+1})}|^2 \zeta_i dx + \int_0^T\int_U |\nabla \bar p^{(k_{i+1})}|^{2-a} \zeta_i dx dt \\
& \le \delta^{-a}\int_0^T\int_U |\bar p^{(k_{i+1})}|^2 (\zeta_{i})_t dxdt
+C\mathcal E \Big (\int_0^T \int_U \chi_{k_{i+1}} \zeta_i dxdt\Big ).
\end{aligned}
\end{equation}
Now, we define
$F_{i} = \sup_{[t_{i+1},T]} \int_U |\bar p^{(k_{i+1})}|^{2} \zeta_i dx
+ \int_{t_{i}}^T\int_U |\nabla \bar p^{(k_{i+1})}|^{2-a} \zeta_i dx dt. $
Then it follows from \eqref{pren1} that
\begin{equation}
F_{i}
\le \delta^{-a} \int_{t_i}^{t_{i+1}} \int_U |\bar p^{(k_{i+1})}|^2 (\zeta_i)_t dxdt
+C \mathcal E \Big (\int_{t_{i+1}}^T\int_U \chi_{k_{i+1}} dx dt\Big ).
\end{equation*}
Therefore,
\begin{equation} \label{firstFn}
F_{i}\le C 2^i (\theta T)^{-1} \delta^{-a} \|\bar{p}^{(k_{i+1})}\|_{L^2(A_{i+1,i})}^2 +C\mathcal E |A_{i+1,i}|.
\end{equation}
Since
$ \|\bar p^{(k_i)}\|_{L^2(A_i)}\ge \|\bar p^{(k_{i})}\|_{L^2(A_{i+1,i})}\ge (k_{i+1}-k_i) |A_{i+1,i}|^{1/2}$,
we see that
\begin{equation} \label{An.n}
|A_{i+1,i}| \le (k_{i+1}-k_i)^{-2} \|\bar p^{(k_i)}\|_{L^2(A_i)}^{2} = 4^{i} M_0^{-2}\|\bar p^{(k_i)}\|_{L^2(A_i)}^{2}.
\end{equation}
From \eqref{firstFn} and \eqref{An.n}, we get
\begin{equation}\label{Fn}
F_{i} \le C \delta^{-a} (\theta T)^{-1} 2^i \|\bar p^{(k_{i+1})}\|_{L^2(A_{i+1,i})}^2
+\mathcal E 4^{i} M_0^{-2}\| \bar p^{(k_i)}\|_{L^2(A_i)}^{2}
\le C C_\delta 4^{i} \|\bar p^{(k_i)}\|_{L^2(A_i)}^{2},
\end{equation}
where
$C_\delta=\delta^{-a}(\theta T)^{-1}+\mathcal E M_0^{-2}.$
Under the Strict Degree Condition, $2<{\mu_1}<(2-a)^*$, then applying Lemma~\ref{ParSob-1} with $r=2-a$ yields
\begin{align*}
& \|\bar p^{(k_{i+1})}\|_{L^{\mu_1}(A_{i+1,i+1})} =\Big(\int_{t_{i+1}}^T \int_U |\bar p^{(k_{i+1})}|^{\mu_1} dx dt \Big)^{1/{\mu_1}}\\
&\le C(1 +(T-t_{i+1})^{1/{\mu_1}} )\Big[ \sup_{[t_{i+1}, T]} \norm{\bar p^{(k_{i+1})} }_{L^2(U)} + \Big(\int_{t_{i+1}}^T \int_U |\nabla \bar p^{(k_{i+1})} |^{2-a} dx)dt\Big)^{1/(2-a)}\Big] \\
&\le C(1 +T)^{1/{\mu_1}} \Big[ \sup_{[t_{i+1}, T]} \norm{\bar p^{(k_{i+1})}\zeta_i }_{L^2(U)} + \Big(\int_{t_{i+1}}^T \int_U |\nabla \bar p^{(k_{i+1})} |^{2-a} \zeta_i dxdt\Big)^{1/(2-a)}\Big].
\end{align*}
Since $t_i\le t_{i+1}$, it follows that
\begin{equation} \label{Lr-p}
\| \bar p^{(k_{i+1})}\|_{L^{\mu_1}(A_{i+1,i+1})}
\le C (T+1)^{1/{\mu_1}} (F_{i}^{1/2}+F_{i}^{1/(2-a)}).
\end{equation}
Then, it follows from H\"{o}lder's inequality and \eqref{Lr-p} that
\begin{align*}
& \|\bar p^{(k_{i+1})}\|_{L^2(A_{i+1,i+1})}
\le \|\bar p^{(k_{i+1})}\|_{L^{\mu_1}(A_{i+1,i+1})} |A_{i+1,i+1}|^{1/2-1/{\mu_1}}\\
& \le \|\bar p^{(k_{i+1})}\|_{L^{\mu_1}(A_{i+1,i+1})} |A_{i+1,i}|^{1/2-1/{\mu_1}}
\le C (T+1)^{1/{\mu_1}} (F_{i}^{1/2}+F_{i}^{1/(2-a)}) |A_{i+1,i}|^{1/2-1/{\mu_1}}.
\end{align*}
This and \eqref{An.n}, \eqref{Fn} give
\begin{align*}
& \|\bar p^{(k_{i+1})}\|_{L^2(A_{i+1,i+1})}\\
&\le 4^i C (T+1)^{1/{\mu_1}} \Big ( C_\delta^{1/2}\|\bar p^{(k_i)}\|_{L^2(A_i)}+C_\delta^{1/(2-a)}\|\bar p^{(k_i)}\|_{L^2(A_i)}^{2/(2-a)}\Big ) M_0^{-1+2/{\mu_1}} \| p^{(k_i)}\|_{L^2(A_i)}^{1-2/{\mu_1}}\\
&\le 4^i C (T+1)^{1/{\mu_1}} M_0^{-1+2/{\mu_1}} \Big (C_\delta^{1/2} \|\bar p^{(k_i)}\|_{L^2(A_i)}^{2-2/{\mu_1}}+ C_\delta^{1/(2-a)}\|\bar p^{(k_i)}\|_{L^2(A_i)}^{1-2/{\mu_1}+2/(2-a)}\Big ).
\end{align*}
Now, let $Y_i=\|\bar p^{(k_i)}\|_{L^2(A_i)}$, we get
\begin{equation} \label{Yn-uni}
Y_{i+1}\le D 4^i \Big (Y_i^{2-2/{\mu_1}}+Y_i^{1-2/{\mu_1}+2/(2-a)} \Big )= D 4^i \Big (Y_i^{1+\mu_2}+Y_i^{1+\mu_3} \Big ),
\end{equation}
where $D=C (T+1)^{1/{\mu_1}} \big( C_\delta^{1/2} + C_\delta^{1/(2-a)} \big) M_0^{-\mu_2}$, and
\begin{equation}\label{muzero}
\mu_2= 1-\frac 2 {\mu_1} =\frac{4-a(n+2)}{(2-a)(n+2)},\quad \mu_3 = 2\Big (\frac{1}{2-a} - \frac{1}{{\mu_1}} \Big)=\frac{4}{(2-a)(n+2)}.
\end{equation}
To be able to apply \eqref{Y0mcond} with $m=2$ in Lemma~\ref{multiseq}, we will choose $M_0$ sufficiently large such that
\begin{equation}\label{condY1}
Y_0 \leq C\min\{D^{-1/\mu_2}, D^{-1/\mu_3} \}.
\end{equation}
Since $Y_0\le \| \bar{p}\|_{L^2(U\times (0,T))}$, it suffices to choose $M_0$ such that
\[
M_0 \ge C (T+1)^\frac1{{\mu_1}\mu_2} (C_\delta^\frac12+ C_\delta^\frac 1{2-a} )^\frac1{\mu_2} \|\bar{p}\|_{L^2(U\times (0,T))},
M_0\ge C (T+1)^\frac1{{\mu_1}\mu_2} (C_\delta^\frac12+ C_\delta^\frac 1{2-a} )^\frac1{\mu_2} \|\bar{p}\|_{L^2(U\times (0,T))}^{\mu_3/\mu_2}.
\]
Observe that, if $M_0\ge \sqrt {\mathcal E}$ then $C_\delta\le 1 + \frac 1 {\delta^a \theta T }\in (1, \infty) $. Thus, \eqref{condY1} holds for
\[
M_0= C\Big\{ \sqrt{\mathcal E}+ (T+1)^\frac1{{\mu_1}\mu_2} \big(1+\frac 1 { \delta^a \theta T}\big)^\frac1{(2-a)\mu_2}\Big( \|\bar p\|_{L^2(U\times (0,T))} + \|\bar p\|_{L^2(U\times (0,T))}^{\mu_3/\mu_2} \Big)\Big\}.
\]
With this choice of $M_0$, applying \eqref{Y0mcond} with $m=2$ in Lemma~\ref{multiseq} to \eqref{Yn-uni}, we obtain
$\displaystyle{\lim_{i\to\infty}}Y_i=0$. Consequently,
$
\int_{\theta T}^T\int_U |\bar p^{(M_0)}|^2 dxdt=0.
$
This implies $\bar p(x,t)\le M_0$ in $U\times (\theta T,T)$.
To prove the lower bound $\bar{p}\ge -M_0$, we replace $p$ by $-p$ and $\psi$ by $-\psi$ and
use the same argument. We conclude
\[
|\bar p(x,t)|\le M_0 \quad \text{in } U\times (\theta T,T).
\]
The proof of \eqref{mainest-1} is complete.
\end{proof}
Our next goal is to combine Propositions \ref{p-T.est} and \ref{local-L-infty} with $L^2$-estimates of $\bar{p}$ in \cite{HI2}
to derive the uniform (in time) $L^\infty$-estimate of $\bar{p}$. For our purpose, we need to introduce some notation.
Firstly, let us define two functions $f(t)$ and $\tilde f(t)$ for $t\ge 0$ by
\begin{equation}\label{b}
\begin{split}
f(t) & =\|\psi(t)\|_{L^\infty}^2+\|\psi(t)\|_{L^\infty}^{\frac{2-a}{1-a}} \quad \text{and}\quad
\tilde f(t)=\|\psi_t(t)\|_{L^\infty}^2+\|\psi_t(t)\|_{L^\infty}^{\frac{2-a}{1-a}}.
\end{split}
\end{equation}
Assume throughout that $\psi(\cdot,t)$ and $\psi_t(\cdot,t)$ belong $C([0,\infty),L^\infty(\Gamma))$, hence $f(t)$ and $\tilde f(t)$ belong to $C([0,\infty))$.
Whenever $f'(t)$ is mentioned, we implicitly assume that $f\in C^1((0,\infty))$.
Let $M_f(t)$ be a continuous, increasing majorant of $f(t)$ on $[0,\infty)$.
Let
\begin{equation}\label{Abeta} A=\limsup_{t\to\infty} f(t)\quad \text{and}\quad \beta=\limsup_{t\to\infty} [f'(t)]^-.\end{equation}
Again, whenever $\beta$ is used in subsequently statements,
it is understood that $f(t)\in C^1((0,\infty))$.
Also, for a function $u(x,t)$ defined on $U\times [0,\infty)$ we denote
\begin{equation} \label{J.def}
J_H[u](t)=\int_U H(|\nabla u(x,t)|)dx,
\end{equation}
where $H$ is the function defined by \eqref{Hxi}. Note from \eqref{Hcompare} that
\begin{equation}\label{JHcom}
c_1\int_U|\nabla u(x,t)|^{2-a} dx - c_3 \le J_H[u](t)\le c_2\int_U|\nabla u(x,t)|^{2-a} dx,\quad c_3>0.
\end{equation}
We now recall relevant estimates from \cite{HI2}.
\begin{theorem}[cf. \cite{HI2}, Theorem 4.3]\label{p-estimate} The following estimates hold
\begin{itemize}
\item[\textup{(i)}] For $t\ge 0$,
\begin{equation}\label{p-bar-bound1}
\norm{\bar p(t)}_{L^2}^2 +\int_0^t J_H[p](\tau)d\tau\le \norm{\bar p_0}_{L^2}^2 +C\int_0^t f(\tau)d\tau.
\end{equation}
\item[\textup{(ii)}] Assume that the Degree Condition holds, then
\begin{equation}\label{p-bar-bound}
\norm{\bar p(t)}_{L^2}^2 \le \norm{\bar p_0}_{L^2}^2 +C(1+ M_f(t)^\frac{2}{2-a})\quad \text{for all } t\ge 0.
\end{equation}
Moreover, if $A<\infty$ then
\begin{equation}\label{nonzerobeta} \limsup_{t\to\infty} \norm{\bar p(t)}_{L^2}^2 \le C (A+A^\frac2{2-a}),\end{equation}
and if $\beta<\infty$ then there is $T>0$ such that
\begin{equation}\label{unboundp} \norm{\bar p(t)}_{L^2}^2 \le C\big(1+\beta^\frac{1}{1-a}+f(t)^\frac{2}{2-a}\big) \quad \text{for all } t>T.\end{equation}
\end{itemize}
\end{theorem}
We finalize our main estimates for $L^\infty$-norm in the following theorem.
The exponent $\mu_2$ appearing below is already defined in \eqref{muzero}, and we also denote
\begin{equation}\label{mu1}
\mu_4= \frac{4}{(2-a)(4-a(n+2))}.
\end{equation}
\begin{theorem}\label{thm-1}
Assume \eqref{strictdeg}. Then there is a constant $C>0$ such that the following statements hold true.
\begin{itemize}
\item[\textup{(i)}] For $t>0$,
\begin{equation}\label{p1}
\|\bar p(t)\|_{L^\infty} \le C\Big(1+t^{-\frac 1{\mu_2(2-a)}}\Big) \Big\{ 1+ \|\bar p_0\|_{L^2}^{\mu_4(2-a)} +M_f(t)^{\mu_4} \Big\},
\end{equation}
\begin{equation} \label{pmax}
\|\bar p(t)\|_{L^\infty} \le C\Big( 1+ \|\bar p_0\|_{L^\infty}+\|\bar p_0\|_{L^2}^{\mu_4(2-a)} +M_f(t)^{\mu_4} \Big).
\end{equation}
\item[\textup{(ii)}] If $A<\infty$ then
\begin{equation}\label{p-2}
\limsup_{t\to\infty} \| \bar p(t)\|_{L^\infty} \le C \big( 1+ A^{\mu_4} \big).
\end{equation}
\item[\textup{(iii)}] If $\beta<\infty$ then there is $T>0$ such that
\begin{equation}\label{p-3} \| \bar p(t)\|_{L^\infty} \le C\Big\{ 1 +\beta^\frac{\mu_4(2-a)}{2(1-a)} + \sup_{[t-2,t]} \|\psi\|_{L^\infty}^\frac{\mu_4(2-a)}{1-a} \Big\} \quad \hbox{for all } t>T.\end{equation}
\end{itemize}
\end{theorem}
\begin{proof}
(i) We prove \eqref{p1} first.
For $t\in(0,1)$, applying \eqref{mainest-1} to $T_0=0$, $T=t$, $\delta=1$ and $\theta =1/2$ we obtain
\begin{align*}
\|\bar p(t)\|_{L^\infty}
&\le C\Big\{ (t+1)^{\frac 1{2}}(1+ \sup_{[0,t]}\|\psi\|_{L^\infty})^\frac{2-a}{2(1-a)})\\
&\quad+ (t+1)^\frac1{\mu_1 \mu_2} \big(1+ 2 t^{-1}\big)^\frac1{(2-a)\mu_2} \Big(\|\bar p\|_{L^2(U\times (0,t))} + \|\bar p\|_{L^2(U\times (0,t))}^{\frac{4}{4-a(n+2)}}\Big) \Big\}\\
&\le C (1+ t^{-1})^\frac1{(2-a)\mu_2} \Big(1+\sup_{[0,t]}\|\psi\|_{L^\infty}^\frac{2-a}{2(1-a)} + \sup_{[0,t]}\|\bar p\|_{L^2}^{\frac{4}{4-a(n+2)}}\Big) \Big\}.
\end{align*}
By estimate \eqref{p-bar-bound},
\begin{align*}
\|\bar p(t)\|_{L^\infty}&\le C t^{-\frac 1{\mu_2(2-a)}}\Big\{1+ M_f(t)^\frac12+ \Big(1+ \norm{\bar p_0}_{L^2} + M_f(t)^{\frac 1{2-a}}\Big)^{\frac{4}{4-a(n+2)}} \Big\}\\
&\le C t^{-\frac 1{\mu_2(2-a)}}\Big\{ 1+ \norm{\bar p_0}_{L^2}^\frac{4}{4-a(n+2)} +M_f(t)^\frac{4}{(2-a)(4-a(n+2))} \Big\}.
\end{align*}
Therefore, we obtain \eqref{p1} for $t\in(0,1)$.
For $t\in[1,\infty)$, applying \eqref{mainest-1} to $T_0=t-1$, $T=1$, $\delta=1$, and $\theta=1/2$, we have
\begin{equation*}
\|\bar p(t)\|_{L^\infty}\le C\Big\{ 1+\sup_{[t-1,t]} \|\psi\|_{L^\infty}^\frac{2-a}{2(1-a)} + \sup_{[t-1,t]} \norm{\bar p}_{L^2} + \sup_{[t-1,t]} \norm{\bar p(t)}_{L^2}^{\mu_4(2-a)} \Big\}.
\end{equation*}
Note that $\mu_4(2-a)> 1$, then H\"older's inequality gives
\begin{equation}\label{opp}
\|\bar p(t)\|_{L^\infty}\le C\Big\{ 1+\sup_{[t-1,t]} \|\psi\|_{L^\infty}^\frac{2-a}{2(1-a)} + \sup_{[t-1,t]} \norm{\bar p}_{L^2}^{\mu_4(2-a)} \Big\}.
\end{equation}
Then, again, combining \eqref{opp} with \eqref{p-bar-bound} and using H\"older's inequality yield \eqref{p1} for $t\in[1,\infty)$.
To prove \eqref{pmax}, we note that in the case $t\ge 1$, \eqref{p1} implies \eqref{pmax}.
Now, for $t\in (0,1),$ we have from inequality \eqref{pbaronly} that
\begin{equation}\label{pp}
\|\bar p(t)\|_{L^\infty}\le C \big( 1+\| \bar p_0\|_{L^\infty} + \sup_{[0,t]} \|\psi\|_{L^\infty}^\frac{1}{1-a} \big).
\end{equation}
Since $\|\psi(t)\|_{L^\infty}\le M_f(t)^\frac{1-a}{2-a}\le M_f(t)^{\mu_4}$, then \eqref{pmax} follows \eqref{pp}.
(ii) Since $A<\infty$, we see that
\begin{equation}\label{limF}
\limsup_{t\to\infty} (\sup_{[t-1,t]} \|\psi\|_{L^\infty})^\frac{2-a}{2(1-a)}\le \limsup_{t\to\infty} f(t)^\frac12 = A^\frac12.
\end{equation}
It follows \eqref{opp}, \eqref{nonzerobeta} and \eqref{limF} that
\begin{equation*}
\limsup_{t\to\infty} \| \bar p(t)\|_{L^\infty} \le C \Big\{ 1+A^\frac12+(A^\frac12+ A^{\frac 1{2-a}})^{\mu_4(2-a)} \Big\},
\end{equation*}
which implies \eqref{p-2}.
(iii) Because $\beta<\infty$, using \eqref{unboundp} in \eqref{opp} there is $T>0$ such that for all $t>T$,
\begin{equation*}\label{p-30}
\| \bar p(t)\|_{L^\infty} \le C\Big\{ 1+ \sup_{[t-1,t]} \|\psi\|_{L^\infty}^\frac{2-a}{2(1-a)}
+\big[1+\beta^\frac{1}{2(1-a)}+\sup_{[t-1,t]} \|\psi\|_{L^\infty}^\frac{1}{1-a}\big]^{\mu_4(2-a)} \Big\},\end{equation*}
and \eqref{p-3} follows.
\end{proof}
Concerning $\bar p(t)$ being small, as $t\to\infty$, rather than just being bounded, we have the following result.
\begin{theorem}\label{thmsplit}
Assume \eqref{strictdeg}. For any $\varepsilon>0$, there is $\delta_0>0$ such that
\begin{equation}\label{limsup:pbar0}
\text{if}\quad \limsup_{t\to\infty}\|\psi(t)\|_{L^\infty}<\delta_0 \quad\text{then}\quad \limsup_{t\to\infty} \|\bar p(t)\|_{L^\infty} < \varepsilon.
\end{equation}
Consequently,
\begin{equation}\label{limsup:pbar}
\text{if}\quad \lim_{t\to\infty}\|\psi(t)\|_{L^\infty}=0\quad\text{then}\quad \lim_{t\to\infty} \|\bar p(t)\|_{L^\infty} =0.
\end{equation}
\end{theorem}
\begin{proof}
Applying \eqref{mainest-1} to $T_0=t-1$, $T=1$ and $\theta=1/2$, we have
\begin{equation*}
\| \bar p(t)\|_{L^\infty} \le C\Big\{ \sqrt{\mathcal E(t)}+ (1+ \delta^{-a})^{\frac 1{(2-a)\mu_2}} \Big(\|\bar p\|_{L^2(U\times (t-1,t))} + \|\bar p\|_{L^2(U\times (t-1,t))}^\frac{4}{4-a(n+2)}\Big) \Big\},
\end{equation*}
where
$\mathcal E(t) = \delta^{2-a}+\delta^{-a}\sup_{[t-1,t]}\|\psi\|_{L^\infty}^2+ \delta^{-\frac {a(2-a)}{1-a}} \sup_{[t-1,t]}\|\psi\|_{L^\infty}^\frac{2-a}{1-a}.$
Then, it follows that
\begin{equation}\label{prelim0}
\|\bar p(t)\|_{L^\infty}\le C\sqrt{\mathcal E(t)} + C (1+\delta^{-a})^\frac1{(2-a)\mu_2} \Big\{\sup_{[t-1, t]} \norm{\bar p}_{L^2} +\sup_{[t-1, t]} \norm{\bar p}_{L^2}^{\frac{4}{4-a(n+2)}}\Big\}.
\end{equation}
Let $\delta_0\in(0,1)$, then
$A\le \delta_0^2+\delta_0^\frac{2-a}{1-a}\le 2\delta_0^2$,
and therefore, by \eqref{nonzerobeta},
\begin{equation*}
\limsup_{t\to\infty} (\sup_{[t-1, t]} \norm{\bar p}_{L^2} )\le C(A^\frac12+A^\frac1{2-a})\le C \delta_0.
\end{equation*}
Hence, we obtain
\begin{align*}
&\limsup_{t\to\infty}(1+\delta^{-a})^\frac1{(2-a)\mu_2} \Big\{\sup_{[t-1, t]} \norm{\bar p}_{L^2} +\sup_{[t-1, t]} \norm{\bar p}_{L^2}^{\frac{4}{4-a(n+2)}}\Big\}
\le C (1+\delta^{-a})^\frac1{(2-a)\mu_2}\delta_0.
\end{align*}
Also,
\begin{equation*}
\limsup_{t\to\infty} \mathcal E(t)\le \delta^{2-a} +\delta^{-a}\delta_0^2 + \delta^{-\frac {a(2-a)}{1-a}} \delta_0^\frac{2-a}{1-a} \le \delta^{2-a} + 2\delta^{-\frac {a(2-a)}{1-a}} \delta_0^2.
\end{equation*}
Therefore, for any $\delta>0$, we have from \eqref{prelim0} that
\begin{equation}\label{limwithdelta}
\limsup_{t\to\infty} \|\bar p(t)\|_{L^\infty} \le C\big(\delta^{1-\frac a 2} +[ \delta^{-\frac {a(2-a)}{2(1-a)}} +(1+\delta^{-a})^\frac1{(2-a)\mu_2} ]\delta_0\big).
\end{equation}
Now, choose $\delta$ sufficiently small such that $2C\delta^{1-a/2} <\varepsilon$. Then we can choose $\delta_0>0$ even sufficiently smaller so that
$
[\delta^{-\frac {a(2-a)}{2(1-a)}} +(1+\delta^{-a})^\frac1{(2-a)\mu_2}]\delta_0 < \delta^{1-a/2}.
$
From this, the desired estimates \eqref{limsup:pbar0} follows \eqref{limwithdelta}.
The statement \eqref{limsup:pbar} obviously is a consequence of \eqref{limsup:pbar0}. The proof is complete.
\end{proof}
\subsection{Estimates for pressure's gradient}
\label{Lsgrad-sec}
In this subsection, we establish the interior $L^s$-estimate of $\nabla p$, for all $s>0$.
We follow the approach in \cite{LadyParaBook68}.
First is a basic estimate for $\nabla p$ which prepares for our iteration later.
\begin{lemma} \label{s.it-L} For each $s\geq 0$, there is a constant $C>0$ depending on $s$ such that for any $T>0$ and smooth cut-off function
$\zeta(x) \in C_c^\infty(U)$, the following estimate holds
\begin{equation}\label{preLs2}
\begin{aligned}
& \sup_{[0,T]} \int_U |\nabla p(x, t)|^{2s+2} \zeta^2 dx
+C \int_0^T\int_U K(|\nabla p|) |\nabla^2 p|^2 |\nabla p|^{2s} \zeta^2 dx dt \\
&\le \int_U |\nabla p_0(x)|^{2s+2} \zeta^2 dx
+ C \int_0^T \int_U K(|\nabla p|) |\nabla p|^{2s+2}|\nabla \zeta|^2 dx dt.
\end{aligned}
\end{equation}
\end{lemma}
\begin{proof}
In this proof, we use Einstein's summation convention, that is, when an index variable appears twice in a single term it implies summation of that term over all the values of the index.
Multiplying the equation \eqref{eqorig} by
$-\nabla \cdot (|\nabla p|^{2s}\nabla p \zeta^2)$ and integrating the resultant over $U$, we obtain
\begin{align*}
\frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla p|^{2s+2} \zeta^2 dx
= -\int_U \partial_i (K(|\nabla p|)\partial_i p) \, \partial_j (|\nabla p|^{2s} \partial_j p \zeta^2) dx.
\end{align*}
This equality and the integration by parts yield
\begin{align*}
& \frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla p|^{2s+2} \zeta^2 dx
= -\int_U \partial_j [K(|\nabla p|)\partial_i p] \, \partial_i [|\nabla p|^{2s} \partial_j p \zeta^2] dx\\
&= -\int_U \partial_j [K(|\nabla p|)\partial_i p] \, \partial_i \partial_j p \, |\nabla p|^{2s} \zeta^2 dx
-\int_U \partial_j [K(|\nabla p|)\partial_i p]\, \partial_j p \, |\nabla p|^{2s} \partial_i [\zeta^2] dx\\
&\quad -\int_U \partial_j [K(|\nabla p|)\partial_i p] \,\partial_j p \, [2s |\nabla p|^{2s-2} \partial_i\partial_m p\partial_m p]\, \zeta^2 dx.
\end{align*}
We rewrite it in the following form
\begin{align*}
& \frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla p|^{2s+2} \zeta^2 dx
= -\int_U \Big[\partial_{y_l} (K(|y|) y_i)\Big|_{y=\nabla p} \partial_j\partial_l p\Big]\partial_j \partial_i p |\nabla p|^{2s} \zeta^2 dx\\
&\quad -2\int_U\Big[ \partial_{y_l} (K(|y|) y_i)\Big|_{y=\nabla p} \partial_j\partial_l p \Big] \partial_j p \, |\nabla p|^{2s} \, \zeta \partial_i \zeta dx\\
&\quad -2s\int_U \Big[\partial_{y_l} (K(|y|) y_i)\Big|_{y=\nabla p} \partial_j\partial_l p \Big]\partial_j p \, (|\nabla p|^{2s-2} \partial_i\partial_m p\partial_m p)\, \zeta^2 dx.
\end{align*}
We denote the three terms on the right-hand side by $I_1$, $I_2$ and $I_3$, and estimate each of them.
By \eqref{K-est-2} one can easily prove for any $y,z\in \ensuremath{\mathbb R}^n$ that
\begin{equation*}
z^T \nabla (K(|y|)y) z\ge (1-a)K(|y|)|z|^2\quad \text{and} \quad |\nabla (K(|y|)y)|\le (1+a)K(|y|).
\end{equation*}
It follows that
\begin{align*}
I_1
&= -\int_U \partial_{y_l} [K(|y|) y_i]\Big|_{y=\nabla p} \partial_l(\partial_j p) \partial_i (\partial_j p) \, |\nabla p|^{2s} \zeta^2 dx\\
&\le -(1-a)\sum_j \int_U K(|\nabla p|) |\nabla(\partial_j p)|^2 |\nabla p|^{2s} \zeta^2 dx\\
& = -(1-a) \int_U K(|\nabla p|) |\nabla^2 p|^2 \, |\nabla p|^{2s} \zeta^2 dx.
\end{align*}
Moreover, we have
\begin{equation*}
|I_2|\le 2(1+a)\int_U K(|\nabla p|) |\nabla^2 p| |\nabla p|^{2s+1} \zeta |\nabla \zeta| dx,
\end{equation*}
and
\begin{align*}
I_3
&= -2s\int_U \partial_{y_l} [K(|y|) y_i]\Big|_{y=\nabla p} (\partial_l\partial_j p\partial_j p) (\partial_i \partial_m p \partial_m p) \, |\nabla p|^{2s-2} \zeta^2 dx\\
&= -2s\int_U \partial_{y_l} [K(|y|) y_i]\Big|_{y=\nabla p} \Big(\partial_l \frac12|\nabla p|^2\Big) \Big(\partial_i \frac12|\nabla p|^2\Big)\, |\nabla p|^{2s-2} \zeta^2 dx\\
&\le -2(1-a)s\int_U K(| \nabla p|)\Big | \nabla \Big(\frac12|\nabla p|^2\Big)\Big |^2 |\nabla p|^{2s-2} \zeta^2 dx\le 0.
\end{align*}
Combining these estimates together with Young's inequality, we see that
\begin{align*}
& \frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla p|^{2s+2} \zeta^2 dx
+(1-a) \int_U K(|\nabla p|) |\nabla^2 p|^2 |\nabla p|^{2s} \zeta^2 dx\\
&\le 2(1+a)\int_U K(|\nabla p|) |\nabla^2 p| |\nabla p|^{2s+1} \zeta|\nabla \zeta| dx\\
& \le \frac{1-a}{2} \int_{U'} K(|\nabla p|) |\nabla^2 p|^2 |\nabla p|^{2s} \zeta^2 dx + C \int_U K(|\nabla p|) |\nabla p|^{2s+2}|\nabla \zeta|^2 dx.
\end{align*}
Thus,
\begin{equation}\label{gradid}
\begin{aligned}
& \frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla p|^{2s+2} \zeta^2 dx
+\frac{1-a}{2} \int_U K(|\nabla p|) |\nabla^2 p|^2 |\nabla p|^{2s} \zeta^2 dx\\
&\le C \int_U K(|\nabla p|) |\nabla p|^{2s+2}|\nabla \zeta|^2 dx.
\end{aligned}
\end{equation}
Integrating this inequality in time yields \eqref{preLs2}.
\end{proof}
In order to iterate \eqref{preLs2} in $s$ we need an embedding similar to Lemma 5.4 on page 93 in \cite{LadyParaBook68}.
For our degenerate equation, the following version has the key weight function $K(|\nabla w|)$.
\begin{lemma}\label{LUK} For each $s\geq1$, there exists a constant $C>0$ depending on $s$ such that for each smooth cut-off function
$\zeta(x) \in C_c^\infty(U)$, the following inequality holds
\begin{equation*}
\begin{aligned}
\int_U K(|\nabla w|) |\nabla w|^{2s+2} \zeta^2 dx
\le C \sup_{\text{\rm supp} \zeta } |w|^2 & \left \{ \int_U K(|\nabla w|)|\nabla w|^{2s-2} |\nabla^2 w|^2 \zeta^2 dx \right. \\
& \quad \quad \quad + \left. \int_U K(|\nabla w|)|\nabla w|^{2s} |\nabla \zeta|^2 dx \right\},
\end{aligned}
\end{equation*}
for every sufficiently regular function $w(x)$ such that the right hand side is well-defined.
\end{lemma}
\begin{proof}
Again, the Einstein's summation convention is used in this proof.
Let $I = \int_U K(|\nabla w|) |\nabla w|^{2s+2} \zeta^2 dx$.
From direct calculations, we see that
\begin{align*}
I
&=\int_U K(|\nabla w|) |\nabla w|^{2s} \partial_i w\partial_i w \zeta^2 dx
= - \int_U \partial_i (K(|\nabla w|) |\nabla w|^{2s} \partial_i w \zeta^2)\cdot w dx\\
& = - \int_U\Big( K'(|\nabla w|)\frac{\partial_i\partial_j w\partial_j w}{|\nabla w|} \Big) |\nabla w|^{2s} \partial_i w\cdot \zeta^2 \cdot w dx - \int_U K(|\nabla w|)|\nabla w|^{2s} \Delta w \cdot \zeta^2 \cdot w dx\\
&\quad - 2s \int_U K(|\nabla w|)\Big(|\nabla w|^{2s-2} \partial_i\partial_j w \partial_j w\Big)\cdot \partial_i w \cdot \zeta^2 \cdot w dx
-2 \int_U K(|\nabla w|)|\nabla w|^{2s} \partial_i w \zeta \partial_i \zeta \cdot w dx.
\end{align*}
From this and \eqref{K-est-2}, it follows that
\begin{align*}
I&\le a \int_U K(|\nabla w|)|\nabla^2 w| |\nabla w|^{2s} \zeta^2 |w| dx
+ \int_U K(|\nabla w|)|\nabla w|^{2s} |\Delta w| \zeta^2 |w| dx\\
&\quad + 2 s \int_U K(|\nabla w|)|\nabla w|^{2s-2} |\nabla^2 w| |\nabla w|^2 \zeta^2 |w| dx
+2 \int_U K(|\nabla w|)|\nabla w|^{2s} |\nabla w| \zeta |\nabla \zeta| |w|dx\\
& \le C \int_U K(|\nabla w|)|\nabla w|^{2s} |\nabla^2 w| \zeta^2 |w| dx
+ C\int_U K(|\nabla w|)|\nabla w|^{2s+1} \zeta |\nabla \zeta| |w| dx.
\end{align*}
This last inequality and the Young's inequality imply that
\begin{equation*}
I \le \frac{1}{2} I
+C \Big \{ \int_U K(|\nabla w|)|\nabla w|^{2s-2} |\nabla^2 w|^2 \zeta^2 |w|^2 dx
+ \int_U K(|\nabla w|)|\nabla w|^{2s} |\nabla \zeta|^2 |w|^2dx \Big \}.
\end{equation*}
Therefore, we obtain
\begin{equation*}
I \le C \sup_{{\rm supp} \zeta} |w|^2 \Big \{ \int_U K(|\nabla w|)|\nabla w|^{2s-2} |\nabla^2 w|^2 \zeta^2 dx
+ \int_U K(|\nabla w|)|\nabla w|^{2s} |\nabla \zeta|^2 dx \Big\}.
\end{equation*}
This completes the proof.
\end{proof}
We combine Lemmas \ref{s.it-L} and \ref{LUK} now. By applying Lemma \ref{LUK} to $\bar p$ with $s+1$ in place of $s$, we have
\begin{equation*}
\begin{aligned}
\int_0^T \int_U K(|\nabla p|) |\nabla p|^{2s+4} \zeta^2 dx dt
\le C \sup_{ \rm{supp}\zeta } |\bar p|^2 & \left \{ \int_0^T\int_U K(|\nabla p|)|\nabla p|^{2s} |\nabla^2 p|^2 \zeta^2 dx dt \right.\\
& + \left. \int_0^T \int_U K(|\nabla p|)|\nabla p|^{2s+2} |\nabla \zeta|^2 dx dt \right \}.
\end{aligned}
\end{equation*}
This last inequality and \eqref{preLs2} imply for $s\ge 0$ that
\begin{equation}\label{Kgrad1}
\begin{aligned}
\int_0^T \int_U K(|\nabla p|) |\nabla p|^{2s+4} \zeta^2 dx dt
& \le C \sup_{{\rm supp}\zeta} |\bar p|^2 \left[ \int_U |\nabla p_0(x)|^{2s+2} \zeta^2 dx \right. \\
& \quad + \left. \int_0^T \int_U K(|\nabla p|) |\nabla p|^{2s+2}|\nabla \zeta|^2 dx dt \right].
\end{aligned}
\end{equation}
Also, from \eqref{preLs2}, we have for $s\ge 2$ that
\begin{equation}\label{preLs3}
\begin{aligned}
& \sup_{[0,T]} \int_U |\nabla p(x, t)|^{s} \zeta^2 dx
\le \int_U |\nabla p_0(x)|^{s} \zeta^2 dx
+ C \int_0^T \int_U K(|\nabla p|) |\nabla p|^{s}|\nabla \zeta|^2 dx dt.
\end{aligned}
\end{equation}
We are ready to iterate the relation \eqref{Kgrad1}. Hereafterward, we consider $U'\Subset U$, that is, $U'$ is an open set compactly contained in $U$.
\begin{proposition}\label{interp}
For $U'\Subset V \subset U$ and $s\ge 2$ there exists a constant $C>0$ depending on $U'$, $V$ and $s$ such that for any $T>0$, we have
\begin{equation}\label{Kgrad50}
\begin{split}
&\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C \Big(1+\sup_{ [0,T]} \|\bar p\|_{L^\infty(V)}\Big)^{s-2} \\
&\quad \cdot \Big\{ \int_{U} |\nabla p_0(x)|^{2} dx +\int_{U} |\nabla p_0(x)|^{s-2} dx + \int_0^T \int_{U} K(|\nabla p|) |\nabla p|^2 dx dt\Big\}.
\end{split}
\end{equation}
\end{proposition}
\begin{proof} Let $m \in \mathbb{N}, m \geq 2$ be a fixed number. Then, let $\{U_k\}_{k=0}^m$ be a family of open, smooth domains in $U$ such that
$U'=U_m\Subset U_{m-1}\Subset U_{m-2}\Subset \ldots \Subset U_1 \Subset U_0=V\subset U$. For each $k =1,2,\cdots, m$, let $\zeta_k(x)$ be a smooth cut-off function
which is equal to one on $U_k$ and zero on $U\setminus U_{k-1}$. There is a positive constant $C>0$ depending on $\zeta_k$, $k=1,2,\ldots,m$, such that
$ |\nabla \zeta_k| \leq C$, for all $k = 1, 2,\cdots, m.$
Also, for each integer $k\ge 1$ and $s_0 \geq 0$, we define
$$X_k=\int_0^T \int_{U_k} K(|\nabla p|) |\nabla p|^{2k+s_0} dxdt, \quad A_k=\int_{U_k} |\nabla p_0(x)|^{2k+s_0} dx.$$
Then, by applying \eqref{Kgrad1} with $\zeta =\zeta_k$ and $2s+2=2k+s_0$, we see that
\begin{equation}\label{Xk1}
X_{k+1}\le C_{k,m,s_0} N_0 (A_k + X_k), \quad k = 1,2,\cdots m-1,
\end{equation}
where $N_0=\sup_{V\times[0,T]} |\bar p|^2=\sup_{[0,T]}\|\bar p\|_{L^\infty(V)}^2$.
Letting $C_m=\max\{C_{k,m,s_0}:k=1,2,\ldots , m-1\}$, we have from \eqref{Xk1} that
\begin{equation*}
X_{k+1}\le C_m N_0 (A_k + X_k), \quad k = 1,2,\cdots m-1,
\end{equation*}
This inequality particularly yields
\begin{equation*}
\begin{split}
X_m&\le (C_m N_0) A_{m-1}+(C_m N_0)^2 A_{m-2}+\cdots+(C_m N_0)^{m-1} A_1 + (C_m N_0)^{m-1} X_1\\
&\le C[( N_0 A_{m-1}+N_0^2 A_{m-2}+\cdots+N_0^{m-1} A_1) + N_0^{m-1} X_1],
\end{split}
\end{equation*}
with $C = C(m)$ depending on all sets $U_k$ and functions $\zeta_k$, for $k=1,2,\ldots,m-1$.
In other words, we have proved that for each integer $m \geq 2$ and real number $s\ge 0$, there is $C = C(m,s_0)$ such that
\begin{equation*}\label{grad1}
\begin{split}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{2m+s_0} dx dt & \le C \sum_{i=1}^{m-1} N_0^i\int_{U_{m-i}} |\nabla p_0(x)|^{2(m-i)+s_0} dx \\
& \quad +C N_0^{m-1}\int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^{2+s_0} dx dt.
\end{split}
\end{equation*}
By using Young's inequality, one can rewrite this inequality as
\begin{multline}\label{grad15}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{2m+s_0} dx dt \le C (N_0+N_0^{m-1}) \\
\cdot \Big\{ \int_{U} |\nabla p_0(x)|^{2+s_0} dx +\int_{U} |\nabla p_0(x)|^{2m+s_0-2} dx
+\int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^{2+s_0} dx dt\Big\}.
\end{multline}
In particular, with $m=2$ and $s_0=0$, \eqref{grad15} becomes
\begin{align}\label{Kgrad3}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^4 dx
&\le C N_0 \Big\{ \int_{U} |\nabla p_0(x)|^2 dx
+ \int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^2 dx dt\Big\}.
\end{align}
This implies \eqref{Kgrad50} when $s =4$. In case $ s \in (2,4)$, let $\alpha$ and $\beta$ be two positive numbers such that
\begin{equation}\label{salbe}
\frac 1 s =\frac \alpha{2}+\frac\beta {4}\quad\text{and} \quad \alpha+\beta=1.
\end{equation}
Then, using interpolation inequality, we get
\begin{align*}
\Big(\int_0^T\int_{U'} K(|\nabla p|)|\nabla p|^s dx dt\Big)^{\frac 1 s} &\le \Big(\int_0^T\int_{U_1} K(|\nabla p|) |\nabla p|^{2} dx dt\Big)^{\frac \alpha {2}} \Big(\int_0^T\int_{U'} K(|\nabla p|) |\nabla p|^{4} dx dt\Big)^{\frac \beta {4}}
\end{align*}
From this and \eqref{Kgrad3}, it follows that
\begin{align*}
\int_0^T\int_{U'} K(|\nabla p|)|\nabla p|^s dx dt
&\le C N_0^{\frac {\beta s} {4}} \Big\{ \int_{U} |\nabla p_0(x)|^2 dx
+ \int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^2 dx\Big\}.
\end{align*}
From \eqref{salbe} follows $\beta s/4=s/2-1$. Thus, we have
\begin{equation*}\label{Kgrad4}
\int_0^T\int_{U'} K(|\nabla p|)|\nabla p|^s dx dt
\le C N_0^{\frac {s} {2}-1} \Big\{ \int_{U} |\nabla p_0(x)|^2 dx
+ \int_0^T \int_{U} K(|\nabla p|) |\nabla p|^2 dx\Big\}.
\end{equation*}
This implies \eqref{Kgrad50} for $s\in(2,4)$. Therefore, we have proved \eqref{Kgrad50} with $s \in (2, 4]$. Now, for $s>4$, there are a number $s_0\in (0, 2]$ and an integer $m\ge 2$ such that $s=s_0+2m$. From \eqref{grad15}, we have
\begin{align*}
&\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C (N_0+N_0^{m-1}) \Big\{ \int_{U} |\nabla p_0(x)|^{2+s_0} dx +\int_{U} |\nabla p_0(x)|^{2m+s_0-2} dx\\
&\quad +\int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^{2+s_0} dx dt\Big\}\\
&\le C (N_0+N_0^\frac{s-s_0-2}{2}) \Big\{ \int_{U} |\nabla p_0(x)|^{2+s_0} dx +\int_{U} |\nabla p_0(x)|^{s-2} dx
+\int_0^T \int_{U_1} K(|\nabla p|) |\nabla p|^{2+s_0} dx dt\Big\}.
\end{align*}
Since $s_0+2\in(2,4]$, it follows from the last inequality and estimate \eqref{Kgrad50} already proved for the case $s \in (2,4]$ with $U_1$ in place of $U'$ and $s_0+2$ in place of $s$ that
\begin{align*}
&\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C (N_0+N_0^\frac{s-s_0-2}{2}) \Big\{ \int_{U} |\nabla p_0(x)|^{2+s_0} dx +\int_{U} |\nabla p_0(x)|^{s-2} dx\\
&\quad + N_0^{\frac {s_0} {2}} \Big( \int_{U} |\nabla p_0(x)|^2 dx
+ \int_0^T \int_{U} K(|\nabla p|) |\nabla p|^2 dx dt \Big)\Big\}\\
&\le C (1+N_0^\frac{s-2}{2}) \Big\{ \int_{U} |\nabla p_0(x)|^{2} dx +\int_{U} |\nabla p_0(x)|^{s-2} dx+ \int_0^T \int_{U} K(|\nabla p|) |\nabla p|^2 dxdt \Big\}.
\end{align*}
From this, the desired estimate \eqref{Kgrad50} follows and the proof is complete.
\end{proof}
Now, we give specific estimates for the $L^s$-norm of $\nabla p$ in terms of initial and boundary data.
We denote
\begin{equation}\label{Nf1}
\mathcal K_{1,T} = 1+\sup_{[0,T]}\|\psi\|_{L^\infty}.
\end{equation}
\begin{theorem} \label{thm312}
Assume \eqref{strictdeg}. Then for $s\ge 2$, there is a positive constant $C$ depending on $U'$ and $s$ such that for any $T >0$
\begin{equation}\label{grad20}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C L_1(s) (T+1)^{\frac{2(s-2)}{2-a}+1} \mathcal K_{1,T}^{\frac{s-a}{1-a}},
\end{equation}
\begin{equation}\label{grad20pw}
\sup_{[0,T]}\int_{U'} |\nabla p(x,t)|^{s} dx
\le C L_2(s) (T+1)^{\frac{2(s-2)}{2-a}+1} \mathcal K_{1,T}^{\frac{s-a}{1-a}},
\end{equation}
where
\begin{align*}
L_1(s)&=L_1(s;[p_0])\mathbin{\buildrel \rm def \over \longequals} (1+ \|\bar p_0\|_{L^\infty})^{s-2} \Big ( 1+\int_U |\nabla p_0(x)|^{\max\{2,s-2\}} dx \Big),\\
L_2(s)&=L_2(s;[p_0])\mathbin{\buildrel \rm def \over \longequals} (1+ \|\bar p_0\|_{L^\infty})^{s-2} \Big ( 1+\int_U |\nabla p_0(x)|^{s} dx \Big).
\end{align*}
\end{theorem}
\begin{proof} From the inequality \eqref{pbaronly} in Proposition \ref{p-T.est}, it follows that
\begin{equation*}\label{Lamdbound}
\sup_{[0,T]} \|\bar p\|_{L^\infty} \le C \Big\{ \|\bar p_0\|_{L^\infty} + (T+1)^\frac2{2-a}\mathcal K_{1,T}^\frac1{1-a} \Big\}.
\end{equation*}
Thanks to \eqref{p-bar-bound1} of Theorem \ref{p-estimate}, we have
\begin{equation}\label{JH2-a}
\int_0^T \int_U K(|\nabla p|) |\nabla p|^2 dxdt\le \norm{\bar p_0}_{L^2}^2 +C\int_0^T f(\tau)d\tau.
\end{equation}
These inequalities together with \eqref{Kgrad50} of Proposition \ref{interp} yield
\begin{multline*
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C \Big ( \|\bar p_0\|_{L^\infty} + (T+1)^\frac2{2-a}\mathcal K_{1,T}^\frac1{1-a} \Big )^{s-2} \\
\cdot\Big (\|\bar p_0\|_{L^2}^2+ \int_{U} \big(|\nabla p_0(x)|^{2}+|\nabla p_0(x)|^{s-2}\big) dx+\int_0^T f(\tau)d\tau\Big).
\end{multline*}
Applying Poincar\'e's inequality to $\bar p_0$, and using Young's inequality, we obtain
\begin{multline}\label{gpr1}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C \Big ( \|\bar p_0\|_{L^\infty} + (T+1)^\frac2{2-a}\mathcal K_{1,T}^\frac1{1-a} \Big )^{s-2} \\
\cdot\Big ( 1+\int_U |\nabla p_0(x)|^{\max\{2,s-2\}} dx+\int_0^T f(\tau)d\tau\Big),
\end{multline}
Then \eqref{grad20} follows.
It follows from \eqref{preLs3} and \eqref{grad20} that
\begin{multline*
\sup_{[0,T]}\int_{U'} |\nabla p(x,t)|^{s} dx
\le \int_{U'} |\nabla p_0(x)|^{s} dx + C \Big ( \|\bar p_0\|_{L^\infty} + (T+1)^\frac2{2-a}\mathcal K_{1,T}^\frac1{1-a} \Big )^{s-2} \\
\cdot\Big (1+\int_{U} \big(|\nabla p_0(x)|^{\max\{2,s-2\}} dx+\int_0^T f(\tau)d\tau\Big),
\end{multline*}
Then by Young's inequality we obtain
\begin{multline}\label{gpr2}
\sup_{[0,T]}\int_{U'} |\nabla p(x,t)|^{s} dx
\le C \Big ( \|\bar p_0\|_{L^\infty} + (T+1)^\frac2{2-a}\mathcal K_{1,T}^\frac1{1-a} \Big )^{s-2} \\
\cdot\Big (1+ \int_{U} |\nabla p_0(x)|^{s} dx+\int_0^T f(\tau)d\tau\Big).
\end{multline}
Then \eqref{grad20pw} follows.
\end{proof}
Recall that $f$ is defined in \eqref{b} and $M_f$ is a continuous increasing majorant of $f$. Our next result is similar to Theorem \ref{thm312} but the estimates do not contain the power growth in $T$.
\begin{theorem}\label{cor311}
Assume \eqref{strictdeg}. For $s\ge 2$, there is a positive constant $C$ depending on $U'$ and $s$ such that for any $T >1$ we have
\begin{equation}\label{Kgrad55}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C L_3(s) \Big(1+M_f(T) \Big)^{\mu_4(s-2)}\Big\{1+ \int_0^T f(t)dt\Big\},
\end{equation}
\begin{equation}\label{Kgrad55pw}
\sup_{[0,T]}\int_{U'} |\nabla p(x,t)|^{s} dx
\le C L_4(s)\Big(1+M_f(T) \Big)^{\mu_4(s-2)}\Big\{1+ \int_0^T f(t)dt\Big\},
\end{equation}
where
\begin{align*}
L_3(s)&= L_3(s;[p_0])\mathbin{\buildrel \rm def \over \longequals} \Big( 1+ \|\bar p_0\|_{L^\infty}+\|\bar p_0\|_{L^2}^{\mu_4(2-a)}\Big)^{s-2} \Big\{ 1+\int_{U} |\nabla p_0(x)|^{\max\{2,s-2\}} dx\Big\},\\
L_4(s)&=L_4(s;[p_0])\mathbin{\buildrel \rm def \over \longequals} \Big( 1+ \|\bar p_0\|_{L^\infty}+\|\bar p_0\|_{L^2}^{\mu_4(2-a)}\Big)^{s-2} \Big\{ 1+\int_{U} |\nabla p_0(x)|^{s} dx\Big\}.
\end{align*}
\end{theorem}
\begin{proof}
The proof is the same as in Theorem \ref{thm312} with the use of estimate \eqref{pmax} in place of \eqref{pbaronly}.
Instead of \eqref{grad20} and \eqref{grad20pw}, we have, respectively,
\begin{equation}\label{gpr3}
\begin{aligned}
\int_0^T \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
&\le C \Big( 1+ \|\bar p_0\|_{L^\infty}+\|\bar p_0\|_{L^2}^{\mu_4(2-a)} +M_f(T)^{\mu_4} \Big)^{s-2}\\
&\quad \cdot \Big\{ 1+\int_{U} |\nabla p_0(x)|^{\max\{2,s-2\}} dx+ \int_0^T f(t)dt\Big\},
\end{aligned}
\end{equation}
\begin{equation}\label{gpr4}
\begin{aligned}
\sup_{[0,T]}\int_{U'} |\nabla p(x,t)|^{s} dx
&\le C \Big( 1+ \|\bar p_0\|_{L^\infty}+\|\bar p_0\|_{L^2}^{\mu_4(2-a)} +M_f(T)^{\mu_4} \Big)^{s-2}\\
&\quad \cdot \Big\{ 1+\int_{U} |\nabla p_0(x)|^{s} dx+ \int_0^T f(t)dt\Big\}.
\end{aligned}
\end{equation}
Then \eqref{Kgrad55} and \eqref{Kgrad55pw} follow, respectively.
\end{proof}
In \eqref{Kgrad55} and \eqref{Kgrad55pw}, letting $T\to\infty$, we obtain the following.
\begin{corollary}\label{corMM}
Under the Strict Degree Condition, if
\begin{equation}\label{xgood}
\Upsilon_1 \mathbin{\buildrel \rm def \over \longequals} 1+\sup_{[0,\infty)}f <\infty \quad \text{and}\quad \Upsilon_2 \mathbin{\buildrel \rm def \over \longequals} 1+\int_0^\infty f(t)dt<\infty,
\end{equation}
then for any $s\ge 2$, there is a constant $C=C(U',s)>0$ such that
\begin{equation} \label{Kgrad60}
\int_0^\infty \int_{U'} K(|\nabla p|) |\nabla p|^{s} dx dt
\le C L_3(s) \Upsilon_1^{\mu_4(s-2)} \Upsilon_2,
\end{equation}
\begin{equation} \label{Kgrad60pw}
\sup_{[0,\infty)}\int_{U'} |\nabla p(x,t)|^{s} dx
\le C L_4(s) \Upsilon_1^{\mu_4(s-2)} \Upsilon_2.
\end{equation}
\end{corollary}
Using the preceding theorems and the properties of function $K(\cdot)$, we can derive the following direct estimates for $\int_0^T\int_{U'}|\nabla p|^s dxdt$.
\begin{theorem} \label{cor312}
Assume \eqref{strictdeg}. For $s\ge 2-a$, there is
a constant $C>0$ such that for any $T>0$ we have
\begin{equation}\label{grad39}
\int_0^T \int_{U'} |\nabla p|^{s} dx dt \le C \le C L_1(s+a) (T+1)^{\frac{2s}{2-a}-1} \mathcal K_{1,T}^{\frac{s}{1-a}},
\end{equation}
and, alternatively,
\begin{equation}\label{Kgrad70}
\int_0^T \int_{U'} |\nabla p|^{s} dx dt
\le CT + C L_3(s+a) \Big(1+M_f(T) \Big)^{\mu_4(s+a-2)}\Big\{1+ \int_0^T f(t)dt\Big\},
\end{equation}
where the positive numbers $L_1(\cdot)$ and $L_3(\cdot)$ are defined in Theorems \ref{thm312} and \ref{cor311}.
\end{theorem}
\begin{proof}
By relation \eqref{Km},
\begin{equation}\label{KgradsRel}
\int_0^T \int_{U'} |\nabla p|^{s} dx dt \le C T + \int_0^T\int_{U'} K(|\nabla p|)|\nabla p|^{s+a} dx dt.
\end{equation}
The last integral is estimated by applying \eqref{grad20} with $s+a$ replacing $s$.
As a result, we obtain \eqref{grad39}.
Now, using relation \eqref{KgradsRel} and applying \eqref{Kgrad55} for $s+a$ in place of $s$, we obtain \eqref{Kgrad70}.
\end{proof}
\subsection{Estimates for pressure's time derivative}
\label{Lpt-sec}
In this subsection, we derive the interior $L^\infty$-norm of $\bar p_t$.
Throughout this subsection $U'\Subset U$.
Under the Strict Degree Condition, let
\begin{equation} \label{sdef}
{\mu_5}=4\Big(1-\frac 1 {(2-a)^*}\Big),\quad
\mu_6=1-\frac2 {\mu_5},\quad
\mu_7=\frac 2 {{\mu_5}(2-a)}.
\end{equation}
Then ${\mu_5}\in (2,(2-a)^*)$ and $\mu_6,\mu_7\in(0,1).$
\begin{proposition} \label{Lqbar} Assume \eqref{strictdeg}.
There is a constant $C =C(U')>0$ such that for any $T_0\ge 0$, $T>0$ and $\theta\in(0,1)$, we have
\begin{multline}\label{qbarbound}
\sup_{[T_0+\theta T,T_0+T]}\|\bar p_t\|_{L^\infty(U')}\le C\Big\{\big[\lambda (1+(\theta T)^{-1/2} )\big]^\frac{1}{\mu_6} +\big[ \lambda T^{1/2} \sup_{[T_0,T_0+T]}\|\psi_t\|_{L^\infty} \big]^\frac1{\mu_6+1}\Big\}\\
\cdot \Big( \|\bar p_t\|_{L^2(U\times (T_0,T_0+T))} + \|\bar p_t\|_{L^2(U\times (T_0,T_0+T))}^\frac{\mu_6}{\mu_6+1}\Big),
\end{multline}
where
$\lambda= 1 + \sup_{[T_0,T_0+T]}\left(\int_U |\nabla p(x,t)|^{2-a}dx \right)^{\mu_7},$
and
\begin{equation}\label{pt:bound}
\sup_{[T_0+\theta T,T_0+T]}\|p_t\|_{L^\infty(U')}\le C\Big(1+\frac 1{\theta T} \Big)^\frac{1}{\mu_6} \Big( 1+\sup_{[T_0,T_0+T]}\norm{\nabla p}_{L^{2-a}(U)}^{\frac 2{{\mu_5}-2}}\Big) \norm{ p_t}_{L^2(U\times(T_0, T_0+T))}.
\end{equation}
\end{proposition}
\begin{proof} Without loss of generality, we assume $T_0=0$.
We prove \eqref{qbarbound} first.
Let $ q=p_t$ and
\begin{equation*}
\bar q=q-\frac1{|U|}\int_U qdx =p_t -\frac1{|U|} \frac d{dt}\int_U pdx=p_t + \frac1{|U|} \int_\Gamma \psi(x,t) d\sigma=\frac{\partial}{\partial t} \bar p=\bar p_t.
\end{equation*}
Then it follows from \eqref{eqorig} that $\bar q$ solves
\begin{equation}\label{eqt}
\frac{\partial \bar q}{\partial t}=\nabla \cdot \big(K(|\nabla p|)\nabla p\big)_t + \frac1{|U|} \int_\Gamma \psi_t d\sigma.
\end{equation}
For $k\ge 0$, let $\bar q^{(k)}=\max\{\bar q-k,0\}$
and $\chi_k(x,t)$ be the characteristic function of set $\{(x,t) \in U\times(0,T): \bar q(x,t)>k\}$.
On $S_k(t)$, we have $(\nabla p)_t=(\nabla\bar p)_t=\nabla\bar q=\nabla\bar q^{(k)}$.
Let $\zeta=\zeta(x,t)$ be the cut-off function on $U\times [0,T]$ satisfying $\zeta(\cdot,0)=0$ and $\zeta(\cdot,t)$ having compact support in $U$.
We will use test function $\bar q^{(k)}\zeta^2$, noting that $\nabla (\bar q^{(k)}\zeta^2) =\zeta [\nabla (\bar q^{(k)}\zeta)+\bar q^{(k)}\nabla \zeta]$.
Multiplying \eqref{eqt} by $\bar q^{(k)}\zeta^2 $ and integrating the resultant on $U$, we get
\begin{align*}
\frac 12\frac{d}{dt} \int_U |\bar q^{(k)}\zeta|^2 dx &
= \int_U |\bar q^{(k)}|^2 \zeta \zeta_t dx - \int_U (K(|\nabla p|))_t\nabla p \cdot[ \nabla (\bar q^{(k)}\zeta)+ \bar q^{(k)}\nabla \zeta]\zeta dx \\
& \quad
- \int_U K(|\nabla p|) (\nabla p)_t \cdot [ \nabla (\bar q^{(k)}\zeta)+ \bar q^{(k)}\zeta\nabla]\zeta dx +\frac1{|U|}\int_\Gamma \psi_t d\sigma \int_U \bar q^{(k)}\zeta^2 dx.
\end{align*}
Put $z=\zeta[\nabla (\bar q^{(k)}\zeta)+ \bar q^{(k)}\nabla \zeta]$.
We simplify the third term on the right-hand side of the last inequality as
\begin{multline*}
(\nabla p)_t \cdot z
=\zeta \nabla \bar q^{(k)}\cdot [\nabla (\bar q^{(k)}\zeta)+ \bar q^{(k)}\nabla \zeta]\\
=[\nabla (\bar q^{(k)}\zeta)- \bar q^{(k)}\nabla \zeta]\cdot [\nabla (\bar q^{(k)}\zeta)+ \bar q^{(k)}\nabla \zeta]
= |\nabla (\bar q^{(k)}\zeta)|^2 - |\bar q^{(k)} \nabla \zeta|^2 .
\end{multline*}
For the second term of right-hand side, using \eqref{K-est-2} we have
\begin{align*}
|(K(|\nabla p|))_t\nabla p \cdot z|\
&=|K'(|\nabla p|)| \frac{|\nabla p\cdot \nabla p_t|}{|\nabla p|} |\nabla p \cdot z|
\le a K(|\nabla p|) |\nabla \bar q||z|.
\end{align*}
Moreover,
\begin{align*}
|\nabla \bar q| |z|
&=|\zeta \nabla\bar q^{(k)}| | \nabla (\bar q^{(k)}\zeta)+ q^{(k)}\nabla \zeta |
\le (|\nabla (\bar q^{(k)}\zeta)| + |\bar q^{(k)}||\nabla \zeta|)^2 \\
&= |\nabla (\bar q^{(k)}\zeta)|^2 + 2|\bar q^{(k)}| |\nabla \zeta| |\nabla (\bar q^{(k)}\zeta)| + |\bar q^{(k)} \nabla \zeta|^2.
\end{align*}
It follows that
\begin{align*}
& \frac 12\frac{d}{dt} \int_U |\bar q^{(k)} \zeta|^2 dx +
(1-a) \int_U K(|\nabla p|) |\nabla (\bar q^{(k)} \zeta)|^2 dx \\
& \le \int_U |\bar q^{(k)}|^2 \zeta |\zeta_t| dx + (1+a) \int_U K(|\nabla p|)|\bar q^{(k)} \nabla \zeta|^2 dx\\
& \quad+2a\int_U K(|\nabla p|)| \bar q^{(k)} \nabla \zeta| |\nabla(\bar q^{(k)}\zeta )|dx
+C\|\psi_t(t)\|_{L^\infty} \int_U |\bar q^{(k)}|\zeta^2 dx.
\end{align*}
Let $\varepsilon>0$. By Cauchy's inequality,
\begin{align*}
2aK(|\nabla p|)| \bar q^{(k)} \nabla \zeta| |\nabla(\bar q^{(k)}\zeta )|
&\le \frac {1-a}{2} K(|\nabla p|)|\nabla(\bar q^{(k)}\zeta )|^2 + \frac{2a^2}{1-a} K(|\nabla p|)|\bar q^{(k)} \nabla \zeta|^2,\\
C \|\psi_t(t)\|_{L^\infty} |\bar q^{(k)}|\zeta^2
&\le \varepsilon |\bar q^{(k)}\zeta|^2 + C\varepsilon^{-1}\|\psi_t(t)\|_{L^\infty}^2 \zeta^2.
\end{align*}
Therefore,
\begin{align*}
& \frac 12\frac{d}{dt} \int_U |\bar q^{(k)} \zeta|^2 dx + \frac{1-a}2 \int_U K(|\nabla p|) |\nabla (\bar q^{(k)} \zeta)|^2 dx
\le \int_U |\bar q^{(k)}|^2 \zeta |\zeta_t| dx\\
&\quad + C \int_U K(|\nabla p|)|\bar q^{(k)} \nabla \zeta|^2 dx + \varepsilon \int_U |\bar q^{(k)}\zeta|^2 dx + C\varepsilon^{-1}\|\psi_t(t)\|_{L^\infty}^2 \int_U \chi_k \zeta^2 dx.
\end{align*}
Now, integrating this inequality in time from $0$ to $t$ and taking supremum on $[0,T]$, we obtain
\begin{multline*}
\sup_{[0,T]} \int_U |\bar q^{(k)}\zeta |^2 dx + \int_0^T \int_U K(|\nabla p|) |\nabla (\bar q^{(k)}\zeta) |^2 dx dt\\
\le C \Big [\int_0^T \int_U |\bar q^{(k)}|^2 \zeta |\zeta_t| dx dt
+ \int_0^T\int_U K(|\nabla p|) |\bar q^{(k)} \nabla \zeta|^2 dxdt \Big]\\
+ C_1\varepsilon T \sup_{[0,T]}\int_U |\bar q^{(k)}\zeta|^2 dx + C\varepsilon^{-1}\sup_{[0,T]}\|\psi_t(t)\|_{L^\infty}^2 \int_0^T \int_U \chi_k \zeta^2 dxdt.
\end{multline*}
Using the fact that the functions $K$ and $\zeta$ are bounded and taking $\varepsilon=1/(2C_1 T)$, we obtain
\begin{equation} \label{m-Lp-t}
\begin{split}
& \sup_{[0,T]} \int_U |\bar q^{(k)} \zeta|^2 dx
+ \int_0^T \int_U K(|\nabla p|)|\nabla (\bar q^{(k)} \zeta)|^2 dx dt\\
&\quad \le C\left [ \int_0^T \int_U |\bar q^{(k)}|^2 (\zeta |\zeta_t| +|\nabla \zeta|^2) dxdt \right]
+ C T \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty}^2 \int_0^T \int_U \chi_k \zeta^2 dxdt.
\end{split}
\end{equation}
Applying inequality \eqref{Wemb} in Lemma~\ref{ParaSob-3} to $\bar q^{(k)} \zeta$ with the weight $W(x,t)=K(|\nabla p(x,t)|)$, we have
\begin{align*}
& \norm{\bar q^{(k)}\zeta}_{L^{\mu_5}(Q_T)}
\le C \left \{ \sup_{[0,T]}\Big(\int_U K(|\nabla p|)^{-\frac {2-a} a} dx \Big)^{\frac 2 {s(2-a)} } \right \}\\
&\quad \cdot \left\{ \sup_{[0,T]} \Big(\int_U |\bar q^{(k)}\zeta|^2 dx \Big)^{1/2}+\Big(\int_0^T \int_U K(|\nabla p|)|\nabla (\bar q^{(k)}\zeta)|^2 dx dt\Big)^{1/2} \right\}\\
& \le C \sup_{[0,T]}\Big(\int_U (1+|\nabla p|)^{2-a} dx \Big)^{\mu_7}
\cdot \left\{\sup_{[0,T]} \int_U |\bar q^{(k)}\zeta |^2 dx + \int_0^T \int_U K(|\nabla p|)|\nabla (\bar q^{(k)}\zeta)|^2 dx dt \right\}^{1/2}.
\end{align*}
Therefore, by \eqref{m-Lp-t}
\begin{equation}\label{q:bound}
\norm{\bar q^{(k)}\zeta}_{L^{\mu_5}(Q_T)} \le C\lambda \Big(\int_0^T \int_U |\bar q^{(k)}|^2 (|\zeta_t|\zeta + |\nabla \zeta|^2)dx dt + T \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty}^2 \int_0^T \int_U \chi_k \zeta^2 dxdt\Big)^{1/2}.
\end{equation}
Let $x_0$ be any given point in $U$. Denote $\rho={\rm dist}(x_0,\partial U)>0$.
Let $M_0 >0$ be fixed which will be determined later.
For $i\ge 0$, define
\[
k_i= M_0(1-2^{-i}),\quad
t_i =\theta T( 1- 2^{-i}),\quad
\rho_i= \frac 1 4\rho(1+ 2^{-i}).
\]
Then
$t_0=0<t_1<\ldots<\theta T$ and $\rho_0=\rho/2>\rho_1>\ldots >\rho/4 >0.$
Note that
\begin{equation*}
\lim_{i\to\infty}t_i=\theta T\quad\text{and}\quad
\lim_{i\to\infty}\rho_i=\rho/4.
\end{equation*}
Let $U_i =\{x : \norm{x-x_0}< \rho_i\} $ then $ U_{i+1} \Subset U_i$ for $i=0,1,2,\ldots$.
For $i,j\ge 0$, we denote
\begin{equation}\label{defQnAnm}
\begin{aligned}
\mathcal Q_i &=\{(x,t): x\in U_i,\ t\in(t_i,T) \}, \quad
A_{i,j} &=\{(x,t)\in\mathcal Q_i: \bar q(x,t)>k_i,\ t\in(t_j, T) \}.
\end{aligned}
\end{equation}
For each $\mathcal Q_i$, we use a cut-off function $\zeta_i(x,t)$ which is piecewise linear in $t$ and satisfies $\zeta_i\equiv 1$ on $\mathcal Q_{i+1}$ and $\zeta_i\equiv 0$ on $Q_T\setminus\mathcal Q_{i}$.
Then there is $C>0$ such that
\begin{equation}\label{der:cutoff}
|(\zeta_i)_t|\le \frac C{t_{i+1}-t_i} = \frac {C2^{i+1}}{\theta T} \quad \text { and } \quad
|\nabla \zeta_i | \le \frac {C}{\rho_{i}-\rho_{i+1}} = \frac {C 2^{i+1}}{4\rho}\quad\text{for all } i\ge 0.\end{equation}
Define
$F_{i} = \norm{\bar q^{(k_{i+1})}\zeta_i }_{L^{\mu_5}(A_{i+1, i})}$.
Applying \eqref{q:bound} with $k=k_{i+1}$ and $\zeta=\zeta_i$ gives
\begin{equation}\label{qre2}
F_i \le C\lambda \Big\{ \int_0^T \int_U |\bar q^{(k_{i+1})}|^2 \Big(\zeta_i |(\zeta_i)_t| +|\nabla \zeta_i|^2\Big) dxdt +T \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty}^2 |A_{i+1,i}|\Big\}^{ 1/2}.
\end{equation}
Using derivative estimates \eqref{der:cutoff} for $\zeta$, we obtain
\begin{equation} \label{Fnbound}
\begin{aligned}
F_{i}
&\le C\lambda \Big( \big [2^i(\theta T)^{-1/2}+ 2^i\big] \|\bar q^{(k_{i+1})}\|_{L^2(A_{i+1,i})} + T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} |A_{i+1,i}|^{1/2}\Big)\\
&\le C\lambda 2^i(1+\frac 1 {\theta T})^{1/2} \|\bar q^{(k_i)} \|_{L^2(A_i)} + C\lambda T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} |A_{i+1,i}|^{1/2}.
\end{aligned}
\end{equation}
Then, it follows from H\"{o}lder's inequality, \eqref{qre2}
and \eqref{Fnbound} that
\begin{equation}\label{qki+1}
\begin{aligned}
\|\bar q^{(k_{i+1})} \|_{L^2(A_{i+1, i+1})}
& \le \|\bar q^{(k_{i+1})}\|_{L^{\mu_5}(A_{i+1,i+1})} |A_{i+1,i+1}|^{1/2-1/{\mu_5}}\\
& \le \|\bar q^{(k_{i+1})}\zeta_i \|_{L^{\mu_5}(A_{i+1,i+1})} |A_{i+1,i}|^{1/2-1/{\mu_5}}
\le C F_{i} |A_{i+1,i}|^{1/2-1/{\mu_5}}.
\end{aligned}
\end{equation}
Note that
$ \|\bar q^{(k_i)}\|_{L^2(A_i)}\ge \|\bar q^{(k_{i})}\|_{L^2(A_{i+1,i})}\ge (k_{i+1}-k_i) |A_{i+1,i}|^{1/2}$.
Thus,
\begin{equation}\label{Aii}
|A_{i+1,i}| \le (k_{i+1}-k_i)^{-2} \|\bar q^{(k_i)}\|_{L^2(A_i)}^{2} \le C 4^{i} M_0^{-2}\|\bar q^{(k_i)}\|_{L^2(A_i)}^{2}.
\end{equation}
Then it follows \eqref{qki+1}, \eqref{Fnbound} and \eqref{Aii} that
\begin{align*}
\|\bar q^{(k_{i+1})}\|_{L^2(A_{i+1,i+1})}&\le C\lambda
\Big\{2^i (1+\frac 1{\theta T} )^{1/2}\|\bar q^{(k_i)}\|_{L^2(A_i)}+ T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty}2^{i}M_0^{-1} \| q^{(k_i)} \|_{L^2(A_i)}\Big\}\\
&\quad \cdot 2^{i-\frac {2i} {\mu_5} }M_0^{-1+2/{\mu_5}} \| \bar q^{(k_i)} \|_{L^2(A_i)}^{1-2/{\mu_5}}\\
&\le C 4^i \lambda \Big\{ (1+\frac 1{\theta T} ) M_0^{-1+2/{\mu_5}} + T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} M_0^{-2+2/{\mu_5}}\Big\}\norm{\bar q^{(k_i)}}_{L^2(A_i)}^{2-2/{\mu_5}}.
\end{align*}
Let $Y_i=\| \bar q^{(k_i)} \|_{L^2(A_i)}$, $B =4$ and
\begin{align*}
D_1&=C\lambda (1+\frac 1{\theta T} )^{1/2} M_0^{-1+2/{\mu_5}}=C\lambda(1+\frac 1{\theta T} )^{1/2}M_0^{-\mu_6},\\
D_2&=C\lambda T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} M_0^{-2+2/{\mu_5}}=C\lambda T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} M_0^{-1-\mu_6}.
\end{align*}
We obtain
$Y_{i+1}\le B^i(D_1 Y_i^{1+\mu_6}+D_1 Y_i^{1+\mu_6})$ for all $i\ge 0$.
We now determine $M_0$ so that
\begin{equation}\label{Y0qcond}
Y_0 \leq (2D_1)^{-1/\mu_6}B^{-1/{\mu_6^2}}, \quad Y_0 \leq (2D_2)^{-1/\mu_6}B^{-1/{\mu_6^2}}.
\end{equation}
This condition is met if
\[ M_0\ge C \big[ \lambda (1+(\theta T)^{-1/2} ) \big]^{1/\mu_6} Y_0,
\quad M_0\ge C \big[ \lambda T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} \big]^\frac1{\mu_6+1} Y_0^\frac{\mu_6}{\mu_6+1}.\]
Since $Y_0\le \| \bar q\|_{L^2(U\times (0,T))}$, it suffices to choose $M_0$ as
\begin{equation}\label{Mzero}
M_0 = C\Big\{\big[\lambda (1+(\theta T)^{-1/2} )\big]^{1/\mu_6} +\big[ \lambda T^{1/2} \sup_{[0,T]}\|\psi_t(t)\|_{L^\infty} \big]^\frac1{\mu_6+1}\Big\}
\Big( \|\bar q\|_{L^2(U\times (0,T))}^\frac{\mu_6}{\mu_6+1}+\|\bar q\|_{L^2(U\times (0,T))}\Big).
\end{equation}
Then by condition \eqref{Y0qcond}, applying \eqref{Y0mcond} in Lemma \ref{multiseq} for $m=2$ gives
$\displaystyle{\lim_{i\to\infty}}Y_i=0$. (Alternatively, Lemma \ref{oriseq} can be used in this case.)
Hence,
\begin{equation}\label{intq}
\int_{\theta T}^T\int_{B(x_0,\rho/4)} |\bar q^{(M_0)}|^2 dxdt=0.
\end{equation}
Since $\bar q(x,t)\in C(U\times (0,\infty)$, it follows from \eqref{intq} that $\bar q(x,t)\le M_0$ in $B(x_0,\rho/4)\times (\theta T,T)$.
Replace $q$ by $-q$ and $\psi$ by $-\psi$ and
use the same argument we obtain $|\bar q(x,t)|\le M_0$ in $B(x_0,\rho/4)\times (\theta T,T)$.
Now by covering $U'$ by finitely many such balls $B(x_0,\rho/4)$, we come to conclusion
\begin{equation}\label{ptM0}
|\bar q(x,t)|\le M_0 \quad \text{in } U'\times (\theta T,T).
\end{equation}
By the choice of $M_0$, we obtain \eqref{qbarbound} from \eqref{ptM0}.
In the above proof of \eqref{qbarbound}, we can work with $q$ instead of $\bar q$, with
\begin{equation}\label{ebarqt}
\frac{\partial q}{\partial t}=\nabla \cdot \big(K(|\nabla p|)\nabla p\big)_t,
\end{equation}
instead of equation \eqref{eqt}, then the term $\|\psi_t\|_{L^\infty}$ can be removed and we obtain the desired estimate \eqref{pt:bound} for $p_t$.
\end{proof}
\begin{remark}
The main difference between estimate of $\bar p_t$ in \eqref{qbarbound} and the estimate of $p_t$ in \eqref{pt:bound} is the involvement of $\psi_t$.
Though we cannot derive \eqref{pt:bound} from \eqref{qbarbound}, in the following development we will focus on $\bar p_t$ only.
\end{remark}
The following estimates can be easily derived (by the mean of Young's inequality) from corresponding ones in \cite{HI2}. They will be used in finding $L^\infty$-estimates for $\bar p_t$ in terms of initial data and boundary data.
We use the following notation:
\begin{align}
&m_1(t)=1+ \norm{\bar p_0}_{L^2}^2 + M_f(t)^\frac2{2-a}+\int_{t-1}^t \tilde f(\tau)d\tau,
&&m_2(t)=1 + A^\frac2{2-a}+\int_{t-1}^t \tilde f(\tau)d\tau,\\
\label{Atildef}
&m_3(t)=1+\beta^\frac1{1-a}
+\sup_{[t-1,t]} f^\frac{2}{2-a} +\int_{t-1}^t \tilde f(\tau)d\tau,
&&\mathcal A_1=A+A^\frac2{2-a}+\limsup_{t\to\infty}\int_{t-1}^t \tilde f(\tau)d\tau.
\end{align}
\begin{theorem}[cf. \cite{HI2} Theorems 4.4 and 4.5]\label{H-bound-lemma}
{\rm (i)} One has for all $t\ge 0$ that
\begin{equation}
\label{H-bound-0}
J_H[p](t)+\norm{\bar p_t}_{L^2(U\times(0,t))}^2
\le C\Big( \norm{\bar p_0}_{L^2}^2 +J_H[p](0)+ (t+1) \sup_{[0,t]} f +\int_{0}^t\tilde f(\tau)d\tau\Big),
\end{equation}
and for all $t\ge 1$ that
\begin{equation} \label{H-bound-6}
J_H[p](t)+\norm{\bar p_t}_{L^2(U\times(t-1/2,t))}^2
\le C\Big( \norm{\bar p(t-1)}_{L^2}^2 + \sup_{[t-1,t]}f+ \int_{t-1}^t \tilde f(\tau)d\tau\Big).
\end{equation}
Now, assume that the Degree Condition holds.
{\rm (ii)} For $0<t_0< 1$ and $t\ge t_0$, one has
\begin{equation}\label{Jpt-boundA0}
\norm{\bar p_t(t)}_{L^2}^2 \le C\, t_0^{-1} L_5(t_0)+ C \int_0^t f(\tau)+ \tilde f(\tau)d\tau,
\end{equation}
where $L_5(t_0)=L_5(t_0;[p_0,\psi])\mathbin{\buildrel \rm def \over \longequals} 1+\norm{\bar p_0}_{L^2}^2 +\|\nabla p_0\|_{L^{2-a}}^{2-a} + M_f(t_0)^\frac2{2-a}+\int_0^{t_0}\tilde f(\tau)d\tau$.
For all $t\ge 1$, one has
\begin{equation}\label{H-bound-b}
J_H[p](t),\ \norm{\bar p_t(t)}_{L^2}^2 \le C m_1(t) \quad \text{for all }t\ge 1.
\end{equation}
{\rm (iii)} If $A<\infty$ then there is $T>1$ such that
\begin{equation}\label{boundedlarge}
J_H[p](t),\ \norm{\bar p_t(t)}_{L^2}^2 \le C m_2(t) \quad \text{for all } t>T,
\end{equation}
\begin{equation}\label{limsupH1}
\limsup_{t\to\infty} J_H[p](t),\ \limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^2}^2 \le C \mathcal A_1.\end{equation}
{\rm (iv)} If $\beta<\infty$ then there is $T>1$ such that
\begin{equation}\label{unbounJ}
J_H[p](t),\ \norm{\bar p_t(t)}_{L^2}^2 \le C m_3(t) \quad \text{for all } t>T.\end{equation}
\end{theorem}
We now state our main estimates of $\bar p_t$.
In addition to $\mathcal K_{1,t}$ defined by \eqref{Nf1}, we will also use
\begin{equation}\label{Nf2}
\mathcal K_{2,t}=1+\int_{0}^t \tilde f(\tau)d\tau.
\end{equation}
\begin{theorem}\label{ptbar} Assume \eqref{strictdeg}.
For $t>0$, one has
\begin{equation}\label{tpos}
\begin{aligned}
\norm{\bar p_t(t)}_{L^\infty(U')}
&\le C \Big\{ \norm{\bar p_0}_{L^2}^2+\|\nabla p_0\|_{L^{2-a}}^{2-a}+(1+t)\mathcal K_{1,t}^\frac{2-a}{1-a} +\mathcal K_{2,t}\Big\}^{\frac{\mu_7}{\mu_6}+\frac 1 2} \\
&\quad \cdot\Big\{1+ t^{-\frac 1{2\mu_6}} + t^{\frac 1 {2(\mu_6+1)}} \sup_{[0,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big\}.
\end{aligned}
\end{equation}
For $t\ge 3/2$, one has
\begin{equation}\label{tlarge}
\begin{aligned}
\norm{\bar p_t(t)}_{L^\infty(U')}
\le C\Big( 1+\norm{\bar p_0}_{L^2}^2 +M_f(t)^{\frac 2{2-a} }+ \int_{t-3/2}^t \tilde f(\tau)d\tau \Big)^{\frac{\mu_7}{\mu_6} +\frac 1 2} \Big(1 +\sup_{[t-1/2,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big).
\end{aligned}
\end{equation}
Above, the positive constant $C$ depends on $U'$.
\end{theorem}
\begin{proof}
Applying \eqref{qbarbound} with $T_0=0$, $T=t$ and $\theta=1/ 2$, an then applying Young's inequality to the term
$\|\bar p_t\|_{L^2(U\times (0,t))}^{\mu_6/(\mu_6+1)}$, we have
\begin{align*}
\norm{\bar p_t(t)}_{L^\infty(U')}
&\le C (1+\sup_{[0,t]}\|\nabla p\|_{L^{2-a}}^{2-a})^\frac{\mu_7}{\mu_6} \\
&\quad \cdot \Big(\Big[1+\sqrt{2/t\,} \Big]^\frac{1}{\mu_6} +\big[ t^{1/2} \sup_{[0,t]}\|\psi_t\|_{L^\infty} \big]^\frac1{\mu_6+1}\Big) \Big( 1+\|\bar p_t\|_{L^2(U\times (0,t))}\Big).
\end{align*}
Using \eqref{H-bound-0} to estimate $\sup_{[0,t]}\|\nabla p\|_{L^{2-a}}^{2-a}$ and $\norm{\bar p_t}_{L^2(U\times(0,t))}$ in the previous inequality, we find that
\begin{align*}
&\norm{\bar p_t(t)}_{L^\infty(U')}
\le C \Big\{ 1+\norm{\bar p_0}_{L^2}^2 +\|\nabla p_0\|_{L^{2-a}}^{2-a} +(1+t)\mathcal K_{1,t}^\frac{2-a}{1-a} +\mathcal K_{2,t}\Big\}^{ \frac{\mu_7}{\mu_6}} \\
&\quad \cdot\Big\{1+t^{-\frac 1{2\mu_6}} + t^{\frac 1 {2(\mu_6+1)}} \sup_{[0,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big\}
\Big(1+\norm{\bar p_0}_{L^2}^2 +\|\nabla p_0\|_{L^{2-a}}^{2-a} + (1+t)\mathcal K_{1,t}^\frac{2-a}{1-a} +\mathcal K_{2,t}\Big)^\frac12 .
\end{align*}
Then inequality \eqref{tpos} follows.
Now, consider $t\ge 3/2$.
Applying estimate \eqref{qbarbound} to the interval $[t-1/2,t]$, i.e., $T_0=t-1/2$ and $T=1/2$, with $\theta=1/2$, we obtain
\begin{align*}
\norm{\bar p_t(t)}_{L^\infty(U')}
&\le C\Big( 1+\sup_{[t-1/2,t]}\norm{\nabla p}_{L^{2-a}(U)}^{2-a}\Big)^{\frac{\mu_7}{\mu_6}} \Big( 1 +\sup_{[t-1/2,t]}\|\psi_t(t)\|_{L^\infty} ^\frac1{\mu_6+1}\Big)\\
&\quad\cdot\Big( \|\bar p_t\|_{L^2(U\times (t-1/2,t))}^\frac{\mu_6}{\mu_6+1}+\|\bar p_t\|_{L^2(U\times (t-1/2,t))}\Big).
\end{align*}
Let $m(t)=\sup_{[t-3/2,t]}f+ \int_{t-3/2}^t \tilde f(\tau)d\tau$.
Utilizing estimate \eqref{H-bound-6}, we have
\begin{multline}\label{t32}
\norm{\bar p_t(t)}_{L^\infty(U')}
\le C\Big( 1+\sup_{\tau\in[t-3/2,t-1]} \norm{\bar p(\tau)}_{L^2}^2 + m(t)\Big)^{\frac{\mu_7}{\mu_6}}
\Big(1 +\sup_{[t-1/2,t]}\|\psi_t(t)\|_{L^\infty}^\frac1{\mu_6+1}\Big)\\
\cdot \Big\{ \Big( \norm{\bar p(t-1)}_{L^2}^2 + m(t) \Big)^\frac{\mu_6}{\mu_6+1}+\Big( \norm{\bar p(t-1)}_{L^2}^2 + m(t) \Big) \Big\}^{\frac12}.
\end{multline}
By Young's inequality,
\begin{align*}
|\bar p_t(x,t)|&\le C\Big\{ 1+\sup_{\tau\in[t-3/2,t-1]} \norm{\bar p(\tau)}_{L^2}^2 + m(t) \Big\}^{\frac{\mu_7}{\mu_6} +\frac 1 2}
\Big\{1 +\sup_{[t-1/2,t]}\|\psi_t(t)\|_{L^\infty}^\frac1{\mu_6+1}\Big\}.
\end{align*}
Then inequality \eqref{tlarge} is obtained by using \eqref{p-bar-bound} to estimate $\| \bar p(\tau)\|_{L^2}$ for $\tau\le t-1$.
\end{proof}
Let
\begin{equation}\label{Ulim2}
\mathcal A_2=\limsup_{t\to\infty} m_2(t)=1+A^\frac{2}{2-a}+\limsup_{t\to\infty}\int_{t-1}^t\tilde f(\tau)d\tau,
\end{equation}
\begin{equation}\label{A3def}
\mathcal A_3=1+A^{\frac2{2-a}}+ \limsup_{t\to\infty} \|\psi_t(t)\|^\frac{2-a}{1-a}.
\end{equation}
We have from \eqref{Atildef}, \eqref{Ulim2} and \eqref{A3def} the relation:
\begin{equation}\label{Urel}
\mathcal A_1\le C\mathcal A_2\le C'\mathcal A_3.
\end{equation}
\begin{theorem}\label{ptbar2}
Assume \eqref{strictdeg}.
{\rm (i)} If $A<\infty$ then
\begin{equation}\label{limsup-pt}
\begin{aligned}
\limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^\infty(U')}
\le C\big(\mathcal A_1 +\mathcal A_1^\frac{\mu_6}{\mu_6+1}\big)^{1/2}
\mathcal A_2^{\frac{\mu_7}{\mu_6}} \Big(1 +\limsup_{t\to\infty}\|\psi_t(t)\|_{L^\infty}^\frac1{\mu_6+1}\Big).
\end{aligned}
\end{equation}
Consequently, denoting $\mu_8=\frac{\mu_7}{\mu_6}+\frac 12+\frac{1-a}{(2-a)(\mu_6+1)}$, one has
\begin{equation}\label{limsup-pt10}
\limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^\infty(U')}\le C \mathcal A_3^{\mu_8}.
\end{equation}
{\rm (ii)} If $\beta<\infty$, then there is $T>0$ such that for all $t>T$,
\begin{equation}\label{ptbeta}
\norm{\bar p_t(t)}_{L^\infty(U')}
\le C\Big(1+\beta^\frac1{1-a}+\sup_{\tau\in[t-3/2,t]} f(\tau)^\frac2{2-a}+\int_{t-3/2}^t \tilde f(\tau)d\tau\Big)^{\frac {\mu_7}{\mu_6} +\frac 1 2}\Big(1 +\sup_{[t-1/2,t]}\|\psi_t(t)\|_{L^\infty}^\frac1{\mu_6+1}\Big) .
\end{equation}
Above, the positive constant $C$ depends on $U'$.
\end{theorem}
\begin{proof}
(i) We follow the proof in Theorem \ref{ptbar}.
Observe that
\begin{equation}\label{limint}
\limsup_{t\to\infty} \int_{t-3/2}^t \tilde f(\tau)d\tau
\le \limsup_{t\to\infty} \int_{t-2}^{t-1} \tilde f(\tau)d\tau + \limsup_{t\to\infty} \int_{t-1}^t \tilde f(\tau)d\tau
\le 2 \limsup_{t\to\infty} \int_{t-1}^t \tilde f(\tau)d\tau.
\end{equation}
The limit estimates \eqref{nonzerobeta} and \eqref{limint} yield
\begin{equation*}
\limsup_{t\to\infty} \Big(\sup_{[t-3/2,t-1]} \|\bar p\|_{L^2}+m(t)\Big)\le C \mathcal A_1.
\end{equation*}
Combining this with \eqref{t32}, we obtain
\begin{equation}
\limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^\infty(U')}\le C\big( 1+\mathcal A_1\big)^\frac{\mu_7}{\mu_6} \Big(1 +\limsup_{t\to\infty}\big[\sup_{[t-1/2,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\big]\Big) \big(\mathcal A_1^\frac{\mu_6}{\mu_6+1} +\mathcal A_1 \big)^{\frac12}.
\end{equation*}
Then \eqref{limsup-pt} follows this and relation \eqref{Urel}.
Elementary calculations will give \eqref{limsup-pt10} from \eqref{limsup-pt}.
(ii) We use \eqref{t32} and Young's inequality again, this time, with estimate \eqref{unboundp}; it results in \eqref{ptbeta}.
\end{proof}
While \eqref{limsup-pt} gives an asymptotic estimate for $\bar p_t$ in the case $A<\infty$, the next result covers the case $A=\infty$.
\begin{theorem}\label{smallqbar}
Assume \eqref{strictdeg}.
Let $L_6=\norm{\bar p_0}_{L^2}^2 $ in general case, and $L_6=\beta^\frac1{1-a}$ in case $\beta<\infty$, and define
\begin{equation}\label{hdef}
N(t)=1+ L_6 +M_f(t)^{\frac 2 {2-a}}+\int_{t-2}^t \tilde f(\tau) d\tau,\quad
h(t)= N(t)^{\frac{\mu_7}{\mu_6}} \Big(1 +\sup_{[t-1,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big).
\end{equation}
If
\begin{equation}\label{hNcond}
\lim_{t\to\infty} h(t)^\frac{2(\mu_6+1)}{\mu_6} e^{-d_1\int_2^t N(\tau)^{-b}d\tau} =0
\quad \text{and} \quad
\lim_{t\to \infty} \frac {h'(t)}{h(t)}N^b(t) =0 ,
\end{equation}
where $b$ is defined in \eqref{ab} and $d_1>0$ appears in \eqref{recallpt} below,
then
\begin{equation}\label{limqbar}
\limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^\infty(U')}
\le C\limsup_{t\to\infty}
\Big\{ h(t) \Big( N(t)^{b}\norm {\psi_t(t)}_{L^\infty} +\Big[ N(t)^{b} \norm { \psi_t(t)}_{L^\infty} \Big]^\frac{\mu_6}{\mu_6+1}\Big)\Big\},
\end{equation}
where $C>0$ depends on $U'$.
\end{theorem}
\begin{proof}
Applying \eqref{qbarbound} to the interval $[t-1,t]$ with $\theta=\frac 1 2$, we have
\begin{multline}\label{ptinterval}
\norm{\bar p_t(t)}_{L^\infty(U')} \le C \Big[ \Big( 1+\sup_{[t-1,t]} J_H[p](\tau)\Big)^\frac{\mu_7}{\mu_6}
\Big(1+\sup_{[0,T]} \|\psi_t(t)\|_{L^\infty}^\frac1{\mu_6+1}\Big)\Big]\\
\cdot \Big( \sup_{[t-1,t]} \|\bar p_t\|_{L^2(U)}^\frac{\mu_6}{\mu_6+1}+\sup_{[t-1,t]} \|\bar p_t\|_{L^2(U)}\Big).
\end{multline}
In the following $T>2$ is sufficiently large. By \eqref{H-bound-b} in the general case, and by \eqref{unbounJ} in case $\beta<\infty$, we have
\begin{equation*}
J_H[p](\tau) \le C\Big(1+ L_6 +M_f(t)^{\frac 2 {2-a}}+\int_{t-1}^t \tilde f(\tau) d\tau\Big), \quad \tau>T.
\end{equation*}
Therefore,
\begin{equation}\label{nablaP1b}
\sup_{[t-1,t]} J_H[p](\tau) \le C N(t).
\end{equation}
Recall inequality (5.13) in \cite{HI2}: there is $d_1>0$ such that for $t>$ we have
\begin{equation}\label{recallpt}
\frac {d} {dt} \norm{\bar p_t(t)}^2 \le -d_1N(t)^{-b} \norm{\bar p_t(t)}^2 +C \norm { \psi_t(t)}^2_{L^\infty} N(t)^b.
\end{equation}
Therefore, for $t'\in (T,\infty)$
\begin{equation}
\norm{\bar p_t(t')}^2 \le e^{-d_1\int_T^{t'} N(\tau)^{-b}d\tau} \norm{\bar p_t(T)}^2+ \int_T^{t'} e^{-d_1 \int_\tau^{t'} N(s)^{-b}ds} \norm { \psi_t(\tau)}^2_{L^\infty} N(\tau)^b d\tau.
\end{equation*}
Then taking supremum in $t'$ over the interval $[t-1,t]$ yields
\begin{align*}
\sup_{[t-1,t]} \norm{\bar p_t}^2
\le e^{-d_1\int_T^{t-1} N^{-b}(\tau)d\tau} \norm{\bar p_t(T)}^2 + e^{-d_1\int_T^{t-1} N^{-b}(\tau)d\tau} \int_T^t e^{d_1 \int_T^\tau N^{-b}(\theta)d\theta} \norm{\psi_t(\tau)}_{L^\infty}^2 N(\tau)^{b} d\tau .
\end{align*}
Thanks to the fact $N(t)\ge 1$, we have
\begin{equation*}
e^{-d_1\int_T^{t-1} N(\tau)^{-b}d\tau}= e^{-d_1\int_T^{t} N(\tau)^{-b}d\tau} e^{d_1\int_{t-1}^t N(\tau)^{-b}d\tau} \le e^{d_1} e^{-d_1\int_T^{t} N(\tau)^{-b}d\tau}.
\end{equation*}
Hence,
\begin{multline}\label{supptint}
\sup_{[t-1,t]} \norm{\bar p_t}^2\\
\le C \Big(e^{-d_1\int_T^{t} N^{-b}(\tau)d\tau} \norm{\bar p_t(T)}^2 + e^{-d_1 \int_T^{t} N^{-b}(\tau)d\tau} \int_T^t e^{d_1 \int_T^\tau N^{-b}(\theta)d\theta} \norm{\psi_t(\tau)}_{L^\infty}^2 N(\tau)^b d\tau\Big)\\
=C\Big(e^{-d_1\int_T^{t} N^{-b}(\tau)d\tau} \norm{\bar p_t(T)}^2 + \int_T^t e^{-d_1\int_\tau^t N^{-b}(\theta)d\theta} \norm{\psi_t(\tau)}_{L^\infty}^2 N(\tau)^b d\tau \Big).
\end{multline}
By \eqref{ptinterval}, \eqref{nablaP1b} and \eqref{supptint}, we obtain
\begin{equation*}
\begin{aligned}
\norm{\bar p_t(t)}_{L^\infty(U')}
& \le C h(t)
\Big\{ \Big(e^{-d_1\int_T^{t} N(\tau)^{-b}d\tau} \norm{\bar p_t(T)}^2+ \int_T^t e^{-d_1 \int_\tau^{t} N(s)^{-b} ds} \norm { \psi_t(\tau)}^2_{L^\infty} N(\tau)^b d\tau \Big)^{\frac{\mu_6}{2(\mu_6+1)}}\\
&\quad+\Big( e^{-d_1\int_T^{t} N(\tau)^{-b}d\tau} \norm{\bar p_t(T)}^2+ \int_T^t e^{-d_1 \int_\tau^{t} N(s)^{-b} ds} \norm { \psi_t(\tau)}^2_{L^\infty} N(\tau)^b d\tau \Big)^\frac12\Big\}.
\end{aligned}
\end{equation*}
Thus,
\begin{align*}
&\norm{\bar p_t(t)}_{L^\infty(U')}
\le C\Big( h(t)^{\frac{2(\mu_6+1)}{\mu_6} }e^{-d_1 \int_T^t N(\tau)^{-b}d\tau} \norm{\bar p_t(T)}^2\\
&\quad + h(t)^{\frac{2(\mu_6+1)}{\mu_6} }\int_T^t e^{-d_1 \int_\tau^t N(s)^{-b} ds} \norm { \psi_t(\tau)}^2_{L^\infty} N(\tau)^b d\tau\Big)^{\frac{\mu_6}{2(\mu_6+1)}} \\
&\quad+C\Big(h(t)^2 e^{- d_1 \int_T^t N(\tau)^{-b}d\tau} \norm{\bar p_t(T)}^2+ h^2(t)\int_T^t e^{-d_1 \int_\tau^t N(s)^{-b} ds} \norm { \psi_t(\tau)}^2_{L^\infty} N(\tau)^b d\tau\Big)^{\frac 12} .
\end{align*}
Under condition \eqref{hNcond}, applying Lemma~\ref{difflem2} we obtain
\begin{equation*}
\begin{aligned}
\limsup_{t\to\infty} \norm{\bar p_t(t)}_{L^\infty } &\le C\Big\{\limsup_{t\to\infty}\Big[ \norm { \psi_t(\tau)}^2_{L^\infty} h(t)^{\frac{2(\mu_6+1)}{\mu_6} } N(t)^{2b}\Big]\Big\}^{\frac{\mu_6} {2(\mu_6+1)} }\\
&\quad+C\Big\{\limsup_{t\to\infty} \Big[\norm { \psi_t(\tau)}^2_{L^\infty} h(t)^2 N(t)^{2b}\Big]\Big\}^{\frac 1 2}.
\end{aligned}
\end{equation*}
Therefore, we obtain \eqref{limqbar}.
\end{proof}
Note from \eqref{limsup-pt} that $\lim_{t\to\infty} \|\bar p_t(t)\|_{L^\infty(U')}= 0$ provided
$$\lim_{t\to\infty}\|\psi(t)\|_{L^\infty}= 0\quad \text{and}\quad \lim_{t\to\infty}\|\psi_t(t)\|_{L^\infty}= 0.$$
In the following, we can drop the first limit condition.
\begin{corollary}\label{ZeroCor}
Under the Strict Degree Condition, if $\|\psi(t)\|_{L^\infty}$ is uniformly bounded on $[0,\infty)$ and $\|\psi_t(t)\|_{L^\infty}\to 0$ as $t\to\infty$, then
\begin{equation}\label{zerolim}
\|\bar p_t(t)\|_{L^\infty(U')}\to 0 \quad \text{as } t\to\infty.
\end{equation}
\end{corollary}
\begin{proof}
In this case, $N(t)$ and $h(t)$ are uniformly bounded on $[2,\infty)$ by a constant $C_0$. In the Theorem \ref{smallqbar}, we can replace $N(t)$ and $h(t)$ by this $C_0$. Then conditions in \eqref{hNcond} are met and \eqref{zerolim} follows \eqref{limqbar}.
\end{proof}
\subsection{Estimates for pressure's second derivatives}
\label{seconderiv}
We estimate the Hessian $\nabla^2 p=\big(\frac{\partial^2 p}{\partial x_i \partial x_j}\big)_{i,j=1,2,\ldots,n}$.
Throughout this subsection $U'\Subset U$.
\begin{proposition}\label{hessprop}
Let $U'\Subset V \Subset U$ and $\delta\in(0,a]$.
Then for $t>0$,
\begin{equation}\label{delunite}
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx \le C \Big(1+ \int_{U} |\nabla p(x,t)|^{(2-\delta)a/\delta} dx\Big) \Big(1+ \|\bar p_t(t)\|_{L^\infty( V )}^2\Big),
\end{equation}
where the positive constant $C$ depends on $U'$, $V$ and $\delta$.
\end{proposition}
\begin{proof}
Young's inequality gives
\begin{multline}\label{del3}
\int_{U'} |\nabla^2 p|^{2-\delta} dx \le \int_{U'} K(|\nabla p|) |\nabla^2 p|^2 dx + \int_{U'} K(|\nabla p|)^{-(2-\delta)/\delta} dx\\
\le \int_{U'} K(|\nabla p|) |\nabla^2 p|^2 dx + C\int_{U'} \big[1+|\nabla p|^{(2-\delta)a/\delta}\big] dx.
\end{multline}
Note that for $s\ge 0$, we have
\begin{equation}\label{silly}
\frac{1}{2s+2} \frac{d}{dt} \int_U |\nabla\bar p|^{2s+2} \zeta^2 dx= - \int_U \bar p_t \nabla \cdot (|\nabla\bar p|^{2s}\nabla\bar p \zeta^2) dx.
\end{equation}
Since $\nabla \bar p =\nabla p$, replacing $p$ by $\bar p$ in \eqref{gradid} and using \eqref{silly} with $s=0$, we have for any $\zeta(x)\in C_c^\infty(U)$ that
\begin{align*}
\int_U K(|\nabla p|) |\nabla^2 p|^2 \zeta^2 dx\le C \int_U K(|\nabla p|) |\nabla p|^{2}|\nabla \zeta|^2 dx
+ C \int_U \bar p_t \nabla \cdot (\nabla p \zeta^2) dx.
\end{align*}
Let $\varepsilon>0$. For the last integral,
\begin{align*}
& \Big| \int_U \bar p_t \nabla \cdot (\nabla p \zeta^2) dx \Big|
\le C \int_U |\bar p_t| |\nabla ^2 p| \zeta^2 + |\bar p_t||\nabla p| \zeta |\nabla \zeta| dx \\
&\le \varepsilon \int_U K(|\nabla p|) |\nabla^2 p|^2 \zeta^2 dx
+ C\varepsilon^{-1}\int_U |\bar p_t|^2 K(|\nabla p|)^{-1} \zeta^2 dx
+C \int_U |\bar p_t||\nabla p| \zeta |\nabla \zeta| dx.
\end{align*}
Therefore,
\begin{equation}\label{gradineq1}
\begin{aligned}
& (1-\varepsilon) \int_U K(|\nabla p|) |\nabla^2 p|^2 \zeta^2 dx
\le C \int_U K(|\nabla p|) |\nabla p|^{2}|\nabla \zeta|^2 dx\\
&\quad\quad\quad\quad+C\varepsilon^{-1}\int_U |\bar p_t|^2 (1+|\nabla p|)^a \zeta^2 dx
+C \int_U |\bar p_t||\nabla p| \zeta |\nabla \zeta| dx.
\end{aligned}
\end{equation}
Constructing appropriate $\zeta$ with $\zeta\equiv 1$ on $U'$ and supp $\zeta \subset V$, we obtain from \eqref{gradineq1} that
\begin{align*}
& \int_{U'} K(|\nabla p|) |\nabla^2 p|^2 dx
\le C\int_{ V }|\nabla p|^{2-a}dx+ C \|\bar p_t(t)\|_{L^\infty( V )}^2 \int_{ V } (1+|\nabla p|)^a dx
\\
&+ C \|\bar p_t(t)\|_{L^\infty( V )} \int_{ V } |\nabla p| dx \le C\int_{ V }|\nabla p|^{2-a}dx+ C \Big(1+ \int_{ V } |\nabla p| dx\Big) (1+ \|\bar p_t\|_{L^\infty(V)}^2).
\end{align*}
Thus, by Young's inequality, we have
\begin{equation}\label{gradinqe2}
\int_{U'} K(|\nabla p|) |\nabla^2 p|^2 dx
\le C \Big(1+ \int_{ V } |\nabla p|^{2-a} dx\Big) (1+ \|\bar p_t\|_{L^\infty(V)}^2).
\end{equation}
Combining \eqref{del3} with \eqref{gradinqe2}, we obtain
\begin{multline*
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx \le C \Big(1+ \int_{U} |\nabla p(x,t)|^{2-a} dx\Big) \Big(1+ \|\bar p_t(t)\|_{L^\infty( V )}^2\Big)\\
+ C\int_{U'} |\nabla p(x,t)|^{(2-\delta)a/\delta} dx.
\end{multline*}
Then \eqref{delunite} follows.
\end{proof}
We consider two cases $\delta=a$ and $\delta<a$ separately.
We define two exponents
\begin{equation}
\mu_9=\frac{2\mu_7}{\mu_6}+2+\frac 1 {\mu_6+1}\quad\text{and}\quad \mu_{10}=\frac{2-a}{1-a}\big(\frac{2\mu_7}{\mu_6}+2\big) + \frac2{\mu_6+1}.
\end{equation}
\begin{theorem}\label{hessest}
Assume \eqref{strictdeg}.
{\rm (i)} If $t>0$ then
\begin{multline}\label{hess1b}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx
\le C L_7 (1+ t^{-\frac 1{\mu_6}})
(t+1)^{\mu_9}
\mathcal K_{1,t}^{\frac{2-a}{1-a}\big(\frac{2\mu_7}{\mu_6}+2\big)} \Big\{ 1 + \sup_{[0,t]}\|\psi_t\|_{L^\infty}\Big\}^{\mu_{10}},
\end{multline}
where $L_7=( 1+\norm{\bar p_0}_{L^2}^2 +\|\nabla p_0\|_{L^{2-a}}^{2-a} )^{\frac{2\mu_7}{\mu_6}+ 2}$ and $\mathcal K_{1,t}$ is defined in \eqref{Nf1}.
If $t\ge 3/2$ then
\begin{equation}\label{h2bcomp}
\begin{aligned}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx
\le C L_8 \{ 1 +M_f(t)\}^{\frac2{2-a}\big(\frac{2\mu_7}{\mu_6}+ 2\big)}
\cdot\Big\{1+\sup_{[t-3/2, t]} \norm{\psi_t}_{L^\infty}\Big\}^{\mu_{10}},
\end{aligned}
\end{equation}
where $L_8=( 1+\norm{\bar p_0}_{L^2}^2 )^{\frac{2\mu_7}{\mu_6}+ 2}$.
{\rm (ii)} If $A<\infty$ then
\begin{equation}\label{limsupHessb}
\limsup_{t\to\infty}\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx
\le C \mathcal A_3^{2\mu_8+1},
\end{equation}
where $\mathcal A_3$ is defined by \eqref{A3def}.
{\rm (iii)} If $\beta<\infty$, then there is $T>0$ such that for all $t>T$,
\begin{equation}\label{Hess3b}
\begin{split}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx &\le C\Big(1+\beta^\frac1{1-a}+\sup_{[t-3/2,t]} f^\frac2{2-a}\Big)^{\frac{2\mu_7}{\mu_6} + 2}
\big(1 +\sup_{[t-3/2,t]}\|\psi_t\|_{L^\infty}\big)^{\mu_{10}}.
\end{split}
\end{equation}
Above, the positive constant $C$ depends on $U'$.
\end{theorem}
\begin{proof}
(i) For $t>0$, we obtain from \eqref{delunite} with $\delta=a$ that
\begin{equation}\label{Hesspb}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx \le C(1+ \norm{\nabla p}^{2-a}_{L^{2-a}})(1 + \|\bar p_t\|_{L^\infty( V )}^{2}).
\end{equation}
Then using estimates \eqref{H-bound-0} and \eqref{tpos} in \eqref{Hesspb}, we obtain
\begin{multline*}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx
\le C \Big\{ 1+\norm{\bar p_0}_{L^2}^2+\|\nabla p_0\|_{L^{2-a}}^{2-a} +(1+t)\sup_{[0,t]} f +\int_{0}^t \tilde f(\tau)d\tau\Big\}^{\frac{2\mu_7}{\mu_6}+2} \\
\quad \cdot\Big\{1+ t^{-\frac 1{2\mu_6}} + t^{\frac 1 {2(\mu_6+1)}} \sup_{[0,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big\}^2.
\end{multline*}
Then \eqref{hess1b} follows.
If $t\ge 3/2$ then using \eqref{H-bound-b} and \eqref{tlarge} in \eqref{Hesspb} we obtain
\begin{align*}
\int_{U'} |\nabla^2 p(x,t)|^{2-a} dx
\le C\Big\{ 1+\norm{\bar p_0}_{L^2}^2 +M_f(t)^\frac2{2-a}+ \int_{t-3/2}^t \tilde f(\tau)d\tau \Big\}^{\frac{2\mu_7}{\mu_6}+ 2}\\
\cdot\Big\{1+\sup_{[t-1/2, t]} \norm{\psi_t}_{L^\infty}^{\frac 1{\mu_6+1}}\Big\}^2.
\end{align*}
Then \eqref{h2bcomp} follows.
(ii) If $A<\infty$ then using \eqref{limsupH1} and \eqref{limsup-pt10} in \eqref{Hesspb} we obtain
\begin{equation*}
\limsup_{t\to\infty}\int_{U'} |\nabla^2 p(x,t) |^{2-a} dx \le C \mathcal A_2 \mathcal A_3^{2\mu_8}.
\end{equation*}
This yields \eqref{limsupHessb}.
(iii) If $\beta<\infty$ then using \eqref{unbounJ} and \eqref{ptbeta} in \eqref{Hesspb} we obtain \eqref{Hess3b}.
\end{proof}
Next, we treat the case $\delta\in(0,a)$ for which we define the following exponents
\begin{align}
&\nu_1= \max\{2,\frac{(2-\delta)a}\delta\},
&&
\nu_2=2+\frac{2\mu_7}{\mu_6}+\frac 1 {\mu_6+1}+\frac{2(\nu_1-2)}{2-a},\\
&\nu_3=\frac{2-a}{1-a}\Big(\frac{2\mu_7}{\mu_6}+1\Big)+\frac{\nu_1-a}{1-a},
&&
\nu_4=\frac 2{2-a}\Big(\frac{2\mu_7}{\mu_6} + 1\Big)+\mu_4(\nu_1-2),
\\
&\mu_{11}=\frac{2-a}{1-a}\Big(\frac{2\mu_7}{\mu_6}+1\Big)+\frac2{\mu_6+1},
&&\mu_{12}=\frac{2\mu_7}{\mu_6} + 1 +\frac{1-a}{2-a}\frac 2{\mu_6+1}.
\end{align}
\begin{theorem}\label{HessThm3}
Assume \eqref{strictdeg}. Let $\delta$ be any number in $(0,a)$.
For $t>0$, we have
\begin{equation}\label{hess1c}
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx
\le C L_9 (1+ t^{-\frac 1{\mu_6}}) (t+1)^{\nu_2} \mathcal K_{1,t}^{\nu_3} ( \sup_{[0,t]}\|\psi_t\|_{L^\infty}+1)^{\mu_{11}},
\end{equation}
where
$L_9=(1+ \norm{\bar p_0}_{L^2}^2+\|\nabla p_0\|_{L^{2-a}}^{2-a}) L_2(\nu_1).$
For $t\ge 3/2$, we have
\begin{equation}\label{hess2c}
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx
\le C L_{10} (1+M_f(t))^{\nu_4} \Big( 1+\sup_{[t-3/2,t]} \tilde f\Big)^{\mu_{12}}\Big\{ 1+ \int_0^t f(\tau)d\tau\Big\},
\end{equation}
where
$L_{10}=( 1+\norm{\bar p_0}_{L^2}^2)L_4(\nu_1).$
Above, the positive constant $C$ depends on $U'$ and $\delta$, and $L_2(\cdot)$, $L_4(\cdot)$ are defined in Theorems \ref{thm312} and \ref{cor311}.
\end{theorem}
\begin{proof}
We use estimate \eqref{delunite} and utilizing \eqref{grad20pw} with $s=\nu_1$ and \eqref{tpos} to obtain
\begin{multline*
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx
\le C \Big\{ 1+\norm{\bar p_0}_{L^2}^2+\|\nabla p_0\|_{L^{2-a}}^{2-a}+(1+t)\mathcal K_{1,t}^\frac{2-a}{1-a} +\mathcal K_{2,t}\Big\}^{\frac{2\mu_7}{\mu_6}+1} \\
\quad \cdot\Big\{1+ t^{-\frac 1{2\mu_6}} + t^{\frac 1 {2(\mu_6+1)}} \sup_{[0,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big\}^2
L_2(\nu_1) (t+1)^{\frac{2(\nu_1-2)}{2-a}+1} \mathcal K_{1,t}^{\frac{\nu_1-a}{1-a}}.
\end{multline*}
Then \eqref{hess1c} follows. For $t\ge 3/2$, using \eqref{Kgrad55pw} for $s=\nu_1$ instead of \eqref{grad20pw} and using \eqref{tlarge} instead of \eqref{tpos}, we obtain
\begin{multline*
\int_{U'} |\nabla^2 p(x,t)|^{2-\delta} dx
\le C \Big( 1+\norm{\bar p_0}_{L^2}^2 +M_f(t)^{\frac 2{2-a} }+ \int_{t-3/2}^t \tilde f(\tau)d\tau \Big)^{\frac{2\mu_7}{\mu_6} + 1 } \Big(1 +\sup_{[t-1/2,t]}\|\psi_t\|_{L^\infty}^\frac1{\mu_6+1}\Big)^2\\
\cdot L_4(\nu_1)\Big(1+M_f(t) \Big)^{\mu_4(\nu_1-2)}\Big\{1+ \int_0^t f(\tau)d\tau\Big\}.
\end{multline*}
Note that $\|\psi_t(t)\|_{L^\infty}\le C(1+\tilde f(t))^\frac{1-a}{2-a}$.
Then we obtain \eqref{hess2c}.
\end{proof}
For asymptotic estimates, we have the following.
\begin{theorem}\label{HessLim}
Assume \eqref{strictdeg}.
Suppose $\Upsilon_1$ and $\Upsilon_2$ defined in Corollary \ref{corMM} are finite numbers.
Then for any $\delta\in (0,a)$, we have
\begin{equation}\label{limhessU}
\limsup_{t\to\infty} \int_{U'} |\nabla p(x,t)|^{2-\delta} dx
\le C L_4(\nu_1) \Upsilon_1^{\mu_4(\nu_1-2)} \Upsilon_2
\mathcal A_3^{2\mu_8},
\end{equation}
where $C>0$ depends on $U'$ and $\delta$.
\end{theorem}
\begin{proof}
Taking limit superior of \eqref{delunite} as $t\to\infty$ and using \eqref{Kgrad60pw} for $s=\nu_1$, and using \eqref{limsup-pt10}, we obtain
\eqref{limhessU}.
\end{proof}
\clearpage
\section{Dependence on initial and boundary data}
\label{datacont}
In this section, we prove the continuous dependence of solutions $p(x,t)$ of the IBVP \eqref{eqorig}, \eqref{BC} and \eqref{In-Cond} with respect to the $L^\infty$-norm on the initial data and boundary data. The results are established for either finite time intervals or at time infinity.
Let $p_1(x,t)$ and $p_2(x,t)$ be two solutions of the IBVP \eqref{eqgamma} having fluxes $\psi_1$ and $\psi_2$, and initial data $p_1(x,0)$ and $p_2(x,0)$, respectively.
Let $\Psi=\psi_1-\psi_2$, $P=p_1-p_2$, and $\bar P=P-|U|^{-1}\int_U P dx$. Then by \eqref{pbargam},
\begin{equation}
\bar P(x,t)
=\bar p_1(x,t)-\bar p_2(x,t)
=P(x,t)- \frac{1}{|U|}\int_U P(x,0) dx + \frac{1}{|U|}\int_0^t\int_\Gamma \Psi(x,\tau) d\sigma d\tau.
\end{equation*}
We will estimate $\norm{\bar P}_{L^\infty(U')}$ where the subset $U'$ satisfies $U'\Subset U$ throughout the section.
From \eqref{eqgamma} follows
\begin{equation}\label{eq-1} \frac{\partial\bar P}{\partial t}=\nabla \cdot \big (K(|\nabla \bar p_1|)\nabla \bar p_1 \big ) - \nabla \cdot \big (K(|\nabla \bar p_2|)\nabla \bar p_2 \big )
+\frac{1}{|U|}\int_\Gamma \Psi(x,t) d\sigma. \end{equation}
The following quantity will be used throughout this section:
\begin{equation*}
\Lambda(t) =1 + \int_U \Big(|\nabla p_1(x,t)|^{2-a} + |\nabla p_2(x,t)|^{2-a}\Big) dx,
\end{equation*}
\subsection{Results for pressure}
\label{sec41}
First we establish $L^\infty$-estimates for $\bar P$ in terms of its $L^2$- and $W^{1,s}$- norms.
We recall that the exponents $\mu_5$, $\mu_6$ and $\mu_7$ are defined in \eqref{sdef}.
\begin{proposition}\label{cont:strictcond} Assume \eqref{strictdeg}.
Let $ U'\Subset V \Subset U$.
Let $\mu> \mu_6^{-1}=\frac {\mu_5} {{\mu_5}-2}$ and denote
\begin{equation} \label{etadef}
{\gamma_1}=\mu_6 -\frac1\mu =1- \frac 2 {\mu_5} -\frac 1 \mu\in(0,1).
\end{equation}
Then we have for any $T_0\ge 0$, $T>0$ and $\theta\in(0,1)$ that
\begin{equation}\label{Pinfty1}
\sup_{[T_0+\theta T,T_0+T]} \|\bar P\|_{L^\infty(U')}\le C\, {\mathcal C}_{T_0,T,\theta}\Big( \norm{\bar P}_{L^2(U\times (T_0,T_0+T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (T_0,T_0+T) )}\Big),
\end{equation}
where the positive constant $C$ is independent of $T_0$, $T$ and $\theta$,
\begin{equation}\label{Cthe3}
\mathcal C_{T_0,T,\theta}= (\lambda T^{1/2} \vartheta)^{ \frac{1}{{\gamma_1}+1}} + \Big(\lambda\Big[1+ \frac1{\theta T}\Big]^{1/2}\Big)^\frac1{\mu_6} + \Big(\lambda T^{1/2}\sup_{[T_0,T_0+T]}\norm{\Psi}\Big)^{ \frac{1}{\mu_6+1}},
\end{equation}
with
\begin{align}
\label{lamdef2}
&\lambda = \lambda_{T_0,T} \mathbin{\buildrel \rm def \over \longequals} \sup_{t\in [T_0,T_0+T]} \Lambda(t) ^{\mu_7},\\
\label{updef}
& \vartheta =\vartheta_{T_0,T} \mathbin{\buildrel \rm def \over \longequals} \Big[ \int_{T_0}^{T_0+T}\int_{V} \Big(|\nabla \bar p_1|^{2\mu(1-a)} + |\nabla \bar p_2|^{2\mu(1-a)} \Big)dx dt\Big]^\frac{1}{2\mu}.
\end{align}
\end{proposition}
\begin{proof}
Without loss of generality, we assume $T_0=0$.
Let $\zeta(x,t)=\phi(x)\varphi(t)$ be a cut-off function with $\varphi(0)=0$ and supp\,$\phi \subset V$.
Same as in Proposition \ref{p-T.est} we define
\begin{align*}
\bar P^{(k)} =\max\{\bar P-k,0\},
\quad S_{k}(t)=\{ x\in U: \bar P^{(k)}(x,t)\ge 0\},
\end{align*}
and denote by $\chi_k(x,t)$ the characteristic function of $S_k(t)$.
Multiplying equation \eqref{eq-1} by $\bar P^{(k)} \zeta^2 $, integrating it over $U$ and using integration by parts, we have
\begin{align*}
&\frac 1 2\frac{d}{dt} \int_U |\bar P^{(k)}\zeta|^2 dx
=\int_U |\bar P^{(k)}|^2\zeta \zeta_t dx \\
&\quad -\int_U \Big(K(\nabla \bar p_1|)\nabla \bar p_1-K(|\nabla \bar p_2|)\nabla \bar p_2 \Big ) \cdot \nabla (\bar P^{(k)} \zeta^2) dx
+\frac{1}{|U|}\int_\Gamma \Psi(x,t) d\sigma\int_U\bar P^{(k)}\zeta^2 dx.
\end{align*}
Elementary calculations yield
\begin{align*}
&\frac 1 2\frac{d}{dt} \int_U |\bar P^{(k)}\zeta|^2 dx
\le \int_U |\bar P^{(k)}|^2\zeta |\zeta_t| dx -\int_U \Big(K(\nabla \bar p_1|)\nabla \bar p_1-K(|\nabla \bar p_2|)\nabla \bar p_2 \Big ) \cdot \nabla \bar P^{(k)} \zeta^2 dx\\
&\quad -2 \int_U \Big(K(\nabla \bar p_1|)\nabla \bar p_1-K(|\nabla \bar p_2|)\nabla \bar p_2 \Big ) \cdot \bar P^{(k)} \zeta \nabla \zeta dx +\frac{1}{|U|}\int_\Gamma \Psi(x,t) d\sigma\int_U\bar P^{(k)}\zeta^2 dx.
\end{align*}
Denoting the last four summands by $I_i$, $i=1,2,3,4$, respectively, we rewrite the above as
\begin{equation}\label{EqPprime}
\frac 1 2\frac{d}{dt} \int_U |\bar P^{(k)}\zeta|^2 dx \le I_1+I_2+I_3+I_4.
\end{equation}
Let $\xi(x,t)=|\nabla \bar p_1|\vee |\nabla \bar p_2|$.
We consider $I_2$. Let $J(x,t)=\Big(K(\nabla \bar p_1|)\nabla \bar p_1-K(|\nabla \bar p_2|)\nabla \bar p_2 \Big ) \cdot \nabla \bar P^{(k)} $.
On the set $U\setminus S_k(t)$, since $\nabla \bar P^{(k)}=0$ a.e., we have $J(x,t)=0$ a.e..
On the set $S_k(t)$, we have $\nabla \bar P^{(k)}=\nabla \bar P$ a.e., and, by the monotonicity \eqref{mono1}, we have for almost all $x\in S_k(t)$ that
\begin{align*}
J(x,t)
& = \big( K(|\nabla \bar p_1|)\nabla \bar p_1-K(\nabla \bar p_2|)\nabla \bar p_2 \big)\cdot (\nabla \bar p_1 -\nabla \bar p_2)\\
&\ge (1-a)K(|\nabla \bar p_1|\vee |\nabla \bar p_2|) |\nabla \bar P|^2
= (1-a)K(\xi) |\nabla P^{(k)}|^2.
\end{align*}
Therefore,
\begin{equation}\label{oriI2}
I_2\le - (1-a) \int_U K(\xi)|\nabla \bar P^{(k)}|^2 \zeta^2dx.
\end{equation}
Note that
\begin{equation*}
|\zeta\nabla \bar P^{(k)}|^2 = |\nabla(\bar P^{(k)} \zeta)- \bar P^{(k)}\nabla \zeta |^2 = |\nabla (\bar P^{(k)} \zeta) |^2 - 2\nabla (\bar P^{(k)} \zeta) \cdot (\bar P^{(k)} \nabla \zeta ) + |\bar P^{(k)} \nabla \zeta|^2 .
\end{equation*}
Thus, \eqref{oriI2} gives
\begin{align*}
I_2
&\le - (1-a) \int_U K(\xi)|\nabla (\bar P^{(k)} \zeta) |^2dx +2(1-a)\int_U K(\xi)|\nabla (\bar P^{(k)} \zeta)| | P^{(k)}\nabla \zeta |dx\\
&\quad -(1-a)\int_U K(\xi)|\bar P^{(k)} \nabla\zeta|^2dx .
\end{align*}
Using Cauchy's inequality for the second term on the right-hand side of the previous inequality gives
\begin{equation}\label{I2}
\begin{aligned}
I_2
&\le - \frac {1-a} 2 \int_U K(\xi)|\nabla (\bar P^{(k)} \zeta) |^2dx + (1-a)\int_U K(\xi)|\bar P^{(k)} \nabla\zeta|^2dx \\
&\le - \frac {1-a} 2 \int_U K(\xi)|\nabla (\bar P^{(k)} \zeta) |^2dx + C\int_U |\bar P^{(k)} \nabla\zeta|^2dx .
\end{aligned}
\end{equation}
For the last inequality, we used the fact that $K(\xi)$ is bounded above.
For $I_3$ and $I_4$, we have for any $\varepsilon>0$ that
\begin{align}
\label{I3} I_3
&\le C \int_U |\bar P^{(k)} | (|\nabla \bar p_1|^{1-a} + |\nabla \bar p_2|^{1-a}) \zeta |\nabla \zeta| dx \nonumber \\
&\le \varepsilon \int_U |\bar P^{(k)} \zeta |^2 dx + C \varepsilon^{-1}\int_U (|\nabla \bar p_1|^{1-a} + |\nabla \bar p_2|^{1-a})^2 \chi_k |\nabla \zeta|^2 dx,\\
\label{I4} I_4&\le\varepsilon \int_U |\bar P^{(k)} \zeta |^2 dx + C\varepsilon^{-1}\|\Psi(t)\|_{L^\infty}^2\int_U \chi_k \zeta^2 dx.
\end{align}
Combining \eqref{EqPprime}, \eqref{I2}, \eqref{I3} and \eqref{I4} yields
\begin{align*}
& \frac 1 2\frac{d}{dt} \int_U |\bar P^{(k)}\zeta |^2 dx
+ \frac{1-a}2 \int_U K(\xi) |\nabla \bar P^{(k)}\zeta|^{2-a}dx
\le C \int_U | \bar P^{(k)} |^2 (|\zeta_t|\zeta + |\nabla \zeta|^2) dx \\
&\quad+2\varepsilon \int_U |\bar P^{(k)}\zeta|^2 dx + C\varepsilon^{-1} \|\Psi(t)\|_{L^\infty}^2 \int_U \chi_k \zeta^2 dx
+C\varepsilon^{-1}\int_U (|\nabla \bar p_1|^{1-a} + |\nabla \bar p_2|^{1-a})^2 \chi_k |\nabla \zeta|^2 dx.
\end{align*}
Integrating in time, taking the supremum in $t$ over $[0,T]$, and selecting $\varepsilon =1/(16T)$ we find that
\begin{multline*}
\sup_{[0,T]}\norm{\bar P^{(k)}\zeta}_{L^2(U)}^2
+ \int_0^T \int_U K(\xi) |\nabla \bar P^{(k)} \zeta |^2dx dt\le C\int_0^T \int_U | \bar P^{(k)}|^2 (|\zeta_t|\zeta +|\nabla \zeta|^2 )dxdt\\
\quad +CT\sup_{[0,T]}\|\Psi\|_{L^\infty}^2 \int_0^T \int_U \chi_k \zeta^2 dx dt+CT \int_0^T\int_U (|\nabla \bar p_1|^{1-a} + |\nabla \bar p_2|^{1-a})^2 \chi_k |\nabla \zeta|^{2} dx dt.
\end{multline*}
Applying H\"older's inequality to the last double integral yields
\begin{equation}\label{Pkpre}
\begin{aligned}
&\sup_{[0,T]}\norm{\bar P^{(k)}\zeta}_{L^2(U)}^2
+ \int_0^T \int_U K(\xi) |\nabla \bar P^{(k)} \zeta |^2dx dt\\
&\le C\int_0^T \int_U | \bar P^{(k)}|^2 (|\zeta_t|\zeta +|\nabla \zeta|^2 )dx
+CT\sup_{[0,T]}\|\Psi\|_{L^\infty}^2 \int_0^T \int_U \chi_k \zeta^2 dxdt\\
&\quad + CT\Big( \int_0^T\int_{V} (|\nabla \bar p_1|^{1-a} + |\nabla \bar p_2|^{1-a})^{2\mu} |\nabla \zeta|^{2\mu} dx dt\Big)^{1/\mu}
\Big(\iint_{Q_T\cap {\rm supp}\zeta}\chi_k dxdt\Big)^{1-1/\mu}.
\end{aligned}
\end{equation}
Under the Strict Degree Condition, by applying Sobolev inequality \eqref{Wemb} with $W(x,t)=K(\xi(x,t))$, we have
\begin{equation*}
\|\bar P^{(k)}\zeta\|_{L^{\mu_5}(Q_T)}
\le C\lambda \Big(\sup_{[0,T]}\norm{\bar P^{(k)}\zeta}_{L^2(U)}^2
+ \int_0^T \int_U K(\xi) |\nabla \bar P^{(k)} \zeta |^2dx dt\Big)^{1/2},
\end{equation*}
where $\lambda$ is defined by \eqref{lamdef2}. Combining with \eqref{Pkpre}, we have
\begin{multline}\label{Pkzeta}
\|\bar P^{(k)}\zeta\|_{L^{\mu_5}(Q_T)}
\le C\lambda \Big[ \Big(\int_0^T \int_U | \bar P^{(k)}|^2 (|\zeta_t|+ |\nabla \zeta|^2)dxdt\Big)^{1/2}\\
+ T^{1/2} \sup_{[0,T]} \|\Psi\|_{L^\infty}\Big(\iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big)^{1/2}
+T^{1/2}\vartheta \Big(\iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big)^{\frac 12(1-1/\mu)}\Big],
\end{multline}
where $\vartheta$ is defined by \eqref{updef}.
Let $M_0 >0$ be fixed which we will determine later.
For integers $i\ge 0$, let $k_i= M_0(1-2^{-i})$ and where $\zeta_i$ be defined as in \eqref{der:cutoff} and the sets $\mathcal Q_i$, $A_{i,j}$ be defined similar to \eqref{defQnAnm} replacing $\bar p_t$ by $\bar P$.
Define $F_{i} = \norm{\bar P ^{(k_{i+1})}\zeta_i}_{L^{\mu_5}(A_{i+1,i}) }$.
Applying \eqref{Pkzeta} with $k=k_{i+1}$ and $\zeta=\zeta_i$ we have
\begin{equation}\label{more1}
\begin{aligned}
F_{i}
&\le C\lambda \Big[ \big(2^{i/2} (\theta T)^{-1/2} + 2^i\big) \norm{ \bar P^{(k_{i+1})}}_{L^2(A_{i+1, i})}\\
&\quad +2^iT^{1/2}\vartheta |A_{i+1,i}|^{\frac 12(1-1/\mu)} + T^{1/2}\sup_{[0,T]}\|\Psi\|_{L^\infty} |A_{i+1,i}|^{1/2} \Big].
\end{aligned}
\end{equation}
Estimating the same way as in \eqref{qki+1}, we have
\begin{equation}\label{more2}
\|\bar P^{(k_{i+1})}\zeta_i \|_{L^2(A_{i+1,i+1})} \le C F_{i} |A_{i+1,i}|^{1/2-1/{\mu_5}}.
\end{equation}
Also, similar to \eqref{Aii}, we have
\begin{equation}\label{more3}
|A_{i+1,i}| \le C 4^{i} M_0^{-2}\|\bar P^{(k_i)}\|_{L^2(A_i)}^{2}.
\end{equation}
From estimates \eqref{more1}, \eqref{more2}, and the boundedness of $\zeta_t$, we obtain
\begin{align*}
&\|\bar P^{(k_{i+1})}\|_{L^2(A_{i+1})}
=\|\bar P^{(k_{i+1})}\zeta_i\|_{L^2(A_{i+1})}
\le C\lambda \Big\{ 2^{i}( (\theta T)^{-1/2} + 1) \norm{ \bar P^{(k_{i+1})}}_{L^2(A_{i+1, i})}\\
&\quad +2^iT^{1/2}\vartheta |A_{i+1,i}|^{\frac 12(1-1/\mu)} + T^{1/2}\sup_{[0,T]}\|\Psi\|_{L^\infty}|A_{i+1,i}|^{1/2} \Big\} |A_{i+1,i}|^{1/2-1/{\mu_5}}.
\end{align*}
Using \eqref{more3} gives
\begin{align*}
\|\bar P^{(k_{i+1})}\|_{L^2(A_{i+1})}
&\le C\lambda \Big\{ 2^{i}( (\theta T)^{-1/2} + 1) \norm{ \bar P^{(k_{i+1})}}_{L^2(A_{i+1, i})}
+ 4^iT^{1/2}\vartheta M_0^{-1+1/\mu} \|\bar P^{(k_i)}\|_{L^2(A_i)}^{1-1/\mu}\\
&\quad + T^{1/2}2^iM_0^{-1}\sup_{[0,T]}\|\Psi\|_{L^\infty} \|\bar P^{(k_i)}\|_{L^2(A_i)} \Big\} 2^{i} M_0^{-(1-\frac 2 {\mu_5})}\|\bar P^{(k_i)}\|_{L^2(A_i)}^{1-\frac 2 {\mu_5}}.
\end{align*}
Put $Y_i=\| \bar P^{(k_i)} \|_{L^2(A_i)}$, the previous inequality gives
\begin{align*}
&Y_{i+1}\le C\lambda M_0^{-1+\frac 2 {\mu_5}}8^i
\Big\{\vartheta T^\frac12 M_0^{-1+\frac1\mu} Y_i^{1-\frac1\mu} + \Big[(1+ \frac1{\theta T})^\frac12+T^\frac12\sup_{[0,T]}\|\Psi\|_{L^\infty} M_0^{-1}\Big]Y_i \Big\} Y_i^{1-\frac 2 s}\\
&\le C 8^i
\Big\{\lambda\vartheta T^\frac12 M_0^{-1-{\gamma_1}} Y_i^{1+{\gamma_1}}
+ \lambda\Big[(1+ \frac1{\theta T})^\frac12 M_0^{-\mu_6}\Big]Y_i^{1+\mu_6}
+\lambda T^\frac12\sup_{[0,T]}\|\Psi\|_{L^\infty}M_0^{-1-\mu_6}Y_i^{1+\mu_6} \Big\}.
\end{align*}
Since $Y_0\le \norm{\bar P}_{L^2(U\times (0,T) )} $, aiming at applying Lemma~\ref{multiseq} with $m=3$ we choose $M_0$ sufficiently large such that
\[
\begin{split}
\norm{\bar P}_{L^2(U\times (0,T) )}
\le C\min\Big\{
&(\lambda T^{1/2} \vartheta)^{- 1/{\gamma_1}}M_0^\frac{1+{\gamma_1}}{{\gamma_1}}, (\lambda(1+ \frac1{\theta T})^{1/2})^{-1/\mu_6}M_0, \\
& (\lambda T^{1/2}\sup_{[0,T]}\|\Psi\|_{L^\infty})^{- 1/\mu_6} M_0^\frac{1+\mu_6}{\mu_6} \Big\}.
\end{split}
\]
Thus, we require
\begin{align*}
M_0&\ge C(\lambda T^{1/2}\vartheta)^{ \frac{1}{{\gamma_1}+1}} \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}},\quad
M_0\ge C (\lambda(1+ \frac1{\theta T})^{1/2})^{1/\mu_6} \norm{\bar P}_{L^2(U\times (0,T) )}, \\
M_0&\ge C(\lambda T^{1/2} \sup_{[0,T]}\|\Psi\|_{L^\infty})^{ \frac{1}{\mu_6+1}} \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{\mu_6}{\mu_6+1}}.
\end{align*}
We choose
\begin{align}\label{M0}
M_0 \ge C \Big((\lambda T^{1/2} \vartheta)^{ \frac{1}{{\gamma_1}+1}} + (\lambda(1+ \frac1{\theta T})^{1/2})^{1/\mu_6} + (\lambda T^{1/2} \sup_{[0,T]}\norm{\Psi}_{L^\infty})^{ \frac{1}{\mu_6+1}} \Big)\nonumber\\
\cdot\Big( \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (0,T) )} + \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{\mu_6}{\mu_6+1}}\Big),
\end{align}
for some constant $C>0$.
Since $\lambda\ge 1$ and ${\gamma_1}<\mu_6<1$, we select
\begin{equation}\label{firstM}
\begin{aligned}
M_0 &= C \Big((\lambda T^{1/2} \vartheta)^{ \frac{1}{{\gamma_1}+1}} + (\lambda(1+ \frac1{\theta T})^{1/2})^{1/\mu_6} + ( \lambda T^{1/2}\sup_{[0,T]}\norm{\Psi})^{ \frac{1}{\mu_6+1}} \Big) \\
&\quad \cdot \Big( \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (0,T) )}\Big).
\end{aligned}
\end{equation}
Applying Lemma \ref{multiseq} and using the same arguments as in Proposition \ref{local-L-infty}, we obtain \eqref{Pinfty1}.
\end{proof}
We now derive a variation of Proposition \ref{cont:strictcond}. Let $T_0=0$.
In the proof of Proposition \ref{cont:strictcond}, let $\zeta =\zeta(x) $ be the cut-off function such that $\zeta$ vanishes in neighborhood of the boundary $\Gamma$.
If $k\ge \|\bar P_0\|_{L^\infty}$, then $\bar P^{(k)}(0)=0$ and, hence, \eqref{Pkzeta} holds true with $\zeta_t=0$.
In the iteration process, we choose $M_0 \geq \|\bar P(\cdot, 0)\|_{L^\infty}$ and $k_i=M_0(2-2^{-i})$ for $i\ge 0$. This way, we have $k_i \geq M_0\ge \|\bar P_0\|_{L^\infty}$, thus, $\bar P^{(k_i)}(0)=0$.
Therefore, we can replace $\frac1{\theta T}$ by $0$ in \eqref{M0}, which becomes
\begin{align*}
M_0 &\ge C \Big((\lambda T^{1/2} \vartheta)^{ \frac{1}{{\gamma_1}+1}} + \lambda^{1/\mu_6} + (\lambda T^{1/2} \sup_{[0,T]}\norm{\Psi})^{ \frac{1}{\mu_6+1}} \Big) \\
&\quad \cdot\Big( \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (0,T) )} + \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{\mu_6}{\mu_6+1}}\Big).
\end{align*}
Therefore, instead of \eqref{firstM}, we can select
\begin{align*}
M_0
&= \|\bar P(0)\|_{L^\infty} \\
&\quad +C \Big((\lambda T^{1/2} \vartheta)^{ \frac{1}{{\gamma_1}+1}} + \lambda^{1/\mu_6}+ (\lambda T^{1/2} \sup_{[0,T]}\norm{\Psi})^{ \frac{1}{\mu_6+1}} \Big)
\Big( \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (0,T) )}\Big).
\end{align*}
Applying Lemma \ref{multiseq}, we have
$\lim_{i \rightarrow \infty}Y_i = 0$
and $\bar p(x,t)\le 2M_0$ a.e. in $B(x_0,\rho/4)\times(\theta T,T)$. Proceeding the proof as in Proposition \ref{local-L-infty}, we obtain the following result.
\begin{proposition}\label{newcont} Assume the same as in Proposition \ref{cont:strictcond}. Then we have for any $T>0$ that
\begin{equation}\label{Pinfty2}
\sup_{[0,T]} \|\bar P\|_{L^\infty(U')}\le 2 \|\bar P(0)\|_{L^\infty}+C\, \mathcal C_{T}\Big( \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{{\gamma_1}}{{\gamma_1}+1}}+ \norm{\bar P}_{L^2(U\times (0,T) )}\Big),
\end{equation}
where
\begin{equation}\label{Cthe10}
\mathcal C_{T}= (\lambda_T T^{1/2} \vartheta_T)^{ \frac{1}{{\gamma_1}+1}} + \lambda_T^\frac{1}{\mu_6}+\Big(\lambda_T T^{1/2}\sup_{[0,T]}\norm{\Psi}_{L^\infty}\Big)^{ \frac{1}{\mu_6+1}},
\end{equation}
with $\lambda_T=\lambda_{0,T}$ and $\vartheta_T=\vartheta_{0,T}$ defined in Theorem \ref{cont:strictcond} for $T_0=0$.
\end{proposition}
We will use Propositions \ref{cont:strictcond} and \ref{newcont} to obtain specific $L^\infty$-estimates of $\bar P$ in terms of initial and boundary data.
First, we introduce some quantities and parameters.
Same as \eqref{b}, we define for $i=1,2$,
\begin{equation*}
f_i(t)=\|\psi_i(t)\|_{L^\infty}^2+\|\psi_i(t)\|_{L^\infty}^{\frac{2-a}{1-a}} \quad\text{and}\quad
\tilde f_i(t)=\|\psi_{it}(t)\|_{L^\infty}^2+\|\psi_{it}(t)\|_{L^\infty}^{\frac{2-a}{1-a}}.
\end{equation*}
For $i=1,2$, we assume $f_i(t),\tilde f_i(t)\in C([0,\infty))$ and when needed $f_i(t)\in C^1((0,\infty))$.
Let
\begin{equation*}
A_i=\limsup_{t\to\infty} f_i(t)\quad \text{and}\quad \beta_i=\limsup_{t\to\infty} [f_i'(t)]^-,
\end{equation*}
\begin{equation*}
\bar A=A_1+A_2\quad \text{and}\quad \bar \beta=\beta_1+\beta_2.
\end{equation*}
Let $M_{f_i}(t)$, $i=1,2$, be a continuous increasing majorant of $f_i(t)$ on $[0,\infty)$.
Set
\begin{equation*}
F(t)= f_1(t)+ f_2(t),\quad \ M_F(t)=M_{f_1}(t)+M_{f_2}(t)\quad\text{and}\quad
\tilde F(t)=\tilde f_1(t)+\tilde f_2(t).
\end{equation*}
For initial data, set
\begin{equation}\label{AB0}
A_0=\|\bar p_1(0)\|_{L^2}^2 +\|\bar p_2(0)\|_{L^2}^2 \quad \text{and} \quad B_0=J_H[p_1](0)+J_H[p_2](0).
\end{equation}
We recall results from \cite{HI2}.
\begin{theorem}[cf. \cite{HI2}, Lemma 5.5 and Theorem 5.6]\label{pgrowth}
{\rm (i)} For $t\ge 0$,
\begin{align}\label{pcont}
\norm{\bar P(t)}_{L^2}^2 \le \norm{\bar P(0)}_{L^2}^2
+C\int_0^t\norm{\Psi(\tau)}_{L^\infty}^2\Lambda(\tau)^b d\tau,
\end{align}
where $b$ is defined in \eqref{ab}.
Consequently, for any $T>0$,
\begin{align}\label{pLip}
\sup_{[0,T]}\norm{\bar P}_{L^2}^2 \le \norm{\bar P(0)}_{L^2}^2
+C\cdot M_{1,T} \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}^2 ,
\end{align}
where $M_{1,T}=A_0+T+\int_0^T\big[ f_1(\tau)+f_2(\tau)\big] d\tau$.
{\rm (ii)} Assume the Degree Condition. Then for $t\ge 0$
\begin{equation}\label{psubt}
\norm{\bar P(t)}_{L^2}^2 \le e^{-d_2\int_0^t \Lambda^{-b}(\tau)d\tau} \norm{\bar P(0)}_{L^2}^2
+\int_0^t e^{-d_2\int_\tau^t \Lambda^{-b}(\theta)d\theta}\norm{\Psi(\tau)}_{L^\infty}^2\Lambda(\tau)^{b} d\tau. \end{equation}
Moreover, if $\bar A<\infty$ and
$\int_1^\infty(1+\int_{\tau-1}^\tau \tilde F(s)ds )^{-b} d\tau=\infty$
then
\begin{equation}\label{psub-bounded}
\limsup_{t\to\infty} \norm{\bar P(t)}_{L^2}^2
\le C \limsup_{t\to\infty}\Big\{\norm{\Psi(t)}_{L^\infty}^2\Big(1+{\bar A}^\frac2{2-a}+\int_{t-1}^t \tilde F(\tau)d\tau\Big)^{2b}\Big\}.\end{equation}
\end{theorem}
We now state the continuous dependence in interior $L^\infty$-norm.
We use similar quantities to $\mathcal K_{1,t}$ in \eqref{Nf1} and $\mathcal K_{2,t}$ \eqref{Nf2}, namely,
\begin{equation}\label{Nfbar}
\bar\mathcal K_{1,T}= 1+\sum_{i=1,2} \sup_{[0,T]}\|\psi_i\|_{L^\infty} \quad\text{and}\quad
\bar\mathcal K_{2,T}= 1 + \int_0^T \tilde F(\tau)d\tau.
\end{equation}
\begin{theorem}\label{thm46}
Assume \eqref{strictdeg}.
Let $\mu$ be any number satisfying
\begin{equation}
\mu>\frac {\mu_5}{{\mu_5}-2}\quad\text{and}\quad \mu\ge \frac{2-a}{2(1-a)}.
\end{equation}
For $T>0$, we have
\begin{equation}\label{contfinite10}
\begin{aligned}
&\sup_{[0,T]}\|\bar P\|_{L^\infty(U')}
\le 2 \|\bar P(0)\|_{L^\infty}\\
&\quad + C L_{11} M_{2,T} \Big(\|\bar P(0)\|_{L^2} + \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}
+\big[ \|\bar P(0)\|_{L^2} +\sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}\big]^{\frac { {\gamma_1}}{{\gamma_1}+1}}\Big),
\end{aligned}
\end{equation}
where ${\gamma_1}$ is defined by \eqref{etadef}, number $L_{11}>0$ depends on the initial data and is defined by \eqref{L13} below, number $M_{2,T}>0$ depends on the boundary data and is defined by \eqref{NT} below.
\end{theorem}
\begin{proof}
Many exponents will be needed in our proof and are defined here:
\begin{align}
\label{mu34}
&\gamma_2=2(1-a)\mu\ge 2-a,
&&\gamma_3=\frac{\gamma_2}{2\mu(1-a)}=1,
\\
\label{mu56}
&\gamma_4=\frac1{2\mu}\Big(\frac{2\gamma_2}{2-a}-1\Big)=\frac{2(1-a)}{2-a}-\frac1{2\mu},
&&\gamma_5=\frac{\mu_7}{\mu_6} + \frac{2\gamma_4+1}{2({\gamma_1}+1)},\\
\label{mu89}
& \gamma_6=\frac{\mu_7(2-a)}{\mu_6(1-a)}+\frac{\gamma_3}{{\gamma_1}+1}.
\end{align}
We will use notation $\gamma_3$ in calculations below instead of its explicit value for the sake of generality which will be needed in section \ref{polycont}.
We prove \eqref{contfinite10} with $M_{2,T}$ explicitly given by
\begin{equation}\label{NT}
M_{2,T}= T^{\mu_{20}} \bar\mathcal K_{1,T}^{\gamma_6+\frac{2-a}{2(1-a)}} \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6},
\end{equation}
where the exponent $\mu_{20}$ is $\frac{{\gamma_1}}{2({\gamma_1}+1)}$ in case $T\le 1$ and is $\gamma_5+1$ in case $T> 1$.
We will apply Proposition \ref{cont:strictcond} for $T_0=0$ and Proposition \ref{newcont}.
Fix a subset $V$ of $U$ such that $U'\Subset V\Subset U$.
First, for $M_{1,T}$ in \eqref{pLip} we note that
\begin{equation}\label{M1plus}
1+M_{1,T} \le C \Big\{ 1+\sum_{i=1}^2 \norm{\bar p_i(0)}_{L^2}^2 + (T+1) \Big[ (\sum_{i=1}^2 \sup_{[0,T]} \|\psi_i\|)^\frac{2-a}{1-a}+1\Big] \Big\}
\le C \ell_0 (T+1)\bar\mathcal K_{1,T}^\frac{2-a}{1-a},
\end{equation}
where $\ell_0=1+A_0$.
By \eqref{pLip} and \eqref{M1plus}, we have
\begin{multline}\label{st0}
\norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1} {{\gamma_1}+1}} +\norm{\bar P}_{L^2(U\times (0,T) )}
\le C \ell_0^{1/2} (T+1)^{1/2}\bar\mathcal K_{1,T}^\frac{2-a}{2(1-a)} \\
\cdot \Big (
[T^{1/2}(\|\bar P(0)\|_{L^2} + \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty})]^{\frac { {\gamma_1}}{{\gamma_1}+1}}
+T^{1/2}(\|\bar P(0)\|_{L^2} + \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty})
\Big).
\end{multline}
Second, we will estimate ${\mathcal C}_{T_0,T,\theta}$ in \eqref{Cthe3} with $T_0=0$, and $\mathcal C_T$ in \eqref{Cthe10}.
To simplify our calculations, we will replace ${\mathcal C}_{0,T,\theta}$ in \eqref{Cthe3} by the following upper bound
\begin{equation}\label{Cthesimple}
{\mathcal C}_{T,\theta}= \lambda_T^\frac1{\mu_6} \Big\{ (T^{1/2} \vartheta_T)^{ \frac{1}{{\gamma_1}+1} } + (1+ \frac1{(\theta T)^{1/2}})^\frac1{\mu_6} + (T^{1/2}\sup_{[0,T]}\norm{\Psi}_{L^\infty})^{ \frac{1}{\mu_6+1}} \Big\},
\end{equation}
and replace ${\mathcal C}_T$ in \eqref{Cthe10} by
\begin{equation}\label{CTsimple}
{\mathcal C}_{T}= \lambda_T^\frac{1}{\mu_6} \Big\{ 1+ (T^{1/2} \vartheta_T)^{ \frac{1}{{\gamma_1}+1}} + (T^{1/2}\sup_{[0,T]}\norm{\Psi}_{L^\infty})^{ \frac{1}{\mu_6+1}} \Big\}.
\end{equation}
We now find bounds for involved quantities in \eqref{Cthesimple} and \eqref{CTsimple}.
We have from estimate \eqref{H-bound-0} that
\begin{align*}
\lambda_T=\sup_{t\in [0,T]} \Lambda(t) ^{\mu_7}
&\le C\sum_{i=1}^2\Big\{ 1 + \norm{\bar p_i(0)}_{L^2}^2 + J_H[\bar p_i](0) + (T+1) \sup_{[0,T]} \|\psi_i\|^\frac{2-a}{1-a} +\int_0^T \tilde f_i(\tau)d\tau \Big \}^{\mu_7}.
\end{align*}
Therefore,
\begin{equation}\label{lamd}
\lambda_T\le C \ell_1 (T+1)^{\mu_7} \bar\mathcal K_{1,T}^\frac{\mu_7(2-a)}{1-a} \bar\mathcal K_{2,T}^{\mu_7},
\end{equation}
where $\ell_1=(1 +A_0+B_0)^{2-a}$.
Applying \eqref{grad39} to $s=\gamma_2$ and $V$ replacing $U'$, we have
\begin{equation}\label{gradInt20}
\vartheta_T
= \Big(\int_0^T \int_{V}\Big[ 1+ |\nabla \bar p_1|^{\gamma_2} + |\nabla \bar p_2|^{\gamma_2}\Big]dx dt\Big)^{\frac 1{2\mu}}
\le C \ell_2 (T+1)^{\gamma_4}\bar\mathcal K_{1,T}^{\gamma_3},
\end{equation}
where $\ell_2=\big\{\sum_{i=1,2} L_1\big(\gamma_2+a;[p_i(0)]\big)\big\}^\frac1{2\mu}$.
Also,
\begin{equation}\label{Psisim}
\sup_{[0,T]} \norm{\Psi}_{L^\infty} \le \Big(\sum_{i=1}^2 \sup_{[0,T]} \norm{\psi_i}_{L^\infty}\Big) \le \bar\mathcal K_{1,T}.
\end{equation}
We denote
$D_T=\|\bar P(0)\|_{L^2}^{\frac { {\gamma_1}}{{\gamma_1}+1}} +\|\bar P(0)\|_{L^2}
+ \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}^{\frac { {\gamma_1}}{{\gamma_1}+1}} + \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}.$
We consider $0<T\le 1$ first. By \eqref{st0} and \eqref{gradInt20}, we respectively have
\begin{equation}\label{st10}
\norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1} {{\gamma_1}+1}} +\norm{\bar P}_{L^2(U\times (0,T) )}
\le C \ell_0^\frac12 T^\frac{{\gamma_1}}{2({\gamma_1}+1)} \bar\mathcal K_{1,T}^\frac{2-a}{2(1-a)} D_T,
\end{equation}
\begin{equation}\label{gradInt}
\vartheta_T \le C \ell_2 \bar\mathcal K_{1,T}^{\gamma_3}.
\end{equation}
By \eqref{CTsimple}, \eqref{lamd}, \eqref{gradInt} and \eqref{Psisim}, we have
\begin{equation}\label{st30}
\begin{aligned}
{\mathcal C}_{T}
&\le C \ell_1^\frac1{\mu_6} (T+1)^\frac{\mu_7}{\mu_6} \bar\mathcal K_{1,T}^{\frac{\mu_7(2-a)}{\mu_6(1-a)}} \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6}
\Big\{ 1+ \ell_2^\frac{1}{{\gamma_1}+1} \bar\mathcal K_{1,T}^\frac{{\gamma_3}}{{\gamma_1}+1} + \ell_2^\frac{1}{\mu_6+1} \bar\mathcal K_{1,T}^\frac{1}{\mu_6+1} \Big\}
\le C \ell_3 \bar\mathcal K_{1,T}^{\gamma_6} \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6} ,
\end{aligned}
\end{equation}
where $\ell_3=\ell_1^\frac1{\mu_6} \ell_2^\frac{1}{{\gamma_1}+1}.$
Note that we used the facts $\gamma_3\ge 1$ and ${\gamma_1}<\mu_6<{\gamma_1}+1$.
Applying \eqref{Pinfty2} with the use of \eqref{st10} and \eqref{st30}, we have
\begin{equation}\label{contfinite0}
\begin{aligned}
\sup_{[0,T]} \|\bar P(t)\|_{L^\infty(U')}
&\le 2 \|\bar P(0)\|_{L^\infty} +C \ell_0^\frac12 \ell_3 T^{\frac { {\gamma_1}}{2({\gamma_1}+1)}}\bar\mathcal K_{1,T}^{\gamma_6+\frac{2-a}{2(1-a)}} \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6} \cdot D_T,
\end{aligned}
\end{equation}
hence, obtaining \eqref{contfinite10} for $T\le 1$ with
\begin{equation}\label{L13}
L_{11}=\ell_0^\frac12 \ell_3=\ell_0^\frac12\ell_1^\frac1{\mu_6} \ell_2^\frac{1}{{\gamma_1}+1}.
\end{equation}
Consider $T>1$ now. By \eqref{st0},
\begin{equation}\label{st1}
\norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1} {{\gamma_1}+1}} +\norm{\bar P}_{L^2(U\times (0,T) )}\le C\ell_0^{1/2} T \bar\mathcal K_{1,T}^\frac{2-a}{2(1-a)} D_T.
\end{equation}
Using $\mathcal C_{T,\theta}$ in \eqref{Cthesimple} with bounds \eqref{lamd}, \eqref{gradInt20} and \eqref{Psisim}, we have
\begin{align*}
{\mathcal C}_{T,\theta}
&\le C \ell_1^\frac1{\mu_6} (T+1)^\frac{\mu_7}{\mu_6} \bar\mathcal K_{1,T}^\frac{\mu_7(2-a)}{\mu_6(1-a)} \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6}
\Big\{
1+ T^{\frac 1 {2({\gamma_1}+1)} } \ell_2^\frac1{{\gamma_1}+1} (T+1)^\frac{\gamma_4}{{\gamma_1}+1} \bar\mathcal K_{1,T}^\frac{\gamma_3}{{\gamma_1}+1}
+ T^{\frac 1 {2(\mu_6+1)} } \bar\mathcal K_{1,T}^\frac{1}{\mu_6+1}
\Big\}.
\end{align*}
Thus,
\begin{equation}\label{st3}
{\mathcal C}_{T,\theta}
\le C \ell_3 T^{\gamma_5} \bar\mathcal K_{1,T}^{\gamma_6}\bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6}.
\end{equation}
Therefore, combining \eqref{Pinfty1}, \eqref{st1} and \eqref{st3} yields
\begin{equation}\label{contfinite}
\sup_{[1,T]}\|\bar P\|_{L^\infty(U')}
\le C L_{11} T^{\gamma_5+1}\bar\mathcal K_{1,T}^{\gamma_6+\frac{2-a}{2(1-a)} } \bar\mathcal K_{2,T}^\frac{\mu_7}{\mu_6} \cdot D_T.
\end{equation}
Then combining \eqref{contfinite0} for $T=1$ with \eqref{contfinite}, noting that $\gamma_5\ge \frac{{\gamma_1}}{2({\gamma_1}+1)}$, we obtain \eqref{contfinite10} for $T>1$.
The proof is complete.
\end{proof}
Now, we derive asymptotic estimates as $t\to\infty$. Let $t>2$. Applying Proposition \ref{cont:strictcond} to $T_0=t-1$, $T=1$ and $\theta=1/2$, we have
\begin{equation}\label{bigPbar}
\|\bar P(t)\|_{L^\infty(U')} \le C \tilde{\mathcal C}(t) \Big( \norm{\bar P}_{L^2(U\times (t-1,t) )}^{\frac {\gamma_1}{{\gamma_1}+1} } +\norm{\bar P}_{L^2(U\times (t-1,t) )} \Big),
\end{equation}
where
\begin{equation}\label{Clam}
\tilde{\mathcal C}(t)= [\tilde \lambda(t)\tilde \vartheta(t)]^{ \frac{1}{{\gamma_1}+1}} + \tilde \lambda(t)^\frac1{\mu_6} +[\tilde \lambda(t) \sup_{[t-1,t]}\norm{\Psi}_{L^\infty}]^{ \frac{1}{\mu_6+1}}
\end{equation}
with
\begin{align}
\label{lamtildefn}
&\tilde \lambda(t) = \sup_{\tau\in [t-1,t]} \Lambda(\tau)^{\mu_7},\\
\label{uptildef}
&\tilde \vartheta(t) = \Big[ \int_{t-1}^t\int_{V}\Big( |\nabla \bar p_1|^{2\mu(1-a)} + |\nabla \bar p_2|^{2\mu(1-a)}\Big) dx dt\Big]^\frac{1}{2\mu}.
\end{align}
Before going into specific estimates, we state a general result on the limit superior of $\|\bar P(t)\|_{L^\infty(U')}$ as $t\to\infty$.
It is of the same spirit as of the $L^2$-result \eqref{psub-bounded}.
\begin{lemma}\label{limPprop}
Assume \eqref{strictdeg}. Suppose
\begin{equation}\label{boundcond}
\sup_{\tau\in[t-1,t]}\Lambda(\tau)\le \kappa_0 g(t),\quad \tilde{\mathcal C}(t)\le C B(t)
\end{equation}
for sufficiently large $t$, with some functions $g(t),B(t)\ge 1$.
Let $\kappa_1=d_2 \kappa_0^{-b}$, where $d_2$ is the positive constant in \eqref{psubt}.
If
\begin{equation}\label{condlim1}
\lim_{t\to\infty} B(t)^\frac{2({\gamma_1}+1)}{{\gamma_1}}e^{-\kappa_1\int_{2}^t g(\tau)^{-b}d\tau} =0
\quad\text{and}\quad
\lim_{t\to\infty} g(t)^b\frac {B'(t)}{B(t)} =0,
\end{equation}
then
\begin{align}\label{limsup00}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
&\le C \limsup_{t\to\infty} \Big\{ B(t) \Big( g(t)^{b} \norm{\Psi(t)}_{L^\infty} + \big[ g(t)^{b} \norm{\Psi(t)}_{L^\infty} \big]^\frac{\gamma_1}{{\gamma_1}+1}\Big)\Big\}.
\end{align}
\end{lemma}
\begin{proof}
Let $T>2$ be sufficiently large such that \eqref{boundcond} holds for all $t>T$.
By \eqref{psubt} and \eqref{boundcond}, we have for $t'>T$ that
\begin{align*}
\norm{\bar P(t')}_{L^2}^2
&\le e^{-\kappa_1\int_T^{t'} g^{-b}(\tau)d\tau} \norm{\bar P(T)}_{L^2}^2 \\
&\quad +e^{-\kappa_1\int_T^{t'} g^{-b}(\tau)d\tau} \int_T^{t'} e^{\kappa_1\int_T^\tau g^{-b}(\theta)d\theta}\norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^{b} d\tau.
\end{align*}
Then similar to \eqref{supptint}, we have
\begin{equation}\label{supPtprime}
\sup_{[t-1,t]} \norm{\bar P(t')}_{L^2}^2
\le C\Big(e^{-\kappa_1\int_T^{t} g^{-b}(\tau)d\tau} \norm{\bar P(T)}_{L^2}^2 + \int_T^t e^{-\kappa_1\int_\tau^t g^{-b}(\theta)d\theta} \norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^b d\tau \Big).
\end{equation}
Combining \eqref{bigPbar} with \eqref{boundcond} and estimate \eqref{supPtprime}, we obtain
\begin{multline*
\norm{\bar P(t)}_{L^\infty(U')}
\le C B(t) \Big\{
e^{-\kappa_1\int_T^{t} g^{-b}(\tau)d\tau} \norm{\bar P(T)}_{L^2}^2 + \int_T^t e^{-\kappa_1\int_\tau^t g^{-b}(\theta)d\theta} \norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^b d\tau
\Big\}^{\varrho_1/2} \\
\quad + C B( t) \Big\{
e^{-\kappa_1\int_T^{t} g^{-b}(\tau)d\tau} \norm{\bar P(T)}_{L^2}^2 + \int_T^t e^{-\kappa_1\int_\tau^t g^{-b}(\theta)d\theta} \norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^b d\tau
\Big\}^{1/2},
\end{multline*}
where $\varrho_1=\frac{\gamma_1}{{\gamma_1}+1}$.
Note that $\varrho_1<1$, then by condition \eqref{condlim1}, we have
\begin{equation}\label{stillg}
\begin{aligned}
\limsup_{t\to\infty} \norm{\bar P(t)}_{L^\infty(U')}
&\le C \limsup_{t\to\infty} \Big\{ B(t)^\frac2{\varrho_1} \int_T^t e^{-\kappa_1\int_\tau^t g^{-b}(\theta)d\theta} \norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^b d\tau
\Big\}^{\varrho_1/2}\\
&\quad + C \limsup_{t\to\infty}\Big\{ B(t)^2 \int_T^t e^{-\kappa_1\int_\tau^t g^{-b}(\theta)d\theta} \norm{\Psi(\tau)}_{L^\infty}^2 g(\tau)^b d\tau\Big\}^{1/2}.
\end{aligned}
\end{equation}
The first condition in \eqref{condlim1} and the fact that $B(t)\ge 1$ imply $\int_2^\infty g(\tau)^{-b}d\tau=\infty$.
With this and the second condition in \eqref{condlim1}, we apply Lemma~\ref{difflem2} to each limit in \eqref{stillg} and obtain
\begin{align*
\limsup_{t\to\infty} \norm{\bar P(t)}_{L^\infty(U')}
&\le C \limsup_{t\to\infty} \Big( B(t)^{\frac 2{\varrho_1}}\norm{\Psi(t)}_{L^\infty}^2 g(t)^{2b}\Big)^{\varrho_1/2} \\
&\quad + C \limsup_{t\to\infty} \Big( B(t)^2 \norm{\Psi(t)}_{L^\infty}^2 g(t)^{2b} \Big)^{1/2},
\end{align*}
thus, \eqref{limsup00} follows.
\end{proof}
Between two required estimates in \eqref{boundcond} of Lemma \ref{limPprop}, the second one needs more work. Therefore, we focus on estimating $\tilde {\mathcal C}(t)$ defined by \eqref{Clam}, which contains $\tilde \vartheta(t)$ given by \eqref{uptildef}. For this one, we have from relation \eqref{Kestn} that
\begin{align}
\nonumber
\tilde \vartheta(t) &\le C+C\Big(\int_{t-1}^t \int_{V} K(|\nabla \bar p_1|)|\nabla \bar p_1|^{\gamma_2+a} + K(|\nabla \bar p_2|)|\nabla \bar p_2|^{\gamma_2+a}dx dt\Big)^\frac1{2\mu}
\\
\label{pregam}
&\le C+C\Big(\int_{0}^t \int_{V} K(|\nabla \bar p_1|)|\nabla \bar p_1|^{\gamma_2+a} + K(|\nabla \bar p_2|)|\nabla \bar p_2|^{\gamma_2+a}dx dt\Big)^\frac1{2\mu}.
\end{align}
By \eqref{pregam} and \eqref{Kgrad55},
\begin{align}\label{newgamtil}
\tilde \vartheta(t)\le C L_{12} B_1(t)^\frac{\mu_4(\gamma_2+a-2)}{2\mu}B_2(t)^\frac1{2\mu},
\end{align}
where $L_{12}=\big\{\sum_{i=1,2} L_3\big(\gamma_2+a;[p_i(0)]\big)\big\}^\frac1{2\mu}$,
\begin{equation}
B_1(t)=1+M_F(t)\quad\text{and}\quad
B_2(t)=1+\int_0^t F(\tau)d\tau.
\end{equation*}
\begin{theorem}\label{newP}
Assume \eqref{strictdeg}. Suppose $\bar A<\infty$.
Define
\begin{equation*}
\bar\Upsilon_3 = 1+\sup_{[0,\infty)}(\|\psi_1\|_{L^\infty}+\|\psi_2\|_{L^\infty}),\quad
\omega(t) =1+ \int_{t-2}^t \tilde F(\tau)d\tau,\quad
B_3(t)=\omega(t)^\frac{\mu_7}{\mu_6} B_2(t)^\frac1{2\mu({\gamma_1}+1)}.
\end{equation*}
Let $d_4=d_2 d_3^{-b}$, where $d_2$ is in \eqref{psubt} and $d_3$ is in \eqref{lamtilde} below.
If
\begin{equation}\label{3limcond}
\lim_{t\to\infty} B_3(t)^\frac{2(\gamma_1+1)}{\gamma_1} e^{-d_4 \bar\Upsilon_3^{-\frac{2b}{1-a}}\int_2^t \omega(\tau)^{-b}d\tau}=0,\quad
\lim_{t\to\infty} (\omega^b(t))'=\lim_{t\to\infty} \frac{\omega^b(t) F(t)}{\int_0^t F(\tau)d\tau}=0 ,
\end{equation}
then
\begin{equation}\label{limsup70}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
\le C \Upsilon_4
\limsup_{t\to\infty} \Big\{ B_3(t) \Big( \omega(t)^b\norm{\Psi(t)}_{L^\infty} + \big[ \omega(t)^b\norm{\Psi(t)}_{L^\infty} \big]^\frac{\gamma_1}{\gamma_1+1}\Big) \Big\},
\end{equation}
where $\Upsilon_4$ is defined by \eqref{c1def} below.
\end{theorem}
\begin{proof}
Since $\bar A<\infty$, we have $\bar\Upsilon_3 <\infty$.
Then $$\bar A\le C (\bar\Upsilon_3^2+\bar\Upsilon_3^\frac{2-a}{1-a})\le C \bar\Upsilon_3^\frac{2-a}{1-a}<\infty .$$
By \eqref{boundedlarge}, there is $T_1>0$ such that for $\tau>T_1$, we have
\begin{equation}
\Lambda(\tau)\le C\Big(\bar\Upsilon_3^\frac2{1-a}+\int_{\tau-1}^\tau \tilde F(\tau)d\tau\Big),
\end{equation*}
hence, for $t>T_1+1$,
\begin{equation}\label{lamtilde}
\sup_{\tau\in[t-1,t]}\Lambda(\tau)\le C\Big(\bar\Upsilon_3^\frac2{1-a}+\int_{t-2}^t \tilde F(\tau)d\tau\Big)
\le d_3 \bar\Upsilon_3^\frac2{1-a} \omega(t),\quad \tilde \lambda(t)\le C \bar\Upsilon_3^\frac{2\mu_7}{1-a} \omega(t)^{\mu_7},
\end{equation}
where $d_3$ is a positive constant.
Note also that $\norm{\Psi(t)}_{L^\infty}\le \bar\Upsilon_3$ for all $t\ge 0$.
Then by \eqref{Clam}, \eqref{lamtilde} and \eqref{newgamtil}, we have
\begin{equation}\label{Ctilbound}
\tilde {\mathcal C}(t)
\le C \bar\Upsilon_3^\frac{2\mu_7}{(1-a)\mu_6} \omega(t)^\frac{\mu_7}{\mu_6} \Big( 1 + L_{12}^\frac1{{\gamma_1}+1} \bar\Upsilon_3^\frac{\mu_4(\gamma_2+a-2)}{2\mu({\gamma_1}+1)} B_2(t)^\frac1{2\mu({\gamma_1}+1)} + \bar\Upsilon_3^\frac1{\mu_6+1}\Big)\\
\end{equation}
Thus,
\begin{equation}\label{ctil2}
\tilde {\mathcal C}(t)\le C \eta_1 \omega(t)^\frac{\mu_7}{\mu_6} B_2(t)^\frac1{2\mu({\gamma_1}+1)}=C \eta_1 B_3(t) ,
\end{equation}
where $\eta_1=L_{12}^\frac1{{\gamma_1}+1} \bar\Upsilon_3^\frac{2\mu_7}{(1-a)\mu_6}\big[ \bar\Upsilon_3^\frac{\mu_4(\gamma_2+a-2)}{2\mu({\gamma_1}+1)} + \bar\Upsilon_3^\frac1{\mu_6+1}\big]$.
Using \eqref{lamtilde} and \eqref{ctil2}, we apply Lemma \ref{limPprop} with $g(t)=\bar\Upsilon_3^\frac{2}{1-a}\omega(t)$
and $B(t) = \eta_1 B_3(t)$. Note that the last two limits in \eqref{3limcond} imply the second limit in \eqref{condlim1}.
As a result, we obtain from \eqref{limsup00} that
\begin{align*}
\limsup_{t\to\infty} \|\bar P(\cdot,t)\|_{L^\infty}
&\le C \limsup_{t\to\infty}\Big\{ \eta_1 B_3(t)
\Big[ \Big( [\bar\Upsilon_3^\frac2{1-a}\omega(t)]^{b} \norm{\Psi(t)}_{L^\infty} \Big)^{\varrho_1} + [\bar\Upsilon_3^\frac2{1-a}\omega(t)]^{b} \norm{\Psi(t)}_{L^\infty} \Big]
\Big\} .
\end{align*}
Hence \eqref{limsup70} follows with
\begin{equation}\label{c1def}
\Upsilon_4= \bar\Upsilon_3^\frac{2b}{1-a} \eta_1=L_{12}^\frac1{{\gamma_1}+1} \bar\Upsilon_3^{\frac{2\mu_7}{(1-a)\mu_6}+\frac{2b}{1-a}}\big[ \bar\Upsilon_3^\frac{\mu_4(\gamma_2+a-2)}{2\mu({\gamma_1}+1)} + \bar\Upsilon_3^\frac1{\mu_6+1}\big].
\end{equation}
The proof is complete.
\end{proof}
Next, we will treat the case $\bar A=\infty$. For that we recall some estimates from \cite{HI2}.
\begin{lemma}[cf. \cite{HI2}, Lemma 5.8]\label{lamda-pt-estimate}
Assume the Degree Condition and $\bar A=\infty$. Define
\begin{align*}
W_1(t) &= 1+M_F(t)^\frac2{2-a}+\int_{t-1}^t \tilde F(\tau)d\tau,\\
W_2(t) &=1+\bar \beta^\frac1{1-a}+F(t-1)^\frac2{2-a} +F(t)+\int_{t-1}^t F(\tau)+\tilde F(\tau)d\tau.
\end{align*}
{\rm (i)} There is $T_1>0$ such that $\Lambda(t)\le CW_1(t)$ for all $t>T_1$.
{\rm (ii)} If $\bar \beta<\infty$ then there is $T_2>0$ such that $\Lambda(t)\le W_2(t)$ for all $t>T_2$.
\end{lemma}
Let $W(t)$ be defined, in the general case, by
\begin{equation}
W(t)= 1+M_F(t)^\frac2{2-a}+\int_{t-2}^t \tilde F(\tau)d\tau,
\end{equation*}
and, in case $\bar \beta<\infty$, by
\begin{equation}
W(t)=1+\bar \beta^\frac1{1-a}+\sup_{[t-2,t]}F(t)^\frac2{2-a} +\int_{t-2}^t \tilde F(\tau)d\tau.
\end{equation*}
Then for large $t$, we have from Lemma \ref{lamda-pt-estimate} that
\begin{equation}\label{lamW}
\sup_{\tau\in[t-1,t]}\Lambda(\tau)\le d_5 W(t)
\quad\text{and}\quad
\tilde \lambda(t)\le CW(t)^{\mu_7},
\end{equation}
where $d_5$ is a positive constant.
With \eqref{lamW}, we restate Lemma \ref{limPprop} as the following.
\begin{lemma}\label{propAbar}
Assume \eqref{strictdeg}. Suppose $\bar A=\infty$ and $\tilde {\mathcal C}(t)\le C \tilde B(t)$ for sufficient large $t$, with some function $\tilde B(t)\ge 1$.
Let $d_6=d_2 d_5^{-b}$, where $d_2$ is in \eqref{psubt} and $d_5$ is in \eqref{lamW}.
If
\begin{equation}\label{condlim15}
\lim_{t\to\infty} \tilde B(t)^\frac{2({\gamma_1}+1)}{{\gamma_1}} e^{-d_6\int_{2}^t W(\tau)^{-b}d\tau} =0
\quad\text{and}\quad
\lim_{t\to\infty} W(t)^b\frac {\tilde B'(t)}{\tilde B(t)} =0,
\end{equation}
then
\begin{equation}\label{limsup50}
\begin{aligned}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
&\le C \limsup_{t\to\infty}\Big\{ \tilde B(t) \Big( W(t)^{b} \norm{\Psi(t)}_{L^\infty}
+ \big[ W(t)^{b} \norm{\Psi(t)}_{L^\infty}\big]^\frac{\gamma_1}{{\gamma_1}+1}\Big)\Big\} .
\end{aligned}
\end{equation}
\end{lemma}
Denote
\begin{equation}
\mu_{13}=\max\Big\{ \frac{\mu_7}{\mu_6} ,\frac{\mu_7}{\mu_6+1}+\frac{1-a}{2(\mu_6+1)}\Big \}\quad \text{and}\quad
\gamma_7=\frac{\mu_4(\gamma_2+a-2)}{2\mu}.
\end{equation}
\begin{theorem}\label{newthmAbar2}
Assume \eqref{strictdeg}. Suppose $\bar A=\infty$ and $\int_0^\infty F(\tau)d\tau=\infty$.
Define
\begin{equation*}
B_4(t)=\int_0^t F(\tau)d\tau
\quad\text{and}\quad
B_5(t)=M_F(t)^{\gamma_7} B_4(t)^\frac1{2\mu}.
\end{equation*}
Let $d_6$ be defined as in Lemma \ref{propAbar}.
If
\begin{equation}\label{newexpCond}
\lim_{t\to\infty} W(t)^{\frac {2\mu_{13}({\gamma_1}+1)}{{\gamma_1}}} B_5(t)^{\frac 2{{\gamma_1}}} e^{-d_6\int_2^{t} W^{-b}(\tau)d\tau}=0,
\end{equation}
\begin{equation}\label{newlemCond}
\lim_{t\to\infty} (W^b(t))' =0,\quad \lim_{t\to\infty} W^b(t)\frac{M_F'(t)} {M_F(t)} =0 \quad\text{and}\quad \lim_{t\to\infty} W^b(t)\frac{F(t)} {\int_0^t F(\tau)d\tau} =0,
\end{equation}
then
\begin{equation}\label{newlimsupP}
\begin{aligned}
&\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}\\
&\le C \limsup_{t\to\infty} \Big\{ W(t)^{\mu_{13}} B_5(t)^\frac{1}{{\gamma_1}+1} \Big( W(t)^{b} \norm{\Psi(t)}_{L^\infty}
+ \big[ W(t)^{b} \norm{\Psi(t)}_{L^\infty} \big]^\frac{{\gamma_1}}{{\gamma_1}+1}\Big) \Big\}.
\end{aligned}
\end{equation}
\end{theorem}
\begin{proof}
Let
$B_6(t)=1+\sup_{[0,t]}(\|\psi_1\|_{L^\infty}+\|\psi_2\|_{L^\infty}).$
Then $\lim_{t\to\infty} B_6(t)=\infty$ and $\lim_{t\to\infty} B_4(t)=\infty$.
Note that $B_6(t)\le (1+M_F(t))^\frac{1-a}{2-a}$, and in both cases of definition of $W(t)$, we have
\begin{equation}\label{maxpsi}
\sup_{[t-1,t]}\|\Psi\|_{L^\infty}\le C W(t)^{\frac{1-a}{2-a}\frac{2-a}{2}}=CW(t)^\frac{1-a}2.
\end{equation}
We estimate $\tilde {\mathcal C}(t)$. By \eqref{pregam} and \eqref{gpr3}, we have for $t$ sufficiently large that
\begin{align}\label{newgamtil2}
\tilde \vartheta(t) &\le C M_F(t)^{\frac{\mu_4(\gamma_2+a-2)}{2\mu}}B_4(t)^\frac1{2\mu}
= C B_5(t).
\end{align}
By \eqref{Clam}, \eqref{lamW}, \eqref{newgamtil2} and \eqref{maxpsi},
\begin{align*}
\tilde {\mathcal C}(t)
&\le C \Big\{ W(t)^\frac{\mu_7}{\mu_6} + B_5(t)^\frac1{{\gamma_1}+1} W(t)^\frac{\mu_7}{{\gamma_1}+1} + W(t)^\frac{\mu_7}{\mu_6+1} W(t)^\frac{1-a}{2(\mu_6+1)}\Big\}
\le C B_7(t),
\end{align*}
where $B_7(t)= W(t)^{\mu_{13}}B_5(t)^\frac{1}{{\gamma_1}+1}$.
We apply Lemma \ref{propAbar} with $\tilde B(t)=B_7(t)$.
The first condition in \eqref{condlim15} is replaced by
\begin{equation}
\lim_{t\to\infty} ( W(t)^{\mu_{13}}B_5(t)^\frac{1}{{\gamma_1}+1} )^{\frac 2{\varrho_1}} e^{-d_6\int_2^{t} W^{-b}(\tau)d\tau}=0,
\end{equation*}
which is \eqref{newexpCond}. The second condition in \eqref{condlim15} is replaced by \eqref{newlemCond}.
Then we obtain \eqref{newlimsupP} directly from \eqref{limsup50}.
\end{proof}
\begin{example}
Suppose that for $t$ sufficiently large, we have
\begin{equation}
F(t)\le M_F(t)\le Ct^{\theta_1},\quad W(t)\le Ct^{\theta_2/b},
\end{equation*}
for some $\theta_1>0$ and $0<\theta_2<1$.
Following the proof of Theorem \ref{newthmAbar2}, we can see that the statements still hold true if the functions $F(t)$, $M_F(t), W(t)$ are replaced by their upper bounds $Ct^{\theta_1}$, $Ct^{\theta_1}$, $Ct^{\theta_2/b}$, respectively.
With such replacements, conditions \eqref{newexpCond} and \eqref{newlemCond} are met, and $B_5(t) =C t^{\theta_3}$, where $\theta_3=\frac{\mu_4(\gamma_2+a-2)\theta_1+\theta_1+1}{2\mu}$.
Therefore, from \eqref{newlimsupP}, it follows
\begin{equation}
\limsup_{t\to\infty} \|\bar P(\cdot,t)\|_{L^\infty(U')}
\le C \limsup_{t\to\infty} \Big( t^{\theta_5} \norm{\Psi(t)}_{L^\infty} + [t^{\theta_4 } \norm{\Psi(t)}_{L^\infty}]^\frac{\gamma_1}{{\gamma_1}+1}\Big),
\end{equation*}
where
$\theta_4=\frac{\mu_4(\gamma_2+a-2)\theta_1+\theta_1+1}{2\mu{\gamma_1}}+\frac{\theta_2\mu_{13}({\gamma_1}+1) }{b{\gamma_1}} +\theta_2 $
and
$\theta_5= \frac{\mu_4(\gamma_2+a-2)\theta_1+\theta_1+1}{2\mu({\gamma_1}+1)}+\frac{\theta_2\mu_{13} }b +\theta_2 $.
\end{example}
\subsection{Results for pressure gradient}
\label{sec42}
In \cite{HI2}, the norm $\norm{ \nabla P(t) }_{L^{2-a}}$ is estimated for all $t>0$. Now we estimate
$\norm{\nabla P(t)}_{L^{s}(U')}$ and $\|\nabla P\|_{L^{s}(U'\times(0,T))}$ for any $s\in(2-a,2)$.
\begin{proposition}\label{PropGrad}
Let $\delta\in (0,a)$.
Then for all $t>0$,
\begin{multline}\label{DG1}
\|\nabla P(t)\|_{L^{2-\delta}(U')}
\le C \|\bar P(t)\|_{L^2}^\frac12 \Big\{1+\sum_{i=1,2}\Big[ \| \bar p_{it}(t)\|_{L^2}^2 + \| \nabla p_i(t)\|_{L^{2-a}}^{2-a}
+\|\psi_i(t)\|_{L^\infty}^2 \Big]\Big\}^\frac14\\
\cdot \Big( 1 + \sum_{i=1,2} \int_{U'} |\nabla p_i(x,t)|^\frac{a(2-\delta)}{\delta}dx \Big)^\frac{\delta}{2(2-\delta)},
\end{multline}
and for any $T> 0$,
\begin{multline}\label{InGradP}
\|\nabla P\|_{L^{2-\delta}(U'\times(0,T))}
\le C \|\bar P\|_{L^2(U\times(0,T))}^{1/2} \\
\cdot \Big[T+ \sum_{i=1,2} \Big(\norm{\bar p_{it}}_{L^2(U\times(0,T))}^2 +\int_0^T \int_U |\nabla p_{i}|^{2-a}dxdt+\int_0^T \norm{\psi_i(t)}_{L^\infty}^2dt\Big) \Big]^{\frac {1}4}\\
\cdot\Big[ T+\sum_{i=1,2}\int_{0}^T\int_{U'} |\nabla p_i|^\frac{a(2-\delta)}{\delta}dx dt\Big]^\frac{\delta}{2(2-\delta)}.
\end{multline}
Here, constant $C>0$ depends on $U$, $U'$ and $\delta$.
\end{proposition}
\begin{proof}
Note that $\nabla \bar P =\nabla P$.
Let $\zeta=\zeta(x)$ be a cut-off function such that $\zeta$ vanishes in neighborhood of $\Gamma$.
Multiplying equation \eqref{eq-1} by $\bar P\zeta^2$ and integrating over $U$, using integration by parts, we have
\begin{align*}
\int_U \bar P_t \bar P \zeta^2 dx
&= -\int_U \Big(K(\nabla p_1|)\nabla p_1-K(|\nabla p_2|)\nabla p_2 \Big ) \cdot (\nabla P \zeta^2+2\bar P \zeta\nabla \zeta) dx \\
&\quad +\frac1{|U|}\int_\Gamma \Psi(x,t)d\sigma \int_U\bar P\zeta^2 dx.
\end{align*}
Let $\xi(x,t)=|\nabla p_1|\vee |\nabla p_2|$. By the monotonicity \eqref{mono1} in Lemma \ref{quasimono-lem}, we obtain
\begin{equation}\label{pregradP}
\begin{aligned}
\int_U \bar P_t \bar P \zeta^2 dx
&\le -(1-a)\int_U K(\xi)|\nabla P\zeta|^2 dx + 2\int_U ( |\nabla p_1|+ |\nabla p_2|)^{1-a}|P| \zeta|\nabla\zeta| dx\\
&\quad + \norm{\Psi(t)}_{L^\infty}\int_U |\bar P| \zeta^2 dx.
\end{aligned}
\end{equation}
Let $V\Subset U$ such that $U'\Subset V$. We select $\zeta$ such that $\zeta\equiv 1$ on $U'$ and supp $\zeta\subset V$.
We obtain from \eqref{pregradP} that
\begin{align*}
&(1-a)\int_U K(\xi)|\nabla P\zeta|^2 dx
\le \int_{U_1} |\bar P_t| |\bar P| dx +C\int_{V} ( |\nabla p_1|+ |\nabla p_2|)^{1-a}|\bar P|dx+ C \norm{\Psi(t)}_{L^\infty} \norm{\bar P}_{L^2}\\
&\le C (\sum_{i=1,2} \norm{\bar p_{it}}_{L^2} )\norm{\bar P}_{L^2}+C\Big(\sum_{i=1,2} \int_U |\nabla p_i|^{2-2a}dx\Big)^{1/2} \norm{\bar P}_{L^2} +C \sum_{i=1,2} \norm{\psi_i(t)}_{L^\infty} \norm{\bar P}_{L^2}.
\end{align*}
Hence
\begin{equation}\label{sub1}
\int_{U'} K(\xi)|\nabla P|^2 dx
\le C \norm{\bar P}_{L^2} \Big[ \sum_{i=1,2} \norm{\bar p_{it}}_{L^2} +\Big(\int_U (1+\sum_{i=1,2}|\nabla p_i|)^{2-a} dx\Big)^{1/2}
+\sum_{i=1,2}\norm{\psi_i(t)}_{L^\infty}\Big].
\end{equation}
By H\"older's inequality and property \eqref{Kesta}, we have
\begin{equation}\label{DG2}
\int_{U'} |\nabla P|^{2-\delta} dx
\le C \Big ( \int_{U'} K(\xi)|\nabla P|^2 dx \Big )^\frac{2-\delta}{2} \Big( \int_{U'} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx \Big)^\frac{\delta}{2}.
\end{equation}
Combining \eqref{DG2} with \eqref{sub1} yields
\begin{multline*
\| \nabla P(t)\|_{L^{2-\delta}(U')} \le C \|\bar P(t)\|_{L^2}^\frac12 \Big\{1 + \sum_{i=1,2}\Big[ \| \bar p_{it}(t)\|_{L^2}^2 + \int_U |\nabla p_i(x,t)|^{2-a} dx +\|\psi_i(t)\|_{L^\infty}^2 \Big] \Big\}^\frac14\\
\cdot \Big( \int_{U'} (1+|\nabla p_1(x,t)|+|\nabla p_2(x,t)|)^\frac{a(2-\delta)}{\delta}dx \Big)^\frac{\delta}{2(2-\delta)},
\end{multline*}
then we obtain \eqref{DG1}.
We now prove \eqref{InGradP}.
Integrating \eqref{sub1} in $t$ over $[0,T]$ and applying H\"older's inequality yield
\begin{align*}
&\int_0^T \int_{U'} K(\xi)|\nabla P|^2 dx
\le C \Big( \int_0^T \norm{\bar P(t)}_{L^2}^2 dt \Big)^{1/2}
\cdot\Big( \int_0^T \norm{\bar p_{1t}(t)}_{L^2}^2 + \norm{\bar p_{2t}(t)}_{L^2}^2 dt \\
&\quad + \int_0^T \int_U (|\nabla p_1|+|\nabla p_2|)^{2-2a} dx dt +\int_0^T \norm{\Psi(t)}_{L^\infty}^2dt \Big)^{1/2}.
\end{align*}
Then
\begin{equation}\label{sub10}
\begin{aligned}
&\int_0^T \int_{U'} K(\xi)|\nabla P|^2 dx
\le C \|\bar P\|_{L^2(U\times(0,T))} \\
&\cdot \Big[ \sum_{i=1,2} \Big(\norm{\bar p_{it}}_{L^2(U\times(0,T))}^2 +\int_0^T \int_U (1+|\nabla p_{i}|^{2-a})dxdt+\int_0^T \norm{\psi_i(t)}_{L^\infty}^2dt\Big) \Big]^{\frac 12}.
\end{aligned}
\end{equation}
Again, by H\"older's inequality and property \eqref{Kesta}, we have
\begin{equation}\label{GradPclose2}
\int_{0}^T \int_{U'} |\nabla P|^{2-\delta} dx dt
\le C \Big ( \int_{0}^T \int_{U'} K(\xi)|\nabla P|^2 dxdt \Big )^\frac{2-\delta}{2} \Big( \int_{0}^T\int_{U'} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx dt\Big)^\frac{\delta}{2}.
\end{equation}
Using \eqref{sub10} in \eqref{GradPclose2} and taking the power $1/(2-\delta)$, we obtain \eqref{InGradP}.
\end{proof}
We now have explicit estimates in terms of initial and boundary data.
\begin{theorem}\label{GradThm1}
Assume \eqref{strictdeg}.
For $\delta\in(0,a)$, $0<t_0<1$ and $T>t_0$ we have
\begin{equation}\label{DG5}
\sup_{[t_0,T]} \|\nabla P(t)\|_{L^{2-\delta}(U')} \le C M_{3,t_0,T}(\|\bar P(0)\|_{L^2} +\sup_{[0,T]}\|\Psi\|_{L^\infty})^\frac12,
\end{equation}
where $M_{3,t_0,T}$ is defined in \eqref{N8def} below.
\end{theorem}
\begin{proof}
We use estimate \eqref{DG1}. We bound $\|\bar P(t)\|_{L^2}$ by \eqref{pLip}:
\begin{equation}\label{pLip2}
\sup_{[0,T]}\norm{\bar P(t)}_{L^2}^2 \le C\cdot M_{1,T}\Big(\norm{\bar P(0)}_{L^2}^2
+ \sup_{[0,T]}\norm{\Psi(t)}_{L^\infty}^2 \Big),
\end{equation}
where $M_{1,T}=A_0+T+\int_0^T\big[ f_1(\tau)+f_2(\tau)\big] d\tau$.
We estimate time derivative by \eqref{Jpt-boundA0}, estimate $\int_U |\nabla p_i|^{2-a} dx$ by \eqref{H-bound-0}, then we have
\begin{equation}\label{pit}
\begin{aligned}
& \sum_{i=1,2}\Big[ \| \bar p_{it}(t)\|_{L^2}^2 + \int_U |\nabla p_i(x,t)|^{2-a} dx
+\|\psi_i(t)\|_{L^\infty}^2 \Big]\\
&\le C\, t_0^{-1}\Big(\sum_{i=1,2} L_5(t_0;[p_i(0),\psi_i]) \Big) + (T+1)\bar\mathcal K_{1,T}^\frac{2-a}{1-a}+\bar\mathcal K_{2,T}\Big).
\end{aligned}
\end{equation}
Recalling $\nu_1=\max\{2,\frac{a(2-\delta)}{\delta}\}$, we apply \eqref{grad20pw} with $s=\nu_1$ and obtain
\begin{multline}\label{gradalot}
\sum_{i=1,2} \int_{U'} |\nabla p_i(x,t)|^\frac{a(2-\delta)}{\delta}dx
\le \sum_{i=1,2} \int_{U'} 1+ |\nabla p_i(x,t)|^{\nu_1}dx \\
\le C\Big[\sum_{i=1,2} L_2(\nu_1;[p_i(0)])\Big] (T+1)^{\frac{2(\nu_1-2)}{2-a}+1} \bar\mathcal K_{1,T}^{\frac{\nu_1-a}{1-a}}.
\end{multline}
Combining \eqref{pLip2}, \eqref{pit} and \eqref{gradalot} with \eqref{DG1}, we have \eqref{DG5} with
\begin{equation}\label{N8def}
M_{3,t_0,T}
=(1+t_0^{-\frac14}) M_{1,T}^\frac14 \Big[\sum_{i=1,2} L_5(t_0;[p_i(0),\psi_i]) \Big]^\frac14
\Big[\sum_{i=1,2} L_2\big(\nu_1;[p_i(0)]\big)\Big]^\frac{\delta}{2(2-\delta)}
(T+1)^{\gamma_8}
\bar\mathcal K_{1,T}^{\gamma_9}
\bar\mathcal K_{2,T}^\frac14,
\end{equation}
where
$\gamma_8= \frac14+\frac{\delta}{2(2-\delta)}\big(\frac{2(\nu_1-2)}{2-a}+1\big)$ and
$\gamma_9= \frac{2-a}{4(1-a)}+\frac{\delta}{2(2-\delta)} \big(\frac{\nu_1-a}{1-a}\big).$
The proof is complete.
\end{proof}
As $t\to\infty$, we have the following asymptotic estimate.
\begin{theorem}\label{GradThm2}
Assume \eqref{strictdeg}. Let
\begin{equation*}\bar\Upsilon_1 = 1+\sup_{[0,\infty)}F,\ \bar\Upsilon_2 = 1+\int_0^\infty F(t)dt
\quad \text{and} \quad
\bar\mathcal A_2 = 1+\bar A^\frac2{2-a}+\limsup_{t\to\infty} \int_{t-1}^t \tilde F(\tau)d\tau.
\end{equation*}
If $\bar\Upsilon_1$, $\bar\Upsilon_2$ and $\bar\mathcal A_2$ are finite then
\begin{equation}\label{DG6}
\limsup_{t\to\infty} \|\nabla P(t)\|_{L^{2-\delta}(U')} \le C \Upsilon_5 \limsup_{t\to\infty}\|\Psi\|_{L^\infty}^{1/2},
\end{equation}
where $\Upsilon_5$ is defined by \eqref{U7} below.
\end{theorem}
\begin{proof}
Taking limsup of \eqref{DG1} and making use the limits \eqref{psub-bounded}, \eqref{limsupH1} and estimate \eqref{Kgrad60pw} give
\begin{multline*}
\limsup_{t\to\infty} \|\nabla P(t)\|_{L^{2-\delta}(U')} \le C \limsup_{t\to\infty}\|\Psi\|_{L^\infty}^{1/2}
\bar\mathcal A_2^\frac b 2 \bar\mathcal A_2^\frac14 \Big\{ \sum_{i=1,2} L_4(\nu_1;[p_i(0)]) \bar\Upsilon_1^{\mu_4(\nu_1-2)} \bar\Upsilon_2 \Big\}^\frac\delta{2(2-\delta)}.
\end{multline*}
Therefore, we obtain \eqref{DG6} with
\begin{equation}\label{U7}
\Upsilon_5= \bar\mathcal A_2^{\frac b 2 +\frac14}
\Big\{ \bar\Upsilon_1^{\mu_4(\nu_1-2)}\bar\Upsilon_2 \sum_{i=1,2} L_4(\nu_1;[p_i(0)]) \Big\}^\frac\delta{2(2-\delta)}.
\end{equation}
\end{proof}
\begin{remark}\label{rmk414}
Even though the limit estimate in \eqref{DG6} still depends on the initial data presented in $\Upsilon_5$, the smallness of the estimate can be controlled by the difference $\Psi(t)$ for large $t$.
\end{remark}
The estimate in Theorem \ref{GradThm1} blows up when $t_0\to 0$.
To overcome this, we consider the Lebesgue norm in both $x$ and $t$.
\begin{theorem}\label{lem412}
Assume \eqref{strictdeg}. Let $\delta\in(0,a)$. For any $ T> 0$, we have
\begin{equation}\label{InGradPt0}
\|\nabla P\|_{L^{2-\delta}(U'\times(0,T))} \le C M_{4,T} \Big( \norm{\bar P(0)}_{L^2}^2+\int_0^T \norm{\Psi(t)}_{L^\infty}^2 dt \Big)^{\frac {1} 4},
\end{equation}
where $ M_{4,T}$ is defined in \eqref{hatN} below.
\end{theorem}
\begin{proof}
Applying \eqref{grad39} to $s=\nu_1\ge 2$ we have
\begin{multline}\label{subs0}
T+\int_0^T\int_{U'} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx dt
\le T+\int_0^T\int_{U'} (1+|\nabla p_1|+|\nabla p_2|)^{\nu_1}dx dt \\
\le C \Big[ \sum_{i=1,2} L_1(\nu_1+a,[p_i(0)]) \Big]
(T+1)^{\frac {2\nu_1}{2-a} - 1} \bar\mathcal K_{1,T}^{\frac{\nu_1}{1-a}}.
\end{multline}
Inequality \eqref{H-bound-0} provides
\begin{equation*}
\sum_{i=1}^2 \norm{\bar p_{it}}_{L^2(U\times(0,T))}^2 \le C(1+A_0+B_0) (T+1) \bar\mathcal K_{1,T}^\frac{2-a}{1-a} \bar\mathcal K_{2,T}.
\end{equation*}
Using \eqref{p-bar-bound1} we see that
\begin{equation*}
T+\sum_{i=1}^2\int_0^T\int_U |\nabla p_i|^{2-a} dxdt \le C(1+A_0) (T+1) \bar\mathcal K_{1,T}^\frac{2-a}{1-a}.
\end{equation*}
Hence,
\begin{equation}\label{subs1}
T+ \sum_{i=1}^2\Big\{ \norm{\bar p_{it}}_{L^2(U\times(0,T))}^2+\int_0^T\int_U |\nabla p_i(x,t)|^{2-2a} dx dt +\int_0^T \norm{\psi_i(t)}_{L^\infty}^2 dt \Big\}
\le C N_{1,T},
\end{equation}
where $N_{1,T}=(1+A_0+B_0) (T+1) \bar\mathcal K_{1,T}^\frac{2-a}{1-a} \bar\mathcal K_{2,T}.$
According to \eqref{pcont} and \eqref{H-bound-0},
\begin{equation}\label{subs2}
\begin{aligned}
\int_0^T \norm{\bar P(t)}_{L^2}^2 dt
& \le C T \sup_{t\in [0,T]} \Lambda(t)^{b} \Big( \norm{\bar P(0)}_{L^2}^2 + \int_0^T \norm{\Psi(t)}_{L^\infty}^2 dt \Big)\\
&\le C T N_{1,T}^b \Big( \norm{\bar P(0)}_{L^2}^2 +\int_0^T \norm{\Psi(t)}_{L^\infty}^2 dt \Big).
\end{aligned}
\end{equation}
Combining \eqref{InGradP}, \eqref{subs0}, \eqref{subs1} and \eqref{subs2} we obtain
\begin{align*}
&\|\nabla P\|_{L^{2-\delta}(U'\times(0,T))}
\le C T^\frac{1}4 N_{1,T}^\frac{b}{4} \Big( \norm{\bar P(0)}_{L^2}^2 +\int_0^T \norm{\Psi(t)}_{L^\infty}^2 dt \Big)^\frac14 \cdot N_{1,T}^\frac14 \\
&\quad \cdot \Big\{\Big[ \sum_{i=1,2} L_1(\nu_1+a,[p_i(0)]) \Big]
(T+1)^{\frac {2\nu_1}{2-a} - 1} \bar\mathcal K_{1,T}^{\frac{\nu_1}{1-a}}\Big\}^\frac{\delta}{2(2-\delta)}.
\end{align*}
Therefore, we obtain \eqref{InGradPt0} with
\begin{equation}\label{hatN}
M_{4,T} = (1+A_0+B_0)^\frac{b+1}4 \Big[ \sum_{i=1,2} L_1(\nu_1+a,[p_i(0)]) \Big]^\frac{\delta}{2(2-\delta)}
T^\frac14 (T+1)^{\gamma_{10}} \bar\mathcal K_{1,T}^{\gamma_{11}} \bar\mathcal K_{2,T}^\frac{b+1}4,
\end{equation}
where
$\gamma_{10}=\frac{\delta}{2(2-\delta)}\Big(\frac {2\nu_1}{2-a} - 1\Big)+\frac{b+1}4$ and
$\gamma_{11}=\frac{\nu_1}{1-a} \frac{\delta}{2(2-\delta)}+\frac{(b+1)(2-a)}{4(1-a)}$.
\end{proof}
\clearpage
\section{Dependence on the Forchheimer polynomial}\label{polycont}
In this section we study the dependence of solutions of IBVP \eqref{eqgamma} on the coefficients of the Forchheimer polynomial $g(s)$ in \eqref{gsa}.
Let $N\ge 1$, the exponent vector $\vec \alpha=(0,\alpha_1,\ldots,\alpha_N)$ and the boundary data $\psi(x,t)$ be fixed.
Let ${\bf D}$ be a compact subset of $\{\vec a=(a_0,a_1,\ldots,a_N):a_0,a_N>0, a_1,\ldots,a_{N-1}\ge 0\}$.
Set
$\hat \chi({\bf D})=\max\{\chi(\vec a):\vec a\in {\bf D}\}.$
Then $\hat \chi ({\bf D})$ is a number in $[1,\infty)$.
Let $g_1(s)=g(s,\vec a^{(1)})$ and $g_2(s)=g(s,\vec a^{(2)})$ be two functions of class FP($N,\vec \alpha$), where $\vec a^{(1)}$ and $\vec a^{(2)}$ belong to ${\bf D}$. For $k=1,2$, let $p_k=p_k(x,t;\vec a^{(k)})$ be the solution of \eqref{eqorig} and \eqref{BC} with $K=K(\xi,\vec{a})$ and the same boundary flux $\psi$.
Let $P=p_1-p_2$ and $\bar P=P-|U|^{-1}\int_U P dx$.
Then
\begin{equation}\label{Erreq}
\bar P_t = \nabla \cdot (K (|\nabla \bar p_1|,\vec{a}^{(1)})\nabla \bar p_1 ) - \nabla \cdot (K (|\nabla \bar p_2|,\vec{a}^{(2)})\nabla \bar p_2 ) \quad \text {in } U\times (0,\infty)
\end{equation}
As shown in \cite{HI1}, all constants $d_j$, $c_j$, $C_j$ and $C$ appearing in estimates in the previous sections when $\vec a$ varies among the vectors $\vec a^{(1)}$, $\vec a^{(2)}$, $\vec a^{(1)}\vee \vec a^{(2)}$ and $\vec a^{(1)}\wedge \vec a^{(2)}$, can be made dependent of $\hat \chi({\bf D})$, but independent of $\vec a$.
We still denote them by $d_j$, $c_j$, $C_j$ and $C$, respectively, in this section.
Throughout this section, we continue to have $U'\Subset U$.
The flux-related quantities $f(t)$, $\tilde f(t)$, $M_f(t)$ $A$ and $\beta$ are defined in section \ref{supestimate}, from \eqref{b}
to \eqref{Abeta}.
Also, $\mathcal K_{1,t}$ and $\mathcal K_{2,t}$ are defined in \eqref{Nf1} and \eqref{Nf2}, respectively.
Regarding initial data, $A_0$ is defined in \eqref{AB0}.
\subsection{Results for pressure}
\label{sec51}
We will estimate the interior $L^\infty$-norm for $\bar P(x,t)$ in terms of $\bar P(x,0)$ and $|\vec a^{(1)}-\vec a^{(2)}|$.
We start with the following general estimates which are counter parts of Propositions \ref{cont:strictcond} and \ref{newcont}.
\begin{proposition}\label{theo49}
We use the same assumptions and notation as in Proposition \ref{cont:strictcond}.
{\rm (i)} There exists a positive constant $C=C(U,V,U')$ such that
\begin{equation}\label{Pcoe}
\sup_{[T_0+\theta T,T_0+T]}\|\bar P\|_{L^\infty(U')}\le C\widehat{\mathcal C}_{T_0,T,\theta}\Big( \norm{\bar P}_{L^2(U\times (T_0,T_0+T) )}+\norm{\bar P}_{L^2(U\times (T_0,T_0+T) )}^{\frac {\gamma_1}{{\gamma_1}+1}}\Big),
\end{equation}
where
\begin{equation}\label{Ctheta5}
\widehat{\mathcal C}_{T_0,T,\theta}= [\lambda(1+ (\theta T)^{-\frac1 2})]^{\frac 1\mu_6} + [\lambda (T^{1/2}\vartheta_1+\vartheta_2)]^{\frac 1 {{\gamma_1}+1}},
\end{equation}
with $\lambda=\lambda_{T_0,T}$ defined in \eqref{lamdef2}, and
\begin{align}
\label{gam1}
\vartheta_1&=\vartheta_{1,T_0,T} \mathbin{\buildrel \rm def \over \longequals} \Big(\int_{T_0}^{T_0+T} \int_{V} |\nabla \bar p_1|^{2(1-a)\mu}+|\nabla \bar p_2|^{2(1-a)\mu} dxdt\Big)^\frac1{2\mu},\\
\label{gam2}
\vartheta_2 &=\vartheta_{2,T_0,T} \mathbin{\buildrel \rm def \over \longequals} \Big(\int_{T_0}^{T_0+T} \int_{V} |\nabla \bar p_1|^{(2-a)\mu}+|\nabla \bar p_2|^{(2-a)\mu} dxdt\Big)^\frac1{2\mu}.
\end{align}
{\rm (ii)} There exists a positive constant $C=C(U,V,U')$ such that
\begin{equation}\label{Pcoe2}
\sup_{[0,T]}\|\bar P\|_{L^\infty(U')}\le 2\|\bar P(0)\|_{L^\infty} + C \widehat{\mathcal C}_{T}\Big( \norm{\bar P}_{L^2(U\times (0,T) )}+\norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1}{{\gamma_1}+1}}\Big) ,
\end{equation}
where
\begin{equation}\label{Ctheta6}
\widehat{\mathcal C}_{T}= \lambda_T^{\frac 1\mu_6} + [\lambda_T (T^{1/2}\vartheta_{1,T}+\vartheta_{2,T})]^{\frac 1 {{\gamma_1}+1}},
\end{equation}
with $\lambda_T=\lambda_{0,T}$, $\vartheta_{1,T}=\vartheta_{1,0,T}$ and $\vartheta_{2,T}=\vartheta_{2,0,T}$.
\end{proposition}
\begin{proof}
(i) Let $\zeta(x,t)$ be a cut-off function such that it is piecewise linear continuous in $t$, $\zeta(\cdot,0)=0$ and for each $t$, supp $\zeta(\cdot,t)\Subset V$.
Let $\bar P^{(k)}=\max\{\bar P-k,0\}$ and $\chi_k$ be characteristic function of the set $\{ (x,t): \bar P^{(k)}>0\}$.
Let $\xi=|\nabla \bar p_1|\vee |\nabla \bar p_2|$.
Multiplying equation \eqref{Erreq} by $\bar P^{(k)} \zeta^2$ and integrating over $U$, we have
\begin{align*}
&\frac12\frac{d}{dt} \int_U |\bar P^{(k)}\zeta|^2 dx- \int_U |\bar P^{(k)}|^2\zeta \zeta_t dx\\
&= \int_U ( K(|\nabla p_1|, \vec{a}^{(1)})\nabla \bar p_1 - K(|\nabla p_2|, \vec{a}^{(2)})\nabla \bar p_2 )\cdot \nabla\bar P^{(k)}\zeta^2 dx\\
&\quad + \int_U ( K(|\nabla p_1|, \vec{a}^{(1)})\nabla \bar p_1 - K(|\nabla p_2|, \vec{a}^{(2)})\nabla \bar p_2 )\cdot [2P^{(k)}\zeta\nabla\zeta ]dx
= I_1+I_2.
\end{align*}
Let $\varepsilon>0$. For $I_2$, we use \eqref{Kesta} to estimate
\begin{align*}
I_2
&\le C \int_U (|\nabla \bar p_1|^{1-a}+|\nabla \bar p_2|^{1-a})|\bar P^{(k)}|\zeta|\nabla \zeta| dx \\
&\le \varepsilon \int_{S_k} |\bar P^{(k)}\zeta|^2 dx
+C\varepsilon^{-1} \int_U \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{2(1-a)} \chi_k |\nabla \zeta|^{2}dx.
\end{align*}
For $I_1$, similar to estimate \eqref{oriI2} of term $I_2$ in Proposition \ref{cont:strictcond} but using monotonicity \eqref{quasimonotone} in place of \eqref{mono1}, we obtain
\begin{align*}
I_1 \le -(1-a) \int_U K(\xi) |\nabla \bar P^{(k)}\zeta|^2dx + C|\vec a^{(1)}-\vec a^{(2)}| \int_U (|\nabla \bar p_1|^{1-a}+|\nabla \bar p_2|^{1-a})|\nabla \bar P^{(k)}| \zeta^2 dx.
\end{align*}
Similar to the way \eqref{I2} was derived from \eqref{oriI2}, one obtains for $\varepsilon>0$ that
\begin{align*}
I_1
&\le - \frac {1-a} 2 \int_U K(\xi)|\nabla (\bar P^{(k)} \zeta) |^2dx + C\int_U |\bar P^{(k)} \nabla\zeta|^2dx \\
&\quad + C|\vec a^{(1)}-\vec a^{(2)}| \int_U (|\nabla \bar p_1|^{2-a}+|\nabla \bar p_2|^{2-a}) \chi_k \zeta^2 dx.
\end{align*}
Therefore,
\begin{align*}
&\frac12\frac{d}{dt} \int_U |\bar P^{(k)}\zeta|^2 dx+\frac{1-a}{2} \int_U K(\xi) |\nabla (\bar P^{(k)}\zeta)|^2dx\\
&\le \int_U |\bar P^{(k)}|^2\zeta |\zeta_t| dx+ C \int_U K(\xi)|P^{(k)}\nabla \zeta|^2 dx+ \varepsilon \int_{S_k} |\bar P^{(k)}\zeta|^2 dx \\
&\quad +C\varepsilon^{-1} \int_U \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{2(1-a)} \chi_k |\nabla \zeta|^{2}dx
+ C|\vec a^{(1)}-\vec a^{(2)}| \int_U (|\nabla \bar p_1|^{2-a}+|\nabla \bar p_2|^{2-a}) \chi_k \zeta^2 dx.
\end{align*}
Let $J=\sup_{[0,T]} \int_U |\bar P^{(k)}\zeta|^2 dx +\int_0^T \int_U K(\xi)|\nabla (\bar P^{(k)}\zeta)|^2dxdt$.
Integrating the preceding inequality in time, taking supremum in $t$ over $[0,T]$, selecting $\varepsilon=1/(8T)$ and applying H\"older's inequality yield
\begin{align*}
&J \le C\int_0^T \int_U |\bar P^{(k)}|^2\zeta |\zeta_t| dx dt+C\int_0^T\int_U K(\xi)|P^{(k)}\nabla \zeta|^2 dxdt\\
& \quad +CT \Big\{\int_0^T \int_U \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{2(1-a)\mu} |\nabla \zeta|^{2\mu} dx dt \Big\}^{\frac 1\mu} \Big\{ \iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big\}^{1-\frac 1\mu} \\
&\quad + C|\vec a^{(1)}-\vec a^{(2)}| \Big\{ \int_0^T \int_{S_k} \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{(2-a)\mu} \zeta^2 dx dt\Big\}^{\frac 1\mu} \Big\{ \int_0^T\int_U \chi_k \zeta^2 dxdt\Big\}^{1-\frac 1\mu}.
\end{align*}
Note that $\vec a^{(i)}$ and $\vec a^{(i)}$ belong to the compact set ${\bf D}$, hence $ |\vec a^{(1)}-\vec{a}^{(2)}| \le C$. Then
\begin{align*}
J&\le C\int_0^T \int_U |\bar P^{(k)}|^2(|\zeta_t|+|\nabla \zeta| ) dx dt\\
& \quad +C T \Big\{ \int_0^T \int_U \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{2(1-a)\mu} |\nabla \zeta|^{2\mu} dx dt \Big\}^{\frac 1\mu} \Big\{ \iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big\}^{1-\frac 1\mu}\\
& \quad +C \Big\{ \int_0^T \int_{S_k} \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{(2-a)\mu} \zeta^2 dx dt\Big\}^{\frac 1\mu} \Big\{ \iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big\}^{1-\frac 1\mu}.
\end{align*}
Using inequality \eqref{Wemb} in Lemma \ref{ParaSob-3} with $W(x,t)=K(\xi(x,t))$, we have
\begin{align*}
& \norm{P^{(k)}\zeta}_{L^{\mu_5}(Q_T)} \le C \lambda J
\le C\lambda\Big\{ \Big( \int_0^T \int_U |\bar P^{(k)}|^2(|\zeta_t|+|\nabla \zeta| ) dx dt\Big)^{1/2} \\
&\quad +T^{1/2}\Big[\int_0^T \int_{V} \Big(|\nabla \bar p_1|+|\nabla \bar p_2| \Big )^{2(1-a)\mu}|\nabla \zeta|^{2\mu} dx dt\Big]^{\frac 1{2\mu}} \Big\{ \iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big\}^{\frac12-\frac 1{2\mu}} \\
&\quad +\Big( \int_0^T\int_{V} (|\nabla \bar p_1|+|\nabla \bar p_2| )^{(2-a)\mu} dxdt \Big)^{\frac 1{2\mu}} \Big\{ \iint_{Q_T\cap{\rm supp}\zeta} \chi_k dxdt\Big\}^{\frac12-\frac 1{2\mu}}.
\end{align*}
Let $Y_i=\| \bar P^{(k_i)} \|_{L^2(A_i)}$. In the same way as in proof of Proposition \ref{cont:strictcond} we find that
\begin{align*}
Y_{i+1}&\le C\lambda M_0^{-1+\frac 2 {\mu_5}}2^i
\Big\{ (1+(\theta T)^{-1} )^\frac 12 Y_i + T^{1/2}\vartheta_1M_0^{-1+\frac 1\mu}Y_i^{1-\frac 1\mu} + \vartheta_2 M_0^{-1+\frac 1\mu} Y_i^{1-\frac 1\mu} \Big\} Y_i^{1-\frac 2 {\mu_5}}\\
&\le C \lambda2^i
\Big\{ M_0^{-\mu_6}(1+(\theta T)^{-1} )^\frac 12 Y_i^{1+\mu_6} +T^{1/2}\vartheta_1 M_0^{-1-{\gamma_1}} Y_i^{1+{\gamma_1}} +\vartheta_2 M_0^{-1-{\gamma_1}} Y_i^{1+{\gamma_1}} \Big\},
\end{align*}
where the exponents ${\gamma_1},\mu_6$ are defined in Theorem \ref{cont:strictcond}.
Since $Y_0\le \norm{\bar P}_{L^2(U\times (0,T) )} $ we choose $M_0$ sufficiently large such that
\[
\norm{\bar P}_{L^2(U\times (0,T) )}\le C\min\Big\{ ((\lambda (1+(\theta T)^{-1} )^\frac12 )^{-\frac 1\mu_6} M_0 , (\lambda T^{1/2} \vartheta_1)^{-\frac 1 {{\gamma_1}}} M_0^{\frac 1{\gamma_1}+1}, M_0^{\frac 1{\gamma_1}+1} (\lambda \vartheta_2)^{-\frac 1 {{\gamma_1}}} \Big \},
\]
thus
\begin{align*}
M_0\ge C\max \Big\{
& (\lambda (1+(\theta T)^{-1} )^\frac12 )^{\frac 1\mu_6} \norm{\bar P}_{L^2(U\times (0,T) )},
(\lambda T^{1/2} \vartheta_1)^{\frac 1 {{\gamma_1}+1}} \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1}{{\gamma_1}+1}} ,\\
& (\lambda \vartheta_2)^{\frac 1 {{\gamma_1}+1}} \norm{\bar P}_{L^2(U\times (0,T) )}^{\frac{\gamma_1}{{\gamma_1}+1}} \Big\}.
\end{align*}
Using the fact that
$ (\lambda (1+(\theta T)^{-1} )^\frac12)^{\frac 1\mu_6}$, $(\lambda T^{1/2} \vartheta_1)^{\frac 1 {{\gamma_1}+1}}$, $(\lambda \vartheta_2)^{\frac 1 {{\gamma_1}+1}}$ are less than or equal to $C \widehat{\mathcal C}_{T_0,T,\theta}$,
we then choose
\begin{equation*}
M_0 =C \widehat{\mathcal C}_{T_0,T,\theta} \Big( \norm{\bar P}_{L^2(U\times (0,T) )}+\norm{\bar P}_{L^2(U\times (0,T) )}^{\frac {\gamma_1}{{\gamma_1}+1}}\Big).
\end{equation*}
Then applying Lemma~\ref{multiseq} for $m=3$ to sequence $\{Y_i\}$, and using the same argument as in Theorem \ref{cont:strictcond}, we obtain $|\bar p(x,t)|\le M_0$ in $U'\times (\theta T,T)$.
(ii) Using same arguments as for deriving Proposition \ref{newcont} from the proof of Proposition \ref{cont:strictcond}, we can modify the proof in part (i) to obtain \eqref{Pcoe2}.
\end{proof}
We recall some results from \cite{HI2} which will be needed in subsequent developments.
\begin{theorem}[cf. \cite{HI2}, Theorems 5.16 and 5.17]\label{supgradthem}
{\rm (i)} For $0< T < \infty$, we have
\begin{equation}\label{coeffJsup1} \sup_{[0,T]} \norm{\bar P(t)}_{L^2}^2 \le \norm{\bar P(0)}_{L^2}^2 + C M_{5,T} |\vec a^{(1)}-\vec a^{(2)}| ,\end{equation}
where $M_{5,T}=A_0+T+\int_0^T f(\tau)d\tau$.
{\rm (ii)} Assume the Degree Condition.
Suppose $\sup_{[0,\infty)} f(t)<\infty$ and $\sup_{[1,\infty)} \int_{t-1}^t \tilde f(\tau)d\tau<\infty$.
Then
\begin{equation}\label{coeffJsup2}
\sup_{[1,\infty)} \norm{\bar P(t)}_{L^2}^2 \le \norm{\bar P(0)}_{L^2}^2 + C \Upsilon_6^{b+1} |\vec a^{(1)}-\vec a^{(2)}| ,
\end{equation}
\begin{equation}\label{coefflim}
\limsup_{t\to\infty} \norm{\bar P(t)}_{L^2}^2 \le C \mathcal A_2^{b+1} |\vec a^{(1)}-\vec a^{(2)}| ,
\end{equation}
where $\Upsilon_6=1+A_0+\sup_{[0,\infty)} f^\frac2{2-a}(t)+\sup_{[1,\infty)} \int_{t-1}^t \tilde f(\tau)d\tau$,
and $\mathcal A_2$ is defined by \eqref{Ulim2}.
\end{theorem}
We will take advantage of calculations in Theorem \ref{thm46}. However the exponents will be changed.
The new counterparts of exponents $\gamma_2,\gamma_3,\ldots,\gamma_6$ in \eqref{mu34}--\eqref{mu89} are
\begin{align*}
&\gamma_2'=(2-a)\mu,
&&\gamma_3'=\frac{\gamma_2'}{2\mu(1-a)}=\frac{2-a}{2(1-a)}\ge 1,\\
&\gamma_4'=\frac1{2\mu}\Big(\frac{2\gamma_2'}{2-a}-1\Big)=1-\frac1{2\mu},
&&\gamma_5'=\frac{\mu_7}{\mu_6}+ \frac{2\gamma_4'+1}{2({\gamma_1}+1)},\\
&\gamma_6'=\frac{\mu_7(2-a)}{\mu_6(1-a)}+\frac{\gamma_3'}{{\gamma_1}+1}.
\end{align*}
Firstly, we replace $\gamma_2$ by $\gamma_2'$ in estimate \eqref{gradInt20} to obtain
\begin{equation}\label{gradInt2}
\vartheta_{2,T}\le C (T+1)^{\gamma_4'}\mathcal K_{1,T}^{\gamma_3'}=C \ell_2'(T+1)^{\gamma_4'}\mathcal K_{1,T}^{\gamma_3'},
\end{equation}
where $\ell_2'=[\sum_{i=1,2} L_1(\gamma_2'+a;[p_i(0)])]^\frac1{2\mu}$.
Secondly, since $\gamma_4'\ge 1/(2\mu)$, it follows
\begin{equation}\label{upUp12}
\vartheta_{1,T}\le CT^\frac1{2\mu}+\vartheta_{2,T}\le C \ell_2'(T+1)^{\gamma_4'}\mathcal K_{1,T}^{\gamma_3'}.
\end{equation}
Thirdly, same as \eqref{lamd}, we have
\begin{equation}\label{lamUp2}
\lambda_T\le C \ell_1' (T+1)^{\mu_7} \mathcal K_{1,T}^\frac{\mu_7(2-a)}{1-a} \mathcal K_{2,T}^{\mu_7},
\end{equation}
where $\ell_1'=1+A_0+\norm{\nabla p_1(0)}_{L^{2-a}}^{2-a}+\norm{\nabla p_2(0)}_{L^{2-a}}^{2-a}$.
For finite time intervals, we have the following estimates.
\begin{theorem}\label{thm54}
Assume \eqref{strictdeg}.
{\rm (i)} For $T\in(0, 1]$, we have
\begin{equation}\label{P-inter-t<1}
\sup_{[0,T]} \norm{\bar P}_{L^\infty(U')} \le 2\|\bar P(0)\|_{L^\infty} +C M_{6,T}\big ( \mathcal A+\mathcal A^{\frac {\gamma_1}{{\gamma_1}+1}}\big),
\end{equation}
where $M_{6,T}$ is defined in \eqref{N5def} below and
$\mathcal A = \norm{\bar P(0)}_{L^2} +|\vec a^{(1)}-\vec a^{(2)}|^{1/2}$.
{\rm (ii)} For $T>1$, we have
\begin{equation}\label{contfinite40}
\begin{aligned}
\sup_{[1,T]}\|\bar P\|_{L^\infty(U')}
&\le C M_{7,T} \Big (\mathcal A+\mathcal A^{\frac {\gamma_1}{{\gamma_1}+1}}\Big),
\end{aligned}
\end{equation}
where $M_{7,T}$ is defined in \eqref{N6def} below.
\end{theorem}
\begin{proof}
(i)
Let $T\in(0,1]$.
Applying \eqref{Pcoe2} of Proposition \ref{theo49} gives
\begin{align*}
\|\bar P(t)\|_{L^\infty(U')}
&\le 2\|\bar P(0)\|_{L^\infty}+C \widehat{\mathcal C}_{T} \Big\{ T^{1/2}\sup_{[0,t]} \norm{\bar P}_{L^2} + (T^{1/2}\sup_{[0,t]} \norm{\bar P}_{L^2} )^{\frac {\gamma_1}{{\gamma_1}+1} }\Big\}\\
&\le 2\|\bar P(0)\|_{L^\infty}+C \widehat{\mathcal C}_{T} T^{\frac {\gamma_1}{2({\gamma_1}+1)} } \Big\{ 1+\sup_{[0,t]} \norm{\bar P}_{L^2}\Big\}.
\end{align*}
Using \eqref{gradInt2}, \eqref{upUp12} and \eqref{lamUp2} to estimate corresponding terms in \eqref{Ctheta6}, we have
\begin{align*}
\widehat{\mathcal C}_{T}\le C \lambda_T^\frac{1}{\mu_6} (\vartheta_{1,T}+\vartheta_{2,T})^\frac1{{\gamma_1}+1}
= C \ell_3' \mathcal K_{1,T}^{\gamma_6'} \mathcal K_{2,T}^\frac{\mu_7}{\mu_6},
\end{align*}
where $\ell_3'=(\ell_1')^\frac1{\mu_6} (\ell_2')^\frac1{{\gamma_1}+1}$.
Moreover, by \eqref{coeffJsup1}
\begin{equation}\label{CalA}
\sup_{[0,T]} \norm{\bar P}_{L^2}
\le C( \norm{\bar P(0)}_{L^2} +|\vec a^{(1)}-\vec a^{(2)}|^\frac12 ) \Big[1+A_0+\int_0^T f(\tau)d\tau\Big]^\frac12
\le C \mathcal A (1+A_0)^\frac12 \mathcal K_{1,T}^\frac{2-a}{2(1-a)}.
\end{equation}
Combining the above, we obtain \eqref{P-inter-t<1} with
\begin{equation}\label{N5def}
M_{6,T}= \ell_3' (1+A_0)^{1/2} T^\frac{{\gamma_1}}{2({\gamma_1}+1)} \mathcal K_{1,T}^{\gamma_6'+\frac{2-a}{2(1-a)}}\mathcal K_{2,T}^\frac{\mu_7}{\mu_6}.
\end{equation}
(ii) Let $T>1$. We apply \eqref{Pcoe} with $T_0=0$ and $\theta T=1$. Note that
\begin{align*}
\widehat{\mathcal C}_{0,T,\theta}
\le C \lambda_T^\frac{\mu_7}{\mu_6} ( 1+ [T^{1/2} (\vartheta_{1,T}+\vartheta_{2,T}) ]^\frac 1{{\gamma_1}+1} )
\le C \ell_3' T^{\gamma_5'}\mathcal K_{1,T}^{\gamma_6'} \mathcal K_{2,T}^\frac{\mu_7}{\mu_6}.
\end{align*}
Then similar calculations to those for proving \eqref{contfinite} in Theorem \ref{thm46} lead to
\eqref{contfinite40} with
\begin{equation}\label{N6def}
M_{7,T}= (1+A_0)^{1/2} \ell_3' T^{\gamma_5'+1} \mathcal K_{1,T}^{\gamma_6'+\frac{2-a}{2(1-a)}} \mathcal K_{2,T}^\frac{\mu_7}{\mu_6}.
\end{equation}
The proof is complete.
\end{proof}
We now consider estimates when $t\to\infty$. We use the same notation as in subsection \ref{sec41}.
For $t\ge 2$, we apply Proposition \ref{theo49}(i) to the interval $[t-1,t]$, that is $T_0=t-1$ and $T=1$, and use $\theta = 1/2$. Then we have
\begin{equation}\label{rawJs}
\|\bar P(t)\|_{L^\infty(U')} \le C \widehat{\mathcal C}(t) \Big\{ \sup_{[t-1,t]} \norm{\bar P}_{L^2}+\sup_{[t-1,t]} \norm{\bar P}_{L^2}^{\frac {\gamma_1}{{\gamma_1}+1} }\Big\},
\end{equation}
where
\begin{equation}\label{newCtil}
\widehat{\mathcal C}(t) = \tilde \lambda(t)^\frac{1}{\mu_6} (1+\widehat\vartheta(t))^\frac1{{\gamma_1}+1},
\end{equation}
with $\tilde\lambda(t)$ defined by \eqref{lamtildefn}, and
\begin{equation}\label{newuptil}
\widehat\vartheta(t) = 1+ \Big(\int_{t-1}^t\int_{V} |\nabla \bar p_1|^{(2-a)\mu}+|\nabla \bar p_2|^{(2-a)\mu} dxdt\Big)^\frac1{2\mu}.
\end{equation}
Note that $\widehat\vartheta(t)$ is the same as $\tilde\vartheta(t)$ in \eqref{uptildef} with exponent $\gamma_2'=(2-a)\mu$ replacing $\gamma_2=2(1-a)\mu$.
By (5.53) in \cite{HI2}, there is $d_7>0$ such that
\begin{align} \label{coeffdP}
\frac 12\frac{d}{dt} \int_U \bar P^2 dx \le - d_7 \Lambda^{-b}(t) \int_U \bar P^2 dx
+C|\vec a^{(1)}-\vec a^{(2)}|\Lambda(t).
\end{align}
Hence, for $t'\ge 1$
\begin{equation}\label{JbarP1}
\begin{aligned}
\norm{\bar P(t')}_{L^2}^2
&\le e^{-d_7\int_1^{t'} \Lambda(\tau)^{-b}d\tau} \norm{\bar P(1)}_{L^2}^2+C|\vec a^{(1)}-\vec a^{(2)}|\int_1^{t'} e^{-d_7\int_\tau^{t'} \Lambda(\theta)^{-b}d\theta}\Lambda(\tau)d\tau.
\end{aligned}
\end{equation}
Similar to \eqref{supPtprime}, we obtain
\begin{equation}\label{JbarP2}
\begin{aligned}
\sup_{[t-1,t]} \norm{\bar P(t')}_{L^2}^2
&\le C e^{-d_7\int_0^{t} \Lambda(\tau)^{-b}d\tau} \norm{\bar P(1)}_{L^2}^2 +C|\vec a^{(1)}-\vec a^{(2)}|\int_1^{t} e^{-d_7\int_\tau^{t} \Lambda(\theta)^{-b}d\theta}\Lambda(\tau)d\tau.
\end{aligned}
\end{equation}
Let $\Upsilon_2$ be defined as in Corollary \ref{corMM}, and
$$\Upsilon_3=1+\sup_{[0,\infty)} \norm{\psi}_{L^\infty},\quad \Upsilon_7 = 1+\sup_{[2,\infty)}\int_{t-2}^t \tilde f(\tau) d\tau.$$
We have the following result for unbounded time intervals.
\begin{theorem}\label{thmsupP}
Assume \eqref{strictdeg}. Suppose $\Upsilon_2$, $\Upsilon_3$ and $\Upsilon_7$ are finite numbers.
Then we have
\begin{equation}\label{supPlarge}
\sup_{[2,\infty)} \|\bar P\|_{L^\infty(U')}\le C \Upsilon_8 ( \mathcal A+\mathcal A^{\frac {\gamma_1}{{\gamma_1}+1}}),
\end{equation}
where $\Upsilon_8$ is defined by \eqref{Up4def} below, and $\mathcal A$ is defined in Theorem \ref{thm54}.
\end{theorem}
\begin{proof}
Use \eqref{lamtilde} to bound $\Lambda(t)$ and $\tilde \lambda(t)$, we have for all $t\ge 1$ that
\begin{equation}\label{supLg0}
\sup_{\tau\in[t-1,t]}\Lambda(\tau) \le C \Upsilon_7 \Upsilon_3^\frac2{1-a},
\quad \tilde \lambda(t) \le C \Upsilon_7^{\mu_7} \Upsilon_3^\frac{2\mu_7}{1-a}.
\end{equation}
Moreover, using \eqref{newgamtil} to bound $\widehat\vartheta(t)$ with $\gamma_2'$ replacing $\gamma_2$, we obtain
\begin{equation}\label{boundnu}
\widehat\vartheta(t)\le C L'_{12} B_1(t)^\frac{\mu_4(\gamma_2'+a-2)}{2\mu}B_2(t)^\frac1{2\mu}
\le C L'_{12} \Upsilon_3^\frac{\mu_4(\gamma_2'+a-2)}{2\mu} \Upsilon_2^\frac1{2\mu},
\end{equation}
where $L'_{12}=\{\sum_{i=1,2} L_3(\gamma_2'+a,[p_i(0)])\}^\frac1{2\mu}$.
It follows from \eqref{newCtil}, \eqref{supLg0} and \eqref{boundnu} that
\begin{equation}\label{preCpsi}
\widehat{\mathcal C}(t) \le C \eta_2,\quad \text{where }
\eta_2=\Upsilon_7^\frac{\mu_7}{\mu_6} \Upsilon_2^\frac1{2\mu({\gamma_1}+1)} (L'_{12})^\frac{1}{{\gamma_1}+1} \Upsilon_3^{ \frac{2\mu_7}{\mu_6(1-a)} +\frac{\mu_4(\gamma_2'+a-2)}{2\mu({\gamma_1}+1)} }.
\end{equation}
Then we have from \eqref{rawJs} and \eqref{preCpsi} that
\begin{equation}\label{rawJt}
\norm{\bar P(t)}_{L^\infty(U')}
\le C \eta_2 \Big\{ \sup_{[t-1,t]} \norm{\bar P}_{L^2}+\sup_{[t-1,t]} \norm{\bar P}_{L^2}^{\frac {\gamma_1}{{\gamma_1}+1} }\Big\}.
\end{equation}
By \eqref{JbarP2} and \eqref{supLg0}, we have
\begin{align*}
\sup_{[t-1,t]} \norm{\bar P}_{L^2}^2\le C \norm{\bar P(1)}_{L^2}^2+ C|\vec a^{(1)}-\vec a^{(2)}|(\Upsilon_7 \Upsilon_3^\frac2{1-a})^{b+1}.
\end{align*}
We estimate $ \norm{\bar P(1)}_{L^2}$ by \eqref{CalA} with $T=1$ and obtain
\begin{align*}
\sup_{[t-1,t]} \norm{\bar P}_{L^2}\le C \eta_3 \mathcal A, \quad \text{where }\eta_3=(1+A_0)^{1/2} \mathcal K_{1,1}^\frac{2-a}{2(1-a)} + (\Upsilon_7 \Upsilon_3^\frac2{1-a})^\frac{b+1}2,
\end{align*}
with $\mathcal K_{1,1}$ defined by \eqref{Nf1}.
Combining this with \eqref{rawJt}, we obtain \eqref{supPlarge} with
\begin{equation}\label{Up4def}
\Upsilon_8 = \eta_2 \eta_3.
\end{equation}
The proof is complete.
\end{proof}
For results as $t\to\infty$, we have:
\begin{theorem}\label{thm55}
Assume the same as in Theorem \ref{thmsupP}. Then
\begin{equation}\label{JbarP4}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
\le C \Upsilon_9 \Big(|\vec a^{(1)}-\vec a^{(2)}| + |\vec a^{(1)}-\vec a^{(2)}|^\frac{{\gamma_1}}{{\gamma_1}+1}\Big)^{1/2},
\end{equation}
where $\Upsilon_9$ is defined by \eqref{Up10def} below.
\end{theorem}
\begin{proof}
Define $B_7(t)=1+A^\frac2{2-a}+\int_{t-2}^t \tilde f(\tau)d\tau$.
Note that $\limsup_{t\to\infty} B_7(t)\le 2\mathcal A_2$, where $\mathcal A_2$ is defined by \eqref{Ulim2}.
By \eqref{boundedlarge}, for $t$ sufficiently large we have
\begin{equation}\label{Lamlim}
\sup_{\tau\in[t-1,t]}\Lambda(\tau)\le C B_7(t)
\quad \text{and}\quad
\tilde \lambda(t)\le C B_7(t)^{\mu_7}.
\end{equation}
Using \eqref{Lamlim} and \eqref{boundnu} to estimate $\widehat{\mathcal C}(t)$ in \eqref{newCtil}, we derive from \eqref{rawJs} for large $t$ that
\begin{equation}\label{rawJg}
\|\bar P(t)\|_{L^\infty(U')} \le C \eta_4 B_7(t)^\frac{\mu_7}{\mu_6}
\Big\{ \sup_{[t-1,t]} \norm{\bar P}_{L^2}+\sup_{[t-1,t]} \norm{\bar P}_{L^2}^{\frac {\gamma_1}{{\gamma_1}+1} }\Big\},
\end{equation}
where
$\eta_4=\big[ L'_{12} \Upsilon_2^\frac1{2\mu} \Upsilon_3^\frac{\mu_4(\gamma_2'+a-2)}{2\mu} \big]^\frac1{{\gamma_1}+1}.$
Similar to \eqref{supPtprime} and \eqref{JbarP2}, for large $T$ and $t>T$, we have
\begin{equation}\label{PbarJ3}
\sup_{[t-1,t]} \norm{\bar P}_{L^2}^2
\le C e^{-C\int_T^{t} B_7(\tau)^{-b}d\tau} \norm{\bar P(T)}_{L^2}^2+C|\vec a^{(1)}-\vec a^{(2)}|\int_T^{t} e^{-C\int_\tau^{t} B_7(\theta)^{-b}d\theta} B_7(\tau)d\tau.
\end{equation}
Since $B_7(t)$ is bounded, one can easily see from \eqref{rawJg} and \eqref{PbarJ3} that
\begin{align*}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
&\le C \eta_4 \Big\{ \limsup_{t\to\infty} B_7(t)^\frac{\mu_7}{\mu_6} \Big\} \Big\{
|\vec a^{(1)}-\vec a^{(2)}|^\frac12 \Big[ \limsup_{t\to\infty} \int_T^{t} e^{-C\int_\tau^{t} B_7(\theta)^{-b}d\theta} B_7(\tau)d\tau\Big ]^\frac12\\
&\quad + |\vec a^{(1)}-\vec a^{(2)}|^\frac{{\gamma_1}}{2({\gamma_1}+1)} \Big[\limsup_{t\to\infty} \int_T^{t} e^{-C\int_\tau^{t} B_7(\theta)^{-b}d\theta} B_7(\tau)d\tau\Big]^\frac{{\gamma_1}}{2({\gamma_1}+1)}\Big\}.
\end{align*}
Applying Lemma \ref{difflem2} with $h\equiv 1$ to the last two limits, we obtain
\begin{align*}
\limsup_{t\to\infty} \|\bar P(t)\|_{L^\infty(U')}
&\le C \eta_4 \mathcal A_2^{\frac{\mu_7}{\mu_6}}\Big\{
|\vec a^{(1)}-\vec a^{(2)}|^\frac12 \mathcal A_2^{\frac{b+1}{2}}
+ |\vec a^{(1)}-\vec a^{(2)}|^\frac{{\gamma_1}}{2({\gamma_1}+1)} \mathcal A_2^{\frac{(b+1){\gamma_1}}{2({\gamma_1}+1)}}\Big\}.
\end{align*}
Then \eqref{JbarP4} follows with
\begin{equation}\label{Up10def}
\Upsilon_9=\eta_4 \mathcal A_2^{\frac{\mu_7}{\mu_6}+\frac{b+1}2}.
\end{equation}
The proof is complete.
\end{proof}
Similar to Remark \ref{rmk414}, estimate \eqref{JbarP4} shows that the smallness of $\|\bar P(t)\|_{L^\infty(U')}$ as $t\to\infty$ can be controlled by $|\vec a^{(1)}-\vec a^{(2)}|$.
\subsection{Results for pressure gradient}
\label{sec52}
We now study the dependence for pressure gradient.
The results are parallel to those in subsection \ref{sec42}.
For $i=1,2$, we denote $H_i(\xi)=H(\xi,a^{(i)})$ defined in \eqref{Hxi}, and recall that the functional $J_H[\cdot]$ is defined by \eqref{J.def}.
\begin{proposition}\label{Grada}
Let $\delta\in (0,a)$ and $U'\Subset V\Subset U$. There exists a constant $C>0$ depending on $U$, $U'$,$V$ and $\delta$ such that for any $t>0$, we have
\begin{multline}\label{DG9}
\|\nabla P(t)\|_{L^{2-\delta}(U')}
\le C\norm{\bar P(t)}_{L^2}^\frac12 \Big(1+\sum_{i=1,2}\norm{\bar p_{it}(t)}_{L^2}^2+\sum_{i=1,2} J_{H_i}[p_i](t)\Big)^\frac12 \\
\cdot \Big(1+ \sum_{i=1,2} \int_{V} |\nabla p_i(x,t)|^\frac{a(2-\delta)}{\delta}dx \Big)^\frac{\delta}{2(2-\delta)},
\end{multline}
and for any $T>0$, we have
\begin{multline}\label{GP}
\| \nabla P\|_{L^{2-\delta}(U'\times(0,T))}
\le C\Big\{ \norm{\bar P}_{L^2(U\times(0,T))}^\frac12 \Big( T+\sum_{i=1,2}\norm{\bar p_{it}}_{L^2(U\times(0,T))}^2 + \sum_{i=1,2} \int_0^T J_{H_i}[p_i](t)dt\Big)^\frac14 \\
+ |\vec a^{(1)}-\vec a^{(2)}|^\frac12 \Big(T+\sum_{i=1,2}\int_0^T J_{H_i}[p_i](t) dt\Big)^{\frac 1 2}\Big\}
\cdot \Big\{ T+\sum_{i=1,2}\int_{0}^T\int_{V} |\nabla p_i(x,t)|^\frac{a(2-\delta)}{\delta}dx dt\Big\}^\frac{\delta}{2(2-\delta)}.
\end{multline}
\end{proposition}
\begin{proof}
Let $\zeta(x,t)$ be the same cut-off function as in Proposition \ref{PropGrad}.
Multiplying equation \eqref{Erreq} by $\bar P\zeta^2 $, integrating over $U$ and using integration by parts, we obtain
\begin{align*}
\int_U \bar P_t \bar P \zeta^2 dx
&= -\int_U \Big(K(\nabla p_1|,\vec a^{(1)})\nabla p_1-K(|\nabla p_2|,\vec a^{(2)})\nabla p_2 \Big ) \cdot (\nabla P \zeta^2+2\bar P \zeta\nabla \zeta) dx.
\end{align*}
Again, the fact $\nabla \bar P =\nabla P$ was used in the above.
Let $\xi(x)=|\nabla p_1|\vee |\nabla p_2|$. By the monotonicity \eqref{quasimonotone},
\begin{align*}
&\int_U \bar P_t \bar P \zeta^2 dx
\le -(1-a) \int_U K(\xi,\vec a^{(1)}\vee a^{(2)})|\nabla P\zeta|^2 dx \\
&\quad + C|\vec a^{(1)}-\vec a^{(2)}| \int_{U} (|\nabla p_1|^{1-a}+|\nabla p_2|^{1-a})|\nabla P|\zeta^2 dx
+C\int_{U} ( |\nabla p_1|+ |\nabla p_2|)^{1-a}|\bar P|\zeta dx.
\end{align*}
Hence,
\begin{align*}
&(1-a) \int_U K(\xi,\vec a^{(1)}\vee a^{(2)})|\nabla P\zeta|^2 dx
\le C \norm{\bar P_{t}}_{L^2}\norm{\bar P}_{L^2}\\
&\quad +C\Big(\int_U ( |\nabla p_1|+ |\nabla p_2|)^{2-2a}dx\Big)^{1/2} \norm{\bar P}_{L^2}
+ C|\vec a^{(1)}-\vec a^{(2)}| \int_U (|\nabla p_1|^{2-a}+|\nabla p_2|^{2-a})dx.
\end{align*}
By constructing appropriate function $\zeta(x)\in[0,1]$ with $\zeta\equiv 1$ on $V$, we have
\begin{multline}\label{s1}
\int_{V} K(\xi,\vec a^{(1)}\vee a^{(2)})|\nabla P|^2 dx
\le \int_{U} K(\xi,\vec a^{(1)}\vee a^{(2)})|\nabla P|^2 \zeta^2 dx \\
\le C \norm{\bar P}_{L^2} \Big\{ \norm{\bar p_{1t}}_{L^2}+ \norm{\bar p_{2t}}_{L^2}+\Big(\int_U ( |\nabla p_1|+ |\nabla p_2|)^{2-2a}dx\Big)^{1/2} \Big\} \\
+ C|\vec a^{(1)}-\vec a^{(2)}| \int_U (|\nabla p_1|^{2-a}+|\nabla p_2|^{2-a})dx.
\end{multline}
Using \eqref{DG2} with $K(\xi)=K(\xi,\vec a^{(1)}\vee a^{(2)})$ and \eqref{s1} we obtain
\begin{multline*
\|\nabla P(t)\|_{L^{2-\delta}(U')}
\le C\norm{\bar P(t)}_{L^2}^{1/2} \Big\{ \Big(1+\norm{\bar p_{1t}(t)}_{L^2}^2+\norm{\bar p_{2t}(t)}_{L^2}^2 + J_{H_1}[p_1]+J_{H_2}[p_2]\Big)^{1/4} \\
\quad + |\vec a^{(1)}-\vec a^{(2)}|^{1/2} \Big( 1+J_{H_1}[p_1]+J_{H_2}[p_2] \Big)^{1/2}\Big\} \Big( \int_{V} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx \Big)^\frac{\delta}{2(2-\delta)}.
\end{multline*}
Then simply due to $|\vec a^{(1)}-\vec a^{(2)}|\le C$, inequality \eqref{DG9} follows.
To prove \eqref{GP}, we have from \eqref{s1} that
\begin{equation}
\begin{aligned}
&\int_0^T\int_{V} K(\xi)|\nabla P|^2 dxdt \\
&\le C \Big(\int_0^T \norm{\bar P}_{L^2}^2 dt\Big)^{1/2} \Big( \int_0^T \norm{\bar p_{1t}}_{L^2}^2+ \norm{\bar p_{2t}}_{L^2}^2 dt+ \int_0^T\int_U ( |\nabla p_1|+ |\nabla p_2|)^{2-2a}dx dt\Big)^{1/2} \\
&\quad + C|\vec a^{(1)}-\vec a^{(2)}| \int_0^T \int_U (|\nabla p_1|^{2-a}+|\nabla p_2|^{2-a})dx dt.
\end{aligned}
\end{equation}
According to \eqref{GradPclose2},
\begin{equation}\label{s2}
\int_{0}^T \int_{U'} |\nabla P|^{2-\delta} dx dt
\le C \Big ( \int_{0}^T \int_U K(\xi)|\nabla P\zeta|^2 dxdt \Big )^\frac{2-\delta}{2} \Big( \int_{0}^T\int_{V} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx dt\Big)^\frac{\delta}{2}.
\end{equation}
Using \eqref{s1} in \eqref{s2}, we obtain
\begin{multline*
\int_{0}^T \int_{U'} |\nabla P|^{2-\delta} dx dt \\
\le C\Big\{\Big(\int_0^T \norm{\bar P}_{L^2}^2 dt\Big)^{1/2} \Big( \int_0^T \norm{\bar p_{1t}}_{L^2}^2+ \norm{\bar p_{2t}}_{L^2}^2 dt+ \int_0^T\int_U ( |\nabla p_1|+ |\nabla p_2|)^{2-2a}dx dt\Big)^{1/2} \\
+ |\vec a^{(1)}-\vec a^{(2)}| \int_0^T \int_U (|\nabla p_1|^{2-a}+|\nabla p_2|^{2-a})dx dt\Big\}^{\frac {2-\delta} 2} \Big( \int_{0}^T\int_{V} (1+|\nabla p_1|+|\nabla p_2|)^\frac{a(2-\delta)}{\delta}dx dt\Big)^\frac{\delta}{2}.
\end{multline*}
Thus, we have \eqref{GP}.
\end{proof}
\begin{theorem}\label{GradThm3}
Assume \eqref{strictdeg}. For $\delta\in(0,a)$, $0<t_0<1$, and $T>t_0$ we have
\begin{equation}\label{DG7}
\sup_{[t_0,T]} \|\nabla P(t)\|_{L^{2-\delta}(U')} \le C M_{8,t_0,T}(\|\bar P(0)\|_{L^2} +|\vec a^{(1)}-\vec a^{(2)}|^{1/2})^\frac12,
\end{equation}
where $M_{8,t_0,T}$ is defined in \eqref{N9def} below.
\end{theorem}
\begin{proof}
Using estimates \eqref{coeffJsup1}, \eqref{Jpt-boundA0}, \eqref{H-bound-0} and \eqref{gradalot} in \eqref{DG9}, we obtain \eqref{DG7} with
\begin{multline}\label{N9def}
M_{8,t_0,T}=M_{5,T}^{1/4} \Big( t_0^{-1} \sum_{i=1,2} L_5(t_0;[p_i(0)]) + (T+1)\mathcal K_{1,T}^\frac{2-a}{1-a} +\mathcal K_{2,T}\Big)^{1/2}\\
\cdot \Big\{ \sum_{i=1,2}L_2(\nu_1;[p_i(0)]) (T+1)^{\frac{2(\nu_1-2)}{2-a}+1} \mathcal K_{1,T}^{\frac{\nu_1-a}{1-a}}\Big\}^\frac{\delta}{2(2-\delta)}.
\end{multline}
\end{proof}
\begin{theorem}\label{GradThm4}
Assume \eqref{strictdeg}.
Suppose $\Upsilon_1$, $\Upsilon_2$ (defined in \eqref{xgood}) and $\mathcal A_2$ (defined by \eqref{Ulim2}) are finite.
Then for any $\delta\in(0,a)$, we have
\begin{equation}\label{DG8}
\limsup_{t\to\infty} \|\nabla P(t)\|_{L^{2-\delta}(U')} \le C \Upsilon_{10} |\vec a^{(1)}-\vec a^{(2)}|^{1/4},
\end{equation}
where $\Upsilon_{10}$ is defined by \eqref{U8} below.
\end{theorem}
\begin{proof}
Taking limsup of \eqref{DG9} and making use the limits \eqref{coefflim}, \eqref{limsupH1} and estimate \eqref{Kgrad60pw} with $s=\nu_1$ yield
\begin{equation*}
\limsup_{t\to\infty} \|\nabla P(t)\|_{L^{2-\delta}}
\le C \Big(\mathcal A_2^{b+1} |\vec a^{(1)}-\vec a^{(2)}|\Big)^{1/4}
\mathcal A_2^{1/2} \Big\{ \sum_{i=1,2} L_4(\nu_1;[p_i(0)]) \Upsilon_1^{\mu_4(\nu_1-2)} \Upsilon_2 \Big\}^\frac\delta{2(2-\delta)}.
\end{equation*}
Therefore, we obtain \eqref{DG8} with
\begin{equation}\label{U8}
\Upsilon_{10}=\mathcal A_2^{\frac {b+3}4} \Big\{ \Upsilon_1^{\mu_4(\nu_1-2)} \Upsilon_2 \sum_{i=1,2}L_4(\nu_1;[p_i(0)]) \Big\}^\frac\delta{2(2-\delta)}.
\end{equation}
\end{proof}
\begin{theorem}\label{lastthm}
Assume \eqref{strictdeg}.
Let $\delta\in (0,a)$ and $T> 0$. Then there exists a constant $C>0$ depending on $U,U', \delta$ such that
\begin{equation}\label{GP2}
\| \nabla P\|_{L^{2-\delta}(U'\times(0,T))}
\le C M_{9,T} \Big( \norm{\bar P(0)}_{L^2}^{1/2} +|\vec a^{(1)}-\vec a^{(2)}|^{1/2} +|\vec a^{(1)}-\vec a^{(2)}|^{1/4}\Big),
\end{equation}
where $M_{9,T}$ is defined by \eqref{N7def} below.
\end{theorem}
\begin{proof}
Define $N_{2,T}=\sum_{i=1}^2 \norm{\bar p_i(0)}_{L^2}^2 + \sum_{i=1}^2 \|\nabla p_i(0)\|_{L^{2-a}}^{2-a } +(T+1)\mathcal K_{1,T}^\frac{2-a}{1-a} +\mathcal K_{2,T}$.
From \eqref{p-bar-bound1} and \eqref{H-bound-0} we have
\begin{equation}\label{s3}
\sum_{i=1}^2 \norm{\bar p_{it}}_{L^2(U\times(0,T))}^2 + \sum_{i=1}^2\int_0^T\int_U |\nabla p_i(x,t)|^{2-a} dx dt \le C N_{2,T}.
\end{equation}
Neglecting the negative term on the right-hand side of \eqref{coeffdP} and integrating in $t$ twice yield
\begin{equation}\label{s4}
\norm{\bar P}_{L^2(U\times(0,T))}^2 \le C T \Big ( \norm{\bar P(0)}_{L^2}^2 +|\vec a^{(1)}-\vec a^{(2)}| \int_0^T \Lambda(t)dt \Big)
\le C T N_{2,T} \Big ( \norm{\bar P(0)}_{L^2}^2 +|\vec a^{(1)}-\vec a^{(2)}|\Big).
\end{equation}
Combining \eqref{s3}, \eqref{s4}, \eqref{subs0} and \eqref{GP} we get
\begin{align*}
&\| \nabla P\|_{L^{2-\delta}(U'\times(0,T))}
\le C \Big\{ T^{1/4} N_{2,T}^{1/4} ( \norm{\bar P(0)}_{L^2}^2 +|\vec a^{(1)}-\vec a^{(2)}|)^{1/4} \cdot N_{2,T}^{1/4}\\
&\quad +|\vec a^{(1)}-\vec a^{(2)}|^{1/2} N_{2,T}^{1/2} \Big\}
\cdot \Big\{\Big[ \sum_{i=1,2} L_1(\nu_1+a;[p_i(0)]) \Big]
(T+1)^{\frac {2\nu_1}{2-a} - 1} \mathcal K_{1,T}^{\frac{\nu_1}{1-a}}\Big\}^\frac\delta{2(2-\delta)}.
\end{align*}
Therefore, we obtain \eqref{GP2} with
\begin{equation}\label{N7def}
M_{9,T}=(T+1)^{\frac14}N_{2,T}^{1/2} \Big\{\Big[ \sum_{i=1,2} L_1(\nu_1+a;[p_i(0)]) \Big]
(T+1)^{\frac {2\nu_1}{2-a} - 1} \mathcal K_{1,T}^{\frac{\nu_1}{1-a}}\Big\}^\frac\delta{2(2-\delta)}.
\end{equation}
The proof is complete.
\end{proof}
\clearpage
|
\section{Introduction}
Strongly-correlated ensembles of ultracold atoms provide an unique platform for simulating dynamics and models predicted for condensed-phase systems, statistical mechanics, as well as to test quantum-field theoretical hypotheses \cite{Lewenstein:07,Bloch:08,Georgescu:14}. Self-organized phases of trapped ions, atoms, and dipolar systems play in this context a prominent role, as they allow one to study and simulate Wigner crystallization \cite{Dubin:99,Astrakharchik:07,Buechler:07}, supersolidity \cite{Goral:02}, and quantum magnetism \cite{Porras:04,Friedenauer:08,Islam:11}, to mention a few examples.
One peculiar instance is the linear-zigzag instability in ion chains. This instability is observed in a linear array of trapped ions by lowering the transverse confinement: Below a critical value the equilibrium configuration is a double array, forming a zigzag chain \cite{Birkl:92}. The transition is continuous and is classically described by a Landau model \cite{Fishman:08}. In the quantum regime, it is a quantum phase transition of the same universality class of the ferromagnetic transition of an Ising chain in a transverse field \cite{Shimshoni:11,Silvi:13}. The spin order is here associated to the transverse displacement of the ions from the chain axis. It thus naturally offers a testbed for studying, amongst others, kink formation after quenches across the structural transition \cite{kinks} and the spin-Peierls instability \cite{Bermudez:12}. Deep in the quantum regime, where the quantum statistical properties are relevant such as in quantum wires, the linear-zigzag instability is characterized by a rich phase diagram \cite{Meyer:07}.
\begin{figure*}
\subfloat[$\omega_t = 1.1 \ \omega_t^{(c)}$]{
\includegraphics[width=0.3\textwidth]{fig1a.png}}
\subfloat[$\omega_t = 0.99 \ \omega_t^{(c)}$]{
\includegraphics[width=0.3\textwidth]{fig1b.png}}
\subfloat[$\omega_t = 0.8 \ \omega_t^{(c)}$]{
\includegraphics[width=0.3\textwidth]{fig1c.png}}
\caption{(Color online) Side view (main panels) and top view (insets) of the various configurations found in the Monte-Carlo simulations: single ring (linear chain) (a), inhomogeneous configuration (b), and double ring (zigzag chain) (c) of classical dipolar particles confined in the plane perpendicular to the polarizing electric field. The different configurations correspond to three decreasing values of the radial confinement in the ring trap. The inhomogeneous configurations as in (b) indicate a coexistence of linear and zigzag structures, and are numerically found using periodic boundary conditions. Similar structures are found as well in a box with hard walls by varying the transverse frequency or the linear density. \label{fig:1}}
\end{figure*}
In this work we analyse linear-zigzag instability in other systems exhibiting repulsive power-law interactions of the type $1/r^\alpha$, focusing in particular on the case $\alpha=3$ corresponding to dipolar gases. For exponent $\alpha>2$ we show that, in absence of external potentials imposing long-range order, the instability becomes of first order due to the coupling between transverse and axial vibrations, which modifies the critical properties. Quite remarkably, this longitudinal-transverse coupling among the modes plays an analogous role as the coupling between spins and phonons for ferromagnetic transitions in compressible lattices \cite{Larkin,Imry}. Evidence for a first-order transition is brought forward by the numerical observation of inhomogenous configurations, indicating that at the instability the chain alternates regions in which the ions exhibit either zigzag or linear order, as shown in Fig. \ref{fig:1}. The regions are separated by kinks whose form is reminiscent of soliton excitations. Such configurations were not reported in previous numerical studies, which analysed the instability for small samples~\cite{Astrakharchik2008,Astrakharchik2009} (composed of about 16 or less dipolar particles), and are observed when the particles number exceeds several tens of particles. Further insight on the nature of the transition is gained by means of a low-energy theory, which shows that the parameter range in which the inhomogeneous configurations are found shrinks in the thermodynamic limit, even though it remains finite. The transition therefore can be considered as "weakly" first-order or nearly second order, using the therminology of Refs. \cite{Larkin,Yurtseven:06}.
This article is organized as follows. In Sec.~\ref{Sec:2} we describe the model and discuss the stability of the ring chain. Monte-Carlo results are presented Sec.~\ref{Sec:3}. In Sec.~\ref{res:low-energy} we compare the numerical results with the analytical predictions of the low-energy theory. Sec.~\ref{res:low-energy} also contains the analysis of the nature of the transition and our predictions for the thermodynamic-limit behaviour. Finally, Sec.~\ref{Sec:5} discusses the role of thermal fluctuations and offers our concluding remarks.
\section{Physical system}
\label{Sec:2}
We consider $N$ classical particles of mass $m$ which are confined by an anisotropic trap on the $x-y$ plane, assuming a very tight confinement along the $z$ direction. The particles interact via a power-law repulsive potential of the form
\begin{equation}
V_{\rm int}({\bf r}_1,\ldots, {\bf r}_N) = \frac{C_D}{2} \sum_{j \neq l}\frac{1}{|{\bf r}_j-{\bf r}_l|^{\alpha}}\,,
\label{eq:basic_model}
\end{equation}
where $C_D$ is the interaction strength and ${\bf r}_j=(x_j,y_j)$ is the position of particle $j=1,\ldots,N$. The generic power-law exponent $\alpha$ describes, for instance, the dipolar interaction for $\alpha=3$ (when the particles possess permanent dipoles and are polarized by an external field orthogonal to the plane), or Van-der-Waals interactions for $\alpha=6$. Moreover, the particles are confined by a ring trap of radius $R_0$, which generates the (radially harmonic) potential
\begin{align}
V_{\rm trap}({\bf r}_1,\ldots, {\bf r}_N) = \frac{1}{2} m \omega_t^2 \sum_{j=1}^N\left( r_j - R_0 \right)^2\,,
\end{align}
with $r_j=|{\bf r}_j|$ and $\omega_t$ the frequency in the radial direction. Such trapping potential is currently realized for quantum gases \cite{Exp,Exp1,Exp2,Exp3,Exp4,Exp7}.
For large radii it approaches a linear trap with periodic boundary conditions.
We will numerically seek in Sec.\ref{Sec:3} for the configuration which minimizes the energy in the total potential
\begin{equation}
\label{V:tot}
V=V_{\rm int}+V_{\rm trap}\,,
\end{equation}
close to the linear-zigzag instability. The regime of stability of the linear configuration is analytically identified by means of a Taylor expansion of the potential about the linear array. This has been performed in Refs. \cite{Astrakharchik2008,Astrakharchik2009}. Below we report the basic steps, here applied to the specific configuration of a ring trap.
\subsection{Taylor expansion about the equilibrium configuration}
In order to analyse the stability properties of the ring chain, we first rewrite the interaction potential $V_{\rm int}$, Eq. \eqref{eq:basic_model}, in terms of polar coordinates, such that
$V_{\rm int}=(1/2)\sum_{j,l\neq j}U(r_j,\phi_j,r_l,\phi_l)$. We then use the center-of-mass and relative coordinates $R_{jl} = (r_j + r_l)/2$, $\rho_{jl} = r_j - r_l$ and $\phi_{jl} = \phi_j - \phi_l$, and cast $U(r_j,\phi_j,r_l,\phi_l)$ into the form
\begin{align}
U\left(R_{jl},\rho_{jl},\phi_{jl}\right) = \frac{C_D}{[ \rho_{jl}^2 \cos^2(\phi_{jl}/2) + 4 R_{jl}^2 \sin^2(\phi_{jl}/2) ]^{\frac{\alpha}{2}}}\,.
\label{interactions}
\end{align}
We then perform a systematic expansion of the interaction energy about the configuration in which the ions form a single ring. We denote by $R$ the ring radius, which results to be $R>R_0$ due to the interparticle repulsion. Moreover, we denote by $a$ the uniform interparticle distance along the ring, such that $a=2\pi R/N$. Assuming that one dipole of the ring is pinned, the single ring is a regular structure which exhibits discrete translational invariance where the particles are located at radial position $r_j=R$ and at angles $\phi_j= 2\pi j/N$ ($j=0,\ldots,N-1$). This configuration corresponds to equilibrium since the first derivatives of the total potential $V$, Eq. \eqref{V:tot}, vanish. In order to verify that the equilibrium is stable, we consider the further terms in the Taylor expansion. Setting $r_j = R + a \Psi_j$ and $\phi_j = 2 \pi j /N + a \Theta_j / R$, the expansion reads
\begin{widetext}
\begin{align}
V_{\text{int}} = \frac 1 2 \sum_{j=1}^N \sum_{l \neq j}^N \sum_{0 \leq n_1 + n_2 + n_3 \leq {\color{black} 6}} \frac{1}{n_1 ! n_2 ! n_3 !} \frac{a^{n_1+n_2+n_3}}{2^{n_1}R^{n_3}}\frac{\partial^n U\left(R, 0, \phi_j^{(0)} - \phi_l^{(0)}\right)}{\partial R^{n_1} \partial \rho^{n_2} \partial \phi^{n_3}} (\Psi_j + \Psi_l)^{n_1}(\Psi_j - \Psi_l)^{n_2}(\Theta_j - \Theta_l)^{n_3}, \label{eq:expansion}
\end{align}
\end{widetext}
where $n_1$, $n_2$, $n_3$ are positive integers. In these derivatives all even-order derivatives in $\rho$ vanish because of the symmetry of the single-ring configuration.
\subsection{Stability of the single ring}
The stability of the linear chain is determined by analysing the Hessian of the second-order derivatives. An analytical expression of the dispersion relation is found using the Fourier modes $\Psi_k$ and $\Theta_k$, such that $ \Psi_j = \frac{1}{\sqrt N} \sum_{k} \tilde{\Psi}_k e^{i k j a } \label{eq:lin-modes-psi}$, $\Theta_j =\frac{1}{\sqrt N} \sum_{k} \tilde{\Theta}_k e^{i k j a}$ with $k=-\pi N/L,\ldots, N\pi/L$ and $L=2\pi R=Na$. Denoting by $V^{(2)}$ the term of the second-order Taylor expansion for $V_{\rm int}$, it takes the form $ V^{(2)} =\sum_k V^{(2)}_k$ with
\begin{align}
V^{(2)}_k =a^2& \sum_{l \neq 0} \left[ \left| \tilde\Psi_k \right|^2 \frac{ 1}{ 4} \frac{\partial^2 U(R,0,2\pi l/N)}{\partial R^2} \cos^2\left( k l a /2 \right) \right. \nonumber \\
&+ \left| \tilde\Psi_k \right|^2 \frac{\partial^2 U(R,0,2\pi l/N)}{\partial \rho^2} \sin^2\left( k l a /2 \right) \nonumber \\
&+ \left| \tilde\Theta_k \right|^2 \frac{1}{R^2} \frac{\partial^2 U(R,0,2\pi l/N)}{\partial \phi^2} \sin^2\left( k l a /2 \right)\nonumber \\
&\left. + \tilde\Theta_{k} \tilde\Psi_{-k}\frac{1}{4R} \frac{\partial^2 U(R,0,2\pi l/N)}{\partial \phi \partial R}\sin\left(k l a \right) \right] \,.
\label{eq:energy-FT}
\end{align}
For $R,N\to\infty$, but keeping $a=2\pi R/N$ constant, the derivatives with respect to $R$ vanish, such that axial and transverse Fourier modes become decoupled \cite{Astrakharchik2009}. In this thermodynamic limit, the linear chain is mechanically unstable at $\omega_t=\omega_t^{(c)}(N)$, with $$\lim_{N\to\infty}\omega_t^{(c)}(N)=\sqrt{(93\zeta(5)/8)C_D/(ma^5)}$$
and $\zeta(5)$ the Riemann's zeta function. At this value of the transverse trap frequency the frequency of the transverse mode with quasi momentum $k_0=\pi/a$, $ \tilde\Psi_{k_0}=\sum_j(-1)^j\Psi_j/\sqrt{N}$, vanishes. The details of the corresponding calculation are reported in Ref. \cite{Astrakharchik2009}. For the Coulomb interaction this instability is a second-order phase transition which is classically described by the Landau model \cite{Fishman:08}. The mode at $k_0$ is then the soft mode driving the instability, and the order parameter the displacement $a\Psi_j$ in the radial direction. In Ref. \cite{Astrakharchik2008,Shimshoni:11,Piacente:10,Altman:12} it has been conjectured that this may hold for any power-law repulsive interaction with $\alpha\ge 1$.
\section{Minimal-energy configurations}
\label{Sec:3}
We first numerically study the linear-zigzag instability, focusing on the case $\alpha=3$ of dipolar interactions. We search for the particle configuration which minimizes the total potential energy $V =V_{\rm trap} + V_{\rm int}$ for different values of the trap frequency $\omega_t$. We determine the classical ground state of a dipolar gas using the Basin-Hopping Monte-Carlo method \cite{MC}, with which we identify the equilibrium configurations corresponding to the global minimum of the potential energy for $N$ ranging from $16$ to $1100$. We note that the configurations we find are expected to reproduce the correct ground state at $T=0$ when the interaction energy exceeds the kinetic energy, hence at sufficiently high densities and for large permanent dipoles \cite{Astrakharchik2008,Citro,Silvi:13}.
For sufficiently large frequencies $\omega_t$ (or, alternatively, small linear densities $1/a$), we find a single array, or linear configuration, as in Fig. \ref{fig:1}(a). Its equilibrium radius $R$ is larger than the confining radius $R_0$ due to the repulsive interactions. For $\omega_t<\omega_t^{(c)}$ and a sufficiently large number of particles the minimal energy configurations determined numerically are inhomogeneous. In particular, they result to be a mixture of single- and two-ring structures, as shown in Fig.~\ref{fig:1}(b). The inhomogeneous configurations appear when the number of dipoles exceeds a certain value $N_0>32$, and they are thus absent for $N=16$, which was the case reported in Ref. \cite{Astrakharchik2008,Astrakharchik2009}.
For this parameter range the homogeneous double ring (zigzag configuration) is metastable, separated by a small {\color{black} energy barrier from the linear chain. Both structures are at higher energy than the inhomogeneous one, which exhibits domains of linear and zigzag configurations.} By further decreasing $\omega_t$ the global minimum is the zigzag configuration, whose equilibrium positions are given by $r_j=R+(-1)^j b$ and $\phi_j= 2\pi j/N$, where $b>0$ is half the radial distance between the two rings.
The zigzag configuration is illustrated in Fig. \ref{fig:1}(c). It is found provided the number of particles is even, while for odd $N$ the structure exhibits topological defects \cite{Cartarius}.
Figure \ref{fig:displacement} displays the average transverse displacement as a function of the trapping frequency as obtained from the Monte-Carlo calculations. The region of inhomogeneous configurations is clearly visible as a deviation from the expected square-root behaviour predicted by the Landau theory for a second-order phase transition \cite{Fishman:08,Astrakharchik2008}. A zoom on the transition region also illustrates how the actual transition occurs quite suddenly (within the numerical accuracy) and at a frequency which is slightly larger than the frequency $\omega_t^{(c)}$. The frequency $\omega_t$ below which inhomogeneous configurations are found tends asymptotically to the value $\omega_t=1.0011(9)\omega_t^{(c)}$. Finite-size corrections scale linearly with $1/N$, as illustrated in Fig.\ref{fig:displacement} (b).
\begin{figure}
\subfloat{\includegraphics[width=0.24\textwidth,valign=t]{fig2a.png}}
\subfloat{ \includegraphics[width=0.24\textwidth,valign=t]{fig2b.png}}
\caption{(Color online) (a) Average transverse displacement $b$ (solid line) along the ring (in units of the interparticle distance along the ring) as a function of $\omega_t/\omega_t^{(c)}$ for N=500 dipoles.
The dashed line indicates the average displacement of a continuous transition, that is obtained by only allowing transverse particle movement.
The inset shows the displacement $b$ close to the transition region for 1100 dipoles.
(b) Trap frequency below which inhomogeneous configurations are the minimal energy solutions in the numerical simulations as a function of 1/$N$, where $N$ is the number of particles along the ring. The red line is a linear fit $\omega_t/\omega_t^{(c)}=a+b/N$ with parameters $a = 1.0011\pm 0.0009$ and $b=-0.77\pm 0.08$.
\label{fig:displacement}}
\end{figure}
The results presented here are {\it not} a peculiarity of the ring geometry and of the power-law exponent $\alpha=3$. We have also run Monte-Carlo simulations for linear traps with hard walls as boundaries, and for particles on a ring with other power-law interactions with $\alpha>2$. In both cases we have found inhomogeneous configurations, similar to those reported here. For Coulomb interactions, on the other hand, we have found a homogeneous ground-state solution, in agreement with the results of Ref.~\cite{Fishman:08}. In the Coulomb case, indeed, the inhomogeneous configurations are excitations \cite{Landa:NJP}, and the linear-zigzag transition is continuous \cite{Fishman:08}. Our numerical results clearly indicate that the structural transition for dipolar gases (and in general for $\alpha>2$) deviates from the behaviour predicted from the Landau theory for second-order phase transitions.
\section{Analysis of the structural transition}
\label{res:low-energy}
Since at the mechanical instability second-order derivatives of the potential energy vanish, the thermodynamic properties in this parameter region can be analytically determined by considering higher-order terms in the Taylor expansion.
For this purpose we derive here an expression of the potential-energy functional at low energies. This then allows us to gain analytical insight of the numerical results.
\subsection{Low-energy model}
To proceed, we recall that close to the structural transition low-energy excitations correspond to normal modes in the longitudinal (tangential) direction with wave numbers $|k|a\ll 1$, and in the transverse (radial) direction with $|k-k_0|a\ll 1$. The latter are long-wavelength excitations of the staggered field $\Psi_{j,st}=(-1)^j\Psi_j$. The procedure is a straightfoward extension
of the one performed for Coulomb interactions in Ref.~\cite{Fishman:08,DeChiara:10}, to which we refer for further details of the derivation.
Keeping just the modes within this low energy cutoff and going back to real space, one can resort to a continuum theory, introducing now the fields as a function of the continuous variable $x$:
\begin{eqnarray}
\Psi(x) \rightarrow & \frac{1}{\sqrt{N}} \sum_k \tilde{ \Psi}_k e^{i k x a} \, ,\\
\Theta(x) \rightarrow & \frac{1}{\sqrt{N}} \sum_k \tilde{ \Theta}_k e^{i k x a} \, ,
\end{eqnarray}
where the coordinate $x$ is in units of the average interparticle distance $a$. With this low-energy cutoff one obtains an expression for the potential energy, $V_0 = V^{\text{eq}} + V_0$, where $V^{{\color{black}\text{eq}}}$ is the equilibrium energy of the single ring and
\begin{align}
V_0\!&=\!\! \frac{C_D}{a^\alpha} \!\!\! \int \!\!\! \dd x
\left[ h_1^2 (\partial_x\Theta)^2 \!\!+ h_2^2 (\partial_x\Psi)^2 \!\!+ \Delta \Psi^2 \!\! + e(\partial_x\Theta) \Psi^2 + f \Psi^4
\right. \nonumber \\
&\left.+
r (\partial_x\Psi )^2 \Psi^2+ \ell (\partial_x\Theta) ^2 \Psi^2
+ t \Psi^6 + p (\partial_x\Theta)^3 + q \Psi^4 \partial_x\Theta \right] \, , \label{eq:cont-en}
\end{align}
and all parameters are dimensionless constants defined in Appendix \ref{App:A}. Expression \eqref{eq:cont-en} differs from the one reported in Ref. \cite{DeChiara:10} since it contains an expansion up to 6th order as well as the coupling between axial and transverse modes. For Coulomb repulsion this coupling leads to a renormalization of the coefficients, such that sufficiently close to the zigzag instability one can reduce the potential to an effective $\phi^4$ model and neglect higher order corrections. The inhomogeneous configuration found numerically, however, suggest that for $\alpha>2$ this coupling may play a relevant role.
\subsection{Minimum energy configurations}
\label{thermo-limit}
In order to get an insight into the nature of the transition,
we now look for uniform solutions for the fields $\Psi$ and $\Theta'=\partial_x\Theta$ minimizing the long-wavelength potential energy (\ref{eq:cont-en}) for different values of $\Delta$, and thus of $\omega_t$.
This allows us to find an analytical solution, with which we can verify whether there exists a parameter regime where the linear and the zigzag configurations are {\color{black} both local minima of the potential energy.}
The solutions are extrema of the potential, satisfying $\partial V_0/\partial \Theta' = 0$ and $\partial V_0/\partial \Psi = 0$ with positive-definite Hessian matrix.
We determine an effective potential for the transverse-displacement field $\Psi$ by eliminating the solution for $\Theta'$, which in the small-$\Psi$ limit reads
\begin{align}
\label{Thetaprime}
\Theta'=- \frac{1}{2 h_1^2} \Psi^2 \left[e +\left(q - \frac{e l}{h_1^2} + \frac{3 e^2 p}{4 h_1^4} \right)\Psi^2\right].
\end{align}
Note that there is a second solution for $\Theta'$, which is finite at small $\Psi$, and thus inconsistent with our initial assumptions. Substitution of Eq. \eqref{Thetaprime} in the expression (\ref{eq:cont-en}) leads to the effective potential density
\begin{equation}
V_{\rm eff}\propto \Delta \Psi^2 + u_{\rm eff} \Psi^4/4 + \lambda \Psi^6 \, ,
\end{equation}
where $u_{\rm eff}=(4 f-e^2/h_1^2)$ and $\lambda= \left( \frac{l e^2}{4 h_1^4} - \frac{e^3 p}{8 h_1^{6}} - \frac{e q}{2 h_1^2} + t \right)$. Using the explicit form of the coefficients for the case of dipolar interactions (see App. \ref{App:A}) we obtain that $u_{\rm eff}<0$ and $\lambda>0$. The effective model thus describes a first-order phase transition at $\Delta=0$. It is interesting to point out that the sign of the quartic term is negative due to the coupling with the axial vibrations. Figure~\ref{fig:psi6} shows the energy of the local minima and the corresponding displacement field $\Psi$ obtained from the low-energy effective model as a function of the control parameter~$\Delta$.
This solution predicts a sudden jump into two stable local minima near the dynamical instability of the single ring, which is characteristic of a first-order transition. {\color{black} Note that this solution is restricted to uniform transverse fields. Numerically, we find that the inhomogeneous solution is at lower energy, corresponding to the coexistence of the zigzag and linear configurations}. Quite remarkably, the parameter region of coexistence of phases is very narrow and close to the frequency $\omega_t^{(c)}$. Therefore, this transition is of 'weakly first-order' or of nearly second order \cite{Larkin,Imry}.
\begin{figure}
\subfloat{\includegraphics[width=0.24\textwidth]{fig3a_ringv2.png}} %
\subfloat{\includegraphics[width=0.24\textwidth]{fig3b_ring.png}}
\caption{(Color online) (a) Local minima of the energy in Eq.~(\ref{eq:cont-en}) for homogeneous solutions ($\Psi' = 0$) and (b) corresponding transverse-displacement field (in units of distance $a$ along the chain) as a function of $\Delta$ (dimensionless) and in the thermodynamic limit. The region of coexistence of phases is in the interval $\Delta\in [0,0.0009]$, corresponding to $\omega_t^{(c)}\le\omega_t \le 1.000075 \ \omega_t^{(c)}$.
\label{fig:psi6}}
\end{figure}
\subsection{Finite-size system}
We now address the predictions of the low-energy model for the displacement fields $\Theta$ and $\Psi$ in a ring of finite size. An analytical solution can be obtained if we keep just the leading order in the transverse-axial coupling, after setting $r,\ell,t,p,q=0$ in Eq. \eqref{eq:cont-en}. This corresponds to a truncation of the effective potential to fourth order. This approach is clearly not capable to describe the nature of the phase in the thermodynamic limit, since it misses the sixth-order terms which stabilize the uniform solution. Nevertheless, in the finite-size ring, the solution is inhomogeneous, stabilized by the presence of the gradient terms in \eqref{eq:cont-en} and can be employed to account for the observed inhomogeneous configurations close to the transition point.
Using the variational principle we determine the equations for ~$\Psi(x)$ and $\Theta(x)$ which minimize Eq.\eqref{eq:cont-en},
\begin{align}
\frac{\dd }{\dd x} \left( 2 h_1^2 (\Theta') + e \Psi^2 \right) =& 0 \label{eq:el-theta} \, , \\
2 h_2^2 \Psi '' - 2 \Delta \Psi - 2 e \Theta ' \Psi - 4 f \Psi^3 =& 0 \, . \label{eq:el-psi}
\end{align}
These equations admit
an inhomogeneous soliton-like solution, of the form \cite{Carr,Cominotti}
\begin{align}
\Psi^2(x) =& y_3 \, \text{cn}^2\left( \frac{\sqrt{g (y_3-y_1)}}{2} x \Big| m \right)\,, \label{eq:soliton} \\
\Theta '(x) =& \frac{1}{2} B - \frac{1}{2} \frac{e}{h_1^2} \Psi^2(x)\label{Theta}\,,
\end{align}
where cn is a Jacobi elliptic function and $y_1$, $y_3$, and $B$ are determined by solving coupled transcendental equations, while $m=y_3/(y_3-y_1)$ and $g=-u_{\rm eff} /h_2^2$ (see Appendix \ref{App:B}).
\begin{figure}
\subfloat{
\includegraphics[width=0.24\textwidth]{fig4a_ring.png}}%
\subfloat{
\includegraphics[width=0.24\textwidth]{fig4b_ring.png}}
\caption{(Color online) (a) Transverse squared displacement $\Psi^2(x)$ and (b) axial displacement $\Theta(x)$ (in units of distance $a$ along the chain) as a function of distance $x$ along the chain (in units of $a$) for the minimal energy configurations on a ring with $N=90$ particles. Numerical Monte-Carlo data (circles) are compared to the solutions of Eqs.~\eqref{eq:soliton}-\eqref{Theta} (solid lines). From top to bottom, the blue, red and black curves correspond to trap frequencies $\omega_t= 0.9915 \omega_t^{(c)}$, $0.99 \omega_t^{(c)}$, $0.985 \omega_t^{(c)}$, respectively.\label{Fig:soliton}}
\end{figure}
Figure~\ref{Fig:soliton} displays the behaviour predicted by Eqs. (\ref{eq:soliton}-\ref{Theta}) along the chain and the corresponding numerical results, showing a very good agreement within the model's regime of validity. The energy of the inhomogeneous configurations is obtained by substituting the corresponding solutions into the potential-energy density. It is found to be smaller than the energy of the zigzag case, in full agreement with the numerical observations.
Inspection of Fig.~\ref{fig:displacement} shows that in the numerical calculations for a finite ring the parameter region of phase coexistence is larger than in the thermodynamic limit, extending to negative values of $\Delta$. This can be explained noticing that boundary effects yield a renormalized control parameter $\Delta_{\rm eff}$ for the transition. Details are reported in Appendix \ref{App:B}.
\section{Discussion and conclusions}
\label{Sec:5}
Our predictions are strictly valid when the effect of fluctuations is negligible. To study the effect of thermal fluctuations on the various configurations found at zero temperature, we have performed a finite temperature Monte-Carlo calculation, and determined the pair correlation function $g_2(r,\phi) = \langle \sum_{i,j\neq i} \delta (r - (r_i-r_j)) \delta (\phi - (\phi_i-\phi_j)) \rangle$ for temperatures which are lower than the difference between the inhomogeneus and zigzag energies. Figure \ref{fig:g2} displays the two-particle correlation functions for different values of $\Delta<0$. The inhomogeneous configurations are clearly visible as the correlation is smeared along the radial direction in a semicircular shape, indicating varying radial displacements (thus, inhomogeneous $\Psi(x)$). This result for the pair-correlation function is considerably different from both the one for the linear configuration, characterized by a periodic structure only along the tangential (axial) direction, and the one for a uniform two-ring configuration, where radially the only possible relative distances allowed are $\pm \Psi$ and 0. The clear distinction between the various configurations is lost for temperatures higher than the energy {\color{black} barrier} between the various configurations. Taking the value of the dipolar moment of LiCs molecules~\cite{LiCs} and typical densities of the ongoing experiments \cite{Wu:2012}, we estimate that the energy gap between the {\color{black} inhomogeneous and uniform configurations} corresponds to a temperature of 0.2 nK. Although this value is still quite challenging from an experimental point of view, it can rapidly increase at increasing the density and the dipolar moment of the gases.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig5_ring.pdf}
\caption{(Color online) Two-particle correlator $g_2(r,\phi)$ of 90 dipoles evaluated numerically and at $T=8 \times 10^{-4} \ C_D/ (a^3 k_B)$. The configurations correspond to a uniform single-ring (top, $\omega_t=1.05 \omega_t^{(c)}$), an inhomogeneous structure (center, $\omega_t=0.98 \omega_t^{(c)}$) and uniform two-ring configuration (bottom, $\omega_t=0.7 \omega_t^{(c)}$).
\label{fig:g2}}
\end{figure}
To estimate the parameter range for which the system is in a classical regime, we can compare the length scale associated with the quantum fluctuations $a$, with the length scale associated with the interactions $r_0$, which can be estimated to be $r_0 = m C_D / \hbar^2$~\cite{Astrakharchik2008}. If $a \ll r_0$, the ground state energy of the system is well approximated by the classical ground state energy. In this regime, the quantum fluctuations have a similar effect as the temperature has in a classical system~\cite{Astrakharchik2009}. For LiCs molecules, the characteristic length is given by $r_0 = 63 \, \mu$m. Taking a Gaussian wave packet of the same size, the kinetic energy of a molecule can be estimated to be $E \approx k_B \cdot 9 \, \mu$K, which is larger than the energy gap of 0.2 nK. Thus, for the parameters of LiCs molecular gases, it is expected that quantum fluctuations will smear the transition.
In conclusion, we have shown that the linear-zigzag instability for power-law interactions $\alpha>2$ is a first-order phase transition, even though weak, whose hallmark is the appearance of inhomogeneous soliton-like structures which minimize the energy of finite systems. The instability is thus not described by a $\phi^4$ model, since the coupling with the axial vibrations substantially modifies the properties of the transition. This is different from Coulomb systems, where the dispersion relation of the axial modes leads just to a renormalization of the coefficient of the $\phi^4$ model in the critical region, without changing its nature \cite{Silvi:unpublished}. The dipolar system therefore realizes an example of Ising model coupled to axial phonons \cite{Larkin,Imry}.
Whether the weakly first-order nature of the transition survives the inclusion of quantum fluctuations is a question for future work.
In the quantum regime, the instability is expected to exhibit the existence of a critical point with enhanced symmetry and nonuniversal critical exponents, in analogy to the model discussed in Ref. \cite{Sitte:2009}.
\begin{acknowledgments}
The authors are grateful to Eugene Demler, Shmuel Fishman, Frank Hekking, Julia Meyer, Efrat Shimshoni, and Pietro Silvi for stimulating discussions and helpful comments. GM acknowledges hospitality by the ion storage group at NIST, Boulder, during completion of this work. Financial support from the European Commission (STREP PICC), the German Research Foundation, the Handy-Q ERC grant N.\;258608 and the ANR project no. ANR-13-JS01-0005-01 is acknowledged.
\end{acknowledgments}
\onecolumngrid
|
\section{Introduction}
In this paper we deal with the quasilinear stationary Kirchhoff equation
$$
\left\{
\begin{array}{rclcc}
-M\left(\displaystyle\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u & = & f(x,u, \nabla u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega
\end{array}
\right.
\eqno{(P)}
$$
where $\Omega \subset \mathbb{R}^{N}, N\geq 1$, is a bounded smooth domain, $f:\Omega \times \mathbb{R} \times \mathbb{R}^{N}\rightarrow [0,+\infty)$ is a continuous function satisfying
\vspace{0.5cm}
\noindent $(f_{1})$ There are a continuous function $h:\overline{\Omega} \times \mathbb{R} \to [0,+\infty)$ and $\eta \in [0,2]$ such that $|f(x,t,y)|\leq h(x,t)(1+|y|^{\eta})$ for all $(x,t,y)\in \Omega \times \mathbb{R} \times \mathbb{R}^{N}$;
\vspace{0.5cm}
\noindent and $M:[0, +\infty ) \rightarrow [0, +\infty ) $ satisfies
\vspace{0.5cm}
\noindent $(M_{1})$ $M$ is continuous and increasing;
\vspace{0.5cm}
\noindent $(M_{2})$ There is a positive constant $m$ such that $M(t) \geq m>0$ for all $t\in \mathbb{R}$.
\vspace{0.5cm}
Problem $(P)$ is a generalization of the classical stationary Kirchhoff equation
\begin{equation}\label{kirchhoff classical}
\left\{
\begin{array}{rclcc}
-M\left(\displaystyle\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u & = & f(x,u) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial\Omega .
\end{array}
\right.
\end{equation}
As it is well known, problem (\ref{kirchhoff classical}) is the general form of the stationary counterpart of the hyperbolic Kirchhoff equation
\begin{equation}\label{kirchhoff hyperbolic}
\rho \frac{\partial^{2}u}{\partial t^{2}}-\left(\frac{P_{0}}{h}+\frac{E}{2L}\int^{L}_{0}\left|\frac{\partial u}{\partial x}\right|^{2}dx\right)\frac{\partial^{2}u}{\partial x^{2}}=0,
\end{equation}
that appeared at the first time in the work of Kirchhoff \cite{kirchhoff}, in 1883. The equation in (\ref{kirchhoff hyperbolic}) is called Kirchhoff Equation and it extends the classical D'Alembert wave equation, by considering the effects of the changes in the length of the strings during the vibrations.
The interest of the mathematicians on the so called nonlocal problems like (\ref{kirchhoff classical}) (nonlocal because of the presence of the term $M(\|u\|^{2})$, which implies that equations in $(P)$ and (\ref{kirchhoff classical}) are no longer pontwise equalities) has increased because they represent a variety of relevant physical and engineering situations and requires a nontrivial apparatus to solve them. It is worthwhile to emphasize that the most of the articles on this subject are concerned with the semilinear case, i.e., $f=f(x,u)$.
In several places we should face nonhomogeneous Kirchhoff term, that is, the function $M$ also depends on the variable $x\in \Omega$. For instance, L\'imaco, Clark and Medeiros \cite{medeiros} attack a biharmonic evolution equation in which the operator is of the form
$$
\mathcal{L}u\equiv a(x)u''+\Delta (b(x)\Delta u)-M\left( x,t, \int_{\Omega}|\nabla u(x,t)|^{2}dx\right)\Delta u
$$
motivated by the problem of vertical flexion of fully clamped beams. In Figueiredo, Morales-Rodrigo, Santos Junior \& Su\'arez \cite{figueiredo et al} consider a problem whose equation is of the form
$$
-M\left(x, \int_{\Omega}|\nabla u|^{2}dx\right)\Delta u =f(x,u) ~~\mbox{in}~~ \Omega ,
$$
under homogeneous Dirichlet boundary condition, by using a bifurcation argument. Note that for $M$ nonhomogeneous we lose the variational structure and the approach we use in the present article can not be used, at least in a direct way.
In this work, we explore the presence of the gradient term $|\nabla u|$, which makes problem $(P)$ nonvariational, by considering the nonlocal term $M$ with the minimal typical assumptions $(M_{1})-(M_{2})$ which, up to now, at least to our knowledge, has not been considered yet. We point out that in the original Kirchhoff equation the term $M$ is of the form $M(t)=a+bt, a,b>0$, which enjoys assumptions $(M_{1})$ and $(M_{2})$.
Our approach was motivated by Cuesta Leon \cite{cuestaleon} and in it the method of sub-supersolution and pseudomonotone operator theory play a key role. We should say that here we have to surmount several technical difficulties provoked by the presence of the nonlocal term $M$.
The method of sub and supersolution for semilinear nonlocal equations has been previously used by some authors. We cite some of them.
In Alves-Corr\^ea \cite{alves-correa} the authors study the problem
\begin{equation} \label{ problem1}
\left\{
\begin{array}{rclcc}
-M\left(\displaystyle\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u & = & f(x,u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega ,
\end{array}
\right.
\end{equation}
via sub-supersolution (monotone iteration) by considering $M:\mathbb{R}^{+}\rightarrow \mathbb{R}^{+}$ nonincreasing and $H(t)=M(t^{2})t$ increasing. Note that the typical Kirchhoff term $M(t)=a+bt, a,b>0$ is increasing, i.e., the result in \cite{alves-correa} does not include such a $M$.
In Corr\^ea \cite{correa} the author studies the problem
\begin{equation} \label{correa problem}
\left\{
\begin{array}{rclcc}
-a\left(\displaystyle\int_{\Omega}|u|^{q}dx\right)\Delta u & = & H(x)f(u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega ,
\end{array}
\right.
\end{equation}
where $a: \mathbb{R} \rightarrow \mathbb{R}^{+}$ is a function satisfying $a(s)\geq a_{0}>0 \; \forall s\in \mathbb{R}$, $s\mapsto s^{\frac{1}{q}}a(s)$ is increasing and $s\mapsto a(s)$ is decreasing. In particular, $a$ is a bounded function. In this work the author uses sub-supersolution combined with fixed point theory.
In Chipot-Corr\^ea \cite{chipot-correa} the authors consider the problem
\begin{equation} \label{chipot-correa problem}
\left\{
\begin{array}{rclcc}
-\mathcal{A}(x,u)\Delta u & = & \lambda f(u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega ,
\end{array}
\right.
\end{equation}
where, among other things, $\mathcal{A}:\Omega \times \mathbb{R} \rightarrow \mathbb{R}$ satisfies
\begin{equation}
0<a_{0}\leq \mathcal{A}(x,u) \leq a_{\infty}, \; \mbox{a.e.}\; x\in \Omega , \forall u\in L^{p}(\Omega ).
\end{equation}
In that work, it is used sub-supersolution via fixed point properties and, again, the nonlocal term is bounded.
Here, we permit, inspired by \cite{cuestaleon}, that the Kirchhoff term $M$ may be of the form of the original one.
\begin{definition}\label{weak solution definition}
We say that $u\in H_{0}^{1}(\Omega )\cap L^{\infty}(\Omega )$ is a weak solution of the problem $(P)$ if
\begin{equation}\label{weak solution}
M\left(\int_{\Omega}|\nabla u|^{2}dx\right)\int_{\Omega}\nabla u \nabla vdx=\int_{\Omega}f(x,u, \nabla u)vdx \; \forall v\in H_{0}^{1}(\Omega ).
\end{equation}
\end{definition}
The main result of this paper is as follows:
\begin{thm}\label{main result}
Assume the hypotheses $(M_{1})-(M_{2})$ and $(f_1)$. Moreover, suppose that there are $\overline{u} \in W^{1,\infty}(\Omega) $ and a family $(\underline{u}_{\delta}) \subset W^{1,\infty}_0(\Omega)$ such that:
\begin{equation}\label{weak supersolution}
\int_{\Omega}\nabla \overline{u}\nabla vdx \geq \int_{\Omega}\frac{1}{m}f(x, \overline{u}, \nabla \overline{u})vdx ~~ \forall v\in H_{0}^{1}(\Omega ), \; v\geq 0 ~~~\mbox{and} ~~ \overline{u} \geq 0 ~~ \mbox{on} ~~ \partial \Omega,
\end{equation}
$$
\|\underline{u}_\delta\|_{1,\infty} \to 0 ~~ \mbox{as} ~~ \delta \to 0,
$$
$$
\underline{u}_\delta \leq \overline{u} ~~ \mbox{in} ~~ \Omega ~~ \mbox{for} ~~ \delta ~~ \mbox{small enough},
$$
and given $\alpha >0$, there is $\delta_0>0$ such that
\begin{equation}\label{weak subsolution}
\int_{\Omega}\nabla \underline{u}_\delta \nabla vdx \leq \int_{\Omega}\frac{1}{\alpha}f(x, \underline{u}_{\delta}, \nabla \underline{u}_\delta)vdx ~~ \forall v\in H_{0}^{1}(\Omega ), \; v\geq 0 ~~~\mbox{for} ~~ \delta \leq \delta_0 .
\end{equation}
Then there is a small enough $\delta >0$ such that problem $(P)$ has a weak solution $u$ satisfying $\underline{u}_\delta \leq u \leq \overline{u}$.
\end{thm}
\section{Preliminary Results}
In this section we introduce some concepts and results in order to attack problem $(P)$. The abstract results concerning monotone operators can be found, for instance, in Lions \cite{lions}, Ne$\check{c}$as \cite{necas} and Pascali \& Sburlan \cite{pascali}
\begin{definition}\label{monotone operator}
Let $E$ be a reflexive Banach space and $E^{\ast}$ its topological dual. A nonlinear mapping $A:D(A)\subset E\rightarrow E^{\ast}$
is said to be monotone if it satisfies
\begin{equation}\label{inequality monotone}
\langle Au-Av,u-v\rangle \geq 0 \;\; u,v\in D(A).
\end{equation}
If the inequality (\ref{inequality monotone}) is strict for $u\neq v$, we say that $A$ is strict monotone. Here, $\langle \cdot , \cdot \rangle$ means the duality pairing between $E^{\ast}$ and $E$.
\end{definition}
\begin{definition}\label{gradient}
If $E$ is a Hilbert space and $\phi : E\rightarrow \mathbb{R}$ is $C^{1}$-functional, the gradient of $\phi $, denoted by $\nabla \phi :E\rightarrow E$, is defined, through the Riesz Representation Theorem, by
$$
\langle \nabla \phi (u), w \rangle =\phi' (u)w \;\; \forall u, w \in E,
$$
where $\langle \cdot \; , \; \cdot \rangle $ is the inner product in $E$.
\end{definition}
\begin{lem}\label{convex-monotone}
If $E$ is a Hilbert space and $\phi \in C^{1}(E, \mathbb{R} )$, then $\phi$ is convex (strictly convex) if, and only if, $\nabla \phi$ is monotone (strictly monotone).
\end{lem}
\begin{definition}\label{S+}
Let $E$ be a Banach space and $\mathcal{C}\subset E$ a closed convex set. An operator $T: \mathcal{C}\rightarrow E^{\ast}$ is said to be of type $(S_{+})$ provided that whenever $x_{n}\rightharpoonup x$ in $E$ and
\begin{equation}\label{condition S+}
\limsup_{n \to +\infty} \; \langle Tx_{n}, x_{n}-x\rangle \leq 0,
\end{equation}
then $x_{n}\rightarrow x$ in $E$.
We remark that the condition (\ref{condition S+}) can be rewritten as
\begin{equation}
\limsup_{n \to +\infty} \; \langle Tx_{n}-Tx, x_{n}-x\rangle \leq 0.
\end{equation}
\end{definition}
\begin{definition}\label{pseudomonotone definition}
Let $E$ be a Banach space and $B: E\rightarrow E^{\ast}$ an operator. We say that $B$ is pseudomonotone if $u_{n}\rightharpoonup u$ in $E$ and
\begin{equation}
\limsup_{n \to +\infty} \; \langle Bu_{n},u_{n}-u \rangle \leq 0,
\end{equation}
then
\begin{equation}
\liminf_{n \to +\infty} \; \langle Bu_{n},u_{n}-v \rangle \geq \langle B(u),u-v \rangle \;\; \forall v\in E.
\end{equation}
\end{definition}
\begin{definition}\label{demicontinuous definition}
We say that $T:E\rightarrow E^{\ast}$ is demicontinuous if $x_{n}\rightarrow x$ in $E$ implies that $Tx_{n}\rightharpoonup Tx$ in $E^{\ast}$.
\end{definition}
\begin{lem}\label{demicontinuous S+}
Any demicontinuous operator $T:E\rightarrow E^{\ast}$ of type $(S_{+})$ is pseudomonotone.
\end{lem}
\begin{thm}\label{B surjective}
Let $E$ be a reflexive and separable Banach space and $B:E \rightarrow E^{\ast}$ an operator satisfying
\begin{description}
\item[(i)] $B$ is coercive, i.e.,
\begin{equation}\label{B coercive}
\frac{\langle B(u),u\rangle}{\|u\|}\rightarrow +\infty \;\;\mbox{as}\;\; \|u\|\rightarrow +\infty \;
\end{equation}
\item[(ii)] $B$ is bounded and continuous;
\item[(iii)] $B$ is pseudomonotone.
\end{description}
Then $B$ is surjective, that is, $B(E)=E^{\ast}$.
\end{thm}
Next, $\| \cdot \|$ will denote the usual norm $\|u\|= \left(\displaystyle\int_{\Omega}|\nabla u|^{2}dx\right)^{\frac{1}{2}}$ in $H_{0}^{1}(\Omega )$.
\begin{lem}\label{L monotone}
The operator $L:H_{0}^{1}(\Omega )\rightarrow H^{-1}(\Omega )$ given by
\begin{equation}\label{L monotone}
\langle Lu,v\rangle =\displaystyle\int_{\Omega}M(\|u\|^{2})\nabla u \nabla vdx
\end{equation}
is strictly monotone.
\end{lem}
\dem Let us consider $G:H_{0}^{1}(\Omega ) \rightarrow \mathbb{R}$ given by
\begin{equation}\label{G definition}
G(u)=\frac{1}{2}\widehat{M}(\|u\|^{2})\;\; \forall u \in H_{0}^{1}(\Omega ),
\end{equation}
where $\widehat{M}(t)=\displaystyle\int_{0}^{t}M(\tau )d\tau$. Because $M$ is positive and continuous, we have that $G$ is strictly convex. Furthermore
\begin{equation}\label{G' = L}
G'(u)v=\langle \nabla G(u),v\rangle \displaystyle\int_{\Omega}M(\|u\|^{2})\nabla u \nabla vdx=\langle Lu,v\rangle \;\; \forall u,v\in H_{0}^{1}(\Omega ),
\end{equation}
that is, $\nabla G =L$ and so, in view of Lemma \ref{convex-monotone}, $L$ is strictly monotone. \hfill\rule{2mm}{2mm}
\begin{lem}\label{L is S+}
$L$ is of type $(S_{+})$.
\end{lem}
\dem Let $(u_{n})$ be a sequence in $H_{0}^{1}(\Omega )$ such that
\begin{equation}\label{weak convergence 1}
u_{n} \rightharpoonup u \;\; \mbox{in} \;\; H_{0}^{1}(\Omega)
\end{equation}
and
\begin{equation}\label{limsup1}
\limsup_{n \to +\infty} \; \langle Lu_{n}, u_{n}-u\rangle \leq 0.
\end{equation}
We have to prove that $u_{n}\rightarrow u$ in $H_{0}^{1}(\Omega )$. For this, we first note that
$$
\langle Lu_n,u_n-u\rangle = M(\|u_n\|^{2})\displaystyle\int_{\Omega}|\nabla u_n|^{2}dx-M(\|u_n\|^{2})\displaystyle\int_{\Omega}\nabla u_n\nabla udx
$$
that is,
$$
\frac{1}{M(\|u_n\|^{2})} \langle Lu_n,u_n-u\rangle = \displaystyle\int_{\Omega}|\nabla u_n|^{2}dx-\displaystyle\int_{\Omega}\nabla u_n\nabla udx
$$
Note that $M(\|u_n\|^{2})\geq m>0$, and so,
$$
0\geq \limsup_{n \to \infty} \|u_n\|^{2}-\|u\|^{2},
$$
which implies
$$
\|u\|^{2}\geq \limsup_{n \to \infty} \|u_n\|^{2} \geq \liminf_{n \to \infty} \|u_n\|^{2} \geq \|u\|^{2},
$$
from where it follows that $\|u_n\|^{2}\rightarrow \|u\|^{2}$. Invoking the weak convergence $u_n\rightharpoonup u$ in $H_{0}^{1}(\Omega )$, we see that $u_n\rightarrow u$ in $H_{0}^{1}(\Omega )$, and the proof of the lemma is over. \hfill\rule{2mm}{2mm}
\section{Proof of the Main Theorem}
From now on, we fix $R>0$ large enough such that
$$
\|\nabla \overline{u}\|_{\infty}, \|\nabla \underline{u}_{\delta}\|_{\infty} \leq R
$$
for all $\delta$ small enough, where $\overline{u}$ and $u_\delta$ were given in Theorem \ref{main result}. We recall that if $\overrightarrow{V}=(V_{1}, \ldots , V_{N})\in (L^{\infty}(\Omega))^{N}$, we have $\|\overrightarrow{V}\|_{\infty}=\displaystyle\max_{1\leq i\leq N}\|V_{i}\|_{\infty}$. Moreover, we set the function $g_R:\mathbb{R} \to \mathbb{R}$ given by
$$
g_R(t)=
\left\{
\begin{array}{l}
t, ~~ \mbox{if} ~~ |t| \leq R,\\
\mbox{}\\
R, ~~ \mbox{if} ~~ t \geq R, \\
\mbox{}\\
-R, ~~ \mbox{if} ~~ t \leq -R.
\end{array}
\right.
$$
Here, we would like to point out that
\begin{equation} \label{E0}
g_R(t)=t ~~ \mbox{if} ~~ |t| \leq R
\end{equation}
and
$$
|g_{R}(t)|= \min\{R, |t|\} ~~ \mbox{for all} ~~ t \in \mathbb{R}.
$$
Hence,
\begin{equation} \label{E1}
|g_{R}(t)|\leq R ~~~\mbox{and} ~~ |g_{R}(t)|\leq |t| ~~~\mbox{for all} ~~ t \in \mathbb{R}.
\end{equation}
Taking into account the above function $g_{R}$ and and their properties, we will consider the following auxiliary function $f_R:\Omega \times \mathbb{R} \times \mathbb{R}^{N} \to [0,+\infty)$ given by
$$
f_R(x,t,y)=f(x,t,\stackrel{\rightarrow}{g_R}(y)),
$$
where $\stackrel{\rightarrow}{g_R}(y)=(g_R(y_1),g_R(y_2),...,g_R(y_N))$. Using the definition of the function $f_R$, it follows the ensuing estimates:
\begin{equation} \label{E2}
|f_R(x,t,y)| \leq h(x,t)(1+|\stackrel{\rightarrow}{g_R}(y)|^{\eta}) \leq h(x,t)(1+R^{\eta}N^{\frac{\eta}{2}})
\end{equation}
and
\begin{equation} \label{E3}
|f_R(x,t,y)| \leq h(x,t)(1+|\stackrel{\rightarrow}{g_R}(y)|^{\eta}) \leq h(x,t)(1+|y|^{\eta}).
\end{equation}
Furthermore, it is crucial observing that
\begin{equation} \label{E4}
f_R(x,t,y)=f(x,t,y) ~~~\mbox{if} ~~ |y| \leq R,
\end{equation}
and so,
$$
f_R(x,\overline{u},\nabla \overline{u})= f(x,\overline{u},\nabla \overline{u}) ~~~\mbox{and} ~~ f_R(x,{u}_\delta,\nabla \overline{u}_\delta)= f(x,\overline{u}_\delta,\nabla \overline{u}_\delta).
$$
Using function $f_R$, we are able to fix the following auxiliary problem
$$
\left\{
\begin{array}{rclcc}
-M\left(\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u & = & f_{R}(x,u, \nabla u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega.
\end{array}
\right.
\eqno{(AP)}
$$
Our intention is proving the existence of a solution $u_R$ for $(AP)$ with $\|\nabla u_R\|_{\infty} \leq R$ if $R$ is large enough and, because of (\ref{E4}), we can guarantee that $u_R$ is a solution of the original problem $(P)$.
\subsection{Supersolution}
In this subsection, we will be concerned on supersolutions of the problem $(AP)$.
\begin{definition}\label{supersolution definition}
We say that $w \in W^{1, \infty}(\Omega)$ is a supersolution of the problem $(AP)$ if
\begin{equation}\label{inequality supersolution}
-M(\|w\|^{2})\Delta w\geq f_R(x, w, \nabla w) \;\;\mbox{in}\;\; \Omega ~~ \mbox{and} ~~ w \geq 0 ~~ \mbox{on} ~~ \partial \Omega,
\end{equation}
\end{definition}
in the weak sense, that is,
\begin{equation}\label{weak supersolution}
M(\|w\|^{2})\int_{\Omega}\nabla w\nabla vdx \geq \int_{\Omega}f_R(x, w, \nabla w)vdx
\end{equation}
$\forall v\in H_{0}^{1}(\Omega ), \; v\geq 0$ a.e. in $\Omega$
How to get a supersolution to the problem $(AP)$? Under the hypotheses of Theorem \ref{main result}, we know that $\overline{u}\in W^{1, \infty}(\Omega)$ verifies
\begin{equation}\label{local supersolution}
-\Delta \overline{u} \geq \frac{1}{m}f(x, \overline{u}, \nabla \overline{u}) \;\;\mbox{in}\;\; \Omega.
\end{equation}
Since $M(t)\geq m>0$ for all $t\geq 0$ and $f(x, \overline{u}, \nabla \overline{u})=f_R(x, \overline{u}, \nabla \overline{u})$, we deduce that
$\overline{u}\in W^{1, \infty}(\Omega)$ is a supersolution of the problem $(AP)$.
We point out that sub and supersolutions for quasilinear local problems like
\begin{equation} \label{local problem}
\left\{
\begin{array}{rclcc}
-\Delta u & = & f(x,u, \nabla u ) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega
\end{array}
\right.
\end{equation}
were studied in \cite{cuestaleon}.
\begin{lem}\label{a priori supersolution}
Let $u_R \in H_{0}^{1}(\Omega )\cap L^{\infty}(\Omega )$ be a weak solution of the problem $(AP)$ with $0<u_R \leq \overline{u}$ a.e. in $\Omega$. Then there is a constant $K=K(\|\overline{u}\|_{\infty},R)$ such that
\begin{equation}
\|u_R\|^{2}\leq K.
\end{equation}
\end{lem}
\dem Setting $T= \|\overline{u}\|_{\infty}$, by condition $(f_{1})$ combined with (\ref{E2}), there is a constant $C=C(T)>0$ such that
\begin{equation}\label{consequence f1}
|f_R(x,t,y)|\leq C(1+R^{\eta}N^{\frac{\eta}{2}})=C_1
\end{equation}
for all $(x,t,y)\in \Omega \times [0,T]\times \mathbb{R}^{N}$. Since $u_R$ is a solution of $(AP)$, we have
\begin{equation}
M(\|u_R\|^{2})\displaystyle\int_{\Omega}\nabla u_R \nabla vdx=\displaystyle\int_{\Omega}f_R(x,u_R, \nabla u_R)vdx \;\; \forall v\in H_{0}^{1}(\Omega )
\end{equation}
and so
\begin{equation}
M(\|u_R\|^{2})\|u_R\|^{2}=\displaystyle\int_{\Omega}f_R(x,u_R, \nabla u_R)u_Rdx.
\end{equation}
Invoking (\ref{consequence f1}), we obtain
\begin{equation}
M(\|u_R\|^{2})\|u_R\|^{2}\leq C_1\displaystyle\int_{\Omega}|u_R|dx \leq C_1\displaystyle\int_{\Omega}|\overline{u}|dx
\end{equation}
leading to
\begin{equation}
m\|u_R\|^{2}\leq C_2,
\end{equation}
from where it follows that there is $K>0$ satisfying $\|u_R\|^{2}\leq K$. \hfill\rule{2mm}{2mm}
\subsection{Subsolution}\label{subsolution section}
In this section we will be concerned on subsolutions of $(AP)$.
\begin{definition}\label{subsolution definition}
We say that $w \in W^{1, \infty}(\Omega)$ is a subsolution of the problem $(AP)$ if
\begin{equation}\label{inequality subsolution}
-M(\|w\|^{2})\Delta w \leq f_R(x, w, \nabla w) \;\;\mbox{in}\;\; \Omega ~~ \mbox{and} ~~ w \leq 0 ~~~\mbox{on} ~~ \partial \Omega,
\end{equation}
\end{definition}
in the weak sense, that is,
\begin{equation}\label{weak subsolution}
M(\|w\|^{2})\int_{\Omega}\nabla w\nabla vdx \leq \int_{\Omega}f_R(x, w, \nabla w)vdx
\end{equation}
$\forall v\in H_{0}^{1}(\Omega ), \; v\geq 0$ a.e. in $\Omega$
In order to construct a subsolution, we consider the family $(\underline{u}_{\delta}) \subset W_{0}^{1, \infty}(\Omega )$ mentioned in Theorem \ref{main result}, we know that there is $\delta^{*}>0$ such that
\begin{equation}
\underline{u}_{\delta}\leq \overline{u} \;\; \forall \delta \in [0, \delta^{\ast}],
\end{equation}
with
\begin{equation}
\|\underline{u}_{\delta}\|_{1, \infty}\rightarrow 0 \; \mbox{as}\; \delta \rightarrow 0^{+}.
\end{equation}
Thereby, fixing $\alpha =\displaystyle\max_{t\in [0, 1 ]}M(t)$, we can reduce if necessary $\delta^{*}$ to get
\begin{equation}
-\Delta \underline{u}_{\delta}\leq \frac{1}{\alpha}f(x, \underline{u}_{\delta}, \nabla \underline{u}_{\delta}) ~~ \mbox{in} ~~ \Omega ~~ \mbox{and} ~~ u_\delta=0 ~~ \mbox{on} ~~ \partial \Omega.
\end{equation}
Once that $f(x, \underline{u}_{\delta}, \nabla \underline{u}_{\delta})=f_R(x, \underline{u}_{\delta}, \nabla \underline{u}_{\delta})$, we can claim that $\underline{u}=u_\delta$ for $\delta \in (0, \delta^{*})$ is a subsolution of $(AP)$.
\subsection{Another Auxiliary Problem}
In what follows, we define
$$
z_R(x,t,y)=\left\{
\begin{array}{rl}
f_R(x,\underline{u}(x),\nabla \underline{u}(x)), & t\leq \underline{u}(x), \\
f_R(x,t,y), & \underline{u}(x)\leq t \leq \overline{u}(x), \\
f_R(x, \overline{u}(x), \nabla \overline{u}(x)), & t\geq \overline{u}(x)
\end{array}
\right.
$$
and for $l\in (0,1)$ we define the function
$$
\gamma_R (x,t)=-(\underline{u}(x)-t)^{l}_{+}+(t-\overline{u}(x))_{+}^{l}.
$$
Using the above functions, we consider below a second auxiliary problem
\begin{equation} \label{auxiliary problem}
\left\{
\begin{array}{rclcc}
-M\left(\displaystyle\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u & = & z_R(x,u, \nabla u )-\gamma_R (x,u) & \mbox{in} & \Omega , \\
u & = & 0 & \mbox{on} & \partial \Omega.
\end{array}
\right.
\end{equation}
Next, our goal is proving the existence of a solution for the problem (\ref{auxiliary problem}). To this end, we will use Theorem \ref{B coercive} to the operator
$$
\begin{array}{rcl}
B: H_{0}^{1}(\Omega ) & \rightarrow & H^{-1}(\Omega ) \\
u & \mapsto & B(u)
\end{array}
$$
where
$$
\begin{array}{rcl}
B(u): H_{0}^{1}(\Omega ) & \rightarrow & \mathbb{R} \\
v & \mapsto & \langle B(u),v\rangle
\end{array}
$$
is given by
$$
\langle B(u),v\rangle =M(\|u\|^{2})\int_{\Omega}\nabla u \nabla v dx -\int_{\Omega}z_R(x,u, \nabla u)vdx +\int_{\Omega}\gamma_R (x,u)vdx.
$$
In what follows, we are going to show that $B$ is onto. So, there exists $u_R\in H_{0}^{1}(\Omega )$ such that $B(u_R)=0$ in $H^{-1}(\Omega )$. Consequently, $u_R$ is a weak solution of the auxiliary problem. If such a solution enjoys $\underline{u}\leq u_R \leq \overline{u}$ a.e. in $\Omega$ we get a solution of problem $(AP)$.
Plainly $B$ is continuous. In what follows, we fix our attention to others properties of $B$ in order to apply Theorem \ref{B coercive}.
\begin{lem}
$B$ is coercive.
\end{lem}
\dem First note that
$$
\langle B(u),u\rangle =M(\|u\|^{2})\|u\|^{2}-\int_{\Omega }z_R(x,u, \nabla u )udx +\int_{\Omega }\gamma_R (x,u)udx.
$$
It follows from the definition of $z_R$ that there exists $C=C(R)>0$ such that
$$
z_R(x,t,y)\leq C ~~~ \forall (x,t,y) \in \Omega \times \mathbb{R} \times \mathbb{R}^{N}
$$
and
$$
|\gamma (x,t)|\leq C_{1}+C_{2}t^{l} ~~~ \forall (x,t) \in \Omega \times \mathbb{R}.
$$
Consequently,
$$
|z_R(x,u, \nabla u)u| \leq C|u|
$$
and
$$
|\gamma (x,u)|\leq C_{1}|u|+C_{2}|u|^{l+1}.
$$
From these last inequalities,
$$
-\int_{\Omega}z_R(x,u, \nabla u)udx \geq -\int_{\Omega}|z_R(x,u, \nabla u)|udx \geq -C_{1}\|u\|
$$
and
$$
\int_{\Omega}\gamma (x,u)udx \geq -\int_{\Omega}|\gamma (x,u)u|dx \geq -C_{3}\int_{\Omega}|u|dx -C_{4}\int_{\Omega}|u|^{l+1}dx,
$$
that is,
$$
\int_{\Omega}\gamma (x,u)udx \geq -C_{5}\|u\|-C_{6}\|u\|^{l+1}.
$$
Since $M(t)\geq m>0$ for all $t\geq 0$, one has
$$
\langle B(u),u \rangle \geq m \|u\|^{2}-C_{7}\|u\|-C_{6}\|u\|^{l+1}
$$
which yields
$$
\frac{\langle B(u),u \rangle}{\|u\|}\geq m\|u\| -C_{7}-C_{6}\|u\|^{l}
$$
and the result follows because $l\in (0,1)$. \hfill\rule{2mm}{2mm}
\begin{lem}
$B$ is pseudomonotone.
\end{lem}
\dem Let $(u_{n}) \subset H_{0}^{1}(\Omega )$ be a sequence satisfying
$$
u_{n}\rightharpoonup u ~~ \mbox{in} ~~ H_{0}^{1}(\Omega ) ~~ \mbox{and} ~~ \limsup_{n \to \infty} \; \langle B(u_{n}), u_{n}-u \rangle \leq 0,
$$
and recall that
\begin{equation}
\langle B(u_{n}), u_{n}-u \rangle =\langle L u_{n}, u_{n}-u\rangle -\int_{\Omega}h(x,u_{n}, \nabla u_{n})(u_{n}-u)dx+\int_{\Omega}\gamma (x,u_{n})(u_{n}-u)dx.
\end{equation}
Note that
$$
\left|\int_{\Omega}z_R(x,u_{n}, \nabla u_{n})(u_{n}-u)dx\right|\leq C|u_{n}-u|_{1} \rightarrow 0
$$
and
$$
\int_{\Omega}\gamma_R (x,u_{n})|u_{n}-u|dx \rightarrow 0.
$$
Consequently,
$$
\limsup_{n \to \infty} \; \langle B(u_{n}),u_{n}-u\rangle =\limsup_{n \to \infty} ~\langle Lu_{n}, u_{n}-u\rangle .
$$
Since $L$ is an operator of the type $(S_{+})$, it follows that $u_{n}\rightarrow u$ in $H_{0}^{1}(\Omega )$ and invoking the continuity of $B$, we obtain
$$
\liminf_{n \to \infty} \; \langle B(u_{n}),u_{n}-u \rangle = \langle B(u),u-v \rangle \;\; \forall v \in H_{0}^{1}(\Omega ),
$$
showing that $B$ is pseudomonotone. \hfill\rule{2mm}{2mm}
\vspace{0.5 cm}
From the above lemmas, the operator $B$ enjoys all the hypotheses of Theorem \ref{B coercive} and so $B$ is onto. Consequently, there is $u_R \in H_{0}^{1}(\Omega )$ such that $B(u_R)=0$.
\subsection{Existence of Solution for $(AP)$}
As we remarked before, it is enough to show that $\underline{u} \leq u_R \leq \overline{u}$. In this section, we will denote $u_R$ by $u$.
\vspace{0.5cm}
\noindent ${ \bf 1^{st}}$ {\bf Step}. $u \leq \overline{u}$.
\vspace{0.5cm}
For this first step, we take $v=(u-\overline{u})_{+}$ as a test function. Then,
$$
M(\|u\|^{2})\displaystyle\int_{\Omega}\nabla u \nabla (u-\overline{u})_{+}dx = \displaystyle \int_{\Omega}z_R(x,u, \nabla u)(u-\overline{u})_{+}- \displaystyle \int_{\Omega}\gamma_R (x,u)(u-\overline{u})_{+}
$$
Thus,
$$
\begin{array}{rcl}
\int_{\Omega}\nabla u \nabla (u-\overline{u})_{+}dx & = & \int_{\Omega}\frac{1}{M(\|u\|^{2})}f_R(x, \overline{u}, \nabla \overline{u})(u-\overline{u})_{+}dx-\frac{1}{M(\|u\|^{2})} \int_{\Omega}(u-\overline{u})_{+}^{l+1}dx \\
\mbox{}\\
& \leq & \frac{1}{m}\int_{\Omega}f_R(x, \overline{u}, \nabla \overline{u})(u-\overline{u})_{+}dx-\frac{1}{M(\|u\|^{2})} \int_{\Omega}(u-\overline{u})_{+}^{l+1}dx \\
\mbox{}\\
& \leq & \int_{\Omega}\nabla \overline{u}\nabla (u-\overline{u})_{+}dx-\frac{1}{M(\|u\|^{2})} \int_{\Omega}(u-\overline{u})_{+}^{l+1}dx.
\end{array}
$$
Combining these inequalities, we get
$$
0\leq \int_{\Omega}|\nabla (u-\overline{u})_{+}|^{2}dx\leq -\frac{1}{M(\|u\|^{2})} \int_{\Omega}(u-\overline{u})_{+}^{l+1}dx \leq 0,
$$
from where it follows that $ u \leq \overline{u}$ in $\Omega$.
\vspace{0.5cm}
\noindent ${\bf 2^{nd}}$ {\bf Step}. $\underline{u}\leq u$.
\vspace{0.5cm}
Firstly, we point out that if $\delta >0$ is small enough, there is $\beta^{*}>0$, independent of $\delta$, such that $\|u\|^{2}\leq \beta^{*}$. Indeed, note that
$$
M(\|u\|^{2})\int_{\Omega}|\nabla u|^{2}dx=\int_{\Omega}z_R(x,u,\nabla u)udx-\int_{\Omega}\gamma_R (x,u)udx.
$$
By the first step, $\gamma_R (x,u)=-(\underline{u}-u)_{+}^{l}$. Then,
$$
m\|u\|^{2}\leq C\int_{\Omega}|u|dx+ \int_{\Omega}(\underline{u}-u)_{+}^{l}|u|
$$
This last inequality gives
$$
m\|u\|^{2} \leq C\|u\|+ C\|\underline{u}\|_{\infty}^{l}\|u\|+C\|u\|^{l+1}.
$$
Thereby, there is $\beta^{*}=\beta^{*}(R,m,l)>0$, independent of $\delta>0$ small enough, such that
$$
\|u\|^{2} \leq \beta^*.
$$
In what follows, we reduce $\delta>0$ if necessary, to get
$$
-\Delta \underline{u}\leq \frac{1}{\alpha^{*}} f_R(x,\underline{u}, \nabla \underline{u})
$$
where $\alpha^{*} =\displaystyle \max_{0\leq t\leq \beta^*}M(t)$. Choosing $v=(\underline{u}-u)_{+}$, we obtain
$$
\begin{array}{rcl}
M(\|u\|^{2})\int_{\Omega}\nabla u \nabla (\underline{u}-u)_{+}dx & = & \int_{\Omega}z_R(x,u,\nabla u)(\underline{u}-u)_{+}dx-\int_{\Omega}\gamma_R (x,u)(\underline{u}-u)_{+}dx \\
& = & \int_{\Omega}z_R(x,\underline{u}, \nabla \underline{u})(\underline{u}-u)_{+}dx+\int_{\Omega}(\underline{u}-u)_{+}^{l+1}dx
\end{array}
$$
and so
$$
\int_{\Omega}\nabla u\nabla (\underline{u}-u)_{+}dx=\int_{\Omega}\frac{1}{M(\|u\|^{2})}z_R(x,\underline{u}, \nabla \underline{u})(\underline{u}-u)_{+}dx+\int_{\Omega}\frac{1}{M(\|u\|^{2})}(\underline{u}-u)_{+}^{l+1}dx.
$$
Hence
$$
\begin{array}{rcl}
\int_{\Omega}\nabla u \nabla (\underline{u}-u)_{+}dx & \geq & \int_{\Omega}\frac{1}{\alpha^{*}}f_R(x, \underline{u} , \nabla \underline{u})(\underline{u}-u)_{+}dx +\frac{1}{M(\|u\|^{2})}\int_{\Omega}(\underline{u}-u)_{+}^{l+1}dx \\
\mbox{}\\
& \geq & \int_{\Omega}\nabla \underline{u}\nabla (\underline{u}-u)_{+}dx+\frac{1}{M(\|u\|^{2})}\int_{\Omega}(\underline{u}-u)_{+}^{l+1}dx.
\end{array}
$$
Then
$$
0\geq \int_{\Omega}|\nabla (\underline{u}-u)_{+}|^{2}dx+\frac{1}{M(\|u\|^{2})}\int_{\Omega}(\underline{u}-u)_{+}^{l+1}dx \geq 0
$$
and this implies that $(\underline{u}-u)_{+}=0$. Thus, $\underline{u}\leq u$ in $\Omega$, and the proof of the existence of solution for $(AP)$ is over.
\subsection{Existence of Solution for $(P)$}
To begin with, we observe that in the last subsection we proved the existence of a solution $u_R$ of $(AP)$ verifying $\underline{u} \leq u_R \leq \overline{u}$ in $\Omega$. Here, we would like point out that $\underline{u}$ and $\overline{u}$ does not depend of $R$, for $R$ large enough. In what follows, we denote $u_R$ by $u$.
Our goal is to show that there is $R^{*}>0$ such that
$$
\|\nabla u\|_{\infty} \leq R ~~~~ \mbox{for} ~~~ R \geq R^{*}.
$$
By Elliptic Regularity,
\begin{equation*}
u \in W^{2,p}(\Omega )\,\,\,\forall p\in \lbrack 1,+\infty ),
\end{equation*}
because $f_R \in L^{\infty }([0,+\infty ))$ and $u\in L^{\infty
}(\Omega )$. From now on, we will
fix $p$ such that
\begin{equation}
W^{2,p}(\Omega )\hookrightarrow C^{1,\alpha }(\overline{\Omega })
\label{CC1}
\end{equation}%
is a continuous embedding. Now, we observe that $u$ is a solution of the
problem
\begin{equation*}
-\Delta u+u=B_{R}(x)(1+|\nabla u|^{2}),
\end{equation*}%
where
\begin{equation*}
B_{R}(x)=\frac{u+\frac{f_R(x,u(x),\nabla u(x))}{M(\|u\|^{2})}}{1+|\nabla
u|^{2} }.
\end{equation*}%
Once that
\begin{equation*}
|f_{R}(x,t,y)|\leq h(x,t)(1+|y|^{\eta}) ~~~\forall (x,t,y) \in \Omega \times \mathbb{R} \times \mathbb{R}^{N},
\end{equation*}%
combining the fact that $\eta \in [0,2]$, $\underline{u} \leq u \leq \overline{u}$ in $\Omega$ and $\|\underline{u}\|_\infty, \|\overline{u}\|_\infty$ does not depend of $R$, for $R$ large enough, the conditions $(f_1)$ and $(M_2)$ guarantee the existence of $C^{\ast }>0$, independent of $R$, such that
\begin{equation*}
|B_{R}(x)|\leq C^{\ast }\,\,\,\forall x\in \Omega, \,\,\,\mbox{for} ~~ R ~~ \mbox{large enough}.
\end{equation*}%
Thereby, there is $R_1>0$ such that
\begin{equation}
\Vert B_{R}\Vert _{\infty }\leq C^{\ast }\,\,\,\forall R>R_1. \label{CC2}
\end{equation}%
By using a result due to Amann \& Crandall \cite[Lemma 4]{Amann}, there is
an increasing function $\gamma _{0}:[0,+\infty )\rightarrow \lbrack 0,\infty
)$, depending only of $\Omega $, $p$ and $N$, and satisfying
\begin{equation*}
\Vert u\Vert _{W^{2,p}(\Omega )}\leq \gamma _{0}(\Vert B_{R}\Vert _{\infty
}).
\end{equation*}%
Combining the last inequality with (\ref{CC1}) and (\ref{CC2}), we get
\begin{equation*}
\Vert u\Vert _{C^{1,\alpha }(\overline{\Omega} )}\leq C\gamma _{0}(C^{\ast
}),
\end{equation*}%
for some $C>0$. Fixing
\begin{equation*}
K_1=C\gamma _{0}(C^{\ast }),
\end{equation*}%
we derive that
\begin{equation*}
|\frac{\partial u(x)}{\partial x_{i}}|\leq K_1\,\,\,\forall x\in \overline{%
\Omega }\,\,\ \mbox{and}\,\,\,i=1,2,..,N.
\end{equation*}%
Thereby,
\begin{equation*}
|\nabla u(x)|\leq NK_1\,\,\,\forall x\in \overline{\Omega },
\end{equation*}%
implying that
\begin{equation*}
\max_{x\in \overline{\Omega }}|\nabla u(x)|\leq NK_1.
\end{equation*}%
Fixing $R_{2}=NK_1$ and $R \geq R^{*}=\max\{R_1,R_2\}$, it follows that
\begin{equation*}
\max_{x\in \overline{\Omega }}|\nabla u(x)|\leq R,
\end{equation*}%
showing that $u$ is a solution of $(P)$ if $R \geq R^{*}$.
\section{Applications}
In this section, we will present two situations in which our main theorem works.
\vspace{0.5 cm}
\noindent {\bf Application 1:} Our first application is the following problem
\begin{equation} \label{example1}
\left\{
\begin{array}{l}
-M\left(\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u = \lambda |u|^{q} + |u|^{p}+\mu |\nabla u|^{q} ~~ \mbox{in} ~~ \Omega , \\
u(x)>0 ~~ \mbox{in} ~~ \Omega \\
u = 0 ~~ \mbox{on} ~~ \partial \Omega .
\end{array}
\right.
\end{equation}
where $\lambda$ is a positive parameter, $0<q<1<p<+\infty$ and $M$ verifies conditions $(M_1)-(M_2)$.
Here, we must observe that the above problem is a nonlocal version of a well known result due to Ambrosetti, Brezis \& Cerami \cite{abc} with an additional gradient term $|\nabla u|^{q}$.
We begin observing that it is easy to find a positive function $\overline{u}$ verifying the inequality
$$
-\Delta \overline{u} \geq \frac{1}{m}(\lambda \overline{u}^{q} + \overline{u}^{p}+\mu |\nabla \overline{u}|^{q})
$$
if $\lambda , \mu$ are small enough. It is enough to follow the ideas found in Ambrosetti, Brezis \& Cerami \cite{abc}. Indeed, let $0<e$ in $\Omega$, $e\in C^{1}(\overline{\Omega})$ the only solution of
\begin{equation} \label{e 1 solution}
\left\{
\begin{array}{rclcc}
-\Delta e & = & 1 & \mbox{in} & \Omega , \\
e & = & 0 & \mbox{on} & \partial \Omega .
\end{array}
\right.
\end{equation}
We now take $S>0$ such that
\begin{equation}\label{S large}
m\geq \frac{1}{S^{1-q}}(\lambda \|e\|_{\infty}^{q}+\mu \||\nabla e|\|_{\infty}^{q})+S^{p-1}\|e\|_{\infty}^{p}.
\end{equation}
A straightforward computation shows that there is $0< \lambda^{\ast}$ such that for $0<\lambda , \mu <\lambda^{\ast}$ there is $S>0$ such that the inequality (\ref{S large}) holds true. Hence we can take $\overline{u}:= Se\in W^{1, \infty}(\Omega )$, $S$ as above, so that the first inequality in the Theorem \ref{main result} is satisfied.
Now, fixed $\lambda , \mu >0$ as before, we consider the family $(u_\delta)$ with $u_\delta=\delta \varphi_1$, $\varphi_1$ is a positive eigenfunction associated with the principal eigenvalue $\lambda_1$ of $(-\Delta , H_{0}^{1}(\Omega))$. A simple computation also gives
for all $\alpha >0$ fixed, there exist $\delta^*>0$ such that
$$
-\Delta u_\delta \leq \frac{1}{\alpha}(\lambda u_\delta^{q}+u_\delta^{p}+\mu |\nabla u_{\delta}|^{q}) ~~~\mbox{in} ~~ \Omega .
$$
As it is well known, we can consider $\delta >0$ sufficiently small such that $u_{\delta}\leq \overline{u}$.
From the above commentaries, we can apply Theorem \ref{main result} to prove the existence of a weak solution $u$ for (\ref{example1}) satisfying $u_{\delta}\leq u\leq \overline{u}$.
\vspace{0.5 cm}
\noindent {\bf Application 2:} ~~ Our next application is concerning the problem
\begin{equation} \label{example2}
\left\{
\begin{array}{l}
-M\left(\int_{\Omega}|\nabla u|^{2}dx\right)\Delta u = Au^{q}(B-u)+|\nabla u|^{\eta} ~~ \mbox{in} ~~ \Omega , \\
u = 0 ~~ \mbox{on} ~~\partial \Omega
\end{array}
\right.
\end{equation}
where $A,B$ are positive constants satisfying some properties which will be established later, $\eta \in (1,2]$ and $q \in (0,1)$. We will find a solution $u$ satisfying $0<u\leq B$ in $\Omega$. First of all, let us consider the continuous function $f:\Omega \times \mathbb{R} \times \mathbb{R}^{N}\rightarrow [0,+\infty)$ defined as
$$
f(x,t,y)=
\left\{
\begin{array}{l}
|y|^{\eta}, ~~ \mbox{if} ~~ t \geq B \\
\mbox{}\\
At^{q}(B-t)+|y|^{\eta}, ~~ \mbox{if} ~~0 \leq t \leq B \\
\mbox{}\\
|y|^{\eta} , ~~ \mbox{if} ~~ t \leq 0.
\end{array}
\right.
$$
It is clear that the function $\overline{u}\equiv B$ belongs to $W^{1, \infty}(\Omega )$ and satisfies the assumption (\ref{weak supersolution}) in the Theorem \ref{main result}.
If $\lambda_{1}$ is the principal eigenvalue of $(-\Delta , H_{0}^{1}(\Omega ))$ associated to the eigenfunction $\varphi_{1}>0$ in $\Omega$, for each $\alpha >0 $, there is $\delta^* >0$ such that
$$
\lambda_{1}\delta \varphi_{1}\leq \frac{A}{\alpha}(\delta \varphi_{1})^q(B-\delta\varphi_{1})+\frac{1}{\alpha}|\nabla (\delta\varphi_{1})|^{\eta} ~~~~ \forall \delta \in (0, \delta^*].
$$
Taking $\underline{u}_{\delta}:= \delta \varphi_{1}$ we get $\|\underline{u}_{\delta} \|_{1, \infty}\rightarrow 0$ as $\delta \rightarrow 0$, $\underline{u}_{\delta} \leq \overline{u}\equiv B $ in $\Omega$, if $\delta >0$ is small enough and a straightforward calculation shows that the inequality (\ref{weak subsolution}) holds true. Hence, for $\delta$ sufficiently small, problem (P) possesses a weak solution $u$ satisfying $\underline{u}_{\delta}\leq u \leq B$. Consequently, such a function is a solution of the problem (\ref{example2}).
\begin{rmk}
For some applications concerning the quasilinear problem (P), with $M\equiv 1$, still using a sub and supersolution approach, the reader may consult Xavier \cite{xavier} and the references therein.
\end{rmk}
|
\section{Introduction}
\label{intro}
Control theory is the branch of engineering that focuses on manipulating `open systems'---systems
with inputs and outputs---to achieve desired goals. In control theory, `signal-flow diagrams' are
used to describe linear ways of manipulating signals, which we will take here to be smooth
real-valued functions of time \cite{Friedland}. For a category theorist, at least, it is natural to
treat signal-flow diagrams as string diagrams in a symmetric monoidal category \cite{JS1,JS2}. This
forces some small changes of perspective, which we discuss below, but more important is the
question: \emph{which symmetric monoidal category?}
We shall argue that the answer is: the category \(\mathrm{FinRel}_k\) of finite-dimensional vector spaces over a
certain field \(k\), but with \emph{linear relations} rather than linear maps as morphisms, and
\emph{direct sum} rather than tensor product providing the symmetric monoidal structure. We use the
field \(k = {\mathbb R}(s)\) consisting of rational functions in one real variable \(s\). This variable has
the meaning of differentation. A linear relation from \(k^m\) to \(k^n\) is thus a system of linear
constant-coefficient ordinary differential equations relating \(m\) `input' signals and \(n\)
`output' signals.
Our main goal is to provide a complete `generators and relations' picture of this symmetric monoidal
category, with the generators being familiar components of signal-flow diagrams. It turns out that
the answer has an intriguing but mysterious connection to ideas that are familiar in the
diagrammatic approach to quantum theory. Quantum theory also involves linear algebra, but it uses
linear maps between Hilbert spaces as morphisms, and the tensor product of Hilbert spaces provides
the symmetric monoidal structure.
We hope that the category-theoretic viewpoint on signal-flow diagrams will shed new light on control
theory. However, in this paper we only lay the groundwork. In Section \ref{sigflow} we introduce
signal-flow diagrams and summarize our main results. In Section \ref{finvect} we use
signal-flow diagrams to give a presentation of \(\mathrm{FinVect}_k\), the symmetric monoidal
category of finite-dimensional vector spaces and linear maps. In Section \ref{finrel} we use them
to give a presentation of \(\mathrm{FinRel}_k\). In Section \ref{example} we discuss a well-known example from
control theory: an inverted pendulum on a cart. Finally, in Section \ref{conclusions} we compare
our results to subsequent work of Bonchi--Soboci\'nski--Zanasi \cite{BSZ1,BSZ2} and Wadsley--Woods
\cite{WW}.
\section{Signal-flow diagrams}
\label{sigflow}
There are several basic operations that one wants to perform when manipulating signals. The
simplest is multiplying a signal by a scalar. A signal can be amplified by a constant factor:
\[ f \mapsto cf \]
where \(c \in {\mathbb R}\). We can write this as a string diagram:
\begin{center}
\begin{tikzpicture}[thick]
\node[coordinate] (in) [label=\(f\)] {};
\node [multiply] (mult) [below of=in] {\(c\)};
\node[coordinate] (out) [below of=mult, label={[shift={(0,-0.8)}]\(cf\)}] {};
\draw (in) -- (mult) -- (out);
\end{tikzpicture}
\end{center}
Here the labels \(f\) and \(cf\) on top and bottom are just for explanatory purposes and not really
part of the diagram. Control theorists often draw arrows on the wires, but this is unnecessary from
the string diagram perspective. Arrows on wires are useful to distinguish objects from their
duals, but ultimately we will obtain a compact closed category where each object is its own dual, so
the arrows can be dropped. What we really need is for the box denoting scalar multiplication to
have a clearly defined input and output. This is why we draw it as a triangle. Control theorists
often use a rectangle or circle, using arrows on wires to indicate which carries the input \(f\) and
which the output \(c f\).
A signal can also be integrated with respect to the time variable:
\[ f \mapsto \int f .\]
Mathematicians typically take differentiation as fundamental, but engineers sometimes prefer
integration, because it is more robust against small perturbations. In the end it will not matter
much here. We can again draw integration as a string diagram:
\begin{center}
\begin{tikzpicture}[thick]
\node[coordinate] (in) [label=\(f\)] {};
\node [integral] (int) [below of=in] {\(\int\)};
\node[coordinate] (out) [below of=int, label={[shift={(0.0,-0.9)}]\(\int f\)}] {};
\draw (in) -- (int) -- (out);
\end{tikzpicture}
\end{center}
Since this looks like the diagram for scalar multiplication, it is natural to extend \({\mathbb R}\) to
\({\mathbb R}(s)\), the field of rational functions of a variable \(s\) which stands for differentiation.
Then differentiation becomes a special case of scalar multiplication, namely multiplication by
\(s\), and integration becomes multiplication by \(1/s\). Engineers accomplish the same effect with
Laplace transforms, since differentiating a signal \(f\) is equivalent to multiplying its Laplace
transform
\[ (\mathcal{L}f)(s) = \int_0^\infty f(t) e^{-st} \,dt \]
by the variable \(s\). Another option is to use the Fourier transform: differentiating \(f\) is
equivalent to multiplying its Fourier transform
\[ (\mathcal{F}f)(\omega) = \int_{-\infty}^\infty f(t) e^{-i\omega t}\, dt \]
by \(-i\omega\). Of course, the function \(f\) needs to be sufficiently well-behaved to justify
calculations involving its Laplace or Fourier transform. At a more basic level, it also requires
some work to treat integration as the two-sided inverse of differentiation. Engineers do this by
considering signals that vanish for \(t < 0\), and choosing the antiderivative that vanishes under
the same condition. Luckily all these issues can be side-stepped in a formal treatment of
signal-flow diagrams: we can simply treat signals as living in an unspecified vector space over the
field \({\mathbb R}(s)\). The field \({\mathbb C}(s)\) would work just as well, and control theory relies heavily on
complex analysis. In most of this paper we work over an arbitrary field \(k\).
The simplest possible signal processor is a rock, which takes the `input' given by the force \(F\)
on the rock and produces as `output' the rock's position \(q\). Thanks to Newton's second law
\(F=ma\), we can describe this using a signal-flow diagram:
\begin{center}
\scalebox{0.8}{
\begin{tikzpicture}[thick]
\node[coordinate] (q) [label={[shift={(0,-0.6)}]\(q\)}] {};
\node [integral] (diff) [above of=q] {\(\int\)};
\node (v) [above of=diff, label={[shift={(0.4,-0.5)}]\(v\)}] {};
\node [integral] (dot) [above of=v] {\(\int\)};
\node (a) [above of=dot, label={[shift={(0.4,-0.5)}]\(a\)}] {};
\node [multiply] (m) [above of=a] {\(\frac{1}{m}\)};
\node[coordinate] (F) [above of=m, label={[shift={(0,0)}]\(F\)}] {};
\draw (F) -- (m) -- (dot) -- (diff) -- (q);
\end{tikzpicture}
}
\end{center}
Here composition of morphisms is drawn in the usual way, by attaching the output wire of one
morphism to the input wire of the next.
To build more interesting machines we need more building blocks, such as addition:
\[ + \maps (f,g) \mapsto f + g \]
and duplication:
\[ \Delta \maps f \mapsto (f,f) \]
When these linear maps are written as matrices, their matrices are transposes of each other. This
is reflected in the string diagrams for addition and duplication:
\begin{center}
\begin{tikzpicture}[thick]
\node[plus] (adder) {};
\node (f) at (-0.5,1.35) {\(f\)};
\node (g) at (0.5,1.35) {\(g\)};
\node (out) [below of=adder] {\(f+g\)};
\draw (f) .. controls +(-90:0.6) and +(120:0.6) .. (adder.left in);
\draw (g) .. controls +(-90:0.6) and +(60:0.6) .. (adder.right in);
\draw (adder) -- (out);
\end{tikzpicture}
\hspace{2cm}
\begin{tikzpicture}[thick]
\node[delta] (dupe){};
\node (o1) at (-0.5,-1.35) {\(f\)};
\node (o2) at (0.5,-1.35) {\(f\)};
\node (in) [above of=dupe] {\(f\)};
\draw (o1) .. controls +(90:0.6) and +(-120:0.6) .. (dupe.left out);
\draw (o2) .. controls +(90:0.6) and +(-60:0.6) .. (dupe.right out);
\draw (in) -- (dupe);
\end{tikzpicture}
\end{center}
The second is essentially an upside-down version of the first. However, we draw addition as a dark
triangle and duplication as a light one because we will later want another way to `turn addition
upside-down' that does \emph{not} give duplication. As an added bonus, a light upside-down triangle
resembles the Greek letter \(\Delta\), the usual symbol for duplication.
While they are typically not considered worthy of mention in control theory, for completeness we
must include two other building blocks. One is the zero map from \(\{0\}\) to our field \(k\),
which we denote as \(0\) and draw as follows:
\begin{center}
\begin{tikzpicture}[thick]
\node (out1) {\(0\)};
\node [zero] (ins1) at (0,1) {};
\draw (out1) -- (ins1);
\end{tikzpicture}
\end{center}
The other is the zero map from \(k\) to \(\{0\}\), sometimes called `deletion', which we denote as
\(!\) and draw thus:
\begin{center}
\begin{tikzpicture}[thick]
\node (in1) {\(f\)};
\node [bang] (del1) at (0,-1) {};
\draw (in1) -- (del1);
\end{tikzpicture}
\end{center}
Just as the matrices for addition and duplication are transposes of each other, so are the matrices
for zero and deletion, though they are rather degenerate, being \(1 \times 0\) and \(0 \times 1\)
matrices, respectively. Addition and zero make \(k\) into a commutative monoid, meaning that the
following relations hold:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.74cm]
\node [plus] (summer) {};
\node [coordinate] (sum) [below of=summer] {};
\node [coordinate] (Lsum) [above left of=summer] {};
\node [zero] (insert) [above of=Lsum, shift={(0,-0.35)}] {};
\node [coordinate] (Rsum) [above right of=summer] {};
\node [coordinate] (sumin) [above of=Rsum] {};
\node (equal) [right of=Rsum, shift={(0,-0.26)}] {\(=\)};
\node [coordinate] (in) [right of=equal, shift={(0,1)}] {};
\node [coordinate] (out) [right of=equal, shift={(0,-1)}] {};
\draw (insert) .. controls +(270:0.3) and +(120:0.3) .. (summer.left in)
(summer.right in) .. controls +(60:0.6) and +(270:0.6) .. (sumin)
(summer) -- (sum) (in) -- (out);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (uradder) {};
\node [plus] (adder) [below of=uradder, shift={(-0.35,0)}] {};
\node [coordinate] (urm) [above of=uradder, shift={(-0.35,0)}] {};
\node [coordinate] (urr) [above of=uradder, shift={(0.35,0)}] {};
\node [coordinate] (left) [left of=urm] {};
\draw (adder.right in) .. controls +(60:0.2) and +(270:0.1) .. (uradder.io)
(uradder.right in) .. controls +(60:0.35) and +(270:0.3) .. (urr)
(uradder.left in) .. controls +(120:0.35) and +(270:0.3) .. (urm)
(adder.left in) .. controls +(120:0.75) and +(270:0.75) .. (left)
(adder.io) -- +(270:0.5);
\node (eq) [right of=uradder, shift={(0,-0.25)}] {\(=\)};
\node [plus] (ulsummer) [right of=eq, shift={(0,0.25)}] {};
\node [plus] (summer) [below of=ulsummer, shift={(0.35,0)}] {};
\node [coordinate] (ulm) [above of=ulsummer, shift={(0.35,0)}] {};
\node [coordinate] (ull) [above of=ulsummer, shift={(-0.35,0)}] {};
\node [coordinate] (right) [right of=ulm] {};
\draw (summer.left in) .. controls +(120:0.2) and +(270:0.1) .. (ulsummer.io)
(ulsummer.left in) .. controls +(120:0.35) and +(270:0.3) .. (ull)
(ulsummer.right in) .. controls +(60:0.35) and +(270:0.3) .. (ulm)
(summer.right in) .. controls +(60:0.75) and +(270:0.75) .. (right)
(summer.io) -- +(270:0.5);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (twadder) {};
\node [coordinate] (twout) [below of=twadder] {};
\node [coordinate] (twR) [above right of=twadder, shift={(-0.2,0)}] {};
\node (cross) [above of=twadder] {};
\node [coordinate] (twRIn) [above left of=cross, shift={(0,0.3)}] {};
\node [coordinate] (twLIn) [above right of=cross, shift={(0,0.3)}] {};
\draw (twadder.right in) .. controls +(60:0.35) and +(-45:0.25) .. (cross)
.. controls +(135:0.2) and +(270:0.4) .. (twRIn);
\draw (twadder.left in) .. controls +(120:0.35) and +(-135:0.25) .. (cross.center)
.. controls +(45:0.2) and +(270:0.4) .. (twLIn);
\draw (twout) -- (twadder);
\node (eq) [right of=twR] {\(=\)};
\node [coordinate] (L) [right of=eq] {};
\node [plus] (adder) [below right of=L] {};
\node [coordinate] (out) [below of=adder] {};
\node [coordinate] (R) [above right of=adder] {};
\node (cross) [above left of=R] {};
\node [coordinate] (LIn) [above left of=cross] {};
\node [coordinate] (RIn) [above right of=cross] {};
\draw (adder.left in) .. controls +(120:0.7) and +(270:0.7) .. (LIn)
(adder.right in) .. controls +(60:0.7) and +(270:0.7) .. (RIn)
(out) -- (adder);
\end{tikzpicture}
}
\end{center}
The equation at right is the commutative law, and the crossing of strands is the `braiding'
\[ B \maps (f,g) \mapsto (g,f) \]
by which we switch two signals. In fact this braiding is a `symmetry', so it does not matter which
strand goes over which:
\begin{center}
\begin{tikzpicture}[thick, node distance=0.5cm]
\node (fstart) {\(f\)};
\node [coordinate] (ftop) [below of=fstart] {};
\node (center) [below right of=ftop] {};
\node [coordinate] (fout) [below right of=center] {};
\node (fend) [below of=fout] {\(f\)};
\node [coordinate] (gtop) [above right of=center] {};
\node (gstart) [above of=gtop] {\(g\)};
\node [coordinate] (gout) [below left of=center] {};
\node (gend) [below of=gout] {\(g\)};
\draw [rounded corners] (fstart) -- (ftop) -- (center) --
(fout) -- (fend) (gstart) -- (gtop) -- (gout) -- (gend);
\end{tikzpicture}
\hspace{1em}
\raisebox{3em}{=}
\hspace{1em}
\begin{tikzpicture}[thick, node distance=0.5cm]
\node (fstart) {\(f\)};
\node [coordinate] (ftop) [below of=fstart] {};
\node (center) [below right of=ftop] {};
\node [coordinate] (fout) [below right of=center] {};
\node (fend) [below of=fout] {\(f\)};
\node [coordinate] (gtop) [above right of=center] {};
\node (gstart) [above of=gtop] {\(g\)};
\node [coordinate] (gout) [below left of=center] {};
\node (gend) [below of=gout] {\(g\)};
\draw [rounded corners] (fstart) -- (ftop) -- (fout) -- (fend)
(gstart) -- (gtop) -- (center) -- (gout) -- (gend);
\end{tikzpicture}
\end{center}
Dually, duplication and deletion make \(k\) into a cocommutative comonoid. This means that if we
reflect the equations obeyed by addition and zero across the horizontal axis and turn dark
operations into light ones, we obtain another set of valid equations:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.74cm]
\node [delta] (dupe) {};
\node [coordinate] (top) [above of=dupe] {};
\node [coordinate] (Ldub) [below left of=dupe] {};
\node [bang] (delete) [below of=Ldub, shift={(0,0.35)}] {};
\node [coordinate] (Rdub) [below right of=dupe] {};
\node [coordinate] (dubout) [below of=Rdub] {};
\node (equal) [right of=Rdub, shift={(0,0.26)}] {\(=\)};
\node [coordinate] (in) [right of=equal, shift={(0,1)}] {};
\node [coordinate] (out) [right of=equal, shift={(0,-1)}] {};
\draw (delete) .. controls +(90:0.3) and +(240:0.3) .. (dupe.left out)
(dupe.right out) .. controls +(300:0.6) and +(90:0.6) .. (dubout)
(dupe) -- (top) (in) -- (out);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [delta] (lrduper) {};
\node [delta] (duper) [above of=lrduper, shift={(-0.35,0)}] {};
\node [coordinate](lrm) [below of=lrduper, shift={(-0.35,0)}] {};
\node [coordinate](lrr) [below of=lrduper, shift={(0.35,0)}] {};
\node [coordinate](left) [left of=lrm] {};
\draw (duper.right out) .. controls +(300:0.2) and +(90:0.1) .. (lrduper.io)
(lrduper.right out) .. controls +(300:0.35) and +(90:0.3) .. (lrr)
(lrduper.left out) .. controls +(240:0.35) and +(90:0.3) .. (lrm)
(duper.left out) .. controls +(240:0.75) and +(90:0.75) .. (left)
(duper.io) -- +(90:0.5);
\node (eq) [right of=lrduper, shift={(0,0.25)}] {\(=\)};
\node [delta] (lldubber) [right of=eq, shift={(0,-0.25)}] {};
\node [delta] (dubber) [above of=lldubber, shift={(0.35,0)}] {};
\node [coordinate] (llm) [below of=lldubber, shift={(0.35,0)}] {};
\node [coordinate] (lll) [below of=lldubber, shift={(-0.35,0)}] {};
\node [coordinate] (right) [right of=llm] {};
\draw (dubber.left out) .. controls +(240:0.2) and +(90:0.1) .. (lldubber.io)
(lldubber.left out) .. controls +(240:0.35) and +(90:0.3) .. (lll)
(lldubber.right out) .. controls +(300:0.35) and +(90:0.3) .. (llm)
(dubber.right out) .. controls +(300:0.75) and +(90:0.75) .. (right)
(dubber.io) -- +(90:0.5);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [coordinate] (twtop) {};
\node [delta] (twdupe) [below of=twtop] {};
\node [coordinate] (twR) [below right of=twdupe, shift={(-0.2,0)}] {};
\node (cross) [below of=twdupe] {};
\node [coordinate] (twROut) [below left of=cross, shift={(0,-0.3)}] {};
\node [coordinate] (twLOut) [below right of=cross, shift={(0,-0.3)}] {};
\draw (twdupe.left out) .. controls +(240:0.35) and +(135:0.25) .. (cross)
.. controls +(-45:0.2) and +(90:0.4) .. (twLOut)
(twdupe.right out) .. controls +(300:0.35) and +(45:0.25) .. (cross.center)
.. controls +(-135:0.2) and +(90:0.4) .. (twROut)
(twtop) -- (twdupe);
\node (eq) [right of=twR] {\(=\)};
\node [coordinate] (L) [right of=eq] {};
\node [delta] (dupe) [above right of=L] {};
\node [coordinate] (top) [above of=dupe] {};
\node [coordinate] (R) [below right of=dupe] {};
\node (uncross) [below left of=R] {};
\node [coordinate] (LOut) [below left of=uncross] {};
\node [coordinate] (ROut) [below right of=uncross] {};
\draw (dupe.left out) .. controls +(240:0.7) and +(90:0.7) .. (LOut)
(dupe.right out) .. controls +(300:0.7) and +(90:0.7) .. (ROut)
(top) -- (dupe);
\end{tikzpicture}
}
\end{center}
There are also relations between the monoid and comonoid operations. For example, adding two
signals and then duplicating the result gives the same output as duplicating each signal and then
adding the results:
\begin{center}
\begin{tikzpicture}[thick]
\node[plus] (adder) {};
\node [coordinate] (f) [above of=adder, shift={(-0.4,-0.325)}, label={\(f\)}] {\(f\)};
\node [coordinate] (g) [above of=adder, shift={(0.4,-0.325)}, label={\(g\)}] {\(g\)};
\node[delta] (dupe) [below of=adder, shift={(0,0.25)}] {};
\node [coordinate] (outL) [below of=dupe, shift={(-0.4,0.325)}, label={[shift={(-0.2,-0.6)}]\(f+g\)}] {};
\node [coordinate] (outR) [below of=dupe, shift={(0.4,0.325)}, label={[shift={(0.15,-0.6)}]\(f+g\)}] {};
\draw (adder.io) -- (dupe.io)
(f) .. controls +(270:0.4) and +(120:0.25) .. (adder.left in)
(adder.right in) .. controls +(60:0.25) and +(270:0.4) .. (g)
(dupe.left out) .. controls +(240:0.25) and +(90:0.4) .. (outL)
(dupe.right out) .. controls +(300:0.25) and +(90:0.4) .. (outR);
\end{tikzpicture}
\raisebox{4em}{=}
\hspace{1em}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (addL) {};
\node (cross) [above right of=addL, shift={(-0.1,-0.0435)}] {};
\node [plus] (addR) [below right of=cross, shift={(-0.1,0.0435)}] {};
\node [delta] (dupeL) [above left of=cross, shift={(0.1,-0.0435)}] {};
\node [delta] (dupeR) [above right of=cross, shift={(-0.1,-0.0435)}] {};
\node [coordinate] (f) [above of=dupeL, label={\(f\)}] {};
\node [coordinate] (g) [above of=dupeR, label={\(g\)}] {};
\node [coordinate] (sum1) [below of=addL, shift={(0,0.2)}, label={[shift={(-0.2,-0.6)}]\(f+g\)}] {};
\node [coordinate] (sum2) [below of=addR, shift={(0,0.2)}, label={[shift={(0.15,-0.6)}]\(f+g\)}] {};
\path
(addL) edge (sum1) (addL.right in) edge (dupeR.left out) (addL.left in) edge [bend left=30] (dupeL.left out)
(addR) edge (sum2) (addR.left in) edge (cross) (addR.right in) edge [bend right=30] (dupeR.right out)
(dupeL) edge (f)
(dupeL.right out) edge (cross)
(dupeR) edge (g);
\end{tikzpicture}
\end{center}
This diagram is familiar in the theory of Hopf algebras, or more generally bialgebras. Here it is
an example of the fact that the monoid operations on \(k\) are comonoid homomorphisms---or
equivalently, the comonoid operations are monoid homomorphisms. We summarize this situation by
saying that \(k\) is a \Define{bimonoid}.
So far all our string diagrams denote linear maps. We can treat these as morphisms in the category
\( \mathrm{FinVect}_k \), where objects are finite-dimensional vector spaces over a field \(k\) and morphisms
are linear maps. This category is equivalent to a skeleton where the only objects are vector spaces
\(k^n\) for \(n \ge 0\), and then morphisms can be seen as \(n \times m\) matrices. The space of
signals is a vector space \(V\) over \(k\) which may not be finite-dimensional, but this does not
cause a problem: an \(n \times m\) matrix with entries in \(k\) still defines a linear map from
\(V^n\) to \(V^m\) in a functorial way.
In applications of string diagrams to quantum theory \cite{BS,CP}, we make \(\mathrm{FinVect}_k\) into a
symmetric monoidal category using the tensor product of vector spaces. In control theory, we
instead make \(\mathrm{FinVect}_k\) into a symmetric monoidal category using the \emph{direct sum} of vector
spaces. In Lemma~\ref{gensvk} we prove that for any field \(k\), \(\mathrm{FinVect}_k\) with direct sum
is generated as a symmetric monoidal category by the one object \(k\) together with these morphisms:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node[coordinate] (in) at (0,2) {};
\node [multiply] (mult) at (0,1) {\(c\)};
\node[coordinate] (out) at (0,0) {};
\draw (in) -- (mult) -- (out);
\end{tikzpicture}
\hspace{3 em}
\begin{tikzpicture}[thick]
\node[plus] (adder) at (0,0.85) {};
\node[coordinate] (f) at (-0.5,1.5) {};
\node[coordinate] (g) at (0.5,1.5) {};
\node[coordinate] (out) at (0,0) {};
\node [coordinate] (pref) at (-0.5,2) {};
\node [coordinate] (preg) at (0.5,2) {};
\draw[rounded corners] (pref) -- (f) -- (adder.left in);
\draw[rounded corners] (preg) -- (g) -- (adder.right in);
\draw (adder) -- (out);
\end{tikzpicture}
\hspace{2em}
\begin{tikzpicture}[thick]
\node[delta] (dupe) at (0,1.15) {};
\node[coordinate] (o1) at (-0.5,0.5) {};
\node[coordinate] (o2) at (0.5,0.5) {};
\node[coordinate] (in) at (0,2) {};
\node [coordinate] (posto1) at (-0.5,0) {};
\node [coordinate] (posto2) at (0.5,0) {};
\draw[rounded corners] (posto1) -- (o1) -- (dupe.left out);
\draw[rounded corners] (posto2) -- (o2) -- (dupe.right out);
\draw (in) -- (dupe);
\end{tikzpicture}
\hspace{3em}
\begin{tikzpicture}[thick]
\node[hole] (heightHolder) at (0,2) {};
\node [coordinate] (out) at (0,0) {};
\node [zero] (del) at (0,1) {};
\draw (del) -- (out);
\end{tikzpicture}
\hspace{2em}
\begin{tikzpicture}[thick]
\node[coordinate] (in) at (0,2) {};
\node [bang] (mult) at (0,1) {};
\node [hole] (heightHolder) at (0,0) {};
\draw (in) -- (mult);
\end{tikzpicture}
}
\end{center}
where \(c \in k\) is arbitrary.
However, these generating morphisms obey some unexpected relations! For example, we have:
\begin{center}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (UpUpLeft) at (-0.4,-0.1) {};
\node [coordinate] (UpLeft) at (-0.4,-0.6) {};
\node (mid) at (0,-1) {};
\node [coordinate] (DownRight) at (0.4,-1.4) {};
\node (DownDownRight) at (0.4,-1.9) {};
\node [coordinate] (UpRight) at (0.4,-0.6) {};
\node (UpUpRight) at (0.4,-0.1) {};
\node [coordinate] (DownLeft) at (-0.4,-1.4) {};
\node (DownDownLeft) at (-0.4,-1.9) {};
\draw [rounded corners=2mm] (UpUpLeft) -- (UpLeft) -- (mid) --
(DownRight) -- (DownDownRight) (UpUpRight) -- (UpRight) -- (DownLeft) -- (DownDownLeft);
\begin{scope}[font=\fontsize{20}{20}\selectfont]
\node (equals) at (1.75,-1) {\scalebox{0.65}{\(=\)}};
\end{scope}
\node [coordinate] (sum2L) at (3.5,0) {};
\node [multiply] (neg) [above of=sum2L] {\(\scriptstyle{-1}\)};
\node [coordinate] (dupe1L) [above of=neg] {};
\node [delta] (dupe1) [above right of=dupe1L, shift={(-0.1,0)}] {};
\node (in1) [above of=dupe1] {};
\node [plus] (sum1) [below right of=dupe1] {};
\node [coordinate] (sum1R) [above right of=sum1, shift={(-0.1,0)}] {};
\node (in2) [above of=sum1R] {};
\node [delta] (dupe2) [below of=sum1, shift={(0,-0.85)}] {};
\node [plus] (sum2) [below right of=sum2L, shift={(-0.1,0)}] {};
\node [coordinate] (dupe2R) [below right of=dupe2, shift={(0.4,-0.6)}] {};
\node [delta] (dupe3) [below of=sum2] {};
\node [coordinate] (dupe3L) [below left of=dupe3, shift={(0.1,0)}] {};
\node [multiply] (neg1) [below right of=dupe3, shift={(-0.28,-0.5)}] {\(\scriptstyle{-1}\)};
\node [plus] (sum3) [below right of=neg1, shift={(-0.2,-0.65)}] {};
\node [coordinate] (sum3R) [above right of=sum3, shift={(0,0.2)}] {};
\node (out2) [below of=sum3] {};
\node (out1) [below of=dupe3L, shift={(0,-1.75)}] {};
\draw (in1) -- (dupe1);
\draw (dupe1.left out) .. controls +(240:0.5) and +(90:0.5) .. (neg.90);
\draw (dupe1.right out) -- (sum1.left in);
\draw (in2) .. controls +(270:0.5) and +(60:0.5) .. (sum1.right in);
\draw (sum1) -- (dupe2);
\draw (dupe2.right out) .. controls (dupe2R) and (sum3R) .. (sum3.right in);
\draw (dupe2.left out) -- (sum2.right in);
\draw (neg.io) .. controls +(270:0.3) and +(120:0.3) .. (sum2.left in);
\draw (sum2) -- (dupe3);
\draw (dupe3.left out) .. controls +(240:0.7) and +(90:1) .. (out1);
\draw (dupe3.right out) .. controls +(300:0.3) and +(90:0.3) .. (neg1.90);
\draw (neg1.io) .. controls +(270:0.2) and +(120:0.2) .. (sum3.left in);
\draw (sum3) -- (out2);
\end{tikzpicture}
\end{center}
Thus, it is important to find a complete set of relations obeyed by these generating
morphisms, thus obtaining a presentation of \(\mathrm{FinVect}_k\) as a symmetric monoidal
category. We do this in Theorem~\ref{presvk}. In brief, these relations say:
\begin{enumerate}
\item \( (k, +, 0, \Delta, !) \) is a bicommutative bimonoid;
\item the rig operations of \(k\) can be recovered from the generating morphisms;
\item all the generating morphisms commute with scalar multiplication.
\end{enumerate}
Here item (2) means that \(+\), \(\cdot\), \(0\) and \(1\) in the field \(k\) can be expressed in
terms of signal-flow diagrams as follows:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (bctop) {};
\node [multiply] (bc) [below of=bctop, shift={(0,-0.59)}] {\(\scriptstyle{b+c}\)};
\node (bcbottom) [below of=bc, shift={(0,-0.59)}] {};
\draw (bctop) -- (bc) -- (bcbottom);
\node (eq) [right of=bc, shift={(0.15,0)}] {\(=\)};
\node [multiply] (b) [right of=eq, shift={(0,0.1)}] {\(\scriptstyle{b}\)};
\node [delta] (dupe) [above right of=b, shift={(-0.2,0.1)}] {};
\node (top) [above of=dupe, shift={(0,-0.2)}] {};
\node [multiply] (c) [below right of=dupe, shift={(-0.2,-0.1)}] {\(\scriptstyle{c}\)};
\node [plus] (adder) [below right of=b, shift={(-0.2,-0.3)}] {};
\node (out) [below of=adder, shift={(0,0.2)}] {};
\draw
(dupe.left out) .. controls +(240:0.15) and +(90:0.15) .. (b.90)
(dupe.right out) .. controls +(300:0.15) and +(90:0.15) .. (c.90)
(top) -- (dupe.io)
(adder.io) -- (out)
(adder.left in) .. controls +(120:0.15) and +(270:0.15) .. (b.io)
(adder.right in) .. controls +(60:0.15) and +(270:0.15) .. (c.io);
\end{tikzpicture}
\hspace{0.8cm}
\begin{tikzpicture}[-, thick]
\node (top) {};
\node [multiply] (c) [below of=top] {\(c\)};
\node [multiply] (b) [below of=c] {\(b\)};
\node (bottom) [below of=b] {};
\draw (top) -- (c) -- (b) -- (bottom);
\node (eq) [left of=b, shift={(0.2,0.5)}] {\(=\)};
\node (bctop) [left of=top, shift={(-0.6,0)}] {};
\node [multiply] (bc) [left of=eq, shift={(0.2,0)}] {\(bc\)};
\node (bcbottom) [left of=bottom, shift={(-0.6,0)}] {};
\draw (bctop) -- (bc) -- (bcbottom);
\end{tikzpicture}
\hspace{0.8cm}
\raisebox{2em}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (top) {};
\node [multiply] (one) [below of=top] {1};
\node (bottom) [below of=one] {};
\draw (top) -- (one) -- (bottom);
\node (eq) [right of=one] {\(=\)};
\node (topid) [right of=top, shift={(0.6,0)}] {};
\node (botid) [right of=bottom, shift={(0.6,0)}] {};
\draw (topid) -- (botid);
\end{tikzpicture}
}
\hspace{0.6cm}
\raisebox{2em}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node [multiply] (prod) {\(0\)};
\node (in0) [above of=prod] {};
\node (out0) [below of=prod] {};
\node (eq) [right of=prod] {\(=\)};
\node [bang] (del) [right of=eq, shift={(-0.2,0.2)}] {};
\node [zero] (ins) [right of=eq, shift={(-0.2,-0.2)}] {};
\node (in1) [above of=del, shift={(0,-0.2)}] {};
\node (out1) [below of=ins, shift={(0,0.2)}] {};
\draw (in0) -- (prod) -- (out0);
\draw (in1) -- (del);
\draw (ins) -- (out1);
\end{tikzpicture}
}
}
\end{center}
Multiplicative inverses cannot be so expressed, so our signal-flow diagrams so far do not know that
\(k\) is a field. Additive inverses also cannot be expressed in this way. And indeed, a version of Theorem~\ref{presvk} holds whenever \(k\) is a commutative rig: that is, a commutative `ring without negatives', such as \({\mathbb N}\). See Section \ref{conclusions} for details.
While Theorem~\ref{presvk} is a step towards understanding the category-theoretic underpinnings of
control theory, it does not treat signal-flow diagrams that include `feedback'. Feedback is one of
the most fundamental concepts in control theory because a control system without feedback may be
highly sensitive to disturbances or unmodeled behavior. Feedback allows these uncontrolled
behaviors to be mollified. As a string diagram, a basic feedback system might look schematically
like this:
\begin{center}
\begin{tikzpicture}[thick]
\node (in) {};
\node [coordinate] (inplus) [below of=in, label={[shift={(2.2em,0)}]reference}] {};
\node [plus] (plus) [below of=inplus, shift={(-0.7,0)}]{};
\node [multiply] (controller) [below of=plus, label={[shift={(3.0em,-0.5)}]controller},
label={[shift={(3.5em,0.15)}]measured error}, label={[shift={(3.1em,-1.5)}]system input}] {\(a\)};
\node [multiply] (system) [below of=controller, shift={(0,-1)},
label={[shift={(2.4em,-0.6)}]system}] {\(b\)};
\node [delta] (split) [below of=system, label={[shift={(5.2em,-1.7)}]system output}] {};
\node [coordinate] (outsplit) [below of=split, shift={(0.7,0)}] {};
\node (out) [below of=outsplit] {};
\node [coordinate] (rcup) [below of=split, shift={(-0.7,0)}] {};
\node [coordinate] (lcup) [left of=rcup, shift={(0.4,0)}] {};
\node [upmultiply] (sensor) [above of=lcup, shift={(0,0.5)}, label={[shift={(-2.2em,-0.5)}]sensor},
label={[shift={(-4.1em,1)}]measured output}] {\(c\)};
\node [upmultiply] (minus) [above of=sensor, shift={(0,1.8)}] {\(\scriptstyle{-1}\)};
\node [coordinate] (rcap) [above of=plus, shift={(-0.7,0)}] {};
\node [coordinate] (lcap) [left of=rcap, shift={(0.4,0)}] {};
\draw[rounded corners=8pt] (in) -- (inplus) -- (plus.right in) (plus) --
(controller) -- (system) -- (split) (split.right out) --
(outsplit) -- (out) (split.left out) -- (rcup) -- (lcup) --
(sensor) -- (minus) -- (lcap) -- (rcap) -- (plus.left in);
\end{tikzpicture}
\end{center}
The user inputs a `reference' signal, which is fed into a controller, whose output is fed into a
system, or `plant', which in turn produces its own output. But then the system's output is
duplicated, and one copy is fed into a sensor, whose output is added (or if we prefer, subtracted)
from the reference signal.
In string diagrams---unlike in the usual thinking on control theory---it is essential to be
able to read any diagram from top to bottom as a composite of tensor products of generating
morphisms. Thus, to incorporate the idea of feedback, we need two more generating morphisms.
These are the `cup':
\begin{center}
\begin{tikzpicture}[thick]
\node (0) [label={[shift={(0,-1.6)}]}] {\(f=g\)};
\node[coordinate] (3) [left of=0] {};
\node[coordinate] (4) [right of=0] {};
\node (1) [above of=3] {\(f\)};
\node (2) [above of=4] {\(g\)};
\path
(1) edge (3)
(2) edge (4)
(3) edge [-, bend right=90] (4);
\end{tikzpicture}
\end{center}
and `cap':
\begin{center}
\begin{tikzpicture}[thick]
\node (0) {\(f=g\)};
\node[coordinate] (3) [left of=0] {};
\node[coordinate] (4) [right of=0] {};
\node (1) [below of=3] {\(f\)};
\node (2) [below of=4] {\(g\)};
\path
(3) edge (1)
(4) edge (2)
(3) edge [bend left=90] (4);
\end{tikzpicture}
\end{center}
These are not maps: they are relations. The cup imposes the relation that its two inputs be equal,
while the cap does the same for its two outputs. This is a way of describing how a signal flows
around a bend in a wire.
To make this precise, we use a category called \(\mathrm{FinRel}_k\). An object of this category is a
finite-dimensional vector space over \(k\), while a morphism from \(U\) to \(V\), denoted \(L \maps
U \nrightarrow V\), is a \Define{linear relation}, meaning a linear subspace
\[ L \subseteq U \oplus V .\]
In particular, when \(k = {\mathbb R}(s)\), a linear relation \(L \maps k^m \to k^n\) is just an arbitrary
system of constant-coefficient linear ordinary differential equations relating \(m\) input
variables and \(n\) output variables.
Since the direct sum \(U \oplus V\) is also the cartesian product of \(U\) and \(V\), a linear
relation is indeed a relation in the usual sense, but with the property that if \(u \in U\) is
related to \(v \in V\) and \(u' \in U\) is related to \(v' \in V\) then \(cu + c'u'\) is related to
\(cv + c'v'\) whenever \(c,c' \in k\). We compose linear relations \(L \maps U \nrightarrow V\) and
\(L' \maps V \nrightarrow W\) as follows:
\[ L'L = \{(u,w) \colon \; \exists\; v \in V \;\; (u,v) \in L \textrm{ and }
(v,w) \in L'\} .\]
Any linear map \(f \maps U \to V\) gives a linear relation \(F \maps U \nrightarrow V\), namely the
graph of that map:
\[ F = \{ (u,f(u)) : u \in U \}. \]
Composing linear maps thus becomes a special case of composing linear relations, so \(\mathrm{FinVect}_k\)
becomes a subcategory of \(\mathrm{FinRel}_k\). Furthermore, we can make \(\mathrm{FinRel}_k\) into a monoidal category
using direct sums, and it becomes symmetric monoidal using the braiding already present in
\(\mathrm{FinVect}_k\).
In these terms, the \Define{cup} is the linear relation
\[ \cup \maps k^2 \nrightarrow \{0\} \]
given by
\[ \cup \; = \; \{ (x,x,0) : x \in k \} \; \subseteq \; k^2 \oplus \{0\}, \]
while the \Define{cap} is the linear relation
\[ \cap \maps \{0\} \nrightarrow k^2 \]
given by
\[ \cap \; = \; \{ (0,x,x) : x \in k \} \; \subseteq \; \{0\} \oplus k^2 .\]
These obey the \Define{zigzag relations}:
\begin{center}
\begin{tikzpicture}[-, thick, node distance=1cm]
\node (zigtop) {};
\node [coordinate] (zigincup) [below of=zigtop] {};
\node [coordinate] (zigcupcap) [right of=zigincup] {};
\node [coordinate] (zigoutcap) [right of=zigcupcap] {};
\node (zigbot) [below of=zigoutcap] {};
\node (equal) [right of=zigoutcap] {\(=\)};
\node (mid) [right of=equal] {};
\node (vtop) [above of=mid] {};
\node (vbot) [below of=mid] {};
\node (equals) [right of=mid] {\(=\)};
\node [coordinate] (zagoutcap) [right of=equals] {};
\node (zagbot) [below of=zagoutcap] {};
\node [coordinate] (zagcupcap) [right of=zagoutcap] {};
\node [coordinate] (zagincup) [right of=zagcupcap] {};
\node (zagtop) [above of=zagincup] {};
\path
(zigincup) edge (zigtop) edge [bend right=90] (zigcupcap)
(zigoutcap) edge (zigbot) edge [bend right=90] (zigcupcap)
(vtop) edge (vbot)
(zagincup) edge (zagtop) edge [bend left=90] (zagcupcap)
(zagoutcap) edge (zagbot) edge [bend left=90] (zagcupcap);
\end{tikzpicture}
\end{center}
Thus, they make \(\mathrm{FinRel}_k\) into a compact closed category where \(k\), and thus every object, is its
own dual.
Besides feedback, one of the things that make the cap and cup useful is that they allow any
morphism \(L \maps U \nrightarrow V \) to be `plugged in backwards' and thus `turned around'. For
instance, turning around integration:
\begin{center}
\begin{tikzpicture}[thick]
\node [integral] (dot) {\(\int\)};
\node [coordinate] (cupout) [below of=dot, shift={(0,0.2)}] {};
\node [coordinate] (cupin) [left of=cupout] {};
\node [coordinate] (capin) [above of=dot, shift={(0,-0.4)}] {};
\node [coordinate] (in) [left of=capin, shift={(0,0.5)}] {};
\node [coordinate] (capout) [right of=capin] {};
\node [coordinate] (out) [right of=cupout, shift={(0,-0.7)}] {};
\draw (capin) -- (dot) -- (cupout);
\path
(in) edge (cupin)
(capout) edge (out)
(cupin) edge [bend right=90] (cupout)
(capin) edge [bend left=90] (capout);
\node (eq) [left of=dot, shift={(-1,0)}] {\(:=\)};
\node [upmultiply] (diff) [left of=eq, shift={(-0.3,-0.2)}] {\(\int\)};
\node [coordinate] (diffin) [above of=diff, shift={(0,0.3)}] {};
\node [coordinate] (diffout) [below of=diff, shift={(0,-0.3)}] {};
\draw (diffin) -- (diff) -- (diffout);
\end{tikzpicture}
\end{center}
we obtain differentiation. In general, using caps and cups we can turn around any linear relation
\(L \maps U \nrightarrow V\) and obtain a linear relation \(L^\dagger \maps V \nrightarrow U\), called the
\Define{adjoint} of \(L\), which turns out to given by
\[ L^\dagger = \{(v,u) : (u,v) \in L \} .\]
For example, if \(c \in k\) is nonzero, the adjoint of scalar multiplication by \(c\) is
multiplication by \(c^{-1}\):
\begin{center}
\begin{tikzpicture}[thick]
\node [multiply] (c) {\(c\)};
\node [coordinate] (cupout) [below of=c, shift={(0,0.4)}] {};
\node [coordinate] (cupin) [left of=cupout] {};
\node [coordinate] (capin) [above of=c, shift={(0,-0.5)}] {};
\node [coordinate] (in) [left of=capin, shift={(0,0.7)}] {};
\node [coordinate] (capout) [right of=capin] {};
\node [coordinate] (out) [right of=cupout, shift={(0,-0.7)}] {};
\draw (capin) -- (c) -- (cupout);
\path
(in) edge (cupin)
(capout) edge (out)
(cupin) edge [bend right=90] (cupout)
(capin) edge [bend left=90] (capout);
\node (eq) [right of=c, shift={(1.1,0)}] {\(=\)};
\node [multiply] (mult) [right of=eq, shift={(0.5,0)}] {\(c^{-1}\!\!\)};
\node [coordinate] (min) [above of=mult, shift={(0,0.2)}] {};
\node [coordinate] (mout) [below of=mult, shift={(0,-0.3)}] {};
\draw (min) -- (mult) -- (mout);
\node (colon) [left of=c, shift={(-1.1,0)}] {\(:=\)};
\node [upmultiply] (adj) [left of=colon, shift={(-0.25,0)}] {\(c\)};
\node [coordinate] (adin) [above of=adj, shift={(0,0.2)}] {};
\node [coordinate] (adout) [below of=adj, shift={(0,-0.3)}] {};
\draw (adin) -- (adj) -- (adout);
\end{tikzpicture}
\end{center}
Thus, caps and cups allow us to express multiplicative inverses in terms of signal-flow diagrams!
One might think that a problem arises when when \(c = 0\), but no: the adjoint of scalar
multiplication by \(0\) is
\[ \{(0,x) : x \in k \} \subseteq k \oplus k .\]
In Lemma~\ref{gensrk} we show that \(\mathrm{FinRel}_k\) is generated, as a symmetric monoidal category, by
these morphisms:
\begin{center}
\scalebox{0.9}{
\begin{tikzpicture}[thick]
\node [coordinate] (in) at (0,2) {};
\node [multiply] (mult) at (0,1) {\(c\)};
\node [coordinate] (out) at (0,0) {};
\draw (in) -- (mult) -- (out);
\end{tikzpicture}
\hspace{3 em}
\begin{tikzpicture}[thick]
\node [plus] (adder) at (0,0.85) {};
\node [coordinate] (f) at (-0.5,1.5) {};
\node [coordinate] (g) at (0.5,1.5) {};
\node [coordinate] (out) at (0,0) {};
\node [coordinate] (pref) at (-0.5,2) {};
\node [coordinate] (preg) at (0.5,2) {};
\draw [rounded corners] (pref) -- (f) -- (adder.left in);
\draw [rounded corners] (preg) -- (g) -- (adder.right in);
\draw (adder) -- (out);
\end{tikzpicture}
\hspace{2em}
\begin{tikzpicture}[thick]
\node[delta] (dupe) at (0,1.15) {};
\node[coordinate] (o1) at (-0.5,0.5) {};
\node[coordinate] (o2) at (0.5,0.5) {};
\node[coordinate] (in) at (0,2) {};
\node [coordinate] (posto1) at (-0.5,0) {};
\node [coordinate] (posto2) at (0.5,0) {};
\draw[rounded corners] (posto1) -- (o1) -- (dupe.left out);
\draw[rounded corners] (posto2) -- (o2) -- (dupe.right out);
\draw (in) -- (dupe);
\end{tikzpicture}
\hspace{3em}
\begin{tikzpicture}[thick]
\node[coordinate] (in) at (0,2) {};
\node [bang] (mult) at (0,1) {};
\node [hole] (heightHolder) at (0,0) {};
\draw (in) -- (mult);
\end{tikzpicture}
\hspace{2em}
\begin{tikzpicture}[thick]
\node[hole] (heightHolder) at (0,2) {};
\node [coordinate] (out) at (0,0) {};
\node [zero] (del) at (0,1) {};
\draw (del) -- (out);
\end{tikzpicture}
\hspace{3em}
\begin{tikzpicture}[thick]
\node [coordinate] (3) at (0,0.375) {};
\node [coordinate] (4) at (1.3,0.375) {};
\node [coordinate] (1) at (0,2) {};
\node [coordinate] (2) at (1.3,2) {};
\path
(1) edge (3)
(2) edge (4)
(3) edge [-, bend right=90] (4);
\end{tikzpicture}
\hspace{3em}
\begin{tikzpicture}[thick]
\node [coordinate] (3) at (0,1.625) {};
\node [coordinate] (4) at (1.3,1.625) {};
\node [coordinate] (1) at (0,0) {};
\node [coordinate] (2) at (1.3,0) {};
\path
(3) edge (1)
(4) edge (2)
(3) edge [bend left=90] (4);
\end{tikzpicture}
}
\end{center}
where \(c \in k\) is arbitrary.
In Theorem~\ref{presrk} we find a complete set of relations obeyed by these generating morphisms,
thus giving a presentation of \(\mathrm{FinRel}_k\) as a symmetric monoidal category. To describe these
relations, it is useful to work with adjoints of the generating morphisms. We have already seen
that the adjoint of scalar multiplication by \(c\) is scalar multiplication by \(c^{-1}\), except
when \(c = 0\). Taking adjoints of the other four generating morphisms of \(\mathrm{FinVect}_k\), we obtain
four important but perhaps unfamiliar linear relations. We draw these as `turned around' versions
of the original generating morphisms:
\begin{itemize}
\item \Define{Coaddition} is a linear relation from \(k\) to \(k^2\) that holds when the two
outputs sum to the input:
\[ +^\dagger \maps k \nrightarrow k^2 \]
\[ +^\dagger = \{(x,y,z) : \; x = y + z \} \subseteq k \oplus k^2 \]
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (adder) at (0,0) {};
\node [coordinate] (sum) at (0,-0.5) {};
\node [coordinate] (sumup) at (0.9,-0.5) {};
\node [coordinate] (input) at (0.9,1.3) {};
\node [coordinate] (highcap) at (-0.6,1.05) {};
\node [coordinate] (outerloop) at (-1.7,0.1) {};
\node [coordinate] (outerout) at (-1.7,-1) {};
\node [coordinate] (innerloop) at (-0.9,-0.2) {};
\node [coordinate] (innerout) at (-0.9,-1) {};
\draw (innerloop) -- (innerout);
\path (sum) edge [bend right=90] (sumup);
\draw (adder.io) -- (sum)
(sumup) -- (input)
(outerloop) -- (outerout)
(adder.left in) .. controls +(120:0.5) and +(0,1) .. (innerloop)
(adder.right in) .. controls +(60:0.75) and +(0.6,0) .. (highcap)
(highcap) .. controls +(-0.6,0) and +(0,0.6) .. (outerloop);
\node (eq) at (-2.75,0.15) {\(:=\)};
\node [coplus] (coadder) at (-4,0.3) {};
\node [coordinate] (topco) at (-4,1.3) {};
\node [coordinate] (leftout) at (-4.5,-1) {};
\node [coordinate] (rightout) at (-3.5,-1) {};
\draw
(coadder.left out) .. controls +(240:0.7) and +(90:0.7) .. (leftout)
(coadder.right out) .. controls +(300:0.7) and +(90:0.7) .. (rightout)
(coadder.io) -- (topco);
\end{tikzpicture}
\end{center}
\item \Define{Cozero} is a linear relation from \(k\) to \(\{0\}\) that holds
when the input is zero:
\[ 0^\dagger \maps k \nrightarrow \{0\} \]
\[ 0^\dagger = \{ (0,0)\} \subseteq k \oplus \{0\} \]
\begin{center}
\begin{tikzpicture}[thick]
\node [zero] (Ze) at (0,0) {};
\node (eq) at (-1,0) {\(:=\)};
\node [zero] (Ro) at (-1.8,-0.5) {};
\draw[rounded corners=7pt]
(Ze) -- (0,-0.5) -- (0.5,-0.5) -- (0.5,0.5);
\draw (Ro) -- (-1.8,0.5);
\end{tikzpicture}
\end{center}
\item \Define{Coduplication} is a linear relation from \(k^2\) to \(k\) that holds when the two
inputs both equal the output:
\[ \Delta^\dagger \maps k^2 \nrightarrow k \]
\[ \Delta^\dagger = \{(x,y,z) : \; x = y = z \} \subseteq k^2 \oplus k \]
\begin{center}
\begin{tikzpicture}[thick]
\node [delta] (copier) at (0,0) {};
\node [coordinate] (original) at (0,0.5) {};
\node [coordinate] (origdown) at (0.9,0.5) {};
\node [coordinate] (output) at (0.9,-1.3) {};
\node [coordinate] (lowcup) at (-0.6,-1.05) {};
\node [coordinate] (outerloop) at (-1.7,-0.1) {};
\node [coordinate] (outerin) at (-1.7,1) {};
\node [coordinate] (innerloop) at (-0.9,0.2) {};
\node [coordinate] (innerout) at (-0.9,1) {};
\draw (innerloop) -- (innerout);
\path (original) edge [bend left=90] (origdown);
\draw (copier.io) -- (original)
(origdown) -- (output)
(outerloop) -- (outerin)
(copier.left out) .. controls +(240:0.5) and +(0,-1) .. (innerloop)
(copier.right out) .. controls +(300:0.75) and +(0.6,0) .. (lowcup)
(lowcup) .. controls +(-0.6,0) and +(0,-0.6) .. (outerloop);
\node (eq) at (-2.75,-0.15) {\(:=\)};
\node [codelta] (pier) at (-4,-0.3) {};
\node [coordinate] (bottomco) at (-4,-1.3) {};
\node [coordinate] (leftin) at (-4.5,1) {};
\node [coordinate] (rightin) at (-3.5,1) {};
\draw
(pier.left in) .. controls +(120:0.7) and +(270:0.7) .. (leftin)
(pier.right in) .. controls +(60:0.7) and +(270:0.7) .. (rightin)
(pier.io) -- (bottomco);
\end{tikzpicture}
\end{center}
\item \Define{Codeletion} is a linear relation from \(\{0\}\) to \(k\) that holds always:
\[ !^\dagger \maps \{0\} \nrightarrow k \]
\[ !^\dagger = \{(0,x) \} \subseteq \{0\} \oplus k \]
\begin{center}
\begin{tikzpicture}[thick]
\node [bang] (Ba) at (0,0) {};
\node (eq) at (-1,0) {\(:=\)};
\node [bang] (ng) at (-1.8,0.5) {};
\draw[rounded corners=7pt]
(Ba) -- (0,0.5) -- (0.5,0.5) -- (0.5,-0.5);
\draw (ng) -- (-1.8,-0.5);
\end{tikzpicture}
\end{center}
\end{itemize}
Since \(+^\dagger,0^\dagger,\Delta^\dagger\) and \(!^\dagger\) automatically obey turned-around
versions of the relations obeyed by \(+,0,\Delta\) and \(!\), we see that \(k\) acquires a
\emph{second} bicommutative bimonoid structure when considered as an object in \(\mathrm{FinRel}_k\).
Moreover, the four dark operations make \(k\) into a \Define{Frobenius monoid}. This means that
\((k,+,0)\) is a monoid, \((k,+^\dagger, 0^\dagger)\) is a comonoid, and the \Define{Frobenius
relation} holds:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [plus] (sum1) at (0.5,-0.216) {};
\node [coplus] (cosum1) at (1,0.216) {};
\node [coordinate] (sum1corner) at (0,0.434) {};
\node [coordinate] (cosum1corner) at (1.5,-0.434) {};
\node [coordinate] (sum1out) at (0.5,-0.975) {};
\node [coordinate] (cosum1in) at (1,0.975) {};
\node [coordinate] (1cornerin) at (0,0.975) {};
\node [coordinate] (1cornerout) at (1.5,-0.975) {};
\draw[rounded corners] (1cornerin) -- (sum1corner) -- (sum1.left in)
(1cornerout) -- (cosum1corner) -- (cosum1.right out);
\draw (sum1.right in) -- (cosum1.left out)
(sum1.io) -- (sum1out)
(cosum1.io) -- (cosum1in);
\node (eq1) at (2,0) {\(=\)};
\node [plus] (sum2) at (3,0.325) {};
\node [coplus] (cosum2) at (3,-0.325) {};
\node [coordinate] (sum2inleft) at (2.5,0.975) {};
\node [coordinate] (sum2inright) at (3.5,0.975) {};
\node [coordinate] (cosum2outleft) at (2.5,-0.975) {};
\node [coordinate] (cosum2outright) at (3.5,-0.975) {};
\draw (sum2inleft) .. controls +(270:0.3) and +(120:0.15) .. (sum2.left in)
(sum2inright) .. controls +(270:0.3) and +(60:0.15) .. (sum2.right in)
(cosum2outleft) .. controls +(90:0.3) and +(240:0.15) .. (cosum2.left out)
(cosum2outright) .. controls +(90:0.3) and +(300:0.15) .. (cosum2.right out)
(sum2.io) -- (cosum2.io);
\node (eq2) at (4,0) {\(=\)};
\node [plus] (sum3) at (5.5,-0.216) {};
\node [coplus] (cosum3) at (5,0.216) {};
\node [coordinate] (sum3corner) at (6,0.434) {};
\node [coordinate] (cosum3corner) at (4.5,-0.434) {};
\node [coordinate] (sum3out) at (5.5,-0.975) {};
\node [coordinate] (cosum3in) at (5,0.975) {};
\node [coordinate] (3cornerin) at (6,0.975) {};
\node [coordinate] (3cornerout) at (4.5,-0.975) {};
\draw[rounded corners] (3cornerin) -- (sum3corner) -- (sum3.right in)
(3cornerout) -- (cosum3corner) -- (cosum3.left out);
\draw (sum3.left in) -- (cosum3.right out)
(sum3.io) -- (sum3out)
(cosum3.io) -- (cosum3in);
\end{tikzpicture}
}
\end{center}
All three expressions in this equation are linear relations saying that the sum of the two inputs
equal the sum of the two outputs.
The operation sending each linear relation to its adjoint extends to a contravariant functor
\[ \dagger \maps \mathrm{FinRel}_k\ \to \mathrm{FinRel}_k ,\]
which obeys a list of properties that are summarized by saying that \(\mathrm{FinRel}_k\) is a
`\(\dagger\)-compact' category \cite{AC,Selinger}. Because two of the operations in the Frobenius
monoid \((k, +,0,+^\dagger,0^\dagger)\) are adjoints of the other two, it is a
\Define{\(\dagger\)-Frobenius monoid}. This Frobenius monoid is also \Define{special}, meaning that
comultiplication (in this case \(+^\dagger\)) followed by multiplication (in this case \(+\)) equals
the identity:
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (sum) at (0.4,-0.5) {};
\node [coplus] (cosum) at (0.4,0.5) {};
\node [coordinate] (in) at (0.4,1) {};
\node [coordinate] (out) at (0.4,-1) {};
\node (eq) at (1.3,0) {\(=\)};
\node [coordinate] (top) at (2,1) {};
\node [coordinate] (bottom) at (2,-1) {};
\path (sum.left in) edge[bend left=30] (cosum.left out)
(sum.right in) edge[bend right=30] (cosum.right out);
\draw (top) -- (bottom)
(sum.io) -- (out)
(cosum.io) -- (in);
\end{tikzpicture}
\end{center}
This Frobenius monoid is also commutative---and cocommutative, but for Frobenius monoids this
follows from commutativity.
Starting around 2008, commutative special \(\dagger\)-Frobenius monoids have become important in the
categorical foundations of quantum theory, where they can be understood as `classical structures'
for quantum systems \cite{CPV,Vicary}. The category \(\mathrm{Fin}\mathrm{Hilb}\) of finite-dimensional Hilbert
spaces and linear maps is a \(\dagger\)-compact category, where any linear map \(f \maps H \to K\)
has an adjoint \(f^\dagger \maps K \to H\) given by
\[ \langle f^\dagger \phi, \psi \rangle = \langle \phi, f \psi \rangle \]
for all \(\psi \in H, \phi \in K \). A commutative special \(\dagger\)-Frobenius monoid in
\(\mathrm{Fin}\mathrm{Hilb}\) is then the same as a Hilbert space with a chosen orthonormal basis. The reason is
that given an orthonormal basis \( \psi_i \) for a finite-dimensional Hilbert space \(H\), we can
make \(H\) into a commutative special \(\dagger\)-Frobenius monoid with multiplication \(m \maps H
\otimes H \to H\) given by
\[ m (\psi_i \otimes \psi_j ) = \left\{ \begin{array}{cl} \psi_i & i = j \\
0 & i \ne j
\end{array}\right. \]
and unit \(i \maps {\mathbb C} \to H\) given by
\[ i(1) = \sum_i \psi_i . \]
The comultiplication \(m^\dagger\) duplicates basis states:
\[ m^\dagger(\psi_i) = \psi_i \otimes \psi_i . \]
Conversely, any commutative special \(\dagger\)-Frobenius monoid in \(\mathrm{Fin}\mathrm{Hilb}\) arises this way.
Considerably earlier, around 1995, commutative Frobenius monoids were recognized as important in
topological quantum field theory. The reason, ultimately, is that the free symmetric monoidal
category on a commutative Frobenius monoid is \(2\mathrm{Cob}\), the category with 2-dimensional oriented
cobordisms as morphisms: see Kock's textbook \cite{Kock} and the many references therein. But the
free symmetric monoidal category on a commutative \emph{special} Frobenius monoid was worked out
even earlier \cite{CW,Kock2,RSW}: it is the category with finite sets as objects, where a morphism
\(f \maps X \to Y\) is an isomorphism class of cospans
\[ X \longrightarrow S \longleftarrow Y .\]
This category can be made into a \(\dagger\)-compact category in an obvious way, and then the
1-element set becomes a commutative special \(\dagger\)-Frobenius monoid.
For all these reasons, it is interesting to find a commutative special \(\dagger\)-Frobenius monoid
lurking at the heart of control theory! However, the Frobenius monoid here has yet another
property, which is more unusual. Namely, the unit \(0 \maps \{0\} \nrightarrow k\) followed by the
counit \(0^\dagger \maps k \nrightarrow \{0\} \) is the identity:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [zero] (Bins) {};
\node [zero] (Tins) [above of=Bins] {};
\path
(Tins) edge (Bins);
\end{tikzpicture}
\quad
\raisebox{1em}{=}
}
\end{center}
We call a special Frobenius monoid that also obeys this extra law \Define{extra-special}. One can
check that the free symmetric monoidal category on a commutative extra-special Frobenius monoid is
the category with finite sets as objects, where a morphism \(f \maps X \to Y\) is an equivalence
relation on the disjoint union \(X \sqcup Y\), and we compose \(f \maps X \to Y\) and \(g \maps Y
\to Z\) by letting \(f\) and \(g\) generate an equivalence relation on \(X \sqcup Y \sqcup Z\) and
then restricting this to \(X \sqcup Z\).
As if this were not enough, the light operations share many properties with the dark ones. In
particular, these operations make \(k\) into a commutative extra-special \(\dagger\)-Frobenius
monoid in a second way. In summary:
\begin{itemize}
\item \((k, +, 0, \Delta, !)\) is a bicommutative bimonoid;
\item \((k, \Delta^\dagger, !^\dagger, +^\dagger, 0^\dagger)\) is a bicommutative bimonoid;
\item \((k, +, 0, +^\dagger, 0^\dagger)\) is a commutative extra-special
\(\dagger\)-Frobenius monoid;
\item \((k, \Delta^\dagger, !^\dagger, \Delta, !)\) is a commutative extra-special
\(\dagger\)-Frobenius monoid.
\end{itemize}
It should be no surprise that with all these structures built in, signal-flow diagrams are a
powerful method of designing processes. However, it is surprising that most of these structures
are present in a seemingly very different context: the so-called `ZX calculus', a diagrammatic
formalism for working with complementary observables in quantum theory \cite{CD}. This arises
naturally when one has an \(n\)-dimensional Hilbert space \(H\) with two orthonormal bases
\(\psi_i, \phi_i \) that are `mutually unbiased', meaning that
\[ |\langle \psi_i, \phi_j\rangle|^2 = \displaystyle{\frac{1}{n}} \]
for all \(1 \le i, j \le n\). Each orthonormal basis makes \(H\) into commutative special
\(\dagger\)-Frobenius monoid in \(\mathrm{Fin}\mathrm{Hilb}\). Moreover, the multiplication and unit of either one
of these Frobenius monoids fits together with the comultiplication and counit of the other to form
a bicommutative bimonoid. So, we have all the structure present in the list above---except
that these Frobenius monoids are only extra-special if \(H\) is 1-dimensional.
The field \(k\) is also a 1-dimensional vector space, but this is a red herring: in \(\mathrm{FinRel}_k\)
\emph{every} finite-dimensional vector space naturally acquires all four structures listed above,
since addition, zero, duplication and deletion are well-defined and obey all the relations we have
discussed. We focus on \(k\) in this paper simply because it generates all the objects \(\mathrm{FinRel}_k\)
via direct sum.
Finally, in \(\mathrm{FinRel}_k\) the cap and cup are related to the light and dark operations as follows:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node (eq) at (0.2,-0.1) {\(=\)};
\node [coordinate] (lcap) at (-1.5,0.5) {};
\node [coordinate] (rcap) at (-0.5,0.5) {};
\node [coordinate] (lcapbot) at (-1.5,-1) {};
\node [coordinate] (rcapbot) at (-0.5,-1) {};
\node [delta] (dub) at (1.25,0) {};
\node [bang] (boom) at (1.25,0.65) {};
\node [coordinate] (Leftout) at (0.75,-1) {};
\node [coordinate] (Rightout) at (1.75,-1) {};
\draw (dub.left out) .. controls +(240:0.5) and +(90:0.5) .. (Leftout)
(dub.right out) .. controls +(300:0.5) and +(90:0.5) .. (Rightout);
\draw (boom) -- (dub) (lcapbot) -- (lcap) (rcap) -- (rcapbot);
\path (lcap) edge[bend left=90] (rcap);
\end{tikzpicture}
\qquad \qquad
\begin{tikzpicture}[thick]
\node [multiply] (neg) at (0,0.1) {\(\scriptstyle{-1}\)};
\node [coordinate] (cupInLeft) at (0,1) {};
\node [coordinate] (Lcup) at (0,-0.5) {};
\node [coordinate] (Rcup) at (1,-0.5) {};
\node [coordinate] (cupInRight) at (1,1) {};
\node (eq) at (1.7,0.1) {\(=\)};
\node [coordinate] (SumLeftIn) at (2.25,1) {};
\node [coordinate] (SumRightIn) at (3.25,1) {};
\node [plus] (Sum) at (2.75,0) {};
\node [zero] (coZero) at (2.75,-0.65) {};
\draw (SumRightIn) .. controls +(270:0.5) and +(60:0.5) .. (Sum.right in)
(SumLeftIn) .. controls +(270:0.5) and +(120:0.5) .. (Sum.left in);
\draw (cupInLeft) -- (neg) -- (Lcup)
(Rcup) -- (cupInRight)
(Sum) -- (coZero);
\path (Lcup) edge[bend right=90] (Rcup);
\end{tikzpicture}
}
\end{center}
Note the curious factor of \(-1\) in the second equation, which breaks some of the symmetry we have
seen so far. This equation says that two elements \(x, y \in k\) sum to zero if and only if \(-x =
y\). Using the zigzag relations, the two equations above give
\begin{center}
\begin{tikzpicture}[thick]
\node (eq) {\(=\)};
\node[delta] (Lup) at (-1,0.216) {};
\node[plus] (Ldn) at (-1.5,-0.216) {};
\node[coordinate] (Lupo) at (-0.5,-0.434) {};
\node[coordinate] (Ldni) at (-2,0.434) {};
\node[bang] (Lupi) at (-1,0.866) {};
\node[zero] (Ldno) at (-1.5,-0.866) {};
\node (Lo) at (-0.5,-1.082) {};
\node (Li) at (-2,1.082) {};
\node [multiply] (neg1) [right of=eq] {\(\scriptstyle{-1}\)};
\node[coordinate] (inR) [above of=neg1] {};
\node[coordinate] (outR) [below of=neg1] {};
\draw[rounded corners] (Li) -- (Ldni) -- (Ldn.left in) (Lo) -- (Lupo) -- (Lup.right out);
\draw (Ldn) -- (Ldno) (Lup) -- (Lupi) (Ldn.right in) -- (Lup.left out);
\path (neg1) edge (inR) edge (outR);
\end{tikzpicture}
\end{center}
We thus see that in \(\mathrm{FinRel}_k\), both additive and multiplicative inverses can be expressed in terms
of the generating morphisms used in signal-flow diagrams.
Theorem~\ref{presrk} gives a presentation of \(\mathrm{FinRel}_k\) based on the ideas just discussed. Briefly,
it says that \(\mathrm{FinRel}_k\) is equivalent to the symmetric monoidal category generated by an object \(k\)
and these morphisms:
\begin{enumerate}
\item addition \(+\maps k^2 \nrightarrow k\)
\item zero \(0 \maps \{0\} \nrightarrow k \)
\item duplication \(\Delta\maps k\nrightarrow k^2 \)
\item deletion \(! \maps k \nrightarrow 0\)
\item scalar multiplication \(c\maps k\nrightarrow k\) for any \(c\in k\)
\item cup \(\cup \maps k^2 \nrightarrow \{0\} \)
\item cap \(\cap \maps \{0\} \nrightarrow k^2 \)
\end{enumerate}
obeying these relations:
\begin{enumerate}
\item \((k, +, 0, \Delta, !)\) is a bicommutative bimonoid;
\item \(\cap\) and \(\cup\) obey the zigzag equations;
\item \((k, +, 0, +^\dagger, 0^\dagger)\) is a commutative extra-special
\(\dagger\)-Frobenius monoid;
\item \((k, \Delta^\dagger, !^\dagger, \Delta, !)\) is a commutative extra-special
\(\dagger\)-Frobenius monoid;
\item the field operations of \(k\) can be recovered from the generating morphisms;
\item the generating morphisms (1)-(4) commute with scalar multiplication.
\end{enumerate}
Note that item (2) makes \(\mathrm{FinRel}_k\) into a \(\dagger\)-compact category, allowing us to mention the
adjoints of generating morphisms in the subsequent relations. Item (5) means that \(+,\cdot, 0,
1\) and also additive and multiplicative inverses in the field \(k\) can be expressed in terms of
signal-flow diagrams in the manner we have explained.
\section{A presentation of \(\mathrm{FinVect}_k\)}
\label{finvect}
Our goal in this section is to find a presentation for the symmetric monoidal category \(\mathrm{FinVect}_k\).
To simplify some technicalities, we shall use Mac Lane's coherence theorem \cite{MacLane} to choose
a symmetric monoidal equivalence \(F \maps \mathrm{FinVect}_k' \to \mathrm{FinVect}_k\) where \(\mathrm{FinVect}_k'\) is strict. This
allows us to avoid mentioning associators and unitors, since in \(\mathrm{FinVect}_k'\) these are identity
morphisms. In what follows, we call \(\mathrm{FinVect}_k'\) simply \(\mathrm{FinVect}_k\), and call objects and morphisms
in \(\mathrm{FinVect}_k\) by the names of their images under \(F\). Colloquially speaking, we `work in a strict
version' of \(\mathrm{FinVect}_k\), and do not bother to indicate that this is a different (though equivalent)
symmetric monoidal category.
We say a strict symmetric monoidal category \(C\) is \Define{generated} by a set \(O\) of objects
and a set \(M\) of morphisms going between tensor products of objects in \(O\) if the smallest
subcategory \(C_0\) of \(C\) containing:
\begin{itemize}
\item the objects in \(O\),
\item the morphisms in \(M\),
\item the tensor products of any objects or morphisms in \(C_0\)
\item the braiding for any pair of objects in \(C_0\)
\end{itemize}
has the property that the inclusion \(i \maps C_0 \to C\) is an equivalence of categories. It
follows that \(i\) extends to an equivalence of symmetric monoidal categories. In this situation we
call the elements of \(O\) \Define{generating objects} for \(C\), and call the elements of \(M\)
\Define{generating morphisms}.
\begin{lemma} \label{gensvk}
For any field \(k\), the object \(k\) together with the morphisms:
\begin{enumerate}
\item scalar multiplication \(c \maps k \to k\) for any \(c \in k\)
\item addition \(+ \maps k \oplus k \to k\)
\item zero \(0 \maps \{0\} \to k\)
\item duplication \(\Delta \maps k \to k \oplus k\)
\item deletion \(! \maps k \to \{0\}\)
\end{enumerate}
generate \(\mathrm{FinVect}_k\), the category of finite-dimensional vector spaces over \(k\) and linear maps, as
a symmetric monoidal category.
\end{lemma}
\begin{proof}
It suffices to show that \(k\) together with the morphisms in (1)--(5) generate the full subcategory
of \(\mathrm{FinVect}_k\) containing only the iterated direct sums \(k^n = k \oplus \cdots \oplus k\), since
this is equivalent to \(\mathrm{FinVect}_k\).
A linear map in \(\mathrm{FinVect}_k\), \(T \maps k^m \to k^n\) can be expressed as \(n\) \(k\)-linear
combinations of \(m\) elements of \(k\). That is, \(T(k_1, \ldots, k_m) = (\sum_j{a_{1j} k_j},
\ldots, \sum_j{a_{nj} k_j})\), \(a_{ij} \in k\). Any \(k\)-linear combination of \(r\) elements can
be constructed with only addition, multiplication, and zero, with zero only necessary when
providing the unique \(k\)-linear combination for \(r=0\). When \(r=1\), \(a_1 (k_1)\) is an
arbitrary \(k\)-linear combination. For \(r>1\), \(+ (S_{r-1}, a_r (k_r))\) yields an arbitrary
\(k\)-linear combination on \(r\) elements, where \(S_{r-1}\) is an arbitrary \(k\)-linear
combination of \(r-1\) elements. The inclusion of duplication allows process of forming
\(k\)-linear combinations to be repeated an arbitrary (finite) positive number of times, and
deletion allows the process to be repeated zero times. When \(n\) \(k\)-linear combinations are
needed, each input may be duplicated \(n-1\) times. Because \(\mathrm{FinVect}_k\) is being generated as a
symmetric monoidal category, the \(mn\) outputs can then be permuted into \(n\) collections of \(m\)
outputs: one output from each input for each collection. Each collection can then form a
\(k\)-linear combination, as above. The following diagrams illustrate the pieces that form this
inductive argument.
\begin{center}
\begin{tikzpicture}[thick]
\node (top) at (0,4) {\(k_1\)};
\node [multiply] (times) at (0,2) {\(a_1\)};
\node (bottom) at (0,0) {\(a_1 k_1\)};
\draw (top) -- (times) -- (bottom);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[thick]
\node (sum) {\(\sum\limits_{j=1}^{r-1} a_j k_j\)};
\node [coordinate] (UL) [below of=sum, shift={(0.1,-0.9)}] {};
\node [plus] (adder) [below right of=UL, shift={(0,-0.1)}] {};
\node [coordinate] (UR) [above right of=adder] {};
\node [multiply] (mult) [above of=UR, shift={(-0.1,-0.1)}] {\(a_r\)};
\node (next) [above of=mult] {\(k_r\)};
\node (combo) [below of=adder, shift={(0,-0.2)}] {\(\sum\limits_{j=1}^{r} a_j k_j\)};
\draw (next) -- (mult) (mult.io) .. controls +(270:0.5) and +(60:0.5) .. (adder.right in);
\draw (sum.270) .. controls +(270:1.2) and +(120:0.5) .. (adder.left in);
\draw (adder) -- (combo);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[thick, node distance=1.1cm]
\node (top) {\(k_1\)};
\node [delta] (dupe) [below of=top, shift={(0,-0.125)}] {};
\node [coordinate] (L) [below left of=dupe] {};
\node [coordinate] (R) [below right of=dupe] {};
\node [multiply] (mult) [below of=R, shift={(-0.2,0.3)}] {\(a_{i1}\)};
\node (prod) [below of=mult, shift={(0,-0.3)}] {\(a_{i1} k_1\)};
\node (id) [below of=L, shift={(0.2,-1.1)}] {\(k_1\)};
\draw (dupe.right out) .. controls +(300:0.5) and +(90:0.5) .. (mult.90) (mult.io) -- (prod);
\draw (dupe.left out) .. controls +(240:0.8) and +(90:2) .. (id);
\draw (top) -- (dupe);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[thick, node distance=1.1cm]
\node (core) {\(\sum\limits_{j=1}^{r-1} a_{ij} k_j\)};
\node [coordinate] (subcore) [below of=core, shift={(-0.2,0)}] {};
\node (cross) [below right of=subcore, shift={(-0.08,0.08)}] {};
\node [coordinate] (out) [below left of=cross, shift={(0.08,-0.08)}] {};
\node (id) [below of=out, shift={(0.2,-0.2)}] {\(k_r\)};
\node [plus] (adder) [below right of=cross, shift={(0.08,-0.38)}] {};
\node (sum) [below of=adder] {\(\sum\limits_{j=1}^{r} a_{ij} k_j\)};
\node [multiply] (mult) [above right of=adder, shift={(-0.4,0.45)}] {\(a_{ir}\)};
\node [delta] (dupe) [above left of=mult, shift={(0.4,0.1)}] {};
\node (in) [above of=dupe, shift={(0,-0.25)}] {\(k_r\)};
\draw (core.270) .. controls +(270:0.5) and +(120:0.5) .. (cross)
(cross) .. controls +(300:0.5) and +(120:0.5) .. (adder.left in)
(dupe.left out) .. controls +(240:0.7) and +(90:1.8) .. (id.90);
\draw (dupe.right out) .. controls +(300:0.2) and +(90:0.1) .. (mult.90)
(mult.io) .. controls +(270:0.1) and +(60:0.2) .. (adder.right in)
(adder) -- (sum) (in) -- (dupe);
\end{tikzpicture}
\end{center}
Since multiplication provides the map \(k_1 \mapsto a_1 k_1\), as in the far left diagram, the
middle-left diagram can be used inductively to form a \(k\)-linear combination of any number of
inputs. In particular, we have any linear map \(S_r \maps k^m \to k\) given by \((k_1, \ldots, k_m)
\mapsto (\sum_j a_{rj} k_j)\). Using duplication as in the middle-right diagram, one can produce
the map \(k_1 \mapsto (k_1, a_{i1} k_1)\), to which the right diagram can be inductively applied.
Thus we can build any linear map, \(T_j \in \mathrm{FinVect}_k\), \(T_j \maps k^m \to k^{m+1}\) given by \((k_1,
\ldots, k_m) \mapsto (k_1, \ldots, k_m, \sum_j a_{ij} k_j)\). If we represent the identity map on
\(k^r\) as \(1^r\), the \(r\)-fold tensor product of the identity map on \(k\), any linear map \(T
\maps k^m \to k^n\) can be given by \((k_1, \ldots, k_m) \mapsto (\sum_j a_{1j} k_j, \ldots, \sum_j
a_{nj} k_j)\), which can be expressed as \(T = (S_1 \oplus 1^{n-1}) (T_2 \oplus 1^{n-2}) \cdots
(T_{n-1} \oplus 1^1) T_n\). The above works as long as the vector spaces are not \(0\)-dimensional.
\(f \maps k^m \to \{0\}\) can be written as an \(m\)-fold tensor product of deletion, \(!^m\), and
\(f \maps \{0\} \to k^n\) can be written as an \(n\)-fold tensor product of zero, \(0^n\). \(f
\maps \{0\} \to \{0\}\) is the empty morphism, which has an empty diagram for its string diagram.
\end{proof}
It is easy to see that the morphisms given in Lemma~\ref{gensvk} obey the following 18 relations:
\vskip 1em \noindent
\textbf{(1)--(3)} Addition and zero make \(k\) into a commutative monoid:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.74cm]
\node [plus] (summer) {};
\node [coordinate] (sum) [below of=summer] {};
\node [coordinate] (Lsum) [above left of=summer] {};
\node [zero] (insert) [above of=Lsum, shift={(0,-0.35)}] {};
\node [coordinate] (Rsum) [above right of=summer] {};
\node [coordinate] (sumin) [above of=Rsum] {};
\node (equal) [right of=Rsum, shift={(0,-0.26)}] {\(=\)};
\node [coordinate] (in) [right of=equal, shift={(0,1)}] {};
\node [coordinate] (out) [right of=equal, shift={(0,-1)}] {};
\draw (insert) .. controls +(270:0.3) and +(120:0.3) .. (summer.left in)
(summer.right in) .. controls +(60:0.6) and +(270:0.6) .. (sumin)
(summer) -- (sum) (in) -- (out);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (uradder) {};
\node [plus] (adder) [below of=uradder, shift={(-0.35,0)}] {};
\node [coordinate] (urm) [above of=uradder, shift={(-0.35,0)}] {};
\node [coordinate] (urr) [above of=uradder, shift={(0.35,0)}] {};
\node [coordinate] (left) [left of=urm] {};
\draw (adder.right in) .. controls +(60:0.2) and +(270:0.1) .. (uradder.io)
(uradder.right in) .. controls +(60:0.35) and +(270:0.3) .. (urr)
(uradder.left in) .. controls +(120:0.35) and +(270:0.3) .. (urm)
(adder.left in) .. controls +(120:0.75) and +(270:0.75) .. (left)
(adder.io) -- +(270:0.5);
\node (eq) [right of=uradder, shift={(0,-0.25)}] {\(=\)};
\node [plus] (ulsummer) [right of=eq, shift={(0,0.25)}] {};
\node [plus] (summer) [below of=ulsummer, shift={(0.35,0)}] {};
\node [coordinate] (ulm) [above of=ulsummer, shift={(0.35,0)}] {};
\node [coordinate] (ull) [above of=ulsummer, shift={(-0.35,0)}] {};
\node [coordinate] (right) [right of=ulm] {};
\draw (summer.left in) .. controls +(120:0.2) and +(270:0.1) .. (ulsummer.io)
(ulsummer.left in) .. controls +(120:0.35) and +(270:0.3) .. (ull)
(ulsummer.right in) .. controls +(60:0.35) and +(270:0.3) .. (ulm)
(summer.right in) .. controls +(60:0.75) and +(270:0.75) .. (right)
(summer.io) -- +(270:0.5);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (twadder) {};
\node [coordinate] (twout) [below of=twadder] {};
\node [coordinate] (twR) [above right of=twadder, shift={(-0.2,0)}] {};
\node (cross) [above of=twadder] {};
\node [coordinate] (twRIn) [above left of=cross, shift={(0,0.3)}] {};
\node [coordinate] (twLIn) [above right of=cross, shift={(0,0.3)}] {};
\draw (twadder.right in) .. controls +(60:0.35) and +(-45:0.25) .. (cross)
.. controls +(135:0.2) and +(270:0.4) .. (twRIn);
\draw (twadder.left in) .. controls +(120:0.35) and +(-135:0.25) .. (cross.center)
.. controls +(45:0.2) and +(270:0.4) .. (twLIn);
\draw (twout) -- (twadder);
\node (eq) [right of=twR] {\(=\)};
\node [coordinate] (L) [right of=eq] {};
\node [plus] (adder) [below right of=L] {};
\node [coordinate] (out) [below of=adder] {};
\node [coordinate] (R) [above right of=adder] {};
\node (cross) [above left of=R] {};
\node [coordinate] (LIn) [above left of=cross] {};
\node [coordinate] (RIn) [above right of=cross] {};
\draw (adder.left in) .. controls +(120:0.7) and +(270:0.7) .. (LIn)
(adder.right in) .. controls +(60:0.7) and +(270:0.7) .. (RIn)
(out) -- (adder);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(4)--(6)} Duplication and deletion make \(k\) into a cocommutative comonoid:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.74cm]
\node [delta] (dupe) {};
\node [coordinate] (top) [above of=dupe] {};
\node [coordinate] (Ldub) [below left of=dupe] {};
\node [bang] (delete) [below of=Ldub, shift={(0,0.35)}] {};
\node [coordinate] (Rdub) [below right of=dupe] {};
\node [coordinate] (dubout) [below of=Rdub] {};
\node (equal) [right of=Rdub, shift={(0,0.26)}] {\(=\)};
\node [coordinate] (in) [right of=equal, shift={(0,1)}] {};
\node [coordinate] (out) [right of=equal, shift={(0,-1)}] {};
\draw (delete) .. controls +(90:0.3) and +(240:0.3) .. (dupe.left out)
(dupe.right out) .. controls +(300:0.6) and +(90:0.6) .. (dubout)
(dupe) -- (top) (in) -- (out);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [delta] (lrduper) {};
\node [delta] (duper) [above of=lrduper, shift={(-0.35,0)}] {};
\node [coordinate](lrm) [below of=lrduper, shift={(-0.35,0)}] {};
\node [coordinate](lrr) [below of=lrduper, shift={(0.35,0)}] {};
\node [coordinate](left) [left of=lrm] {};
\draw (duper.right out) .. controls +(300:0.2) and +(90:0.1) .. (lrduper.io)
(lrduper.right out) .. controls +(300:0.35) and +(90:0.3) .. (lrr)
(lrduper.left out) .. controls +(240:0.35) and +(90:0.3) .. (lrm)
(duper.left out) .. controls +(240:0.75) and +(90:0.75) .. (left)
(duper.io) -- +(90:0.5);
\node (eq) [right of=lrduper, shift={(0,0.25)}] {\(=\)};
\node [delta] (lldubber) [right of=eq, shift={(0,-0.25)}] {};
\node [delta] (dubber) [above of=lldubber, shift={(0.35,0)}] {};
\node [coordinate] (llm) [below of=lldubber, shift={(0.35,0)}] {};
\node [coordinate] (lll) [below of=lldubber, shift={(-0.35,0)}] {};
\node [coordinate] (right) [right of=llm] {};
\draw (dubber.left out) .. controls +(240:0.2) and +(90:0.1) .. (lldubber.io)
(lldubber.left out) .. controls +(240:0.35) and +(90:0.3) .. (lll)
(lldubber.right out) .. controls +(300:0.35) and +(90:0.3) .. (llm)
(dubber.right out) .. controls +(300:0.75) and +(90:0.75) .. (right)
(dubber.io) -- +(90:0.5);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [coordinate] (twtop) {};
\node [delta] (twdupe) [below of=twtop] {};
\node [coordinate] (twR) [below right of=twdupe, shift={(-0.2,0)}] {};
\node (cross) [below of=twdupe] {};
\node [coordinate] (twROut) [below left of=cross, shift={(0,-0.3)}] {};
\node [coordinate] (twLOut) [below right of=cross, shift={(0,-0.3)}] {};
\draw (twdupe.left out) .. controls +(240:0.35) and +(135:0.25) .. (cross)
.. controls +(-45:0.2) and +(90:0.4) .. (twLOut)
(twdupe.right out) .. controls +(300:0.35) and +(45:0.25) .. (cross.center)
.. controls +(-135:0.2) and +(90:0.4) .. (twROut)
(twtop) -- (twdupe);
\node (eq) [right of=twR] {\(=\)};
\node [coordinate] (L) [right of=eq] {};
\node [delta] (dupe) [above right of=L] {};
\node [coordinate] (top) [above of=dupe] {};
\node [coordinate] (R) [below right of=dupe] {};
\node (uncross) [below left of=R] {};
\node [coordinate] (LOut) [below left of=uncross] {};
\node [coordinate] (ROut) [below right of=uncross] {};
\draw (dupe.left out) .. controls +(240:0.7) and +(90:0.7) .. (LOut)
(dupe.right out) .. controls +(300:0.7) and +(90:0.7) .. (ROut)
(top) -- (dupe);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(7)--(10)} The monoid and comonoid structures on \(k\) fit together to form a bimonoid:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node [plus] (adder) {};
\node [coordinate] (f) [above of=adder, shift={(-0.4,-0.325)}] {};
\node [coordinate] (g) [above of=adder, shift={(0.4,-0.325)}] {};
\node [delta] (dupe) [below of=adder, shift={(0,0.25)}] {};
\node [coordinate] (outL) [below of=dupe, shift={(-0.4,0.325)}] {};
\node [coordinate] (outR) [below of=dupe, shift={(0.4,0.325)}] {};
\draw (adder.io) -- (dupe.io)
(f) .. controls +(270:0.4) and +(120:0.25) .. (adder.left in)
(adder.right in) .. controls +(60:0.25) and +(270:0.4) .. (g)
(dupe.left out) .. controls +(240:0.25) and +(90:0.4) .. (outL)
(dupe.right out) .. controls +(300:0.25) and +(90:0.4) .. (outR);
\end{tikzpicture}
\raisebox{2.49em}{=}
\hspace{1em}
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node [plus] (addL) {};
\node (cross) [above right of=addL, shift={(-0.1,-0.0435)}] {};
\node [plus] (addR) [below right of=cross, shift={(-0.1,0.0435)}] {};
\node [delta] (dupeL) [above left of=cross, shift={(0.1,-0.0435)}] {};
\node [delta] (dupeR) [above right of=cross, shift={(-0.1,-0.0435)}] {};
\node [coordinate] (f) [above of=dupeL] {};
\node [coordinate] (g) [above of=dupeR] {};
\node [coordinate] (sum1) [below of=addL, shift={(0,0.2)}] {};
\node [coordinate] (sum2) [below of=addR, shift={(0,0.2)}] {};
\path
(addL) edge (sum1) (addL.right in) edge (dupeR.left out) (addL.left in) edge [bend left=30] (dupeL.left out)
(addR) edge (sum2) (addR.left in) edge (cross) (addR.right in) edge [bend right=30] (dupeR.right out)
(dupeL) edge (f)
(dupeL.right out) edge (cross)
(dupeR) edge (g);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[thick]
\node [zero] (z) at (0,1) {};
\node [delta] (dub) at (0,0.2) {};
\node [coordinate] (oL) at (-0.35,-0.6) {};
\node [coordinate] (oR) at (0.35,-0.6) {};
\node (eq) at (1,0.42) {=};
\node [zero] (zleft) at (2,1) {};
\node [zero] (zright) at (2.7,1) {};
\node [coordinate] (Lo) at (2,-0.6) {};
\node [coordinate] (Ro) at (2.7,-0.6) {};
\draw (z) -- (dub)
(dub.left out) .. controls +(240:0.3) and +(90:0.5) .. (oL)
(dub.right out) .. controls +(300:0.3) and +(90:0.5) .. (oR)
(zleft) -- (Lo)
(zright) -- (Ro);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[thick]
\node [bang] (b) at (0,-1) {};
\node [plus] (sum) at (0,-0.2) {};
\node [coordinate] (oL) at (-0.35,0.6) {};
\node [coordinate] (oR) at (0.35,0.6) {};
\node (spacemaker) at (0,-1.38) {};
\node (eq) at (1,-0.47) {=};
\node [bang] (bleft) at (2,-1) {};
\node [bang] (bright) at (2.7,-1) {};
\node [coordinate] (Lo) at (2,0.6) {};
\node [coordinate] (Ro) at (2.7,0.6) {};
\draw (b) -- (sum)
(sum.left in) .. controls +(120:0.3) and +(270:0.5) .. (oL)
(sum.right in) .. controls +(60:0.3) and +(270:0.5) .. (oR)
(bleft) -- (Lo)
(bright) -- (Ro);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[thick]
\node [zero] (z) at (0,0.11) {};
\node [bang] (b) at (0,-1) {};
\node (spacemaker) at (0,-1.38) {};
\node (eq) at (0.7,-0.47) {=};
\draw (z) -- (b);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(11)--(14)} The rig structure of \(k\) can be recovered from the generating morphisms:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick]
\node (top) {};
\node [multiply] (c) [below of=top] {\(c\)};
\node [multiply] (b) [below of=c] {\(b\)};
\node (bottom) [below of=b] {};
\draw (top) -- (c) -- (b) -- (bottom);
\node (eq) [left of=b, shift={(0.2,0.5)}] {\(=\)};
\node (bctop) [left of=top, shift={(-0.6,0)}] {};
\node [multiply] (bc) [left of=eq, shift={(0.2,0)}] {\(bc\)};
\node (bcbottom) [left of=bottom, shift={(-0.6,0)}] {};
\draw (bctop) -- (bc) -- (bcbottom);
\end{tikzpicture}
\hspace{1.0cm}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (bctop) {};
\node [multiply] (bc) [below of=bctop, shift={(0,-0.59)}] {\(\scriptstyle{b+c}\)};
\node (bcbottom) [below of=bc, shift={(0,-0.59)}] {};
\draw (bctop) -- (bc) -- (bcbottom);
\node (eq) [right of=bc, shift={(0.15,0)}] {\(=\)};
\node [multiply] (b) [right of=eq, shift={(0,0.1)}] {\(\scriptstyle{b}\)};
\node [delta] (dupe) [above right of=b, shift={(-0.2,0.1)}] {};
\node (top) [above of=dupe, shift={(0,-0.2)}] {};
\node [multiply] (c) [below right of=dupe, shift={(-0.2,-0.1)}] {\(\scriptstyle{c}\)};
\node [plus] (adder) [below right of=b, shift={(-0.2,-0.3)}] {};
\node (out) [below of=adder, shift={(0,0.2)}] {};
\draw
(dupe.left out) .. controls +(240:0.15) and +(90:0.15) .. (b.90)
(dupe.right out) .. controls +(300:0.15) and +(90:0.15) .. (c.90)
(top) -- (dupe.io)
(adder.io) -- (out)
(adder.left in) .. controls +(120:0.15) and +(270:0.15) .. (b.io)
(adder.right in) .. controls +(60:0.15) and +(270:0.15) .. (c.io);
\end{tikzpicture}
\hspace{1.0cm}
\raisebox{1.6em}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (top) {};
\node [multiply] (one) [below of=top] {1};
\node (bottom) [below of=one] {};
\draw (top) -- (one) -- (bottom);
\node (eq) [right of=one] {\(=\)};
\node (topid) [right of=top, shift={(0.6,0)}] {};
\node (botid) [right of=bottom, shift={(0.6,0)}] {};
\draw (topid) -- (botid);
\end{tikzpicture}
}
\hspace{1.0cm}
\raisebox{1.6em}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node [multiply] (prod) {\(0\)};
\node (in0) [above of=prod] {};
\node (out0) [below of=prod] {};
\node (eq) [right of=prod] {\(=\)};
\node [bang] (del) [right of=eq, shift={(-0.2,0.2)}] {};
\node [zero] (ins) [right of=eq, shift={(-0.2,-0.2)}] {};
\node (in1) [above of=del, shift={(0,-0.2)}] {};
\node (out1) [below of=ins, shift={(0,0.2)}] {};
\draw (in0) -- (prod) -- (out0);
\draw (in1) -- (del);
\draw (ins) -- (out1);
\end{tikzpicture}
}
}
\end{center}
\vskip 1em \noindent
\textbf{(15)--(16)} Scalar multiplication commutes with addition and zero:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node [plus] (adder) {};
\node (out) [below of=adder, shift={(0,0.2)}] {};
\node [multiply] (L) [above left of=adder, shift={(0.15,0.45)}] {\(c\)};
\node [multiply] (R) [above right of=adder, shift={(-0.15,0.45)}] {\(c\)};
\node [coordinate] (RIn) [above of=R] {};
\node [coordinate] (LIn) [above of=L] {};
\draw (adder.right in) .. controls +(60:0.2) and +(270:0.2) .. (R.io) (R) -- (RIn);
\draw (adder.left in) .. controls +(120:0.2) and +(270:0.2) .. (L.io) (L) -- (LIn);
\draw (out) -- (adder);
\end{tikzpicture}
\hspace{0.1cm}
\begin{tikzpicture}[node distance=1.15cm]
\node (eq){\(=\)};
\node [below of=eq] {};
\end{tikzpicture}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (out) {};
\node [multiply] (c) [above of=out] {\(c\)};
\node [plus] (adder) [above of=c] {};
\node [coordinate] (L) [above left of=adder, shift={(0.15,0.25)}] {};
\node [coordinate] (R) [above right of=adder, shift={(-0.15,0.25)}] {};
\draw (R) .. controls +(270:0.2) and +(60:0.4) .. (adder.right in) (adder) -- (c) -- (out);
\draw (L) .. controls +(270:0.2) and +(120:0.4) .. (adder.left in);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node [multiply] (prod) {\(c\)};
\node (out0) [below of=prod] {};
\node [zero] (ins0) [above of=prod] {};
\node (eq) [right of=prod] {\(=\)};
\node (out1) [below right of=eq] {};
\node [zero] (ins1) [above of=out1, shift={(0,0.2)}] {};
\draw (out0) -- (prod) -- (ins0);
\draw (out1) -- (ins1);
\end{tikzpicture}
\hspace{0.7cm}
}
\end{center}
\vskip 1em \noindent
\textbf{(17)--(18)} Scalar multiplication commutes with duplication and deletion:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (top) {};
\node [delta] (dupe) [below of=top, shift={(0,0.2)}] {};
\node [multiply] (L) [below left of=dupe, shift={(0.15,-0.45)}] {\(c\)};
\node [multiply] (R) [below right of=dupe, shift={(-0.15,-0.45)}] {\(c\)};
\node [coordinate] (ROut) [below of=R] {};
\node [coordinate] (LOut) [below of=L] {};
\draw (dupe.left out) .. controls +(240:0.2) and +(90:0.2) .. (L.90) (L) -- (LOut);
\draw (dupe.right out) .. controls +(300:0.2) and +(90:0.2) .. (R.90) (R) -- (ROut);
\draw (top) -- (dupe);
\end{tikzpicture}
\hspace{0.1cm}
\begin{tikzpicture}[node distance=1.15cm]
\node (eq){\(=\)};
\node [below of=eq] {};
\end{tikzpicture}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node (top) {};
\node [multiply] (c) [below of=top] {\(c\)};
\node [delta] (dupe) [below of=c] {};
\node [coordinate] (L) [below left of=dupe, shift={(0.15,-0.25)}] {};
\node [coordinate] (R) [below right of=dupe, shift={(-0.15,-0.25)}] {};
\draw (dupe.left out) .. controls +(240:0.2) and +(90:0.2) .. (L);
\draw (dupe.right out) .. controls +(300:0.2) and +(90:0.2) .. (R);
\draw (top) -- (c) -- (dupe);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[-, thick, node distance=0.85cm]
\node [multiply] (prod) {\(c\)};
\node (in0) [above of=prod] {};
\node [bang] (del0) [below of=prod] {};
\node (eq) [right of=prod] {\(=\)};
\node (in1) [above right of=eq] {};
\node [bang] (del1) [below of=in1, shift={(0,-0.2)}] {};
\node [below of=del1, shift={(0,-0.2)}] {};
\draw (in0) -- (prod) -- (del0);
\draw (in1) -- (del1);
\end{tikzpicture}
}
\end{center}
In fact, these relations are enough. That is, together with the generating objects and morphisms,
they give a `presentation' of \(\mathrm{FinVect}_k\) as a symmetric monoidal category. However, we need to
make this concept precise.
Suppose \(C\) is generated by a set \(O\) of objects and a set \(M\) of morphisms going between
tensor products of objects in \(O\). Define a \Define{formal morphism} to be a formal expression
built from symbols for morphisms in \(M\) via composition, identity morphisms, tensor product, the
unit object and the braiding. Any formal morphism \(f\) can be \Define{evaluated} to obtain a
morphism \(\mathrm{ev}(f)\) in \(C\), which actually lies in \(C_0\).
Define a \Define{relation} to be a pair \(f,g\) of formal morphisms. We say the relation
\Define{holds} in \(C\) if \(\mathrm{ev}(f) = \mathrm{ev}(g) \). Suppose \(R\) is a set of relations that hold in
\(C\). We say \((O,M,R)\) is a \Define{presentation} of \(C\) if given any two formal morphisms
\(j,k\) that evaluate to the same morphism, then we can go from \(j\) to \(k\) via a finite
sequence of moves of these kinds:
\begin{enumerate}
\item replacing an instance of a generating morphism \(f\) in a formal morphism by the generating
morphism \(g\), where \((f,g) \in R\),
\item applying an equational law in the definition of strict symmetric monoidal category to a
formal morphism.
\end{enumerate}
In intuitive terms, this means that there are enough relations to prove all the equations
that hold in \(C\)---or more precisely, in the equivalent category \(C_0\).
\begin{thm} \label{presvk} The symmetric monoidal category \(\mathrm{FinVect}_k\) is presented by the object
\(k\), the morphisms given in Lemma~\ref{gensvk}, and relations {\bf (1)--(18)} as listed above.
\end{thm}
\begin{proof}
To prove this, we show that these relations suffice to rewrite any formal morphism into a
standard form, with all formal morphisms that evaluate to the same morphism \(T \maps k^m \to k^n\)
in \(\mathrm{FinVect}_k\) having the same standard form. To deal with moves of type (2), we draw formal
morphisms as string diagrams built from generating morphisms and the braiding. Two formal
morphisms that differ only by equational laws in the definition of strict symmetric monoidal
category will have topologically equivalent string diagrams. It suffices, then, to show that any
string diagram built from generating morphisms and the braiding can be put into a standard form
using topological equivalences and relations \textbf{(1)--(18)}.
A qualitative description of this standard form will be helpful for understanding how an arbitrary
string diagram can be rewritten in this form. By way of example, consider the linear
transformation \(T \maps {\mathbb R}^3 \to {\mathbb R}^2\) given by
\[ (x_1, x_2, x_3) \mapsto (y_1, y_2) = (3x_1 + 7x_2 +
2x_3, 9x_1 + x_2).\]
Its standard form looks like this:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node (i1) at (-2.5,3.4) {\(x_1\)};
\node (i2) at (-0.5,3.4) {\(x_2\)};
\node (i3) at (1.5,3.4) {\(x_3\)};
\node[multiply] (a11) at (-3,1.732) {\(3\)};
\node[multiply] (a21) at (-2,1.732) {\(9\)};
\node[multiply] (a12) at (-1,1.732) {\(7\)};
\node[multiply] (a22) at (0,1.732) {\(1\)};
\node[multiply] (a13) at (1,1.732) {\(2\)};
\node[multiply] (a23) at (2,1.732) {\(0\)};
\node[delta] (D1) at (-2.5,2.6) {};
\node[delta] (D2) at (-0.5,2.6) {};
\node[delta] (D3) at (1.5,2.6) {};
\node[plus] (S1T) at (-2,0.31) {};
\node[plus] (S1B) at (-1,-0.556) {};
\node[plus] (S2T) at (-1,0.31) {};
\node[plus] (S2B) at (0,-0.556) {};
\node (o1) at (-1,-1.24) {\(y_1\)};
\node (o2) at (0,-1.24) {\(y_2\)};
\draw
(a11.io) .. controls +(270:0.2) and +(120:0.32) .. (S1T.left in)
(a21.io) .. controls +(270:0.2) and +(120:0.32) .. (S2T.left in)
(a22.io) .. controls +(270:0.2) and +(60:0.32) .. (S2T.right in)
(S1T.io) .. controls +(270:0.2) and +(120:0.2) .. (S1B.left in)
(i1) -- (D1) (i2) -- (D2) (i3) -- (D3)
(S1B) -- (o1) (S2B) -- (o2)
(S2T.io) .. controls +(270:0.2) and +(120:0.2) .. (S2B.left in)
(a23.io) .. controls +(270:0.5) and +(60:0.5) .. (S2B.right in);
\node [hole] (cross2) at (-0.59,-0.2) {};
\node [hole] (cross1) at (-1.5,0.75) {};
\draw
(a12.io) .. controls +(270:0.2) and +(60:0.32) .. (S1T.right in)
(a13.io) .. controls +(270:0.5) and +(60:0.5) .. (S1B.right in)
(D1.left out) .. controls +(240:0.2) and +(90:0.2) .. (a11.90)
(D1.right out) .. controls +(300:0.2) and +(90:0.2) .. (a21.90)
(D2.left out) .. controls +(240:0.2) and +(90:0.2) .. (a12.90)
(D2.right out) .. controls +(300:0.2) and +(90:0.2) .. (a22.90)
(D3.left out) .. controls +(240:0.2) and +(90:0.2) .. (a13.90)
(D3.right out) .. controls +(300:0.2) and +(90:0.2) .. (a23.90);
\end{tikzpicture}
}
\end{center}
This is a string diagram picture of the following equation:
\[ T x =
\left( \begin{array}{ccc}
3 & 7 & 2 \\
9 & 1 & 0 \end{array} \right)
\left( \begin{array}{c} x_1 \\ x_2 \\ x_3 \end{array} \right) =
\left( \begin{array}{c} y_1 \\ y_2 \end{array} \right)
\]
In general, given a \(k\)-linear transformation \(T \maps k^m \to k^n\), we can describe it using an
\(n \times m\) matrix with entries in \(k\). The case where \(m\) and/or \(n\) is zero gives a
matrix with no entries, so their standard form will be treated separately. For positive values of
\(m\) and \(n\), the standard form has three distinct layers. The top layer consists of \(m\)
clusters of \(n-1\) instances of \(\Delta\). The middle layer is \(mn\) multiplications. The \(n\)
outputs of the \(j\)th cluster connect to the inputs of the multiplications \(\{a_{1j}, \dotsc,
a_{nj}\}\), where \(a_{ij}\) is the \(ij\) entry of \(A\), the matrix for \(T\). The bottom layer
consists of \(n\) clusters of \(m-1\) instances of \(+\). There will generally be braiding in this
layer as well, but since the category is being generated as symmetric monoidal, the locations of the
braidings doesn't matter so long as the topology of the string diagram is preserved. The topology
of the sum layer is that the \(i\)th sum cluster gets its \(m\) inputs from the outputs of the
multiplications \(\{a_{i1}, \dotsc, a_{im}\}\). The arrangement of the instances of \(\Delta\) and
\(+\) within their respective clusters does not matter, due to the associativity of \(+\) via
relation \textbf{(2)} and coassociativity of \(\Delta\) via relation \textbf{(5)}. For the sake of
making the standard form explicit with respect to these relations, we may assume the right output of
a \(\Delta\) is always connected to a multiplication input, and the right input of a \(+\) is always
connected to a multiplication output. This gives a prescription for drawing the standard form of a
string diagram with a corresponding matrix \(A\).
The standard form for \(T \maps k^0 \to k^n\) is \(n\) zeros (\(0 \oplus \dotsb \oplus 0\)), and the
standard form for \(T \maps k^m \to k^0\) is \(m\) deletions (\(! \oplus \dotsb \oplus\, !\)).
Each of the generating morphisms can easily be put into standard form: the string diagrams for zero,
deletion, and multiplication are already in standard form. The string diagram for duplication
(resp. addition) can be put into standard form by attaching a multiplication by \(1\), relation
\textbf{(13)}, to each of the outputs (resp. inputs).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[delta] (dupe) {};
\node[multiply] (o1) at (-0.5,-0.65) {\(1\)};
\node[multiply] (o2) at (0.5,-0.65) {\(1\)};
\node (in) [above of=dupe] {};
\node (posto1) [below of=o1] {};
\node (posto2) [below of=o2] {};
\draw[rounded corners] (posto1) -- (o1) -- (dupe.left out);
\draw[rounded corners] (posto2) -- (o2) -- (dupe.right out);
\draw (in) -- (dupe);
\node (equals) at (-1,0) {\(=\)};
\node[delta] (dub) at (-2,0) {};
\node (dubin) [above of=dub] {};
\node (L) at (-2.5,-0.65) {};
\node (R) at (-1.5,-0.65) {};
\draw (L) -- (dub.left out) (R) -- (dub.right out) (dubin) -- (dub);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[thick]
\node[plus] (summer) {};
\node[multiply] (o1) at (-0.5,1.02) {\(1\)};
\node[multiply] (o2) at (0.5,1.02) {\(1\)};
\node (out) [below of=summer] {};
\node (posto1) [above of=o1] {};
\node (posto2) [above of=o2] {};
\draw (posto1) -- (o1) -- (o1.io) -- (summer.left in);
\draw (posto2) -- (o2) -- (o2.io) -- (summer.right in);
\draw (out) -- (summer);
\node (equals) at (-1,0) {\(=\)};
\node[plus] (adder) at (-2,0) {};
\node (addout) [below of=adder] {};
\node (L) at (-2.5,0.65) {};
\node (R) at (-1.5,0.65) {};
\draw (L) -- (adder.left in) (R) -- (adder.right in) (addout) -- (adder);
\end{tikzpicture}
}
\end{center}
The braiding morphism is just as basic to our argument as the generating morphisms, so we will need
to write the string diagram for \(B\) in standard form as well. The matrix corresponding to
braiding is
\[\left( \begin{array}{cc}
0 & 1 \\
1 & 0 \end{array} \right),\]
so its standard form is as follows:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node (UpUpLeft) at (-3.7,2.2) {};
\node [coordinate] (UpLeft) at (-3.7,1.7) {};
\node (mid) at (-3.3,1.3) {};
\node [coordinate] (DownRight) at (-2.9,0.9) {};
\node (DownDownRight) at (-2.9,0.4) {};
\node [coordinate] (UpRight) at (-2.9,1.7) {};
\node (UpUpRight) at (-2.9,2.2) {};
\node [coordinate] (DownLeft) at (-3.7,0.9) {};
\node (DownDownLeft) at (-3.7,0.4) {};
\draw [rounded corners=2mm] (UpUpLeft) -- (UpLeft) -- (mid) --
(DownRight) -- (DownDownRight) (UpUpRight) -- (UpRight) -- (DownLeft) -- (DownDownLeft);
\node (eq) at (-2.2,1.3) {\(=\)};
\node[coordinate] (i1) at (-1,3) {};
\node[coordinate] (i2) at (1,3) {};
\node[multiply] (a11) at (-1.5,1.832) {\(0\)};
\node[multiply] (a21) at (-0.5,1.832) {\(1\)};
\node[multiply] (a12) at (0.5,1.832) {\(1\)};
\node[multiply] (a22) at (1.5,1.832) {\(0\)};
\node[delta] (D1) at (-1,2.4) {};
\node[delta] (D2) at (1,2.4) {};
\node (cross) at (0,0.75) {};
\node[plus] (S1) at (-0.5,0.2) {};
\node[plus] (S2) at (0.5,0.2) {};
\node[coordinate] (o1) at (-0.5,-0.4) {};
\node[coordinate] (o2) at (0.5,-0.4) {};
\draw
(i1) -- (D1) (i2) -- (D2)
(S1.io) -- (o1) (S2.io) -- (o2)
(a11.io) -- (S1.left in) (a22.io) -- (S2.right in)
(a21.io) -- (cross) -- (S2.left in) (a12.io) -- (S1.right in)
(D1.left out) -- (a11) (D1.right out) -- (a21)
(D2.left out) -- (a12) (D2.right out) -- (a22);
\end{tikzpicture}
}
\end{center}
For \(n > 1\), any morphism built from \(n\) copies of the \Define{basic morphisms}---that is,
generating morphisms and the braiding---can be built up from a morphism built from \(n-1\) copies by
composing or tensoring with one more basic morphism. Thus, to prove that any string diagram built
from basic morphisms can be put into its standard form, we can proceed by induction on the number of
basic morphisms.
Furthermore, because strings can be extended using the identity morphism, relation \textbf{(13)} can
be used to show tensoring with any generating morphism is equivalent to tensoring with \(1\),
followed by a composition: \(\Delta = \Delta \circ 1\), \(+ = 1 \circ +\), \(c = 1 \circ c\), \(! =\, !
\circ 1\), \(0 = 1 \circ 0\). In the case of braiding, the step of tensoring with \(1\) is repeated
once before making the composition: \(B = (1 \oplus 1) \circ B\).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[blackbox] (bb1) {};
\node (tensor1) [right of=bb1] {\(\oplus\)};
\node[sqnode] (gen1) [right of=tensor1] {G};
\node (eq1) [right of=gen1] {\(=\)};
\node[blackbox] (bb2) [right of=eq1] {};
\node (tensor2) [right of=bb2] {\(\oplus\)};
\node[sqnode] (gen2) [right of=tensor2, shift={(0,1)}] {G};
\node (eq2) [right of=gen2, shift={(0,-1)}] {\(=\)};
\node[blackbox] (bb3) [right of=eq2] {};
\node (tensor3) [right of=bb3] {\(\oplus\)};
\node[sqnode] (gen3) [right of=tensor3, shift={(0,-1)}] {G};
\node[multiply] (times2) [below of=gen2] {\(1\)};
\node[multiply] (times3) [above of=gen3] {\(1\)};
\node (bb1in) [above of=bb1] {};
\node (bb1out) [below of=bb1] {};
\node (bb2in) [above of=bb2] {};
\node (bb2out) [below of=bb2] {};
\node (bb3in) [above of=bb3] {};
\node (bb3out) [below of=bb3] {};
\node (gen1in) [above of=gen1] {};
\node (gen1out) [below of=gen1] {};
\node (gen2in) [above of=gen2] {};
\node (gen2out) [below of=times2] {};
\node (gen3in) [above of=times3] {};
\node (gen3out) [below of=gen3] {};
\draw (bb1in) -- (bb1) -- (bb1out) (gen1in) -- (gen1) -- (gen1out)
(bb2in) -- (bb2) -- (bb2out) (gen2in) -- (gen2) -- (times2) -- (gen2out)
(bb3in) -- (bb3) -- (bb3out) (gen3in) -- (times3) -- (gen3) -- (gen3out);
\end{tikzpicture}
}
\end{center}
Thus there are 11 cases to consider for this induction: \(\oplus 1\), \(+ \circ\), \(\circ \Delta\),
\(\Delta \circ\), \(\circ +\), \(\circ c\), \(c \circ\), \(\circ 0\), \(! \circ\), \(B \circ\), \(\circ B\).
Without loss of generality, the string diagram \(S\) to which a generating morphism is added will be
assumed to be in standard form already. Labels \(ij\) on diagrams illustrating these cases
correspond to strings incident to the multiplications \(a_{ij}\).
\begin{itemize}[leftmargin=1em]
\item \Define{\(\oplus 1\)}\\
When tensoring morphisms together, the matrix corresponding to \(C \oplus D\) is the block diagonal
matrix
\[\left( \begin{array}{cc}
C & 0 \\
0 & D \end{array} \right),\]
where, by abuse of notation, the block \(C\) is the matrix corresponding to morphism \(C\), and
respectively \(D\) with \(D\). Thus, when tensoring \(S\) by \(1\), we write the matrix for \(S\)
with one extra row and one extra column. Each of these new entries will be \(0\) with the exception
of a \(1\) at the bottom of the extra column. The string diagram corresponding to the new matrix
can be drawn in standard form as prescribed above. Using relations \textbf{(14)}, \textbf{(4)}, and
\textbf{(1)}, the standard form reduces to \(S \oplus 1\). The process is reversable (\(\mathrm{ev}(f) =
\mathrm{ev}(g)\) implies \(\mathrm{ev}(g) = \mathrm{ev}(f)\)), so if the string diagram \(S\) can be drawn in standard form,
the string diagram \(S \oplus 1\) can be drawn in standard form, too. The diagrams below show the
relevant strings before they are reduced.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[delta] (D1) {};
\node[delta] (D2) at (-0.5,-0.866) {};
\node[delta] (D3) at (-1,-1.732) {};
\node[multiply] (zero) at (0.5,-.65) {\(0\)};
\node (i1) [above of=D1] {};
\node (o1) at (0.5,-1.516) {\(\quad\scriptstyle{n+1,j}\)};
\node (o2) at (0,-1.516) {\(\scriptstyle{nj}\)};
\node (o3) at (-0.5,-2.382) {\(\scriptstyle{2j}\)};
\node (o4) at (-1.5,-2.382) {\(\scriptstyle{1j}\)};
\draw (i1) -- (D1) (D1.left out) -- (D2.io) (D1.right out) -- (zero)
-- (o1) (D2.right out) -- (o2) (D3.right out) -- (o3) (D3.left out) -- (o4);
\draw[dotted] (D2.left out) -- (D3.io);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[thick]
\node[delta] (D1) {};
\node[delta] (D2) at (-0.5,-0.866) {};
\node[delta] (D3) at (-1,-1.732) {};
\node[multiply] (one) at (0.5,-.65) {\(1\)};
\node (o1) at (0.5,-1.516) {\(\quad\qquad\scriptstyle{n+1,m+1}\)};
\node (i1) [above of=D1] {};
\node[multiply] (z2) at (0,-1.516) {\(0\)};
\node (o2) at (0,-2.382) {\(\qquad\scriptstyle{n,m+1}\)};
\node[multiply] (z3) at (-0.5,-2.382) {\(0\)};
\node (o3) at (-0.5,-3.248) {\(\quad\scriptstyle{2,m+1}\)};
\node[multiply] (z4) at (-1.5,-2.382) {\(0\)};
\node (o4) at (-1.5,-3.248) {\(\scriptstyle{1,m+1}\)};
\draw (i1) -- (D1) (D1.left out) -- (D2.io) (D1.right out) -- (one)
-- (o1) (D2.right out) -- (z2) -- (o2) (D3.right out) -- (z3) --
(o3) (D3.left out) -- (z4) -- (o4);
\draw[dotted] (D2.left out) -- (D3.io);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[thick]
\node[plus] (P1) {};
\node[plus] (P2) at (-0.5,0.866) {};
\node[plus] (P3) at (-1,1.732) {};
\node (zero) at (0.5,.65) {\(\quad\scriptstyle{i,m+1}\)};
\node (o1) [below of=P1] {};
\node (i2) at (0,1.516) {\(\scriptstyle{im}\)};
\node (i3) at (-0.5,2.382) {\(\scriptstyle{i2}\)};
\node (i4) at (-1.5,2.382) {\(\scriptstyle{i1}\)};
\draw (o1) -- (P1) (P1.left in) -- (P2.io) (P1.right in) -- (zero)
(P2.right in) -- (i2) (P3.right in) -- (i3) (P3.left in) -- (i4);
\draw[dotted] (P2.left in) -- (P3.io);
\end{tikzpicture}
}
\end{center}
Note that for \(i = n+1\) the multiplications \(a_{i2}, \dotsc, a_{im}\) going to the sum cluster
will be multiplication by zero, and \(a_{i,m+1} = 1\). Otherwise \(a_{i,m+1} = 0\), and the rest
depend on the matrix corresponding to \(S\). When \(S = (! \oplus \dotsb \oplus !)\), the matrix
corresponding to \(S \oplus 1\) has a single row, \((0 \cdots 0 \; 1)\), and the standard form
generated is just the middle diagram above. When the same simplifications are applied, no sum
cluster exists to eliminate the zeros, so the standard form still simplifies to \(S \oplus 1\).
Dually, when \(S = (0 \oplus \dotsb \oplus 0)\), the matrix representation of \(S \oplus 1\) is a
column matrix. No duplication cluster exists in the standard form for this matrix, so the same
simplifications again reduce to \(S \oplus 1\).
\item \Define{\(+ \circ\)}\\
If we compose the string diagram for addition with \(S\), first consider only the affected clusters
of additions: two clusters are combined into a larger cluster. Without loss of generality we can
assume these are the first two clusters, or formally, \((+ \oplus 1^{n-2})(S)\). We can rearrange
the sums using the associative law, relation \textbf{(2)}, and permute the inputs of this
large cluster using the commutative law, relation \textbf{(3)}. After several iterations of these
two relations, the desired result is obtained:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[plus] (base) {};
\node (out) [below of=base] {};
\node (Laim) at (-0.25,0.217) {};
\node (Lbase) at (-0.87,0.433) {};
\node[plus] (L1) at (-1.0,0.866) {};
\node[plus] (L2) at (-1.5,1.732) {};
\node[plus] (L3) at (-2.0,2.598) {};
\node (Raim) at (0.25,0.217) {};
\node (Rbase) at (0.87,0.433) {};
\node[plus] (R1) at (1.0,0.866) {};
\node[plus] (R2) at (0.5,1.732) {};
\node[plus] (R3) at (0.0,2.598) {};
\node (iL1) at (-0.5,1.516) {\(\scriptstyle{1m}\)};
\node (iL2) at (-1.0,2.382) {};
\node (iL3) at (-1.5,3.248) {\(\scriptstyle{12}\)};
\node (iL4) at (-2.5,3.248) {\(\scriptstyle{11}\)};
\node (iR1) at (1.5,1.516) {\(\scriptstyle{2m}\)};
\node (iR2) at (1.0,2.382) {};
\node (iR3) at (0.5,3.248) {\(\scriptstyle{22}\)};
\node (iR4) at (-0.5,3.248) {\(\scriptstyle{21}\)};
\draw (base.left in) .. controls (Laim) and (Lbase) .. (L1.io);
\draw (base.right in) .. controls (Raim) and (Rbase) .. (R1.io);
\draw (R1.left in) -- (R2.io) (L1.left in) -- (L2.io) (out) -- (base)
(L1.right in) -- (iL1) (L2.right in) -- (iL2) (L3.right in) -- (iL3) (L3.left in) -- (iL4)
(R1.right in) -- (iR1) (R2.right in) -- (iR2) (R3.right in) -- (iR3) (R3.left in) -- (iR4);
\draw[dotted] (R2.left in) -- (R3.io) (L2.left in) -- (L3.io);
\node (eq) at (2.5,1.3) {\(=\)};
\node[plus] (base2) at (6,0) {};
\node (basement) [below of=base2] {};
\node[plus] (R) at (7,0.866) {};
\node[plus] (L) at (5,0.866) {};
\node[plus] (LR) at (6,1.732) {};
\node[plus] (LL) at (4,1.732) {};
\node[plus] (LLL) at (3,2.598) {};
\node[plus] (LLR) at (5,2.598) {};
\node (iRl) at (6.5,1.516) {\(\scriptstyle{1m}\)};
\node (iRr) at (7.5,1.516) {\(\scriptstyle{2m}\)};
\node (iLRl) at (5.5,2.382) {};
\node (iLRr) at (6.5,2.382) {};
\node (iLLRl) at (4.5,3.248) {\(\scriptstyle{12}\)};
\node (iLLRr) at (5.5,3.248) {\(\scriptstyle{22}\)};
\node (iLLLl) at (2.5,3.248) {\(\scriptstyle{11}\)};
\node (iLLLr) at (3.5,3.248) {\(\scriptstyle{21}\)};
\draw (base2.left in) .. controls (5.75,0.217) and (5.13,0.433) .. (L.io)
(base2.right in) .. controls (6.25,0.217) and (6.87,0.433) .. (R.io)
(base2) -- (basement) (R.right in) -- (iRr) (R.left in) -- (iRl)
(LR.right in) -- (iLRr) (LR.left in) -- (iLRl)
(LLR.right in) -- (iLLRr) (LLR.left in) -- (iLLRl)
(LLL.right in) -- (iLLLr) (LLL.left in) -- (iLLLl)
(L.right in) .. controls (5.25,1.083) and (5.87,1.299) .. (LR.io)
(LL.left in) .. controls (3.75,1.949) and (3.13,2.165) .. (LLL.io)
(LL.right in) .. controls (4.25,1.949) and (4.87,2.165) .. (LLR.io);
\draw[dotted] (L.left in) .. controls (4.75,1.083) and (4.13,1.299) .. (LL.io);
\end{tikzpicture}
}
\end{center}
Now the right side of relation \textbf{(12)} appears in the diagram \(m\) times with \(a_{1j}\) and
\(a_{2j}\) in place of \(b\) and \(c\). Relation \textbf{(12)} can therefore be used to simplify
to the multiplications \(a_{1j}+a_{2j}\).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[delta] (dub) {};
\node (top) [above of=dub] {};
\node[multiply] (a1j) at (-0.5,-0.65) {\(\!\scriptstyle{a_{1j}}\!\)};
\node[multiply] (a2j) at (0.5,-0.65) {\(\!\scriptstyle{a_{2j}}\!\)};
\node[plus] (sum) at (0,-1.516) {};
\node (bottom) [below of=sum] {};
\draw (top) -- (dub) (dub.left out) -- (a1j)
(dub.right out) -- (a2j) (sum) -- (bottom)
(a1j.io) .. controls (-0.47,-1.15) and (-0.25,-1.299) .. (sum.left in)
(a2j.io) .. controls (0.47,-1.15) and (0.25,-1.299) .. (sum.right in);
\node (eq) at (1.2,-0.65) {\(=\)};
\node[multiply] (added) at (2.4,-0.65) {\(\!\!\!{}^{a_{1j}+a_{2j}}\!\!\!\)};
\node (in) at (2.4,1) {};
\node (out) at (2.4,-2.516) {};
\draw (in) -- (added) -- (out);
\end{tikzpicture}
}
\end{center}
The simplification removes one instance of \(\Delta\) from each of the \(m\) clusters of \(\Delta\)
and \(m\) instances of \(+\) from the large addition cluster. There will remain \((m-1) + (m-1) +
(1) - (m) = m-1\) instances of \(+\), which is the correct number for the cluster. I.e. the
composition has been reduced to standard form.
\\
The argument is vastly simpler if \(S = (0 \oplus \dotsb \oplus 0)\). In that case relation
\textbf{(1)} deletes the addition and one of the \(0\) morphisms, and \(S\) is still in the same
form.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node [plus] (sum) at (0,0) {};
\node [zero] (zero1) at (-0.5,0.65) {};
\node [zero] (zero2) at (0.5,0.65) {};
\node (eq) at (1,0.15) {\(=\)};
\node [zero] (zero3) at (1.5,0.65) {};
\node (sumout) at (0,-0.65) {};
\node (zeroout) at (1.5,-0.65) {};
\draw (zero1) -- (sum.left in) (sum.right in) -- (zero2) (sum.io)
-- (sumout) (zero3) -- (zeroout);
\end{tikzpicture}
}
\end{center}
\item \Define{\(\circ \Delta\)}\\
The argument for \(S \circ (\Delta \oplus 1^{m-2})\) is dual to the above argument, using the light
relations \textbf{(4)}, \textbf{(5)} and \textbf{(6)} instead of the dark relations \textbf{(1)},
\textbf{(2)} and \textbf{(3)}.
\item \Define{\(\Delta \circ\)}\\
For \((\Delta \oplus 1^{n-1}) \circ S\), relation \textbf{(7)} can be used iteratively to ``float''
the \(\Delta\) layer above each of the two \(+\) clusters formed by the first iteration.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[plus] (base) {};
\node[delta] (dub) at (0,-0.866) {};
\node (oLa) at (-0.5,-1.516) {};
\node (oRa) at (0.5,-1.516) {};
\node (1ma) at (0.5,0.65) {\(\scriptstyle{1m}\)};
\node[plus] (top) at (-0.5,0.866) {};
\node (11a) at (-1,1.516) {\(\scriptstyle{11}\)};
\node (12a) at (0,1.516) {\(\scriptstyle{12}\)};
\draw[dotted] (base.left in) -- (top.io);
\draw (oLa) -- (dub.left out) (oRa) -- (dub.right out) (dub) -- (base)
(base.right in) -- (1ma) (top.right in) -- (12a) (top.left in) -- (11a);
\node (eq1) at (0.875,0) {\(=\)};
\node (oLb) at (1.625,-1.516) {};
\node[plus] (baseLb) at (1.625,-0.866) {};
\node[delta] (dubLb) at (1.625,0) {};
\node[plus] (topb) at (1.625,0.866) {};
\node (11b) at (1.125,1.516) {\(\scriptstyle{11}\)};
\node (12b) at (2.125,1.516) {\(\scriptstyle{12}\)};
\node (oRb) at (2.5,-1.516) {};
\node[plus] (baseRb) at (2.5,-0.866) {};
\node[delta] (dubRb) at (2.5,0) {};
\node (1mb) at (2.5,0.65) {\(\scriptstyle{1m}\)};
\node (crossb) at (2.0625,-0.433) {};
\draw[dotted] (dubLb) -- (topb);
\draw (11b) -- (topb.left in) (topb.right in) -- (12b) (dubRb) -- (1mb)
(dubLb.right out) -- (crossb) -- (baseRb.left in)
(dubRb.left out) -- (baseLb.right in)
(baseRb) -- (oRb) (baseLb) -- (oLb);
\path (baseLb.left in) edge [bend left=30] (dubLb.left out)
(baseRb.right in) edge [bend right=30] (dubRb.right out);
\node (eq2) at (3.25,0) {\(=\)};
\node (11c) at (4.125,1.516) {\(\scriptstyle{11}\)};
\node[delta] (dubLc) at (4.125,0.866) {};
\node[plus] (topLc) at (4.125,0) {};
\node (ucross) at (4.5625,0.433) {};
\node (12c) at (5,1.516) {\(\scriptstyle{12}\)};
\node[delta] (dubRc) at (5,0.866) {};
\node[plus] (topRc) at (5,0) {};
\node[plus] (baseLc) at (4.625,-0.866) {};
\node (oLc) at (4.625,-1.516) {};
\node (1mc) at (6,1.516) {\(\scriptstyle{1m}\)};
\node (dots) at (5.5,1.516) {\(\cdots\)};
\node[delta] (dubmc) at (6,0.866) {};
\node[plus] (baseRc) at (5.5,-0.866) {};
\node (oRc) at (5.5,-1.516) {};
\draw[dotted] (baseLc.left in) -- (topLc.io) (baseRc.left in) -- (topRc.io);
\draw (oRc) -- (baseRc) (oLc) -- (baseLc) (dubmc) -- (1mc) (dubLc) -- (11c) (dubRc) -- (12c)
(topLc.right in) -- (dubRc.left out)
(topRc.left in) -- (ucross) -- (dubLc.right out)
(baseLc.right in) -- (dubmc.left out)
(baseRc.right in) .. controls (5.875,-0.433) and (6.625,0) .. (dubmc.right out);
\path (topLc.left in) edge [bend left=30] (dubLc.left out)
(topRc.right in) edge [bend right=30] (dubRc.right out);
\end{tikzpicture}
}
\end{center}
Each of these instances of \(\Delta\) can pass through the multiplication layer to \(\Delta\)
clusters using relation \textbf{(17)}.
\\
As before, we consider the subcase \(S = (0 \oplus \dotsb \oplus 0)\) separately. Relation
\textbf{(8)} removes the duplication and creates a new zero, so \(S\) remains in the same form.
\item \Define{\(\circ +\)}\\
For \(S(+ \oplus 1^{m-1})\), the argument is dual to the previous one: relation \textbf{(7)} is used
to ``float'' the additions down, relation \textbf{(15)} sends the additions through the
multiplications, and relation \textbf{(9)} removes the addition and creates a new deletion in the
subcase \(S = (! \oplus \dotsb \oplus !)\).
\item \Define{\(\circ c\)}\\
We can iterate relation \textbf{(17)} when a multiplication is composed on top, as in \(S(c \oplus
1^{m-1})\).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[multiply] (topc) {\(c\)};
\node (top1) at (0,0.866) {};
\node[delta] (dub1) [below of=topc] {};
\node[delta] (dub2) at (-0.5,-1.866) {};
\node (11a) at (-1,-2.516) {\(\scriptstyle{11}\)};
\node (21a) at (0,-2.516) {\(\scriptstyle{21}\)};
\node (n1a) at (0.5,-1.65) {\(\scriptstyle{n1}\)};
\draw[dotted] (dub1.left out) -- (dub2.io);
\draw (top1) -- (topc) -- (dub1) (dub1.right out) -- (n1a)
(dub2.left out) -- (11a) (dub2.right out) -- (21a);
\node (eq) at (1,-1) {\(=\)};
\node (top2) at (2.5,0.866) {};
\node[delta] (delt1) at (2.5,-0.134) {};
\node[delta] (delt2) at (2,-1) {};
\node[multiply] (c1) at (1.5,-1.65) {\(c\)};
\node[multiply] (c2) at (2.5,-1.65) {\(c\)};
\node[multiply] (c3) at (3.5,-1.65) {\(c\)};
\node (11b) at (1.5,-2.516) {\(\scriptstyle{11}\)};
\node (21b) at (2.5,-2.516) {\(\scriptstyle{21}\)};
\node (dots) at (3,-2.516) {\(\cdots\)};
\node (n1b) at (3.5,-2.516) {\(\scriptstyle{n1}\)};
\draw[dotted] (delt1.left out) -- (delt2.io);
\draw (delt2.left out) -- (c1) -- (11b) (delt2.right out) -- (c2) -- (21b)
(delt1.right out) -- (c3) -- (n1b) (top2) -- (delt1);
\end{tikzpicture}
}
\end{center}
The double multiplications in the multiplication layer reduce to a single multiplication via
relation \textbf{(11)}, \(c \circ a_{ij} = ca_{ij}\), which leaves the diagram in standard form. The
composition does nothing when \(S = (! \oplus \dotsb \oplus !)\), due to relation \textbf{(18)}.
\item \Define{\(c \circ\)}\\
A dual argument can be made for \((c \oplus 1^{n-1}) \circ S\) using relations \textbf{(15)},
\textbf{(11)} and \textbf{(16)}.
\item \Define{\(\circ 0\)}\\
For \(S(0 \oplus 1^{m-1})\), relations \textbf{(8)} and \textbf{(16)} eradicate the first \(\Delta\)
cluster and all the multiplications incident to it, leaving behind \(n\) zeros. Relation
\textbf{(1)} erases each of these zeros along with one addition per addition cluster, leaving a
diagram that is in standard form.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[zero] (top) {};
\node[delta] (apex) at (0,-0.866) {};
\node[delta] (dub) at (-0.5,-1.732) {};
\node (11a) at (-1,-2.382) {\(\scriptstyle{11}\)};
\node (21a) at (0,-2.382) {\(\scriptstyle{21}\)};
\node (n1a) at (0.5,-1.516) {\(\scriptstyle{n1}\)};
\draw[dotted] (apex.left out) -- (dub.io);
\draw (top) -- (apex) (apex.right out) -- (n1a)
(dub.left out) -- (11a) (dub.right out) -- (21a);
\node (eq) at (1,-1.3) {\(=\)};
\node[zero] (z1) at (1.5,-0.866) {};
\node[zero] (z2) at (2,-0.866) {};
\node[zero] (z3) at (3,-0.866) {};
\node (dots) at (2.5,-1.3) {\(\cdots\)};
\node (11b) at (1.5,-1.732) {\(\scriptstyle{11}\)};
\node (21b) at (2,-1.732) {\(\scriptstyle{21}\)};
\node (n1b) at (3,-1.732) {\(\scriptstyle{n1}\)};
\draw (z1) -- (11b) (z2) -- (21b) (z3) -- (n1b);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[thick]
\node[zero] (zip) {};
\node[multiply] (c) [below of=zip] {\(\scriptstyle{a_{i1}}\)};
\node (out) [below of=c] {};
\node (eq) [right of=c] {\(=\)};
\node[zero] (nada) [above right of=eq] {};
\node (bot) [below of=nada] {};
\draw (zip) -- (c) -- (out) (nada) -- (bot);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[thick]
\node[zero] (zip) at (-0.5,0.65) {};
\node[plus] (topa) {};
\node[plus] (mida) at (0.5,-0.866) {};
\node[plus] (bota) at (1,-1.732) {};
\node (outa) at (1,-2.382) {};
\node (i2a) at (0.5,0.65) {\(\scriptstyle{i2}\)};
\node (i3a) at (1,-0.216) {\(\scriptstyle{i3}\)};
\node (ima) at (1.5,-1.082) {\(\scriptstyle{im}\)};
\draw[dotted] (mida.io) -- (bota.left in);
\draw (bota) -- (outa) (bota.right in) -- (ima)
(mida.right in) -- (i3a) (mida.left in) -- (topa.io)
(topa.left in) -- (zip) (topa.right in) -- (i2a);
\node (eq) at (2,-0.95) {\(=\)};
\node[plus] (top) at (2.75,-0.866) {};
\node[plus] (bot) at (3.25,-1.732) {};
\node (out) at (3.25,-2.382) {};
\node (i2) at (2.25,-0.216) {\(\scriptstyle{i2}\)};
\node (i3) at (3.25,-0.216) {\(\scriptstyle{i3}\)};
\node (im) at (3.75,-1.082) {\(\scriptstyle{im}\)};
\draw[dotted] (top.io) -- (bot.left in);
\draw (bot.right in) -- (im) (top.right in) -- (i3)
(top.left in) -- (i2) (bot) -- (out);
\end{tikzpicture}
}
\end{center}
When \(S = (! \oplus \dotsb \oplus !)\), the zero annihilates one of the deletions via relation
\textbf{(10)}.
\item \Define{\(! \circ\)}\\
A dual argument erases the indicated output for the composition \((! \oplus 1^{n-1}) \circ S\) using
relations \textbf{(9)}, \textbf{(18)}, and \textbf{(4)}. Again, relation \textbf{(10)} annihilates
the deletion and one of the zeros if \(S = (0 \oplus \dotsb \oplus 0)\).
\item \Define{\(B \circ\)}\\
Since this category of string diagrams is symmetric monoidal, an appended braiding will naturally
commute with the addition cluster morphisms. The principle that only the topology matters means the
composition \((B \oplus 1^{n-2}) \circ S\) is in standard form. Braiding will similarly commute with
deletion morphisms.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick,node distance=0.5cm]
\node [coordinate] (fstart) {};
\node [coordinate] (ftop) [below of=fstart] {};
\node (center) [below right of=ftop] {};
\node [coordinate] (fout) [below right of=center] {};
\node [bang] (fend) [below of=fout] {};
\node [coordinate] (gtop) [above right of=center] {};
\node [coordinate] (gstart) [above of=gtop] {};
\node [coordinate] (gout) [below left of=center] {};
\node [bang] (gend) [below of=gout] {};
\draw [rounded corners=0.25cm] (fstart) -- (ftop) -- (center) --
(fout) -- (fend) (gstart) -- (gtop) -- (gout) -- (gend);
\node (eq1) [below right of=gtop, shift={(0.5,0)}] {\(=\)};
\node [bang] (ftop1) [above right of=eq1, shift={(0.5,0)}] {};
\node [coordinate] (fstart1) [above of=ftop1] {};
\node (center1) [below right of=ftop1] {};
\node [coordinate] (gtop1) [above right of=center1] {};
\node [coordinate] (gstart1) [above of=gtop1] {};
\node [coordinate] (gout1) [below left of=center1] {};
\node [bang] (gend1) [below of=gout1] {};
\draw [rounded corners=0.25cm] (fstart1) -- (ftop1)
(gstart1) -- (gtop1) -- (gout1) -- (gend1);
\node (eq2) [below right of=gtop1, shift={(0.35,0)}] {\(=\)};
\node [bang] (big) [right of=eq2, shift={(0.2,-0.5)}] {};
\node [bang] (bang) [right of=big] {};
\draw (big) -- +(0,1) (bang) -- +(0,1);
\end{tikzpicture}
}
\end{center}
\item \Define{\(\circ B\)}\\
Composing with \(B\) on the top, braiding commutes with duplication, multiplication and zero, so \(S
\circ (B \oplus 1^{m-2})\) almost trivially comes into standard form.
\end{itemize}
\end{proof}
An interesting exercise is to use these relations to derive a relation that expresses the braiding
in terms of other basic morphisms. One example of such a relation appeared in
Section~\ref{sigflow}. Here is another:
\begin{center}
\scalebox{0.75}{
\begin{tikzpicture}[-, thick]
\node (Lin) {};
\node [delta] (Ldub) [below of=Lin] {};
\node [coordinate] (ULang) [below of=Ldub, shift={(-0.5,0)}] {};
\node [multiply] (Lmin) [below of=ULang, shift={(0,0.65)}] {\(\scriptstyle{-1}\)};
\node [coordinate] (BLang) [below of=Lmin, shift={(0,0.35)}] {};
\node [plus] (Lsum) [below of=BLang, shift={(0.5,0)}] {};
\node [plus] (Usum) [below of=Ldub, shift={(0.9,0)}] {};
\node [delta] (Bdub) [below of=Usum] {};
\node [delta] (Rdub) [above of=Usum, shift={(0.9,0)}] {};
\node (Rin) [above of=Rdub] {};
\node [coordinate] (URang) [below of=Rdub, shift={(0.5,0)}] {};
\node [multiply] (Rmin) [below of=URang, shift={(0,0.65)}] {\(\scriptstyle{-1}\)};
\node [coordinate] (BRang) [below of=Rmin, shift={(0,0.35)}] {};
\node [plus] (Rsum) [below of=BRang, shift={(-0.5,0)}] {};
\node (Lout) [below of=Lsum] {};
\node (Rout) [below of=Rsum] {};
\draw (Lin) -- (Ldub) (Ldub.right out) -- (Usum.left in) (Usum) --
(Bdub) (Bdub.left out) -- (Lsum.right in) (Lsum) -- (Lout)
(Rin) -- (Rdub) (Rdub.left out) -- (Usum.right in) (Bdub.right out)
-- (Rsum.left in) (Rsum) -- (Rout);
\draw (Ldub.left out) .. controls +(240:0.5) and +(90:0.3) .. (Lmin.90)
(Lmin.io) .. controls +(270:0.3) and +(120:0.5) .. (Lsum.left in)
(Rdub.right out) .. controls +(300:0.5) and +(90:0.3) .. (Rmin.90)
(Rmin.270) .. controls +(270:0.3) and +(60:0.5) .. (Rsum.right in);
\node (eq) [left of=Lmin, shift={(-0.3,0)}] {\(=\)};
\node (center) [left of=eq, shift={(-0.2,0)}] {};
\node [coordinate] (ftop) [above left of=center, shift={(0.35,-0.35)}] {};
\node (fstart) [above of=ftop, shift={(0,-0.5)}] {};
\node [coordinate] (fout) [below right of=center, shift={(-0.35,0.35)}] {};
\node (fend) [below of=fout, shift={(0,0.5)}] {};
\node [coordinate] (gtop) [above right of=center, shift={(-0.35,-0.35)}] {};
\node (gstart) [above of=gtop, shift={(0,-0.5)}] {};
\node [coordinate] (gout) [below left of=center, shift={(0.35,0.35)}] {};
\node (gend) [below of=gout, shift={(0,0.5)}] {};
\draw [rounded corners=7pt] (fstart) -- (ftop) -- (center) --
(fout) -- (fend) (gstart) -- (gtop) -- (gout) -- (gend);
\end{tikzpicture}
}
\end{center}
With a few more relations, \(\mathrm{FinVect}_k\) can be presented as merely a monoidal category. Lafont
\cite{Lafont} did this in the special case where \(k\) is the field with two elements.
\section{A presentation of \(\mathrm{FinRel}_k\)}
\label{finrel}
Now we give a presentation for the symmetric monoidal category \(\mathrm{FinRel}_k\). As we did in the
previous section for \(\mathrm{FinVect}_k\), we work in a strict version of the symmetric monoidal category
\(\mathrm{FinRel}_k\).
\begin{lemma} \label{gensrk}
For any field \(k\), the object \(k\) together with the morphisms:
\begin{itemize}
\item addition \(+ \maps k \oplus k \nrightarrow k\)
\item zero \(0 \maps \{0\} \nrightarrow k\)
\item duplication \(\Delta \maps k \nrightarrow k \oplus k\)
\item deletion \(! \maps k \nrightarrow \{0\}\)
\item multiplication \(c \maps k \nrightarrow k\) for any \(c \in k\)
\item cup \(\cup \maps k \oplus k \nrightarrow \{0\} \)
\item cap \(\cap \maps \{0\} \nrightarrow k \oplus k \)
\end{itemize}
generate \(\mathrm{FinRel}_k\), the category of finite-dimensional vector spaces over \(k\) and linear
relations, as a symmetric monoidal category.
\end{lemma}
\begin{proof}
A morphism of \(\mathrm{FinRel}_k\), \(R\maps k^m \nrightarrow k^n\) is a subspace of \(k^m \oplus k^n \cong
k^{m+n}\). It can be expressed as a system of \(k\)-linear equations in \(k^{m+n}\).
Lemma~\ref{gensvk} tells us any number of arbitrary \(k\)-linear combinations of the inputs may be
generated. Any \(k\)-linear equation of those inputs can be formed by setting such a \(k\)-linear
combination equal to zero. In particular, if caps are placed on each of the outputs to make them
inputs and all the \(k\)-linear combinations are set equal to zero, any \(k\)-linear system of
equations of the inputs and outputs can be formed. Expressed in terms of string diagrams,
\begin{center}
\begin{tikzpicture}[-, thick, node distance=0.708cm]
\node (mn) {\(k_{m+n}\)};
\node (dots) [right of=mn] {\(\dots\)};
\node (m1) [right of=dots] {\(k_{m+1}\)};
\node [coordinate] (inbend1) [above of=m1] {};
\node [coordinate] (inbend2) [above of=dots] {};
\node [coordinate] (inbend3) [above of=mn] {};
\node (blank) [left of=inbend3] {};
\node [coordinate] (midbend3) [above of=blank] {};
\node [coordinate] (midbend2) [above of=midbend3] {};
\node [coordinate] (midbend1) [above of=midbend2] {};
\node [coordinate] (outbend3) [left of=blank] {};
\node [coordinate] (outbend2) [left of=outbend3] {};
\node [coordinate] (outbend1) [left of=outbend2] {};
\node (dota) [right of=mn, shift={(0.16,0)}] {};
\node [coordinate] (inbenda2) [above of=dota] {};
\node [coordinate] (midbenda2) [above of=midbend3, shift={(0,0.16)}] {};
\node [coordinate] (outbenda2) [left of=outbend3, shift={(-0.16,0)}] {};
\node (dotd) [right of=mn, shift={(-0.16,0)}] {};
\node [coordinate] (inbendd2) [above of=dotd] {};
\node [coordinate] (midbendd2) [above of=midbend3, shift={(0,-0.16)}] {};
\node [coordinate] (outbendd2) [left of=outbend3, shift={(0.16,0)}] {};
\draw[loosely dotted,out=-45,in=-135,relative]
(dota) -- (inbenda2) to (midbenda2) to (outbenda2)
(dotd) -- (inbendd2) to (midbendd2) to (outbendd2)
(dots) -- (inbend2) to (midbend2) to (outbend2);
\draw[out=-45,in=-135,relative]
(m1) -- (inbend1) to (midbend1) to (outbend1)
(mn) -- (inbend3) to (midbend3) to (outbend3);
\end{tikzpicture}
\hspace{1cm}
\begin{tikzpicture}[-, thick, node distance=1cm]
\node (f_i) at (0,1.5) {\(f_i\)};
\node [zero] (zero) at (0,0) {};
\draw (f_i) -- (zero);
\end{tikzpicture}
\end{center}
The left diagram turns the \(n\) outputs into inputs by placing caps on all of them. The morphism
zero gives the \(k\)-linear combination zero, so an arbitrary \(k\)-linear combination in
\(k^{m+n}\) is set equal to zero (\(f_i=0\)) via the cozero morphism. These elements can be
combined with Lemma~\ref{gensvk} to express any system of \(k\)-linear equations in \(k^{m+n}\).
\end{proof}
Putting these elements together, taking the \(\mathrm{FinVect}_k\) portion as a black box and drawing a single
string to denote zero or more copies of \(k\), the picture is fairly simple:
\begin{center}
\begin{tikzpicture}[thick]
\node [blackbox] (blackbox) {};
\node [zero] (zilch) at (0,-0.7) {};
\node (outs) at (0.7,-1) {};
\node [coordinate] (capR) at (0.7,0.5) {};
\node [coordinate] (capL) at (0.15,0.5) {};
\node (ins) at (-0.15,1) {};
\path (capL) edge[bend left=90] (capR);
\draw (ins) -- (-0.15,0) (blackbox) -- (zilch) (capL) -- (0.15,0)
(capR) -- (outs);
\end{tikzpicture}
\end{center}
To obtain a presentation of \(\mathrm{FinRel}_k\) as a symmetric monoidal category, we need to find enough
relations obeyed by the generating morphisms listed in Lemma~\ref{gensrk}. Relations \textbf{(1)--(18)} from
Theorem~\ref{presvk} still apply, but we need more.
For convenience, in the list below we draw the adjoint of any generating morphism by rotating it by
\(180^\circ\). It will follow from relations (19) and (20) that the cap is the adjoint of the cup,
so this convenient trick is consistent even in that case, where \emph{a priori} there might have
been an ambiguity.
\vskip 1em \noindent
\textbf{(19)--(20)} \(\cap\) and \(\cup\) obey the zigzag relations, and thus give a
\(\dagger\)-compact category:
\begin{center}
\begin{tikzpicture}[-, thick, node distance=1cm]
\node (zigtop) {};
\node [coordinate] (zigincup) [below of=zigtop] {};
\node [coordinate] (zigcupcap) [right of=zigincup] {};
\node [coordinate] (zigoutcap) [right of=zigcupcap] {};
\node (zigbot) [below of=zigoutcap] {};
\node (equal) [right of=zigoutcap] {\(=\)};
\node (mid) [right of=equal] {};
\node (vtop) [above of=mid] {};
\node (vbot) [below of=mid] {};
\node (equals) [right of=mid] {\(=\)};
\node [coordinate] (zagoutcap) [right of=equals] {};
\node (zagbot) [below of=zagoutcap] {};
\node [coordinate] (zagcupcap) [right of=zagoutcap] {};
\node [coordinate] (zagincup) [right of=zagcupcap] {};
\node (zagtop) [above of=zagincup] {};
\path
(zigincup) edge (zigtop) edge [bend right=90] (zigcupcap)
(zigoutcap) edge (zigbot) edge [bend right=90] (zigcupcap)
(vtop) edge (vbot)
(zagincup) edge (zagtop) edge [bend left=90] (zagcupcap)
(zagoutcap) edge (zagbot) edge [bend left=90] (zagcupcap);
\end{tikzpicture}
\end{center}
\vskip 1em \noindent
\textbf{(21)--(22)} \( (k, +, 0, +^\dagger, 0^\dagger)\) is a Frobenius monoid:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [plus] (sum1) at (0.5,-0.216) {};
\node [coplus] (cosum1) at (1,0.216) {};
\node [coordinate] (sum1corner) at (0,0.434) {};
\node [coordinate] (cosum1corner) at (1.5,-0.434) {};
\node [coordinate] (sum1out) at (0.5,-0.975) {};
\node [coordinate] (cosum1in) at (1,0.975) {};
\node [coordinate] (1cornerin) at (0,0.975) {};
\node [coordinate] (1cornerout) at (1.5,-0.975) {};
\draw[rounded corners] (1cornerin) -- (sum1corner) -- (sum1.left in)
(1cornerout) -- (cosum1corner) -- (cosum1.right out);
\draw (sum1.right in) -- (cosum1.left out)
(sum1.io) -- (sum1out)
(cosum1.io) -- (cosum1in);
\node (eq1) at (2,0) {\(=\)};
\node [plus] (sum2) at (3,0.325) {};
\node [coplus] (cosum2) at (3,-0.325) {};
\node [coordinate] (sum2inleft) at (2.5,0.975) {};
\node [coordinate] (sum2inright) at (3.5,0.975) {};
\node [coordinate] (cosum2outleft) at (2.5,-0.975) {};
\node [coordinate] (cosum2outright) at (3.5,-0.975) {};
\draw (sum2inleft) .. controls +(270:0.3) and +(120:0.15) .. (sum2.left in)
(sum2inright) .. controls +(270:0.3) and +(60:0.15) .. (sum2.right in)
(cosum2outleft) .. controls +(90:0.3) and +(240:0.15) .. (cosum2.left out)
(cosum2outright) .. controls +(90:0.3) and +(300:0.15) .. (cosum2.right out)
(sum2.io) -- (cosum2.io);
\node (eq2) at (4,0) {\(=\)};
\node [plus] (sum3) at (5.5,-0.216) {};
\node [coplus] (cosum3) at (5,0.216) {};
\node [coordinate] (sum3corner) at (6,0.434) {};
\node [coordinate] (cosum3corner) at (4.5,-0.434) {};
\node [coordinate] (sum3out) at (5.5,-0.975) {};
\node [coordinate] (cosum3in) at (5,0.975) {};
\node [coordinate] (3cornerin) at (6,0.975) {};
\node [coordinate] (3cornerout) at (4.5,-0.975) {};
\draw[rounded corners] (3cornerin) -- (sum3corner) -- (sum3.right in)
(3cornerout) -- (cosum3corner) -- (cosum3.left out);
\draw (sum3.left in) -- (cosum3.right out)
(sum3.io) -- (sum3out)
(cosum3.io) -- (cosum3in);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(23)--(24)} \( (k, \Delta^\dagger, !^\dagger, \Delta, !) \) is a Frobenius monoid:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [codelta] (sum1) at (0.5,-0.216) {};
\node [delta] (cosum1) at (1,0.216) {};
\node [coordinate] (sum1corner) at (0,0.434) {};
\node [coordinate] (cosum1corner) at (1.5,-0.434) {};
\node [coordinate] (sum1out) at (0.5,-0.975) {};
\node [coordinate] (cosum1in) at (1,0.975) {};
\node [coordinate] (1cornerin) at (0,0.975) {};
\node [coordinate] (1cornerout) at (1.5,-0.975) {};
\draw[rounded corners] (1cornerin) -- (sum1corner) -- (sum1.left in)
(1cornerout) -- (cosum1corner) -- (cosum1.right out);
\draw (sum1.right in) -- (cosum1.left out)
(sum1.io) -- (sum1out)
(cosum1.io) -- (cosum1in);
\node (eq1) at (2,0) {\(=\)};
\node [codelta] (sum2) at (3,0.325) {};
\node [delta] (cosum2) at (3,-0.325) {};
\node [coordinate] (sum2inleft) at (2.5,0.975) {};
\node [coordinate] (sum2inright) at (3.5,0.975) {};
\node [coordinate] (cosum2outleft) at (2.5,-0.975) {};
\node [coordinate] (cosum2outright) at (3.5,-0.975) {};
\draw (sum2inleft) .. controls +(270:0.3) and +(120:0.15) .. (sum2.left in)
(sum2inright) .. controls +(270:0.3) and +(60:0.15) .. (sum2.right in)
(cosum2outleft) .. controls +(90:0.3) and +(240:0.15) .. (cosum2.left out)
(cosum2outright) .. controls +(90:0.3) and +(300:0.15) .. (cosum2.right out)
(sum2.io) -- (cosum2.io);
\node (eq2) at (4,0) {\(=\)};
\node [codelta] (sum3) at (5.5,-0.216) {};
\node [delta] (cosum3) at (5,0.216) {};
\node [coordinate] (sum3corner) at (6,0.434) {};
\node [coordinate] (cosum3corner) at (4.5,-0.434) {};
\node [coordinate] (sum3out) at (5.5,-0.975) {};
\node [coordinate] (cosum3in) at (5,0.975) {};
\node [coordinate] (3cornerin) at (6,0.975) {};
\node [coordinate] (3cornerout) at (4.5,-0.975) {};
\draw[rounded corners] (3cornerin) -- (sum3corner) -- (sum3.right in)
(3cornerout) -- (cosum3corner) -- (cosum3.left out);
\draw (sum3.left in) -- (cosum3.right out)
(sum3.io) -- (sum3out)
(cosum3.io) -- (cosum3in);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(25)--(26)} The Frobenius monoid \( (k, +, 0, +^\dagger, 0^\dagger)\) is extra-special:
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (sum) at (0.4,-0.5) {};
\node [coplus] (cosum) at (0.4,0.5) {};
\node [coordinate] (in) at (0.4,1) {};
\node [coordinate] (out) at (0.4,-1) {};
\node (eq) at (1.3,0) {\(=\)};
\node [coordinate] (top) at (2,1) {};
\node [coordinate] (bottom) at (2,-1) {};
\path (sum.left in) edge[bend left=30] (cosum.left out)
(sum.right in) edge[bend right=30] (cosum.right out);
\draw (top) -- (bottom)
(sum.io) -- (out)
(cosum.io) -- (in);
\end{tikzpicture}
\qquad
\qquad
\qquad
\begin{tikzpicture}[thick]
\node [zero] (Bins) at (0,-0.35) {};
\node [zero] (Tins) at (0,0.35) {};
\node (eq) at (0.7,0) {\(=\)};
\node [hole] at (0,-0.865) {};
\draw (Tins) -- (Bins);
\end{tikzpicture}
\end{center}
\vskip 1em \noindent
\textbf{(27)--(28)} The Frobenius monoid \( (k, \Delta^\dagger, !^\dagger, \Delta, !)\) is
extra-special:
\begin{center}
\begin{tikzpicture}[thick]
\node [codelta] (sum) at (0.4,-0.5) {};
\node [delta] (cosum) at (0.4,0.5) {};
\node [coordinate] (in) at (0.4,1) {};
\node [coordinate] (out) at (0.4,-1) {};
\node (eq) at (1.3,0) {\(=\)};
\node [coordinate] (top) at (2,1) {};
\node [coordinate] (bottom) at (2,-1) {};
\path (sum.left in) edge[bend left=30] (cosum.left out)
(sum.right in) edge[bend right=30] (cosum.right out);
\draw (top) -- (bottom)
(sum.io) -- (out)
(cosum.io) -- (in);
\end{tikzpicture}
\qquad
\qquad
\qquad
\begin{tikzpicture}[thick]
\node [bang] (Bins) at (0,-0.35) {};
\node [bang] (Tins) at (0,0.35) {};
\node (eq) at (0.7,0) {\(=\)};
\node [hole] at (0,-0.865) {};
\draw (Tins) -- (Bins);
\end{tikzpicture}
\end{center}
\vskip 1em \noindent
\textbf{(29)} \(\cup\) with a factor of \(-1\) inserted can be expressed in terms of \(+\) and
\(0\):
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [multiply] (neg) at (0,0.1) {\(\scriptstyle{-1}\)};
\node [coordinate] (cupInLeft) at (0,1) {};
\node [coordinate] (Lcup) at (0,-0.5) {};
\node [coordinate] (Rcup) at (1,-0.5) {};
\node [coordinate] (cupInRight) at (1,1) {};
\node (eq) at (1.7,0.1) {\(=\)};
\node [coordinate] (SumLeftIn) at (2.25,1) {};
\node [coordinate] (SumRightIn) at (3.25,1) {};
\node [plus] (Sum) at (2.75,0) {};
\node [zero] (coZero) at (2.75,-0.65) {};
\draw (SumRightIn) .. controls +(270:0.5) and +(60:0.5) .. (Sum.right in)
(SumLeftIn) .. controls +(270:0.5) and +(120:0.5) .. (Sum.left in);
\draw (cupInLeft) -- (neg) -- (Lcup)
(Rcup) -- (cupInRight)
(Sum) -- (coZero);
\path (Lcup) edge[bend right=90] (Rcup);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(30)} \(\cap\) can be expressed in terms of \(\Delta\) and \(!\):
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node (eq) at (0.2,-0.1) {\(=\)};
\node [coordinate] (lcap) at (-1.5,0.5) {};
\node [coordinate] (rcap) at (-0.5,0.5) {};
\node [coordinate] (lcapbot) at (-1.5,-1) {};
\node [coordinate] (rcapbot) at (-0.5,-1) {};
\node [delta] (dub) at (1.25,0) {};
\node [bang] (boom) at (1.25,0.65) {};
\node [coordinate] (Leftout) at (0.75,-1) {};
\node [coordinate] (Rightout) at (1.75,-1) {};
\draw (dub.left out) .. controls +(240:0.5) and +(90:0.5) .. (Leftout)
(dub.right out) .. controls +(300:0.5) and +(90:0.5) .. (Rightout);
\draw (boom) -- (dub) (lcapbot) -- (lcap) (rcap) -- (rcapbot);
\path (lcap) edge[bend left=90] (rcap);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(31)} For any \(c \in k\) with \(c \ne 0\), scalar multiplication by \(c^{-1}\) is the
adjoint of scalar multiplication by \(c\):
\begin{center}
\begin{tikzpicture}[thick]
\node[upmultiply] (c) {\(c\)};
\node[coordinate] (in1) [above of=c] {};
\node[coordinate] (out1) [below of=c] {};
\draw (in1) -- (c) -- (out1);
\node (eq) [right of=c] {\(=\)};
\node[multiply] (mult) [right of=eq, shift={(0.5,0)}] {\(c^{-1}\!\!\)};
\node[coordinate] (in) [above of=mult] {};
\node[coordinate] (out) [below of=mult] {};
\draw (in) -- (mult) -- (out);
\end{tikzpicture}
\end{center}
\vskip 1em
Some curious identities can be derived from relations \textbf{(1)--(31)}, beyond those already
arising from \textbf{(1)--(18)}. For example:
\vskip 1em \noindent
\textbf{(D1)--(D2)} Deletion and zero can be expressed in terms of other generating morphisms:
\begin{center}
\begin{tikzpicture}[thick]
\node [bang] (bang) at (0.2,0) {};
\node (eq1) at (1,1.15) {\(=\)};
\node (rel1) at (1,0.8) {(27)};
\node [bang] (buck) at (2,0) {};
\node [codelta] (codub) at (2,0.65) {};
\node [delta] (dub) at (2,1.65) {};
\node (eq2) at (3,1.15) {\(=\)};
\node (rel2) at (3,0.8) {(30)\({}^\dagger\)};
\node [delta] (dupe) at (4,1) {};
\path
(codub.left in) edge[bend left=30] (dub.left out)
(codub.right in) edge[bend right=30] (dub.right out);
\draw
(bang) -- +(up:2.3) (dub.io) -- (2,2.3) (dupe.io) -- (4,2.3)
(codub.io) -- (buck)
(dupe.left out) .. controls +(240:1.2) and +(300:1.2) .. (dupe.right out);
\end{tikzpicture}
\hskip 2em
\begin{tikzpicture}[thick]
\node [zero] (Zero1) at (0,1) {};
\node (eq1) at (0.8,-0.15) {\(=\)};
\node (rel1) at (0.8,-0.5) {(28)};
\node [bang] (cobang) at (1.6,1) {};
\node [bang] (bang) at (1.6,0.2) {};
\node [zero] (Zero2) at (1.6,-0.5) {};
\node (eq2) at (2.4,-0.15) {\(=\)};
\node (rel2) at (2.4,-0.5) {(14)};
\node [multiply] (times) at (3.4,-0.15) {\(0\)};
\node [bang] (cobuck) at (3.4,1) {};
\node (eq3) at (4.4,-0.15) {\(=\)};
\node (rel3) at (4.4,-0.5) {(D1)\({}^\dagger\)};
\node [multiply] (oh) at (5.4,-0.5) {\(0\)};
\node [codelta] (cod) at (5.4,0.2) {};
\draw
(Zero1) -- +(down:2.3) (cobang) -- (bang) (Zero2) -- +(down:0.8)
(cobuck) -- (times) -- (3.4,-1.3) (cod.io) -- (oh) -- (5.4,-1.3)
(cod.left in) .. controls +(120:1.2) and +(60:1.2) .. (cod.right in);
\end{tikzpicture}
\end{center}
This does not diminish the role of deletion and zero. Indeed, regarding these generating morphisms
as superfluous buries some of the structure of \(\mathrm{FinRel}_k\).
\vskip 1em \noindent
\textbf{(D3)} Addition can be expressed in terms of coaddition and scalar multiplication by \(-1\),
and the cup:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [multiply] (neg1) at (0,0) {\(\scriptstyle{-1}\)};
\node [coplus] (cosum1) at (1,-0.21) {};
\draw
(neg1) -- (0,1) (cosum1.io) -- (1,1)
(cosum1.right out) .. controls +(300:0.5) and +(90:0.5) .. +(0.3,-1.3)
(neg1.io) .. controls +(270:1) and +(240:0.5) .. (cosum1.left out);
\node (eq2) at (2,-0.216) {\(=\)};
\node (rel2) at (2,-0.566) {(29)};
\node [plus] (sum3) at (3,-0.432) {};
\node [zero] (coz3) at (3,-1) {};
\node [coplus] (cosum3) at (3.5,0) {};
\draw
(sum3.left in) .. controls +(120:0.5) and +(270:0.5) .. +(-0.3,1.3)
(sum3) -- (coz3) (sum3.right in) -- (cosum3.left out)
(cosum3.right out) .. controls +(300:0.5) and +(90:0.5) .. +(0.3,-1.3)
(cosum3.io) -- +(0,0.76);
\node (eq3) at (4.5,-0.216) {\(=\)};
\node (rel3) at (4.5,-0.566) {(21)};
\node [plus] (sum4) at (5.5,0.159) {};
\node [coplus] (cosum4) at (5.5,-0.591) {};
\node [zero] (coz4) at (5.1,-1.35) {};
\draw
(sum4.io) -- (cosum4.io)
(cosum4.left out) .. controls +(240:0.2) and +(90:0.25) .. (coz4)
(sum4.left in) .. controls +(120:0.5) and +(270:0.5) .. +(-0.3,1)
(sum4.right in) .. controls +(60:0.5) and +(270:0.5) .. +(0.3,1)
(cosum4.right out) .. controls +(300:0.5) and +(90:0.5) .. +(0.3,-1);
\node (eq4) at (6.5,-0.216) {\(=\)};
\node (rel4) at (6.5,-0.566) {(1)\({}^\dagger\)};
\node [plus] (sum5) at (7.5,-0.4) {};
\draw
(sum5.io) -- +(0,-0.5)
(sum5.left in) .. controls +(120:0.5) and +(270:0.5) .. +(-0.3,1)
(sum5.right in) .. controls +(60:0.5) and +(270:0.5) .. +(0.3,1);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(D4)} Duplication can be expressed in terms of coduplication and the cap:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [coordinate] (neg1) at (0,0) {};
\node [codelta] (codub) at (1,0) {};
\draw
(neg1) -- (0,-0.91) (codub.io) -- (1,-0.91)
(codub.right in) .. controls +(60:0.5) and +(270:0.5) .. +(0.3,1)
(neg1) .. controls +(90:1) and +(120:0.5) .. (codub.left in);
\node (eq) at (2,0) {\(=\)};
\node [delta] (dub) at (3,0.2) {};
\draw
(dub.io) -- +(0,0.69)
(dub.left out) .. controls +(240:0.5) and +(90:0.5) .. +(-0.3,-1)
(dub.right out) .. controls +(300:0.5) and +(90:0.5) .. +(0.3,-1);
\end{tikzpicture}
}
\end{center}
where the proof is similar to that of \textbf{(D3)}.
\vskip 1em \noindent
\textbf{(D5)--(D7)} We can reformulate the bimonoid relations (7)--(9) using adjoints:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick, node distance=0.7cm]
\node (in1z) {};
\node (in2z) [right of=in1z, shift={(0.2,0)}] {};
\node (in3z) [right of=in2z, shift={(0.45,0)}] {};
\node [codelta] (nabzip) [below right of=in2z, shift={(0.1,-0.3)}] {};
\node [plus] (add) [below left of=nabzip, shift={(0.05,-0.3)}] {};
\node (outz) [below of=add] {};
\node (equal) [below right of=nabzip, shift={(0.2,0)}] {\(=\)};
\node [plus] (addl) [right of=equal, shift={(0.2,0)}] {};
\node (cross) [above right of=addl, shift={(-0.1,0)}] {};
\node [delta] (delta) [above left of=cross, shift={(0.1,0)}] {};
\node (in1u) [above of=delta] {};
\node [plus] (addr) [below right of=cross, shift={(-0.1,0)}] {};
\node [codelta] (nablunzip) [below left of=addr, shift={(0.1,-0.3)}] {};
\node (outu) [below of=nablunzip] {};
\node (in2u) [right of=in1u, shift={(0.4,0)}] {};
\node (in3u) [right of=in2u, shift={(0.1,0)}] {};
\draw (in1z) -- (add.left in) (add) -- (outz) (in2z) --
(nabzip.left in) (in3z) -- (nabzip.right in) (nabzip.io) -- (add.right in);
\path
(delta.left out) edge [bend right=30] (addl.left in);
\draw (in1u) -- (delta) (delta.right out) -- (cross) -- (addr.left in);
\draw (in2u) -- (addl.right in) (in3u) -- (addr.right in);
\draw (addl.io) -- (nablunzip.left in) (addr.io) -- (nablunzip.right in)
(nablunzip) -- (outu);
\end{tikzpicture}
}
\end{center}
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [zero] (coz1) at (0,0) {};
\node [delta] (dub) at (0.4,0.75) {};
\draw
(coz1) .. controls +(90:0.25) and +(240:0.2) .. (dub.left out)
(dub.right out) .. controls +(300:0.5) and +(90:0.5) .. +(0.3,-1)
(dub.io) -- +(0,0.5);
\node (eq) at (1.4,0.552) {\(=\)};
\node [zero] (coz2) at (2,0.867) {};
\node [zero] (zero) at (2,0.237) {};
\draw (coz2) -- +(0,0.6) (zero) -- +(0,-0.6);
\end{tikzpicture}
\hskip 3em
\begin{tikzpicture}[thick]
\node [bang] (cobang1) at (0,0) {};
\node [plus] (add) at (0.4,-0.75) {};
\draw
(cobang1) .. controls +(270:0.25) and +(120:0.2) .. (add.left in)
(add.right in) .. controls +(60:0.5) and +(270:0.5) .. +(0.3,1)
(add.io) -- +(0,-0.5);
\node (eq) at (1.4,-0.552) {\(=\)};
\node [bang] (cobang2) at (2,-0.867) {};
\node [bang] (bang) at (2,-0.237) {};
\draw (bang) -- +(0,0.6) (cobang2) -- +(0,-0.6);
\end{tikzpicture}
}
\end{center}
\vskip 1em \noindent
\textbf{(D8)--(D9)} When \(c \ne 1\), we have:
\begin{center}
\begin{tikzpicture}[thick]
\node [coplus] (cosum) at (0,1.8) {};
\node [multiply] (times) at (0.38,0.95) {\(c\)};
\node [plus] (sum) at (0,0) {};
\draw
(cosum.io) -- +(0,0.3) (sum.io) -- +(0,-0.3)
(cosum.left out) .. controls +(240:0.5) and +(120:0.5) .. (sum.left in)
(cosum.right out) .. controls +(300:0.15) and +(90:0.15) .. (times.90)
(times.io) .. controls +(270:0.15) and +(60:0.15) .. (sum.right in);
\node (eq) at (1.2,0.9) {\(=\)};
\node [bang] (bang) at (1.8,1.3) {};
\node [bang] (cobang) at (1.8,0.5) {};
\draw (bang) -- +(0,1.02) (cobang) -- +(0,-1.02);
\end{tikzpicture}
\hskip 3em
\begin{tikzpicture}[thick]
\node [delta] (dub) at (0,1.8) {};
\node [multiply] (times) at (0.38,0.95) {\(c\)};
\node [codelta] (codub) at (0,0) {};
\draw
(dub.io) -- +(0,0.3) (sum.io) -- +(0,-0.3)
(dub.left out) .. controls +(240:0.5) and +(120:0.5) .. (codub.left in)
(dub.right out) .. controls +(300:0.15) and +(90:0.15) .. (times.90)
(times.io) .. controls +(270:0.15) and +(60:0.15) .. (codub.right in);
\node (eq) at (1.2,0.9) {\(=\)};
\node [zero] (coz) at (1.8,1.3) {};
\node [zero] (zero) at (1.8,0.5) {};
\draw (coz) -- +(0,1.02) (zero) -- +(0,-1.02);
\end{tikzpicture}
\end{center}
\noindent
We leave the derivation of \textbf{(D5)--(D9)} as exercises for the reader.
Next we show that relations \textbf{(1)--(31)} are enough to give a presentation of \(\mathrm{FinRel}_k\) as a symmetric
monoidal category. As before, we do this by giving a standard form that any morphism can be written
in and use induction to show that an arbitrary diagram can be rewritten in its standard form using
the given relations.
\begin{thm} \label{presrk} The symmetric monoidal category \(\mathrm{FinRel}_k\) is presented by the object
\(k\), the morphisms given in Lemma~\ref{gensrk}, and relations {\bf (1)--(31)} as listed above.
\end{thm}
\begin{proof}
We prove this theorem by using the relations \textbf{(1)--(31)} to put any string diagram built from
the generating morphisms and braiding into a standard form, so that any two string diagrams
corresponding to the same morphism in \(\mathrm{FinRel}_k\) have the same standard form.
As before, we induct on the number of \Define{basic morphisms} involved in a string diagram, where
the basic morphisms are the generating morphisms together with the braiding. If we let \(R \maps
k^m \nrightarrow k^n\) be a morphism in \(\mathrm{FinRel}_k\), we can build a string diagram \(S\) for \(R\) as in
Lemma~\ref{gensrk}. Each output of \(S\) is capped, and, together with the inputs of \(S\), form
inputs for a \(\mathrm{FinVect}_k\) block, \(T\). For some \(r \leq m+n\), there are \(r\) outputs of
\(T\)--linear combinations of the \(m+n\) inputs--each set equal to zero via \((0^\dagger)^r\).
When \(T\) is in standard form for \(\mathrm{FinVect}_k\), we say \(S\) is in \Define{prestandard form}, and can
be depicted as follows:
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node [blackbox] (blackbox) {};
\node [zero] (zilch) at (0,-0.7) {};
\node (outs) at (0.7,-1) {};
\node [coordinate] (capR) at (0.7,0.5) {};
\node [coordinate] (capL) at (0.15,0.5) {};
\node (ins) at (-0.15,1) {};
\path (capL) edge[bend left=90] (capR);
\draw (ins) -- (-0.15,0) (blackbox) -- (zilch) (capL) -- (0.15,0)
(capR) -- (outs);
\end{tikzpicture}
}
\end{center}
While the linear subspace of \(k^{m+n}\) defined by \(R\) is determined by a system of \(r\) linear
equations, the converse is not true, meaning there may be multiple prestandard string diagrams for a
single morphism \(R\). The second stage of this proof collapses all the prestandard forms into a
standard form using some basic linear algebra. The standard form will correspond to when the matrix
representation of \(T\) is written in row-reduced echelon form. For this stage it will suffice to
show all the elementary row operations correspond to relations that hold between diagrams. By
Theorem \ref{presvk}, an arbitrary \(\mathrm{FinVect}_k\) block can be rewritten in its standard form, so the
\(\mathrm{FinVect}_k\) blocks here need not be demonstrated in their standard form.
\\
When there is one basic morphism, there are eight cases to consider, one per basic morphism. In
each of these basic cases, the block of the diagram equivalent to a morphism in \(\mathrm{FinVect}_k\) is
denoted by a dashed rectangle. We first consider \(\cup\).
\vskip 1em \noindent
\textbf{(D10)}
\begin{center}
\scalebox{1}{
\begin{tikzpicture}[thick]
\node (Lin0) at (0,1) {};
\node [coordinate] (Lcup) at (0,0) {};
\node [coordinate] (Rcup) at (1,0) {};
\node (Rin0) at (1,1) {};
\node (eq1) at (1.75,0.25) {\(=\)};
\node (rel1) at (1.75,-0.15) {(13)};
\node (rel2) at (1.75,-0.65) {(11)};
\node (Lin1) at (2.75,1.7) {};
\node [multiply] (neg1a) at (2.75,0.9) {\(\scriptstyle{-1}\)};
\node [multiply] (neg1b) at (2.75,-0.1) {\(\scriptstyle{-1}\)};
\node (Rin1) at (3.75,1.7) {};
\node (eq2) at (4.25,0.25) {\(=\)};
\node (rel3) at (4.25,-0.15) {(29)};
\node (Lin4) at (5.35,1.7) {};
\node [multiply] (neg) at (5.35,0.7) {\(\scriptstyle{-1}\)};
\node (Rin4) at (6.35,1.7) {};
\node [coordinate] (Rcor4) at (6.35,0.15) {};
\node [plus] (S4) at (5.85,-0.5) {};
\node [zero] (coZ4) at (5.85,-1.15) {};
\draw[rounded corners]
(Rin4) -- (Rcor4) -- (S4.right in);
\draw (Lin0) -- (Lcup) (Rcup) -- (Rin0)
(Lin1) -- (neg1a) -- (neg1b)
(neg1b.io) .. controls +(270:0.5) and +(270:0.5) .. +(1,0) -- (Rin1)
(Lin4) -- (neg) (neg.io) -- (S4.left in) (S4) -- (coZ4);
\draw [color=red, dashed, thin] (4.75,1.2) rectangle (6.75,-0.9);
\path (Lcup) edge[bend right=90] (Rcup);
\end{tikzpicture}
}
\end{center}
Capping each of the inputs turns this into the standard form of \(\cap\). Aside from deletion, the
remaining generating morphisms can be formed by introducing a zigzag at each output and rewriting
the resulting cups as above. The standard forms for \(0\) and \(!\) have simpler expressions.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node (Lout0) at (-2.75,-0.5) {};
\node [coordinate] (Lcap0) at (-2.75,0.5) {};
\node [coordinate] (Rcap0) at (-1.75,0.5) {};
\node (Rout0) at (-1.75,-0.5) {};
\node (eq) at (-1,0.2) {\(=\)};
\node [coordinate] (Lin) at (0,1.4) {};
\node [multiply] (neg) at (0,0.7) {\(\scriptstyle{-1}\)};
\node [coordinate] (Rin) at (1,1.4) {};
\node [coordinate] (Rcor) at (1,0.15) {};
\node [plus] (S) at (0.5,-0.5) {};
\node [zero] (coZ) at (0.5,-1.15) {};
\node [coordinate] (Rcap) at (2.5,1.4) {};
\node [coordinate] (Lcap) at (1.5,1.4) {};
\node (Rout) at (2.5,-1.3) {};
\node (Lout) at (1.5,-1.3) {};
\draw (Lout0) -- (Lcap0) (Rcap0) -- (Rout0) (Lin) -- (neg) (S) -- (coZ)
(neg.io) -- (S.left in) (Rcap) -- (Rout) (Lcap) -- (Lout);
\path (Lcap0) edge[bend left=90] (Rcap0) (Lin) edge[bend left=90]
(Rcap) (Rin) edge[bend left=90] (Lcap);
\draw[rounded corners] (Rin) -- (Rcor) -- (S.right in);
\draw[color=red, dashed, thin] (-0.5,1.2) rectangle (1.25,-0.9);
\end{tikzpicture}
\hspace{0.69cm}
\begin{tikzpicture}[thick]
\node (in) at (-1.8,1.7) {};
\node [multiply] (times) at (-1.8,0.2) {\(\scriptstyle{c}\)};
\node (out) at (-1.8,-1.3) {};
\node (eq) at (-1,0.2) {\(=\)};
\node [coordinate] (Lin) at (0,1.7) {};
\node [multiply] (neg) at (0,0.7) {\(\scriptstyle{-c}\)};
\node [coordinate] (Rin) at (1,1.4) {};
\node [coordinate] (Rcor) at (1,0.15) {};
\node [plus] (S) at (0.5,-0.5) {};
\node [zero] (coZ) at (0.5,-1.15) {};
\node [coordinate] (Lcap) at (1.5,1.4) {};
\node (Lout) at (1.5,-1.3) {};
\draw (in) -- (times) -- (out) (Lin) -- (neg) (S) -- (coZ)
(neg.io) -- (S.left in) (Lcap) -- (Lout);
\path (Rin) edge[bend left=90] (Lcap);
\draw[rounded corners] (Rin) -- (Rcor) -- (S.right in);
\draw[color=red, dashed, thin] (-0.5,1.2) rectangle (1.25,-0.9);
\end{tikzpicture}
\hspace{0.69cm}
\begin{tikzpicture}[thick]
\node (in1) at (-2.3,0.85) {};
\node (in2) at (-1.3,0.85) {};
\node [plus] (S) at (-1.8,0.2) {};
\node (out) at (-1.8,-0.45) {};
\node (eq) at (-1,0.2) {\(=\)};
\node [coordinate] (Lin) at (-0.5,1.016) {};
\node (Linup) at (-0.5,1.7) {};
\node [coordinate] (Rin) at (0.5,1.016) {};
\node (Rinup) at (0.5,1.7) {};
\node [multiply] (neg) at (1,0.7) {\(\scriptstyle{-1}\)};
\node [coordinate] (Lcap) at (1,1.3) {};
\node [plus] (Stop) at (0,0.366) {};
\node [plus] (Sbot) at (0.5,-0.5) {};
\node [zero] (coZ) at (0.5,-1.15) {};
\node [coordinate] (Rcap) at (2,1.3) {};
\node (Lout) at (2,-1.3) {};
\draw[rounded corners] (Linup) -- (Lin) -- (Stop.left in)
(Rinup) -- (Rin) -- (Stop.right in);
\draw (in1) -- (S.left in) (S.right in) -- (in2) (S) -- (out)
(Stop.io) -- (Sbot.left in)
(Sbot) -- (coZ) (neg.io) -- (Sbot.right in) (neg) -- (Lcap)
(Rcap) -- (Lout);
\path (Lcap) edge[bend left=90] (Rcap);
\draw[color=red, dashed, thin] (-0.65,1.2) rectangle (1.65,-0.9);
\end{tikzpicture}
}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node [zero] (zero1) at (-0.2,1) {};
\node (eq) at (0.4,0) {\(=\)};
\node [zero] (coz) at (1.55,-0.8) {};
\node [hole] (placeholder) at (0,-1.9) {};
\draw[color=red, dashed, thin] (2,0.5) rectangle (1.1,-0.5);
\draw (zero1) -- +(down:2.2) (coz) -- (1.55,0.5)
.. controls +(90:0.75) and +(90:0.75) .. (2.5,0.5) -- (2.5,-1.2);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[thick]
\node (out1) at (-2.3,0) {};
\node (out2) at (-1.3,0) {};
\node [delta] (D) at (-1.8,0.65) {};
\node (in) at (-1.8,1.3) {};
\node (eq) at (-1,0.65) {\(=\)};
\node [coordinate] (in1) at (0,2.5) {};
\node [multiply] (neg) at (0,1.7) {\(\scriptstyle{-1}\)};
\node [delta] (D1) at (0,0.65) {};
\node [plus] (S1) at (0,-0.65) {};
\node [plus] (S2) at (1,-0.65) {};
\node [zero] (coZ1) at (0,-1.3) {};
\node [zero] (coZ2) at (1,-1.3) {};
\node (cross) at (0.5,0) {};
\node [coordinate] (Loutup) at (1,0.866) {};
\node [coordinate] (Routup) at (1.5,0) {};
\node [coordinate] (oLcap) at (1,2.3) {};
\node [coordinate] (iLcap) at (1.5,2.3) {};
\node [coordinate] (iRcap) at (2,2.3) {};
\node [coordinate] (oRcap) at (2.5,2.3) {};
\node (Lout) at (2,-1.3) {};
\node (Rout) at (2.5,-1.3) {};
\draw[rounded corners=10pt] (S1.right in) -- (Loutup) -- (oLcap)
(iLcap) -- (Routup) -- (S2.right in);
\draw (out1) -- (D.left out) (D.right out) -- (out2) (D) -- (in)
(in1) -- (neg) -- (D1) (D1.right out) -- (cross) -- (S2.left in)
(S2) -- (coZ2) (S1) -- (coZ1) (iRcap) -- (Lout) (oRcap) -- (Rout);
\path (oLcap) edge[bend left=90] (oRcap) (iLcap) edge[bend left=90]
(iRcap) (D1.left out) edge[bend right=30] (S1.left in);
\draw[color=red, dashed, thin] (-0.65,2.2) rectangle (1.75,-1);
\end{tikzpicture}
\hspace{0.7cm}
\begin{tikzpicture}[thick]
\node [bang] (del1) at (0,-1) {};
\node (eq) at (0.7,0) {\(=\)};
\node (rel1) at (0.7,-0.4) {(26)};
\node [bang] (del2) at (1.7,0.3) {};
\node [zero] (zero) at (1.7,-0.3) {};
\node [zero] (coz) at (1.7,-1) {};
\node [hole] (placeholder) at (0,-1.9) {};
\draw[color=red, dashed, thin] (1.25,-0.5) rectangle (2.15,0.5);
\draw (del1) -- +(up:2) (del2) -- +(up:0.7) (coz) -- (zero);
\end{tikzpicture}
}
\end{center}
Braiding is two copies of multiplication by 1 that have been braided together.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node (UpUpLeft) at (-3.7,2) {};
\node [coordinate] (UpLeft) at (-3.7,1.5) {};
\node (mid) at (-3.3,1.1) {};
\node [coordinate] (DownRight) at (-2.9,0.7) {};
\node (DownDownRight) at (-2.9,0.2) {};
\node [coordinate] (UpRight) at (-2.9,1.5) {};
\node (UpUpRight) at (-2.9,2) {};
\node [coordinate] (DownLeft) at (-3.7,0.7) {};
\node (DownDownLeft) at (-3.7,0.2) {};
\draw [rounded corners=2mm] (UpUpLeft) -- (UpLeft) -- (mid) --
(DownRight) -- (DownDownRight) (UpUpRight) -- (UpRight) -- (DownLeft) -- (DownDownLeft);
\node (eq) at (-2.2,1.1) {\(=\)};
\node [plus] (sum1) at (0,0) {};
\node [plus] (sum2) at (0.7,0) {};
\node [zero] (coz1) at (0,-0.65) {};
\node [zero] (coz2) at (0.7,-0.65) {};
\node [multiply] (neg1) at (-1,2.066) {\(\scriptstyle^{-1}\)};
\node [multiply] (neg2) at (0,2.066) {\(\scriptstyle{-1}\)};
\draw[color=red, dashed, thin] (-1.6,-0.3) rectangle (1.9,2.5);
\draw (sum2.left in) .. controls +(120:1) and +(270:0.2) .. (neg2.io);
\node [hole] (hole) at (0.35,0.39) {};
\draw (sum1.right in) .. controls +(60:1) and +(270:1) .. (1,2.5)
.. controls +(90:1) and +(90:1) .. (2.8,2.5) -- (2.8,-1)
(sum2.right in) .. controls +(60:1) and +(270:1) .. (1.7,2.5)
.. controls +(90:0.3) and +(90:0.3) .. (2.1,2.5) -- (2.1,-1)
(sum1.io) -- (coz1) (sum2.io) -- (coz2)
(sum1.left in) .. controls +(120:1) and +(270:0.2) .. (neg1.io)
(neg1) -- (-1,3.5) (neg2) -- (0,3.5);
\end{tikzpicture}
}
\end{center}
Assuming any string diagram with \(j\) basic morphisms can be written in prestandard form, we show
an arbitrary diagram with \(j+1\) basic morphisms can be written in prestandard form as well. Let
\(S\) be a string diagram on \(j\) basic morphisms, rewritten into prestandard form, with a
maximal \(\mathrm{FinVect}_k\) subdiagram \(T\). Several cases are considered: those putting a basic morphism
above \(S\), beside \(S\), and below \(S\).
\begin{itemize}[leftmargin=1em]
\item \Define{\(S \circ G\) for a basic morphism \(G \neq \cap\)}\\
If a diagram \(G\) is composed above \(S\), \(G\) can combine with \(T\) to make a larger \(\mathrm{FinVect}_k\)
subdiagram if \(G\) is \(c\), \(\Delta\), \(+\), \(B\), or \(0\), as these are morphisms in
\(\mathrm{FinVect}_k\). The generating morphisms \(\cap\), \(\cup\) and \(!\) are not on this list, though a
composition with \(\cup\) (resp. \(!\)) would be equivalent to tensoring by \(\cup\) (resp. \(!\)).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.1,0.1) rectangle (1.5,-0.8);
\node (eq) at (2.6,0) {\(=\)};
\node [sqnode] (G) at (0.7,0.8) {\(G\)};
\filldraw[fill=black,draw=black] (3.2,0.1) rectangle (4.8,-0.8);
\node [zero] (Z) at (0.7,-1.2) {};
\node [zero] (ero) at (4,-1.2) {};
\draw (0.2,0) -- (0.2,1.5) (0.7,1.5) -- (G) -- (0.7,0) (0.7,-0.7) -- (Z)
(ero) -- (4,-0.7) (3.5,0) -- (3.5,1.5) (4,0) -- (4,1.5);
\draw (4.5,0) .. controls +(90:0.6) and +(90:0.6) .. (5.2,0) -- (5.2,-1.7)
(1.2,0) .. controls +(90:0.6) and +(90:0.6) .. (1.9,0) -- (1.9,-1.7);
\end{tikzpicture}
}\\
\scalebox{0.80}{
for \begin{tikzpicture}[thick]
\node [sqnode] (G) at (0,0) {\(G\)};
\draw (0,0.7) -- (G) -- (0,-0.7);
\node (eq) at (0.7,0) {\(=\)};
\end{tikzpicture}
\begin{tikzpicture}[thick]
\node [multiply] (c) at (0,0) {\(\scriptstyle{c}\)};
\draw (0,0.55) -- (c) -- (0,-0.65);
\end{tikzpicture},
\begin{tikzpicture}[thick]
\node [delta] (dub) at (0,0) {};
\draw (0,0.6) -- (dub) (dub.left out) -- (-0.4,-0.5) (dub.right out) -- (0.4,-0.5);
\end{tikzpicture} ,
\begin{tikzpicture}[thick]
\node [plus] (sum) at (0,0) {};
\draw (0,-0.6) -- (sum) (sum.left in) -- (-0.4,0.5) (sum.right in) -- (0.4,0.5);
\end{tikzpicture},
\begin{tikzpicture}[thick]
\draw (0.45,-0.45) -- (-0.45,0.45);
\node [hole] (hole) at (0,0) {};
\draw (0.45,0.45) -- (-0.45,-0.45);
\end{tikzpicture} , or
\begin{tikzpicture}[thick]
\node [zero] (zero) at (0,0.4) {};
\draw (0,-0.4) -- (zero);
\end{tikzpicture} .
}
\end{center}
Putting these morphisms on top of \(S\) reduces to performing those compositions on \(T\). The
maximal \(\mathrm{FinVect}_k\) subdiagram now includes \(T\) and \(G\), with \(S\) unchanged outside the
\(\mathrm{FinVect}_k\) block.
\item \Define{\(B \circ S\)}\\
\(B\) commutes with caps because the category is symmetric monoidal, so capping the braiding is
equivalent to putting the braiding on top of \(T\). \(B\) is ``absorbed'' into \(T\), just as in
the \(S \circ G\) case.
\item \Define{\(S \oplus G\) for any basic morphism \(G\)}\\
If any two prestandard string diagrams \(S\) and \(S'\) are tensored together, the result combines
into one prestandard diagram. This is evident because the category of string diagrams is symmetric
monoidal, and the \(\mathrm{FinVect}_k\) blocks can be placed next to each other as the tensor of two \(\mathrm{FinVect}_k\)
blocks. These combine into a single \(\mathrm{FinVect}_k\) block, and absorbing all the braidings into this
block as above brings the diagram into prestandard form. Since each basic morphism can be written
as a prestandard diagram, the tensor \(S \oplus G\) is a special case of this.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.1,0.1) rectangle (0.8,-0.8);
\node (oplus) at (1.8,-0.4) {\(\bigoplus\)};
\filldraw[fill=black,draw=black] (2.4,0.1) rectangle (3.3,-0.8);
\node (eq) at (4.3,-0.4) {\(=\)};
\filldraw[fill=black,draw=black] (4.9,0.1) rectangle (6.4,-0.8);
\node [zero] (Z) at (0.35,-1.2) {};
\node [zero] (e) at (2.85,-1.2) {};
\node [zero] (ro) at (5.65,-1.2) {};
\draw (0.2,0) -- (0.2,1) (0.35,-0.7) -- (Z) (ro) -- (5.65,-0.7)
(e) -- (2.85,-0.7) (2.7,0) -- (2.7,1) (5.2,0) -- (5.2,1)
(5.8,0) -- (5.8,1);
\draw (3,0) .. controls +(90:0.6) and +(90:0.6) .. (3.7,0) -- (3.7,-1.7)
(0.5,0) .. controls +(90:0.6) and +(90:0.6) .. (1.2,0) -- (1.2,-1.7)
(6.1,0) .. controls +(90:0.9) and +(90:0.9) .. (7.2,0) -- (7.2,-1.7);
\node [hole] (ho) at (5.8,0.6) {};
\node [hole] (le) at (6.46,0.63) {};
\draw (5.5,-0.1) .. controls +(90:1.1) and +(90:1.1) .. (6.8,-0.1) -- (6.8,-1.7);
\end{tikzpicture}
}
\end{center}
\item \Define{\(c \circ S\) for \(c \neq 0\)}\\
Because the outputs of \(S\) are capped, putting any morphism on the bottom of \(S\) is equivalent
(via relations \textbf{(19)} and \textbf{(20)}) to putting its adjoint on top of \(T\). Putting \(c
\neq 0\) below \(S\) reduces to putting \(c^{-1}\) on top of \(T\) by relation \textbf{(31)}. The
case of \(c = 0\) will be considered below. The other cases of adjoints of generating morphisms
that need to be considered more carefully are the ones that put \(\Delta^\dagger\), \(+^\dagger\)
and \(\cap = \cup^\dagger\) on top of \(T\).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.45,0.45) rectangle (0.45,-0.45);
\node[multiply] (c) at (1.05,0) {\(c\)};
\node[zero] (coz1) at (0,-0.8) {};
\draw[out=-90,in=-90,relative,looseness=2]
(c.io) -- (1.05,-1.85) (c.90) -- (1.05,0.45) to (0.15,0.45)
(coz1) -- +(0,0.5) (-0.15,0.35) -- +(0,1.5);
\node (eq) at (1.8,0) {\(=\)};
\filldraw[fill=black,draw=black] (2.3,-0.05) rectangle (3.7,-0.95);
\node[zero] (coz2) at (3,-1.3) {};
\node[multiply] (cinv) at (3.4,0.8) {\(c^{-1}\!\!\)};
\draw[out=90,in=90,relative,looseness=2]
(2.6,-0.15) -- +(0,2) (cinv.io) -- +(0,-0.5) (coz2) -- +(0,0.5)
(cinv.90) to +(0.9,0) -- (4.3,-1.85);
\end{tikzpicture}
}
\end{center}
\item \Define{\(\Delta \circ S\)}\\
When putting \(\Delta^\dagger\) on top of \(T\), the idea is to make it ``trickle down.'' If there
is a nonzero multiplication incident to the \(\Delta\) cluster, \(\Delta^\dagger\) can slide through
the \(\Delta\)s using relation \textbf{(23)} to the first nonzero multiplication, switching to
relation \textbf{(24)}. When it encounters this \(c\), relation \textbf{(31)} turns \(c\) into
\((c^{-1})^\dagger\), relation \textbf{(17)}\({}^\dagger\) allows \(\Delta^\dagger\) to pass through
\((c^{-1})^\dagger\). Both copies of \((c^{-1})^\dagger\) can return to being \(c\) by another
application of relation \textbf{(31)}, and the \(\Delta^\dagger\) moves on to the next layer.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[codelta] (cod1) at (0,0) {};
\node[delta] (dub1) at (0,-0.734) {};
\node[multiply] (zip1) at (0.5,-1.5) {\(0\)};
\draw (cod1.left in) -- +(120:0.5) (cod1.right in) -- + (60:0.5) (cod1.io) -- (dub1.io)
(dub1.right out) -- (zip1.90) (zip1.io) -- +(0,-0.3) (dub1.left out) -- +(240:0.5);
\node (eq) at (1.3,-0.9) {\(=\)};
\node (rel) at (1.3,-1.2) {(23)};
\node[delta] (dub2) at (2.6,-0.468) {};
\node[codelta] (cod2) at (2.1,-0.9) {};
\node[multiply] (zip2) at (3.1,-1.234) {\(0\)};
\draw (dub2.io) -- +(90:0.5) (dub2.right out) -- (zip2.90) (zip2.io) -- +(0,-0.3)
(dub2.left out) -- (cod2.right in) (cod2.left in) -- +(120:0.5) (cod2.io) -- +(0,-0.5);
\end{tikzpicture}
\qquad
\begin{tikzpicture}[thick]
\node[codelta] (cod1) at (0,0) {};
\node[delta] (dub1) at (0,-0.734) {};
\node[multiply] (zip1) at (0.5,-1.5) {\(c\)};
\draw (cod1.left in) -- +(120:0.5) (cod1.right in) -- + (60:0.5) (cod1.io) -- (dub1.io)
(dub1.right out) -- (zip1.90) (zip1.io) -- +(0,-0.3) (dub1.left out) -- +(240:0.5);
\node (eq) at (1.3,-0.9) {\(=\)};
\node (rel) at (1.3,-1.2) {(24)};
\node[delta] (dub2) at (2.2,-0.268) {};
\node[codelta] (cod2) at (2.7,-0.7) {};
\node[multiply] (zip2) at (2.7,-1.5) {\(c\)};
\draw (dub2.io) -- +(90:0.5) (dub2.left out) -- +(240:0.5) (cod2.io) -- (zip2.90)
(zip2.io) -- +(0,-0.3) (dub2.right out) -- (cod2.left in) (cod2.right in) -- +(60:0.5);
\end{tikzpicture}
}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[multiply] (c) at (0,0) {\(c\)};
\node[codelta] (cod1) at (0,0.75) {};
\node (eq1) at (0.8,0.55) {\(=\)};
\node (rel1) at (0.8,0.25) {(31)\({}^\dagger\)};
\draw (c.90) -- (cod1.io) (cod1.left in) -- +(120:0.5) (cod1.right in) -- +(60:0.5)
(c.io) -- +(0,-0.3);
\node[upmultiply] (cinv2) at (1.8,-0.1) {\(\!c^{-1}\!\)};
\node[codelta] (cod2) at (1.8,1) {};
\node (eq2) at (2.9,0.55) {\(=\)};
\node (rel2) at (2.9,0.25) {(17)\({}^\dagger\)};
\draw (cinv2.io) -- (cod2.io) (cod2.left in) -- +(120:0.5) (cod2.right in) -- +(60:0.5)
(cinv2.270) -- +(0,-0.4);
\node[upmultiply] (cinv3a) at (4,1.1) {\(\!c^{-1}\!\)};
\node[upmultiply] (cinv3b) at (5.2,1.1) {\(\!c^{-1}\!\)};
\node[codelta] (cod3) at (4.6,-0.15) {};
\node (eq3) at (6.3,0.55) {\(=\)};
\node (rel3) at (6.3,0.25) {(31)\({}^\dagger\)};
\draw (cinv3a.io) -- +(0,0.4) (cinv3b.io) -- +(0,0.4) (cod3.io) -- +(0,-0.5)
(cod3.left in) .. controls +(120:0.5) and +(270:0.3) .. (cinv3a.270)
(cod3.right in) .. controls +(60:0.5) and +(270:0.3) .. (cinv3b.270);
\node[multiply] (c4a) at (7,1.1) {\(c\)};
\node[multiply] (c4b) at (8,1.1) {\(c\)};
\node[codelta] (cod4) at (7.5,0.05) {};
\draw (c4a.90) -- +(0,0.4) (c4b.90) -- +(0,0.4) (cod4.io) -- +(0,-0.5)
(cod4.left in) -- (c4a.270) (cod4.right in) -- (c4b.270);
\end{tikzpicture}
}
\end{center}
When the codelta gets to a \(+\) cluster, derived relation \textbf{(D5)} has a net effect of
bringing it to the bottom of the subdiagram, as the other morphisms involved all belong to
\(\mathrm{FinVect}_k\). This allows the process to be repeated on the next addition until \(\Delta^\dagger\)
reaches the bottom of the \(+\) cluster. Once there, codelta interacts with the cozero layer below
\(T\); relation \textbf{(8)}\({}^\dagger\) reduces it to a pair of cozeros.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node (in1z) {};
\node (in2z) [right of=in1z, shift={(0.2,0)}] {};
\node (in3z) [right of=in2z, shift={(0.45,0)}] {};
\node [codelta] (nabzip) [below right of=in2z, shift={(0.1,-0.3)}] {};
\node [plus] (add) [below left of=nabzip, shift={(0.05,-0.3)}] {};
\node (outz) [below of=add] {};
\node (equal) [below right of=nabzip, shift={(0.2,0)}] {\(=\)};
\node (rel) [below of=equal, shift={(0,0.4)}] {(D5)};
\node [plus] (addl) [right of=equal, shift={(0.2,0)}] {};
\node (cross) [above right of=addl, shift={(-0.1,0)}] {};
\node [delta] (delta) [above left of=cross, shift={(0.1,0)}] {};
\node (in1u) [above of=delta] {};
\node [plus] (addr) [below right of=cross, shift={(-0.1,0)}] {};
\node [codelta] (nablunzip) [below left of=addr, shift={(0.1,-0.3)}] {};
\node (outu) [below of=nablunzip] {};
\node (in2u) [right of=in1u, shift={(0.4,0)}] {};
\node (in3u) [right of=in2u, shift={(0.1,0)}] {};
\draw (in1z) -- (add.left in) (add) -- (outz) (in2z) --
(nabzip.left in) (in3z) -- (nabzip.right in) (nabzip.io) -- (add.right in);
\path
(delta.left out) edge [bend right=30] (addl.left in);
\draw (in1u) -- (delta) (delta.right out) -- (cross) -- (addr.left in);
\draw (in2u) -- (addl.right in) (in3u) -- (addr.right in);
\draw (addl.io) -- (nablunzip.left in) (addr.io) -- (nablunzip.right in)
(nablunzip) -- (outu);
\end{tikzpicture}
\qquad
\begin{tikzpicture}[thick]
\node [codelta] (nabla) at (0,-0.65) {};
\node [zero] (Z1) at (0,-1.3) {};
\node [coordinate] (il1) at (-0.5,0) {};
\node [coordinate] (ir1) at (0.5,0) {};
\node (eq) at (0.9,-0.6) {\(=\)};
\node (rel) at (0.9,-0.9) {(8)\({}^\dagger\)};
\node [coordinate] (il2) at (1.5,0) {};
\node [coordinate] (ir2) at (2.2,0) {};
\node [zero] (ZL) [below of=il2] {};
\node [zero] (ZR) [below of=ir2] {};
\node [hole] (space) at (0,-2) {};
\draw (il1) -- (nabla.left in) (nabla) -- (Z1) (nabla.right in) -- (ir1)
(il2) -- (ZL) (ir2) -- (ZR);
\end{tikzpicture}
}
\end{center}
If all the multiplications incident to the \(\Delta\) cluster are by \(0\), rather than trickling
down, \(\Delta^\dagger\) composes with \(!\) (due to relation \textbf{(14)}), which gives \(\cup\)
by relation \textbf{(30)}\({}^\dagger\). By the zigzag identities, this cup becomes a cap that is
tensored with a subdiagram of \(S\) that is in prestandard form.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[codelta] (top2) at (1,0.866) {};
\node[delta] (delt1) at (1,-0.134) {};
\node[delta] (delt2) at (0.5,-1) {};
\node[multiply] (c1) at (0,-1.65) {\(0\)};
\node[multiply] (c2) at (1,-1.65) {\(0\)};
\node[multiply] (c3) at (2,-1.65) {\(0\)};
\node (11b) at (0,-2.516) {};
\node (21b) at (1,-2.516) {};
\node (n1b) at (2,-2.516) {};
\draw[dotted] (delt1.left out) -- (delt2.io);
\draw (delt2.left out) -- (c1) -- (11b) (delt2.right out) -- (c2) -- (21b)
(delt1.right out) -- (c3) -- (n1b) (top2.io) -- (delt1.io)
(top2.left in) -- +(120:0.4) (top2.right in) -- +(60:0.4);
\node (eq) at (2.7,-0.65) {\(=\)};
\node (rel1) at (2.7,-0.95) {(14)};
\node (rel2) at (2.7,-1.3) {(4)};
\node[codelta] (cod) at (3.8,0.5) {};
\node[bang] (bang) at (3.8,-0.2) {};
\node[zero] (ze) at (3.4,-1) {};
\node[zero] (ro) at (4.2,-1) {};
\node (dots) at (3.8,-1.5) {\(\cdots\)};
\draw (cod.io) -- (bang) (ze) -- +(0,-1) (ro) -- +(0,-1)
(cod.left in) -- +(120:0.4) (cod.right in) -- +(60:0.4);
\node (eq2) at (4.8,-0.65) {\(=\)};
\node (rel3) at (4.8,-0.95) {(30)\({}^\dagger\)};
\node[zero] (ze1) at (5.4,-1) {};
\node[zero] (ro1) at (6.2,-1) {};
\node (dots1) at (5.8,-1.5) {\(\cdots\)};
\draw[out=-90,in=-90,relative,looseness=2]
(5.4,1) -- ++(0,-1) to +(0.8,0) -- +(0,1)
(ze1) -- +(0,-1) (ro1) -- +(0,-1);
\end{tikzpicture}
}
\end{center}
\item \Define{\(+ \circ S\)}\\
There is a similar trickle down argument for \(+^\dagger\). First rewriting all multiplications by
zero via relation \textbf{(14)}, the two \(\Delta\) clusters incident to the coaddition can either
reduce to \(\Delta\) clusters that are incident only to nonzero multiplications or reduce to a
single deletion, as above, if none of the incident multiplications were nonzero. There are three
cases of what can happen from here.
\begin{itemize}[leftmargin=1em]
\item \Define{Both \(\Delta\) clusters were incident to only zero-multiplications}\\
In the first case, as above, the \(\Delta\) clusters will reduce to \(!\) incident to the outputs of
\(+^\dagger\). Relations \textbf{(D7)} and \textbf{(28)} delete the coaddition.
\item \Define{One \(\Delta\) cluster was incident to only zero-multiplications}\\
Without loss of generality, the \(!\) incident to \(+^\dagger\) is on the left. Relation
\textbf{(D7)} replaces \(!\) and \(+^\dagger\) with \(!^\dagger \circ !\), and relation \textbf{(30)}
replaces \(\Delta\) and \(!^\dagger\) with a cap. The \(\Delta\) was -- and the cap is -- incident
to some multiplication by \(c \ne 0\). Without loss of generality, \(c\) is incident to the bottom
addition in the cluster. Relation \textbf{(29)} replaces the addition and cozero with a cup and
multiplication by \(-1\), which combines with \(c\) by relation \textbf{(11)}. The cup and cap turn
\(-c\) around to its adjoint, which is \(-c^{-1}\) by relation \textbf{(31)}.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[coplus] (cop) at (0.5,0.866) {};
\node[delta] (dub) at (1,0) {};
\node[bang] (cobig) at (0,0.2) {};
\node[multiply] (c1) at (0.5,-1) {\(c\)};
\node (eq) at (2.1,0) {\(=\)};
\node (rel1a) at (2.1,-0.3) {(D7)\({}^\dagger\)};
\node (rel1b) at (2.1,-0.75) {(30)};
\draw (cop.io) -- +(0,0.5) (cop.left out) -- (cobig) (cop.right out) -- (dub.io)
(dub.left out) .. controls +(240:0.2) and +(90:0.2) .. (c1.90);
\draw[dotted] (c1.io) -- +(0,-0.5) (dub.right out) -- +(300:0.5);
\node[bang] (cobang) at (3.5,0.8) {};
\node[multiply] (c2) at (3,-1) {\(c\)};
\draw[out=90,in=90,relative,looseness=2] (cobang) -- +(0,0.5) (c2.90) -- ++(0,0.5) to +(1,0);
\draw[dotted] (c2.io) -- +(0,-0.5) (c2.90) ++(1,0.5) -- +(0,-0.7);
\end{tikzpicture}
\qquad
\begin{tikzpicture}[thick]
\node[multiply] (c) at (1,-1) {\(c\)};
\node[plus] (sum) at (0.5,-2) {};
\node[zero] (coz) at (0.5,-2.65) {};
\node (eq1) at (2.5,-1) {\(=\)};
\node (rel1) at (2.5,-1.3) {(29)};
\node (rel1b) at (2.5,-1.7) {(11)};
\draw[out=90,in=90,relative,looseness=2] (c.90) -- ++(0,0.5) to +(1,0) (sum.io) -- (coz)
(c.io) .. controls +(270:0.2) and +(60:0.2) .. (sum.right in);
\draw[dotted] (sum.left in) -- +(120:0.5) (c.90) ++(1,0.5) -- +(0,-0.7);
\node[multiply] (negc) at (4,-1) {\(\scriptstyle{-}\)\(c\)};
\node (eq2) at (5.5,-1) {\(=\)};
\node (rel2) at (5.5,-1.3) {(31)};
\draw[out=90,in=90,relative,looseness=2] (negc.90) -- ++(0,0.4) to +(1,0)
(negc.io) to (negc.io) -- ++(0,-0.4) to +(-1,0);
\draw[dotted] (negc.90) ++(1,0.4) -- +(0,-0.7) (negc.io) ++(-1,-0.4) -- +(0,0.7);
\node[multiply] (inv) at (6.5,-1) {\(\!\scriptstyle{-}\)\(c^{-1}\!\!\)};
\draw[dotted] (inv.90) -- +(0,0.7) (inv.io) -- +(0,-0.7);
\end{tikzpicture}
}
\end{center}
An addition cluster is above \(-c^{-1}\) and a duplication cluster is below, but because those
clusters are not otherwise connected to each other, there is a vertical arrangement of the morphisms
in the \(\mathrm{FinVect}_k\) block of the string diagram such that no cups or caps are present.
\item \Define{Both \(\Delta\) clusters are incident to at least one nonzero multiplication}\\
Using relation \textbf{(D5)}\({}^\dagger\), a \(+^\dagger\) will pass through one \(\Delta\) at a
time. A new \(\Delta^\dagger\) is created each time, but this can trickle down as before.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[-, thick, node distance=0.7cm]
\node (out1z) {};
\node (out2z) [right of=out1z, shift={(0.2,0)}] {};
\node (out3z) [right of=out2z, shift={(0.45,0)}] {};
\node [delta] (delzip) [above right of=out2z, shift={(0.1,0.3)}] {};
\node [coplus] (coadd) [above left of=delzip, shift={(0.05,0.3)}] {};
\node (inz) [above of=coadd] {};
\node (equal) [above right of=delzip, shift={(0.2,0)}] {\(=\)};
\node [coplus] (coaddl) [right of=equal, shift={(0.2,0)}] {};
\node (cross) [below right of=coaddl, shift={(-0.1,0)}] {};
\node [codelta] (nabla) [below left of=cross, shift={(0.1,0)}] {};
\node (out1u) [below of=nabla] {};
\node [coplus] (coaddr) [above right of=cross, shift={(-0.1,0)}] {};
\node [delta] (deltunzip) [above left of=coaddr, shift={(0.1,0.3)}] {};
\node (inu) [above of=deltunzip] {};
\node (out2u) [right of=out1u, shift={(0.4,0)}] {};
\node (out3u) [right of=out2u, shift={(0.1,0)}] {};
\draw (out1z) -- (coadd.right in) (coadd) -- (inz) (out2z) --
(delzip.left out) (out3z) -- (delzip.right out) (delzip.io) -- (coadd.left in);
\path
(nabla.left in) edge [bend left=30] (coaddl.right in);
\draw (out1u) -- (nabla) (nabla.right in) -- (coaddr.right in);
\draw (out2u) -- (cross) -- (coaddl.left in) (out3u) -- (coaddr.left in);
\draw (coaddl.io) -- (deltunzip.left out) (coaddr.io) -- (deltunzip.right out) (deltunzip) -- (inu);
\end{tikzpicture}
}
\end{center}
Once the \(\Delta^\dagger\) trickles down, there are two possibilities for what is directly beneath
each \(+^\dagger\): either the same scenario will recur with a \(\Delta\) connected to one or both
outputs, which can only happen finitely many times, or two nonzero multiplications will be below the
\(+^\dagger\). A multiplication by any unit in \(k\), \(c \ne 0\), can move through a coaddition by
inserting \(c c^{-1}\) on the top branch and applying relation \textbf{(15)}\({}^\dagger\):
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[coplus] (cop1) at (0.5,1) {};
\node[multiply] (c1) at (0,0) {\(c\)};
\node (eq) at (1.7,0.5) {\(=\)};
\node[multiply] (c2) at (2.8,1.8) {\(c\)};
\node[coplus] (cop2) at (2.8,1) {};
\node[multiply] (cinv) at (3.3,-0.1) {\(c^{-1}\!\!\)};
\draw (cop1.io) -- +(0,0.5) (cop1.right out) .. controls +(300:0.5) and +(90:0.9) .. +(0.5,-1.5)
(cop1.left out) .. controls +(240:0.2) and +(90:0.2) .. (c1.90) (c1.io) -- +(0,-0.3)
(c2) -- +(0,0.5) (c2.io) -- (cop2.io) (cop2.left out) .. controls +(240:0.5) and +(90:0.9) .. +(-0.5,-1.9)
(cop2.right out) .. controls +(300:0.2) and +(90:0.2) .. (cinv.90) (cinv.io) -- +(0,-0.3);
\end{tikzpicture}
}
\end{center}
This allows one of the outputs of the coaddition to connect directly to a \(+\) cluster.
\begin{itemize}[leftmargin=1em]
\item \Define{If both branches go to different \(+\) clusters}, Frobenius relations
\textbf{(21)--(22)} slide the \(+^\dagger\) down the \(+\) cluster on one side until it gets to the
end of that cluster.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[plus] (P1a) {};
\node[plus] (P1b) at (-0.5,0.866) {};
\node[coplus] (cop1) at (0.2,1.61) {};
\node[multiply] (c1) at (0.9,0.41) {\(c\)};
\node[zero] (out1) at (0,-0.5) {};
\node (eq) at (1.7,0.75) {\(=\)};
\node (rel1) at (1.7,0.45) {(21)};
\node (rel2) at (1.7,0.05) {(22)};
\draw (P1a.io) -- (out1) (c1.io) -- +(0,-0.3) (cop1.io) -- +(0,0.3) (P1a.right in) -- +(60:0.4)
(cop1.right out) -- (c1.90) (P1b.right in) -- (cop1.left out) (P1b.left in) -- +(120:0.5);
\draw[dotted] (P1a.left in) -- (P1b.io);
\node[plus] (P2a) at (2.8,1) {};
\node[plus] (P2b) at (2.3,1.866) {};
\node[coplus] (cop2) at (2.8,0.42) {};
\node[multiply] (c2) at (3.3,-0.42) {\(c\)};
\node[zero] (out2) at (2.3,-0.23) {};
\draw (P2b.left in) -- +(120:0.5) (c2.io) -- +(0,-0.3) (P2a.io) -- (cop2.io) (cop2.left out) -- (out2)
(cop2.right out) -- (c2.90) (P2b.right in) -- +(60:0.5) (P2a.right in) -- +(60:0.5);
\draw[dotted] (P2a.left in) -- (P2b.io);
\end{tikzpicture}
}
\end{center}
The only morphisms added to the \(\mathrm{FinVect}_k\) block that are not from \(\mathrm{FinVect}_k\) were the coaddition
and the cozero. Since these reduce to an identity morphism string by relation
\textbf{(1)}\({}^\dagger\), the \(\mathrm{FinVect}_k\) block is truly a \(\mathrm{FinVect}_k\) block again.
\item \Define{If both branches go to the same \(+\) cluster}, relation \textbf{(3)} and the
Frobenius relation \textbf{(21)} take both branches to the same addition.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[coplus] (cop1) at (0,1) {};
\node[plus] (add1a) at (-0.25,-0.433) {};
\node[plus] (add1b) at (0.25,-1.3) {};
\node[multiply] (c1) at (0.65,-0.104) {\(c\)};
\node (eq) at (1.3,-0.15) {\(=\)};
\node (rel1) at (1.3,-0.45) {(3)};
\node (rel2) at (1.3,-0.85) {(21)};
\draw (add1a.left in) .. controls +(120:0.5) and +(240:0.5) .. (cop1.left out)
(cop1.io) -- +(0,0.3) (cop1.right out) -- (c1.90) (add1b.io) -- +(0,-0.3)
(c1.io) .. controls +(270:0.2) and +(60:0.2) .. (add1b.right in);
\node[hole] (cross) at (0.42,0.5) {};
\draw (add1a.right in) -- +(60:2);
\draw[dotted] (add1a.io) -- (add1b.left in);
\node[plus] (topadd) at (2.3,1.2) {};
\node[coplus] (cop2) at (2.3,0.42) {};
\node[multiply] (c2) at (2.75,-0.42) {\(c\)};
\node[plus] (botadd) at (2.3,-1.5) {};
\draw (topadd.left in) -- +(120:0.5) (topadd.right in) -- +(60:0.5) (botadd.io) -- +(0,-0.4)
(cop2.right out) .. controls +(300:0.15) and +(90:0.15) .. (c2.90)
(c2.io) .. controls +(270:0.2) and +(60:0.2) .. (botadd.right in)
(botadd.left in) .. controls +(120:0.8) and +(240:0.8) .. (cop2.left out);
\draw[dotted] (cop2.io) -- (topadd.io);
\end{tikzpicture}
}
\end{center}
Depending on whether the remaining multiplication is by \(1\), either relation \textbf{(25)} reduces
the coaddition and the given addition to an identity string or relation \textbf{(D8)} applies. In
the former case we are done, and in the latter case relations \textbf{(D7)} and
\textbf{(10)}\({}^\dagger\) remove the \(!^\dagger\) introduced by applying relation (D8).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[bang] (cob1) at (0,1) {};
\node[plus] (sum1a) at (0.5,0.35) {};
\node[plus] (sum1b) at (1,-0.516) {};
\node (eq) at (2.2,0) {\(=\)};
\node (rel) at (2.2,-0.3) {(D7)};
\draw (cob1) -- (sum1a.left in) (sum1a.right in) -- +(60:0.5) (sum1a.io) -- (sum1b.left in)
(sum1b.right in) -- +(60:0.5) (sum1b.io) -- +(0,-0.5);
\node[bang] (cob2) at (2.9,0.134) {};
\node[bang] (bang) at (2.9,0.65) {};
\node[plus] (sum2) at (3.4,-0.516) {};
\draw (bang) -- +(0,0.5) (cob2) -- (sum2.left in) (sum2.right in) -- +(60:0.5)
(sum2.io) -- +(0,-0.5);
\end{tikzpicture}
\qquad
\begin{tikzpicture}[thick]
\node[bang] (cob1) at (0,0.65) {};
\node[plus] (sum1) at (0.5,0) {};
\node[zero] (coz1) at (0.5,-0.65) {};
\node (eq1) at (1.3,0) {\(=\)};
\node (rel1) at (1.3,-0.3) {(D7)};
\draw (cob1) -- (sum1.left in) (sum1.right in) -- +(60:0.5) (sum1.io) -- (coz1);
\node[bang] (cob2) at (1.9,-0.325) {};
\node[bang] (bang) at (1.9,0.325) {};
\node[zero] (coz2) at (1.9,-0.975) {};
\node (eq2) at (2.5,0) {\(=\)};
\node (rel2) at (2.5,-0.3) {(10)\({}^\dagger\)};
\draw (bang) -- +(0,0.65) (cob2) -- (coz2);
\node[bang] (cob3) at (3.1,-0.5) {};
\draw (cob3) -- +(0,1);
\end{tikzpicture}
}
\end{center}
\end{itemize}
\end{itemize}
\item \Define{\(\cup \circ S\) and \(S \circ \cap\)}\\
Composing with a cup below \(S\) is equivalent to composing with cap above \(T\), since \(\cap =
\cup^\dagger\). Using relation \textbf{(D10)}\({}^\dagger\), this cap can be replaced by multiplication by
\(-1\), coaddition, and zero. By the arguments above, \(-1\), \(+^\dagger\), and \(0\) can each be
absorbed into the \(\mathrm{FinVect}_k\) block.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.4,-0.8) rectangle (0.5,-0.35);
\draw[out=90,in=90,relative,looseness=2] (-0.25,-0.45) -- ++(0,0.9) to +(0.6,0) -- +(0,-0.9);
\node (eq1) at (1.15,0.225) {\(=\)};
\node (rel) at (1.15,-0.1) {(D10)\({}^\dagger\)};
\filldraw[fill=black,draw=black] (1.8,-1.4) rectangle (2.8,-0.95);
\node[multiply] (neg) at (2.65,-0.1) {\(\scriptstyle{-1}\)};
\node[coplus] (cops) at (2.3,1) {};
\node[zero] (zero1) at (2.3,1.7) {};
\node (eq2) at (3.5,0.225) {\(=\)};
\draw (zero1) -- (cops.io) (neg.io) -- (2.65,-1.05)
(cops.right out) .. controls +(300:0.2) and +(90:0.2) .. (neg.90)
(cops.left out) .. controls +(240:0.2) and +(90:0.2) .. (1.95,0) -- (1.95,-1.05);
\filldraw[fill=black,draw=black] (3.95,-0.8) rectangle (4.85,-0.35);
\node[coplus] (robbers) at (4.4,0.225) {};
\node[zero] (zero2) at (4.4,0.9) {};
\node (eq3) at (5.3,0.225) {\(=\)};
\draw (zero2) -- (robbers.io)
(robbers.left out) .. controls +(240:0.2) and +(90:0.2) .. +(255:0.5)
(robbers.right out) .. controls +(300:0.2) and +(90:0.2) .. +(285:0.5);
\filldraw[fill=black,draw=black] (5.75,0.2) rectangle (6.65,-0.25);
\node[zero] (zero3) at (6.2,0.65) {};
\draw (zero3) -- +(0,-0.5);
\node (eq4) at (7.1,0.225) {\(=\)};
\filldraw[fill=black,draw=black] (7.55,0.45) rectangle (8.45,0);
\end{tikzpicture}
}
\end{center}
The compositions with zero and multiplication by \(-1\) expand the \(\mathrm{FinVect}_k\) block, thus have no
effect on whether the diagram can be written in prestandard form.
\item \Define{\(! \circ S\)}\\
When composing \(!^\dagger\) above \(T\), two possibilities arise, depending on whether there is a
layer of \(\Delta\)s in the \(\mathrm{FinVect}_k\) block. If there is such a layer, relation \textbf{(30)} combines the
\(!^\dagger\) with a \(\Delta\), making a cap on top of \(T\). As we have just seen, this can be
rewritten in prestandard form.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.45,0.45) rectangle (0.45,0);
\node[bang] (cob1) at (0,0.9) {};
\node (eq1) at (0.9,0.225) {\(=\)};
\node[delta] (dub) at (1.5,0.225) {};
\node[bang] (cob2) at (1.5,0.9) {};
\filldraw[fill=black,draw=black] (1.05,-0.8) rectangle (1.95,-0.35);
\node (eq2) at (2.5,0.225) {\(=\)};
\node (rel) at (2.5,-0.075) {(30)};
\filldraw[fill=black,draw=black] (2.85,-0.8) rectangle (3.75,-0.35);
\draw[out=90,in=90,relative,looseness=2] (3,-0.45) -- ++(0,0.9) to +(0.6,0) -- +(0,-0.9)
(cob1) -- +(0,-0.5) (cob2) -- (dub.io)
(dub.left out) .. controls +(240:0.2) and +(90:0.2) .. +(255:0.5)
(dub.right out) .. controls +(300:0.2) and +(90:0.2) .. +(285:0.5);
\end{tikzpicture}
}
\end{center}
If no layer of \(\Delta\)s exists, relations \textbf{(31)}\({}^\dagger\) and
\textbf{(18)}\({}^\dagger\) pass the codeletion through a nonzero multiplication. Then relations
\textbf{(D7)} and \textbf{(10)}\({}^\dagger\) can be used to remove \(!^\dagger\), as we have
already seen. This leaves only the basic morphisms of \(\mathrm{FinVect}_k\) within the \(\mathrm{FinVect}_k\) block.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[bang] (cob1) at (0,1) {};
\node[multiply] (c) at (0,0) {\(c\)};
\node (eq1) at (0.8,0) {\(=\)};
\node (rel1) at (0.8,-0.3) {(31)\({}^\dagger\)};
\draw (cob1) -- (c) -- +(0,-1);
\node[bang] (cob2) at (1.8,1) {};
\node[upmultiply] (cinv) at (1.8,0) {\(\!c^{-1}\!\)};
\node (eq2) at (2.8,0) {\(=\)};
\node (rel2) at (2.8,-0.3) {(18)\({}^\dagger\)};
\draw (cob2) -- (cinv) -- +(0,-1);
\node[bang] (cob3) at (3.4,0.5) {};
\draw (cob3) -- +(0,-1);
\end{tikzpicture}
}
\end{center}
If the multiplication is \(c=0\), relation \textbf{(14)} converts \(c=0\) to \(0 \circ !\), allowing
relation \textbf{(28)} to remove the \(!^\dagger\), with the same conclusion.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node[bang] (cob1) at (0.3,1) {};
\node[multiply] (c0) at (0.3,0) {\(0\)};
\node (eq1) at (1,0) {\(=\)};
\node (rel1) at (1,-0.3) {(14)};
\draw (cob1) -- (c0) -- +(0,-1);
\node[bang] (cob2) at (1.6,0.925) {};
\node[bang] (bang) at (1.6,0.325) {};
\node[zero] (zero1) at (1.6,-0.325) {};
\node (eq2) at (2.2,0) {\(=\)};
\node (rel2) at (2.2,-0.3) {(28)};
\draw (cob2) -- (bang) (zero1) -- +(0,-0.65);
\node[zero] (zero2) at (2.8,0.5) {};
\draw (zero2) -- +(0,-1);
\end{tikzpicture}
}
\end{center}
\item \Define{\(c \circ S\) for \(c=0\)}\\
Composing with multiplication by \(c = 0\) below \(S\) is equivalent to composing with codeletion,
followed by tensoring with zero. Codeletion is the \(! \circ S\) case, and zero can be written in a
prestandard form, so this reduces to tensoring two diagrams that are in prestandard form.
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\filldraw[fill=black,draw=black] (-0.45,0.45) rectangle (0.45,0);
\node[multiply] (c) at (1.05,0) {\(0\)};
\draw[out=-90,in=-90,relative,looseness=2]
(c.io) -- (1.05,-1.35) (c.90) -- (1.05,0.45) to (0.15,0.45);
\node (eq) at (1.8,0) {\(=\)};
\node (rel) at (1.8,-0.3) {(14)};
\filldraw[fill=black,draw=black] (2.3,0.45) rectangle (3.2,0);
\node[zero] (ins2) at (3.8,-0.3) {};
\node[bang] (del2) at (3.8,0.3) {};
\draw[out=-90,in=-90,relative,looseness=2]
(ins2) -- (3.8,-1.35) (del2) -- (3.8,0.45) to +(-0.9,0);
\end{tikzpicture}
}
\end{center}
\end{itemize}
Finally, we need to show the prestandard forms can be rewritten in standard form. We need to show
what elementary row operations look like in terms of string diagrams. We also need to show for an
arbitrary prestandard string diagram \(S\) with \(\mathrm{FinVect}_k\) block \(T\) that if \(T\) is replaced
with \(T'\), the diagram where an elementary row operation has been performed on \(T\), the
resulting diagram \(S'\) can be built from \(S\) using relations \textbf{(1)--(31)}.
Because the \(i\)th output of a \(\mathrm{FinVect}_k\) diagram is a linear combinations of the inputs, with the
coefficients coming from the \(i\)th row of its matrix, rows of the matrix correspond to outputs of
the \(\mathrm{FinVect}_k\) block. Because of this, the row operation subdiagrams in \(S'\) will have
\(0^\dagger\)s immediately beneath them. Showing \(S'\) can be built from \(S\) reduces to showing
composition of row operations with \(0^\dagger\)s builds the same number of \(0^\dagger\)s.
\begin{itemize}[leftmargin=1em]
\item Add a multiple \(c\) of one row to another row:\\
If we want to add a multiple of the \(\beta\) row to the \(\alpha\) row, we need a map \((y_\alpha,
y_\beta) \mapsto (y_\alpha + c y_\beta, y_\beta)\). By the naturality of the braiding in a
symmetric monoidal category, we can ignore any intermediate outputs:
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (plus) at (0,0) {};
\node [multiply] (c) at (0.5,1.1) {\(c\)};
\node [delta] (dub) at (1,2) {};
\node (in1) at (-0.5,3) {\(y_\alpha\)};
\node (in2) at (1,3) {\(y_\beta\)};
\node (out1) at (0,-1) {\(y_\alpha + c y_\beta\)};
\node (out2) at (1.5,-1) {\(y_\beta\)};
\draw
(in1) .. controls +(270:2) and +(120:0.5) .. (plus.left in)
(plus.io) -- (out1) (dub.io) -- (in2)
(out2) .. controls +(90:2) and +(300:0.5) .. (dub.right out)
(dub.left out) .. controls +(240:0.2) and +(90:0.2) .. (c.90)
(c.io) .. controls +(270:0.2) and +(60:0.2) .. (plus.right in);
\end{tikzpicture}
\end{center}
When two cozeros are composed on the bottom of this diagram, the result is two cozeros:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node [plus] (plus) at (-0.3,0) {};
\node [multiply] (c) at (0.2,1.1) {\(c\)};
\node [delta] (dub) at (0.7,2) {};
\node [coordinate] (in1) at (-0.8,2.7) {};
\node [coordinate] (in2) at (0.7,2.7) {};
\node [zero] (out1) at (-0.3,-0.7) {};
\node [zero] (out2) at (1.2,-0.7) {};
\draw
(in1) .. controls +(270:2) and +(120:0.5) .. (plus.left in)
(plus.io) -- (out1) (dub.io) -- (in2)
(out2) .. controls +(90:2) and +(300:0.5) .. (dub.right out)
(dub.left out) .. controls +(240:0.2) and +(90:0.2) .. (c.90)
(c.io) .. controls +(270:0.2) and +(60:0.2) .. (plus.right in);
\node (eq1) at (1.85,1) {\(=\)};
\node (rel1) at (1.85,0.7) {(D10)};
\node [multiply] (c2) at (3.3,1.1) {\(c\)};
\node [multiply] (neg2) at (3.3,0.3) {\(\scriptstyle{-1}\)};
\node [delta] (dub2) at (3.8,2) {};
\node [zero] (coz2) at (4.3,-0.7) {};
\draw
(dub2) -- (3.8,2.7) (c2) -- (neg2)
(dub2.left out) .. controls +(240:0.2) and +(90:0.2) .. (c2.90)
(coz2) .. controls +(90:2) and +(300:0.5) .. (dub2.right out)
(neg2.io) .. controls +(270:0.5) and +(270:0.5) .. +(-0.8,0) -- (2.5,2.7);
\node (eq2) at (4.8,1) {\(=\)};
\node (rel2) at (4.8,0.7) {(11)};
\node [multiply] (c3) at (6.1,1.1) {\(\scriptstyle{-}\)\(c\)};
\node [delta] (dub3) at (6.6,2) {};
\node [zero] (coz3) at (7.1,-0.7) {};
\draw
(dub3) -- (6.6,2.7)
(dub3.left out) .. controls +(240:0.2) and +(90:0.2) .. (c3.90)
(coz3) .. controls +(90:2) and +(300:0.5) .. (dub3.right out)
(c3.io) .. controls +(270:0.5) and +(270:0.5) .. +(-0.8,0) -- (5.3,2.7);
\node (eq3) at (7.6,1) {\(=\)};
\node (rel3) at (7.6,0.7) {(D6)};
\node [multiply] (c4) at (8.9,0.2) {\(\scriptstyle{-}\)\(c\)};
\node [zero] (zero) at (8.9,1) {};
\node [zero] (coz4) at (8.9,1.5) {};
\draw
(zero) -- (c4) (coz4) -- (8.9,2.2)
(c4.io) .. controls +(270:0.5) and +(270:0.5) .. +(-0.8,0) -- (8.1,2.2);
\node (eq4) at (9.9,1) {\(=\)};
\node (rel4) at (9.9,0.7) {(16)};
\node [zero] (coz5a) at (10.5,0.5) {};
\node [zero] (coz5b) at (11,0.5) {};
\draw (coz5a) -- (10.5,1.5) (coz5b) -- (11,1.5);
\end{tikzpicture}
}
\end{center}
\item Swap rows:\\
If we want to swap the \(\beta\) row with the \(\alpha\) row, we need a map \((y_\alpha, y_\beta)
\mapsto (y_\beta, y_\alpha)\), which is the braiding of two outputs. Again, intermediate outputs
may be ignored:
\begin{center}
\begin{tikzpicture}[thick,node distance=0.5cm]
\node (fstart) {\(y_\alpha\)};
\node [coordinate] (ftop) [below of=fstart] {};
\node (center) [below right of=ftop] {};
\node [coordinate] (fout) [below right of=center] {};
\node (fend) [below of=fout] {\(y_\alpha\)};
\node [coordinate] (gtop) [above right of=center] {};
\node (gstart) [above of=gtop] {\(y_\beta\)};
\node [coordinate] (gout) [below left of=center] {};
\node (gend) [below of=gout] {\(y_\beta\)};
\draw [rounded corners] (fstart) -- (ftop) -- (center) --
(fout) -- (fend) (gstart) -- (gtop) -- (gout) -- (gend);
\end{tikzpicture}
\end{center}
When two cozeros are composed at the bottom of this diagram, the cut strings untwist by the
naturality of the braiding:
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick,node distance=0.5cm]
\node [coordinate] (fstart) {};
\node [coordinate] (ftop) [below of=fstart] {};
\node (center) [below right of=ftop] {};
\node [coordinate] (fout) [below right of=center] {};
\node [zero] (fend) [below of=fout] {};
\node [coordinate] (gtop) [above right of=center] {};
\node [coordinate] (gstart) [above of=gtop] {};
\node [coordinate] (gout) [below left of=center] {};
\node [zero] (gend) [below of=gout] {};
\draw [rounded corners=0.25cm] (fstart) -- (ftop) -- (center) --
(fout) -- (fend) (gstart) -- (gtop) -- (gout) -- (gend);
\node (eq1) [below right of=gtop, shift={(0.5,0)}] {\(=\)};
\node [zero] (ftop1) [above right of=eq1, shift={(0.5,0)}] {};
\node [coordinate] (fstart1) [above of=ftop1] {};
\node (center1) [below right of=ftop1] {};
\node [coordinate] (gtop1) [above right of=center1] {};
\node [coordinate] (gstart1) [above of=gtop1] {};
\node [coordinate] (gout1) [below left of=center1] {};
\node [zero] (gend1) [below of=gout1] {};
\draw [rounded corners=0.25cm] (fstart1) -- (ftop1)
(gstart1) -- (gtop1) -- (gout1) -- (gend1);
\node (eq2) [below right of=gtop1, shift={(0.35,0)}] {\(=\)};
\node [zero] (coz1) [right of=eq2, shift={(0.2,-0.5)}] {};
\node [zero] (coz2) [right of=coz1] {};
\draw (coz1) -- +(0,1) (coz2) -- +(0,1);
\end{tikzpicture}
}
\end{center}
\item Multiply a row by \(c \neq 0\):\\
The third row operation is multiplying an arbitrary row by a unit, but since \(k\) is a field, that
means any \(c \neq 0\). This is just the multiplication map on one of the outputs:
\begin{center}
\begin{tikzpicture}[thick]
\node [multiply] (c) at (0,0) {\(c\)};
\node (in) at (0,1) {\(y_\alpha\)};
\node (out) at (0,-1) {\(c y_\alpha\)};
\draw (in) -- (c) -- (out);
\end{tikzpicture}
\end{center}
Because \(c\) is a unit, \(c^{-1} \in k\), so the multiplication by \(c\) can be replaced by the
adjoint of multiplication by \(c^{-1}\).
\begin{center}
\scalebox{0.80}{
\begin{tikzpicture}[thick]
\node [multiply] (c) at (0,0) {\(c\)};
\node (in) at (0,1) {};
\node [zero] (out) at (0,-1) {};
\draw (in) -- (c) -- (out);
\node (eq1) at (0.8,0) {\(=\)};
\node (rel1) at (0.8,-0.35) {(31)\({}^\dagger\)};
\node [zero] (coz1) at (1.9,-1) {};
\node [upmultiply] (c1) at (1.9,-0.1) {\(\!c^{-1}\!\)};
\draw (coz1) -- (c1) -- (1.9,1);
\node (eq2) at (3,0) {\(=\)};
\node (rel2) at (3,-0.35) {(16)\({}^\dagger\)};
\node [zero] (coz2) at (3.7,-0.5) {};
\draw (coz2) -- +(0,1);
\end{tikzpicture}
}
\end{center}
\end{itemize}
\end{proof}
\section{An example}
\label{example}
A famous example in control theory is the `inverted pendulum': an upside-down
pendulum on a cart \cite{Friedland}. The pendulum naturally tends to fall over, but we
can stabilize it by setting up a feedback loop where we observe its position and move
the cart back and forth in a suitable way based on this observation. Without introducing
this feedback loop, let us see how signal-flow diagrams can be used to describe the pendulum
and the cart. We shall see that the diagram for a system made of parts is built
from the diagrams for the parts, not merely by composing and tensoring, but also with the
help of duplication and coduplication, which give additional ways to set variables equal to one another.
Suppose the cart has mass \(M\) and can only move back and forth in one direction, so its
position is described by a function \(x(t)\). If it is acted on by a total force
\(F_{\mathrm{net}}(t)\) then Newton's second law says
\[
F_{\mathrm{net}}(t) = M \ddot{x}(t) .
\]
We can thus write a signal-flow diagram with the force as input and the cart's position as output:
\begin{center}
\begin{tikzpicture}[thick]
\node[coordinate] (q) [label={[shift={(0,-0.6)}]\(x\)}] {};
\node [integral] (diff) [above of=q] {\(\int\)};
\node (v) [above of=diff, label={[shift={(0.4,-0.5)}]\(\dot x\)}] {};
\node [integral] (dot) [above of=v] {\(\int\)};
\node (a) [above of=dot, label={[shift={(0.4,-0.5)}]\(\ddot{x}\)}] {};
\node [multiply] (M) [above of=a] {\(\frac{1}{M}\)};
\node[coordinate] (F) [above of=m, label={[shift={(0,0)}]\(F_{\mathrm{net}}\)}] {};
\draw (F) -- (M) -- (dot) -- (diff) -- (q);
\end{tikzpicture}
\end{center}
The inverted pendulum is a rod of length \(\ell\) with a mass \(m\) at its end, mounted
on the cart and only able to swing back and forth in one direction, parallel to the cart's movement. If its angle from vertical, \(\theta(t)\), is small, then its equation of motion
is approximately linear:
\[
\ell \ddot{\theta}(t) = g \theta(t) - \ddot{x}(t)
\]
where \(g\) is the gravitational constant. We can turn this equation
into a signal-flow diagram with \(\ddot{x}\) as input and \(\theta\) as output:
\begin{center}
\begin{tikzpicture}[thick]
\node [multiply] (linverse) at (0,0) {\(-\frac{1}{l}\)};
\node [plus] (adder) at (-0.5,-2) {};
\node [integral] (int1) at (-0.5,-3) {\(\int\)};
\node [upmultiply] (goverl) at (-2,-3.5) {\(\frac{g}{l}\)};
\node [integral] (int2) at (-0.5,-4.3) {\(\int\)};
\node [delta] (split) at (-0.5,-5.5) {};
\node at (0 em,1.3) {\(\ddot{x}\)};
\node at (0 em,-7.4) {\(\theta\)};
\draw (linverse) -- (0,1)
(linverse.io) .. controls +(270:0.8) and +(60:0.7) .. (adder.right in)
(adder.io) -- (int1) -- (int2) -- (split)
(goverl.io) .. controls +(90:1.5) and +(120:1) .. (adder.left in)
(split.right out) .. controls +(300:0.7) and +(90:1.3) .. (0,-7)
(split.left out) .. controls +(240:1) and +(270:1.5) .. (goverl.270);
\end{tikzpicture}
\end{center}
Note that this already includes a kind of feedback loop, since the pendulum's angle affects
the force on the pendulum.
Finally, there is an equation describing the total force on the cart:
\[ F_{\mathrm{net}}(t) = F(t) - m g \theta(t) \]
where \(F(t)\) is an externally applied force and \(-mg\theta(t)\) is the force due to the
pendulum. It will be useful to express this as follows:
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (differ) at (0,0) {};
\node [upmultiply] (mg) at (1.5,-1.5) {\(\scriptstyle{-mg}\)};
\node at (0,-3.5) {\(F_{\mathrm{net}}\)};
\node at (1.5,-3.5) {\(\theta\)};
\node at (-0.5,1.8) {\(F\)};
\draw (mg.io) .. controls +(90:1.5) and +(60:1) .. (differ.right in)
(differ.left in) .. controls +(120:.7) and +(270:.7) .. (-0.5,1.5)
(differ.io) -- (0,-3.2)
(mg) -- (1.5,-3.2);
\end{tikzpicture}
\end{center}
Here we are treating \(\theta\) as an output rather than an input, with the help of a cap.
The three signal-flow diagrams above describe the following linear relations:
\begin{eqnarray}
x &=& \int \int \frac{1}{M} F_{\mathrm{net}} \label{eq.1} \\
\theta &=& \int \int
\left(\frac{g}{l} \,\theta - \frac{1}{l}\, \ddot{x}\right) \label{eq.2} \\
F_{\mathrm{net}} + mg \theta &=& F \label{eq.3}
\end{eqnarray}
where we treat \eqref{eq.1} as a relation with \(F_{\mathrm{net}}\) as input and
\(x\) as output, \eqref{eq.2} as a relation with \(\ddot{x}\) as input and \(\theta\)
as output, and \eqref{eq.3} as a relation with \(F\) as input and \((F_{\mathrm{net}},
\theta)\) as output.
To understand how the external force affects the position of the cart and the angle of the pendulum,
we wish to combine all three diagrams to form a signal-flow diagram that has the external force
\(F\) as input and the pair \((x, \theta)\) as output. This is not just a simple matter of
composing and tensoring the three diagrams. We can take \(F_{\mathrm{net}}\), which is an output of
\eqref{eq.3}, and use it as an input for \eqref{eq.1}. But we also need to duplicate \(\ddot{x}\),
which appears as an intermediate variable in \eqref{eq.1} since \(\ddot{x} = \frac{1}{M}
F_{\mathrm{net}}\), and use it as an input for \eqref{eq.2}. Finally, we need to take the variable
\(\theta\), which appears as an output of both \eqref{eq.2} and \eqref{eq.3}, and identify the two
copies of this variable using coduplication. Following traditional engineering practice, we shall
write coduplication in terms of duplication and a cup, as follows:
\begin{center}
\begin{tikzpicture}[thick]
\node [codelta] (split) at (-2,0) {};
\node [] at (-1,0) {\(=\)};
\node [delta] (theta) at (0,0) {};
\draw (theta.left out) .. controls +(240:0.5) and +(90:0.4) .. (-0.4,-1)
(theta.right out) .. controls +(300:1) and +(270:2) .. (1,1)
(theta.io) -- (0,1)
(split.left in) .. controls +(120:0.6) .. (-2.5,1)
(split.right in) .. controls +(60:0.6) .. (-1.5,1)
(split.io) -- (-2,-1);
\end{tikzpicture}
\end{center}
The result is this signal-flow diagram:
\vfill \eject
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (differ) at (0,0) {};
\node [upmultiply] (mg) at (1.5,-1.5) {\(\scriptstyle{-mg}\)};
\node [delta] (split) at (0,-2.5) {};
\node [multiply] (Minv) at (0,-1) {\(\frac{1}{M}\)};
\node at (-0.5,1.8) {\(F\)};
\draw (mg.io) .. controls +(90:1.5) and +(60:1) .. (differ.right in)
(differ.left in) .. controls +(120:.7) and +(270:.7) .. (-0.5,1.5)
(differ.io) -- (Minv.90)
(Minv.io) -- (split.io)
(mg.270) .. controls +(270:0.6) and +(90:0.6) .. (3.5,-4);
\node [integral] (int1) at (-0.5,-5.5) {\(\int\)};
\node [integral] (int2) at (-0.5,-8) {\(\int\)};
\node at (-0.5 cm,-13.3) {\(x\)};
\draw (split.left out) .. controls +(240:0.7) and +(90:1) .. (int1.90)
(int1.io) -- (int2.90)
(int2.io) -- (-0.5,-13);
\node [multiply] (linverse) at (2.5,-5) {\(-\frac{1}{l}\)};
\node [plus] (adder) at (2,-7) {};
\node [integral] (int3) at (2,-8) {\(\int\)};
\node [upmultiply] (goverl) at (0.75,-8.5) {\(\frac{g}{l}\)};
\node [integral] (int4) at (2,-9.3) {\(\int\)};
\node [delta] (split2) at (2,-10.5) {};
\node [delta] (theta) at (2.5,-11.5) {};
\node at (2 cm,-13.3) {\(\theta\)};
\draw (split.right out) .. controls +(300:1) and +(90:1) .. (linverse.90)
(linverse.io) .. controls +(270:0.6) and +(60:0.6) .. (adder.right in)
(adder.io) -- (int3) -- (int4) -- (split2)
(goverl.io) .. controls +(90:1.5) and +(120:1) .. (adder.left in)
(split2.right out) .. controls +(300:0.5) and +(90:0.5) .. (theta)
(theta.left out) .. controls +(240:0.7) and +(90:0.7) .. (2,-13)
(theta.right out) .. controls +(300:1) and +(270:1.5) .. (3.5,-10.5)
(3.5,-10.5) -- (3.5,-4)
(split2.left out) .. controls +(240:1) and +(270:1.5) .. (goverl.270);
\end{tikzpicture}
\end{center}
This is not the signal-flow diagram for the inverted pendulum that one sees in
Friedland's textbook on control theory \cite{Friedland}. We leave it as an exercise to the reader to
rewrite the above diagram using the rules given in this paper, obtaining Friedland's diagram:
\vfill \eject
\begin{center}
\begin{tikzpicture}[thick]
\node [delta] (F) at (0,10) {};
\node at (0,11.3) {\(F\)};
\node [multiply] (Minv) at (-1.75,8) {\(\frac{1}{M}\)};
\node [multiply] (Mlinv) at (1.75,8) {\(\frac{-1}{Ml}\)};
\node [plus] (xsum) at (-1.25,6) {};
\node [plus] (thsum) at (2.25,6) {};
\node [integral] (i1) at (-1.25,5) {\(\int\)};
\node [integral] (i2) at (-1.25,3.5) {\(\int\)};
\node at (-1.25,-1.3) {\(x\)};
\node [integral] (i3) at (2.25,5) {\(\int\)};
\node [integral] (i4) at (2.25,3.5) {\(\int\)};
\node [upmultiply] (mgM) at (0.5,4.25) {\(\frac{mg}{M}\)};
\node [upmultiply] (mess) at (4.5,3.25) {\(\!\!\frac{(M+m)g}{Ml}\!\!\)};
\node [delta] (theta1) at (2.25,2) {};
\node [delta] (theta2) at (2.75,0.5) {};
\node at (2.25,-1.3) {\(\theta\)};
\draw (F) -- (0,11)
(F.left out) .. controls +(240:0.7) and +(90:0.7) .. (Minv.90)
(F.right out) .. controls +(300:0.7) and +(90:0.7) .. (Mlinv.90)
(Minv.io) .. controls +(270:0.7) and +(120:0.7) .. (xsum.left in)
(Mlinv.io) .. controls +(270:0.7) and +(120:0.7) .. (thsum.left in)
(mgM.io) .. controls +(90:1.7) and +(60:1) .. (xsum.right in)
(mess.io) .. controls +(90:2.5) and +(60:1) .. (thsum.right in)
(xsum) -- (i1) -- (i2) -- (-1.25,-1)
(thsum) -- (i3) -- (i4) -- (theta1)
(mgM.270) .. controls +(270:1.7) and +(240:1.7) .. (theta1.left out)
(mess.270) .. controls +(270:1.7) and +(300:1.7) .. (theta2.right out)
(theta2.io) .. controls +(90:0.7) and +(300:0.7) .. (theta1.right out)
(theta2.left out) .. controls +(240:0.7) and +(90:0.7) .. (2.25,-1)
;
\end{tikzpicture}
\end{center}
As a start, one can use Theorem \ref{presrk} to prove that it is indeed possible to do this rewriting.
To do this, simply check that both signal-flow diagrams define the same linear relation. The proof of
the theorem gives a method to actually do the rewriting---but not necessarily the fastest method.
\section{Conclusions}
\label{conclusions}
We conclude with some remarks aimed at setting our work in context.
In particular, we would like to compare it to some other recent papers. On April 30th, 2014, after
most of this paper was written, Soboci\'nski told the first author about some closely related
papers that he wrote with Bonchi and Zanasi \cite{BSZ1,BSZ2}. These provide
interesting characterizations of symmetric monoidal categories equivalent to \(\mathrm{FinVect}_k\) and
\(\mathrm{FinRel}_k\). Later, while this paper was being refereed, Wadsley and Woods \cite{WW} generalized the first of
these results to the case where \(k\) is any commutative rig. We discuss Wadsley and Woods'
work first, since doing so makes the exposition simpler.
A particularly tractable sort of symmetric monoidal category is a PROP: that is, a strict symmetric
monoidal category where the objects are natural numbers and the tensor product of objects is given
by ordinary addition. The symmetric monoidal category \(\mathrm{FinVect}_k\) is equivalent to the PROP
\(\mathrm{Mat}(k)\), where a morphism \(f \maps m \to n\) is an \(n \times m\) matrix with entries in
\(k\), composition of morphisms is given by matrix multiplication, and the tensor product of
morphisms is the direct sum of matrices.
Wadsley and Woods gave an elegant description of the algebras of \(\mathrm{Mat}(k)\). Suppose \(C\) is a
PROP and \(D\) is a strict symmetric monoidal category. Then the \Define{category of algebras} of
\(C\) in \(D\) is the category of strict symmetric monoidal functors \(F \maps C \to D\) and
natural transformations between these. If for every choice of \(D\) the category of algebras of
\(C\) in \(D\) is equivalent to the category of algebraic structures of some kind in \(D\), we say
\(C\) is the PROP for structures of that kind.
In this language, Wadsley and Woods proved that \(\mathrm{Mat}(k)\) is the PROP for `bicommutative
bimonoids over \(k\)'. To understand this, first note that for any bicommutative bimonoid \(A\) in
\(D\), the bimonoid endomorphisms of \(A\) can be added and composed, giving a rig \(\mathrm{End}(A)\). A
bicommutative bimonoid \Define{over \(k\)} in \(D\) is one equipped with a rig homomorphism
\(\Phi_A \maps k \to \mathrm{End}(A)\). Bicommutative bimonoids over \(k\) form a category where a
morphism \(f \maps A \to B\) is a bimonoid homomorphism compatible with this extra structure,
meaning that for each \(c \in k\) the square
\[
\xymatrix{ A \ar[dd]_f \ar[rr]^{\Phi_A(c)} && A \ar[dd]^f \\ \\
B \ar[rr]_{\Phi_B(c)} && B
}
\]
commutes. Wadsley and Woods proved that this category is equivalent to the category of algebras of
\(\mathrm{Mat}(k)\) in \(D\).
This result amounts to a succinct restatement of Theorem \ref{presvk}, though technically the
result is a bit different, and the style of proof much more so. The fact that an algebra of
\(\mathrm{Mat}(k)\) is a bicommutative bimonoid is equivalent to our relations \textbf{(1)--(10)}. The
fact that \(\Phi_A(c)\) is a bimonoid homomorphism for all \(c \in k\) is equivalent to relations
\textbf{(15)--(18)}, and the fact that \(\Phi\) is a rig homomorphism is equivalent to relations
\textbf{(11)--(14)}.
Even better, Wadsley and Woods showed that \(\mathrm{Mat}(k)\) is the PROP for bicommutative bimonoids over
\(k\) whenever \(k\) is a commutative rig. Subtraction and division are not required to define the
PROP \(\mathrm{Mat}(k)\), nor are they relevant to the definition of bicommutative bimonoids over \(k\).
Working with commutative rigs is not just generalization for the sake of generalization: it
clarifies some interesting facts.
For example, the commutative rig of natural numbers gives a PROP \(\mathrm{Mat}({\mathbb N})\). This is equivalent
to the symmetric monoidal category where morphisms are isomorphism classes of spans of finite sets,
with disjoint union as the tensor product. Lack \cite[Ex.\ 5.4]{Lack} had already shown that this
is the PROP for bicommutative bimonoids. But this also follows from the result of Wadsley and
Woods, since every bicommutative bimonoid \(A\) is automatically equipped with a unique rig
homomorphism \(\Phi_A \maps {\mathbb N} \to \mathrm{End}(A)\).
Similarly, the commutative rig of booleans \({\mathbb B} = \{F,T\}\), with `or' as addition and `and' as
multiplication, gives a PROP \(\mathrm{Mat}({\mathbb B})\). This is equivalent to the symmetric monoidal category
where morphisms are relations between finite sets, with disjoint union as the tensor product.
Mimram \cite[Thm.\ 16]{Mimram} had already shown this is the PROP for \Define{special}
bicommutative bimonoids, meaning those where comultiplication followed by multiplication is the
identity:
\begin{center}
\begin{tikzpicture}[thick]
\node [plus] (sum) at (0.4,-0.5) {};
\node [delta] (cosum) at (0.4,0.5) {};
\node [coordinate] (in) at (0.4,1) {};
\node [coordinate] (out) at (0.4,-1) {};
\node (eq) at (1.3,0) {\(=\)};
\node [coordinate] (top) at (2,1) {};
\node [coordinate] (bottom) at (2,-1) {};
\path (sum.left in) edge[bend left=30] (cosum.left out)
(sum.right in) edge[bend right=30] (cosum.right out);
\draw (top) -- (bottom)
(sum.io) -- (out)
(cosum.io) -- (in);
\end{tikzpicture}
\end{center}
But again, this follows from the general result of Wadsley and Woods.
Finally, taking the commutative ring of integers \({\mathbb Z}\), Wadsley and Woods showed that \(\mathrm{Mat}({\mathbb Z})\)
is the PROP for bicommutative Hopf monoids. The key here is that scalar multiplication by \(-1\)
obeys the axioms for an antipode, namely:
\begin{center}
\begin{tikzpicture}[thick]
\node [delta] (cosum) at (0,1.8) {};
\node [multiply] (times) at (-0.34,0.97) {\tiny \(-1\)};
\node [plus] (sum) at (0,0) {};
\draw
(cosum.io) -- +(0,0.3) (sum.io) -- +(0,-0.3)
(cosum.right out) .. controls +(300:0.5) and +(60:0.5) .. (sum.right in)
(cosum.left out) .. controls +(240:0.15) and +(90:0.15) .. (times.90)
(times.io) .. controls +(270:0.15) and +(120:0.15) .. (sum.left in);
\node (eq) at (1.2,0.9) {\(=\)};
\node [bang] (bang) at (2.1,1.3) {};
\node [zero] (cobang) at (2.1,0.5) {};
\draw (bang) -- +(0,1.02) (cobang) -- +(0,-1.02);
\node (eq) at (3,0.9) {\(=\)};
\node [delta] (cosum) at (4.2,1.8) {};
\node [multiply] (times) at (4.54,0.97) {\tiny \(-1\)};
\node [plus] (sum) at (4.2,0) {};
\draw
(cosum.io) -- +(0,0.3) (sum.io) -- +(0,-0.3)
(cosum.left out) .. controls +(240:0.5) and +(120:0.5) .. (sum.left in)
(cosum.right out) .. controls +(300:0.15) and +(90:0.15) .. (times.90)
(times.io) .. controls +(270:0.15) and +(60:0.15) .. (sum.right in);
\end{tikzpicture}
\end{center}
More generally, whenever \(k\) is a commutative ring, the presence of \(-1 \in k\) guarantees that
a bimonoid over \(k\) is automatically a Hopf monoid over \(k\). So, when \(k\) is a commutative
ring, Wadsley and Woods' result implies that \(\mathrm{Mat}(k)\) is the PROP for Hopf monoids over \(k\).
Earlier, Bonchi, Soboci\'nski and Zanasi gave an elegant and very different proof that \(\mathrm{Mat}(R)\)
is the PROP for Hopf monoids over \(R\) when \(R\) is a principal ideal domain
\cite[Prop.\ 3.7]{BSZ1}. The advantage of their argument is that they build up the PROP for Hopf
monoids over \(R\) from smaller pieces, using some ideas developed by \cite{Lack}.
These authors also described a PROP that is equivalent to \(\mathrm{FinRel}_k\) as a symmetric monoidal
category whenever \(k\) is a field. In this PROP, which they call \({\mathbb {SV}}_k\), a morphism \(f \maps
m \to n\) is a linear relation from \(k^m\) to \(k^n\). They proved that \({\mathbb {SV}}_k\) is a pushout in
the category of PROPs and strict symmetric monoidal functors:
\[
\xymatrix{ \mathrm{Mat}(R) + \mathrm{Mat}(R)^{\mathrm{op}} \ar[dd] \ar[rr] && \mathrm{Span}(\mathrm{Mat}(R)) \ar[dd] \\ \\
\mathrm{Cospan}(\mathrm{Mat}(R)) \ar[rr] && {\mathbb {SV}}_k
}
\]
This pushout square requires a bit of explanation.
Here \(R\) is any principal ideal domain whose field of fractions is \(k\). For example, we
could take \(R = k\), though Bonchi, Soboci\'nski and Zanasi are more interested in the example
where \(R = {\mathbb R}[s]\) and \(k = {\mathbb R}(s)\). A morphism in \(\mathrm{Span}(\mathrm{Mat}(R))\) is an isomorphism
class of spans in \(\mathrm{Mat}(R)\). There is a covariant functor
\[ \begin{array}{ccc} \mathrm{Mat}(R) &\to& \mathrm{Span}(\mathrm{Mat}(R)) \\
m \stackrel{f}{\to} n &\mapsto & m \stackrel{1}{\leftarrow} m \stackrel{f}{\to} n \end{array} \]
and also a contravariant functor
\[ \begin{array}{ccc} \mathrm{Mat}(R) &\to& \mathrm{Span}(\mathrm{Mat}(R)) \\
m \stackrel{f}{\to} n &\mapsto & n \stackrel{f}{\leftarrow} m \stackrel{1}{\to} m. \end{array} \]
Putting these together we get the functor from \(\mathrm{Mat}(R) + \mathrm{Mat}(R)^{\mathrm{op}}\) to
\(\mathrm{Span}(\mathrm{Mat}(R)) \) that gives the top edge of the square. Similarly, a morphism in
\(\mathrm{Cospan}(\mathrm{Mat}(R))\) is an isomorphism class of cospans in \(\mathrm{Mat}(R)\), and we have both a
covariant functor
\[ \begin{array}{ccc} \mathrm{Mat}(R) &\to& \mathrm{Cospan}(\mathrm{Mat}(R)) \\
m \stackrel{f}{\to} n &\mapsto & m \stackrel{f}{\rightarrow} n \stackrel{1}{\leftarrow} n \end{array} \]
and a contravariant functor
\[ \begin{array}{ccc} \mathrm{Mat}(R) &\to& \mathrm{Cospan}(\mathrm{Mat}(R)) \\
m \stackrel{f}{\to} n &\mapsto & n \stackrel{1}{\rightarrow} n \stackrel{f}{\leftarrow} m. \end{array} \]
Putting these together we get the functor from \( \mathrm{Mat}(R) + \mathrm{Mat}(R)^{\mathrm{op}} \) to
\( \mathrm{Cospan}(\mathrm{Mat}(R)) \) that gives the left edge of the square.
Bonchi, Soboci\'nski and Zanasi analyze this pushout square in detail, giving explicit
presentations for each of the PROPs involved, all based on their presentation of \(\mathrm{Mat}(R)\). The
upshot is a presentation of \({\mathbb {SV}}_k\) which is very similar to our presentation of the equivalent
symmetric monoidal category \(\mathrm{FinRel}_k\). Their methods allow them to avoid many, though not all, of
the lengthy arguments that involve putting morphisms in `normal form'.
\subsection*{Acknowledgements}
We thank Jamie Vicary for pointing out the relevance of the ZX calculus when the first author gave
a talk on this material at Oxford in February 2014 \cite{B}. Discussions with Brendan Fong have
also been very useful.
\bibliographystyle{plain}
|
\section{Model and
results}
\subsection{Introduction}
There is an increasing interest in materials whose
Fermi surface is not extended, as it is usually the case, but it consists of disconnected points.
In several of such materials the charge carriers admit at low energies an effective description in terms
of Dirac massless particles. This opens the exiting possibility that high energy phenomena have a counterpart
at low energies in real materials.
{\it Graphene} is probably the most known example of such systems;
it was pointed out in \cite{W1},\cite{S1}
that fermions on the honeycomb lattice
behave as massless Dirac fermions in $2+1$ dimensions
and indeed the experimental realization of graphene \cite{GN}, a monolayer sheet of graphite,
offered a spectacular physical realization of such a system.
As a next step, it is natural to look
for materials with
electronic bands touching in couples of points and with an emerging description in terms
of $3+1$ massless Dirac (or Weyl) particles, the
same appearing in the standard model to describe quarks and leptons. Such systems have been called
{\it Weyl semimetals}, and their existence has been predicted in several systems
\cite{1},\cite{1a},\cite{1b},\cite{J},\cite{DC}; this has generated an intense experimental research,
see for instance \cite{36}, \cite{37} (and the review \cite{VV1}).
It is of course important to understand the effect of the interactions, which are usually analyzed
in effective relativistic
models neglecting lattice effects \cite{40},\cite{41},
\cite{RL},\cite{MN}. Perturbative considerations suggest that short range interactions
can generate instabilities only at strong coupling, but in order to exclude non perturbative effects one has to prove the convergence of the expansions. It is also known that in such class of systems
the effective relativistic description
misses important features; for instance in the case of graphene the universality of the optical conductivity emerges only taking into account the lattice \cite{GMPcond}.
In this paper we consider
a three dimensional interacting fermionic {\it lattice} model
with a point-like Fermi surface \cite{DC} (see also \cite{J}), in presence
of an Hubbard interaction. We construct the zero temperature correlations for couplings
not too large, proving the persistence of the Weyl semimetallic phase in presence of interactions.
In the non interacting case the semimetallic phase, in which
the elementary excitations are well described in terms of Weyl fermions, is present in an extended region
of the parameters; outside such a region an insulating behavior is present
and a quantum critical point discriminates between the two phases.
In the semimetallic phase the Fermi surface consists of two points; close to the critical point the two points are very close and the Fermi velocity is arbitrarily small (and vanishes at the boundary).
The effective relativistic description coincides with a system of massless Dirac fermions
in $3+1$ dimensions with an ultraviolet cut-off like the Gross-Neveu model or QED with massive photon:
in such models
the interaction is irrelevant and the convergence of the renormalized perturbative expansion has been established, see \cite{PPO} and \cite{M2}.
However the convergence radius in such models is {\it vanishing} with the particle velocity; therefore
such results give essentially no information for lattice Weyl semimetals close to the boundary of the semimetallic phase where the Fermi velocity is very small.
One may suspect that
even an extremely weak interaction could produce some quantum instability close to the boundary of the semimetallic phase, where the parameters correspond to a strong coupling regime in the effective description.
This is however excluded by the present paper: we can prove the persistence of the Weyl semimetallic phase
in presence of interaction in all the semimetallic region, even arbitrarily close to the boundary
where the Fermi velocity vanishes. This result is achieved writing the correlations
in terms of a renormalized expansion with a radius of convergence which is independent from
the Fermi velocity, and in order to get this
one needs to exploit the non linear corrections to the dispersion relation due to the lattice. The proof is indeed based
on two different multiscale analysis in two regions of
the energy momentum space; in the smaller energy region the effective relativistic description is valid while in the larger energy region
the quadratic corrections due to the lattice are dominating. In both regimes the interaction is irrelevant
but the scaling dimensions are different; after the integration of the first regime one gets gain factors
which compensate
exactly the velocities at the denominator produced in the second regime, so that uniformity is achieved.
Such a phenomenon is completely absent in Graphene, in which the the Fermi velocity
is essentially constant. Another important phenomenon present here (and absent either in Graphene
and in the effective relativistic description) is
the movement of Weyl points due to the interaction.
The analysis is based
on the {\it Renormalization Group} (RG) method of {\it Wilson}
and its approach to the effective action \cite{W}, in the form implemented in \cite{G} and \cite{Po}
in the context of perturbative renormalization. It was realized in the eighties that such methods
can be indeed used to get a
full {\it non-perturbative} control of certain fermionic Quantum Field Theories in $d=1+1$, \cite{GK}, \cite{L}
using Gram bounds and Brydges formula for truncated expectation \cite{B}.
A very natural development was then to apply such techniques to condensed matter models,
\cite{BG},\cite{FMRT} with the final aim at obtaining a full non perturbative control of the ground state properties of interacting systems. However, while the interaction in the models considered in
\cite{GK} or \cite{L} is marginally irrelevant or dimensionally irrelevant, this is not the case
in interacting non relativistic fermionic models
in one dimension, or in dimensions greater than one with extended Fermi surface.
This is due to the fact
that the ground state properties of the interacting system are generically different with respect to the non interacting case. In one dimension
it was finally obtained a full control of the zero temperature properties of interacting fermions, in the spinless \cite{BGPS},\cite{BM1} or repulsive spinning case \cite{BFM}; this was achieved
by combining
RG methods with Ward Identities based on the emerging chiral symmetries.
In systems in higher dimensions with extended symmetric Fermi surface, rigorous results
were obtained, see
\cite{DR}, \cite{BGM}, for temperatures {\it above} an
exponentially small scale setting the onset of (possible) quantum instabilities.
Only in the case of an {\it asymmetric} Fermi surface (a condition preventing the formation
of Cooper pairs) the convergence of the renormalized expansion up to zero temperature for a interacting fermionic system was achieved
\cite{FKT}, proving the existence of a Fermi liquid phase.
In
systems with point-like Fermi surfaces in two or three dimensions the interaction is irrelevant and this allows
the proof of the convergence of the renormalized expansion up to zero temperature, as in the case of Graphene \cite{GM}
or the case discussed in the present paper. In the case of Graphene, the combination of
non perturbative bounds with lattice Ward Identities
allows to establish remarkable physical conclusions, like the universality of the optical conductivity \cite{GMPcond}.
Similarly, Ward Identities combined with the results obtained in the present paper
can be used to establish
a weak form of universality for the optical conductivity in Weyl semimetals, see
\cite{M4}.
\subsection{The model}
We consider the interacting version of the
tight binding model introduced in
\cite{DC}, describing fermions on a three dimensional lattice,
with nearest and next to nearest neighbor hopping and with a properly defined magnetic flux density, whose effect is to decorate the hopping with phase factors \cite{Ha}.
We consider two cubic sublattices $\L_A=\L$ and $\L_B$, where
$\L_B=\L_a+\vec\d_+$ and $\L=\{n_1\vec\d_1+n_2\vec\d_2+n_3\vec\d_3, n_1,n_2,n_3=0,1,...,L-1\}$
with
$\vec\d_1=(1,0,0)$,
$\vec\d_2=(0,1,0)$, $\vec\d_3=(0,0,1)$ and $\vec\d_{\pm}={\vec\d_1\pm\vec\d_2\over 2}$.
We introduce creation and annihilation
fermionic operators for electrons sitting at the sites of the A- and B- sublattices; if $\vec x\in\L$
\be
a^\pm_{\vec x}={1\over |\L|}\sum_{\vec k\in\DD_L} e^{\pm i \vec k\vec x}\hat a^\pm_{\vec k}\quad
b^\pm_{\vec x+\vec \d_+}={1\over |\L|}\sum_{\vec k\in\DD_L} e^{\pm i \vec k\vec x}\hat b^\pm_{\vec k}
\ee
with $\DD_L=\{\vec k={2\pi\over L}\vec n$, $\vec n=(n_1,n_2,n_3)$, $n_i=(0,1,...,L-1)\}$
and
\be
\{\hat a^\e_{\vec k},\hat a^{-\e'}_{\vec k'} \}=|\L|\d_{\vec k,\vec k'}\d_{\e,\e'}\quad\quad \{\hat b^\e_{\vec k},\hat b^{-\e'}_{\vec k'} \}=|\L|\d_{\vec k,\vec k'}\d_{\e,\e'}
\ee
and $\{\hat a^\e_{\vec k}, \hat b^{\e'}_{\vec k}\}=0$.
The hopping (or non-interacting) Hamiltonian is given by, if $\hat\psi^\pm_{\vec k}=(\hat a^\pm_{\vec k},\hat b^\pm_{\vec k})$
\be
H_0={1\over |\L|}\sum_{\vec k\in D_L}(\hat\psi^+_{\vec k},\EE(\vec k)\hat\psi^-_{\vec k})
\ee
where, if $k_\pm=\vec k\vec\d_\pm$
\be\EE(\vec k)=t\sin (k_+)\s_1+t\sin (k_-)\s_2+
\s_3(\m+t_\perp \cos k_3- {1\over 2}t'(\cos k_1+\cos k_2))
\ee
and
$$\s^1=
\begin{pmatrix}&0&1\\
&1&0\end{pmatrix}
\quad
\s^2=\begin{pmatrix}
&0&-i\\ &i&0
\end{pmatrix}
\quad\s^3=\begin{pmatrix}&1&0\\
&0&-1\end{pmatrix}$$
The hopping parameters $t,t_\perp,t'$ are assumed $O(1)$ and positive; in coordinate space $t_\perp$
describes the hopping between fermions living in different horizontal layers, $t$
the nearest neighbor hopping in the same layer, $t'$ the next-to-nearest neighbor hopping, while $\m$
the difference of energy between $a$ and $b$ fermions. The hopping terms are multiplied by suitable phases
to take into account a magnetic flux pattern applied to the lattice.
The electrons on the lattice can interact through a short range (or Hubbard) two body interaction, so that the total Hamiltonian is
\be
H=H_0+V\label{ham}
\ee
where
\be
V=U\sum_{\vec x,\vec y} v(\vec x-\vec y)
[a^+_{\vec x} a^-_{\vec x}+
b^+_{\vec x+\vec \d_+} b^-_{\vec x+\vec \d_+}][
a^+_{\vec y} a^-_{\vec y}+ b^+_{\vec y+\vec \d_+} b^-_{\vec y+\vec \d_+}
]\ee
and $|v(\vec x)|\le C e^{-\k|\vec x|}$ is a short-range interaction ($C,\k$ positive constants).
Defining $\psi^\pm_{\vec x}=(a^\pm_{\xx},b^\pm_{\xx+\d_+})$,
we consider the operators
\be
\psi^\pm_{\xx}=e^{x_{0}H}\psi^\pm_{\vec
x}e^{-x_{0}H}\ee with $\xx=(x_{0},\vec x)$ and $0<x_0<\b$
and $\b^{-1}$ is the temperature; on $x_0$ antiperiodic boundary conditions are imposed.
The 2-point {\it Schwinger function} is defined as
\be
S_{U}(\xx-\yy)={\Tr\{e^{-\b H} \psi^{-}_{\xx}\psi^{+}_{\yy}
\}\over
\Tr e^{-\b H}}\equiv \media{{\bf T}\{ \psi^{-}_{\xx_1}\psi^{+}_{\yy}\}
}_{\b,\L}
\label{sf}
\ee
where ${\bf T}$ is the fermionic time ordering operation.
\subsection{The non interacting case}
The Hamiltonian in the non interacting $U=0$ case can be easily written in diagonal form
\be
H_0={1\over |\L|}\sum_{\vec k\in \DD_L}[\l(\vec k)
\hat \a^+_{\vec k} \a^-_{\vec k}-\l(\vec k)\hat\b^+_{\vec k}\b^-_{\vec k}]
\ee
where
\bea
&&\l(\vec k)=\\
&&\sqrt{t^2(\sin^2 (k_+)+\sin^2 (k_-))+
(\m+t_\perp \cos k_3- {1\over 2}t'(\cos k_1+\cos k_2))^2}\nn
\eea
where $\hat\a^\pm_{\vec k},\hat \b^\pm_{\vec k}$ are sitable linear combinations of $\hat a^\pm_{\vec k},\hat b^\pm_{\vec k}$.
If $-\b<x_0-y_0\le \b$, the 2-point Schwinger function is given by
\bea &&\media{{\bf T}\{\a_{\xx}^-\a_{\yy}^+\}}_{\b,\L}=
{1\over|\L|}\sum_{\vec k\in\DD}e^{-i\vec k( \vec x-\vec
y)}\Big[\chi\big(x_0-y_0>0\big)\frac{e^{(x_0-y_0) \l(\vec
k)}} {1+e^{\b\l(\vec k)
}}-\nn
\\
&&\chi\big(x_0-y_0\le
0\big)
\frac{e^{(x_0-y_0+\b) \l(\vec k)}} {1+e^{\b\l(
\vec
k)}} \Big]
\label{A.19aa}
\eea
\bea
&&\media{{\bf T}\{\b_{\xx}^-\b_{\yy}^+\}}_{\b,\L}=
{1\over|\L|}\sum_{\vec k\in\DD}e^{-i\vec k( \vec x-\vec
y)}\Big[\chi\big(x_0-y_0>0\big)\frac{e^{-(x_0-y_0) \l(\vec
k)}} {1+e^{-\b\l(
\vec k)}}-\nn\\
&&\chi\big(x_0-y_0\le
0\big)\frac{e^{-(x_0-y_0+\b) \l(\vec k)}} {1+e^{-\b\l(
\vec
k)}} \Big] \label{A.19}\eea
and $\media{{\bf T}\{\a_{\xx}^-\b_{\yy}^+\}}_{\b,\L}=
\media{{\bf T}\{\b_{\xx}^-\a_{\yy}^+\}}_{\b,\L}=0$. A priori
Eq.(\ref{A.19aa}) and (\ref{A.19}) are defined only for
$-\b<x_0-y_0\le \b$, but we can extend them periodically over the
whole real axis; the periodic extension of the propagator is
continuous in the time variable for $x_0-y_0\not\in\b \ZZZ$, and it has
jump discontinuities at the points $x_0-y_0\in\b\ZZZ$.
Note that at $x_0-y_0=\b n$, the difference between the right and left
limits is equal to
$(-1)^n\d_{\vec x,\vec y}$, so that the propagator is discontinuous only
at $\xx-\yy=\b\ZZZ\times \vec 0$. For $\xx-\yy\not\in\b\ZZZ\times \vec 0$,
we can write, defining $\DD=D_L\times \DD_\b$, $\DD_\b=\{k_0={2\pi\over\b}(n_0+{1\over 2}), n_0\in \ZZZ\}$,
$\kk=(k_0,k)$
\bea &&\media{{\bf T}\{\a_{\xx}^-\a_{\yy}^+\}}_{\b,\L}=
\frac{1}
{\b|\L|}\sum_{\kk\in\DD_{\b,L}}e^{-i\kk(\xx-\yy)}
\frac1{-ik_0-
\l(\vec k)}\;, \label{A.19a1}\\
&&\media{{\bf
T}\{\b_{\xx}^-\b_{\yy}^+\}}_{\b,\L}=\frac{1}
{\b|\L|}\sum_{\kk\in\DD_{\b,L}}e^{-i\kk(\xx-\yy)}
\frac1{-ik_0+ \l(\vec k)}\;. \label{A.19a2}\eea
If we now re-express
$\a_{\xx}^\pm$ and $\b^\pm_{\xx,}$ in terms of
$a^\pm_{\xx,}$ and $b^\pm_{\xx+\dd_1,}$ we get
for $\xx-\yy\not\in\b\ZZZ\times \vec 0$
\be
S_0(\xx-\yy)={1\over |\L|\b}\sum_{\kk\in\DD_{L,\b}}
e^{i\kk(\xx-\yy)}A^{-1}(\kk)
\ee
where
\be
A(\kk)=-i k_0 I+t\s_1\sin k_++t\s_2 \sin k_-+(\m-t'+t_\perp \cos k_3+E(\vec k)
)\s_3
\ee
with
\be
E(\vec k)=t'(\cos k_+\cos k_--1)
\ee
The {\it Fermi surface} is defined as the singularity of the Fourier transform of the 2-point function $\hat S_0(\kk)=A^{-1}
(\kk)$ at zero temperature and $k_0=0$. Note that the functions $\sin (k_+)$ and $\sin (k_-)$ vanish in correspondence of two points
$(k_1,k_2)=(0,0)$ and $(k_1,k_2)=(\pi,\pi)$ and we will assume from now on
\be
\m+t'>2 t_{\perp}\label{a1}
\ee
so that
the only possible singularities are when
$\m-t'+t_\perp \cos k_3=0$. Therefore if
\be
{|\m-t'|\over t_\perp}<1\label{a2}
\ee
than $\hat S_0(\kk)$ is singular in correspondence of two points, called
{\it Weyl points} and denoted by
$\pm \vec p_F$, with
\be
\vec p_F=(0,0,\cos^{-1}({t'-\m\over t_\perp}))\label{w}
\ee
Close to such points
the 2-point function has the following form, if $\kk=\kk'\pm \pp_F$
and $|\kk'|<<|\sin p_F|$
\bea
\hat S_{0}(\kk'\pm \pp_F)\sim
\begin{pmatrix}&-i k_0\pm v_{3,0} k'_{3}
& v_{0}(k_+-i k_-)\\
&
v_{0}(k_++i k_-)&
-i k_0-(\pm) v_{3,0} k'_3
\end{pmatrix}^{-1}\label{rel}
\eea
with
\be
v_{0}=t\quad\quad v_{3,0}=t_\perp \sin p_F
\ee
The two $2\times 2$ matrices $\hat S_0(\kk'+\pp_F)$ and
$\hat S_0(\kk'-\pp_F)$ can be combined in a $4\times 4$ matrix coinciding with the propagator of a massless
Dirac (or Weyl) particle in $D=3+1$ dimension, with an {\it anisotropic} light velocity.
In coordinate space, the 2-point function has a power law
decay times an oscillating factor, denoting a metallic behavior (or semimetallic, as the conductivity computed via Kubo formula vanishes at zero frequency) under the conditions \pref{a1} and \pref{a2}.
On the contrary for ${|\m-t'|\over t_\perp}>1$
the 2-point function decays exponentially for large distances ($\hat S_0(\kk)$ is non singular)
and the system has an {\it insulating} behavior.
Close to the boundaries of the semimetallic phase, the Fermi velocity $t_{\perp}\sin p_F$ becomes arbitrarily small and the Weyl points are very close; the relativistic behavior \pref{rel} emerges only in a very small region
$|\kk'|<<t_\perp \sin p_F$
around the Fermi points as
the linear dispersion relation $v_{3,0} k'_3$ is dominating over the quadratic correction only in that region.
There is a {\it quantum critical point} ${|\m-t'|\over t_\perp}=1$ discriminating the metallic and the insulating region.
We ask now the question if Weyl semimetallic behavior, present under the conditions \pref{a1} and \pref{a2},
survives to the presence of the interaction.
\subsection{Grassmann Integral representation}
The analysis of the interacting case is done by a rigorous implementation of RG techniques. The starting point is a functional integral representation of the Schwinger functions which is quite suitable for such methods.
We want to establish the persistence of Weyl semimetallic behavior with Weyl points given by
\pref{w}. However, even if the semimetallic phase persists in presence of interaction, there is no reason {\it a priori} for which the value of $p_F$ should be the same in the free or interacting case.
Therefore it is convenient to proceed
in two steps. The first consists in writing
$\m=\m-\n+\n=\bar\m+\n$ and in proving that one can choose $\n=\n(\bar\m,\l)$ so that there is Weyl semimetallic behavior under the condition ${|\bar\m-t'|\over t_\perp}<1$, and that in such region the Weyl points are given by $(0,0,\pm p_F)$ with $\cos p_F={|\bar\m-t'|\over t_\perp}$;
in this way the location of the singularity of the two point function does not move, and this is technically convenient as we construct the interacting function as series starting from the non interacting one.
Once that this is
(possibly) done the second step consists in solving the inversion problem $\bar\m+\n(\bar\m,\l)=\m$,
so that one can determine the location of the Weyl points as function of the initial parameters. We will
deal here with the first step only, which is the substantial one; the inversion problem
can be done via standard methods once the first step is done, see for instance Lemma 2.8 of \cite{BFM}
for a similar problem.
We will introduce a set of {\it Grassmann variables}
$\hat\psi^\pm_{\kk}=(\hat a^\pm_\kk,\hat b_{\kk})$, $\kk\in \DD_{\b,L}$
by the same symbol as the fermionic fields. We also define a "regularized" propagator
$g_M(\xx-\yy)$ ($2^{M}$ is an ultraviolet cut-off)
with
\be
g_M(\xx-\yy)=
{1\over |\L|\b}\sum_{\kk\in\DD_{L,\b}}
e^{i\kk(\xx-\yy)}A^{-1}(\kk)\bar\chi(2^{-M}|k_0|)
\ee
with
$\bar\chi(t):\RRR^+\to \RRR$ is a smooth compact support function equal to $1$ for $0<t<1$ and $=0$ for $t>2$. Note that, contrary to the function $S_0(\xx-\yy)$, the sum $\sum_{\kk\in\DD_{L,\b}}$
is restricted over a finite number of elements. Note also that for $\xx-\yy\not=(0,n\b)$, $n\in \ZZZ$ than
\be
\lim_{M\to\io} g_M(\xx-\yy)=S_0(\xx-\yy)\label{fff}
\ee
The above equality is however not true for $\xx-\yy=(0,n\b)$; indeed the r.h.s. of \pref{fff} is discontinuous while the l.h.s. is equal to ${1\over 2}[S_0(0,0^+)+ S_0(0,0^-)]$.
We introduce the {\it generating functional}
\be
e^{\WW_M(\phi)}=\int P(d\psi)e^{\VV(\psi)+(\psi,\phi)}\label{gf}
\ee
where $P(d\psi)$ is the fermionic "measure" with propagator
$g_M(\xx-\yy)$ and
$\VV$ is the interaction given by
\be
\VV=(\n+\n_C) N+ V
\ee
where, if $\int d\xx= \int dx_0\sum_{\vec x}$
\bea
&&N=\int d\xx\psi^+_{\xx}\s_3
\psi^-_{\xx}\\
&&V=U
\int d\xx d\yy
v(\xx-\yy)(\psi^+_{\xx}I\psi^-_{\xx})(\psi^+_{\yy} I\psi^-_{\yy})\nn
\eea
if $v(\xx)=\d(x_0) v(x)$. Moreover
$(\psi,\phi)=\int d\xx [\psi^+_{\xx}\s_0\phi^-_{\xx}+\psi^-_{\xx}\s_0\phi^+_{\xx}]$
and $\n_C=U \hat v(0)[S_0(0,0^+) -S_0(0,0^-)]$. We define
\be
S_{2}(\xx-\yy)=\lim_{M\to\io} S_{M}(\xx-\yy)=\lim_{M\to\io}
{\partial^2 \WW_M\over\partial\phi^+_{\xx}\partial\phi^-_{\yy}}\Big|_0\label{sds}
\ee
It is easy to check order by order in perturbation theory that
$\lim_{M\to\io} S_{M,U}(\xx-\yy)$ coincides in the $M\to\io $ limit with $S_{U}(\xx-\yy)$ by
\pref{sf} with $\m$ replaced by $\m+\n$. Note indeed that both functions can be expressed in terms of
the same Feynman diagrams with propagator respectively $S_0(\xx-\yy)$ and $g_M(\xx-\yy)$. Therefore the equality is trivial except in the graphs containing a {\it tadpole}, involving a propagator computed at $(0,0)$;
the presence of the countertern $\n_C$ ensures than the equality, see \S 2.1 of \cite{BFM} for more details in a similar case.
One can prove more; if $S_{M}(\xx-\yy)$
given by \pref{sds}
is analytic and bounded in $|U|\le U_0$ with $U_0$ independent of $\b,L$
and uniformly convergent as $M\to\io$, then
$\lim_{M\to\io}S_{M}(\xx-\yy)=S_{U}(\xx-\yy)
$ where $S_{U}(\xx-\yy)$ is given by \pref{sf};
the proof of this fact is rather standard (it is an application of
Weierstarss theorem and of properties of analytic functions) and it will be not repeated here (see
Lemma 1 of \cite{GM} or prop 2.1 of \cite{BFM} for an explicit proof in similar cases). This ensures that
one can study directly the Grassmann integral \pref{sds}
to construct the
Schwinger function
\pref{sf}.
\subsection{Main results}
Our main result is the following.
\begin{theorem}
Let us consider $S_2(\xx)$ given by \pref{sds}
with $
\m+t'>2 t_{\perp}\label{a1}
$
and ${|\m-t'|\over t_\perp}<{3\over 2}$.
There exists $U_0>0$, independent of $\b,L $, such that if $|U|\le U_0$, it is possible to find a $\n$, analytic in $U$,
such that $S_2(\xx)$ exists and is analytic uniformly in $\b,L$ as $\b\to\io,L\to\io$. Moreover the Fourier transform of $S_2(\xx)$ in the $\b\to\io, L\to\io $ limit, denoted by $\hat S_2(\kk)$, in the case
${|\m-t'|\over t_\perp}<1$
is singular only at
$\pm \pp_F$, with $\cos p_F={|\m-t'|\over t_\perp}$, $v_{3,0}=t_\perp\sin p_F$ and close to the singularity,
\bea
\hat S_{2}(\kk'\pm \pp_F)=
{1\over Z}\begin{pmatrix}&-i k_0\pm v_{3} k'_{3}
& v (k_+-i k_-)\\
&
v (k_++i k_-)&
-i k_0-(\pm) v_{3} k'_3
\end{pmatrix}^{-1}(1+R(\kk'))\label{rel1}
\eea
with $|R(\kk)|\le {|\kk'|\over v_{3,0}}$ and \be Z=1+O(U),
\quad {v_3-v_{3,0}\over v_{3,0}}=O(U),\quad v=v_{0}+O(U)\;.\ee On the other hand for
${|\m-t'|\over t_\perp}>1$ the 2-point function is bounded for any $\kk$.
\end{theorem}
\vskip.3cm
{\bf Remarks}
\begin{enumerate}
\item The above theorem establishes analyticity in $U$ for values of the parameters including
either the semimetallic and the insulating phase, and proves for the first time
the existence of a Weyl semimetallic phase in an interacting system
with short range interactions. The effect
of the interaction is to generically
modify the location of the Weyl points (the counterterm $\n$ takes
this into account) and to change the parameters of the emerging relativistic description, like
the wave function renormalization and the "light" velocity.
\item Note that close to the boundary of the semimetallic phase the (third component) of the Fermi velocity $v_3$ is small, and vanishes
continuously at the
{\it quantum critical point} ${|\m-t'|\over t_\perp}=1$ discriminating between insulating and semimetallic phase. The estimated radius of convergence is uniform in $v_3$; this is remarkable as small $v_3$
correspond to a strong coupling regime in the effective relativistic description.
The main idea in order to achieve that is to perform a different multiscale analyis in two regions of the energy space, discriminated by an energy scale measuring the distance from the critical point.
\item The Renormalization Group analysis performed here to prove the above theorem can be
used to determine the large distance behavior of the current-current correlations.
As a consequence, in combination with
Ward Identities, some universality properties of the optical conductivity in the semimetallic phase can be proved, see \cite{M4}.
\end{enumerate}
\section{Renormalization Group analysis: First regime}
We find convenient the introduction of a parameter measuring the distance from the boundary of the semimetallic phase; therefore
we define
\be {\m-t'\over t_\perp}=-1+r\ee
with $|r|\le {1\over 2}$; the case $r={1\over 2}$ corresponds, in the non interacting case, to the semimetal with the highest velocity $v_3$, while at $r=0$ the Fermi velocity $v_{3}$ vanishes.
The starting point of the analysis of \pref{sds}
is the decomposition of the propagator in the following way
\be
g_M(\xx-\yy)=g^{(\le 0)}(\xx-\yy)+g^{(> 0)}(\xx-\yy)
\ee
where
\bea
&&\hat g^{(\le 0)}(\kk)=\bar\chi(\g^{-M}|k_0|)\chi_{0}(\kk)\hat A^{-1}(\kk)\\
&&\hat g^{(> 0)}(\kk)=\bar\chi(\g^{-M}|k_0|)(1-\chi_{0}(\kk))\hat A^{-1}(\kk)
\eea
and
$\chi_{<0}(\kk)=\bar\chi(a_0^{-1}|\det A(\kk)|^{1\over 2})$, with $a_0={t_\perp\over 10}$.
The above decomposition corresponds to a decomposition in the Grassmann variables
$\psi=\psi^{(\le 0)}+\psi^{(>0)}$ with propagators respectively
$g^{(\le 0)}(\xx)$ (the {\it infrared} propagator)
and $g^{(> 0)}(\xx)$ (the {\it ultraviolet} propagator).
We can write
\bea
&&e^{\WW(\phi)}=\int P(d\psi^{(> 0)}) P(d\psi^{(\le 0)})
e^{\VV(\psi^{(> 0)}+\psi^{(\le 0)}
)+(\psi^{(> 0)}+\psi^{(\le 0)}
,\phi)}=\nn\\
&&=e^{\b |\L| E_0}
\int P(d\psi^{(\le 0)})
e^{\VV^{(0)}(\psi^{(\le 0},\phi)
}\label{ss}
\eea
with
\be \VV^{(0)}(\psi,\phi)=\sum_{n,m\ge 0}
\int d\underline\xx \int d\underline\yy
[\prod_{i=1}^n\psi^{\e_i}_{\xx_i}][\prod_{i=1}^m\phi^{\e_i}_{\xx_i}]
W_{n,m}(\underline\xx,\underline\yy)\label{irr}
\ee
with $[\prod_{i=1}^n\psi^{\e_i}_{\xx_i}]=1$ if $n=0$ and $[\prod_{i=1}^m\phi^{\e_i}_{\yy_i}]=1$ if $m=0$, and for $U,\n$ smaller than a constant (independent from $L,\b,M$)
\be
{1\over \b |\L|}\int d\underline\xx \underline d\yy
|W^{(0)}_{n,m}(\underline\xx,\underline\yy)|\le \b |\L| C^{n+m}|U|^{max[1,n-1]}
\ee
Moreover $\lim_{M\to\io}W^{(0)}_{n,m}(\underline\xx,\underline\yy)$ and is reached uniformly.
The above properties follow from
Lemma 2 of \cite{BFM} (app. B) or Lemma 2.2 of \cite{BFM}; the proofs in such papers
are written for $d=2$ or $d=1$ lattice system, but the adaptation to the present case is straightforward
(due to the presence of a spatial lattice the ultraviolet problem
is essentially independent from dimension).
The infrared negative scales are
divided in two different regimes, which have to be analyzed
differently as they have different scaling properties.
They are discriminated by a scale \be h^*=[min(\log_2 a_0^{-1}10 |r|,0)]\ee
which discriminates the region where the non-linear corrections to the dispersion relation are dominating with the region where the energy is essentially linear; if $h^*=0$ the first regime described here
is absent. We describe now the integration of the scales $h\ge h^*$ inductively. Assume that we have integrated already
the scales $0,-1,..,h+1$
showing that \pref{gf} can be written as (in the $\phi=0$ for definiteness)
\be
e^{|\L| \b E_h}\int P(d\psi^{(\le h)}) e^{\VV^{(h)}(\sqrt{Z_{h}}\psi^{(\le h)})}\label{ef1}
\ee
where $P(d\psi^{(\le h)})$ has propagator given by
\be
g^{(\le h)}(\xx)=\int d\kk e^{i\kk\xx}{\chi_h(\kk)\over Z_h} A_h^{-1}(\kk)
\ee
where
\bea
&&A_h(\kk)=\\
&&\begin{pmatrix}&-i k_0+v_{3,h}(\cos k_3-1+r+E(\vec k)&
v_{h}(\sin k_+-i\sin k_-)\\
&
v_{h}(\sin k_++i \sin k_-)&
-i k_0-v_{3,h}(\cos k_3-1+r-E(\vec k)
\end{pmatrix}\label{cond2}\nn
\eea
$
\chi_h(\kk)=\bar\chi(a_0^{-1} 2^{-h}|\det A_h(\kk)|^{1\over 2})$ and
\bea
&&\VV^{(h)}(\psi)=\sum_{n\ge 1}
\int d\xx_1...\int d\xx_n
\prod_{i=1}^n\psi^{\e_i}_{\xx_i}
W^{(h)}_{n}(\underline\xx)=\nn\\
&&{1\over (|\L|\b)^n}\sum_{\kk_1,...,\kk_n}\prod_{i=1}^n\hat\psi^{\e_i}_{\kk_i}
\hat W^{(h)}_{n}(\kk_1,..,\kk_{n-1})\d(\sum_{i=1}^n\e_i \kk_i)\nn
\eea
We introduce a {\it localization operator} acting on the effective potential as
\be \VV^{(h)}=\tilde \LL \VV^{(h)}+\RR \VV^{(h)}\label{loc} \ee
with $\RR=1-\tilde \LL$ and $\tilde\LL$ is a linear operator acting on the kernels
$\hat W^{(h)}_{n}(\kk_1,..,\kk_{n-1})$ in the following way:
\begin{enumerate}
\item $\tilde\LL \hat W^{(h)}_{n}(\kk_1,..,\kk_{n-1})=0$ if $n>2$.
\item If $n=2$
\bea
&&\tilde\LL \hat W^{(h)}_{2}(\kk)= \\
&&\hat W^{(h)}_{2}(0)+k_0
\hat W^{(h)}_{2}(0)+\sum_{i=+,-,3} \sin k_i
\partial_i \hat W^{(h)}_{2}(0)+(\cos k_3-1)\partial^2_3 \hat W^{(h)}_{2}(0)\nn
\eea
\end{enumerate}
The definition of $\tilde\LL$ is written in the $L=\b=\io$ limit for definiteness but its expression
for $L,\b$ finite is straightforward.
By symmetry
\bea
&&\hat W^{(h)}_{2}(0)=\s_3 n_h\quad \partial_+ \hat W^{(h)}_{2}(0)=\s_1 b_{+,h}\quad
\partial_- \hat W^{(h)}_{2}(0)=\s_2 b_{-,h}
\nn\\
&& \partial_3 \hat W^{(h)}_{2}(0)=0\quad
\quad \partial_3^2 \hat W^{(h)}_{2}(0)=\s_3 b_{3,h}
\eea
Note also that, by definition $\LL\RR=0$.
We can include the local part of the effective potential in the fermionic integration, so
that \pref{ef1} can be rewritten as
\be
e^{|\L| \b \tilde E_h}\int \tilde P(d\psi^{(\le h)}) e^{\LL\VV^{(h)}(\sqrt{Z_{h-1}}\psi^{(\le h)}
+\RR \VV^{(h)}(\sqrt{Z_{h-1}}\psi^{(\le h)}
)}\label{ef2}
\ee
where
\be
\LL\VV^{(h)}(\psi^{(\le h)})=2^h \n_h \int d\xx\psi^{+(\le h)}_\xx\s_3\psi^{-(\le h)}_\xx
\ee
and $\tilde P(d\psi^{(\le h)})$ is the Grassmann integration with propagator
similar to \pref{cond2} with $Z_{h-1}(\kk),v_{h-1}(\kk),v_{3,h-1}(\kk)$ replacing
$Z_{h},v_{h},v_{3,h}$, where
\bea
&&
Z_{h-1}(\kk)=Z_h[1+\chi_h^{-1}(\kk) b_{0,h}]\nn\\
&&v_{h-1}(\kk)={Z_h\over Z_{h-1}(\kk)}
[v_h+\chi_h^{-1}(\kk) b_{+,h}]\label{bbe}\\
&&
v_{3, h-1}(\kk)={Z_h\over Z_{h-1}(\kk)}[v_{3,h}+\chi_h^{-1}(\kk) b_{3,h}]\nn
\eea
Now we write $\tilde P(d\psi^{(\le h)})=P(d\psi^{(\le h-1)}) P(d\psi^{(h)})$ where $P(d\psi^{(h)})$
has propagator similar to $\hat g^{(\le h)}$ \pref{cond2} with the following differences: a) $A_h$ is replaced by $A_{h-1}$, where
$Z_{h-1}\equiv Z_{h-1}(0)$, $v_{h-1}\equiv v_{h-1}(0)$, $v_{3,h-1}=v_{3, h-1}(0)$; b) $\chi_h$ is replaced by $f_h$, a smooth compact support function with support in $c_1 2^{h-1}\le |\det A_{h-1}(\kk)|^{1\over 2}\le c_2 2^{h+1}$, with $c_1<c_2$
positive constants.
Assuming that $Z_h,v_{3,h},v_{\pm, h}$ are close $O(U)$ to their value at $h=0$, one has
for any $N$ the following bound
\be
|g^{(h)}(\xx)|\le {1\over Z_h}
{2^{5h\over 2}\over 1+[2^h (|x_0|+|x_+|+|x_-|)+2^{h\over 2}|x_3|]^N}
\label{b}
\ee
where we have used that $a_0\g^h-|r|\le a_0\g^{h}(1-{1\over 10})$.
Therefore $k_0,k_\pm= O(2^h)$, $k_3=O(2^{h\over 2})$ for large negative $h$ and the bound \pref{b} follows by integration by parts.
Finally we perform the integration over $\psi^{(h)}$ obtaining
\be
e^{\b |\L|\tilde e_h+\VV^{(h-1}(\sqrt{Z_{h-1}}\psi^{(\le (h-1)})}=\int P(d\psi^{(h)}) e^{\LL\VV^{(h)}(\sqrt{Z_{h-1}}\psi^{(\le h)})+\RR\VV^{(h)}(\sqrt{Z_{h-1}}\psi^{(\le h)})}
\label{2.40cc}
\ee
obtaining an expression identical to \pref{ef1} with $h-1$ replacing $h$, $E_{h-1}=\tilde E_h+\bar
e_h$,
so that the procedure can be iterated.
\subsection{Tree expansion for the effective potentials.}
The effective potential
$\VV^{(h)}(\psi^{(\le h)})$ can be written in terms of a tree expansion, defined as
follows.
\insertplot{300}{150}{ \ins{30pt}{85pt}{$r$} \ins{50pt}{85pt}{$v_0$}
\ins{130pt}{100pt}{$v$} \ins{35pt}{-5pt}{$h$} \ins{52pt}{-5pt}{$h+1$}
\ins{135pt}{-5pt}{$h_{v}$} \ins{215pt}{-5pt}{$-1$} \ins{235pt}{-5pt}{0}
\ins{255pt}{-5pt}{$1$}} {fig50} {A renormalized tree for
$\VV^{(h)}$\lb{h2a}}{0}
\0 1) Let us consider the family of all trees which can be constructed
by joining a point $r$, the {\it root}, with an ordered set of $n\ge 1$
points, the {\it endpoints} of the {\it unlabeled tree},
so that $r$ is not a branching point. $n$ will be called the
{\it order} of the unlabeled tree and the branching points will be called
the {\it non trivial vertices}.
The unlabeled trees are partially ordered from the root
to the endpoints in the natural way; we shall use the symbol $<$
to denote the partial order.
Two unlabeled trees are identified if they can be superposed by a suitable
continuous deformation, so that the endpoints with the same index coincide.
It is then easy to see that the number of unlabeled trees with $n$ end-points
is bounded by $4^n$.
We shall also consider the {\it labeled trees} (to be called
simply trees in the following); they are defined by associating
some labels with the unlabeled trees, as explained in the
following items.
\0 2) We associate a label $h\le -1$ with the root and we denote
$\TT_{h,n}$ the corresponding set of labeled trees with $n$
endpoints. Moreover, we introduce a family of vertical lines,
labeled by an integer taking values in $[h,1]$, and we represent
any tree $\t\in\TT_{h,n}$ so that, if $v$ is an endpoint or a non
trivial vertex, it is contained in a vertical line with index
$h_v>h$, to be called the {\it scale} of $v$, while the root $r$ is on
the line with index $h$.
In general, the tree will intersect the vertical lines in set of
points different from the root, the endpoints and the branching
points; these points will be called {\it trivial vertices}.
The set of the {\it
vertices} will be the union of the endpoints, of the trivial
vertices and of the non trivial vertices; note that the root is not a vertex.
Every vertex $v$ of a
tree will be associated to its scale label $h_v$, defined, as
above, as the label of the vertical line whom $v$ belongs to. Note
that, if $v_1$ and $v_2$ are two vertices and $v_1<v_2$, then
$h_{v_1}<h_{v_2}$.
\0 3) There is only one vertex immediately following
the root, which will be denoted $v_0$ and cannot be an endpoint;
its scale is $h+1$.
\0 4) Given a vertex $v$ of $\t\in\TT_{h,n}$ that is not an endpoint,
we can consider the subtrees of $\t$ with root $v$, which correspond to the
connected components of the restriction of
$\t$ to the vertices $w\ge v$. If a subtree with root $v$ contains only
$v$ and an endpoint on scale $h_v+1$,
it will be called a {\it trivial subtree}.
\0 5) With each endpoint $v$ we associate one of the monomials
contributing to $\RR {\cal
V}^{(0)}(\psi^{(\le h_v-1)})$, corresponding to the terms
in the r.h.s. of (\ref{irr}) (with $\psi^{(\le 0)}$
replaced by $\psi^{(\le h_v-1)}$) and a set $\xx_v$ of space-time
points (the corresponding integration variables in the $\xx$-space
representation); or a term corresponding to $\LL\VV^{(h_v-1)}(\psi^{(\le h_v-1)})$.
\0 6) We introduce a {\it field label} $f$ to distinguish the field variables
appearing in the terms associated with the endpoints described in item 5);
the set of field labels associated with the endpoint $v$ will be called $I_v$;
note that $|I_v|$ is the order of the monomial contributing
to $\RR {\cal V}^{(0)}(\psi^{(\le h_v-1)})$ or $\LL\VV^{(h_v-1)}(\psi^{(\le h_v-1)})$
and associated to $v$.
Analogously, if $v$ is not an endpoint, we shall
call $I_v$ the set of field labels associated with the endpoints following
the vertex $v$; $\xx(f)$ will denote the
space-time point of the Grassmann field variable with label $f$.
In terms of these trees, the effective potential ${\cal V}^{(h)}$, $h\le -1$,
can be written as
\be {\cal V}^{(h)}(\psi^{(\le h)}) + \b|\L| \lis e_{k+1}=
\sum_{n=1}^\io\sum_{\t\in\TT_{h,n}}
{\cal V}^{(h)}(\t,\psi^{(\le h)})\;,\label{2.41}\ee
where, if $v_0$ is the first vertex of $\t$
and $\t_1,\ldots,\t_s$ ($s=s_{v_0}$)
are the subtrees of $\t$ with root $v_0$,
${\cal V}^{(h)}(\t,\psi^{(\le h)})$
is defined inductively as follows:\\
i) if $s>1$, then
\be {\cal V}^{(h)}(\t,\psi^{(\le h)})={(-1)^{s+1}\over s!} \EE^T_{h+1}
\big[\bar{\cal V}^{(h+1)}(\t_1,\psi^{(\le h+1)});\ldots; \bar{\cal V}^{(h+1)}
(\t_{s},\psi^{(\le h+1)})\big]\;,\label{2.42}\ee
where $\EE^T_{h+1}$ denotes the {\it truncated expectation} with propagator $g^{(h)}$ and
$\bar{\cal V}^{(h+1)}(\t_i,\psi^{(\le h+1)})$ is equal to $\RR{\cal
V}^{(h+1)}(\t_i,\psi^{(\le h+1)})$ if the subtree $\t_i$ contains
more than one end-point, or if it contains one end-point but it is
not a trivial subtree;
it is equal to $\RR{\cal V}^{(0)}(\psi^{(\le h+1)})$
or $\LL\VV^{(h+1)}(\psi^{(\le h+1)})$
if $\t_i$ is a trivial subtree;\\
ii) if $s=1$, then ${\bar V}^{(h+1)}(\t,\psi^{(\le h)})$ is equal to
$\big[\RR{\cal V}^{(h+1)}(\t_1,\psi^{(\le h+1)})\big]$
if $\t_1$ is not a trivial
subtree; it is equal to $\RR {\cal
V}^{(0)}(\psi^{(\le h+1)})- \RR{\cal V}^{(0)}(\psi^{(\le h)})\big]$
if $\t_1$ is a trivial subtree (and therefore its end-point $v$ has scale $h_v=h+2$).
\vskip.4cm
Using its inductive definition, the right hand side of (\ref{2.41}) can be
further expanded, and in order to describe the resulting expansion we need some
more definitions. We associate with any vertex $v$ of the tree a subset $P_v$ of $I_v$,
the {\it external fields} of $v$. These subsets must satisfy various
constraints. First of all, if $v$ is not an endpoint and $v_1,\ldots,v_{s_v}$
are the $s_v\ge 1$ vertices immediately following it, then
$P_v \subseteq \cup_i
P_{v_i}$; if $v$ is an endpoint, $P_v=I_v$.
If $v$ is not an endpoint, we shall denote by $Q_{v_i}$ the
intersection of $P_v$ and $P_{v_i}$; this definition implies that $P_v=\cup_i
Q_{v_i}$. The union ${\cal I}_v$ of the subsets $P_{v_i}\setminus Q_{v_i}$
is, by definition, the set of the {\it internal fields} of $v$,
and is non empty if $s_v>1$.
Given $\t\in\TT_{h,n}$, there are many possible choices of the
subsets $P_v$, $v\in\t$, compatible with all the constraints. We
shall denote ${\cal P}_\t$ the family of all these choices and ${\bf P}$
the elements of ${\cal P}_\t$.
With these definitions, we can rewrite
${\cal V}^{(h)}(\t,\psi^{(\le h)})$
in the r.h.s. of (\ref{2.41}) as:
\bea &&{\cal V}^{(h)}(\t,\psi^{(\le
h)})=\sum_{{\bf P}\in{\cal P}_\t}
{\cal V}^{(h)}(\t,\PP)\;,\nn\\
&&{\cal V}^{(h)}(\t,\PP)=\int d\xx_{v_0}
\widetilde\psi^{(\le h)}(P_{v_0})
K_{\t,\PP}^{(h+1)}(\xx_{v_0})\;,\label{2.43}\eea
where
\be \widetilde\psi^{(\le h)}
(P_{v})=\prod_{f\in P_v}\psi^{ (\le
h)\e(f)}_{\xx(f)}\label{2.44}\ee
and $K_{\t,\PP}^{(h+1)}(\xx_{v_0})$ is defined inductively by
the equation, valid for any $v\in\t$ which is not an endpoint,
\be K_{\t,\PP}^{(h_v)}(\xx_v)={1\over s_v !}
\prod_{i=1}^{s_v} [K^{(h_v+1)}_{v_i}(\xx_{v_i})]\; \;\EE^T_{h_v}[
\widetilde\psi^{(h_v)}(P_{v_1}\setminus Q_{v_1}),\ldots,
\widetilde\psi^{(h_v)}(P_{v_{s_v}}\setminus
Q_{v_{s_v}})]\;,\label{2.45}\ee
where $\widetilde\psi^{(h_v)}(P_{v_i}\setminus Q_{v_i})$ has a definition
similar to (\ref{2.44}). Moreover, if $v_i$ is an endpoint
$K^{(h_v+1)}_{v_i}(\xx_{v_i})$ is equal to one
of the kernels of the monomials contributing to
$\RR{\cal V}^{(0)}(\psi^{(\le h_v)})$ or
$\LL{\cal V}^{(h_v)}(\sqrt{Z_{h_v}}\psi^{(\le h_v)})$; if $v_i$ is not an
endpoint, $K_{v_i}^{(h_v+1)}=K_{\t_i,\PP_i}^{(h_v+1)}$,
where ${\bf P}_i=\{P_w, w\in\t_i\}$.
We further decompose ${\cal V}^{(h)}(\t,\PP)$, by using the
following representation of the truncated expectation in the r.h.s. of
(\ref{2.45}). Let us put $s=s_v$, $P_i\=P_{v_i}\setminus Q_{v_i}$;
moreover we order in an arbitrary way the sets $P_i^\pm\=\{f\in
P_i,\e(f)=\pm\}$, we call $f_{ij}^\pm$ their elements and we
define $\xx^{(i)}=\cup_{f\in P_i^-}\xx(f)$, $\yy^{(i)}=\cup_{f\in
P_i^+}\xx(f)$, $\xx_{ij}=\xx(f^-_{ij})$,
$\yy_{ij}=\xx(f^+_{ij})$. Note that $\sum_{i=1}^s
|P_i^-|=\sum_{i=1}^s |P_i^+|\=n$, otherwise the truncated
expectation vanishes.
Then, we use the {\it Brydges-Battle-Federbush} formula \cite{L} saying that, up to a sign, if $s>1$,
\be \EE^T_{h}(\widetilde\psi^{(h)}(P_1),\ldots,
\widetilde\psi^{(h)}(P_s))=\sum_{T}\prod_{l\in T}
g^{(h)}(\xx_l-\yy_l)
\int dP_{T}({\bf t})\; {\rm det}\, G^{h,T}({\bf t})\;,\label{2.46}\ee
where $T$ is a set of lines forming an {\it anchored tree graph} between
the clusters of points $\xx^{(i)}\cup\yy^{(i)}$, that is $T$ is a
set of lines, which becomes a tree graph if one identifies all the
points in the same cluster. Moreover ${\bf t}=\{t_{ii'}\in [0,1],
1\le i,i' \le s\}$, $dP_{T}({\bf t})$ is a probability measure with
support on a set of ${\bf t}$ such that $t_{ii'}={\bf u}_i\cdot{\bf u}_{i'}$
for some family of vectors ${\bf u}_i\in \RRR^s$ of unit norm. Finally
$G^{h,T}({\bf t})$ is a $(n-s+1)\times (n-s+1)$ matrix, whose elements
are given by
\be G^{h,T}_{ij,i'j'}=t_{ii'}g^{(h)}(\xx_{ij}-\yy_{i'j'})\;,
\label{2.48}\ee
with $(f^-_{ij}, f^+_{i'j'})$ not belonging to $T$.
In the following we shall use (\ref{2.44}) even for $s=1$, when $T$
is empty, by interpreting the r.h.s. as equal to $1$, if
$|P_1|=0$, otherwise as equal to ${\rm det}\,G^{h}=
\EE^T_{h}(\widetilde\psi^{(h)}(P_1))$.
It is crucial to note that $G^{h,T}$ is a Gram matrix, i.e.,
the matrix elements in (\ref{2.48}) can be written in terms of scalar products, and therefore it can be bounded by the Gram-Hadamard inequality.
If we apply the expansion (\ref{2.46}) in each vertex of
$\t$ different from the endpoints, we get an expression of the form
\be {\cal V}^{(h)}(\t,\PP) = \sum_{T\in {\bf T}} \int d\xx_{v_0}
\widetilde\psi^{(\le h)}(P_{v_0}) W_{\t,\PP,T}^{(h)}(\xx_{v_0})
\;,\label{2.49}\ee
where ${\bf T}$ is a special family of graphs on the set of points
$\xx_{v_0}$, obtained by putting together an anchored tree graph
$T_v$ for each non trivial vertex $v$. Note that any graph $T\in
{\bf T}$ becomes a tree graph on $\xx_{v_0}$, if one identifies
all the points in the sets $\xx_v$, with $v$ an endpoint.
Given $\t\in\TT_{h,n}$ and the labels $\PP,T$,
calling $I_R$ the endpoints of $\t$
to which is associated $\RR\VV^{(0)}$ and
$I_\n$ the end-points associated to $\LL\VV^{(h_v-1)}$.
, the explicit representation of $W_{\t,\PP,T}^{(h)}
(\xx_{v_0})$ in (\ref{2.49}) is
\bea &&
W_{\t,\PP, T}^{(h)}(\xx_{v_0}) =
\left[\prod_{v\in I_R}K_{v}^{(0)} (\xx_{v})\right]\prod_{v\in I_v} 2^{h_v}\n_{h_v}
\;\cdot\label{2.50aaa}\\ &&\cdot\;
\Bigg\{\prod_{v\,\atop\hbox{\ottorm not e.p.}}{1\over s_v!} \int
dP_{T_v}({\bf t}_v)\;{\rm det}\,\tilde G^{h_v,T_v}({\bf t}_v)\Biggl[
\prod_{l\in T_v}
\big[(\xx_l-\yy_l)^{\a_l}\partial^{\b_l}g^{(h_v)}(\xx_l-\yy_l)\big]\,\Biggr]
\Bigg\}\;,\nn\eea
where $K_{v}^{(0)} (\xx_{v})$ are the kernels of $\RR\VV^{(0)}$; the factors
$(\xx_l-\yy_l)^\a$ and the derivatives $\partial^\b$ in the above expression are produced by the $\RR$ operation and finally $\tilde G^{h_v,T_v}$ differs
from $G^{h_v,T_v}$ for the presence of extra derivatives due to the $\RR$ operation
(see \S 3 of \cite{BM} for more details in a similar case).
The functions
appearing in the r.h.s. of \pref{bbe}, namely
$b_{0,h}$, $b_{\pm,h}$, $b_{3,h}$ can be written as derivatives of
\be\bar W^{(h)}_l=\sum_{n=2}^\io\sum_{\t\in\bar \TT_{n,h}} \sum_{{\bf P}\in{\cal P}_\t}\sum_{T\in {\bf T}}{1\over |\L|\b}\int d\xx_{v_0} W_{\t,\PP, T}(\xx_{v_0})\label{bar}
\ee
where $\bar\TT_{n,h}$ is the subset of $\TT_{h,h}$ such that a)$h_{v^*}=h+1$
where $v^*$ is the first non trivial vertex; b)there is at least an end-point associated to $\VV^{(0)}$.
Condition a) is due to the fact that, by construction, $\LL\RR=0$; condition b) is due to the fact that
the contributions with only $\n$-vertices are vanishing, as it can be easily verified by an explicit computation in momentum space (they are chain graphs and $\hat g^{(k)}(0)=0$).
The next goal is the proof of the following result.\\
\begin{lemma}
There exists a constant $\e_0$ independent of
of $\b$, $L$ and $r$, such that for $|U|\le \e_0$
and $\max_{k\ge h}[|\n_k|, |Z_k-1|,|v_{k,i}-v_{0,i}]\le \e_0$, $i=\pm,3$ then for $h\ge h^*$
\be \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le
2^{h (7/2-5 l/4)} \,(C \e_0)^{max(1,l/2-1)}\;.\label{2.52sz}\ee
Moreover, if $\bar W^{(h)}_{l}$ is given by
\pref{bar}
\be \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|\bar W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le
2^{h (7/2-5 l/4)}2^{{1\over 8} h} \,(C \e_0)^{max(1,l/2-1)}\;.\label{2.52sz1}\ee
with $C$ a suitable constant.
\end{lemma}
\vskip.3cm
{\it Proof.}
Using the tree expansion described above
we find that the l.h.s. of (\ref{2.50aaa}) can be bounded from above
by
\bea && \sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t}\sum_{T\in{\bf T}}
\int\prod_{l\in T^*}
d(\xx_l-\yy_l) \left[\prod_{v\in I_R}K_{v}^{(0)} (\xx_{v})\right]\prod_{v\in I_v} 2^{h_v}|\n_{h_v}|
\cdot
\label{2.53}\\
&&\cdot\Bigg[\prod_{v\ {\rm not}\ {\rm e.p.}}{1\over s_v!}
\max_{{\bf t}_v}\big|{\rm det}\, G^{h_v,T_v}({\bf t}_v)\big|
\prod_{l\in T_v}
\big[|\xx_l-\yy_l|^{\a_l}|\partial^{\b_l}g^{(h_v)}(\xx_l-\yy_l)|\big]\,\Biggr]
\nn\eea
where $T^*$ is a tree graph
obtained from $T=\cup_vT_v$, by adding
in a suitable (obvious) way, for each endpoint $v_i^*$,
$i=1,\ldots,n$, one or more lines connecting the space-time points
belonging to $\xx_{v_i^*}$.
A standard application of Gram--Hadamard inequality, combined with
the dimensional bound on $g^{(h)}(\xx)$ given by (\ref{b}), implies that
\be |{\rm det} G^{h_v,T_v}({\bf t}_v)| \le
c^{\sum_{i=1}^{s_v}|P_{v_i}|-|P_v|-2(s_v-1)}\cdot\;
2^{{5\over 4}{h_v}
\left(\sum_{i=1}^{s_v}|P_{v_i}|-|P_v|-2(s_v-1)\right)}\;.\label{2.54}\ee
By the decay properties of $g^{(h)}(\xx)$ given by \pref{b}, it
also follows that
\be \prod_{v\ {\rm not}\ {\rm e.p.}}
{1\over s_v!}\int \prod_{l\in T_v} d(\xx_l-\yy_l)\,
||g^{(h_v)}(\xx_l-\yy_l)||\le c^n \prod_{v\ {\rm not}\ {\rm e.p.}}
{1\over s_v!} 2^{-h_v(s_v-1)}\;.\label{2.55}\ee
The bound on the kernels produced by the ultraviolet integration
implies that
\bea &&\int\prod_{l\in T^*\setminus\cup_v T_v}d(\xx_l-\yy_l)
\left[\prod_{v\in I_R}K_{v}^{(0)} (\xx_{v})\right]\prod_{v\in I_v} 2^{h_v}|\n_{h_v}|\nn\\
&&\le C^n \e_0^n \Big[\prod_{v\
{\rm e.p.}, |I_v|=2} 2^{h_{v'}(1+\d_v)} \Big] \;,\label{2.56}\eea
where $\d_v={1\over 2}$ if $v\in I_R$ and $|I_v|=2$
and $\d_v=0$ otherwise
; the factors
$\g^{(1+\d_v)h_{v'}}$ are due to fact that $\RR$ acts on the terms with $|I_v|=2$.
Therefore
the l.h.s. of (\ref{2.52sz}) can be bounded from above
by
\bea &&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=l}\sum_{T\in{\bf T}}
C^n \e_0^n \Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}
2^{{h_v}\left(\sum_{i=1}^{s_v}{5|P_{v_i}|\over 4}-{5|P_v|\over 4}-{7\over 2}(s_v-1)\right)}\Big]]\cdot\nn\\
&&\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-(h_v-h_{v'})z(P_v)}\Big] \Big[\prod_{v\
{\rm e.p.}, |I_v|=2} 2^{h_{v'}(1+\d_v)} \Big] \label{2.57aab}\eea
where
$z(P_v)={3\over 2}$ for $|P_v|=2$; the factor
$\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-(h_v-h_{v'})z(P_v)}$ takes into account the presence of the $\RR$ operation
on the vertices.
Once that the bound \pref{2.57aab} is obtained, we have to see if we can sum over the scales and the trees.
Let us define $n(v)=\sum_{i: v_i^*>v}\,1$ as the number of endpoints following
$v$ on $\t$ and $v'$ as
the vertex immediately preceding $v$ on $\t$. Recalling that $|I_v|$
is the number of field labels associated to the endpoints following $v$
on $\t$ (note that $|I_v|\ge 2 n(v)$) and using that
\bea && \sum_{v\ {\rm not}\ {\rm e.p.}}\Big[\big(\sum_{i=1}^{s_v}
|P_{v_i}|\big)-|P_v|\Big]=|I_{v_0}|-|P_{v_0}|\;,\nn\\
&&
\sum_{v\ {\rm not}\ {\rm e.p.}}(s_v-1)=n-1\label{2.58}\\
&& \sum_{v\ {\rm not}\ {\rm e.p.}}
(h_v-h)\Big[\big(\sum_{i=1}^{s_v}
|P_{v_i}|\big)-|P_v|\Big]=\sum_{v\ {\rm not}\ {\rm e.p.}}
(h_v-h_{v'})(|I_v|-|P_v|)\;\nn\\
&&\sum_{v\ {\rm not}\ {\rm e.p.}}(h_v-h)(s_v-1)=
\sum_{v\ {\rm not}\ {\rm e.p.}}(h_v-h_{v'})(n(v)-1)\;,\nn\eea
we find that \pref{2.57aab} can be bounded above by
\bea &&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=2l}\sum_{T\in{\bf T}}
C^n \e_0^n 2^{h({7\over 2}-{5\over 4}|P_{v_0}|+{7\over 4}|I_{v_0}|-{7\over 2}n)} \\
&&
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}
2^{(h_v-h_{v'})({7\over 2}-{5 |P_v|\over 4}+{5|I_v|\over 4}-{7\over 2}n(v)+z(P_v))}\Big]
\Big[\prod_{v\ {\rm e.p.}, |I_v|=2} 2^{ h_{v'}(1+\d_v)} \Big]\nn
\eea
Using the identities
\bea &&2^{h n}
\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{(h_v-h_{v'}
) n(v)}=\prod_{v\ {\rm e.p.}}
2^{h_{v'}}\;,\nn\\
&& \g^{h |I_{v_0}|}
\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{(h_v-h_{v'}) |I_v|}=\prod_{v\ {\rm e.p.}}
2^{h_{v'} |I_v|}\;,\label{2.60}\eea
we finally obtain
\bea&& \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le \sum_{n\ge 1}\sum_{\t\in {\cal
T}_{h,n}} \sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=l}\sum_{T\in{\bf
T}} C^n \e_0^n 2^{h({7\over 2}-{5 l\over 4})}
\label{2.61}\nn\\
&&\cdot \Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
\frac{1}{s_v!} 2^{-(h_v-h_{v'})(5{|P_v|\over 4}-{7\over
2}+z(P_v))}\Big]\Big[\prod_{v\ {\rm e.p.}} 2^{h_{v'}({5|I_v|/4-7/2})}
\Big]\\
&& \Big[\prod_{v\ {\rm e.p.}, |I_v|=2} 2^{h_{v'}(1+\d_v)} \Big]\nn
\eea
Note that, if $v$ is not an endpoint, $5{|P_v|\over 4}-{7\over
2}+z(P_v)
\ge {1\over 2}$ by the
definition of $\RR$.
Now, note that the number of terms in $\sum_{T\in {\bf T}}$ can be
bounded by $C^n\prod_{v\ {\rm not}\ {\rm e.p.}} s_v!$. Using also
that $5{|P_v|\over 4}-{7\over
2}+z(P_v)\ge 1/2$ and $|P_v|-3\ge|P_v|/4$, we find
that the l.h.s. of (\ref{2.52sz}) can be bounded as
\bea&& \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le 2^{h({7\over 2}-{5 l\over 4})}
\sum_{n\ge 1}C^n\e_0^n\sum_{\t\in {\cal T}_{h,n}}\cdot\label{2.61b}\\
&&\cdot\big(
\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-(h_v-h_{v'})/4}\big)
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=2l}\big(\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-|P_v|/8}\big)
\;.\nn \eea
The sum over $\PP$ can be bounded using the following combinatorial
inequality: let $\{p_v, v\in \t\}$,
with $\t\in\TT_{h,n}$, a set of integers such that
$p_v\le \sum_{i=1}^{s_v} p_{v_i}$ for all $v\in\t$ which are not endpoints;
then $\prod_{\rm v\;not\; e.p.} \sum_{p_v}
2^{-p_v/8}\le C^n$.
Finally
$$\sum_{\t\in {\cal T}_{h,n}}
\prod_{v\ {\rm not}\ {\rm e.p.}}2^{{1\over 2}(h_v-h_{v.})}
\le C^n\;,$$
as it follows by the fact that the number of non trivial vertices in $\t$
is smaller than $n-1$ and that the number of trees in ${\cal T}_{h,n}$ is
bounded by ${\rm const}^n$, and collecting all the previous bounds, we obtain
\pref{2.52sz}.
In order to derive \pref{2.52sz1} we note that, for any tree with no $\n$ end-points
\be
\Big[\prod_{v\ {\rm e.p.}}
2^{h_{v'}(5|I_v|/4-7/2)}
\Big] \Big[\prod_{v\ {\rm e.p.}, |I_v|=2} 2^{h_{v'}(1+\d_v)} \Big]\le
\prod_{v\ {\rm e.p.}} 2^{{1\over 2}h_{v'}}
\ee
so that we can replace \pref{2.61} by
\bea&& \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le \sum_{n\ge 1}\sum_{\t\in {\cal
T}_{h,n}} \sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=l}\sum_{T\in{\bf
T}} C^n 2^{h({7\over 2}-{5 l\over 4})}
C^n \e_0^n\cdot\label{2.61a}\nn\\
&&\hskip.2truecm \cdot \Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
\frac{1}{s_v!} 2^{-{1\over 2}(h_v-h_{v'})(5{|P_v|\over 4}-{7\over
2}+z(P_v))}\Big]
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-{1\over 4}(h_v-h_{v'})}\Big]
\prod_{v\ {\rm e.p.}} 2^{{1\over 2}h_{v'}}
\eea
and
\be
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
2^{-{1\over 4}(h_v-h_{v'})}\Big]
\prod_{v\ {\rm e.p.}} 2^{{1\over 2}h_{v'}}\le 2^{h\over 8}
\ee
This concludes the proof of Lemma 1.
\qed
\subsection{The flow of the effective parameters}
The previous lemma provides convergence of the renormalized expansion
provided that the effective parameters remain close $O(U)$ to their initial value and $U$ is chosen
small enough.
The flow equation for $\n_h$ can be written as
\be
\n_{h-1}=2^h \n_h+\b^{(h)}_\n
\ee
with, from \pref{2.52sz1}, $\b^{(h)}_\n=O(U 2^{h\over 8})$. By iteration we get
\be
\n_h=2^{-h+1}[\n+\sum_{k=h+1}^1 2^{k-2}\b^{(k)}_\n]\ee
If we choose $\n$ so that
\be
\n=-\sum_{k=h^*+1}^1 2^{k-2}\b^{(h)}_\n+2^{h^*-1}\n_{h^*}
\ee
then
\be
\n_h=-2^{-h}\sum_{k=h^*+1}^h 2^{k-2}\b^{(h)}_\n+2^{h^*-h}\n_{h^*}\label{11}
\ee
By a fixed point argument one can prove, see for instance Lemma 4.2 of \cite{BM}, that it is possible
to find a sequence of $\n_h$ solving \pref{11}. Moreover, $v_{i,h-1}-v_{i,h}=O(U 2^{h/8})$,
${Z_{h-1}\over Z_h}=1+O(U 2^{h/8})$, from
\pref{2.52sz1}, so that $v_{i,h-1}-v_{i,0}=O(U)$,
$Z_{h-1}=1+O(U)$.
\section{Renormalization Group integration: the second regime}
\subsection{Tree expansion and convergence}
While the analysis of the scales grater than $h^*$ are insensitive to the sign of $r$, the integration of the
scales smaller than $h^*$ depends on it. The case $r<0$
corresponds to the insulating phase; all the scales $\le h^*$ can be integrated in a single step (setting $\n_{h^*}=0$)
as the
propagator of $\psi^{(\le h^*)}$ has the same asymptotic behavior of the single scale propagator for $h\ge h^*$, that is
\be
|g^{(\le h^*)}(\xx)|\le {1\over Z_{h^*}}
{2^{5h^*\over 2}\over 1+[2^{h^*} (|x_0|+|x_+|+|x_-|)+2^{h^*\over 2}|x_3|]^N}
\label{baa}
\ee
Similarly the case $r=0$ correspond to the case $h^*=-\io$ and it can be analyzed as in the previous section.
Let us consider now the case $r>0$, corresponding to the metallic phase:
the Fourier transform of
the
propagator vanishes now in correspondence of two Fermi momenta and
we need a multiscale decomposition. We write
\be
\hat g^{(\le h^*)}(\kk)=\bar g^{(h^*)}(\kk)+\hat g^{(< h^*)}(\kk)
\ee
where $\hat g^{(< h^*)}(\kk)$ is equal to $\hat g^{(\le h^*)}(\kk)$ with
$\chi_{h^*-1}(\kk)$ replaced by
\be
\sum_{\o=\pm }\theta(\o k_3) \bar\chi_{h^*-1}(\kk)\label{sasas}
\ee
and $\chi_{h^*-1}(\kk)
=\bar \chi(
b_0 2^{-h^*}|\det A_{h^*}(k)|^{1\over 2})$
with
$b_0$ chosen so that $\chi_{h^*-1}(\kk)$
has support in two disconnected regions around $\pm \pp_F$.
The propagator $\bar g^{(h^*)}(\kk)$, with support in $\chi_{h^*}-\chi_{h^*-1}$, verifies the same bound as \pref{b} with $h=h^*$; in fact
the denominator of $\hat g^{(h^*)}(\kk)$ is $O(r)$; moreover, if $k=k'+\o p_F$,
one has $\cos(k'+\o p_F)-\cos p_F=v_{3,0} k'+{1\over 2} k'^2+O(k'^3)$,
with $v_{3,0}=O(\sqrt{r})$ for small $r$.
Therefore each derivative with respect to $k'$ produces an extra $r^{-{1\over 2}}$. We can decompose the Grassmann variables
as
\be
\psi^{\pm(\le h^*)}_{\xx}=\psi^{\pm(h^*)}_\xx+\sum_{\o=\pm} e^{\pm i\o \pp_F \xx}\psi^{\pm(<h^*)}_{\o,\xx}
\ee
where $\psi^{\pm(<h^*)}_{\xx}$ has propagator
\be
g^{(<h^*)}_{\o}(\xx)=\int d\xx e^{i\kk'\xx} \hat g^{(< h^*)}(\kk'+\o\pp_F)
\ee
We can therefore integrate
$\psi^{(h^*)}$ so that
\bea
&&e^{|\L|\b E_{h^*}}\int P(d\psi^{(h^*)})\int \prod_{\o=\pm} P(d\psi_\o^{(< h^*)})e^{\VV^{(h^*)}(Z_{h^*}\psi^{(\le h^*)})}=\\
&&\int \prod_{\o=\pm} P(d\psi_\o^{(< h^*)})e^{\VV^{(h^*-1)}(Z_{h^*-1}\psi^{(< h^*)})}\nn
\eea
where
\be \VV^{(h^*-1)}(\psi)=\sum_{n\ge 1}
\int d\xx_1...\int d\xx_n
W^{(h)}_{n}(\underline\xx)[\prod_{i=1}^n e^{i\e_i\o_i\pp_F\xx_i}\psi^{\e_i(\le h)}_{\o_i,\xx_i}]
\label{coco}\ee
and $W^{(h)}_{n}(\underline\xx)$ is translation invariant.
We describe the integration of the scales $h< h^*$ inductively. Assume that we have integrate the scale $h^*,-,..,h+1$
showing that \pref{gf} can be written as
\be
e^{|\L| \b E_h}\int \prod_{\o=\pm}P(d\psi_\o^{(\le h)}) e^{\tilde \VV^{(h)}(\sqrt{Z_{h}}\psi^{(\le h)})}\label{ef1}
\ee
where $P(d\psi^{(\le h)})$ has propagator given by
\bea
&&g^{(\le h)}(\xx)_\o=\\
&&\int d\kk e^{i\kk\xx}{\chi_h(\kk)\over Z_h}
\begin{pmatrix}&-i k_0+\o v_{3,h}\sin k'_3+E'(\kk)&
v_{h}(\sin k_+-i\sin k_-)\\
&
v_{h}(\sin k_++i \sin k_-)&
-i k_0-\o v_{3,h}\sin k'_3- E(kk)
\end{pmatrix}^{-1}\label{cond2}\nn
\eea
where $E'(\kk)=\cos p_F (\cos k_3-1)+E(\vec k)$.
and $V^{(h)}(\psi)$ is similar to \pref{coco}
We introduce a {\it localization operator} acting on the effective potential as in \pref{loc}
acting on the kernels
$\hat W^{(h)}_{n}(\kk_1,..,\kk_{n-1})$ in the following way:
\begin{enumerate}
\item $\LL \hat W^{(h)}_{n}(\kk_1,..,\kk_{n-1})=0$ if $n>2$.
\item If $n=2$
\be
\LL \hat W^{(h)}_{2}(\kk)= \hat W^{(h)}_{2}(\o\pp_F)+k_0
\hat W^{(h)}_{2}(\o\pp_F)+\sum_{i=+,-,3} \sin k'_i
\partial_i \hat W^{(h)}_{2}(\o\pp_F)
\ee
\end{enumerate}
Note that, by symmetry
\bea
&&\hat W^{(h)}_{2}(\o p_F)=\s_3 n_h\quad \partial_+ \hat W^{(h)}_{2}(0)=\s_1 b_{+,h}\nn\\
&&
\partial_- \hat W^{(h)}_{2}(\o p_F)=\s_2 b_{-,h}\quad \partial_3\hat W^{(h)}_{2}(\o p_F)=\s_3 b_{3,h}
\eea
We can include the quadratic part in the free integration; the single scale propagator verifies the following bound
\bea
&&|g^h_\o(\xx)|\le {1\over v_{3,h}}{2^{3h}\over 1+2^h (|
x_0|+|x_+|+|x_-|)+v_{3,h}^{-1} |x|)^N}\nn\\
&&
\int d\xx |g^h(\xx)|\le C 2^h\quad\quad \max |g^h(\xx)|\le {2^h\over v_{3,h}}
\label{x1}\eea
and $v_{3,h^*}=O(\sqrt{r})$. Note also that
\be
g^{(h)}_\o(\xx)=g^{(h)}_{rel,\o}(\xx)+r^{(h)}_{\o}(\xx)\label{dec}
\ee
where
\bea
&&g^{(\le h)}(\xx)_\o=\\
&&\int d\kk e^{i\kk\xx}{\chi_h(\kk)\over Z_h}
\begin{pmatrix}&-i k_0+\o v_{3,h}\sin k'_3&
v_{h}(\sin k_+-i\sin k_-)\\
&
v_{h}(\sin k_++i \sin k_-)&
-i k_0-\o v_{3,h}\sin k'_3-
\end{pmatrix}^{-1}\label{cond2}\nn
\eea
and $r^{(h)}$ verifies a similar bound with an extra $2^h$.
\begin{lemma}
If $r>0$ there exists a constant $\e_0$ independent of
of $\b$, $L$ and $r$, such that for $|U|\le \e_0$
and $\max_{k\ge h}[|\n_k|, |Z_k-1|,|{v_{k,i}\over v_{0,i}}-1|\le \e_0$, $i=\pm,3$ then for $h\le h^*$
\be \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le
2^{h (4-3 l/2)} \,(C \e_0)^{max(1,l/2-1)}\;.\label{2.52sza}\ee
and, if $\bar W^{(h)}_{l}$ is given by
\pref{bar}
\be \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|\bar W^{(h)}_{l}
(\xx_1,\ldots,\xx_{l})|\le
2^{h (4-3 l/2)}2^{{1\over 8} (h-h^*)} \,(C\e_0)^{max(1,l/2-1)}\;.\label{2.52sz1a}\ee
with $C$ a suitable constant.
\end{lemma}
\vskip.3cm
{\it Proof.}
Again the effective potential can be written as a sum over trees similar to the previous ones
..but with some modifications ( see Fig 2).
\insertplot{300}{150}{ \ins{30pt}{85pt}{$r$} \ins{50pt}{85pt}{$v_0$}
\ins{130pt}{100pt}{$v$} \ins{35pt}{-5pt}{$h$} \ins{52pt}{-5pt}{$h+1$}
\ins{135pt}{-5pt}{$h_{v}$} \ins{215pt}{-5pt}{$-1$} \ins{235pt}{-5pt}{$h^*$}
\ins{255pt}{-5pt}{$h^*$+1}} {fig50} {A renormalized tree for
$\VV^{(h)}$ in the second regime\lb{h2aa}}{0}
The scales are
$\le h^*$ and 5) in the previous definition is replaced by:
\vskip.3cm
\0 5') With each endpoint $v$ we associate one of the monomials
with four or more Grassmann fields contributing to $\RR {\cal
V}^{(h^*)}(\psi^{(\le h_v-1)})$ and a set $\xx_v$ of space-time
points (the corresponding integration variables in the $\xx$-space
representation); or a term corresponding to $\LL\VV^{(h_v-1)}$.
In terms of these trees, the effective potential ${\cal V}^{(h)}$, $h\le -1$
is defined as
i) if $s>1$, then
\be {\cal V}^{(h)}(\t,\psi^{(\le h)})={(-1)^{s+1}\over s!} \EE^T_{h+1}
\big[\bar{\cal V}^{(h+1)}(\t_1,\psi^{(\le h+1)});\ldots; \bar{\cal V}^{(h+1)}
(\t_{s},\psi^{(\le h+1)})\big]\;,\label{2.42}\ee
where $\bar{\cal V}^{(h+1)}(\t_i,\Psi^{(\le h+1)})$ is equal to $\RR{\cal
V}^{(h+1)}(\t_i,\psi^{(\le h+1)})$ if the subtree $\t_i$ contains
more than one end-point, or if it contains one end-point but it is
not a trivial subtree;
it is equal to $\RR{\cal V}^{(h^*)}(\t_i,\Psi^{(\le h+1)})$
or $\g^{h+1}\n_{h+1} F_\n(\psi^{(\le h+1)})$
if $\t_i$ is a trivial subtree;\\
ii) if $s=1$, then ${\bar V}^{(h+1)}(\t,\psi^{(\le h)})$ is equal to
$\big[\RR{\cal V}^{(h+1)}(\t_1,\psi^{(\le h+1)})\big]$
if $\t_1$ is not a trivial
subtree; it is equal to $\big[\RR {\cal
V}^{(h^*)}(\psi^{(\le h+1)})- \RR{\cal V}^{(h^*)}(\psi^{(\le h)})\big]$ or
if $\t_1$ is a trivial subtree.
As before, we
we get
\be {\cal V}^{(h)}(\t,\PP) = \sum_{T\in {\bf T}} \int d\xx_{v_0}
\widetilde\psi^{(\le h)}(P_{v_0}) W_{\t,\PP,T}^{(h)}(\xx_{v_0})
\= \sum_{T\in {\bf T}}
{\cal V}^{(h)}(\t,\PP,T)\;,\label{2.49}\ee
where,
given $\t\in\TT_{h,n}$ and the labels $\PP,T$,
calling $I_R$ the endpoints of $\t$
to which is associated $\RR\VV^{(h^*)}$ and
$I_\n$ the end-points associated to $\LL\VV^{(h_v-1)}$, the explicit representation of $W_{\t,\PP,T}^{(h)}
(\xx_{v_0})$ in (\ref{2.49}) is
\bea &&
W_{\t,\PP, T}(\xx_{v_0}) =\left[\prod_{v\in I_R}
K_{v}^{(h^*)} (\xx_{v})\right]\prod_{v\in I_v}\g^{h_v}\n_{h_v};
\Bigg\{\prod_{v\,\atop\hbox{\ottorm not e.p.}}{1\over s_v!} \int
dP_{T_v}({\bf t}_v)\nn\\
&&\;{\rm det}\, \tilde G^{h_v,T_v}({\bf t}_v)
\Biggl[
\prod_{l\in T_v} \d_{\o^-_l,\o^+_l}\,
\big[(\xx_l-\yy_l)^{\a_l}\partial^{\b_l}g^{(h_v)}_{\o_l}(\xx_l-\yy_l)\big]\,\Biggr]
\Bigg\}\;\eea
where $K_{v}^{(h^*)} (\xx_{v})$ are the kernels of $\RR\VV^{(h^*)}$.
By using the bounds obtained in the previous regime \pref{2.52sz}
\bea
&&\int\prod_{l\in T^*\setminus\cup_v T_v}d(\xx_l-\yy_l)
\left[\prod_{v\in I_R}K_{v}^{(h^*)} (\xx_{v})\right]\prod_{v\in I_v} 2^{h_v}|\n_{h_v}|
\le\nn\\
&& C^n\e_0^n
\Big[\prod_{v\ {\rm
e.p.}; |I_v|=2} 2^{h_{v'}+\d_v(h_{v'}-2 h^*))}\Big]
\Big[\prod_{v\ {\rm
e.p.}\in I_R, |I_v|\ge 4}2^{h^*({7\over 2}-{5|I_v|\over 4})}\Big]
\eea
where $\d_v=1$ if $v\in I_R$ (again if $v\in I_R$ the factor $2^{2(h_{v'}- h^*)}$ comes from the definition of $\RR$) .
Therefore
\bea && \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_l)|\le C^n \e_0^n\\
&&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=
l}\sum_{T\in{\bf T}}
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}\nn\\
&&[2^{h_v}]^{\left(\sum_{i=1}^{s_v}{3|P_{v_i}|\over 2}-3{|P_v|\over 2}-4(s_v-1)\right)}
[{1\over v_{3,0}}]^{\left(\sum_{i=1}^{s_v}{|P_{v_i}|\over 2}-{|P_v|\over 2}-(s_v-1)\right)}\Big]\Big]\nn
\\
&&\Big[\prod_{v\ {\rm
e.p.}; |I_v|=2} 2^{h_{v'}+\d_v(h_{v'}-2 h^*))}\Big]
\Big[\prod_{v\ {\rm
e.p.}\in I_R, |I_v|\ge 4}2^{h^*({7\over 2}-{5|I_v|\over 4})} \Big]\nn
\label{2.57a}\nn\eea
and
by using (67)
\bea && \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_l)|\le C^n \e_0^n\nn\\
&&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=
l}\sum_{T\in{\bf T}}
2^{h({4-{3\over 2}|P_{v_0}|+{3\over 2}|I_{v_0}|-4 n)}}\nn\\
&&
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}
2^{(h_v-h_{v'})(4-{3|P_v|\over 2}+{3|I_v|\over 2}- 4 n(v))}\Big]\\
&&\Big[\prod_{v\ {\rm
e.p.}; |I_v|=2} 2^{h_{v'}(1+\d_v(h_{v'}-2 h^*))}\Big]
\Big[\prod_{v\ {\rm
e.p.}\in I_R, |I_v|\ge 4}2^{h^*({7\over 2}-{5|I_v|\over 4})} \Big]\nn\\
&&\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
[{1\over v_{3,0}}]^{\left(\sum_{i=1}^{s_v}{|P_{v_i}|\over 2}-{|P_v|\over 2}-(s_v-1)\right)}\Big]]\nn
\eea
and finally using \pref{2.60}
\bea && \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_l)|\le\\
&&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=l}\sum_{T\in{\bf T}}
C^n \e_0^n 2^{h({4-{3\over 2}|P_{v_0}|)}}\nn\\
&&
\Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}
2^{(h_v-h_{v'})(4-{3|P_v|\over 2})}\Big]\Big[\prod_{v\ {\rm
e.p.}; |I_v|=2, v\in I_R} 2^{2(h_{v'}-h^*) }\Big]\nn\\&&
\Big[\prod_{v\ {\rm
e.p.}; v\in I^{R},|I_v|\ge 4} 2^{h_{v'}(-4+{3|I_v|\over 2})}
2^{h^*({7\over 2}-{5|I_v|\over 4})} ] \nn\\
&&\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
[{1\over v_{3,0}}]^{\left(\sum_{i=1}^{s_v}{|P_{v_i}|\over 2}-{|P_v|\over 2}-
(s_v-1)\right)}\Big]]\nn
\eea
By writing
\bea
&&\Big[\prod_{v\ {\rm
e.p.}; v\in I^{R}} 2^{h_{v'}(-4+{3|I_v|\over 2})}
2^{h^*({7\over 2}-{5|I_v|\over 4})} ] =\nn\\
&&\Big[\prod_{v\ {\rm
e.p.}; v\in I^{R}} 2^{(h_{v'}-h^*)(-4+{3|I_v|\over 2})}
2^{h^*(-{1\over 2}+{|I_v|\over 4})} ]
\eea
and using that $2^{h^*(-{1\over 2}+{|I_v|\over 4})}\le C (v_{3,0})
^{-1+{|I_v|\over 2}})$
we get
\bea && \frac1{\b |\L|}\int d\xx_1\cdots d\xx_{l}|W^{(h)}_{l}
(\xx_1,\ldots,\xx_l)|\le C^n \e_0^n\\
&&\sum_{n\ge 1}\sum_{\t\in {\cal T}_{h,n}}
\sum_{\PP\in{\cal P}_\t\atop |P_{v_0}|=l}\sum_{T\in{\bf T}}
C^n \e_0^n \g^{h({4-{3\over 2}|P_{v_0}|)}}\nn\\
&&\Big[\prod_{v\ {\rm not}\ {\rm e.p.}} \frac{1}{s_v!}
2^{(h_v-h_{v'})(4-{3|P_v|\over 2})}\Big]
\Big[\prod_{v\ {\rm
e.p.}; |I_v|=2, v\in I_R} 2^{2(h_{v'}-h^*) }\Big]\nn\\
&&\Big[\prod_{v\ {\rm not}\ {\rm e.p.}}
[{1\over v_{3,0}}]^{\left(\sum_{i=1}^{s_v}{|P_{v_i}|\over 2}-{|P_v|\over 2}-
(s_v-1)\right)}\Big]]\nn\\
&&\Big[\prod_{v\ {\rm
e.p.}, v\in I_R}2^{-(h_{v'}-h^*)(4-{3|I_v|\over 2})} \Big]
\Big[\prod_{v\ {\rm
e.p.}, v\in I_R}
v_{3,0}^{-1+{|I_v|\over 2}}
\Big]
\label{82}\nn
\eea
Using that
\be
\Big[\prod_{v\ {\rm
e.p.}, v\in I^R} v_{3,0}^{-1+{|I_v|\over 2})}\Big]=\Big[\prod_{v\ {\rm
e.p.}} v_{3,0}^{-1+{|I_v|\over 2})}\Big]
\ee
which follows from the fact that for $v\in I^\n$ one has $|I_v|=2$ so that $
v_{3,0}^{-1+{|I_v|\over 2}}=1$, we can write
\be
\Big[\prod_{v\ {\rm
e.p.}} v_{3,0}^{-1+{|I_v|\over 2})}\Big]\le v_{3,0}^{-n+\sum_{v e.p }|I_v|/2}
\ee
and using $\sum (s_v-1)=n-1$ where $n$ is the number of
end-points we get
\be
\prod_{v\ {\rm
e.p.},} v_{3,0}^{-1}
\prod_{v\ {\rm not}\ {\rm e.p.}} [{1\over v_{3,0}}]^{-
(s_v-1)}\le C v_{3,0}^{-n}
v_{3,0}^{n-1}\le C v_{3,0}^{-1}\ee
Moreover $\sum_{v e.p }|I_v|=l+\sum_v[\sum_{i=1}^{s_v}|P_{v_i}|
-|P_v|]$
\be
\Big[\prod_{v\ {\rm
e.p.}} (v_{3,0})^{|I_v|\over 2}\Big]
\prod_{v\ {\rm not}\ {\rm e.p.}} [{1\over
v_{3,0}}]^{\left(\sum_{i=1}^{s_v}{|P_{v_i}|\over 2}-{|P_v|\over 2}\right)}\Big]\le C
(v_{3,0})^{l/2}\ee
so that in total we get $v_{3,0}^{l/2-1}$ in agreement with
\pref{2.52sza}. Note that the small divisors proportional to $v_{3,0}^{-1}$, which could in principle spoil convergence, are exactly compensated from the factors due to the different scaling of the two regions.
\qed
The flow of the effective coupling can be analyzed as before, noting
that the beta function is $O(U 2^{h-h^*})$ by the above estimate and we get
\bea
&&Z_{h}\to_{h\to-\io} Z=1+O(U^2)\label{aa}\\
&&
v_{3,h}\to_{h\to-\io} v_{3}=t_\perp\sin(p_F)+a_{3} U+O(U^2)
\nn\\
&&v_{\pm, h-1}\to_{h\to-\io} v_{\pm}=t+a_\pm U+O(U^2)\nn\\
\eea
where
\be
a_{3}\s_3=\int d\kk \hat v(\kk) \partial_3 \hat g (\kk)\quad a_{+}\s_1=\int d\kk \hat v(\kk) \partial_+ \hat g(\kk)
\ee
Moreover
\be
\n=U v(0) \hat S_0(0,0^-)+\int d\kk v(\kk) \hat g(\kk)
\ee
and this concludes the proof of Lemma 2.\qed
From Lemma 1 and Lemma 2 the proof of the main theorem follows easily.
|
\section*{REFERENCES}}
\newcommand{{\mathcal E}}{{\mathcal E}}
\newcommand{{\Phi_1}}{{\Phi_1}}
\newcommand{{\Phi_0}}{{\Phi_0}}
\newcommand{{\mathbb I}}{{\mathbb I}}
\newcommand{{\mathcal F}}{{\mathcal F}}
\newcommand{{\mathcal G}}{{\mathcal G}}
\newcommand{{\mathbb C}}{{\mathbb C}}
\newcommand{{\mathbb E}}{{\mathbb E}}
\newcommand{{\mathcal D}}{{\mathcal D}}
\newcommand{{\mathcal L}}{{\mathcal L}}
\newcommand{{\mathcal P}}{{\mathcal P}}
\newcommand{{\mathcal S}}{{\mathcal S}}
\newcommand{{\mathcal K}}{{\mathcal K}}
\newcommand{{\mathfrak D}}{{\mathfrak D}}
\newcommand{{\mathfrak H}}{{\mathfrak H}}
\newcommand{{\mathfrak M}}{{\mathfrak M}}
\newcommand{{\mathcal M}}{{\mathcal M}}
\newcommand{{\mathcal C}}{{\mathcal C}}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{{\mathbb D}}{{\mathbb D}}
\newcommand{{\mD}}{{{\mathbb D}}}
\newcommand{{\pi^o\hspace*{-4.5pt}}}{{\pi^o\hspace*{-4.5pt}}}
\newcommand{{\Pi^o\hspace*{-4.5pt}}}{{\Pi^o\hspace*{-4.5pt}}}
\newcommand{{p^o\hspace*{-4.5pt}}}{{p^o\hspace*{-4.5pt}}}
\newcommand{{\Pi_x}}{{\Pi_x}}
\newcommand{{\Pi_y}}{{\Pi_y}}
\newcommand{{\frac12}}{{\frac12}}
\newcommand{\operatorname{\rm trace}}{\operatorname{\rm trace}}
\newcommand{\mathop{\mathrm{diag}}}{\mathop{\mathrm{diag}}}
\newcommand{\operatorname{\rm argmax}}{\operatorname{\rm argmax}}
\newcommand{{\hat\phi}}{{\hat\phi}}
\newcommand{\|}{\|}
\newcommand{{\tilde \sigma}}{{\tilde \sigma}}
\newcommand{{\tilde x}}{{\tilde x}}
\newcommand{{\rm dist}}{{\rm dist}}
\newcommand{{\chi}}{{\chi}}
\newcommand{{x_0}}{{x_0}}
\newcommand{{\hat\phi_0}}{{\hat\phi_0}}
\newcommand{{\hat\phi_1}}{{\hat\phi_1}}
\newcommand{{\mathbf \Pi}}{{\mathbf \Pi}}
\newcommand{{{\mathcal E}\hspace*{-6.5pt}{\mathcal E}}}{{{\mathcal E}\hspace*{-6.5pt}{\mathcal E}}}
\newcommand{{\rm Supp}}{{\rm Supp}}
\definecolor{llgrey}{rgb}{0.9,0.9,0.9}
\definecolor{lgrey}{rgb}{0.6,0.6,0.6}
\definecolor{lred}{rgb}{0.9,0.7,0.7}
\newtheorem{thm}{Theorem}
\newtheorem{conj}{Conjecture}
\newtheorem{cor}{Corollary}
\newtheorem{lemma}{Lemma}
\newtheorem{prop}{Proposition}
\newtheorem{problem}{Problem}
\newtheorem{remark}{Remark}
\newtheorem{defn}{Definition}
\newtheorem{ex}{Example}
\setlength{\parindent}{20pt}
\parskip 4.5pt
\newcommand{\color{magenta}}{\color{magenta}}
\begin{document}
\title{Positive contraction mappings for classical and quantum Schr\"{o}dinger systems}
\author{Tryphon T. Georgiou}\email{<EMAIL>}\affiliation{University of Minnesota}
\author{Michele Pavon}\email{<EMAIL>}\affiliation{University of Padova}
\begin{abstract}
{\bf Abstract.} The classical Schr\"{o}dinger bridge seeks the most likely probability law for a diffusion process, in path space, that matches marginals at two end points in time; the likelihood is quantified by the relative entropy between the sought law and a prior, and the law dictates a controlled path that abides by the specified marginals.
Schr\"odinger proved that the optimal steering of the density between the two end points is effected by a multiplicative functional transformation of the prior; this transformation represents an automorphism on the space of probability measures and has since been studied by Fortet, Beurling and others.
A similar question can be raised for processes evolving in a discrete time and space as well as for processes defined over non-commutative probability spaces.
The present paper builds on earlier work by Pavon and Ticozzi and begins with the problem of steering a Markov chain between given marginals. Our approach is based on the Hilbert metric and leads to an alternative proof which, however, is constructive. More specifically, we show that the solution to the Schr\"{o}dinger bridge is provided by the fixed point of a contractive map.
We approach in a similar manner the steering of a quantum system across a quantum channel. We are able to establish existence of quantum transitions that are multiplicative {functional} transformations of a given Kraus map, but only for the case of uniform marginals.
As in the Markov chain case, and for uniform density matrices, the solution of the quantum bridge can be constructed from the fixed point of a certain contractive map.
For arbitrary marginal densities, extensive numerical simulations indicate that iteration of a similar map leads to fixed points from which we can construct a quantum bridge. For this general case, however, a proof of convergence remains elusive.
\end{abstract}
\keywords{Schr\"odinger systems, Schr\"odinger bridge, quantum Schr\"{o}dinger bridge, quantum control, quantum channel}
\maketitle
\section{Introduction}
In 1931 Erwin Schr\"odinger published a manuscript on ``the reversal of the laws of nature'' (``\"Uber die Umkehrung der Naturgesetze''). In it, he raised the following ``new and unorthodox'' \cite{beurling1960automorphism} question regarding Brownian motion. Suppose that the density of Brownian particles is observed at two points in time, $t_0$ and $t_1$, and that the two end-point densities differ from the initial and final marginals of the prior path-space distribution. Schr\"odinger then asked for the most likely random evolution that the particles have taken so as to reconcile the observed ``improbable but still possible outcome.'' In modern probabilistic language, as observed by F\"{o}llmer some fifty years later \cite{foellmer1988randomfields}, Schr\"{o}dinger was posing (and, to some extent, solving) a problem of large deviations of the empirical distribution. He was working in an abstract setting, although the very foundations of probability theory were still missing! The solution of the large deviations problem requires, in view of Sanov's theorem \cite{sanov1957largedeviations}, solving a maximum entropy problem.
Schr\"{o}dinger's 1931/32 papers were followed soon afterwards by works by N. Kolmogoroff on ``the reversibility of the statistical laws of nature'' (``Zur Umkehrbarkeit der statistischen Naturgesetze'') and by Fortet, Beurling and many others on the mathematical issues that
Schr\"odinger's paper raised.
In the present paper, following Pavon and Ticozzi \cite{pavon2010discrete,ticozzi2010time},
we consider discrete-time and discrete-space {classical evolutions} as well as quantum Markovian evolutions. More precisely, we first consider {discrete random vectors} with given prior distribution and endpoint marginals. We derive a constructive proof of the existence of multiplicative {functional} transformations of the prior {initial-final time joint distribution} that allows connecting the given end-point marginals. The unique solution is in fact the closest law to the prior in a relative entropy sense amongst all probability laws that are in agreement with the two marginals. A key concept is that of the Hilbert metric --this is a metric which is suitable for quantifying distances in homogeneous positive spaces. A similar approach allows a constructive proof for matching uniform marginal density matrices via a multiplicative {functional} transformation of any given prior Kraus map. In essence, this result extends to the non-commutative case a result of Sinkhorn \cite{sinkhorn1974diagonal,Sinkhorn1964} that any strictly positive stochastic matrix can be transformed into a doubly stochastic matrix via a multiplicative {functional} transformation. Thus, for the quantum case, we establish that any strictly positive Kraus map can be transformed into a doubly stochastic quantum map via a multiplicative {functional} transformation.
{Further, extensive simulations have convinced the authors that the approach works in complete generality, i.e., when the specified marginal densities are not necessarily uniform. However, a rigorous proof as well as a variational principle, in analogy with the classical case, is not available at present.}
The paper is structured as follows. {Section \ref{sec:hilbert} provides an exposition of the Hilbert metric. The corresponding geometry is key in studying the Schr\"odinger bridge in the classical and in the quantum case, in Sections \ref{sec:classical} and \ref{sec:quantum}, respectively. More specifically,
Sections \ref{discreterandom}-\ref{reduction} explain Schr\"odinger's bridges for Markov chains. Then, in Section \ref{solonestep} the Hilbert metric is used to provide a constructive solution to Schr\"odinger's bridge problem for Markov chains. Section \ref{sec:quantumchannels} overviews the formalism of quantum mechanics followed by a description of a quantum analog of the Schr\"odinger bridge problem in Section \ref{sec:quantumbridge}. Section \ref{sec:doublystochastic} presents a solution of the quantum Schr\"odinger bridge for the special case of uniform marginals. The mathematical statement for this special case represents a generalization of a result of Sinkhorn on the existence of doubly stochastic maps to a corresponding quantum probabilistic analog. This result is followed by a discussion and a conjecture about the general quantum Schr\"odinger problem, namely, that a fixed point of a certain map which is used to construct doubly stochastic maps, suitably modified, has a fixed point for general marginal density matrices as well.
}
\section{The Hilbert metric}\label{sec:hilbert}
This metric was introduced by David Hilbert in 1895 \cite{hilbert1895gerade} while exploring the foundations of geometry. Earlier special cases, its importance and several subsequent developement are being discussed by Bushell \cite{bushell1973hilbert}. More recently, the underlying geometry has proven timely on a range of problems in the study of communication and computations over networks (see \cite{tsitsiklis1986distributed} and in particular the work of Sepulchre and collaborators \cite{sepulchre2010consensus,sepulchre2011contraction} on consensus in non-commutative spaces, as well as the references therein) {and in quantum information theory \cite{reeb2011hilbert}}. A recent survey on the applications in analysis is \cite{lemmens2013birkhoff}. A key result that enables the metric to be used for establishing existence of solutions to various equations was {proved} by Garrett Birkhoff in 1957 \cite{birkhoff1957extensions}. Following \cite{sepulchre2010consensus}, we highlight the basic elements of the theory.
Let ${\mathcal S}$ be a real Banach space and let ${\mathcal K}$ be a closed solid cone in ${\mathcal S}$, i.e., ${\mathcal K}$ is closed with nonempty interior and is such that ${\mathcal K}+{\mathcal K}\subseteq {\mathcal K}$, ${\mathcal K}\cap -{\mathcal K}=\{0\}$ as well as $\lambda {\mathcal K}\subseteq {\mathcal K}$ for all $\lambda\geq 0$. Define the partial order
\[
x\preceq y \Leftrightarrow y-x\in{\mathcal K},
\]
and for $x,y\in{\mathcal K}\backslash \{0\}$, define
\begin{eqnarray*}
M(x,y)&:=&\inf\, \{\lambda\,\mid x\preceq \lambda y\}\\
m(x,y)&:=&\sup \{\lambda \mid \lambda y\preceq x \}.
\end{eqnarray*}
Then, the Hilbert metric is defined on ${\mathcal K}\backslash\{0\}$ by
\[
d_H(x,y):=\log\left(\frac{M(x,y)}{m(x,y)}\right).
\]
Strictly speaking, it is a {\em projective} metric since it remains invariant under scaling by positive constants, i.e.,
$d_H(x,y)=d_H(\lambda x,y)=d_H(x,\lambda y)$ for any $\lambda>0$ and, thus, it actually measures distance between rays and not elements.
Birkhoff's theorem, which was originally stated for the linear case and suitably extended by Bushell \cite{bushell1973hilbert}, provides bounds on the induced gain of positive maps. More specifically, a map ${\mathcal E}$ from ${\mathcal S}$ to ${\mathcal S}$ is said to be {\em positive} provided it takes the interior of ${\mathcal K}$ into itself, i.e.,
\[
{\mathcal E}\;:\;{\mathcal K}\backslash\{0\} \to {\mathcal K}\backslash\{0\}.
\]
For such a map define its {\em projective diameter}
\begin{eqnarray*}
\Delta({\mathcal E}):=\sup\{d_H({\mathcal E}(x),{\mathcal E}(y))\mid x,y\in {\mathcal K}\backslash\{0\}\}
\end{eqnarray*}
and the {\em contraction ratio}
\begin{eqnarray*}
\|{\mathcal E}\|_H:=\inf\{\lambda \mid d_H({\mathcal E}(x),{\mathcal E}(y))\leq \lambda d_H(x,y),\mbox{ for all }x,y\in{\mathcal K}\backslash\{0\}\}.
\end{eqnarray*}
The Birkhoff-Bushell theorem states the following.
\begin{thm}[\cite{{birkhoff1957extensions},bushell1973hilbert,bushell1973projective}]\label{BBcontraction}
{Let ${\mathcal E}$ be a positive map as above. If ${\mathcal E}$ is monotone and homogeneous of degree $m$,
i.e., if
\[
x\preceq y \Rightarrow {\mathcal E}(x)\preceq {\mathcal E}(y),
\]
and
\[
{\mathcal E}(\lambda x)=\lambda^m {\mathcal E}(x),
\]
then it holds that}
\[
\|{\mathcal E}\|_H\leq m.
\]
For the special case where ${\mathcal E}$ is also linear, the (possibly stronger) bound
\[
\|{\mathcal E}\|_H=\tanh(\frac{1}{4}\Delta({\mathcal E}))
\]
also holds.
\end{thm}
Birkoff's result provides a far-reaching generalization of the celebrated Perron-Frobenius theorem \cite{birkhoff1962Perron-Frobenius}. Various other applications of the Birkhoff-Bushell result have been developed such as to positive integral operators and to positive definite matrices \cite{bushell1973hilbert, lemmens2013birkhoff}.
We will use Theorem \ref{BBcontraction} to establish existence of solutions to certain equations involving Markovian evolutions on a finite time interval. More specifically, ${\mathcal S}$ will either be $\mathbb{R}^n$ or the space of symmetric/Hermitian $n\times n$ matrices with elements in $\mathbb{R}$ or $\mathbb{C}$, accordingly. Then ${\mathcal K}$ will either be the positive orthant of vectors with non-negative entries or the cone of non-negative definite matrices, respectively. The former setting will be brought in to give an independent proof of existence of the solution to the Schr\"{o}dinger bridge problem for Markov chains. The latter will be called in for studying {\em quantum operations} (Kraus maps) in Section \ref{sec:quantum}.
\section{Schr\"{o}dinger's problem for discrete random vectors}\label{sec:classical}
Following quite closely Schr\"{o}dinger's original derivation, we give below a rather self-contained presentation for the purpose of later reference and comparison when we deal with the more complex quantum case. First we discuss general discrete random vectors and then ``time windows" of a Markov chain. A paper dealing with the discrete time, continuous state space setting is \cite{beghi1996relative}. Schr\"{o}dinger bridges for Markov chains have been discussed in \cite{pavon2010discrete}. A nice survey with extensive bibliography for the diffusion case is \cite{wakolbinger1992bridges}.
\subsection{Discrete random vectors}\label{discreterandom}
Given a finite\footnote{{The case of a finite state space $\mathcal X$ is chosen for simplicity of exposition. Results extend in a straightforward way to the case of a countable $\mathcal X$.}} set $\mathcal X=\{1,\ldots,N\}$ , we are concerned with probability distributions $P$ on ``trajectories" $x=(x_0,x_1,\ldots,x_T)$ in $\mathcal X^{T+1}$. We write $\mathcal P$ for the simplex of all such distributions. Let us introduce the {\em coordinate mapping process} $X=\{X(t), 0\le t\le T\}$ by $X(t)(x)=x_t$. {For brevity, we often write $X(t)=x_t$ instead of $X(t)(x)=x_t$.} For $P\in \mathcal P$, we denote by
\[
p(s,x_s;t,x_t):=P(X(t)=x_t \mid X(s)=x_s), \quad 0\le s < t\le T, \quad x_s, x_t\in{\cal X}.
\]
its transition probabilities. We also use $p$, but indexed, to denote marginals. Thus,
\[
p_{t}(x_t):=P(X(t)=x_t),
\]
and similarly, for two-time marginals,
\[
p_{st}(x_s,x_t):=P(X(s)=x_s, X(t)=x_t).
\]
Schr\"{o}dinger's original formulation was set in continuous time and {with continuous state space}. The formulation herein represents the case where Schr\"{o}dinger's problem has undergone ``coarse graining" in Boltzmann's style in its phase space and where time has also been discretized. Thus, the {\em a priori} model is now given by a distribution $P\in\mathcal P$ and suppose that in experiments an initial and a final marginal $\mathbf p_0$ and $\mathbf p_T$, respectively, have been observed that differ from the marginals $ p_0$ and $p_T$ of the prior distribution $P$. We denote by ${\mathcal P}(\mathbf p_0,\mathbf p_T)\subset \mathcal P$ the family of distributions having the observed marginals and seek a distribution in ${\mathcal P}(\mathbf p_0,\mathbf p_T)$ which is close to the given prior $P$. Large deviation reasoning \cite{sanov1957largedeviations} requires that we employ as ``distance'' the {\em relative entropy}:
\begin{defn}{\em Let $P,Q\in \mathcal P$, that is, they belong to the simplex of probability distributions on ${\cal X}^{T+1}$, and let $x=(x_0,x_1,\ldots,x_T)$. If $P(x)=0\Rightarrow Q(x)=0$, we say that the {\em support} of $Q$ is contained in the support of $P$ and write \[{\rm Supp} (Q)\subseteq {\rm Supp} (P).\]
The {\em Relative Entropy} of $Q$ from $P$ is defined to be
\begin{equation}\label{KLdist}{\mD}(Q\|P)=\left\{\begin{array}{ll} \sum_{x\in{\cal X}^{T+1}}Q(x)\log\frac{Q(x)}{P(x)}, & {\rm Supp} (Q)\subseteq {\rm Supp} (P),\\
+\infty , & {\rm Supp} (Q)\not\subseteq {\rm Supp} (P).\end{array}\right.,
\end{equation}
where, by definition, $0\cdot\log 0=0$.}
\end{defn}
The relative entropy is also known as the {\em information} or {\em Kullback-Leibler divergence}. {As is well known \cite{cover_thomas2006information}, ${\mD}(Q\|P)\ge 0$ and ${\mD}(Q\|P)=0$ if and only if $Q=P$. }Given this notion of distance,
we seek
a probability law $Q^\circ\in {\mathcal P}(\mathbf p_0,\mathbf p_T)$
which is closest to the prior distribution $P$ in this sense. That is,
we seek a solution to the following problem.
\begin{problem}\label{prob:optimization}{\em Assume that $p(0,\cdot\,; T,\cdot)$ is everywhere positive on its domain. Determine
\begin{eqnarray}\label{eq:optimization}
Q^\circ={\rm argmin}\{ {\mD}(Q\|P)&\mid& Q\in {\mathcal P}(\mathbf p_0,\mathbf p_T)
\}.\nonumber
\end{eqnarray}}
\end{problem}
It turns out that if there is at least one $Q$ in
$\mathcal P(\mathbf p_0,\mathbf p_T)$ such that
${\mD}(Q\|P)<\infty$,
there exists a unique minimizer $Q^\circ$ called
{\em the Schr\"{o}dinger bridge} from $\mathbf p_0$ to $\mathbf p_T$ over $P$.
Now let
$$Q_{x_0,x_T}=Q\left[\,\cdot\,|X(0)=x_0,X(T)=x_T\right]$$
be the disintegration of Q with respect to the initial and final positions. Then, we have
$$Q(x_0,x_1,\ldots,x_T)=Q_{x_0,x_T}(x_0,x_1,\ldots,x_T)q_{0T}(x_0,x_T),
$$
where we have assumed that $q_{0T}$ is everywhere positive on $\mathcal X\times \mathcal X$.
We get
\begin{equation}
{\mD}(Q\|P)=\sum_{x_0x_T}q_{0T}(x_0,x_T)\log \frac{q_{0T}(x_0,x_T)}{p_{0T}(x_0,x_T)}+\sum_{x\in{\cal X}^{T+1}}Q_{x_0,x_T}(x)\log \frac{Q_{x_0,x_T}(x)}{P_{x_0,x_T}(x)} q_{0T}(x_0,x_T).
\end{equation}
This is the sum of two nonnegative quantities. The second becomes zero if and only if $Q_{x_0,x_T}(x)=P_{x_0,x_T}(x)$ for all $x\in{\cal X}^{T+1}$. Thus, $Q^\circ_{x_0,x_T}(x)=P_{x_0,x_T}(x)$. As already observed by Schr\"{o}dinger, Problem \ref{prob:optimization} then reduces to minimizing
\begin{equation}{\mD}(q_{0T}\|p_{0T})=\sum_{x_0x_T}p_{0T}(x_0,x_T)\log \frac{q_{0T}(x_0,x_T)}{p_{0T}(x_0,x_T)}
\end{equation}
with respect to $q_{0T}$ subject to the (linear) constraints
\begin{eqnarray}\sum_{x_T}q_{0T}(x_0,x_T)&=&\mathbf p_0(x_0),\quad x_0\in{\cal X},
\\
\sum_{x_0}q_{0T}(x_0,x_T)&=&\mathbf p_T(x_T),\quad x_T\in{\cal X}.
\end{eqnarray}
The Lagrangian function has the form
\begin{eqnarray}\nonumber&&{\cal L}(q_{0T})=\sum_{x_0x_T}q_{0T}(x_0,x_T)\log \frac{q_{0T}(x_0,x_T)}{p_{0T}(x_0,x_T)}\\&&+\sum_{x_0}\lambda(x_0)\left[\sum_{x_T}q_{0T}(x_0,x_T)-\mathbf p_0(x_0)\right]+\sum_{x_T}\mu(x_T)\left[\sum_{x_0}q_{0T}(x_0,x_T)-\mathbf p_T(x_T)\right].\nonumber
\end{eqnarray}
Setting the first variation equal to zero, we get the (sufficient) optimality condition
$$1+\log q_{0T}^\circ(x_0,x_T)-\log p(0,x_0; T,x_T)-\log p_0(x_0)+\lambda(x_0)+\mu(x_T)=0,
$$
where we have used the expression $p_{0T}(x_0,x_T)=p_0(x_0)p(0,x_0;T,x_T)$. Hence, the ratio $q_{0T}^\circ(x_0,x_T)/p(0,x_0; T,x_T)$ factors into a function of $x_0$ times a function of $x_T$; {these are denoted $\hat{\varphi}(x_0)$ and $\varphi(x_T)$, respectively.} We can then write the optimal $q_{0T}^\circ(\cdot,\cdot)$ in the form
\begin{equation}\label{optimaljoint} q_{0T}^\circ(x_0,x_T)=\hat{\varphi}(x_0) p(0,x_0;T,x_T)\varphi(x_T),
\end{equation} where $\varphi$ and $\hat{\varphi}$ {must satisfy}
\begin{eqnarray}\hat{\varphi}(x_0)\sum_{x_T}p(0,x_0;T,x_T)\varphi(x_T)&=&\mathbf p_0(x_0),\label{opt1}\\
\varphi(x_T)\sum_{x_0}p(0,x_0;T,x_T)\hat{\varphi}(x_0)&=&\mathbf p_T(x_T).\label{opt2}
\end{eqnarray}
Let us define $\hat{\varphi}(0,x_0)=\hat{\varphi}(x_0)$, $\quad \varphi(T,x_T)=\varphi(x_T)$ and
$$\hat{\varphi}(T,x_T)=\sum_{x_0}p(0,x_0;T,x_T)\hat{\varphi}(0,x_0),\quad \varphi (0,x_0):=\sum_{x_T}p(0,x_0;T,x_T)\varphi(T,x_T).
$$
Then, (\ref{opt1})-(\ref{opt2}) can be replaced by the system
\begin{eqnarray}\label{Schonestep1}
\hat{\varphi}(T,x_T)=\sum_{x_0}p(0,x_0;T,x_T)\hat{\varphi}(0,x_0),\\\label{Schonestep2}\quad \varphi (0,x_0):=\sum_{x_T}p(0,x_0;T,x_T)\varphi(T,x_T)
\end{eqnarray}
with the boundary conditions
\begin{equation}\label{BConestep}
\varphi(0,x_0)\cdot\hat{\varphi}(0,x_0)=\mathbf p_0(x_0),\quad \varphi(T,x_T)\cdot\hat{\varphi}(T,x_T)=\mathbf p_T(x_T),\quad \forall x_0, x_T\in\mathcal X.
\end{equation}
The question of existence and uniqueness of functions $\hat{\varphi}(x_0)$, $\varphi(x_T)$ satisfying (\ref{Schonestep1})-(\ref{Schonestep2})-(\ref{BConestep}) will be established in Section \ref{solonestep}. Before we do that, however, we investigate what else can be said about the solution when the prior random vector happens to be a ``time window" of a Markov chain.
\subsection{Markovian prior}\label{sec:prior}
Consider the special case where $X=\{X(t), 0\le t\le T\}$ is a window of a Markov chain. Let us introduce, in the language of Doob, the {\em space-time harmonic} function
\begin{equation}\label{defvarphi}\varphi (t,x_t):=\sum_{x_T}p(t,x_t;T,x_T)\varphi(x_T),\quad 0\le t\le T,
\end{equation}
and the {\em space-time co-harmonic} function
\begin{equation}\label{defvarphihat}\hat{\varphi} (t,x_t):=\sum_{x_0}p(0,x_0;t,x_t)\hat{\varphi}(x_0), \quad 0\le t\le T.
\end{equation}
Because of the Markov property, for $0\le s<t<u\le T$, we have
$$
p(s,x_s;u,x_u)=\sum_{x_t}p(s,x_s;t,x_t)p(t,x_t;u,x_u).
$$
For compactness, we often write $\pi_{x_t,x_{t+1}}(t)$ and $\pi_{x_t,x_{t+n}}^{(n)}(t)$ of instead of $p(t,x_t;t+1,x_{t+1})$ and $p(t,x_t;t+n,x_{t+n})$, respectively. Hence, $\varphi$ and $\hat{\varphi}$ satisfy the backward and forward equation, respectively,
\begin{eqnarray}
&&\varphi(t,x_t)=\sum_{x_{t+1}}\pi_{x_t,x_{t+1}}(t)\varphi(t+1,x_{t+1}),\\&&\hat{\varphi}(t+1,x_{t+1})=\sum_{x_t}\pi_{x_t,x_{t+1}}(t)\hat{\varphi}(t,x_t).
\end{eqnarray}
Let $q^\circ_t$ denote the distribution of the Schr\"{o}dinger bridge at time $t$. We get
\begin{eqnarray}
\nonumber q^\circ_t(x_t)=\sum_{x_0}\sum_{x_T}p(0,x_0;t,x_t;T,x_T)q_{0T}^\circ(x_0,x_T)=\\\sum_{x_0}\sum_{x_T}\frac{p(0,x_0;t,x_t)p(t,x_t;T,x_T)}{p(0,x_0;T,x_T)}\hat{\varphi}(x_0)p(0,x_0;T,x_T)\varphi(x_T)\nonumber\\=\hat{\varphi} (t,x_t)\cdot\varphi (t,x_t).\label{factor}
\end{eqnarray}
Similarly, one gets
$$q_{st}^\circ(x_s,x_t)=\sum_{x_0}\sum_{x_T}\hat{\varphi}(x_0)p(0,x_0;s,x_s)p(s,x_s;t,x_t)p(t,x_t;T,x_T)\varphi(x_T)
$$
which yields, using (\ref{defvarphi})-(\ref{defvarphihat}) and (\ref{factor}), the new transition probabilities
\begin{equation}\label{TRA}q^\circ(s,x_s;t,x_t)=\frac{q_{st}^\circ(x_s,x_t)}{q^\circ_s(x_s)}=p(s,x_s;t,x_t)\frac{\varphi(t,x_t)}{\varphi(s,x_s)}.
\end{equation}
Notice, in particular, that the Schr\"{o}dinger bridge is also a {\em Markov chain} which is obtained from the {\em a priori} model via a suitable {\em multiplicative functional transformation} just like in the diffusion case \cite{jamison1975Markovprocesses}. We have therefore established the following result \cite[Theorem 4.1]{pavon2010discrete}:
\begin{thm} \label{fund theorem}Assume that $p(0,\cdot; T,\cdot)$ is everywhere positive on $\mathcal X\times \mathcal X$. Suppose there exist positive functions $\varphi$ and $\hat{\varphi}$ defined on $[0,T]\times\mathcal X$ satisfying for $t\in[0,T-1]$ the system
\begin{eqnarray}\label{Schroedingersystem1}
\varphi(t,x_t)&=&\sum_{x_{t+1}}\pi_{x_t,x_{t+1}}(t)\varphi(t+1,x_{t+1}),\\\hat{\varphi}(t+1,x_{t+1})&=&\sum_{x_t}\pi_{x_t,x_{t+1}}(t)\hat{\varphi}(t,x_t).\label{Schroedingersystem2}
\end{eqnarray}
with the boundary conditions
\begin{equation}\label{bndconditions}
\varphi(0,x_0)\cdot\hat{\varphi}(0,x_0)=\mathbf p_0(x_0),\quad \varphi(T,x_T)\cdot\hat{\varphi}(T,x_T)=\mathbf p_T(x_T),\quad \forall x_0, x_T\in\mathcal X.
\end{equation}
Then the Markov distribution $Q^\circ$ in ${\mathcal P}(\mathbf p_0,\mathbf p_T)$ with one-step transition probabilities
\begin{equation}\label{opttransition}\pi_{x_t,x_{t+1}}^\circ(t)=\pi_{x_t,x_{t+1}}(t)\frac{\varphi(t+1,x_{t+1})}{\varphi(t,x_t)}
\end{equation}
is the unique solution of Problem \ref{prob:optimization}.
\end{thm}
Notice that, if a pair $\varphi,\hat{\varphi}$ solves (\ref{Schroedingersystem1})-(\ref{Schroedingersystem2})-(\ref{bndconditions}), so does the pair $a\varphi,a^{-1}\hat{\varphi}$ for $a>0$. This arbitrariness in the scaling, however, does not affect the transition probabilities (\ref{opttransition}).
Schr\"{o}dinger and Kolmogorov were struck by the intrinsic time-reversibility of the solution: Swapping $\mathbf p_0$ and $\mathbf p_T$ leads to a solution bridge which is simply the time reversal of the original one. Moreover, the factorization (\ref{factor}) resembles Born's relation between the probability density and the wave function in quantum mechanics $\rho_t(x)=\psi(x,t)\psi^\dagger(x,t)$ where $\dagger$, throughout the paper, denotes conjugation/adjoint in a complex Hilbert space \footnote{Schr\"{o}dinger:``Merkw\"{u}rdige Analogien zur
Quantenmechanik, die mir sehr des Hindenkens wert erscheinen" (remarkable analogies to quantum mechanics which appear to me very worth of reflection). Recall that Schr\"{o}dinger never accepted the so-called orthodox theory of measurement in quantum mechanics and was looking for a more classical probabilistic reformulation.}.
At this point, it should be apparent that the bottleneck of the theory of Schr\"{o}dinger bridges is the existence and uniqueness (up to multiplication by a positive constant) of the pair $\varphi,\hat{\varphi}$ solving the Schr\"{o}dinger system (\ref{Schonestep1})-(\ref{Schonestep2})-(\ref{BConestep}) (in the Markov case, (\ref{Schroedingersystem1})-(\ref{Schroedingersystem2})-(\ref{bndconditions})). Schr\"{o}dinger thought that existence and uniqueness should be guaranteed in the diffusion case ``except possibly for very nasty $\rho_0$, $\rho_T$, since the question leading to the pair of equations is so reasonable". This problem turns out to be quite nontrivial and was settled in various degrees of generality by Beurling, Jamison and F\"{o}llmer \cite{beurling1960automorphism,jamison1974reciprocal,foellmer1988randomfields} establishing the feasibility of the dual optimization problem of Problem \ref{prob:optimization}. There is, however, an alternative approach based on proving convergence of successive approximations by Fortet \cite {fortet1940}. As it turns out, this more algorithmic approach had independent counterparts called ``iterative fitting algorithms" in the statistical literature on contingency tables \cite{demingstephan1940}. These were later shown to converge to a ``minimum discrimination information" \cite{irelandkullback1968,fienberg1970}, namely to a minimum entropy distance, see also \cite {csiszar1975divergence}. It will be apparent below that the approach pioneered by Fortet features a very desirable property: While establishing existence and uniqueness for the Schr\"{o}dinger system, it also provides a computationlly efficient {\em algorithm} to actually {\em compute} the space-time harmonic function $\varphi$ and therefore the solution.
Our approach consists in showing that a certain iterative scheme for the Schr\"{o}dinger system is a contraction mapping on the positive orthant with respect to the projective metric which was introduced in Section \ref{sec:hilbert}. We argue that this is the {\em natural} metric in which to cast the iteration for these problems both in the classical and in the quantum case. Before we turn to that, however, we show that in the Markov case the new transition mechanism may be obtained via the solution of a two consecutive times interval problem.
\subsection{Reduction to the one-step bridge problem}\label{reduction}
Let $\Pi(t)=\left(\pi_{x_t,x_{t+1}}(t)\right)_{x_t,x_{t+1}=1}^N$ and $\Pi^\circ(t)=\left(\pi^\circ_{x_t,x_{t+1}}(t)\right)_{x_t,x_{t+1}=1}^N$ be the transition matrices of the prior distribution and of the bridge distribution. It is interesting to express (\ref{opttransition}) in matrix form. To this end, we introduce the notation
\[
\phi(t) = \mathop{\mathrm{diag}}\left(\varphi(t,x_1),\varphi(t,x_2),\ldots,\varphi(t,x_N)\right)
\]
for a diagonal matrix formed out of the entries of $\varphi(t,\cdot)$.
We now have
\begin{equation}\label{opttransitionmatrix}
\Pi^\circ(t)=\phi(t)^{-1}\Pi(t)\phi(t+1).
\end{equation}
Consider now for $n\ge 1$ $\Pi^{(n)}(t)=\left(\pi^{(n)}_{x_t,x_{t+n}}(t)\right)_{x_t,x_{t+n}=1}^N$ the $n$-step transition probabilities matrix. By the Markov property, we have
$$\Pi^{(n)}(t)=\Pi(t)\cdot\Pi(t+1)\cdots\Pi(t+n-1).
$$
Being the product of stochastic matrices, $\Pi^{(n)}(t)$ is also {\em stochastic}\footnote{The elements of $\Pi^{(n)}(t)$ are nonnegative as sum of products of nonnegative numbers. Moreover, let $\mathds{1}^\dagger=(1,1,\ldots,1)$. Then the fact that a matrix $Q$ has rows summing to one can be expressed as $Q \mathds{1}=\mathds{1}$.
As each $\Pi(t)$ has rows summing to one , we have
$$\Pi^{(n)}(t)\mathds{1}=\Pi(t)\cdot\Pi(t+1)\cdots\Pi(t+n-1)\mathds{1}
=\Pi(t)\cdot\Pi(t+1)\cdots\Pi(t+n-2)\mathds{1}
=\cdots=\Pi(t)\mathds{1}
=\mathds{1}.
$$}.
Consider now $\Pi^{\circ^{(n)}}(t)$. By (\ref{opttransitionmatrix}), we get
\begin{eqnarray}\nonumber
&&\Pi^{\circ^{(n)}}(t)=\Pi^\circ(t)\cdot \Pi^\circ(t+1)\cdots \Pi^\circ(t+n-1)\\&&=\phi(t)^{-1}\Pi^{(n)}(t)\phi(t+n).\nonumber
\end{eqnarray}
Thus, we get the following remarkable generalization of formula (\ref{opttransition})
\begin{equation}\label{opttransitionnstep}
\pi_{x_tx_{t+n}}^{\circ^{(n)}}(t)=\pi^{(n)}_{x_t,x_{t+n}}(t)\frac{\varphi(t+n,x_{t+n})}{\varphi(t,x_t)}.
\end{equation}
Consider now $\Pi^{(T)}(0)$ as the transition matrix of a prior ``stroboscopic" evolution. Consider also the Schr\"{o}dinger bridge problem with the same marginals $\mathbf p_0$ and $\mathbf p_T$ as before at the two {\em consecutive} times $0$ and $T$. Specializing Theorem \ref{fund theorem} to this simple situation we get that the new transition probabilities are precisely given by (\ref{opttransitionnstep}) with $n=T$ where $\varphi$, $\hat{\varphi}$ satisfy (\ref{Schonestep1})-(\ref{Schonestep2})-(\ref{BConestep}), namely
\begin{eqnarray}\label{Schroedingersystemonestep}
\varphi(0,x_0)&=&\sum_{x_T} \pi^{(T)}_{x_0,x_T}(0)\varphi(T,x_T),\\\hat{\varphi}(T,x_T)&=&\sum_i\pi^{(T)}_{x_0,x_T}(0)\hat{\varphi}(0,x_0).
\end{eqnarray}
with the boundary conditions
\begin{equation}\label{bndconditionsonestep}
\varphi(0,x_0)\cdot\hat{\varphi}(0,x_0)=\mathbf p_0(x_0),\quad \varphi(T,x_T)\cdot\hat{\varphi}(T,x_T)=\mathbf p_T(x_T),\quad \forall x_0, x_T\in\mathcal X.
\end{equation}
We see that the solution of Problem \ref{prob:optimization} yields, as a by-product, the solution of the one-step problem. The converse, however, is also true. Solving (\ref{Schroedingersystemonestep})-(\ref{bndconditionsonestep}), yields the correct terminal value $\varphi(T,\cdot)$ from which $\varphi$ can be computed at all times through the iteration (\ref{Schroedingersystem1}). From it, the new transition probabilities are obtained through (\ref{opttransition}). We already know from Section \ref {discreterandom} that
solving the one-step problem suffices to characterize the optimal distribution $Q^\circ$. The argument above shows that it also permits to obtain the more explicit description, namely the transition mechanism of $Q^\circ$, which is desirable in the Markov case.
\subsection{The solution to the one-step bridge problem}\label{solonestep}
Motivated by what we have seen in Sections \ref{discreterandom} and \ref{reduction}, we study the one-step bridge problem with possibly non Markovian prior. The following result establishes existence and uniqueness for the system (\ref{Schonestep1})-(\ref{Schonestep2})-(\ref{BConestep}).
\begin{thm} \label{fundtheorem} Given a $N\times N$ stochastic matrix
\[
\Pi=\left[\pi_{x_0,x_T}\right]_{x_0,x_T=1}^N
\]
with strictly positive entries {\rm (}$\pi_{x_0,x_T}>0${\rm )} and probability distributions ${\mathbf p}_0$, ${\mathbf p}_T$, there exist four vectors
$
\varphi(0,x_0),\,\varphi(T,x_T),\,\hat\varphi(0,x_0),\,\hat\varphi(T,x_T)$, indexed by $x_0, x_T\in\mathcal X,
$
with positive entries such that
\begin{subequations}\label{eq:iteration}
\begin{eqnarray}\label{eq:iterationa}
\varphi(0,x_0)&=&\sum_{x_T}\pi_{x_0,x_T}\varphi(T,x_T),\\
\hat\varphi(T,x_T)&=&\sum_{x_0}\pi_{x_0,x_T}\hat\varphi(0,x_0),\label{eq:iterationb}\\
\label{eq:iterationc}
\varphi(0,x_0)\hat\varphi(0,x_0)&=&{\mathbf p}_0(x_0),\\\label{eq:iterationd}
\varphi(T,x_T)\hat\varphi(T,x_T)&=&{\mathbf p}_T(x_T).
\end{eqnarray}
\end{subequations}
The four vectors are unique up to multiplication of $\varphi(0,x_0)$ and $\varphi(T,x_T)$ by the same positive constant and division of $\hat\varphi(0,x_0)$ and $\hat\varphi(T,x_T)$ by the same constant.
\end{thm}
Rather than relying on the results of Beurling and Jamison as in \cite{pavon2010discrete,ticozzi2010time}, we give an independent proof which also yields an effective algorithm. The proof relies on showing that a certain iteration is strictly contractive in the Hilbert metric\footnote{Fortet's proof, which we find very difficult to follow, is apparently based on establishing monotonicity of two sequences of functions produced by the iteration.}.
\begin{lemma}\label{keylemma} Consider the following circular diagram of maps
\begin{equation
\begin{array}{ccccc}
&\hat\varphi(0,x_0) & \overset{{\mathcal E}^\dagger}{\longrightarrow} & \hat\varphi(T,x_T) & =\sum_{x_0}\pi_{x_0,x_T}\hat \varphi(0,x_0)
\\\\
\hat\varphi(0,x_0)= \frac{{\mathbf p}_0(x_0)}{\varphi(0,x_0)}& \uparrow & & \downarrow &\varphi(T,x_T)= \frac{{\mathbf p}_T(x_T)}{\hat\varphi(T,x_T)}\\\\
\sum_{x_N}\pi_{x_0,x_N}\varphi(T,x_T)=&\varphi(0,x_0) &\overset{{\mathcal E}}{\longleftarrow} & \varphi(T,x_T) &
\end{array}
\end{equation}
where
\begin{eqnarray*}
{\hat {\mathcal D}}_0\;:\;\varphi(0,x_0)&\mapsto& \hat\varphi(0,x_0)= \frac{{\mathbf p}_0(x_0)}{\varphi(0,x_0)}\\
{\mathcal D}_T\;:\;\hat\varphi(T,x_T)&\mapsto& \varphi(T,x_T)=\frac{{\mathbf p}_T(x_N)}{\hat\varphi(T,x_T)}
\end{eqnarray*}
represent componentwise division of vectors.
Then the composition
\begin{equation}\label{eq:composition}
\hat\varphi(0,x_0) \overset{{\mathcal E}^\dagger}{\longrightarrow} \hat\varphi(T,x_T)\overset{{\mathcal D}_{T}}{\longrightarrow}\varphi(T,x_T)
\overset{{\mathcal E}}{\longrightarrow} \varphi(0,x_0)
\overset{{\mathcal D}_{0}}{\longrightarrow}\left(\hat\varphi(0,x_0)\right)_{\rm next}
\end{equation}
is contractive in the Hilbert metric.
\end{lemma}
Before proceeding with the proof
we provide a note about notation: The map
\[
{\mathcal E}^\dagger \;:\; \hat\varphi(0,x_0) \mapsto \hat\varphi(T,x_T)=\sum_{x_0}\pi_{x_0,x_N}\hat \varphi(0,x_0)
\]
is the adjoint
of the backward evolution
\[
{\mathcal E}\;:\; \varphi(T,x_T)\mapsto \varphi(0,x_0)=\sum_{x_N}\pi_{x_0,x_T}\varphi(T,x_T),
\]
which is consistent with the standard notation in diffusion processes where the Fokker-Planck (forward) equation involves the adjoint of the {\em generator} appearing in the backward Kolmogorov equation.
Also notice that the componentwise divisions of ${\hat {\mathcal D}}_0$ and ${\mathcal D}_T$ are well defined. Indeed, even when $\hat\varphi(0)$ ($\varphi(T)$) has zero entries, $\hat\varphi(T)$ ($\varphi(0)$) has all positive entries since the elements of $\Pi$ are all positive.
\noindent{\bf Proof of Lemma \ref{keylemma}:}
The diameter of the range of ${\mathcal E}$ is
\begin{eqnarray*}
\Delta({\mathcal E})&=&\sup\{d_H({\mathcal E}(x),{\mathcal E}(y)) \mid x_i>0,\,y_i>0\}\\
&=&\sup\{\log\left(\frac{\pi_{ij}\pi_{k,\ell}}{\pi_{i,\ell}\pi_{k,j}}\right)\mid1\leq i,j,k,\ell\leq n\}
\end{eqnarray*}
is finite since all entries $\pi_{i,j}$'s are positive.
Birkhoff's theorem {(Theorem \ref{BBcontraction})} provides a contraction coefficient for linear positive maps and, in this case,
{we have}
\[
\|{\mathcal E}\|_H=\tanh(\frac{1}{4}\Delta({\mathcal E}))<1.
\]
For the adjoint map ${\mathcal E}^\dagger$ we only need to note that it is homogeneous of degree $1$ and therefore, by Birkhoff's theorem,
\[
\|{\mathcal E}^\dagger\|_H\leq 1.
\]
{Next we note that provided
${\mathbf p}_0(x_0)$ and ${\mathbf p}_T(x_N)$ have positive entries, both ${\hat {\mathcal D}}_0$ and ${\mathcal D}_T$
are isometries in the Hilbert metric since inversion and element-wise scaling are both isometries. Indeed, for vectors $[x_i]_{i=1}^N$, $[y_i]_{i=1}^N$, it holds that
\begin{eqnarray*}
d_H([x_i],[y_i])&=&\log\left((\max_i (x_i/y_i))\frac{1}{\min_i (x_i/y_i)}\right)\\
&=&\log\left(\frac{1}{\min_i ((x_i)^{-1}/(y_i)^{-1})}\max_i ((x_i)^{-1}/(y_i)^{-1})\right)\\&=&d_H([(x_i)^{-1}],[(y_i)^{-1}])
\end{eqnarray*}
and
\begin{eqnarray*}
d_H([{\mathbf p}_i x_i],[{\mathbf p}_i y_i])&=&\log\frac{\max_i (({\mathbf p}_i x_i)/({\mathbf p}_i y_i))}{\min_i (({\mathbf p}_i x_i)/({\mathbf p}_i y_i))}\\
&=&\log\frac{\max_i (x_i/y_i)}{\min_i (x_i/y_i)}=d_H([x_i],[y_i]).
\end{eqnarray*}
If ${\mathbf p}_0(x_0)$ and ${\mathbf p}_T(x_N)$ have any zero entries, then
${\hat {\mathcal D}}_0$ and ${\mathcal D}_T$
are in fact contractions.}
Finally, we observe that
\[
\|{\hat {\mathcal D}}_0\circ {\mathcal E}\circ {\mathcal D}_T\circ{\mathcal E}^\dagger\|_H\leq \|{\hat {\mathcal D}}_0\|_H\cdot\|{\mathcal E}\|_H\cdot\|{\mathcal D}_T\|_H\cdot\|{\mathcal E}^\dagger\|_H<1,
\]
where $\circ$ denotes composition. Therefore, the composition ${\mathcal C}:={\hat {\mathcal D}}_0\circ {\mathcal E}\circ {\mathcal D}_T\circ{\mathcal E}^\dagger$ is contractive: Namely,
\[
d_H({\mathcal C}(x),{\mathcal C}(y))\le\|{\mathcal C}\|_Hd_H(x,y), \quad 0\le\|{\mathcal C}\|_H<1 .
\]
This completes the proof of the lemma.\hfill $\Box$
\noindent{\bf Proof of Theorem \ref{fundtheorem}:}
{Since ${\mathcal C}$ is contractive in the Hilbert metric, there is a unique positive $\hat\varphi(0,\cdot)=[\hat\varphi(0,x_0)]$ so that the corresponding ray is invariant under ${\mathcal C}$. That is, in the notation of \eqref{eq:composition},
\begin{eqnarray*}
(\hat\varphi(0,\cdot))_{\rm next}&=&{\mathcal C}(\hat\varphi(0,\cdot))\\
&=& \lambda \hat\varphi(0,\cdot),
\end{eqnarray*}
for the composition ${\mathcal C}:={\hat {\mathcal D}}_0\circ {\mathcal E}\circ {\mathcal D}_T\circ{\mathcal E}^\dagger$. From this we can obtain
\begin{eqnarray*}
\hat\varphi(T,\cdot)&=&{\mathcal E}^\dagger(\hat\varphi(0,\cdot)),\\
\varphi(0,\cdot)&=&{\mathcal E}(\varphi(T,\cdot)),
\end{eqnarray*}
while
\begin{eqnarray*}
\hat\varphi(T,\cdot)\varphi(T,\cdot)&=& {\mathbf p}_T(\cdot)\mbox{ and}\\
\lambda\hat\varphi(0,\cdot)\varphi(0,\cdot)&=&{\mathbf p}_0(\cdot).
\end{eqnarray*}
However, since
\begin{eqnarray*}
1&=& \lambda\langle \hat\varphi(0,\cdot),\varphi(0,\cdot)\rangle\\
&=& \lambda\langle \hat\varphi(0,\cdot),{\mathcal E}(\varphi(T,\cdot))\rangle\\
&=&\lambda\langle {\mathcal E}^\dagger(\hat\varphi(0,\cdot)),\varphi(T,\cdot)\rangle\\
&=&\lambda\langle \hat\varphi(T,\cdot),\varphi(T,\cdot)\rangle\\
&=&\lambda,
\end{eqnarray*}
we conclude that
these satisfy the Schr\"{o}dinger system (\ref{eq:iterationa})-(\ref{eq:iterationb})-(\ref{eq:iterationc})-(\ref{eq:iterationd}).} \hfill $\Box$
\begin{remark} {\bf Numerical algorithm:} The step above suggest that starting from any positive vector, e.g., $x(0)=\mathds{1}$, the unique fixed point is given by
\begin{eqnarray}\label{eq:algorithm}
\hat\varphi(0) &=& \lim_{k\to\infty} x(k),
\end{eqnarray}
where $x(k+1)={\mathcal C}(x(k))$ for $k=1,2,\ldots$.
\end{remark}
By the variational analysis of Section \ref{discreterandom}, we finally get the following result.
\begin{thm} Assume that $\mathbf p_0$ and $\mathbf p_T$ are positive on $\mathcal X$ and that $p(0,\cdot; T,\cdot)$ is everywhere positive on $\mathcal X\times \mathcal X$. Then, the unique solution to Problem \ref{prob:optimization} with prior $P\in\mathcal P$ is given by
$$Q^\circ(x_0,x_1,\ldots,x_T)=P_{x_0,x_T}(x_0,x_1,\ldots,x_T)q^\circ_{0T}(x_0,x_T),$$
where $q^\circ_{0T}(x_0,x_T)$ solving the problem for the two consecutive times $0,T$ is
given by
\begin{subequations}\label{eq:solutiononestep}
\begin{equation}
q_{0T}^\circ(x_0,x_T)=\hat{\varphi}(0,x_0) p(0,x_0;T,x_T)\varphi(T,x_T).
\end{equation}
Here $\hat{\varphi}(0,x_0)$, $\varphi(T,x_T)$ are as in Theorem \ref{fundtheorem} with
$$\Pi=\left[\pi_{x_0,x_T}\right]_{x_0,x_T=1}^N, \quad \pi_{x_0,x_T}=p(0,x_0;T,x_T).$$
\end{subequations}
\end{thm}
{The corresponding transition probability is given
by
\begin{equation}\label{eq:transition_q0T}
q^\circ(0,x_0;T,x_T)= p(0,x_0;T,x_T)\frac{\varphi(T,x_T)}{\varphi(0,x_0)}.
\end{equation}}\\
In fact, when $P$ is Markovian, we also get the new one-step transition probabilities in a similar manner, i.e., in the way described at the end of Section \ref{reduction}. {Finally, as already observed, extending the results of this section to the case of a countable state space is straightforward.}
\section{The Schr\"{o}dinger bridge problem for quantum evolutions}\label{sec:quantum}
We begin with some background on quantum probability and stochastic maps, {see e.g. \cite{nielsen-chuang2000quantum,holevo2001statistical,petz2008quantum}.}
In quantum probability there is neither a notion of a random variable nor of a probability space. Instead,
random variables are replaced by Hermitian matrices/operators, referred to as {\em observables}, while the expected outcome of an experiment is quantified by a suitable functional that provides the {\em expectation}.
Throughout, we restrict our attention to the case {of finite-dimensional quantum systems with associated Hilbert space
isomorphic to ${\mathbb C}^n$. Here} experiments take values from a finite alphabet and, thereby, the mathematical framework relies on algebras of finite matrices. In this case, the state of the underlying system which defines the expectation is represented by a non-negative definite matrix with trace one which is referred to as a {\em density matrix}.
Notation and basic terminology are explained next, followed by the formulation the quantum bridge problem, namely, the problem of identifying a quantum channel that is consistent with given quantum states at its two ends and has a specific relation to a given prior quantum channel.
For potential applications of quantum bridges and links to topics in quantum information we refer to \cite[Section VII]{pavon2010discrete}.
\subsection{Quantum channels}\label{sec:quantumchannels}
We use
${\mathfrak M}:=\{X\in {\mathbb C}^{n\times n} \}$ to denote square matrices, ${\mathfrak H}:=\{X\in {\mathbb C}^{n\times n} \mid X=X^\dagger\}$ to denote the set of Hermitian matrices, ${\mathfrak H}_+$ and ${\mathfrak H}_{++}$ the cones of nonnegative and positive definite ones, respectively, and ${\mathfrak D}:=\{\rho\in{\mathfrak H}_+\mid \operatorname{\rm trace} \rho=1\}$ the set of density matrices. The latter represents possible ``states'' of a quantum system. In turn, a state $\rho$ defines an {\em expectation} functional on observables $X\in{\mathfrak H}$
\[
{\mathbb E}_\rho(X)=\operatorname{\rm trace}(\rho X).
\]
The standard model for a quantum experiment as well as for a quantum channel is provided by a linear {\em trace-preserving completely-positive} (TPCP) map between density matrices (which may possibly be of different size). This is the quantum counterpart of a Markov evolution and is referred to as a {\em Kraus map}. For the {finite-dimensional} case treated herein, where quantum systems are represented by density matrices and where, for notational simplicity, matrices are all of the same size, a Kraus map assumes the representation
\begin{subequations}\label{eq:krausrepresentation}
\begin{equation}\label{eq:kraus}
{\mathcal E}^\dagger \;:\; {\mathfrak D}\to{\mathfrak D}\;:\;\rho \longrightarrow \sigma=\sum_{i=1}^{n_{\mathcal E}} E_i\rho E_i^\dagger,
\end{equation}
with $E_i\in{\mathfrak M}$ such that
\begin{equation}\label{eq:I}
\sum_{i=1}^{n_{\mathcal E}}E_i^\dagger E_i=I,
\end{equation}
\end{subequations}
see \cite[Chapter 2]{petz2008quantum}. Throughout, $I$ denotes the identity matrix.
Condition \eqref{eq:I} ensures that the map preserves the trace, i.e., that $\operatorname{\rm trace}(\sigma)=\operatorname{\rm trace}(\rho)=1$. {\em Completely-positive} refers to the property that, besides the fact that ${\mathcal E}^\dagger(\rho)\geq 0$ for all $\rho\geq 0$ (i.e., being a {\em positive} map), if ${\mathbb I}_k$ denotes the identity map on ${\mathbb C}^{k\times k}$ and $\otimes$ the tensor product, then ${\mathbb I}_k\otimes {\mathcal E}^\dagger$ is a positive map for all $k$. The Kraus-representation (\ref{eq:kraus}-\ref{eq:I}) characterizes completely positive maps (see, e.g., \cite[Chapter 2]{petz2008quantum}).
Herein, a further condition will be imposed on Kraus maps which we
refer to as {\em positivity improving}. This is the property that
\begin{equation}\label{eq:strictpositivity}
{\mathcal E}^\dagger (\rho) >0 \mbox{ for any }\rho\in{\mathfrak D}.
\end{equation}
Thus, positivity improving amounts to having the range of ${\mathcal E}^\dagger$ contained in the interior of ${\mathfrak D}$ or, in other words and in view of \eqref{eq:kraus}, that there is no pair of vectors $w,\,v\in{\mathbb C}^n$ such that $w^\dagger E_i v=0$ for all $i\in\{1,\ldots,n_{\mathcal E}\}$.
A necessary condition\footnote{To see this assume that $n_{{\mathcal E}}\leq n$ and take $v$ to be an eigenvector of the matrix pencil $E_1-\lambda E_2$. Then $E_1v$ and $E_2v$ are linearly dependent and the span of $E_1v$ \ldots $E_nv$ has dimension less than $n$. This implies that there exists a $w$ such that $w^\dagger E_iv=0$ for all $i$. Determining whether a given Kraus map actually has the property of positivity improving, in general, is NP-hard \cite{gaubert2014checking,hillar2009most}
}
for positivity improving is that $n_{{\mathcal E}}> n$.
Our notation in using ${\mathcal E}^\dagger$ for the Kraus map is \underline{not} \underline{standard} and differs from that in e.g., \cite{petz2008quantum} where ${\mathcal E}$ is used instead. However, our choice maintains consistency with the standard convention throughout probability theory {where the (forward) Fokker-Planck equation involves} the adjoint of the generator of a Markov semigroup.
A model for quantum measurement amounts to an instantiation of a quantum channel \cite[Chapter 2]{petz2008quantum}. Typically an observable $X\in{\mathfrak H}$ is specified having a decomposition
\begin{equation}\nonumber
X=\sum_i \lambda_i {\chi}_i^\dagger {\chi}_i
\end{equation}
with ${\chi}_i\in{\mathfrak M}$ such that
\[
\sum_i {\chi}_i^\dagger {\chi}_i=I.
\]
A particular case is that of a spectral decomposition of $X$. When the outcome $i$ is being recorded and the value $\lambda_i$ registered for a quantum system that is initially in a state $\rho$, then, after the measurement, the system finds itself in the new state
\begin{equation}\label{eq:averaging}
\left(\frac{1}{\operatorname{\rm trace} ({\chi}_i\rho {\chi}^\dagger_i)}\right) {\chi}_i\rho {\chi}^\dagger_i.
\end{equation}
The likelihood of this particular outcome associated with recording $\lambda_i$ is $\operatorname{\rm trace} ({\chi}_i\rho {\chi}^\dagger_i)$. Therefore,
\begin{align*}
{\mathbb E}_\rho(X)&=\sum_i \lambda_i\operatorname{\rm trace} ({\chi}_i\rho {\chi}^\dagger_i)\\
&=\operatorname{\rm trace}(\rho X)
\end{align*}
in agreement with the earlier statement. On the other hand, if the experiment takes place but no particular outcome is recorded or, equivalently, all possible states in \eqref{eq:averaging} are weighed in by the corresponding probabilities $\operatorname{\rm trace} ({\chi}_i\rho {\chi}^\dagger_i)$, the new state becomes
\[
\sum_i {\chi}_i \rho {\chi}_i^\dagger.
\]
This is precisely a Kraus map acting on $\rho$ and this type of measurement is referred to as {\em nonselective}.
In the sequel we will encounter observables $\phi\in{\mathfrak H}$
factored as $\phi={\chi}^\dagger {\chi}$. These represent quantum analogues of the random variables $\varphi$ in the classical case. The expression
${\chi} \rho {\chi}^\dagger$
may be thought of as an ``unnormalized'' state that results in after a nonzero outcome has been recorded when a measurement performed. The factors of observables $\phi$ will give rise to the multiplicative {functional} transformations on Kraus maps sought in addressing the quantum bridge problem. These specify Kraus maps that link a set of two given {``marginal''} density matrices.
While there are many similarities between quantum channels and Markov evolution, there are also stark differences. A fundamental departure from classical probability arises in that, in general, there is no notion of joint probablitity between ``measurements'' at the two ends of a quantum channel. In this, the order with which measurements take place matters.
\subsection{The bridge problem}\label{sec:quantumbridge}
Consider a reference quantum evolution given by an initial density matrix $\sigma_0$ and by a sequence of TPCP maps $\{{\mathcal E}^\dagger_t; 0\le t\le T-1\}$, each admitting a Kraus representation with matrices $\{E_{t,i}\}$, so that
\[
{\mathcal E}^\dagger_t : \sigma_t \mapsto \sigma_{t+1}=\sum_i E_{t,i} \sigma_t E_{t,i}^\dagger, \quad t=0,1,\ldots, T-1.
\]
As usual,
\[
\sum_i E_{t,i}^\dagger E_{t,i} = I.
\]
Consider also the composition map
\[
{\mathcal E}_{0:T}^\dagger:={\mathcal E}^\dagger_{T-1}\circ\cdots \circ{\mathcal E}^\dagger_1\circ{\mathcal E}^\dagger_0.
\]
Observe that ${\mathcal E}_{0:T}^\dagger$ is a TPCP map. For instance, it is immediate that ${\mathcal E}^\dagger_1\circ{\mathcal E}^\dagger_0$ admits a (double-indexed) Kraus representation with matrices $\{E_{1,j}E_{0,i}\}$. Hence, as thoroughly argued in the classical case, we only need to consider the one-step situation. Besides the reference quantum evolution, we are also given an initial and a final positive definite density matrices $\rho_0$ and $\rho_T$, respectively.
We consider the problem of finding a new quantum evolution ${\mathcal F}_{0:T}^\dagger={\mathcal F}^\dagger_{T-1}\circ\cdots \circ{\mathcal F}^\dagger_1\circ{\mathcal F}^\dagger_0$, where the ${\mathcal F}^\dagger_t$ are TPCP maps, such that the new evolution is close to the reference one and
\begin{equation}\label{eq:matchmarginals}
{\mathcal F}_{0:T}^\dagger(\rho_0)=\rho_T.
\end{equation}
Providing a mathematical formulation for the selection of ${\mathcal F}_t$'s is nontrivial. In \cite{pavon2010discrete}, which deals with the simpler problem where only the initial or final density matrix is prescribed, {\em quantum paths} were introduced via a family of observables $\{X_t; 0\le t\le T\}$ with spectral decomposition
\[X_t=\sum_{i_t=1}^{m_t}x_{i_t}\Pi_{i_t}(t).
\]
Then, paths are defined as sequences of ``events'' $\left(\Pi_{i_0}(0),\Pi_{i_1}(1),\ldots,\Pi_{i_T}(T)\right)$ and thereby, one can define path-conditioned density evolutions and formulate and solve maximum entropy problems in a somewhat classical-like fashion. However, this approach does not seem to work when both initial and final densities are given, specifically because it is not clear what notion of relative entropy one should use to compare the two Markovian evolutions associated to ${\mathcal E}$ and ${\mathcal F}$.
Thus, herein we formulate the quantum Shr\"odinger bridge problem as that of seeking a suitable multiplicative functional transformation of the prior Kraus evolution so as to meet the marginal conditions. This is explained next.
Given the sequence of TPCP maps $\{{\mathcal E}^\dagger_t, \mbox{ for }0\le t\le T-1\}$ and the two marginals $\rho_0,\rho_T$ we seek a sequence of invertible matrices ${\chi}_0,\ldots,{\chi}_T$ such that for $t\in\{0,\ldots,T-1\}$,
\begin{equation}\label{eq:multtrans}
{\mathcal F}^\dagger_t(\cdot)={\chi}_{t+1}\left({\mathcal E}_t^\dagger({\chi}_t^{-1}(\cdot){\chi}_t^{-\dagger})\right){\chi}_{t+1}^\dagger
\end{equation}
is a Kraus map (i.e., totally positive and trace preserving) and the evolution
\begin{eqnarray}
{\mathcal F}_{0:T}^\dagger(\cdot)&=&{\mathcal F}^\dagger_{T-1}\circ\cdots \circ{\mathcal F}^\dagger_1\circ{\mathcal F}^\dagger_0(\cdot)\\&=&
{\chi}_T\left({\mathcal E}_{0:T}^\dagger({\chi}_0^{-1}(\cdot){\chi}_0^{-\dagger})\right){\chi}_T^\dagger\label{eq:multiplicative}
\end{eqnarray}
is consistent with the given marginals (i.e., \eqref{eq:matchmarginals} holds).
Conditions \eqref{eq:multtrans} and \eqref{eq:multiplicative} represent the quantum analogue of a multiplicative functional transformation.
Very much as in the classical case, the solution of the Shr\"odinger bridge in the multi-step case reduces to solving the one-step bridge problem:
given the triple $({\mathcal E}_{0:T}^\dagger,\rho_0,\rho_T)$, determine invertible matrices ${\chi}_0,{\chi}_T$ such that \eqref{eq:multiplicative} is Kraus map and satisfies
\begin{eqnarray}\label{eq:agreement}
{\mathcal F}_{0:T}^\dagger(\rho_0)&=&\rho_T.
\end{eqnarray}
The {\em unital} property of ${\mathcal F}$, namely the fact that ${\mathcal F}_{0:T}(I)=I$, when expressed in terms of ${\mathcal E}$ and the factors in \eqref{eq:multiplicative}, implies that
\begin{eqnarray}\label{st-harmonic}
{\mathcal E}_{0:T}(\phi_T)&=&\phi_0
\end{eqnarray}
for the Hermitian, positive definite matrices
\begin{equation}\label{FACTOR}\phi_0 = {\chi}_0^\dagger {\chi}_0,\quad
\phi_T = {\chi}_T^\dagger {\chi}_T.
\end{equation}
Likewise, in view of \eqref{eq:multtrans} and the requirement that ${\mathcal F}^\dagger_t$ is trace preserving, we set
\begin{equation}\label{FACTOR2}
{\mathcal E}_t(\phi_{t+1})=\phi_t,
\end{equation}
for $t\in\{T-1,\ldots,0\}$ and factor
\begin{equation}
\phi_t = {\chi}_t^\dagger {\chi}_t.\label{eq:spacetime}
\end{equation}
Conditions (\ref{st-harmonic}) and \eqref{eq:spacetime} indicate that the $\phi_t$'s are {\em space-time harmonic} matrices with respect to ${\mathcal E}_{0:T}^\dagger$ and ${\mathcal E}_t^\dagger$, respectively. On the other hand,
\begin{eqnarray}\label{eq:agreement2}
{\mathcal E}_{0:T}^\dagger(\underbrace{{\chi}_0^{-1}(\rho_0){\chi}_0^{-\dagger}}_{\hat\phi_0})&=&\underbrace{{\chi}_T^{-1}(\rho_T){\chi}_T^{-\dagger}}_{\hat\phi_T},
\end{eqnarray}
where the matrices $\hat\phi_t$ in \eqref{eq:agreement2} can be thought of as unnormalized density matrices in a dual Heisenberg picture.
The concept of space-time harmonic functions for quantum channels was introduced and studied in \cite{pavon2010discrete,ticozzi2010time}.
Thus, following \cite{pavon2010discrete,ticozzi2010time}, when two Kraus maps are related as in (\ref{eq:multiplicative}-\ref{st-harmonic}-\ref{FACTOR}), we say that ${\mathcal F}^\dagger$ is obtained from ${\mathcal E}^\dagger$ through a {\em multiplicative functional transformation induced by a space-time harmonic function}. In \cite{pavon2010discrete}, it was shown that the solution of the Schr\"{o}dinger problem where only the initial or final density is assigned is indeed given by a multiplicative functional transformation of the prior providing the first such example of a non-commutative counterpart of the classical results. A variational characterisation of the quantum bridge as a critical point of a suitable entropic functional for the present general setting is unknown.
\subsection{Doubly stochastic Kraus maps}\label{sec:doublystochastic}
The results we present next are for the special case where the two marginal densities are uniform, i.e., when both
\[
\rho_0=\frac{1}{n}I \mbox{ as well as }\rho_T=\frac{1}{n}I.
\]
This special case is interesting in its own right as, already for the classical case, it represents a well known basic result in the statistics literature (Sinkhorn's theorem) \cite{Sinkhorn1964,sinkhorn1974diagonal}.
Sinkhorn's result states that for any stochastic matrix $[\pi_{x_0,x_T}]_{x_0,x_T=1}^N$ {\em having all entries strictly positive}, there exists a unique multiplicative {functional} transformation
\begin{eqnarray*}
{\pi^o\hspace*{-4.5pt}}_{x_0,x_T}&=&\pi_{x_0,x_T}
\frac{\varphi(T,x_T)}{\varphi(0,x_0)}
\end{eqnarray*}
so that ${\pi^o\hspace*{-4.5pt}}\;$ is doubly stochastic, that is, it has nonnegative elements and satisfies
\[
\sum_{x_0}{\pi^o\hspace*{-4.5pt}}_{x_0,x_T}=\sum_{x_T}{\pi^o\hspace*{-4.5pt}}_{x_0,x_T}=1.
\]
{In view of \eqref{eq:transition_q0T}, in this classical setting, Sinkhorn's result is a direct corollary of Theorem
\ref{fundtheorem}.}
The property of a stochastic matrix to have all entries strictly positive corresponds to the positivity improving property \eqref{eq:strictpositivity} of a Kraus map. {We proceed to derive the quantum counterpart of Sinkhorn's result.}
\begin{thm}\label{thm:doublystochastic0}
Given a positivity improving Kraus map ${\mathcal E}_{0:T}^\dagger$, i.e., satisfying \eqref{eq:strictpositivity}, there exists a pair of observables
$\phi_0,\,\phi_T\in{\mathfrak H}_{++}$ unique up to multiplication by a positive constant {and related as in (\ref{st-harmonic})} such that,
for any factorization
\begin{eqnarray*}
\phi_0 &=& {\chi}_0^\dagger {\chi}_0,\mbox{ and}\\
\phi_T &=& {\chi}_T^\dagger {\chi}_T,
\end{eqnarray*}
the map ${\mathcal F}_{0:T}^\dagger:{\mathfrak D}\to{\mathfrak D}$ defined by
\begin{equation}\label{eq:doublystochastic2}
{\mathcal F}^\dagger(\cdot):= {\chi}_T\left({\mathcal E}_{0:T}^\dagger({\chi}_0^{-1}(\cdot){\chi}_0^{-\dagger})\right){\chi}_T^\dagger
\end{equation}
is a {\em doubly stochastic} Kraus map, in that ${\mathcal F}(I)=I$ as well as ${\mathcal F}^\dagger(I)=I$.
\end{thm}
Theorem \ref{thm:doublystochastic0} follows immediately from the following result that we establish first.
\begin{thm}\label{thm:doublystochastic}
Given a Kraus map ${\mathcal E}_{0:T}^\dagger$ satisfying \eqref{eq:strictpositivity}, there exist observables $\phi_0$, $\phi_T$ in ${\mathfrak H}_{++}$ {unique up to multiplication by a positive constant}
such that
\begin{align*}
{\mathcal E}_{0:T}(\phi_T) &= \phi_0\\
{\mathcal E}_{0:T}^\dagger(\phi_0^{-1}) &= \phi_T^{-1}.
\end{align*}
\end{thm}
\noindent{\bf Proof of Theorem \ref{thm:doublystochastic}:}
The claim in Theorem \ref{thm:doublystochastic} amounts to the existence of a unique fixed point for the following circular diagram of maps
\begin{equation}\label{eq:circular}
\begin{array}{ccccc}
&\hat\phi_0 & \overset{{\mathcal E}^\dagger_{0,T}}{\longrightarrow} & \hat\phi_T &
\\\\
\hat\phi_0= \phi_0^{-1}& \uparrow & & \downarrow &\phi_T= \hat\phi_T^{-1}\\\\
&\phi_0 &\overset{{\mathcal E}_{0,T}}{\longleftarrow} & \phi_T &
\end{array}
\end{equation}
Thus, it suffices to show that the composition map
\begin{equation}\label{eq:contractivecomposition}
{\mathcal C}\;:\;\left(\hat\phi_0\right)_{\rm starting} \overset{{\mathcal E}^\dagger_{0,T}}{\longrightarrow} \hat\phi_T\overset{(\cdot)^{-1}}{\longrightarrow}\phi_T
\overset{{\mathcal E}_{0,T}}{\longrightarrow} \phi_0
\overset{(\cdot)^{-1}}{\longrightarrow}\left(\hat\phi_0\right)_{\rm next}
\end{equation}
from ${\mathfrak H}_{++}\to{\mathfrak H}_{++}$ is contractive in the Hilbert metric. It should be noted, that ``points'' here are defined up to a scaling factor, thus, they in essence represent rays.
Once again, we use Birkhoff's theorem to determine the contraction coefficient. Here, we are working on the positive cone ${\mathfrak H}_+$. This has a nonempty interior ${\mathfrak H}_{++}$ and the partial order defined by nonnegative definiteness. Accordingly,
\begin{eqnarray*}
M(X,Y)&=&\inf\{\lambda \mid X\leq \lambda Y\}\\
&=&\max\{{\rm eig}(Y^{-1/2}XY^{-1/2})\}\\
&=& \max\{{\rm eig}(XY^{-1})\},
\end{eqnarray*}
\begin{eqnarray*}
m(X,Y)&=&\sup\{\lambda \mid \lambda y\leq x\}\\
&=& \min\{{\rm eig}(XY^{-1})\},
\end{eqnarray*}
where ${\rm eig}(\cdot)$ denotes the ``eigenvalues of,'' and in this case
the Hilbert metric is
\begin{eqnarray*}
d_H(X,Y)&=&d_H(Y^{-1/2}XY^{-1/2},I)\\
&=&\log(\kappa(XY^{-1}))
\end{eqnarray*}
where $\kappa(\cdot)$ is the ``conditioning number'' of a $Z\in{\mathfrak M}$,
\begin{eqnarray*}
\kappa(Z)&=&\frac{\max\{{\rm eig}(Z)\}}{\min\{{\rm eig}(Z)\}}.
\end{eqnarray*}
From the Birkhoff-Bushell theorem, we have that both $\|{\mathcal E}\|_H\leq 1$
as well as $\|{\mathcal E}^\dagger\|_H\leq 1$ (from linearity together with monotonicity). Furthermore,
the diameter of the range of ${\mathcal E}$,
\begin{eqnarray*}
\Delta({\mathcal E})&=&\sup\{d_H({\mathcal E}(X),{\mathcal E}(Y)) \mid X,\,Y\in{\mathfrak H}_{++}\},
\end{eqnarray*}
is finite.
To see this first note that
\[
\Delta({\mathcal E})\leq 2 \sup\{d_H({\mathcal E}(X),I) \mid X\in{\mathfrak H}_{++}\},
\]
utilizing the metric property and the fact that ${\mathcal E}(I)=I$, and then
note that $d_H({\mathcal E}(X),I)$ is invariant under scaling of $X$ by a scalar. Therefore we can restrict our attention to $X\in{\mathfrak D}$ instead, and since ${\mathcal E}(X)>0$ for all $X\in{\mathfrak D}$, by compactness of ${\mathfrak D}$,
\begin{eqnarray*}
\Delta({\mathcal E})&\leq&2\max\{d_H({\mathcal E}(X),I) \mid X\in{\mathfrak D}\}<\infty.
\end{eqnarray*}
Therefore (see Theorem \ref{BBcontraction}),
\[
\|{\mathcal E}\|_H=\tanh(\frac{1}{4}\Delta({\mathcal E}))<1.
\]
Finally, we note that the induced Hilbert-gain of the inversion
\[
Z\mapsto Z^{-1}
\]
is $1$ since
\[
d_H(X,Y)=d_H(X^{-1},Y^{-1}).
\]
We conclude that with ${\mathcal C}$ as in \eqref{eq:contractivecomposition}, $\|{\mathcal C}\|_H<1$. {Again,
contractiveness in the Hilbert metric, there exists a unique fixed ray, i.e.,
\[
\hat\phi_0=\lambda{\mathcal C}(\hat\phi_0),
\]
for some $\lambda>0$, and corresponding
\begin{eqnarray*}
\hat\phi_T&=&{\mathcal E}^\dagger_{0:T}(\hat\phi_0)\\
\phi_0&=&{\mathcal E}_{0:T}(\phi_T)
\end{eqnarray*}
while
\begin{eqnarray*}
\phi_T&=&\hat\phi_T^{-1}\mbox{ and}\\
\lambda \hat\phi_0&=&\phi_0^{-1}.\\
\end{eqnarray*}
Since,
\begin{eqnarray*}
n&=&\lambda\operatorname{\rm trace}(\hat\phi_0\phi_0)\\
&=&\lambda\operatorname{\rm trace}(\hat\phi_0{\mathcal E}_{0:T}(\phi_T))\\
&=&\lambda\operatorname{\rm trace}({\mathcal E}_{0:T}^\dagger(\hat\phi_0)\phi_T)\\
&=&\lambda\operatorname{\rm trace}(\hat\phi_T\phi_T)\\
&=&\lambda n,
\end{eqnarray*}
and therefore $\lambda=1$. This completes the proof of the theorem.}\hfill $\Box$
\noindent{\bf Proof of Theorem \ref{thm:doublystochastic0}:}
This is truly a corollary to Theorem \ref{thm:doublystochastic}. Since $\phi_0,\,\phi_T$ in the proof of Theorem \ref{thm:doublystochastic} are in ${\mathfrak H}_{++}$, we take any factorization
\begin{eqnarray*}
\phi_0 &=& {\chi}_0^\dagger {\chi}_0,\mbox{ and}\\
\phi_T &=& {\chi}_T^\dagger {\chi}_T.
\end{eqnarray*}
Then, with ${\mathcal F}^\dagger$ as in \eqref{eq:doublystochastic2},
\begin{eqnarray*}
{\mathcal F}_{0:T}^\dagger(I)&=& {\chi}_T\left[{\mathcal E}_{0:T}^\dagger\left({\chi}_0^{-1}(I){\chi}_0^{-\dagger}\right)\right]{\chi}_T^\dagger\\
&=&{\chi}_T\left[{\mathcal E}_{0:T}^\dagger(\phi_0^{-1})\right]{\chi}_T^\dagger\\
&=&{\chi}_T\left(\phi_T^{-1}\right){\chi}_T^\dagger\\
&=&I.
\end{eqnarray*}
Likewise,
\begin{eqnarray*}
{\mathcal F}_{0:T}(I)&=& {\chi}_0^{-\dagger}\left[{\mathcal E}_{0:T}\left({\chi}_T^{\dagger}(I){\chi}_T\right)\right]{\chi}_0^{-1}\\
&=&{\chi}_0^{-\dagger}\left[{\mathcal E}_{0:T}(\phi_T)\right]{\chi}_0^{-1}\\
&=&{\chi}_0^{-\dagger}\left(\phi_0\right){\chi}_0^{-1}\\
&=&I.
\end{eqnarray*}
Thus, ${\mathcal F}^\dagger$ is doubly stochastic as claimed.
\hfill $\Box$
While in the classical case (Sinkhorn's theorem) there is a unique doubly stochastic map obtained from $\pi_{x_0,x_T}$ via a multiplicative {functional} transformation, in the quantum case this is clearly false. Uniqueness in Theorem \ref{thm:doublystochastic0} is claimed for the observables $\phi_0,\phi_T$. Hence, ${\mathcal F}$ is unique modulo corresponding unitary factors; obviously, if ${\mathcal F}^\dagger$ is a Kraus so that ${\mathcal F}(I)=I$ and ${\mathcal F}^\dagger(I)=I$, then for any unitary matrices $U_0,U_T$,
\begin{equation}\label{eq:nonuniqueness}
U_T{\mathcal F}^\dagger(U_0(\cdot)U_0^\dagger)U_T^\dagger
\end{equation}
is also a doubly stochastic Kraus map. {It is easy to see that the totality of doubly stochastic Kraus maps that relate to ${\mathcal E}^\dagger$ via a multiplicative transformation are of this form.}
\subsection{The quantum bridge for general marginals and a conjecture}\label{conjecture}
Consider now the situation of the previous section with general initial and final density matrices $\rho_0$ and $\rho_T$. We are namely seeking a quantum bridge, as defined in Section \ref{sec:quantumbridge}, for the triple $({\mathcal E}_{0:T}^\dagger,\rho_0,\rho_T)$ . Theorem \ref{thm:doublystochastic0} admits the following generalization:
\begin{thm}\label{quantumbridge}
Given a Kraus map ${\mathcal E}_{0:T}^\dagger$ and two density matrices $\rho_0$ and $\rho_T$, suppose there exist observables
$\phi_0,\,\phi_T,\,\hat\phi_0,\,\hat\phi_T\in{\mathfrak H}_{++}$ solving the Schr\"{o}dinger system
\begin{eqnarray}\label{SchrSyst1}
{\mathcal E}_{0:T}(\phi_T) &=& \phi_0,\\
{\mathcal E}_{0:T}^\dagger(\hat{\phi}_0) &=& \hat{\phi}_T,\label{SchrSyst2}\\\rho_0&=&{\chi}_0 \hat\phi_0{\chi}_0^\dagger,\label{SchrSyst3a}\\
\rho_T&=&{\chi}_T \hat\phi_T {\chi}_T^\dagger.\label{SchrSyst3b}
\end{eqnarray}
Then, for any factorization
\begin{eqnarray*}
\phi_0 &=& {\chi}_0^\dagger {\chi}_0,\mbox{ and}\\
\phi_T &=& {\chi}_T^\dagger {\chi}_T,
\end{eqnarray*}
the map ${\mathcal F}_{0:T}^\dagger:{\mathfrak D}\to{\mathfrak D}$ defined by
\begin{equation}\label{eq:doublystochastic2a}
{\mathcal F}^\dagger(\cdot):= {\chi}_T\left({\mathcal E}_{0:T}^\dagger({\chi}_0^{-1}(\cdot){\chi}_0^{-\dagger})\right){\chi}_T^\dagger
\end{equation}
is a quantum bridge for $({\mathcal E}_{0:T}^\dagger,\rho_0,\rho_T)$, namely ${\mathcal F}(I)=I$ and ${\mathcal F}^\dagger(\rho_0)=\rho_T$.
\end{thm}
\noindent{\bf Proof:}
$$
{\mathcal F}_{0:T}(I)= {\chi}_0^{-\dagger}\left[{\mathcal E}_{0:T}\left({\chi}_T^{\dagger}(I){\chi}_T\right)\right]{\chi}_0^{-1}={\chi}_0^{-\dagger}\left({\mathcal E}_{0:T}(\phi_T)\right){\chi}_0^{-1}={\chi}_0^{-\dagger}\left(\phi_0\right){\chi}_0^{-1}=I.
$$
Moreover,
$$
{\mathcal F}_{0:T}^\dagger(\rho_0)= {\chi}_T\left[{\mathcal E}_{0:T}^\dagger\left({\chi}_0^{-1}\rho_0{\chi}_0^{-\dagger}\right)\right]{\chi}_T^\dagger={\chi}_T\left[{\mathcal E}_{0:T}^\dagger(\hat\phi_0)\right]{\chi}_T^\dagger={\chi}_T\left(\hat\phi_T\right){\chi}_T^\dagger=\rho_T.
$$
\hfill $\Box$
\noindent
The case of initial and final pure states is worthwhile writing out.
\begin{cor}\label{purestates}
Given a positivity improving Kraus map ${\mathcal E}_{0:T}^\dagger$ and two pure states
\[
\rho_0=v_0v_0^\dagger \mbox{ and } \rho_T=v_Tv_T^\dagger
\]
(i.e., $v_0,v_T$ are unit norm vectors),
define
\begin{align*}
\phi_0&:={\mathcal E}(v_Tv_T^\dagger)\\
\phi_T&:=v_Tv_T^\dagger,
\end{align*}
and
\[
{\mathcal F}^\dagger(\cdot):= \phi_T^{1/2}{\mathcal E}^\dagger(\phi_0^{-1/2}(\cdot)\phi_0^{-1/2})\phi_T^{1/2}
\]
(where, clearly, $\phi_T^{1/2}=\phi_T=v_Tv_T^\dagger$). Then, ${\mathcal F}^\dagger$ is TPTP and satisfies the marginal conditions
\begin{align*}
\rho_T={\mathcal F}^\dagger(\rho_0).
\end{align*}
\end{cor}
\noindent{\bf Proof:} We readily verify that
\begin{align*}
{\mathcal F}(I)&=\phi_0^{-1/2}{\mathcal E}(\phi_T)\phi_0^{-1/2}\\
&=\phi_0^{-1/2}{\mathcal E}(v_Tv_T^\dagger)\phi_0^{-1/2}\\
&=I.
\end{align*}
Next, we observe that
\begin{align*}
{\mathcal F}^\dagger(\rho_0)&=\phi_T^{1/2}{\mathcal E}^\dagger(\phi_0^{-1/2}(v_0v_0^\dagger)\phi_0^{-1/2})\phi_T^{1/2}\\
&=v_Tv_T^\dagger{\mathcal E}^\dagger(\phi_0^{-1/2}(v_0v_0^\dagger)\phi_0^{-1/2})v_Tv_T^\dagger\\
&=v_Tv_T^\dagger\\
&=\rho_T.
\end{align*}
To see this, consider a representation ${\mathcal E}^\dagger(\cdot)=\sum_i E_i(\cdot)E_i^\dagger$ and note that
\begin{align*}
v_T^\dagger{\mathcal E}^\dagger(\phi_0^{-1/2}(v_0v_0^\dagger)\phi_0^{-1/2})v_T&=\sum_i
\left(v_T^\dagger E_i \phi_0^{-1/2}v_0\right)^2\\
&=v_0^\dagger\phi_0^{-1/2}{\mathcal E}(v_Tv_T^\dagger)\phi_0^{-1/2}v_0\\
&=v_0^\dagger I v_0\\
&=1.
\end{align*}
\hfill $\Box$
Thus, precisely as in the classical case, the key challenge is to establish existence and uniqueness for the Schr\"{o}dinger system (\ref{SchrSyst1}-\ref{SchrSyst3b}). At present, proving the natural generalization of Theorem \ref{thm:doublystochastic} appears nontrivial.
Thus, below, we give the relevant statement as a conjecture since its proof remains elusive.
\begin{conj}\label{conj:doublystochastic}
Given a positivity-improving Kraus map ${\mathcal E}_{0:T}^\dagger$, i.e., a Kraus map satisfying \eqref{eq:strictpositivity}, and given two density matrices $\rho_0$ and $\rho_T$, there exist observables $\phi_0$, $\phi_T$, $\hat{\phi}_0$, $\hat{\phi}_T$ in ${\mathfrak H}_{++}$ such that
\begin{align*}
{\mathcal E}_{0:T}(\phi_T) &= \phi_0\\
{\mathcal E}_{0:T}^\dagger(\hat{\phi}_0) &= \hat{\phi}_T
\end{align*}
with
\begin{subequations}
\begin{align}\label{eq:conj:a}
\rho_0&={\chi}_0 \hat\phi_0{\chi}_0^\dagger\\
\rho_T&={\chi}_T \hat\phi_T {\chi}_T^\dagger\label{eq:conj:b}
\end{align}
and
\begin{align}\label{eq:conj:c}
\phi_0 &= {\chi}_0^\dagger {\chi}_0\\
\phi_T &= {\chi}_T^\dagger {\chi}_T.\label{eq:conj:d}
\end{align}
\end{subequations}
In particular, ${\chi}_0,{\chi}_T$ can be taken to be
Hermitian, i.e., for $i\in\{0,T\}$, ${\chi}_i=(\phi_i)^{1/2}$ is the Hermitian square root of $\phi_i$.
\end{conj}
Equations \eqref{eq:conj:a} and \eqref{eq:conj:c} together, represent a non-commutative analogue of the relationship between $\hat\varphi(0,x_0),{\mathbf p}_0(x_0),\varphi(0,x_0)$ in
\[
\hat\varphi(0,x_0)= \frac{{\mathbf p}_0(x_0)}{\varphi(0,x_0)},
\]
while equations \eqref{eq:conj:b} and \eqref{eq:conj:d} represent the analogue of
\[
\hat\varphi(T,x_T)= \frac{{\mathbf p}_T(x_T)}{\varphi(T,x_T)}.\]
By taking ${\chi}_0=(\phi_0)^{1/2}$, i.e., the Hermitian square root, clearly
\[
\hat\phi_0= (\phi_0)^{1/2}\rho_0(\phi_0)^{1/2}.
\]
On the other, taking ${\chi}_T$ to be the Hermitian square root of $\phi_T$ and solving for $\phi_T$ in terms of $\hat\phi_T$ and $\rho_T$ using (\ref{eq:conj:b}-\ref{eq:conj:d}) gives
\[
\hat\phi_T\mapsto \phi_T=\left(
\rho_T^{1/2}\left(\rho_T^{-1/2}\hat\phi^{-1}\rho_T^{-1/2}
\right)^{1/2}\rho_T^{1/2}
\right)^2
\]
Thus, the conjecture claims the validity of the following correspondence,
\[
\begin{array}{ccccccc}
\rho_0&
\overset{{\chi}_0^{-1}(\cdot){\chi}_0^{-\dagger}
}{\longrightarrow}&{\hat\phi_0} & \overset{{\mathcal E}_{0:T}^\dagger}{\longrightarrow} & \hat\phi_T & \overset{{\chi}_T(\cdot){\chi}_T^{\dagger}
}{\longrightarrow} & \rho_T \\\\
\\
I & \overset{{\chi}_0^{-\dagger}(\cdot){\chi}_0^{-1}}{\longleftarrow} &\phi_0={\chi}_0^\dagger {\chi}_0 &\overset{{\mathcal E}_{0:T}}{\longleftarrow} & \phi_T={\chi}_T^\dagger {\chi}_T &\overset{{\chi}_T^{\dagger}(\cdot){\chi}_T
}{\longleftarrow} & I
\end{array}
\]
and therefore, that the Kraus map
\[
{\mathcal F}_{0:T}^\dagger(\cdot):={\chi}_T\left({\mathcal E}_{0:T}^\dagger({\chi}_0^{-1}(\cdot){\chi}_0^{-\dagger})\right){\chi}_T^{\dagger}
\]
solves the one-step quantum Sch\"odinger bridge problem for general marginal density matrices.
Extensive simulations have convinced the authors that the composition of maps
\begin{equation}\label{eq:composition:quantum}
\hat\phi_0 \overset{{\mathcal E}_{0:T}^\dagger}{\longrightarrow} \hat\phi_T\overset{D_{T}}{\longrightarrow}\phi_T
\overset{{\mathcal E}_{0:T}}{\longrightarrow} \phi_0
\overset{{\hat D}_0}{\longrightarrow}\left(\hat\phi_0\right)_{\rm next}
\end{equation}
where
\begin{eqnarray*}
D_T&:& \hat\phi_T\mapsto \phi_T=\left(
\rho_T^{1/2}\left(\rho_T^{-1/2}\hat\phi^{-1}\rho_T^{-1/2}
\right)^{1/2}\rho_T^{1/2}
\right)^2\\
\hat{D}_0&:& \phi_0\mapsto \hat\phi_0=(\phi_0)^{1/2}\rho (\phi_0)^{1/2}
\end{eqnarray*}
has an attractive fixed point \footnote{In this, ${\chi}_i$ for $i\in\{0,T\}$ are taken Hermitian for specificity. Simulation shows that the iteration converges to a fixed point for a variety of other normalizations for the factors ${\chi}_i$ of $\phi_i$, as well as when the boundary conditions are replaced by $\rho_0=(\hat\phi_0)^{1/2}\phi_0(\hat\phi_0)^{1/2}$ and $\rho_T=(\hat\phi_T)^{1/2}\phi_T(\hat\phi_T)^{1/2}$ and the iteration is modified accordingly.}. Unfortunately, $\hat D_0$ and $D_T$ are not isometries in the Hilbert metric as their commutative analogues.
{For a bridge with many intermediary time steps, very much as in the commutative classical case, the solution of the one-step bridge via multiplicative functional transformation permits solving the general bridge in a similar manner.} More specifically, starting from $\hat\phi_0,\phi_T$ that correspond to a fixed point of \eqref{eq:composition:quantum}, define for $i\in\{1,\ldots,T\}$
\begin{align*}
\hat\phi_i &:= {\mathcal E}_{i-1}^\dagger\hat\phi_{i-1}\\
\phi_{i-1} &:= {\mathcal E}_{i-1}\phi_i\\
\rho_i &:= (\phi_i)^{1/2}\hat\phi_i(\phi_i)^{1/2}.
\end{align*}
{The sought sequence of Kraus maps is
\[
{\mathcal F}_{i+1}^\dagger(\cdot)=(\phi_{i+1})^{1/2}\left({\mathcal E}_i^\dagger( (\phi_i)^{-1/2}(\cdot)(\phi_i)^{-1/2} )\right)(\phi_{i+1})^{1/2},
\]
since, clearly,
\[
{\mathcal F}_{0:T}^\dagger(\cdot)=(\phi_{T})^{1/2}\left({\mathcal E}_{0:T}^\dagger( (\phi_0)^{-1/2}(\cdot)(\phi_0)^{-1/2} )\right)(\phi_{T})^{1/2}
\]
and, assuming the validity of the conjecture, satisfies
\begin{eqnarray*}
{\mathcal F}_{0:T}^\dagger(\rho_0)&=& (\phi_{T})^{1/2}\left({\mathcal E}_{0:T}^\dagger( \hat\phi_0 )\right)(\phi_{T})^{1/2}\\
&=&(\phi_{T})^{1/2} \hat\phi_T (\phi_{T})^{1/2}\\
&=&\rho_T.
\end{eqnarray*}}
\section{Examples of doubly stochastic Kraus maps}
Perhaps the simplest nontrivial example of a (self-adjoint) positivity improving doubly stochastic Kraus map is
\begin{equation}\label{ex:kraus}
{\mathcal E}^\dagger(\cdot)=E_1(\cdot)E_1^\dagger+E_2(\cdot)E_2^\dagger+E_3(\cdot)E_3^\dagger
\end{equation}
with
\[
E_1=\left[\begin{matrix}\sqrt{\frac{1}{2}} &0\\0&0\end{matrix}\right],\;
E_2=\left[\begin{matrix}0 &0\\0&\sqrt{\frac{1}{2}}\end{matrix}\right],\;
E_3=\left[\begin{matrix}0 &\sqrt{\frac{1}{2}}\\\sqrt{\frac{1}{2}}&0\end{matrix}\right].
\]
A second example is
\[
E_1=\left[\begin{matrix}\phantom{-}\sqrt{\frac{2}{3}} & \phantom{xxxx}0\\ \phantom{xxx}0 & \phantom{xxxx}0\end{matrix}\right],\;
E_2=\left[\begin{matrix}\phantom{-}\sqrt{\frac{1}{24}} & -\sqrt{\frac{1}{8}}\\ -\sqrt{\frac{1}{8}} & \phantom{-}\sqrt{\frac{3}{8}}\end{matrix}\right],\;
E_3=\left[\begin{matrix}\phantom{-}\sqrt{\frac{1}{24}} & \phantom{-}\sqrt{\frac{1}{8}}\\ \phantom{-}\sqrt{\frac{1}{8}} & \phantom{-}\sqrt{\frac{3}{8}}\end{matrix}\right]
\]
where all coefficient matrices are again symmetric but they all now of rank one.
In general, neither the symmetry (Hermitian-ness) of the coefficients nor any constraint on the rank is essential. The following example is constructed numerically. For this, we
start with
\begin{eqnarray*}
E_1&=&\left[\begin{matrix}1 & 1\\ 0 & 0\end{matrix}\right] M^{-1/2}\\
E_2&=&\left[\begin{matrix}0 & 1 \\ 0 & 1\end{matrix}\right] M^{-1/2}\\
E_2&=&\left[\begin{matrix}0 & 1 \\ 1 & 0\end{matrix}\right]M^{-1/2}
\end{eqnarray*}
where
\[M=\left[\begin{matrix}2&1\\1&4\end{matrix}\right]
\]
so that $E_1'E_1+E_2'E_2+E_3'E_3=I$.
It can be shown that the corresponding Kraus map ${\mathcal E}^\dagger$ is positivity improving, i.e., it satisfies \eqref{eq:strictpositivity}.
We then compute the fixed point of \eqref{eq:contractivecomposition}, and this is
\begin{eqnarray*}
\phi_0 &=&
\left[\begin{matrix}\phantom{-}1.1448 & -0.1350\\ -0.1350 & \phantom{-}0.8749\end{matrix}\right]\\
\phi_1 &=&
\left[\begin{matrix}\phantom{-}0.8411 & -0.2362\\ -0.2362 & \phantom{-}1.3134\end{matrix}\right]\\
\hat\phi_0 &=&
\left[\begin{matrix}\phantom{-}0.8897 & \phantom{-}0.1372\\ \phantom{-}0.1372 & \phantom{-}1.1642\end{matrix}\right]\\
\hat\phi_1 &=&
\left[\begin{matrix}\phantom{-}1.2521 & \phantom{-}0.2251\\ \phantom{-}0.2251 & \phantom{-}0.8018\end{matrix}\right].
\end{eqnarray*}
These space-time harmonics give rise to coefficients
\begin{eqnarray*}
F_1&=&\left[\begin{matrix}\phantom{-}0.5690 & \phantom{-}0.4411\\ -0.0720 & -0.0558\end{matrix}\right]\\
F_2&=&\left[\begin{matrix}-0.0558 & \phantom{-}0.4411 \\ -0.0720 & \phantom{-}0.5690\end{matrix}\right]\\
F_3&=&\left[\begin{matrix}-0.1441 & \phantom{-}0.5131 \\ \phantom{-}0.8013 & -0.1441\end{matrix}\right]
\end{eqnarray*}
for a corresponding doubly stochastic Kraus map.
When $\rho_0$ and/or $\rho_T$ are in general different from the identity, extensive numerical experimentation suggests the validity of Conjecture \ref{conj:doublystochastic} which, in conjuction with Theorem \ref{quantumbridge}, provides
solutions of the quantum Schr\"odinger bridge problem. In particular, for our first example \eqref{ex:kraus} and nonuniform marginals
\[\rho_0=\left[\begin{matrix}1/4 & 0\\ 0& 3/4\end{matrix}\right] \mbox{ and }
\rho_1=\left[\begin{matrix}2/3 & 0\\ 0& 1/3\end{matrix}\right]
\]
by iterating \eqref{eq:composition:quantum} we obtain
\begin{eqnarray*}
\phi_0 &=&
\left[\begin{matrix}1/2 & 0\\ 0 & 1/2\end{matrix}\right]\\
\phi_1 &=&
\left[\begin{matrix}2/3 & 0\\ 0 & 1/3\end{matrix}\right]\\
\hat\phi_0 &=&
\left[\begin{matrix}1/2 & 0\\ 0 & 3/2\end{matrix}\right]\\
\hat\phi_1 &=&
\left[\begin{matrix}1 & 0\\ 0 & 1\end{matrix}\right]
\end{eqnarray*}
and
the Kraus map with coefficients
\begin{eqnarray*}
F_1&=&\left[\begin{matrix}\sqrt{2/3} & \phantom{xx}0\phantom{xx}\\\phantom{xx}0\phantom{xx} & \phantom{xx}0\phantom{xx}\end{matrix}\right]\\
F_2&=&\left[\begin{matrix}\phantom{xx}0\phantom{xx} &\phantom{xx}0\phantom{xx} \\ \phantom{xx}0\phantom{xx}& \sqrt{1/3}\end{matrix}\right]\\
F_3&=&\left[\begin{matrix} \phantom{xx}0\phantom{xx}& \sqrt{2/3} \\ \sqrt{1/3} & \phantom{xx}0\phantom{xx} \end{matrix}\right].
\end{eqnarray*}
This Kraus map is no longer Hermitian, it is of the form \eqref{eq:doublystochastic2} and, as can be readily verified, it satisfies the required condition ${\mathcal F}^\dagger(\rho_0)=\rho_1$.
Software for numerical experimentation can be found at \url{http://www.ece.umn.edu/~georgiou/papers/schrodinger_bridge/}.
\section{Concluding remarks}
{In this paper, we introduced a new approach to studying Schr\"{o}dinger's systems.
In particular, we establish new proofs for existence and uniqueness of solutions.
In contrast to earlier treatments, our approach provides a direct computational procedure for obtaining the space-time harmonic function and the corresponding solution of the Schr\"odinger bridge in finite-dimensions. Space-time harmonics are obtained as fixed points of a certain map. Convergence is established in a suitable {\em projective geometry} as convergence of {\em rays} using the Hilbert metric. That is, in the classical case of discrete random vectors that is treated herein, our approach provides a direct new proof of existence and uniqueness by a contraction mapping principle. Since the approach also provides a computational scheme, it appears to have considerable potential for applications. Indeed, models for stochastic evolution, which include Markov chains, are ubiquitous. The bridge evolution may be viewed as a controlled {\em steering} problem --a facet that will be thoroughly explored elsewhere. In the case of quantum channels, the solution of an analogous Schr\"{o}dinger system corresponds to a steering between two given density matrices. We prove convergence of an iterative algorithm in a corresponding projective metric in the case of uniform marginals, i.e., identity matrices, thereby establishing existence of doubly stochastic Kraus maps that can be derived from a given reference map via multiplicative functional transformations. Extensive simulations have convinced the authors of a more general result regarding the general quantum bridge problem which, however, is stated as a conjecture since at present a proof is not available.}
\section*{Acknowledgment}
The authors are very grateful to Yongxin Chen for a correction in the proof of Theorems 3 and 6.
\bibliographystyle{siam}
|
\section{\label{sec:intro}Introduction and Summary}
It has been a matter of lively debate whether the standard description
of a large black hole with a smooth horizon is quantum mechanically
consistent, and is, in fact, consistent with AdS/CFT. While the
firewall hypothesis \cite{Almheiri:2012rt, Almheiri:2013hfa}
\footnote{See also \cite{Braunstein:2009my}.} argues
against the validity of the standard description, Maldacena and
Susskind \cite{Maldacena:2013xja} have suggested that the region inside
the horizon is a geometric representation of quantum mechanical
entanglement. Both the above proposals, and related issues, are
discussed in a number of papers; for a partial list, related to the
discussion in this paper, see
\cite{Almheiri:2012rt,Almheiri:2013hfa,Shenker:2013pqa,
VanRaamsdonk:2013sza,Marolf:2013dba,
Papadodimas:2013jku,Papadodimas:2013wnh,Shenker:2013yza,
Avery:2013bea,Balasubramanian:2014gla}. The proposal of
\cite{Maldacena:2013xja}, summarized by the symbolic equation ER =
EPR, \footnote{Einstein-Rosen (wormhole) = Einstein-Podolsky-Rosen
(entangled state).} is illustrated by the eternal black hole
geometry which is dual to the thermofield state
\cite{Maldacena:2001kr}.\footnote{See \cite{Hartman:2013qma} for an
AdS/CFT check on the dynamical entanglement entropy which involves
the wormhole region, and \cite{Caputa:2013eka} for generalization to
include angular momentum and charge.} It has been argued in several
papers (see, e.g., \cite{Marolf:2013dba, Balasubramanian:2014gla})
that although the proposal holds for this illustrative case, it does
not hold in general. One of the objectives of the present work is to
explicitly construct a general class of two-sided
geometries \footnote{By {\it two-sided}, we mean geometries which have
two asymptotically AdS regions.} which represent entangled CFT's.
A useful approach to construct the geometric dual to a CFT state is by
using a Fefferman-Graham (FG) expansion, with boundary data provided
by the CFT state. To begin with, let us consider the case of a single
CFT. Since we are primarily interested in the metric, let us focus,
for simplicity, on states in which only the stress tensor is excited.
The dual geometry would then be given by the solution to the
appropriate Einstein equations subject to the boundary data provided
by the stress tensor. This approach has been particularly fruitful in
the context of the AdS$_3$/CFT$_2$ duality where the Fefferman-Graham
expansion has been shown, for pure gravity, to terminate
\cite{Banados:1998gg} , yielding the following exact
metric \footnote{In \eq{banados}, $x_\pm= t\pm x$, with $x \in
{\mathbb R}$. For $L, \bar L$ constant, this corresponds to the BTZ
black string.}
\begin{align}
ds^2 = \frac{dz^2}{z^2} - dx_+ dx_-\left(\frac1{z^2} +
z^2 \frac{L(x_+) {\bar L}(x_-)}{16} \right) +\frac14 \left(L(x_+) dx_+^2 +
{\bar L}(x_-) dx_-^2 \right)
\label{banados}
\end{align}
The boundary data ($z\to 0$) is represented by the following
holographic stress tensors (we choose $-\Lambda =1/\ell^2=1$)
\begin{align}
8\pi G_3 T_{++}(x_+)= \frac{L(x_+)}4, \,
8\pi G_3 T_{--}(x_-)= \frac{{\bar L}(x_-)}4
\label{banados-stress}
\end{align}
The above metric becomes singular at the horizon
\begin{equation}
z= z_0 \equiv 2\left(L(x_+) {\bar L}(x_-)\right)^{-1/4},
\label{banados-horizon}
\end{equation} and therefore the metric \eq{banados}, describes only an exterior
geometry. \footnote{The inverse metric $g^{MN}$ blows up at the
horizon, as in case of Schwarzschild geometry. However, unlike
there, here the other region $z>z_0$ does {\it not} represent the
region behind the horizon; rather it gives a second coordinatization
of the exterior region again. In this paper, we will use a
different set of coordinate systems to probe the interior and a
second exterior region.}
How does one carry out such a construction with two boundaries, with
two sets of boundary data? Indeed, it is not even clear, {\it a
priori}, whether simultaneously specifying two independent pieces of
boundary data can always lead to a consistent solution in the bulk
(this question has been raised in several recent papers, e.g.
see \cite{VanRaamsdonk:2013sza}). A possible approach to this problem
is suggested by the fact that the eternal BTZ solution, which contains
\eq{banados} with constant stress tensors, admits a maximal extension
with two exteriors, which are joined to an interior region across a
smooth horizon. The maximal extension is constructed by transforming,
e.g., to various Eddington-Finkelstein (EF) coordinate patches
(described in Appendix \ref{charts}). A naive generalization of such a
procedure in case of variable $L, {\bar L}$, of transforming the metric
\eq{banados} to EF type coordinates, does not seem to work since it
leads to a complex metric in the interior region \footnote{Such a
coordinate transformation has been discussed in \cite{Gupta:2008th}
in an asymptotic series near the boundary.}. A second approach
could be to solve Einstein's equations, by using the constant $L, {\bar L}$
(eternal BTZ) solution as a starting point and, incorporate the effect
of variable $L, {\bar L}$ perturbatively, either in a derivative expansion
or an amplitude expansion. While this method may indeed work, at the
face of it, it is far from clear how the variation in $L, {\bar L}$ can be
chosen to be different at the two boundaries.
In this paper, we will use the method of solution generating
diffeomorphisms (SGD). In gauge theory terms, these are asymptotically
nontrivial gauge transformations which correspond to global charge
rotations; the use of these objects was introduced in
\cite{Regge:1974zd,Wadia:1979yu, Gervais:1976ec}, and used crucially
by Brown and Henneaux\cite{Brown:1986nw} to generate `Virasoro
charges' through asymptotically nontrivial SGDs that reduced at the
AdS boundary to conformal transformations. (We discuss these in more
detail in Section \ref{behind}). Brown and Henneaux had discussed only
the asymptotic form of the SGDs. We apply two independent, exact
Brown-Henneaux SGDs \footnote{\label{ftnt:gauge} It has been shown by
Roberts \cite{Roberts:2012aq} that the exterior metric \eq{banados}
can be obtained by an exact Brown-Henneaux type diffeomorphism
applied to the Poincare metric. See Appendix \ref{roberts} for a
discussion on this and a different, new, transformation which is
closer to the ones we use in this paper.} to different coordinate
patches of the eternal BTZ geometry, yielding a black hole spacetime
with two completely general stress tensors on the two boundaries. In
other words, our strategy for solving the boundary value problem can
be summarized as: given arbitrary boundary data in terms of stress
tensors $T_R, \bar T_R$, and $T_L, \bar T_L$, we (i) find the two
specific sets of conformal transformations (which we are going to call
$G_+, G_-$ and $H_+, H_-$) which, when acting on a constant stress
tensor, gives rise to these stress tensors, (ii) find the SGD's which
reduce to these conformal transformations and (iii) apply the SGD's to
the eternal BTZ metric.
\gap1
\noindent{\it This solves the boundary value problem we
posed above.}
\gap1
The results in this paper are organized as follows:
(1) \underbar{The new solutions}: In Section \ref{behind} we describe
the explicit {\it solution generating diffeomorphisms} (SGDs) and
construct the resulting two-sided black hole geometries. The
diffeomorphisms reduce to conformal transformations at each boundary,
parameterized by functions $G_\pm$ on the right and $H_\pm$ on the
left. The SGD parameterized by $G_\pm$ is applied to the
Eddington-Finkelstein coordinate chart EF1 (which covers the right
exterior and the black hole interior, see Figure \ref{fig-ef1234}) and
to EF4 (right exterior + white hole interior), whereas the SGD
parameterized by $H_\pm$ is applied to the Eddington-Finkelstein
coordinate chart EF2 (left exterior + black hole interior) and to EF3
(left exterior + white hole interior). To cover the entire spacetime
we also use a Kruskal chart K5 which covers an open neighbourhood of
the bifurcate Killing horizon; here we leave the original Kruskal
metric unaltered. The effect of the above SGDs is that we have a
description of different metric tensors in different charts. In
Section \ref{full-metric} we show that all these can be pieced
together to give a single (pseudo-)Riemannian manifold; we prove this
by showing that in the pairwise overlap of any two charts $N_1 \cap
N_2$ the different metrics constructed above differ only by a trivial
diffeomorphism (see the definition \ref{def-nontrivial}); the
full metric, specified with the help of the various
charts, is schematically represented in Figure \ref{fig-nontrivial}.
An important
manifestation of the asymptotic nontriviality of the SGDs is to move
and warp the infra-red regulator surface (see Figure \ref{fig-ef});
the change in the boundary properties, as found in later sections, can
be directly attributed to this.
The new spacetime so constructed inherits the original causal
structure, with the event horizon, the bifurcation surface, and the
two exterior and interior regions (see also footnotes \ref{inner-a}
and \ref{inner}). The horizon is, therefore, regular by
construction. In the new EF coordinates (the {\it tilded} coordinates)
the horizon consists of smoothly undulating surfaces (see Fig
\ref{fig-horizons}).
(2) \underbar{The CFT duals}: In section \ref{CFT-dual} we use the
observation that the SGDs reduce asymptotically to conformal
transformations to assert that the CFT duals to our geometries are
given by unitary transformations $U_L \otimes U_R$ to the thermofield
double state. Since the unitary transformations implement conformal
transformations, AdS/CFT implies that CFT correlators in the
transformed state are holographically computed by the new spacetime
geometry. We posit this as a test of the proposed AdS/CFT
correspondence.
(3) \underbar{The AdS/CFT checks}: In section \ref{sec-T} we carry out
this test for the stress tensor. We compute the holographic stress
tensor \cite{Balasubramanian:1999re,Skenderis:1999nb} in the new
geometry and show that it exactly matches with the expectation value
of the conformally transformed (including the Schwarzian derivative)
stress tensor in the thermofield double state. In section \ref{2-pt}
we compare AdS and CFT results for both $\langle O_L O_R \rangle$ and $\langle
O_R O_R \rangle$ types of correlators. The holographic two-point function
is found by computing geodesic lengths in the new geometries and we
find that it correctly matches with the two-point function of
transformed operators. This can be regarded as an evidence for the
ER=EPR relation in the presence of probes.
(4) \underbar{Entanglement entropy}: As a further check, in section
\ref{entang} we apply the above result for two-point functions to show
that the entanglement entropy EE in CFT matches the holographic EE
\cite{Ryu:2006bv,Hubeny:2007xt} including when the Ryu-Takayanagi
geodesic passes through the wormhole. This constitutes a direct proof
of the ER=EPR conjecture for the entire class of geometries constructed in this
paper. We work out the dynamical entanglement entropy in an example
(see fig \ref{HEEplot}).
(5) \underbar{Holographic entropy from horizon}: In section
\ref{sec-S}, we make crucial use of the existence of smooth horizons
on both sides to compute a holographic entropy along the lines of
\cite{Bhattacharyya:2008xc}. We are able to compute the entropy in the
CFT by using the Cardy formula and an adiabatic limit (which allows
the use of the `instantaneous' energy eigenvalues to compute
degeneracies); the holographic entropy agrees with this. The entropy
turns out to be divergenceless, reflecting the dissipationless nature
of 2D CFT. There is, however, a nontrivial local flow of entropy (see
fig \ref{entropy}).
(6) \underbar{ER=EPR}: In Section \ref{discuss} we discuss some
implications of our solutions {\it vis-a-vis} the ER=EPR relation of
Maldacena and Susskind \cite{Maldacena:2013xja}. Our solutions
establish an infinite family of quantum states entangling two CFTs
which are represented in the bulk by wormhole geometries. We show, in
particular, that out of a given set of quantum states we consider, all
characterized by the same energy, there are states with low
entanglement entropies, which nevertheless are still represented by
wormhole geometries; this is in keeping with the picture of geometric
entanglement suggested in \cite{Maldacena:2013xja}.
\section{The solutions\label{behind}}
In this section we obtain the new solutions by carrying out the
procedure outlined in the Introduction. As explained in Section
\ref{charts}, for constant $L, {\bar L}$, the metric \eq{banados}
represents a BTZ black hole of constant mass and angular momentum
\eq{btz-mass-ang}. In that case, one can construct EF coordinates (see
Section \ref{charts}) to extend the spacetime to include the region
behind the horizon and a second exterior. We will, in fact, use five
charts to cover the extended geometry (see Fig \ref{fig-ef1234}).
\gap{-3}
\begin{figure}[H]
\centerline{\includegraphics[width=550pt,height=150pt]{ef1234}}
\caption{The (green parts of) the five figures on the right depict the
five coordinate charts used in this paper to cover the eternal BTZ
solution.\protect\footnotemark The coordinate chart K5 is needed
to cover the ``bifurcation surface'' where the past and future
horizons meet (it is a point in the Penrose diagram). The leftmost
diagram (in blue) represents the coordinate chart used in
\eq{banados}. Each of the coordinate charts is shown, for facility
of comparison, within a Penrose diagram where the parts not within
the chart are shown in gray.}
\label{fig-ef1234}
\end{figure}
\footnotetext{\label{inner-a} This is the entire geometry for
the non-spinning BTZ; for spinning BTZ solutions, we do not
attempt to cover the region beyond the inner horizon, since in
this paper we are interested in the asymptotic properties in the
two exteriors mentioned above. See also footnote \ref{inner}.}
\subsection{\label{eternal}The eternal BTZ geometry}
We will now briefly review some properties of the eternal BTZ
geometry. The maximal extension of the eternal BTZ geometry, starting
from \eq{banados} is described in detail in Section \ref{charts}. We
will briefly reproduce some of the formulae relevant to the
coordinate system (``EF1'') describing the right exterior and the
interior. The EF1 coordinates are obtained from the coordinates of
\eq{banados} by the transformations
\begin{align}
\frac{z}{z_0} &= \sqrt{\frac1{\lambda_0} \left( \lambda - \sqrt{\lambda^2- \lambda_0^2}
\right)}
\\\nonumber
x_+ &= v- \frac1{2\sqrt L}\ln\left(\frac{\lambda -\lambda_0}{\lambda + \lambda_0}\right),\;
x_- = w- \frac1{2\sqrt{{\bar L}}}\ln\left(\frac{\lambda -\lambda_0}{\lambda + \lambda_0}\right)
\end{align}
The metric, in these coordinates, becomes
\begin{eqnarray}
ds^2=\frac{d\lambda^2}{4(\lambda+\lambda_0)^2}+\frac{L}{4}dv^2+\frac{{\bar L}}{4}dw^2-\lambda\ dv dw
+\frac{\sqrt{L}}{2(\lambda+\lambda_0)}dv d\lambda+\frac{\sqrt{\overline{L}}}{2(\lambda+\lambda_0)}dw d\lambda
\label{EF}
\end{eqnarray}
The event horizon $\lambda_H$, the inner horizon
$\lambda_i$, and the singularity $\lambda_s$ are at
\begin{align}
\lambda_H &= \lambda_0 \equiv \frac{\sqrt{L{\bar L}}}2, \; \lambda_i = - \lambda_0,\;
\lambda_s = - \frac14 (L +{\bar L})
\label{lam-horizon}
\end{align}
Note that for BTZ black holes without angular momentum ${\bar L}=L$
and $\lambda_i = \lambda_s$. The location of the event horizon corresponds to
\eq{banados-horizon}.
In order to regulate IR divergences coming from $\lambda \to \infty$,
we define a cut-off surface $\Sigma_B$ at a constant large
$\lambda= \lambda_{ir}$; the metric \eq{EF} on $\Sigma_B$ turns out
to be
\begin{equation}
\lambda=\lambda_{ir}= 1/\epsilon^2 \Rightarrow
ds^2|_{\Sigma_B} = -(1/\epsilon^2)\ dv\ dw (1+
O(\epsilon^2))
\label{lam-ir}
\end{equation}
By the usual AdS/CFT correspondence the leading
term defines the boundary metric (see Section \ref{uv-ir})
\begin{equation}
ds^2_{bdry}= - dv\ dw
\label{bdry-lam}
\end{equation}
The subleading term in the metric corresponds to the normalizable
metric fluctuation, which gives the expectation value of the stress tensor;
this is the holographic stress tensor \cite{Balasubramanian:1999re}, and
is given here by
\begin{align}
8\pi G_3 T_{vv}(x_+)= \frac{L}4, \,
8\pi G_3 T_{ww}(x_-)= \frac{{\bar L}}4
\label{lam-stress}
\end{align}
It is easy to see that we will get the same boundary metric and stress
tensor from an analysis of the coordinate chart EF4. It is also
straightforward to derive similar results for the left exterior
(which represent a state with the same mass and angular momentum)
using EF2 and EF3.
\subsection{\label{SGD}Solution generating diffeomorphisms (SGD)}
We will now proceed to construct new solutions with arbitrary boundary
data at the two boundaries (represented by two arbitrary holographic
stress tensors $T_{R, \mu\nu}(x)$ and $T_{L,\mu\nu}(x)$) by applying the
method of solution generating diffeomorphisms to the above geometry,
as explained in the introduction.
The solution generating diffeomorphisms can be described as
follows. Suppose we start with a certain metric $g_{MN}(x) dx^M
dx^N$ \footnote{Notation: $x^M=\{ \lambda, x^\mu \}$,\; $x^\mu = \{v,w\}$.}
in a certain coordinate chart ${\cal U}_P$ containing a point P. The
new metric $\tilde g_{MN}$, in this coordinate chart, is given in
terms of a diffeomorphism (active coordinate transformation) $f:{\tilde x}^M=
{\tilde x}^M(x)$, by the definition
\begin{equation}
g\to \tilde g\equiv f^*g: \quad
\tilde g_{MN}(\tilde x) \equiv \frac{\partial x^P}{\partial {\tilde x}^M}
\frac{\partial x^Q}{\partial {\tilde x}^N} g_{PQ}(x)
\label{diffeo-gen}
\end{equation}
In the above, $f^*g$ is a standard mathematical notation for
the pullback of the metric $g$ under the diffeomorphism $f$.
For diffeomorphisms differing infinitesimally from the identity
map: $\tilde x^M = x^M - \xi^M(x)$, we, of course, have the
familiar relation
\begin{equation}
\delta g_{MN}(x)= D_M \xi_N + D_N \xi_M
\label{killing}
\end{equation}
Normally, a diffeomorphism is considered giving rise to a physically
indistinguishable solution; this, however, is not true
when the diffeomorphism is non-trivial at infinity (this
is explained in more detail in Section \ref{nontrivial}).
As explained in Section \ref{charts}, we use five charts to cover the entire
eternal BTZ geometry (see Fig \ref{fig-ef1234}). These charts are labelled as
EF1, EF2, EF3, EF4 and K5.
We use a nontrivial diffeomorphism in each of EF1, EF2, EF3 and EF4, which
overlap with the boundary and the identity transformation in the
Kruskal patch K5.
\subsubsection{The metric in the coordinate chart EF1}
The diffeomorphism in the EF1 coordinate chart is given by
\begin{equation}
\lambda=\frac{{\tilde \lambda}}{G_+'({\tilde v}) G_-'({\tilde w})}, \; v= G_+({\tilde v}), \;
w= G_-({\tilde w})
\label{diffeo}
\end{equation}
The new metric $\tilde g_{MN}$, written in terms of
${\tilde x}^M = ({\tilde \lambda}, {\tilde v}, {\tilde w})$, is
\begin{align}
\tilde g_{MN}(\tilde x) d{\tilde x}^M d{\tilde x}^N
\equiv
ds^2&=\frac{1}{B^2}
\left[
d{\tilde \lambda}^2 +
A_+^2 d{\tilde v}^2+
A_-^2 d{\tilde w}^2+
2A_+ d{\tilde v} d{\tilde \lambda} +
2A_- d{\tilde w} d{\tilde \lambda}
\right.
\nonumber\\
&\kern-30pt - \left.
{\tilde \lambda} \bigg(
B^2+2\bigg(A_+\frac{G''_-({\tilde w})}{G_-'({\tilde w})}+A_-\frac{G_+''({\tilde v})}{G_+'({\tilde v})}+{\tilde \lambda} \frac{G_+''({\tilde v})G_-''({\tilde w})}{G_+'({\tilde v})G_-'({\tilde w})}
\bigg)\bigg)d{\tilde v} d{\tilde w}
\right]
\label{newmetric}
\end{align}
where
\begin{eqnarray}
A_+=\sqrt{L} G_+'({\tilde v})({\tilde \lambda}+{\tilde \lambda}_0)-{\tilde \lambda} \frac{G_+''({\tilde v})}{G_+'({\tilde v})},
\;
A_-=\sqrt{{\bar L}} G_-'({\tilde w})({\tilde \lambda}+{\tilde \lambda}_0)-{\tilde \lambda} \frac{G_-''({\tilde w})}{G_-'({\tilde w})}, \;
B=2({\tilde \lambda}+{\tilde \lambda}_0)\nonumber
\end{eqnarray}
For infinitesimal transformations $G_\pm(x) \equiv x + \epsilon_\pm(x)$,
this amounts to an asymptotically nontrivial
diffeomorphism $\xi^M$ (see \eq{killing})\footnote{The subscript
in $\xi_1^M$ refers to the chart EF1.}
\begin{equation}
\xi_1^v= \epsilon_+(v),\; \xi_1^w= \epsilon_-(w), \; \xi_1^\lambda
=- \lambda \left( \epsilon_+'(v) + \epsilon_-'(w) \right)
\label{killing-a}
\end{equation}
The behaviour of the metric \eq{newmetric} at a constant
large $\lambda$ surface is given by
\begin{equation}
ds^2= - {\tilde \lambda} \ d{\tilde v} d{\tilde w}\ (1 + O(1/{\tilde \lambda}))
\label{leading-lamt}
\end{equation}
This, by following arguments similar to the previous case
(see Section \ref{eternal}), identifies the IR cutoff
surface as
\begin{equation}
{\tilde \lambda}_{ir}= (1/\epsilon^2)
\label{lamt-ir}
\end{equation}
and the boundary metric as
\begin{align}
ds^2_{bdry} &= - d{\tilde v} d{\tilde w}\
\label{boundary-metric}
\end{align}
The subleading term in
\eq{leading-lamt}, as explored in Section \ref{sec-T}, gives the
holographic stress tensor. We will see there that the subleading term
depends on the SGD functions $G_\pm$; this feature is what makes the
SGD's asymptotically \emph{nontrivial} (see Section \ref{nontrivial} for a
more detailed discussion on this).
In terms of the old $\lambda$-coordinate, the surface \eq{lamt-ir} is
\begin{equation}
\lambda= 1/(\epsilon^2 G_+'({\tilde v}) G_-'({\tilde w}))
\label{lamt-lam-ir}
\end{equation}
Note that this surface is different from \eq{lam-ir}, and is
nontrivially warped, as in Figure \ref{fig-ef}. This is another
manifestation of the asymptotic non-triviality of the diffeomorphism
\eq{diffeo}, which is responsible for nontrivial transformation of
bulk quantities, such as geodesic lengths.
\begin{figure}[H]
\centerline{\includegraphics[scale=.20]{ef4}}
\caption{This figure shows the IR cut-off \eq{lamt-ir} in the new
geometries. The effect of the SGDs, in the old (un-tilded)
coordinates, is to deform the IR cut-off surfaces. The surface
deformation on the right exterior is given by the change from
\eq{lam-ir} to \eq{lamt-lam-ir}; there is a similar surface
deformation on the left exterior.}
\label{fig-ef}
\end{figure}
We note that the leading large ${\tilde \lambda}$ behaviour of \eq{newmetric} is
that of AdS$_3$
\begin{equation}
ds^2 = \frac{d{\tilde \lambda}^2}{4 {\tilde \lambda}^2} - {\tilde \lambda}\ d{\tilde v}\ d{\tilde w} + ...
\label{asym-ads3}
\end{equation}
As mentioned before, and will be explored in detail in
Section \ref{sec-T}, the subleading terms, represented by
the ellipsis $...$, are nontrivially different from that of AdS$_3$.
\subsubsection{The metric in the coordinate chart EF2}
The diffeomorphism (SGD) used in the coordinate chart EF2
(see Fig \ref{fig-ef1234}), which is independent of
the one above used in EF1, is given by
\begin{equation}
\lambda_1=\frac{{\tilde \lambda}_1}{H_+'({\tilde u}) H_-'({\tilde\omega})},\; u=H_+({\tilde u}),\;\omega=H_-({\tilde v})
\end{equation}
which leads to the metric
\begin{align}
ds^2&=\frac{1}{B^2}
\left[
d{\tilde \lambda}_1^2+
A_+^2 d{\tilde u}^2+
A_-^2 d{\tilde\omega}^2
-2 A_+ d{\tilde u} d{\tilde \lambda}_1
-2 A_-d{\tilde\omega} d{\tilde \lambda}_1
\right.
\nonumber\\
& - \left.
{\tilde \lambda}_1 \bigg(
B^2-2\bigg(
A_+\frac{H_-''({\tilde\omega})}{H_-'({\tilde\omega})}+A_-\frac{H_+''({\tilde u})}{H_+'({\tilde u})}-{\tilde \lambda}_1\frac{H_+''({\tilde u})H_-''({\tilde\omega})}{H_+'({\tilde u})H_-'({\tilde\omega})}
\bigg)\bigg)
d{\tilde\omega} d{\tilde u}
\right]
\label{newmetric-2}
\end{align}
where
\begin{eqnarray}
A_+=\sqrt{L} H_+'({\tilde u})({\tilde \lambda}_1+{\tilde \lambda}_0)+ {\tilde \lambda}_1 \frac{H_+''({\tilde u})}{H_+'({\tilde u})},\;
A_-=\sqrt{{\bar L}} H_-'({\tilde\omega})({\tilde \lambda}_1+{\tilde \lambda}_0)+ {\tilde \lambda}_1 \frac{H_-''({\tilde\omega})}{H_-'({\tilde\omega})},\;
B=2({\tilde \lambda}_1+{\tilde \lambda}_0)\nonumber
\end{eqnarray}
For infinitesimal transformations $H_\pm(x)= x + \varepsilon_\pm(x)$, this
implies a diffeomorphism $\xi_2^M$ where
\begin{equation}
\xi_2^u=- \varepsilon_+(u),\; \xi_2^\omega=-
\varepsilon_-(\omega), \; \xi_2^\lambda
= -\lambda \left( \varepsilon_+'(u) + \varepsilon_-'(\omega) \right)
\label{killing-b}
\end{equation}
Note, once again, the asymptotic nontriviality of the above diffeomorphism.
\subsection{\label{full-metric}The full metric}
In a manner similar to the above, we apply the SGD characterized by
$G_\pm$ on EF4 (which shares the right exterior with EF1, see Appendix
\ref{EFs}): and the SGD characterized by $H_\pm$ on EF3 (which shares
the left exterior with EF2):
\begin{align}
\hbox{EF4}:\kern10pt &\lambda=\frac{{\tilde \lambda}}{G_+'({\tilde u}_1) G_-'({\tilde\omega}_1)}, \;
u_1=G_+({\tilde u}_1),\;\omega_1=G_-({\tilde\omega}_1)
\nonumber\\
&\kern20pt \hbox{infinitesimally}~~
\left(\xi_4^\lambda,\xi_4^{u_1}, \xi_4^{\omega_1}\right)
= \left(-\lambda (\epsilon_+'(u_1) + \epsilon_-'(\omega_1)), \epsilon_+(u_1), \epsilon_-(\omega_1)
\right)
\nonumber\\
\hbox{EF3}:\kern10pt &\lambda=\frac{{\tilde \lambda}_1}{H_+'({\tilde v}_1) H_-'({\tilde w}_1)}, \;
v_1=H_+({\tilde v}_1),\;w_1=H_-({\tilde w}_1)
\nonumber\\
&\kern20pt \hbox{infinitesimally}~~
\left(\xi_4^\lambda,\xi_4^{v_1}, \xi_4^{w_1}\right)
= \left(-\lambda (\varepsilon_+'(v_1) +
\varepsilon_-'(w_1)), \varepsilon_+(v_1), \varepsilon_-(w_1)
\right)
\end{align}
The infinitesimal transformations are similar to those in
eqs. \eq{killing-a} and \eq{killing-b}. As mentioned above, we use
the identity diffeomorphism of Kruskal patch K5 (with $\xi_5^M=0$).
The expressions for the metric in various coordinate charts
are given in \eq{newmetric}, \eq{newmetric-2}, \eq{newmetric-3},
\eq{newmetric-4} and \eq{Kruskal}.
We will now show that the five different metrics in the five
coordinate charts define a single metric in the entire spacetime. To
see this, note that although the SGD's applied on the five charts
are different, (equivalently, for infinitesimal transformations, the
diffeomorphisms $\xi_i^M$ in the five charts differ from each other),
they satisfy the following sufficient criteria:
\begin{enumerate}
\item [(i)] At both the right (and left) exterior boundary, the
diffeomorphisms coincide. For example, in case of the right exterior (see
\eq{rt-exterior}), as $\lambda \to \infty$, $u_1 \to v $, $\omega_1 \to w$.
Hence ${\tilde u}_1 = G_+^{-1}(u_1) \to G_+^{-1}(v) = {\tilde v}$. In other words,
for infinitesimal transformations $\xi_4^M(P)\to \xi_1^M(P)$ for a
given point $P$ with $\lambda \to \infty$. This implies that the metric
\eq{newmetric} coincides at the right boundary with the similar
metric\eq{newmetric-3} obtained by applying the $G_\pm$ transformations
on the coordinate chart EF4. Similarly, the metric \eq{newmetric-2} obtained by the
$H_\pm$ transformations in EF2 and the similar metric \eq{newmetric-4} obtained by the
$H_\pm$ transformations in EF3 coincide at the left exterior boundary.
\item [(ii)] Away from the boundary, the metrics obtained in the
various EF coordinate charts differ from each other only by trivial
diffeomorphisms which become the identity transformation at
infinity. Since the physical content of each of these metrics is
represented only by the boundary data, the above point (i) ensures
that all the different metrics represent the same single spacetime
metric in different charts (see Figure \ref{fig-nontrivial}).
\item[(iii)] It is clear that the SGDs lead to a {\it smooth metric}
in each chart, provided $G_\pm(x), H_\pm(x)$ are differentiable and
invertible functions. In the rest of the paper, we will only
consider such functions. It can be verified that such a class of
functions is sufficiently general to generate (through
transformations such as \eq{stress}) any pair of physically sensible
holographic stress tensors at both boundaries.
\end{enumerate}
\begin{figure}[ht]
\centerline{\includegraphics[width=400pt, height=200pt]{nontrivial-a}}
\caption{A schematic illustration of metrics in our paper related by
trivial and nontrivial diffeomorphisms (see the definition
\ref{def-nontrivial}). The metrics \eq{EF}, \eq{EF2}, \eq{EF3} and
\eq{EF4}, represented by the blue lines, define the eternal BTZ
geometry; they are all related by trivial diffeomorphisms, which
either do not extend to the boundaries or when they do, they become
identity asymptotically. The metrics \eq{newmetric},
\eq{newmetric-2}, \eq{newmetric-3} and \eq{newmetric-4}, represented
by the green lines, define our new solution characterized by the
functions $G_\pm, H_\pm$. These are also all related by trivial
diffeomorphisms, which satisfy the same criteria as above. The two
sets however represent physically different metrics since they are
related to each other by nontrivial diffeomorphisms; for instance,
\eq{EF} and \eq{newmetric} are related by a diffeomorphism,
schematically represented by their separation, which does not vanish
(become identity) asymptotically.}
\label{fig-nontrivial}
\end{figure}
\subsubsection{\label{Dirac}Analogy with the Dirac monopole}
It is important to note that our new solutions can only be specified
in terms of a different metric in different coordinate charts which
are equivalent to each other. This is analogous to case of the Dirac
monopole: the gauge field $A_\mu$ for a static U(1) magnetic monopole
of charge $q_m$ at the origin needs to be specified separately on two
separate coordinate charts:
\begin{align}
F= q_m \sin\theta\ d\theta\ d\phi: \;
A_N = q_m (1- \cos\theta)\ d\phi,\;
A_S = q_m (-1- \cos\theta)\ d\phi
\label{dirac}
\end{align}
Here ${\mathbb R}^3 - \{0\}$ is viewed as ${\mathbb R} \times S^2$
where $S^2$ is described by two coordinate charts $N_N$ and $N_S$
(such as obtained by a stereographic projection on to the plane) which
include all points of $S^2$ minus the south and north pole
respectively. $A_N^\theta$ vanishes (and is hence regular) at the
north pole $\theta=0$, but develops a string singularity at the south
pole $\theta=\pi$ (for each $r>0$). Similarly, $A_S$ is regular at
the south pole, but has a string singularity at the north pole. The
important point to note is that in spite of appearances, $A_N$ and
$A_S$ describe the same gauge field in the region of overlap
$N_N \cap N_S$. This is because in this region, $A_N = A_S + d\chi$ where $\chi=2 q_m d\phi$
represents a pure gauge transformation for appropriately quantized
$q_m$ (Dirac quantization condition).
In the present case the metric \eq{newmetric} written in EF1, although
non-singular on the future horizon, is singular on the past horizon for
general $G_\pm$. In order to describe the metric in a neighbourhood of the past horizon, we must
switch to the metric in EF4. Similarly, in order to describe the
diffeomorphism at the bifurcation surface, we must use the metric
\eq{Kruskal} in the K5 coordinate chart.
\subsubsection{Summary of this subsection:} The metrics \eq{newmetric},
\eq{newmetric-2}, \eq{newmetric-3}, \eq{newmetric-4} and \eq{Kruskal},
valid in the coordinate charts EF1, EF2, EF3, EF4 and K5 respectively, define a
spacetime with a regular metric. The metrics are asymptotically
AdS$_3$ at both the right and left boundaries; the subleading terms in
the metric are determined by the solution generating diffeomorphisms
$G_\pm, H_\pm$ and can be chosen to fit boundary data specified by
arbitrary holographic stress tensors. A schematic representation
of our solution is presented in Figure \ref{fig-nontrivial}.
\subsection{\label{horizon}Horizon}
In Section \ref{SGD} we viewed the SGDs as a coordinate
transformation. Alternatively, however, we can also view the
diffeomorphism as an active movement of points: $x^M \to {\tilde x}^M$
$= x^M + \xi^M$. In this viewpoint, the future horizon $\lambda= \lambda_H=
\lambda_0$ (see \eq{lam-horizon}) on the right moves to
\begin{align}
{\tilde \lambda}_H = G_+'({\tilde v})\ G_-'({\tilde w}) \lambda_0,\;
{\tilde \lambda}_{1,H} = H_+'({\tilde u})\ H_-'({\tilde\omega}) \lambda_0
\label{warped-horizon}
\end{align}
\begin{figure}[h]
\centerline{\includegraphics[width=180pt,height=150pt]{hLa}\kern15pt
\includegraphics[width=180pt,height=150pt]{hRa}}
\caption{The figure on the right shows the location of the horizon on
the right in the ${\tilde \lambda}, {\tilde v}, {\tilde w}$ coordinates. The figure on the left
shows the location of the horizon on the left in the $\tilde \lambda_1,
{\tilde u}, {\tilde\omega}$ coordinates. These are described by \eq{warped-horizon}.
These surfaces are diffeomorphic to the undeformed horizon
\eq{lam-horizon} depicted in Figure \ref{fig-ef}. Although the horizon
has an undulating shape in our coordinate system, the
expansion parameter, measured by the divergence of the area-form,
vanishes (see Eq. \eq{no-divergence}).}
\label{fig-horizons}
\end{figure}
Similar statements can be made in the other coordinate charts. The
horizons represented this way are smooth but undulating (see figure
\ref{fig-horizons}).
The geometry of warped horizons in
\cite{Bhattacharyya:2008jc,Bhattacharyya:2008xc} was used to yield a
holographic prescription for computing local entropy current of a
fluid. In Section \ref{sec-S} we use a similar technology to compute a
holographic entropy in our case.
\subsection{On the nontriviality of solution generating
diffeomorphisms \label{nontrivial}}
It is natural to wonder how a metric such as \eq{newmetric} provides a
new solution since it is obtained by a diffeomorphism from \eq{EF};
however, the fact that the diffeomorphism \eq{diffeo} is
asymptotically nontrivial makes the new solution physically
distinct. Thus, in \eq{diffeo} ${\tilde \lambda}$ remains different from $\lambda$ in
the asymptotic region. Indeed, as we will see, the first subleading
term in the metric \eq{newmetric} carries nontrivial data about a
holographic stress tensor \eq{stress} on the right boundary.
Asymptotically AdS$_3$ diffeomorphisms were first discussed by Brown
and Henneaux \cite{Brown:1986nw} who showed that such transformation
led to an additional surface contribution to conserved charges of the
system. These observations were preceded by a general discussion of
such surface charges in the context of gauge theories and gravity in
\cite{Regge:1974zd,Wadia:1979yu,Gervais:1976ec}. These
authors identified asymptotically non-vanishing pure gauge transformations
as global charge rotations.
In the current AdS/CFT context, the surface charges are encapsulated
by the holographic stress tensors on the two boundaries. As we will
see shortly, they change nontrivially under the solution generating
diffeomorphisms (SGD's). In fact, the SGD's reduce to conformal
transformations on the boundary. As a result, the `global charge
rotations' mentioned above correspond to a {\it conformal}
transformation of the stress tensor. The important point is that
starting from a given constant stress tensor on each boundary,
the two independent SGD's can generate two {\it independent} and
{\it completely general} stress tensors by this method.
We should note that the diffeomorphisms define a new theory in which
the appropriate choice of the IR cutoff surface is \eq{lamt-ir}. In
this description, the horizon becomes an undulating surface as in Fig
\ref{fig-horizons}. An equivalent (`active') viewpoint is to describe
the new geometry in terms of the old coordinates \eq{EF}, but to
change the IR-cutoff surface from \eq{lam-ir} to \eq{lamt-ir}. In
either case, the holographic stress tensor changes.
We conclude this section with the following definition of a nontrivial
diffeomorphism, which has been implicit in much of the above
discussion.
\subsubsection{\label{def-nontrivial}Definition}
A local diffeomorphism which does not extend to either boundary (left
or right), or a diffeomorphism which extends to a boundary but
asymptotically approaches the identity diffeomorphism there, is called
a `trivial' diffeomorphism. Contrarily, a diffeomorphism which
extends to a boundary where it does not approach the identity
diffeomorphism, is called `nontrivial'. Quantitatively, a nontrivial
diffeomorphism ($f$) is one under which the holographic stress tensor
computed from the existing metric $g$ at the boundary is different
from that computed from the pulled back metric $f^*g$.
\section{The Dual Conformal Field Theory \label{CFT-dual}}
As we saw above, the SGD's reduce to conformal transformations at the
boundary. We will construct the CFT-dual to the new solutions
using the above idea.
Note that the eternal BTZ black hole geometry,
described by \eq{EF} and \eq{EF2}, corresponds to the following
thermofield double state
\cite{Maldacena:2001kr,Hartman:2013qma,Hubeny:2007xt,Caputa:2013eka}
\begin{equation}
| \psi_0 \rangle = Z(\beta_+, \beta_-)^{-1/2} \sum_{n} \exp[-\beta_+ E_{+,n}/2
- \beta_- E_{-,n}/2] | n\rangle | n \rangle
\label{thermofield-double}
\end{equation}
The states $| n\rangle
\in {\cal H}$ denote all simultaneous eigenstates of
$H_\pm= (H \pm J)/2$ with eigenvalues $E_{\pm,n}$.
$| \psi_0 \rangle$ here is a pure state in $ {\cal H} \otimes {\cal
H}$ obtained by the `purification' of the thermal state
\eq{thermal-state}. \footnote{For definiteness, we will sometimes
call the two Hilbert spaces ${\cal H}_L$ and ${\cal H}_R$, where
$L, R$ represent `left' and `right', corresponding to the two
exterior boundaries of the eternal BTZ. Indeed, $L, R$ also
have an alternative meaning. The left/right boundary of the eternal BTZ
geometry maps to the left/right Rindler wedge of the boundary of
Poincare coordinates, respectively.\label{left-right}}
\begin{equation}
Z(\beta_+, \beta_-) = \text{Tr} \rho_{\beta_+, \beta_-}\hspace{1cm}\mbox{with} \kern20pt
\rho_{\beta_+, \beta_-}= \exp[-\beta_+ H_+ - \beta_- H_-] = \exp[-\beta(H
+ \Omega J)]
\label{thermal-state}
\end{equation}
represents the grand canonical ensemble in ${\cal H}$ with inverse
temperature $\beta$ and angular velocity $\Omega$ (which can be viewed as
the thermodynamic conjugate to the angular momentum $J$). Also $\beta_\pm =
\beta(1\pm \Omega)$. \footnote{The thermal state $\rho_{\beta_+, \beta_-}$ (see
\eq{thermal-state}) implies a field theory geometry where the light
cone directions have periods $\beta_\pm$.}
Note that $| \psi_0 \rangle$ is a pure state in $ {\cal H} \otimes {\cal
H}$ , and is a `purification' of the thermal state
\eq{thermal-state}.
\noindent\underbar{The non-spinning BTZ:}
The CFT dual for the more familiar case of non-spinning eternal BTZ
black hole ($\Omega=0= J$) is the standard thermofield double:
\begin{equation}
| \psi_{0,0} \rangle = Z(\beta)^{-1/2} \sum_{n} \exp[-\beta E_n/2]
| n\rangle | n \rangle
\label{standard-thermofield-double}
\end{equation}
where $| n \rangle$ now denotes all eigenstates of $H$.
\footnote{An entanglement entropy for this state was calculated in
\cite{Hartman:2013qma} and matched with a bulk geodesic calculation.
This was generalized to the spinning eternal BTZ black hole in
\cite{Caputa:2013eka}}
\paragraph{CFT duals of our solutions}
Following the arguments above
\eq{thermofield-double}, we claim
that the CFT-duals to the new solutions described in
Section \ref{full-metric}
are described by the following pure states in ${\cal H}
\otimes {\cal H}$:
\begin{align}
| \psi \rangle = U_L U_R | \psi_0 \rangle=
Z(\beta_+, \beta_-)^{-1/2}\sum_{n} \exp[-\beta_+ E_{+,n}/2
- \beta_- E_{-,n}/2] U_L | n \rangle U_R | n \rangle
\label{thermofield-general}
\end{align}
where $U_R$ is the unitary transformation which implements the
conformal transformations on the CFT on the right boundary
(characterized by
$G_\pm$), and $U_L$ is the unitary transformation which implements the
conformal transformations on the CFT on the left boundary (characterized by
$H_\pm$). See Appendix \ref{unitary} for an explicit construction of a
unitary transformations $U_R$.
\gap1
\noindent In the following sections, we will provide many checks for this proposal.
However, first we shall discuss how to compute various correlators
in the above state \eq{thermofield-general}.
\subsection{\label{corr-gen}Correlators}
Let us first consider correlators in the standard
thermofield double state \eq{thermofield-double}.
It is known that correlators of one-sided CFT observables,
say $O_R$, satisfy an AdS/CFT relation of the form \footnote{We
will mostly use unprimed labels, $P_1, P_2,...$ for points on
the spacetime of the `right' CFT, and primed labels, $P'_1, P'_2,...
$ for the space of the `left' CFT.}
\begin{equation}
\langle \psi_0 | O_R(P_1) O_R(P_2) ... O_R(P_n) | \psi_0 \rangle
\equiv \text{Tr}\left( \rho_{\beta_+, \beta_-}
O_R(P_1) O_R(P_2) ... O_R(P_n) \right)
= G_{\rm bulk}({\bf P}_1, {\bf P}_2,...{\bf P}_n)
\label{RR-correlator}
\end{equation}
where the bulk correlator $G_{bulk}$
is computed from the (right exterior region of)
a dual black hole geometry with
temperature $T=1/\beta$ and angular velocity $\Omega$.
Two-sided correlators, similarly, satisfy a relation like
\begin{equation}
\langle \psi_0 | O_R(P_1) O_R(P_2) ... O_R(P_m)
O_L(P'_1)...O_L(P'_n) | \psi_0 \rangle
= G_{\rm bulk}({\bf P}_1, {\bf P}_2,...{\bf P}_m; {\bf P}'_1,..., {\bf P}'_n)
\label{RL-correlator}
\end{equation}
where the bulk correlator on the RHS is computed from the two-sided geometry
of the eternal BTZ black hole \cite{Maldacena:2001kr,
Hartman:2013qma,Hubeny:2007xt,Caputa:2013eka},
represented in this paper by \eq{EF} and \eq{EF2}. The bold-faced label ${\bf P}$
above represents an image of the field theory point $P$ on a
cut-off surface in the bulk under the usual
AdS/CFT map. E.g. in the coordinates of \eq{EF}, the map is given by
\begin{equation}
P \mapsto {\bf P}\equiv (\lambda=\lambda_{ir}=1/\epsilon^2,P)
\label{bold-map}
\end{equation}
where $\epsilon$ is the UV cut-off in the CFT, cf. \eq{lam-ir}).
There is a similar map for the {\it left} boundary.\\
In particular, the holographic correspondence for
the two point functions of scalar operators
can be written simply as \cite{Louko:2000tp}:
\begin{align}
\langle \psi_0 | O_R(P) O_R(Q) | \psi_0 \rangle
&= \text{Tr} ( \rho_{\beta_+, \beta_-} O_R(P) O_R(Q)) = \exp[-2h L({\bf P},{\bf Q})]
\nonumber\\
\langle \psi_0 | O_R(P) O_L(Q') | \psi_0 \rangle
& = \exp[-2h L({\bf P},{\bf Q}')]
\label{geodesics}
\end{align}
where $L({\bf P},{\bf Q})$ is the length of the extremal geodesic
connecting ${\bf P}$ and ${\bf Q}$ (similarly with $L({\bf P},{\bf Q}')$).\\
It is easy to see that correlators in the new, transformed, state
$|\psi \rangle$ \eq{thermofield-general} can be understood as correlators
of transformed operators in the old state $|\psi_0 \rangle$, i.e.
\begin{align}
\langle \psi | O_R(P_1)...O_R(P_m)O_L(P'_1)... O_L(P'_n) | \psi \rangle
= \langle\psi_0 |
\tilde O_R(P_1) ...\tilde O_R(P_m) \tilde O_L(P'_1)
... \tilde O_L(P'_n) | \psi_0 \rangle
\label{transform-corr}
\end{align}
where
\begin{align}
\tilde O_R(P) \equiv U^\dagger_R O_R(P) U_R, \kern20pt
\tilde O_L(P') \equiv U^\dagger_L O_L(P') U_L
\label{transform-op}
\end{align}
For a primary field $O_R$ with conformal dimensions $(h,\bar h)$, the
conformally transformed operator satisfies the relation
\begin{equation}
\tilde O_R(\tilde v, \tilde w)
= O_R(v,w) \left(\frac{dv}{d{\tilde v}}\right)^h
\left(\frac{dw}{d{\tilde w}}\right)^{\bar h}
\label{primary}
\end{equation}
\subsection{\label{checking}Strategy for checking AdS/CFT}
To check the claim that the states \eq{thermofield-general} are
CFT-duals to the new bulk geometries found in Section \ref{full-metric},
we need to show a relation of the form
(cf. \eq{RL-correlator})
\begin{align}
\langle\psi_0 |
\tilde O_R(P_1) ...\tilde O_R(P_m) \tilde O_L(P'_1)
... \tilde O_L(P'_n) | \psi_0 \rangle
= \tilde G_{\rm bulk}(\tilde {\bf P}_1,
\tilde {\bf P}_2,...\tilde {\bf P}_m; \tilde {\bf P}'_1,...,
\tilde {\bf P}'_n)
\label{matching}
\end{align}
where the RHS is computed in the new geometries. Here $\tilde {\bf P}$
represents the image of the CFT point $P$, under AdS/CFT, on the
cut-off surface \eq{lamt-ir} in the new geometry. In the
language of \eq{newmetric}, the map is
\begin{equation}
P \mapsto \tilde {\bf P}= ({\tilde \lambda}={\tilde \lambda}_{ir}=1/\epsilon^2,P)
\label{new-bold-map}
\end{equation}
\noindent\underbar{Two-point correlators:} In the
particular case of two-point functions
\begin{align}
\langle \psi_0 | \tilde O_R(P) \tilde
O_R(Q) | \psi_0 \rangle
&= \text{Tr} ( \rho_{\beta_+, \beta_-} \tilde O_R(P) \tilde O_R(Q)) =
\exp[-2h \tilde L(\tilde {\bf P},\tilde {\bf Q})]
\nonumber\\
\langle \psi_0 | \tilde O_R(P) \tilde O_L(Q') | \psi_0 \rangle
& = \exp[-2h \tilde L(\tilde {\bf P}, \tilde {\bf Q}')]
\label{geodesics-new}
\end{align}
where $\tilde L(\tilde {\bf P},\tilde {\bf Q})$ is the length of the extremal
geodesic connecting $P$ and $Q$ in the new geometry (similarly with
$\tilde L(\tilde {\bf P}, \tilde {\bf Q}')$). The discerning reader may
justifiably wonder how a geodesic length in the new geometry can be
different from that in the original, eternal BTZ black hole geometry,
since the former is obtained by a diffeomorphism
from the latter; the point is that the bulk points $\tilde {\bf P}$,
given by \eq{new-bold-map} are {\it not} the same as the bulk
points ${\bf P}$ given by \eq{bold-map}. For example, a geodesic with endpoints
at a fixed IR cut-off ${\tilde \lambda}= 1/\epsilon^2$ (both
on the right exterior) corresponds, in the
eternal BTZ black hole, to a geodesic with two end-points
at \eq{lamt-lam-ir} $\lambda= 1/(\epsilon^2 G_+'({\tilde v}) G_-'({\tilde w}))$.
As we will see below, it is this shift which ensures the equality
in \eq{geodesics-new}. This is one more instance of how our geometries
are nontrivially different from the original BTZ solution although
they are obtained by diffeomorphisms (see Section \ref{nontrivial}
for more detail).
\section{Holographic Stress Tensor\label{sec-T}}
In this section we will discuss our first observable $O$: the stress
tensor. We will first consider the stress tensor of the boundary
theory on the right. The generalization to the stress tensor on the
left is trivial. The equation \eq{matching} now implies that we
should demand the following equality
\begin{align}
\langle\psi| T_{vv}(P) | \psi \rangle
\equiv \text{Tr} \left( \rho_{\beta_+, \beta_-} U_R^\dagger T_{vv}(P) U_R \right)
= \tilde T_{{\rm bulk},{\tilde v}\vt}(\tilde {\bf P})
\label{matching-stress}
\end{align}
and a similar equation for the right-moving stress tensor $T_{ww}(w)$.
\paragraph{Bulk} The RHS of this equation is simply the
holographic stress tensor, computed in the new geometry
\eq{newmetric}. We use the definition of holographic stress tensor in
\cite{Balasubramanian:1999re,Skenderis:1999nb}:\footnote{We
drop the subscript {\tiny bulk} from the bulk stress
tensor, as it should be obvious from the context whether we
are talking about the CFT stress tensor or the holographic
stress tensor.}
\begin{equation}
8\pi G_3 T_{\mu\nu}= \lim_{\epsilon\to 0}\left(
K_{\mu\nu}- K h_{\mu\nu} - h_{\mu\nu} \right)
\label{t-ren}
\end{equation}
where $h_{\mu\nu}$ is the induced metric on the cut-off surface
$\Sigma:{\tilde \lambda}= {\tilde \lambda}_{ir}=1/\epsilon^2$, chosen in accordance with
\eq{new-bold-map} which is the natural one in the new
geometry (note that it is different from
the cut-off surface implied by \eq{bold-map}).
$K_{\mu\nu}$ and $K$ are respectively
the extrinsic curvature and its trace on $\Sigma$.
It is straightforward to do the explicit calculation; we find
that
\begin{align}
8\pi G_3 T_{{\tilde v}\vt} &=\frac{L}{4} G_+'({\tilde v})^2 + \frac{3 G_+''({\tilde v})^2-2 G_+'({\tilde v})
G_+'''({\tilde v})}{4G_+'({\tilde v})^2},
\nonumber\\
8\pi G_3 T_{{\tilde w}\wt} &=\frac{{\bar L}}4 G_-'({\tilde w})^2 + \frac{3 G_-''({\tilde w})^2-2 G_-'({\tilde w})
G_-'''({\tilde w})}{4G_-'({\tilde w})^2}
\label{stress}
\end{align}
This clearly looks like a conformal transformation of the
original stress tensor \eq{lam-stress}. We will explicitly verify
below that it agrees with the CFT calculation. The generalization
to $T_{ww}$ and to the stress tensors of the second CFT is
straightforward.This clearly has the form of a conformal transformation of the
original stress tensor \eq{lam-stress}. We will explicitly verify
below in the CFT that it indeed is precisely a conformal
transformation, as demanded by \eq{matching-stress}. The
generalization of \eq{stress} to the stress tensors
$T_{{\tilde u}\ut}, T_{{\tilde\omega}\omt}$ of the second CFT
is straightforward.
In this paper, we will sometimes use the notation $T_R, \bar T_R$ for
$T_{{\tilde v}\vt}, T_{{\tilde w},{\tilde w}}$, and $T_L, \bar T_L$ \footnote{$T_R, \bar
T_R$ represent the left-moving and right-moving stress tensors on
the Right CFT; similarly for $T_L, \bar T_L$.} for $T_{{\tilde u}\ut},
T_{{\tilde\omega}\omt}$ respectively. It is clear that by appropriately choosing
the functions $G_\pm$ and $H_\pm$, any set of boundary stress tensors
$T_{R,L}, \bar T_{R,L}$ can be generated. This is how our solutions
described in Section \ref{full-metric} solve the boundary
value problem mentioned in the Introduction.
\paragraph{CFT}
The unitary transformation in the LHS of \eq{matching-stress},
implements, by definition, the following conformal transformation
(see Appendix \ref{unitary} for more details) on the quantum
operator
\begin{equation}
U_R^\dagger T_{vv}(P) U_R
= \bigg(\frac{\partial {\tilde v}}{\partial v}\bigg)^{-2}[T_{{\tilde v}\vt}({\tilde v})-
\frac{c}{12}S(v,{\tilde v})]
\label{stress-transfmn}
\end{equation}
From \eq{diffeo}, the relevant conformal transformation
here is $v= G_+({\tilde v})$. Using this, the definition
\eq{schwarzian} of the Schwarzian derivative $S(v, {\tilde v})$,
and the identification \cite{Brown:1986nw}
\begin{equation}
G_3= 3/(2c),
\label{newton}
\end{equation}
we find that \eq{stress-transfmn} exactly agrees with \eq{stress}.
\gap1
\noindent This {\emph proves} the AdS-CFT equality \eq{matching-stress} for the
stress tensor.
\section{General two-point correlators\label{2-pt}}
In this section we will discuss general two-point correlators, both
from the bulk and CFT viewpoints following the steps outlined in
Section \ref{corr-gen}.
\subsection{Boundary-to-Boundary Geodesics}
As mentioned in \eq{geodesics}, the holographic calculation of a
two-point correlator reduces to computing the geodesic length between
the corresponding boundary points. We will first calculate correlators
in the thermofield double state \eq{thermofield-double}, which
involves computing geodesics in the eternal BTZ geometry \eq{EF}.
\subsubsection*{In the eternal BTZ geometry}
\noindent\underbar{RL geodesic:} Let us consider a geodesic running from a
point $\bold{P}(1/\epsilon^2_R,v,w)$ on the right boundary to a
point $\bold{Q'}=(1/\epsilon^2_L,u,\omega)$ on the left boundary.\footnote{For the
calculation at hand we need to put $\epsilon_L = \epsilon_R= \epsilon$; however,
we keep the two cutoffs independent for later convenience.} As
shown in Section \ref{sec-poincare} (see \cite{Hartman:2013qma}) both the
right exterior ($\subset$ EF1) and the left exterior ($\subset$ EF2)
can be mapped to a single coordinate chart in Poincare coordinates.
Let the Poincare coordinates for ${\bf P}$ and ${\bf Q'}$, be
$(X_{+R},X_{-R},\zeta_R)$ and $(X_{+L},X_{-L},\zeta_L)$ respectively.
By using the coordinate transformations given in
(\ref{poincare-to-ef1}) and (\ref{poincare-to-ef2}), we find, upto the
first subleading order in $\epsilon_R$ and $\epsilon_L$,
\begin{eqnarray}
\label{bp}
&&X_{+R}=e^{\sqrt{L}v},\quad X_{-R}=-e^{-\sqrt{L}w}+L\epsilon_R^2 e^{-\sqrt{L}w},\quad\zeta_R^2=L\epsilon_R^2\,e^{\sqrt{L}(v-w)}\\
&&X_{+L}=-e^{\sqrt{L}u}+L\epsilon_L^2 e^{\sqrt{L}u},\quad X_{-L}=e^{-\sqrt{L}\omega},\quad\zeta_L^2=L\epsilon_L^2\,e^{\sqrt{L}(u-\omega)}\nonumber\
\end{eqnarray}
with $L=\bar{L}$.\footnote{For simplicity, we present the calculation
here for $L={\bar L}$; the generalization to the spinning BTZ is
straightforward.} The geodesic in Poincare coordinates is given by
\begin{eqnarray}
X_+=A\tanh\tau+C,\quad X_-=B\tanh\tau+D,\quad \zeta=\frac{\sqrt{-AB}}{\cosh\tau}\nonumber\
\end{eqnarray}
where $\tau$ is the affine parameter, which takes the
values $\tau_R$ and $\tau_L$ at ${\bf P}$ and ${\bf Q'}$ respectively. The
constants $A,B,C,D,\tau_L$ and $\tau_R$ are fixed by the endpoint coordinates
given above.
In the limit $\epsilon_R,\epsilon_L\rightarrow0$, we obtain
\begin{eqnarray}
\tau_{R}&=&\log\Big{[}\frac{e^{-(\sqrt{L}v+\sqrt{L}\omega)/2}}{\sqrt{2}}\sqrt{\frac{(e^{\sqrt{L}v}+e^{\sqrt{L}u})(e^{\sqrt{L}w}+e^{\sqrt{L}\omega})}{\lambda_0\epsilon_R^2}}\Big{]}\nonumber\\
\tau_{L}&=&-\log\Big{[}\frac{e^{-\sqrt{L}(u+w)/2}}{\sqrt{2}}\sqrt{\frac{(e^{\sqrt{L}v}+e^{\sqrt{L}u})(e^{\sqrt{L}w}+e^{\sqrt{L}\omega})}{\lambda_0\epsilon_L^2}}\Big{]}\nonumber\
\end{eqnarray}
where $\lambda_0=L/2$ (see \eq{lam-horizon}). The geodesic length is now simply
given by the affine parameter length
\begin{eqnarray}
L(\bold{P},\bold{Q'})=\tau_R-\tau_L=\log\left[\frac{4\cosh[\sqrt{L}(v-u)/2]\cosh[\sqrt{L}(w-\omega)/2]}{L\epsilon_R\epsilon_L}\right]\
\label{geodesic}
\end{eqnarray}
For comparison with CFT correlators in the thermofield double, we will
put, in the above expression, $\epsilon_L= \epsilon_R =\epsilon$, where $\epsilon$ is
the (real space) UV cut-off in the CFT.
\noindent\underbar{RR geodesic:}
If we take the two boundary points on the same
exterior region, say on the right, $\bold{P_1}(1/\epsilon_1^2,v_1,w_1)$ and
$\bold{P_2}(1/\epsilon_2^2,v_2,w_2)$, then the corresponding Poincare
coordinates are (using (\ref{poincare-to-ef1}))
\begin{eqnarray}
\label{bp12}
X_{+1}=e^{\sqrt{L}v_1}, &&\quad X_{-1}=-e^{-\sqrt{L}w_1}+L\epsilon_1^2 e^{-\sqrt{L}w_1},\quad\zeta_1^2=L\epsilon_1^2\,e^{\sqrt{L}(v_1-w_1)}\\
X_{+2}=e^{\sqrt{L}v_2}, &&\quad X_{-2}=-e^{-\sqrt{L}w_2}+L\epsilon_2^2 e^{-\sqrt{L}w_2},\quad\zeta_2^2=L\epsilon_2^2\,e^{\sqrt{L}(v_2-w_2)}\nonumber\
\end{eqnarray}
Following steps similar to above, we have, in the $\epsilon_1,\epsilon_2\rightarrow0$ limit,
\begin{eqnarray}
\tau_{1}&=&\log\Big{[}\frac{e^{-(v_1+w_2)/2}}{\sqrt{2}}\sqrt{\frac{(e^{v_1}-e^{v_2})(-e^{w_1}+e^{w_2})}{\lambda_0\epsilon_1^2}}\Big{]}\nonumber\\
\tau_{2}&=&-\log\Big{[}\frac{e^{-(v_1+w_1)/2}}{\sqrt{2}}\sqrt{\frac{(-e^{v_1}+e^{v_2})(e^{w_1}-e^{w_2})}{\lambda_0\epsilon_2^2}}\Big{]}\nonumber\
\end{eqnarray}
The geodesic length is then
\begin{eqnarray}
L(\bold{P_1},\bold{P_2})=\tau_{+1}-\tau_{+2}=\log\left[\frac{4\,\sinh[(v_1-v_2)/2]\sinh[(w_1-w_2)/2]}{L\epsilon_1\epsilon_2}\right]\
\label{geodesic12}
\end{eqnarray}
For comparison with CFT, we will put $\epsilon_1= \epsilon_2= \epsilon$.
\subsubsection*{In the new geometries}
As explained in Section \ref{behind}, the IR boundary in the new solutions, obtained by
the SGDs, is given by the equation \eq{lamt-ir} or
equivalently by \eq{lamt-lam-ir}, and analogous equations on the left.
This is encapsulated by the CFT-to-bulk map \eq{new-bold-map}. In
case of the {\it RL geodesic}, the CFT endpoints $(P, Q')$
now translate to new boundary points $({\bf{\tilde P},{\tilde Q'}})$
with the following new values of the old $(\lambda$, $\lambda_1)$ coordinates:
\begin{equation}
\lambda \equiv
\frac{1}{\epsilon_R^2} = \frac{1}{\epsilon^2 G'_+(\tilde{v}) G'_-(\tilde{w})},\; \quad \lambda_1 \equiv \frac{1}{\epsilon_L^2} = \frac{1}{\epsilon^2 H'_+(\tilde{u}) H'_-(\tilde{\omega})}\;
\end{equation}
which just has the effect of conformally transforming the
boundary coordinates// $\epsilon_R=\epsilon \to
\epsilon_R=\epsilon \sqrt{ G'_+(\tilde{v})G'_-(\tilde{w})}$, $\epsilon_L=
\epsilon \to \epsilon_L= \epsilon \sqrt{H'_+(\tilde{u}) H'_-(\tilde{\omega})}$. Using
these new values of $\epsilon_{L,R}$, we get
\begin{eqnarray}
L(\bold{\tilde P},\bold{\tilde Q'})&=&\log\left[\frac{4\cosh[\sqrt{L}(G_+(\tilde{v})-H_+(\tilde{u}))/2]}{\sqrt{L}\epsilon\sqrt{G'_+(\tilde{v})H'_+(\tilde{u})}}\frac{\cosh[\sqrt{L}(G_-(\tilde{w})-H_-(\tilde{\omega}))/2]}{\sqrt{L}\epsilon\sqrt{ G'_-(\tilde{w})H'_-(\tilde{\omega})}}\right]\
\label{geodesic1}
\end{eqnarray}
Similarly,
\begin{eqnarray}
L(\bold{\tilde P_1},\bold{\tilde P_2})
&=&\log\left[\frac{4\sinh[\sqrt{L}(G_{+}(\tilde{v}_1)-G_{+}(\tilde{v}_2))/2]}{\sqrt{L}\epsilon\sqrt{G'_+(\tilde{v}_1)G'_{+}(\tilde{v}_2)}}\frac{\sinh[\sqrt{L}(G_{-}(\tilde{w}_1)-G_{-}(\tilde{w}_2))/2]}{\sqrt{L}\epsilon\sqrt{ G'_{-}(\tilde{w}_1)G'_{-}(\tilde{w}_2)}}\right]\nonumber\\
\label{geodesic121}
\end{eqnarray}
\subsection{General two-point correlators from CFT}
\subsubsection*{In the thermofield double state}
\underbar{RL correlator:} For the eternal BTZ string, the coordinate
transformations from the EF to Poincare (see Appendix
\ref{sec-poincare}) reduce, at the boundary, to a conformal
transformation from the Rindler to Minkowski coordinates, so that the
boundary of the right (left) exterior maps to the right (left)
Rindler wedge \cite{Hartman:2013qma}. It is expedient
to compute the CFT correlations first in the Minkowski plane, and then
conformally transform the result to Rindler coordinates. Using this
method of \cite{Hartman:2013qma}, we get the following result
\begin{eqnarray}
\label{RL-corr}
\langle \psi_0| O(X_{+R},X_{-R})\, O(X_{+L},X_{-L})
| \psi_0 \rangle &=&\frac{(\sqrt{L}e^{\sqrt{L}v})^h(\sqrt{L}e^{-\sqrt{L}w})^{\bar{h}}(-\sqrt{L}e^{\sqrt{L}u})^h(-\sqrt{L}e^{-\sqrt{L}\omega})^{\bar{h}}}{(\frac{e^{\sqrt{L}v}+e^{\sqrt{L}u}}{\epsilon})^{2h}(\frac{-e^{-\sqrt{L}w}-e^{-\sqrt{L}\omega}}{\epsilon})^{2\bar{h}}}\nonumber\\
&=&\Big{(}\frac{4\cosh\,[\sqrt{L}(v-u)/2]\,\cosh\,[\sqrt{L}(w-\omega)/2]}{L\epsilon^2}\Big{)}^{-2h}\nonumber\
\end{eqnarray}
where the operator $O$ is assumed to have dimensions $(h,\bar{h})$ and
we have used a real space field theory cut-off $\epsilon$. We have related
the temperature of the CFT to $L(={\bar L})$ by the equation
$\sqrt{L}=2\pi/\beta$. //
It is easy to see that this correlator satisfies the relation \eq{geodesics}
\begin{eqnarray}
\label{2pgeoRL}
\langle \psi_0| O(X_{+R},X_{-R})\, O(X_{+L},X_{-L})
| \psi_0 \rangle =e^{-2h L(\bold{P},\bold{Q})}\
\end{eqnarray}
where in the expression on the right hand side for the geodesic length
(\ref{geodesic}), we use $\epsilon_R=\epsilon_L=\epsilon$ as explained before.
\underbar{RR correlator:} By
following steps similar to the above, the two-point
correlator between the points (\ref{bp12}) is given by
\begin{eqnarray}
\langle \psi_0|\mathcal{O}(X_{+1},X_{-1})\, \mathcal{O}(X_{+2},X_{-2})| \psi_0 \rangle
&=&\frac{(\sqrt{L}e^{\sqrt{L}v_1})^h(\sqrt{L}e^{-\sqrt{L}w_1})^{\bar{h}}(\sqrt{L}e^{\sqrt{L}v_2})^h(\sqrt{L}e^{-\sqrt{L}w_2})^{\bar{h}}}{(\frac{(e^{\sqrt{L}v_1}-e^{\sqrt{L}v_2}}{\epsilon})^{2h}(\frac{-e^{-\sqrt{L}w_1}+e^{-\sqrt{L}w_2}}{\epsilon})^{2\bar{h}}}\nonumber\\
&=&\Big{(}\frac{4\sinh\,[\sqrt{L}(v_1-v_2)/2]\,\sinh\,[\sqrt{L}(w_1-w_2)/2]}{L\epsilon^2}\Big{)}^{-2h}\nonumber\
\end{eqnarray}
It follows, therefore, that
\begin{eqnarray}
\label{2pgeo12}
\langle \psi_0|\mathcal{O}(X_{+1},X_{-1})\, \mathcal{O}(X_{+2},X_{-2})
| \psi_0 \rangle&=&e^{-2hL(\bold{P_1},\bold{P_2})}\
\end{eqnarray}
where, again, the geodesic length on the right hand side is read off
from \eq{geodesic1} with $\epsilon_1=\epsilon_2=\epsilon$.
\subsubsection*{In the new states}
As explained in \eq{transform-corr}, correlators in the state $| \psi
\rangle$ \eq{thermofield-general} can be computed by using a conformal
transformation \eq{primary} of the operators. The new correlator is,
therefore, found from the old one \eq{RL-corr} by a conformal
transformation of the boundary coordinates and an inclusion of the
Jacobian factors. The latter has, in fact, the effect of the
replacement $\epsilon^2 \to \epsilon^2 \sqrt{ G'_+(\tilde{v})
G'_-(\tilde{w}) H'_+(\tilde{u})
H'_-(\tilde{\omega})}$. With these ingredients, it is
straightforward to verify that \eq{geodesics-new} is satisfied.
Similar arguments apply to {\it RR} and {\it LL} correlators.
\section{Entanglement entropy\label{entang}}
We define an entangling region $A= A_R\cup A_L$, where $A_R$ is a half
line $(v-w)/2>x_R$ on the right boundary at `time' $(v+w)/2=t_R$ and
$A_L$ is a half line $(u-\omega)/2>x_L$ of the left boundary at `time'
$(u+\omega)/2=t_L$. The boundary of the region $A$ consists of a
point $P(v_{\partial A},w_{\partial A})$ on the right and a point
$Q'(u_{\partial A},\omega_{\partial A})$ on the left, with coordinates
\begin{eqnarray}
\label{bpent}
P:\kern10pt v_{\partial A}&=&t_R+x_R,\quad w_{\partial A}=t_R-x_R\\
Q':\kern10pt
u_{\partial A}&=&t_L+x_L,\quad \omega_{\partial A}=t_L-x_L\nonumber\
\end{eqnarray}
\subsection*{Bulk calculations}
\subsubsection*{In the BTZ geometry}
We calculate the entanglement entropy $S_A$ of the region A
using the holographic entanglement formula of \cite{Ryu:2006bv,Hubeny:2007xt}.
The HEE is given in terms of the geodesic length $L({\bf P,
Q'})$. The geodesic length, as calculated in (\ref{geodesic}), is
\begin{eqnarray}
L(\bold{P},\bold{Q'})=\log\left[\frac{4\cosh[\sqrt{L}(v_{\partial A}-u_{\partial A})/2]\cosh[\sqrt{L}(w_{\partial A}-\omega_{\partial A})/2]}{M\epsilon^2}\right]\nonumber\\
\end{eqnarray}
The HEE is then given by $S_A=L(\bold{P},\bold{Q'})/4G_3$. Using \eq{newton},
we get
\begin{eqnarray}
S_A=\frac{c}{6}\log\left[\frac{4\cosh[\sqrt{L}((t_R+x_R)-(t_L+x_L))/2]\cosh[\sqrt{L}((t_R-x_R)-(t_L-x_L))/2]}{M\epsilon^2}\right]\
\label{holoEE}
\end{eqnarray}
Note that for $x_R=x_L=0$ and $t=t_R=-t_L$ (which correspond to a
non-trivial time evolution in the geometry) the HEE \eq{holoEE}
reduces to
\begin{eqnarray}
S_{A}= \frac{c}{3}\log\left[\cosh\,\frac{2\pi t}{\beta}\Big{]}+\frac{c}{3}\log\Big{[}\frac{\beta/\pi}{\epsilon}\right]\
\label{HMresult}
\end{eqnarray}
which reproduces the result for the HEE in
\cite{Hartman:2013qma}.\footnote{The UV cutoff in
\cite{Hartman:2013qma} is half of the cutoff, $\epsilon$ used here.}
\subsubsection*{In the new geometries}
The HEE corresponding to the conformally transformed state
\eq{thermofield-general} is given by the length $L({\bf \tilde P,
\tilde Q'})$ connecting the end-points $P$ and $Q'$ in the
new geometries described in Section \ref{full-metric}. Working on
lines similar to the derivation of \eq{geodesic12}, the HEE
is given by
\begin{eqnarray}
S_A&=&\frac{c}{6}\log\Big{[}\frac{4\cosh[\sqrt{L}(G_+(\tilde{t}_R+\tilde{x}_R)-H_+(\tilde{t}_L+\tilde{x}_L))/2]}{\sqrt{L}\epsilon\sqrt{G'_+(\tilde{t}_R+\tilde{x}_R)H'_+(\tilde{t}_L+\tilde{x}_L)}}\nonumber\\
&&\quad\quad\quad\quad\frac{\cosh[\sqrt{L}(G_-(\tilde{t}_R-\tilde{x}_R)-H_-(\tilde{t}_L-\tilde{x}_L))/2]}{\sqrt{L}\epsilon\sqrt{ G'_-(\tilde{t}_R-\tilde{x}_R)H'_-(\tilde{t}_L-\tilde{x}_L)}}\Big{]}\
\label{holoEE1}
\end{eqnarray}
\subsection*{CFT calculations}
\subsubsection*{In the thermofield double state}
The technique of calculating the entanglement entropy in the
thermofield double state is well-known \cite{Cardy:1986ie}. The Renyi entanglement
entropy $S_A^{(n)}$ of the region A (\ref{bpent}) is given by the trace
of the $n^{th}$ power of the reduced density matrix $\rho^n_A$. The
latter can be shown to be a Euclidean path integral on an $n$-sheeted Riemann
cylinder. This can then be calculated in terms of the two point correlator, on a complex
plane, of certain twist fields $\mathcal{O}$, with conformal dimensions
\begin{eqnarray}
h=\frac{c}{24}(n-1/n),\quad \bar{h}=\frac{c}{24}(n-1/n)\nonumber\\
\end{eqnarray}
inserted at the end-points $(P, Q')$ of A. The two-point correlator is
given by a calculation similar to that in the previous section. Thus,
\begin{eqnarray}
S_A^{(n)} &=&
\langle
\mathcal{O}_R(v_{\partial A},w_{\partial A})\,\mathcal{O}_L(u_{\partial A},
\omega_{\partial A}) \rangle \nonumber\\
&=&\frac{(\sqrt{L})^{2h+2\bar{h}}}{(4\cosh[\sqrt{L}((t_R+x_R)-(t_L+x_L))/2]/\epsilon)^{2h}(\cosh[\sqrt{L}((t_R-x_R)-(t_L-x_L))/2]/\epsilon)^{2\bar{h}}}\nonumber\
\end{eqnarray}
The entanglement entropy $S_A=-\partial_n S_A^{(n)}|_{n=1}$ is
\begin{eqnarray}
S_A=\frac{c}{6}\log\left[\frac{4\cosh[\sqrt{L}((t_R+x_R)-(t_L+x_L))/2]\cosh[\sqrt{L}((t_R-x_R)-(t_L-x_L))/2]}{L\epsilon^2}\right]\
\label{CFTEE}
\end{eqnarray}
This proves that the CFT entanglement entropy and holographic entanglement
entropy(\ref{holoEE}) are equal.
\subsubsection*{In the new states}
The EE of the region A, computed in the new state
\eq{thermofield-general}, is given in terms of the
conformally transformed two-point function described in
\eq{transform-corr}. The conformally transformed points are given by
\begin{eqnarray}
v_{\partial A}=G_+(\tilde{v}_{\partial A})=G_+(\tilde{t}_R+\tilde{x}_R), &&\quad w=G_-(\tilde{w}_{\partial A})=G_-(\tilde{t}_R-\tilde{x}_R)\nonumber\\
u_{\partial A}=H_+(\tilde{u}_{\partial A})=H_+(\tilde{t}_L+\tilde{x}_L), &&\quad \omega=H_-(\tilde{\omega}_{\partial A})=H_-(\tilde{t}_L-\tilde{x}_L)\nonumber\
\end{eqnarray}
It follows that the entanglement entropy is
\begin{eqnarray}
S_{A,CFT}&=&\frac{c}{6}\log\Big{[}\frac{4\cosh[\sqrt{L}(G_+(\tilde{t}_R+\tilde{x}_R)-H_+(\tilde{t}_L+\tilde{x}_L))/2]}{\epsilon\sqrt{L}\sqrt{G'_+(\tilde{t}_R+\tilde{x}_R)H'_+(\tilde{t}_L+\tilde{x}_L)}}\nonumber\\
&&\quad\quad\quad\quad\frac{\cosh[\sqrt{L}(G_-(\tilde{t}_R-\tilde{x}_R)-H_-(\tilde{t}_L-\tilde{x}_L))/2]}{\epsilon\sqrt{L}\sqrt{ G'_-(\tilde{t}_R-\tilde{x}_R)H'_-(\tilde{t}_L-\tilde{x}_L)}}\Big{]}\
\label{CFTEE1}
\end{eqnarray}
which matches with the HEE (\ref{holoEE1}).
\subsection{\label{example}Dynamical entanglement entropy
in a specific new geometry}
We now compute the entanglement entropy in an illustrative geometry
specified by a particular choice of the functions $G_\pm$ and $H_\pm$.
In this example, we take
\begin{eqnarray}
x_R=0,\quad t_R=t,\quad x_L=0,\quad t_L=-t\nonumber\
\end{eqnarray}
For simplicity, we consider $G_\pm$ and $H_\pm$ which satisfy
\begin{eqnarray}
G_+(x)\equiv G_-(x)\equiv G(x), \quad H_+(x)\equiv H_-(x)\equiv H(x)\nonumber\
\end{eqnarray}
With the transformations given above, we have
\begin{eqnarray}
\tilde{x}_R=0,\quad\tilde{v}_{\partial A}=\tilde{w}_{\partial A}=\tilde{t}_R=\tilde{t},\quad\quad \tilde{x}_L=0,\quad\tilde{u}_{\partial A}=\tilde{\omega}_{\partial A}=\tilde{t}_L=-\tilde{t}\
\label{ttilde}
\end{eqnarray}
The expression for the HEE (\ref{holoEE1}) then reduces to
\begin{eqnarray}
S_A=\frac{c}{3}\log\left[\frac{2\cosh[\sqrt{L}(G(\tilde{t})+H_1(\tilde{t}))/2]}{\epsilon\sqrt{L}\sqrt{G'(\tilde{t})H'_1(\tilde{t})}}\right]\
\label{holoEEHMG}
\end{eqnarray}
where we have defined the notation $-H(-\tilde{t})=H_1(\tilde{t})$.
\begin{figure}[H]
\centerline{\includegraphics[scale=.2]{HEE_plot}}
\caption{Time evolution of HEE. The red-line represents the linear growth of HEE for a region consisting of spatial half-lines of both sides of a constant 2-sided BTZ geometry.
The blue-line represents the HEE growth of the region consisting of half-lines of both sides of the SGD transformed geometry, for $G(\tilde{t})=\tilde{t}+\frac{1}{6}\cos(3\tilde{t})$ and $H_1(\tilde{t})=\tilde{t}+\frac{3}{5}\sin(\tilde{t})$. The undulating curve can be explained in terms of the quasiparticle
picture
of \cite{Calabrese:2005in}; the entanglement
entropy departs from its usual linear behaviour as the
quasiparticle pairs locally go out and back in to the entangling region
as the region is subjected
to a conformal transformation.}
\label{HEEplot}
\end{figure}
\section{Entropy\label{sec-S}}
As discussed in previous sections, our solutions of Section
\ref{full-metric} are characterized by a smooth, albeit undulating,
horizon (see Figure \ref{fig-horizons}). This allows us, following
\cite{Bhattacharyya:2008xc}, to define a holographic entropy
current. We will first review the equilibrium situation (static black
string), and then describe the calculation for the general,
time-dependent solution. We will include a comparison with CFT
calculations in both cases.
\subsection{Equilibrium}
\paragraph{Bulk calculation:}
In case $L= {\bar L}=$ constant, our solutions represent BTZ black strings
\eq{EF} with a horizon at $\lambda=\lambda_0$. The horizon ${\cal H}$
is a two-dimensional null surface, described by the metric
\begin{equation}
ds^2|_{{\cal H}} \equiv H_{\mu\nu}dx^{\mu}dx^{\nu}
= \left(\sqrt{L}dv/2 - \sqrt{{\bar L}} dw/2\right)^2
\label{induced}
\end{equation}
Since the normal to ${\cal H}$ at any point, given by
$n^M = \partial^M \lambda (M=\{\lambda, v,w\})$, also lies on ${\cal H}$,
${\cal H}$ possesses a natural coordinate system $(\tau, \alpha)$
where $\alpha$ labels the one-parameter family of null geodesics,
and $\tau$ measures the affine distance along the geodesics.
In such a coordinate system, we get, by construction
\begin{equation}
ds^2|_{{\cal H}} = g d\alpha^2
\label{natural}
\end{equation}
The area 1-form and the entropy current on the horizon are defined
by the equations \cite{Bhattacharyya:2008xc}
\footnote{Our convention for $\epsilon_{\mu\nu}$ is $\epsilon_{vw}=-1$.},
\begin{align}
a \equiv 4 G_3\epsilon_{\mu\nu}J_S^{\mu}dx^{\nu} = \sqrt{g} d\alpha,
\label{area}
\end{align}
By inspection, from \eq{induced} and \eq{natural}, we find the
following expressions for the area-form and the entropy current
\begin{align}
a &= \sqrt{L}dv/2 - \sqrt{{\bar L}} dw/2
\nonumber \\
J_s^{v} &= \frac{1}{8G_3} \sqrt{{\bar L}},\;
J_S^{w} = \frac{1}{8G_3} \sqrt L
\label{area-val}
\end{align}
The holographic entropy current on the boundary ${\cal B}$ is obtained
by using a map $f:{\cal B} \to {\cal H}$ and pulling back the
area-form (or alternatively the entropy current $J_{S,\mu}$) from
the horizon to the boundary. It turns out \footnote{The map $f$
is defined by shooting `radial' null geodesics inwards from the
boundary, and is found to be of the form $f:(\lambda_{ir}, v, w)
\mapsto$ $(\lambda_{ir}, v+ C_1, w+C_2)$.}
that the natural pull back retains the form of the area-form
or entropy current, namely the expressions \eq{area-val}
still hold at the boundary.
To find the entropy density, we define the boundary coordinates $t=
(v+w)/2, x= (v-w)/2$ (see Section \ref{entang}), (so that
\eq{bdry-lam} has the canonical form $-dt^2 + dx^2$). With this the
entropy density becomes
\begin{equation}
s \equiv J_S^T = \frac{1}{8G_3} \left( \sqrt L +
\sqrt{{\bar L}} \right)
\label{cardy-b}
\end{equation}
\paragraph{CFT calculation:}
The entropy density from the Cardy formula is \footnote{Recall that
both $T_{vv}, T_{ww}$ are constant in this case. The more familiar
form of \eq{cardy}, for a circular spatial direction of length
$2\pi$, is obtained by putting $S= 2\pi s$, $L_0 = 2\pi T_{vv}$, and
$\bar L_0= 2\pi T_{ww}$, which gives $S= 2\pi(\sqrt{c L_0/6}
+ \sqrt{c {\bar L}_0/6}).$ }
\begin{equation} s= \sqrt{c \pi T_{vv}/3} + \sqrt{c \pi T_{ww}/3}
\label{cardy}
\end{equation}
Using the identification \eq{newton} and \eq{lam-stress},
we can easily see that the two expressions \eq{cardy-b}
and \eq{cardy} exactly match.
\subsection{New metrics: non-equilibrium entropy}
\paragraph{Bulk calculation:} We will now follow a similar procedure
as above, for the general solution in Section \ref{full-metric}.
We find that (in coordinate chart EF1)
\begin{equation}
ds^2|_\mathcal{H}=\frac14 d\alpha^2=\frac14 (\sqrt{L}G_+'({\tilde v})d{\tilde v}-\sqrt{{\bar L}}G_-'({\tilde w})d{\tilde w})^2
\end{equation}
leading to the following area one form on the horizon
\begin{equation}
a= \frac12 \sqrt{L}G_+'({\tilde v})d{\tilde v}-\frac12\sqrt{{\bar L}}G_-'({\tilde w})d{\tilde w}
\label{areaform}
\end{equation}
Note that this could alternatively be obtained from the area
form in \eq{area-val} by a diffeomorphism.
The resulting expression for the entropy current, following the
steps above, is
\begin{equation}
{\tilde J}_s^{{\tilde v}} = \frac{1}{8G_3} \sqrt{{\bar L}} G'_-({\tilde w}),\;
{\tilde J}_S^{{\tilde w}} = \frac{1}{8G_3} \sqrt L G'_+({\tilde v})
\end{equation}
Let us define, as before, the spacetime coordinates as $\tilde x,
\tilde t$ with $(\tilde v, \tilde w)$ = $\tilde t \pm \tilde x$.
The entropy density is then given by
\begin{equation}
\tilde s= {\tilde J}_S^{\tilde t}=\frac{1}{4G_3 }\bigg(\frac12
\sqrt L G_+'({\tilde v})+\frac12 \sqrt {\bar L} G_-'({\tilde w})\bigg)
\label{current}
\end{equation}
Note that the entropy current is divergenceless
\begin{equation}
\partial_\mu {{\tilde J}}^{\mu}_S=
\partial_{{\tilde v}}{{\tilde J}}^{{\tilde v}}_S+\partial_{{\tilde w}}{\tilde J}^{{\tilde w}}_S=0
\label{no-divergence}
\end{equation}
This has two implications:
\begin{enumerate}
\item \underbar{No dissipation:}
We have entropy transfers between different regions with no
net entropy loss or production (see Figure \ref{entropy}).
\begin{figure}[H]
\centerline{\includegraphics[scale=.5]{entropy-a}}
\caption{The undulating horizon of Figure \ref{fig-ef} leads to the
non-trivial entropy current \eq{current}. In this figure, we plot the
entropy density $\tilde s$ as a function of ${\tilde v}, {\tilde w}$ for the right
CFT. Note that although the entropy density fluctuates, the
entropy flow here is such that there is no net entropy
production (or destruction) (see Eq. \eq{no-divergence}).}
\label{entropy}
\end{figure}
\item \underbar{Total entropy is not changed by the
conformal transformation:}
The other implication is that the integrated entropy
over a space-like (or null) slice $\Sigma$
\begin{equation}
\tilde S= \int_{\Sigma} \epsilon_{\mu\nu} J_S^\mu d\sigma^\nu
\end{equation}
is independent of the choice of the slice. In particular,
choosing the slice to be $\Sigma_0: t=v+w= 0$, we get
\begin{align}
\tilde S &=
\frac1{8G_3}\ \int_{\Sigma_0} \left(\sqrt L G'_+({\tilde v}) d{\tilde v} - \sqrt {{\bar L}} G_-'({\tilde w})
d{\tilde w} \right)
=\frac1{8G_3}\ \int_{\Sigma_0} \left(\sqrt L dv - \sqrt {{\bar L}}
dw \right)
\\
\;& =\frac1{8G_3}\ \int dx \left(\sqrt L + \sqrt {{\bar L}} \right)
= \int dx\ s = S
\label{manipulate-S}
\end{align}
Hence although the entropy density is clearly transformed,
the total entropy is not changed by the
conformal transformation.
\end{enumerate}
\subsubsection*{CFT calculation:}
In a non-equilibrium situation, there is no natural notion of an
entropy. However under the adiabatic approximation, the instantaneous
eigenstates of a time-dependent Hamiltonian are a fair representation
of the actual time-dependent wave functions. The consequent energy
level density can thus be used to define an approximate time-dependent
entropy. Generalizing this principle to slow time {\it and}
space variations, and applying this to the stress tensor, one
expects a space-time dependent version of \eq{cardy}, namely
\begin{equation}
\tilde s=\sqrt{\frac{\pi c}{3}\tilde T_{{\tilde v}\vt}}+
\sqrt{\frac{\pi c}{3}\tilde T_{{\tilde w}\wt}}
\label{adiab}
\end{equation}
where the stress tensors are given by \eq{stress}.
Since we have made the adiabatic approximation, we expect
the above formula to be valid only up to the leading order
of space and time derivatives. Under this approximation,
we have
\begin{align}
8\pi G_3 T_{{\tilde v}\vt} &=\frac{L}{4} G_+'({\tilde v})^2,
\quad
8\pi G_3 T_{{\tilde w}\wt} =\frac{{\bar L}}4 G_-'({\tilde w})^2
\label{stress-adiab}
\end{align}
which exactly agrees with the holographic entropy density
in \eq{current}. \footnote{Note that
throughout this paper, we have not used the adiabatic
approximation anywhere else. Thus, it is unsatisfactory
to use this approximation here. It is, in fact,
tempting to believe that the entropy density in \eq{current},
and not that in \eq{adiab}, actually gives the CFT entropy in general;
however, this requires more investigation.}
\gap1
\underbar{Total entropy for ${\cal H}_R$ is unchanged
by the conformal transformation:}
\gap1
Under the conformal transformation \eq{transform-op}, the reduced
density matrix $\rho_R$ is changed by a unitary transformation:
\begin{equation}
\rho_R = \text{Tr}_{{\cal H}_L} | \psi \rangle \langle \psi | = U_R\, \rho_{0,R}\,
U_R^\dagger, \kern10pt \rho_{0,R}= \text{Tr}_{{\cal H}_L} | \psi_0 \rangle \langle \psi_0 |
\label{rho-r}
\end{equation}
The total entropy of the system after the transformation is given by
the von Neumann entropy $\tilde S= -\text{Tr} \rho_R \ln \rho_R$ which,
therefore, is equal to the entropy before; it is unchanged by the
unitary transformation.
\section{Conclusion and open questions\label{discuss}}
In this paper we have solved the boundary value problem for 3D gravity
(with $\Lambda<0$) with independent boundary data on two
asymptotically AdS$_3$ exterior geometries. The boundary data,
specified in the form of arbitrary holographic stress tensors,
yields spacetimes with wormholes, {\it i.e.} with exterior regions
connected across smooth horizons. The explicit
metrics are constructed by the technique of solution generating
diffeomorphisms (SGD) from the eternal BTZ black string. By using the
fact that the SGD's reduce to conformal transformations at both
boundaries, we claim that the dual CFT states are specific
time-dependent entangled states which are conformal transformations of the
standard thermofield double. We compute various correlators and a
dynamical entanglement entropy, in the bulk and in the CFT, to provide
evidence for the duality. We also arrive at an expression for a
non-equilibrium entropy function from the area-form on the horizon of
these geometries.
Our work has implications for a number of other issues. We briefly
discuss two of them below; a detailed study of these is left to future
work.
\subsection{\label{EPR}ER=EPR}
As mentioned above, our work constructs an infinite family of AdS-CFT
dual pairs in which quantum states entangling two CFTs are
holographically dual to spacetimes containing a wormhole region which
connects the two exteriors. Both the quantum states and the wormhole
geometries are explicitly constructed (see eqns.
\eq{thermofield-general} and (\ref{newmetric},\ref{newmetric-2})).
Our examples generalize the construction in
\cite{Maldacena:2001kr,Hartman:2013qma,Caputa:2013eka}\footnote{See \cite{Maldacena:2013xja,
Shenker:2013pqa,Shenker:2013yza}}
(for other remarks on unitary transformations of the
thermofield double and related geometries see \cite{Shenker:2013pqa,Marolf:2013dba,
Shenker:2013yza,Avery:2013bea,Balasubramanian:2014gla}) and provide
an infinite family of examples of the relation ER=EPR, proposed in
\cite{Maldacena:2013xja}. Since this relation has been extensively
discussed and debated in the literature (\cite{Shenker:2013pqa,Shenker:2013yza,
Avery:2013bea,Balasubramanian:2014gla}), we would like to make
some specific points pertaining to some of these discussions.
\subsubsection*{\label{no-suppression}RR correlators vs RL correlators}
It has been argued in \cite{Balasubramanian:2014gla},\cite{Shenker:2013pqa} and
\cite{Shenker:2013yza} that for typical
entangled states connecting two CFTs, $\cal{H}_R$ and $\cal{H}_L$,
correlators involving operators on the left and the right are
suppressed relative to those involving operators all on the right. In
particular, according to \cite{Balasubramanian:2014gla}, correlators of the form $\langle O_R
O_L \rangle$ are of the order $e^{-S} \langle O_R O_R \rangle$, where $S$ is
the entropy of the right sided Hilbert space.
In Section \ref{2-pt} we have computed general two-point functions,
both of the kind $\langle O_R(P) O_R(Q) \rangle$ and $\langle O_R(P) O_L(Q')
\rangle$.\footnote{We use unprimed labels for operators on the right and
primed labels for those on the left.} In case of the eternal BTZ
(dual to the standard thermofield double), an inspection of
\eq{geodesic} and \eq{geodesic12} suggests that as the boundary point
${\bf P}$ goes off to infinity, the $\cosh$ and $\sinh$ factors tend to be
equal, thus $L({\bf P},{\bf Q}) \approx L({\bf P}, {\bf Q}')$, thus there is no extra
suppression in the two-sided correlator $\langle O_R\,O_L \rangle$. Of
course, such a statement, regarding the standard thermofield double,
has been regarded as somewhat of a special nature.
We are therefore naturally led to ask: what happens in case of the new
solutions found in this paper? The geodesic lengths
$L({\bf P}, {\bf Q})$ and $L({\bf P}, {\bf Q}')$ are now given by \eq{geodesic1} and
\eq{geodesic121}. Once again, if the point P goes off towards the
boundary of the Poincare plane, ${\tilde v} \to \infty$. Hence $G_+({\tilde v}) \to
\infty$ (since $G_+$ is a monotonically increasing function). Hence,
both the geodesic lengths approach each other. Thus, we do not see any
peculiar additional suppression, even for our general entangled state,
arising when the second point of the correlation function is moved
from the right to the left CFT.
\subsubsection*{On the genericity of our family of examples}
We start with the following Lemma.
\noindent {\it Lemma:}
Any state $\in {\cal H} \otimes {\cal H}$,
\begin{equation}
| \Psi \rangle = \sum_{i,j} C_{ij} | i \rangle | j \rangle, \kern5pt C_{ij}\in
{\mathbb C},
\label{entang-gen}
\end{equation}
can be expressed in the form
\begin{equation}
| \Psi \rangle = \sum_{i,j,n} e^{-\lambda_n} U_{L,in} U_{R,jn} | n \rangle | n \rangle
\label{entang-gen-a}
\end{equation}
where $U_R, U_L$ are two unitary operators and
$\lambda_n \ge 0$.
\gap1
\noindent {\it Proof:} Using the canonical map ${\cal H} \otimes {\cal H}$ $
\to {\cal H} \otimes {\cal H}^* $, we can regard the above state $|
\Psi \rangle$ as an operator ${\bf \Psi}$ in ${\cal H}$, with matrix
elements $C_{ij}$. Using the singular value decomposition theorem on a
general complex matrix, we can write $C= U_L D U_R^\dagger$ where $D$
is a diagonal matrix with real, non-negative entries. By denoting $D$
as diag[$e^{-\lambda_n}$], we get \eq{entang-gen-a}.
\gap1
The state \eq{entang-gen-a} can be regarded as a thermofield double
with Hamiltonian $H= $$ \sum_n \lambda_n/\beta$$ | n \rangle \langle n |$
transformed by unitary operators $U_L$ on the left and by $U_R$ on the
right. Thus, the above Lemma suggests that the most general entangled
state \eq{entang-gen} can be written as a unitary transformation of
{\it some} thermofield double state. Now, note that the state
\eq{entang-gen-a} is of the same general form as that of
\eq{thermofield-general} discussed in this paper. {\it Are our states
\eq{thermofield-general} the most general entangled states then?}
The answer is no, since the $U_{L,R}$ we use are made of Virasoro
generators (see Appendix \ref{unitary}), and are not the most general
unitaries of \eq{entang-gen-a}. However, in spite of this restriction,
it is clear that the states \eq{thermofield-general} do form a fairly
general class. Furthermore, if the states \eq{entang-gen-a} are states
in which only the stress tensor is excited, then indeed these states
are all contained in our class of states \eq{thermofield-general}.
\footnote{If a CFT dual to pure gravity were to exist, then our states
\eq{thermofield-general} in such a theory would indeed be the most
general state of the form \eq{entang-gen}. However, such a unitary
theory is unlikely to exist \cite{Maloney:2007ud,Gaberdiel:2007ve},
although chiral gravity theories which are dual to CFTs with only
the Virasoro operator have been suggested (see,
e.g. \cite{Afshar:2014rwa}). We would like to thank Justin David
for illuminating discussions on this point.}
\subsubsection*{Weakly entangled states}
To assess the genericity of our states, we ask a different question
now: do our set of states \eq{thermofield-general}, which are {\it
all} explicitly dual to wormholes, include those with a very small
entanglement entropy $S$ for a given energy $E$?\footnote{This
question was suggested to us by Sandip Trivedi.} The answer to this
question turns out to be yes. As we have noted in the remarks around
\eq{manipulate-S} and \eq{rho-r}, the entropy $S$, which is actually the
entanglement entropy of the right Hilbert space, is the same for all
our states. However, the same manipulations as in \eq{manipulate-S}
shows that the energy of these states are {\it not} the same; indeed
by choosing the derivatives $G_\pm'$ to be large, we can make the
energy of the transformed state to be much larger than that of the
standard thermofield double. Stated in another way, for states of a
given energy, our set of states includes states with entanglement
entropy much less than that of the thermofield double. This is
consistent with the proposal of \cite{Maldacena:2013xja} that even a small
entanglement is described by a wormhole geometry.
\subsection{Generalizations and open questions}
It would be interesting to rephrase the results in this paper in terms
of the $SL(2,R) \times SL(2,R)$ Chern-Simons formulation
\cite{Witten:1988hc} of three-dimensional gravity. By the arguments in
\cite{Witten:1988hc}, all diffeomorphisms (together with appropriate
local Lorentz rotations) can be understood as gauge transformations of
the Chern-Simons theory. The Chern-Simons formulation has been
extended to the gauge group $SL(N,R)$ $\times$ $SL(N,R)$ to describe
higher spin theories \cite{Banados:1998gg,Campoleoni:2010zq}. It would
be interesting to see whether the nontrivial gauge transformations in
our paper generalizes to these higher gauge groups, and hence to
higher spin theories. A possible application of our methods in this
case would be to compute HEE by the prescriptions in
\cite{Ammon:2013hba} and \cite{deBoer:2013vca} in the nontrivial
higher spin geometries\footnote{We thank Rajesh Gopakumar for a
discussion on this issue.}. We hope to come back to this issue
shortly.
The solutions presented in this paper are generated by SGDs which can
be regarded as forming a group ($\widetilde{\rm Vir}$
$\times$ $\widetilde{\rm Vir}$)$_L$ $\times$($\widetilde{\rm Vir}$
$\times$ $\widetilde{\rm Vir}$)$_R$. Here the first $\widetilde{\rm
Vir}$ denotes a group of SGDs which is parametrized by the function
$G_+$, and so on. As we emphasized in \eq{rho-r}, the reduced density
matrix on the right $\rho_R$ undergoes a unitary transformation under
this group of transformations, leaving the entropy unaltered. The
family of pure states \eq{thermofield-general} considered in this
paper can, therefore, be considered as an infinite family of
purifications of the class of density matrices $\rho_R$; it would be
interesting to see if these can be regarded as `micro-states' which can
`explain' the entropy of $\rho_R$. We hope to return to this issue
shortly.
It would also be interesting to use our work to explicitly study
various types of holographic quantum quenches involving quantum states
entangling two CFTs.\footnote{For a single CFT, a similar computation
was done in, e.g., \cite{Roberts:2012aq,Ugajin:2013xxa}.} It would
be of particular interest to study limiting cases of our solutions
which correspond to shock-wave geometries.
\subsection*{Acknowledgement}
We would like to thank Atish Dabholkar, Justin David, Avinash Dhar,
Rajesh Gopakumar, Juan Maldacena, Shiraz Minwalla, Suvrat Raju, Ashoke
Sen, Lenny Susskind, Sandip Trivedi, Tomonori Ugajin and Spenta Wadia
for many useful discussions during the course of this work. R.S. would
like to thank Sachin Jain, Nilay Kundu and V. Umesh for
discussions. We are grateful to Justin David, Avinash Dhar, Rajesh
Gopakumar, Juan Maldacena, Lenny Susskind and Spenta Wadia for
important feedbacks on the manuscript.
\section*{Appendix}
|
\section{Introduction}
\subsection{Background}
Let $d\geq1$. A Random Walk in a Random Environment (RWRE) on $\mathbb{Z}^{d}$
is defined by the following procedure: Let $\mathcal{M}_{d}$ denote
the space of all probability measures on $\mathcal{E}_{d}=\left\{ \pm e_{i}\right\} _{i=1}^{d}$
(the standard unit coordinate vectors), and define $\Omega=\left(\mathcal{M}_{d}\right)^{\mathbb{Z}^{d}}$.
An \textbf{environment} is a point $\omega\in\Omega$. For $x\in\mathbb{Z}^{d}$
and $e\in\mathcal{E}_{d}$ we denote by $\omega\left(x,e\right)$
the probability the measure $\omega\left(x\right)$ gives to $e$.
Let $P$ be an i.i.d measure on $\Omega$, in the sense that $P=\nu^{\mathbb{Z}^{d}}$
for some probability measure $\nu$ on $\mathcal{M}_{d}$. Throughout
this paper we assume that $P$ is uniformly elliptic, i.e., there
exists some $\eta>0$ such that for every $e\in\mathcal{E}_{d}$
\begin{equation}
P\left(\left\{ \omega\in\Omega\,:\,\omega\left(0,e\right)<\eta\right\} \right)=\nu\left(\left\{ \omega\,:\,\omega\left(e\right)<\eta\right\} \right)=0.\label{eq:elliptic_constant}
\end{equation}
For a given, fixed, environment $\omega\in\Omega$ and $x\in\mathbb{Z}^{d}$,
the \textbf{quenched }random walk on it (or the \emph{quenched law})
is a time homogeneous Markov chain on $\mathbb{Z}^{d}$ with transition
probabilities
\[
P_{\omega}^{x}\left(X_{n+1}=y+e\Big|X_{n}=y\right)=\omega\left(y,e\right),\quad\forall y\in\mathbb{Z}^{d},\, e\in\mathcal{E}_{d}
\]
and initial distribution $P_{\omega}^{x}\left(X_{0}=x\right)=1$.
We let $\mathbf{P}^{x}=P\otimes P_{\omega}^{x}$ be the joint law
of the environment and the walk, and define the \textbf{annealed }(or\textbf{
}\emph{averaged}) law as its marginal on the space of trajectories
\[
\mathbb{P}^{x}\left(\cdot\right)=\int_{\Omega}P_{\omega}^{x}\left(\cdot\right)dP\left(\omega\right).
\]
We use the notations $E$, $E_{\omega}^{x}$ and $\mathbb{E}^{x}$
for the expectations of the measures $P$, $P_{\omega}^{x}$ and $\mathbb{P}^{x}$
respectively.
In \cite{SZ99,Ze02} Sznitman and Zerner proved that the limiting
velocity of the random walk
\[
\mathbbm{v}=\lim_{n\to\infty}\frac{X_{n}}{n}
\]
exists for $P$ almost every environment and $P_{\omega}^{0}$ any
random walk on it. A question to remain open, which in fact is one
the most important open question in the field, is whether the limiting
velocity is an almost sure constant.
An important family of measures $P$ for the model is given by the
following definition:
\begin{defn}
The RWRE is said to be ballistic if the limiting velocity is a non
zero almost sure constant.
\end{defn}
\subsection{Conditions for ballisticity}
In \cite{Sz01,Sz02} Sznitman introduced two criteria for ballisticity
of RWRE, called conditions $(T)$ and $(T')$. In order to give a
formal definition of these conditions we need some preliminary definitions.
\begin{defn}
Let $\ell\in S^{d-1}:=\left\{ x\in\mathbb{R}^{d}\,:\,\left\Vert x\right\Vert _{2}=1\right\} $
be a direction in $\mathbb{R}^{d}$.
\begin{enumerate}
\item For $L>0$ and a sequence $\left\{ X_{n}\right\} $ (in $\mathbb{Z}^{d}$),
define
\[
T_{L}=T_{L}^{\left\{ \ell\right\} }\left(\left\{ X_{n}\right\} \right)=\inf\left\{ n\geq0\,:\,\left\langle X_{n},\ell\right\rangle \geq L\right\} ,
\]
where $\left\langle \cdot,\cdot\right\rangle $ is the standard inner
product in $\mathbb{R}^{d}$.
\item Similarly, for a set $A\subset\mathbb{Z}^{d}$ and a sequence $\left\{ X_{n}\right\} $
(in $\mathbb{Z}^{d}$), denote
\[
T_{A}=T_{A}\left(\left\{ X_{n}\right\} \right)=\inf\left\{ n\geq0\,:\, X_{n}\in A\right\}
\]
\end{enumerate}
\end{defn}
We can now state the definition of Sznitman's ballisticity conditions:
\begin{defn}
$ $~
\begin{enumerate}
\item Given $0<\gamma\leq1$, we say that $P$ satisfy condition $\left(T_{\gamma}\right)$
in direction $\ell_{0}\in S^{d-1}$ if for every $\ell\in S^{d-1}$
in some neighborhood of $\ell_{0}$ there exists a finite constant
$C$ such that
\[
\mathbb{P}^{0}\left(T_{L}^{\left(-\ell\right)}<T_{L}^{\left(\ell\right)}\right)<Ce^{-L^{\gamma}}.
\]
\item $P$ is said to satisfy condition $\left(T\right)$ if it satisfies
condition $\left(T_{1}\right)$.
\item $P$ is said to satisfy condition $\left(T'\right)$ if it satisfies
condition $\left(T_{\gamma}\right)$ for some $\frac{1}{2}<\gamma<1$.
\end{enumerate}
\end{defn}
\begin{rem}
It was shown in \cite{Sz02} that all the conditions $\left(T_{\gamma}\right)$
for $\frac{1}{2}<\gamma<1$ are equivalent.
\end{rem}
The relation between ballisticity and the above definition is given
by the following theorem and conjecture:
\begin{thm}[Sznitman \cite{Sz02}]
\label{thm:Sznitman_ballisticity_criterion} If condition $\left(T'\right)$
holds for some direction $\ell_{0}\in S^{d-1}$, then the RWRE is
ballistic, and the limiting velocity $\mathbbm{v}$ satisfies $\left\langle \mathbbm{v},\ell_{0}\right\rangle >0$.
In addition, under this assumption condition $\left(T'\right)$ holds
for all $\ell\in S^{d-1}$ satisfying $\left\langle \mathbbm{v},\ell\right\rangle >0$. \end{thm}
\begin{conjecture*}[Sznitman]
Condition $\left(T'\right)$ is equivalent to ballisticity.
\end{conjecture*}
In recent years several improvements of Theorem \ref{thm:Sznitman_ballisticity_criterion}
have been proved: In \cite{DR11} Drewitz and Ramírez showed that
for some $\gamma_{d}\in\left(0.366,0.388\right)$ which is dimension
dependent $\left(T_{\gamma}\right)$ for $\gamma\in\left(\gamma_{d},1\right)$
are all equivalent. In \cite[Theorem 1.4]{berger2008slowdown} Berger
showed that in dimension $d\geq4$ the condition $\left(T_{\gamma}\right)$
for $\gamma\in\left(0,1\right)$ implies ballisticity. In an additional
work \cite{DR12}, Drewitz and Ramírez showed that in dimension $d\geq4$
all the conditions $\left(T_{\gamma}\right)$ for $\gamma\in\left(0,1\right)$
are equivalent. In \cite{BDR12} Berger, Drewitz and Ramírez showed
that in fact (fast enough) polynomial decay (see Definition \ref{def:P_condition}
below) is equivalent to condition $\left(T_{\gamma}\right)$ for any
$0<\gamma<1$. Finally, in \cite{CR13} Campos and Ramírez proved
ballisticity for some non uniformly elliptic environments satisfying
(fast enough) polynomial decay.
\begin{defn}[Condition $\left(\mathscr{P}\right)$]
\label{def:P_condition} Let $N_{0}$ be an even integer larger than
$c_{0}=\exp\left(100+4d\left(\ln\eta\right)^{2}\right)$, where $\eta$
is the ellipticity constant defined in (\ref{eq:elliptic_constant}).
For a coordinate direction $\ell=\ell_{1}$ let $\ell_{2},\ldots,\ell_{d}$
be any fixed completion of $\ell_{1}$ to an orthonormal basis of
$\mathbb{R}^{d}$ and define
\[
\mbox{Box}_{x}=\left\{ y\in\mathbb{Z}^{d}\,:\,-\frac{N_{0}}{2}<\left\langle y-x,\ell\right\rangle <N_{0}\,,\,\left\langle y-x,\ell_{j}\right\rangle <25N_{0}^{3}\,\,\forall2\leq j\leq d\right\} ,
\]
\[
\widetilde{\mbox{Box}}_{x}=\left\{ y\in\mathbb{Z}^{d}\,:\,\frac{1}{3}N_{0}\leq\left\langle y-x,\ell\right\rangle <N_{0}\,,\,,\left\langle y-x,\ell_{j}\right\rangle <N_{0}^{3}\,\,\forall2\leq j\leq d\right\} .
\]
\[
\partial\mbox{Box}_{x}=\left\{ y\in\mathbb{Z}^{d}\backslash\mbox{Box}_{x}\,:\,\exists y\in\mbox{Box}_{x}\mbox{ such that }\left\Vert y-x\right\Vert _{1}=1\right\}
\]
and
\[
\partial_{+}\mbox{Box}_{x}=\left\{ y\in\partial\mbox{Box}_{x}\,:\,\left\langle y-x,\ell\right\rangle \geq N_{0}\,,\,\left|\left\langle y-x,\ell_{j}\right\rangle \right|<25N_{0}^{3}\,\,\forall2\leq j\leq d\right\} .
\]
Fix $M>0$ and $\ell\in S^{d-1}$. We say that condition $\mathscr{P}_{M}|\ell$
is fulfilled if
\[
\sup_{x\in\widetilde{\mbox{Box}}_{0}}\mathbb{P}^{x}\left(T_{\partial\mbox{Box}_{0}}\neq T_{\partial_{+}\mbox{Box}_{0}}\right)<\frac{1}{N_{0}^{M}}
\]
holds for some $N_{0}\geq c_{0}$. We say that condition $\mathscr{P}$
holds in direction $\ell$ if condition $\mathscr{P}_{M}|\ell$ holds
for some $M>15d+5$.
\end{defn}
\begin{defn}
Throughout this paper we denote by $\left(\mathscr{P}\right)$ the
following equivalent conditions:
\begin{itemize}
\item $\left(T'\right)$.
\item $\left(T_{\gamma}\right)$ for some $\gamma\in\left(0,1\right)$.
\item $\left(T_{\gamma}\right)$ for all $\gamma\in\left(0,1\right)$.
\item $\left(\mathscr{P}\right)$.
\end{itemize}
\end{defn}
\subsection{The environment viewed from the particle}
Let $\left\{ X_{n}\right\} $ be a RWRE. The \textbf{environment viewed
from the particle} is the discrete time process $\left\{ \overline{\omega}_{n}\right\} $
defined on $\Omega$ by
\[
\overline{\omega}_{n}=\sigma_{X_{n}}\omega,
\]
where for $x\in\mathbb{Z}^{d}$ we denote by $\sigma_{x}$ the shift
in direction $x$ in $\omega$, i.e. $\sigma_{x}\omega\left(y,\cdot\right)=\omega\left(x+y,\cdot\right)$
for every $y\in\mathbb{Z}^{d}$.
Beside the fact that the environment viewed from the particle process
takes values in a compact space it has the advantage of being Markovian,
c.f. \cite{bolthausen2002satic}, with respect to the transition kernel
\begin{equation}
\mathfrak{R}g\left(\omega\right)=\sum_{e\in\mathcal{E}_{d}}\omega\left(0,e\right)g\left(\sigma_{e}\omega\right),\label{eq:Transition kernel}
\end{equation}
defined for every bounded measurable function $g:\Omega\to\mathbb{R}$.
It is natural to ask what are the invariant measures of the Markov
chain $\left\{ \overline{\omega}_{n}\right\} $.
\begin{defn}
A probability measure $Q$ on $\Omega$ is said to be invariant (or
invariant with respect to the point of view of the particle), if for
every bounded continuous function $g:\Omega\to\mathbb{R}$
\begin{equation}
\int_{\Omega}\mathfrak{R}g\left(\omega\right)dQ\left(\omega\right)=\int_{\Omega}g\left(\omega\right)dQ\left(\omega\right).\label{eq:invariant_probability_measure}
\end{equation}
Equivalently, $Q$ is invariant if for every measurable event $A$
\[
Q\left(A\right)=\sum_{e\in\mathcal{E}_{d}}\omega\left(0,e\right)Q\left(\sigma_{-e}A\right).
\]
\end{defn}
One can find many examples for invariant measures with respect to
the process $\left\{ \overline{\omega}_{n}\right\} $. For example
every Dirac probability measure of any translation invariant environment
provides such an example. One additional method to obtain invariant
measures is by taking any sub-sequential limit of the Céasro means
$\left\{ \frac{1}{n}\sum_{k=0}^{n-1}\mathfrak{R}^{k}\nu\right\} $
where $\nu$ is any probability measure on $\Omega$ and $\mathfrak{R}\nu$
is the measure defined by the identity $\int_{\Omega}f\left(\omega\right)d\left(\mathfrak{R}\nu\right)\left(\omega\right)=\int_{\Omega}\mathfrak{R}f\left(\omega\right)d\nu\left(\omega\right)$
for every bounded measurable function $f:\Omega\to\mathbb{R}$.
As it turns out an invariant measure $Q$ is particularly useful when
it is also equivalent to the original measure $P$. In this case we
say that the static point of view (the one related to $P$) is equivalent
to the dynamic point of view (the one related to $Q$). If such a
measure exists it can be used to prove law of large numbers and central
limit theorem type results, see for example \cite{Ko85,KV86,bolthausen2002ten,SS04,Ze04,BB07,MP07,DR13}
and the references therein.
The existence of an invariant equivalent measure was proved in several
cases. In the one-dimensional case the existence of an equivalent
measure was proved by Alili \cite{Al99}. In the reversible case,
also known as random conductance model, the existence of an invariant
equivalent measure is a well known fact for most cases. For balanced
RWRE the existence of such a measure was proved by Lawler in \cite{La87}.
Later on this was strengthen to the case of balanced elliptic RWRE
by Guo and Zeitouni in \cite{GZ12} and even further to the non elliptic
case (for genuinely $d$-dimensional measures) by Berger and Deuschel
in \cite{BD14}. For Dirichlet random walks a classification for the
cases where such a measure exists was proved by Sabot in \cite{Sa13}.
The following result was proved by Kozlov in \cite{Ko85} (for the
proof see also \cite{bolthausen2002ten,DR13})
\begin{thm}[Kozlov \cite{Ko85}]
\label{thm:Kozlov's_theorem} Assume $P$ is elliptic%
\footnote{$P$ is called elliptic if $P\left(\min_{e\in\mathcal{E}_{d}}\omega\left(0,e\right)>0\right)=1$%
} and ergodic with respect to $\left\{ \sigma_{x}\right\} _{x\in\mathbb{Z}^{d}}$.
Assume there exists an invariant probability measure $Q$ for the
environment seen from the point of view of the particle which is absolutely
continuous with respect to $P$. Then the following hold:
\begin{enumerate}
\item $Q$ is equivalent to $P$.
\item The environment viewed from the particle with initial law $Q$ is
ergodic.
\item $Q$ is the unique invariant probability measure for the point of
view of the particle which is absolutely continuous with respect to
$P$.
\item The Céasro means $\left\{ \frac{1}{N+1}\sum_{k=0}^{N}\mathfrak{R}^{k}P\right\} $
converge weakly to $Q$.
\end{enumerate}
\end{thm}
\subsection{Main Goal}
This paper has two purposes. The first is to prove the equivalence
of the dynamic and static point of views under condition $\left(\mathscr{P}\right)$,
uniform ellipticity and the additional assumption of $d\geq4$. The
second purpose of this paper is to prove a certain type of local limit
theorem relating the quenched and annealed laws by a prefactor.
\subsection{Known results}
Let $d\geq2$. In \cite{Sz01} Sznitman proved an annealed CLT under
the assumption of $\left(T'\right)$. The ideas he presented may also
be used to prove an annealed local CLT. For completeness, we present
the proof of the annealed local CLT in the Appendix. In \cite{bolthausen2002satic}
Bolthausen and Sznitman proved the equivalence of the static and dynamic
point of views for certain (non nestling) ballistic random walks in
random environment, when $d\geq4$ and the disorder is low. In \cite{BZ08}
Berger and Zeitouni and in \cite{RAS05,RAS07,RAS09} Rassoul-Agha
and Sepp\"al\"ainen proved quenched invariance principle under moments
assumptions for the first regeneration time. In particular a quenched
CLT follows under condition $\left(T'\right)$.
\subsection{Main results}
Our two main results are the following:
\begin{thm}
\label{thm:Absolutely_continuous_invariant_measure} Let $d\geq4$
and assume $P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Then there exists a unique probability measure $Q$ on the space of
environments which is invariant w.r.t the point of view of the random
walk and is equivalent to the original measure $P$. In addition $E\left[\left(\frac{dQ}{dP}\right)^{k}\right]<\infty$
for every $k\in\mathbb{N}$.
\end{thm}
\begin{thm}
\label{thm:Prefactor_thm} Let $d\geq4$ and assume $P$ is uniformly
elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$. Then, there
exists a unique measurable, non-negative function $f\in L^{1}\left(\Omega,P\right)$
such that for $P$ almost every $\omega\in\Omega$
\begin{equation}
\lim_{n\to\infty}\sum_{x\in\mathbb{Z}^{d}}\left|P_{\omega}^{0}\left(X_{n}=x\right)-\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)\right|=0.\label{eq:Prefactor_thm}
\end{equation}
This unique function $f$ is the Radon-Nikodym derivative $\frac{dQ}{dP}$
of the measure $Q$ constructed in Theorem \ref{thm:Absolutely_continuous_invariant_measure}
.
\end{thm}
\subsection{Remarks about lower dimensions}
\textcolor{black}{In this paper we only prove Theorem \ref{thm:Absolutely_continuous_invariant_measure}
and Theorem \ref{thm:Prefactor_thm} in dimension $4$ or higher.
Here we wish to remark about the situation in lower dimensions. }
\textcolor{black}{For $d=1$ the existence of an equivalent measure
which is invariant with respect to the point of view of the particle
was proved by Alili, see \cite{Al99}.}
We conjecture that similar results should hold in dimension 3. In
fact the only place in the proof where we directly use the condition
$d\geq4$ is in \cite[Lemma 4.10]{berger2008slowdown}, see also Lemma
\ref{lem:Intersection_lemma} below. On the other hand, we believe
that in dimension $2$ an equivalent measure which is invariant with
respect to the point of view of the particle does not exist.
\subsection{Structure of the paper and general remarks}
In Section 2 we recall some of the notation from \cite{berger2008slowdown}
as well as some of its results. In addition we prove a slightly different
version of \cite[Lemma 4.2]{berger2008slowdown} giving annealed estimations
for a fixed time. In Section 3 we generalize \cite[Proposition 4.5]{berger2008slowdown},
which gives an upper bound on the difference between the annealed
and quenched distribution, to include estimations on the exit time
of the box. Section 4 is devoted to converting the estimation obtained
in Section 3 for $\left(d-1\right)$-dimensional cubes plus time into
a result about $d$-dimensional cubes for a fixed time. In Section
5 we bootstrap the result for large $d$-dimensional cubes obtained
in Section 4 all the way to boxes of finite size. Section 6 is devoted
to the proof of the first main result, the existence of an equivalent
probability measure on the space of environments which is invariant
with respect to the point of view of the particle. Finally, in Section
7 we prove the second main result regarding the existence of a prefactor.
Throughout this paper the value of constants $c$ and $C$ may change
from one line to the next. Numbered constants, such as $c_{1},c_{2},\ldots$
are fixed according to their first appearance in the text. Expectation
with respect to a measure $\mu$ which is not $P,P_{\omega}$or $\mathbb{P}$
is denoted by $E_{\mu}$. Finally, some of the inequalities may only
hold for large enough values of $N,n$ and $M$.
\section{Notation and other preliminary results }
We start by recalling some of the notation and results from \cite{berger2008slowdown}
to be used throughout the paper. In addition we cite an inequality
by McDiarmid for future use and state analogous result to \cite[Lemma 4.2]{berger2008slowdown}
for the annealed measure in a fixed time.
For $k,N\geq0$ define $R_{k}\left(N\right)=\left\lfloor e^{\left(\log\log N\right)^{k+1}}\right\rfloor $
and denote $R\left(N\right)=R_{1}\left(N\right)$. Note that $R_{0}\left(N\right)=\left\lfloor \log N\right\rfloor $
and that for every $k,n\geq0$, and every large enough $N$,
\[
R_{k}^{n}\left(N\right):=\left(R_{k}\left(N\right)\right)^{n}\leq R_{k+1}\left(N\right)<N.
\]
Let
\[
\vartheta=\lim_{n\to\infty}\frac{X_{n}}{\left\Vert X_{n}\right\Vert _{2}}
\]
be the direction of the speed. We assume without loss of generality
that $\left\langle \vartheta,e_{1}\right\rangle >0$ and note that
due to the results of \cite{Sz01,Sz02}, this implies that $\left(\mathscr{P}\right)$
holds both in direction $\vartheta$ and in direction $e_{1}$.
\begin{defn}
For $k\in\mathbb{N}$ define $H_{k}$ to be the hyperplane $H_{k}=\left\{ x\in\mathbb{Z}^{d}\,:\,\left\langle x,e_{1}\right\rangle =k\right\} $.
\end{defn}
\begin{defn}
By the term $N^{-\xi\left(1\right)}$ we mean a non-negative function
of $N\in\mathbb{N}$ which decays faster than any polynomial, i.e.,
if $f\left(N\right)=N^{-\xi\left(1\right)}$, then for every $k\in\mathbb{N}$
\[
\lim_{N\to\infty}N^{k}f\left(N\right)=0.
\]
Note that $N^{-\xi\left(1\right)}$ is independent of the environment
unless otherwise stated.
\end{defn}
\begin{defn}
For two non empty sets $A,B\subset\mathbb{Z}^{d}$ we define $\mbox{dist}\left(A,B\right)=\min\left\{ \left\Vert x-y\right\Vert _{1}\,:\, x\in A\,,\, y\in B\right\} $.
If $A=\left\{ x\right\} $ we write $\mbox{dist}\left(x,B\right)$
instead of $\mbox{dist}\left(\left\{ x\right\} ,B\right)$.
\end{defn}
\begin{defn}
For $x=\left(x_{1},\ldots,x_{d}\right)\in\mathbb{Z}^{d}$ and $n\in\mathbb{N}$
we denote $x\leftrightarrow n$ if $x$ and $n$ have the same parity,
i.e., $\sum_{i=1}^{d}x_{i}+n$ is even. In a similar way for $x,y\in\mathbb{Z}^{d}$
we denote $x\leftrightarrow y$ if $\sum_{i=1}^{d}x_{i}+y_{i}$ is
even.
\end{defn}
\begin{defn}
Recall that for $x\in\mathbb{Z}^{d}$ we denote by $\sigma_{x}$ the
shift in direction $x$ in $\omega$, i.e. $\sigma_{x}\omega\left(y,\cdot\right)=\omega\left(x+y,\cdot\right)$
for every $y\in\mathbb{Z}^{d}$.
\end{defn}
\begin{defn}
For $z\in\mathbb{Z}^{d}$ and $N\in\mathbb{N}$ we define (see also
Figure \ref{fig:The-basic-block})
\begin{enumerate}
\item the parallelogram of size $N$ and center $z$ to be
\[
\mathcal{P}\left(z,N\right)=\left\{ x\in\mathbb{Z}^{d}\,:\,\left|\left\langle x-z,e_{1}\right\rangle \right|<N^{2}\,,\,\left\Vert x-z-\vartheta\cdot\frac{\left\langle x-z,e_{1}\right\rangle }{\left\langle \vartheta,e_{1}\right\rangle }\right\Vert _{\infty}<NR_{5}\left(N\right)\right\} .
\]
\item The middle third of $\mathcal{P}\left(z,N\right)$
\[
\tilde{\mathcal{P}}\left(z,N\right)=\left\{ x\in\mathbb{Z}^{d}\,:\,\left|\left\langle x-z,e_{1}\right\rangle \right|<\frac{1}{3}N^{2}\,,\,\left\Vert x-z-\vartheta\cdot\frac{\left\langle x-z,e_{1}\right\rangle }{\left\langle \vartheta,e_{1}\right\rangle }\right\Vert _{\infty}<\frac{1}{3}NR_{5}\left(N\right)\right\} .
\]
\item The boundary of $\mathcal{P}\left(z,N\right)$
\[
\partial\mathcal{P}\left(z,N\right)=\left\{ x\in\mathbb{Z}^{d}\backslash\mathcal{P}\left(z,N\right)\,:\,\exists y\in\mathcal{P}\left(z,N\right)\, s.t.\,\left\Vert x-y\right\Vert _{1}=1\right\} .
\]
\item The right boundary of $\mathcal{P}\left(z,N\right)$
\[
\partial^{+}\mathcal{P}\left(z,N\right)=\left\{ x\in\partial\mathcal{P}\left(z,N\right)\,:\,\left\langle x-z,e_{1}\right\rangle =N^{2}\right\} .
\]
\end{enumerate}
\end{defn}
\begin{figure}[h]
\centering{}\includegraphics[scale=0.4]{pic/basic_block}\protect\caption{The basic block $\mathcal{P}\left(z,N\right)$: The box $\tilde{\mathcal{P}}\left(z,N\right)$
is in gray and the right boundary $\partial^{+}\mathcal{P}\left(z,N\right)$
is the striped line.\label{fig:The-basic-block}}
\end{figure}
\subsection{\label{sub:Regeneration-times}Regeneration times}
\begin{defn}
Let $\left\{ X_{n}\right\} $ be a nearest-neighbor sequence in $\mathbb{Z}^{d}$,
and let $\ell\in S^{d-1}$ be a direction. We say that $t$ is a regeneration
time for $\left\{ X_{n}\right\} $ in direction $\ell$ if the following
holds:
\begin{enumerate}
\item $\left\langle X_{s},\ell\right\rangle <\left\langle X_{t},\ell\right\rangle $
for every $s<t$.
\item $\left\langle X_{t+1},\ell\right\rangle >\left\langle X_{t},\ell\right\rangle $
\item $\left\langle X_{s},\ell\right\rangle >\left\langle X_{t+1},\ell\right\rangle $
for every $s>t+1$.
\end{enumerate}
\end{defn}
The following theorem summarize the results on the regeneration time
structure:
\begin{thm}
\label{thm:regeneration_times_thm}\cite{SZ99,Sz02} Assume that $P$
satisfies $\left(T_{\gamma}\right)$ in direction $\ell_{0}$ for
some $\gamma>0$. Then
\begin{enumerate}
\item With probability one, there exist infinitely many regeneration times,
which we denote by $\tau_{1}<\tau_{2}<\ldots$.
\item The ensemble
\[
\left\{ \left(\tau_{n+1}-\tau_{n},X_{\tau_{n+1}}-X_{\tau_{n}}\right)\right\} _{n\geq1}
\]
is an i.i.d ensemble under the annealed measure.
\item There exists $C>0$ such that for every $n\in\mathbb{N}$
\[
\mathbb{P}\left(\tau_{2}-\tau_{1}=n\right)\leq C\mathbb{P}\left(\tau_{1}=n\right),
\]
and for every $y\in\mathbb{Z}^{d}$
\[
\mathbb{P}\left(X_{\tau_{2}}-X_{\tau_{1}}=y\right)\leq C\mathbb{P}\left(X_{\tau_{1}}=y\right).
\]
\item There exists $c>0$ such that for every $n$,
\[
\mathbb{P}\left(\exists k\leq\tau_{1}\,:\,\left\Vert X_{k}\right\Vert _{\infty}>n\right)\leq e^{-cn^{\gamma}}.
\]
\end{enumerate}
\end{thm}
The following is the main technical statement from \cite{berger2008slowdown}:
\begin{thm}
\label{thm:Regeneration_times_time_estimate}\cite[Proposition 2.2]{berger2008slowdown}
For any $\gamma>0$, if $d\geq4$, and $P$ satisfies condition $\left(\mathscr{P}\right)$
in one of the $2d$-principal directions, then for every $\alpha<d$
\[
\mathbb{P}\left(\tau_{1}>k\right)\leq\exp\left(-\left(\log t\right)^{\alpha}\right).
\]
\end{thm}
\begin{cor}
\label{cor:length_of_regenerations}For $N\in\mathbb{N}$ denote by
$B_{N}=B_{N}\left(\left\{ X_{n}\right\} \right)$ the event
\[
B_{N}\left(\left\{ X_{n}\right\} \right)=\left\{ \forall1\leq k\leq N^{2}\,:\,\tau_{k}-\tau_{k-1}\leq R\left(N\right)\right\} ,
\]
where $\tau_{0}=0$. Then $\mathbb{P}\left(B_{N}\right)\geq1-N^{-\xi\left(1\right)}$.\end{cor}
\begin{rem}
\label{Rem:The_event_A_N}Note that the event $B_{N}\left(\left\{ X_{n}\right\} \right)$
implies the event that the distance traveled between two regeneration
times is bounded by $R\left(N\right)$ as well, i.e.,
\[
A_{N}\left(\left\{ X_{n}\right\} \right)=\left\{ \forall1\leq k\leq N^{2}\,:\,\max\left\{ \left\Vert X_{t}-X_{\tau_{n-1}}\right\Vert _{\infty}\,:\,\tau_{n-1}\leq t\leq\tau_{n}\right\} \leq R\left(N\right)\right\}
\]
satisfies $B_{N}\left(\left\{ X_{n}\right\} \right)\subset A_{N}\left(\left\{ X_{n}\right\} \right)$
and in particular $\mathbb{P}\left(A_{N}\right)\geq1-N^{-\xi\left(1\right)}$.
\end{rem}
\subsection{Intersections of paths of random walks}
The following lemma estimates the number of intersections of two independent
random walks in dimension $d\geq4$. This is in fact the the only
place in the proof where the assumption $d\geq4$ is used explicitly.
Denote by $P_{\omega,\omega}^{z,z}$, $E_{\omega,\omega}^{z,z}$ the
law (respectively expectation) of two independent random walks with
quenched law starting from $z$.
\begin{lem}
\label{lem:Intersection_lemma}\cite[Lemma 4.10]{berger2008slowdown}
Let $d\geq4$ and assume $P$ is uniformly elliptic, i.i.d and satisfies
$\left(\mathscr{P}\right)$. Let $X^{\left(1\right)}=\left\{ X_{n}^{\left(1\right)}\right\} $
and $X^{\left(2\right)}=\left\{ X_{n}^{\left(2\right)}\right\} $
be two independent random walks running in the same environment $\omega$.
For $i\in\left\{ 1,2\right\} $ let $\left[X^{\left(i\right)}\right]$
be the set of points visited by $X^{\left(i\right)}$. Then
\[
P\left(\left\{ \omega\in\Omega\,:\, E_{\omega,\omega}^{0,0}\left[\left|\left[X^{\left(1\right)}\right]\cap\left[X^{\left(2\right)}\right]\cap\mathcal{P}\left(0,N\right)\right|{\mathbbm{1}}_{A_{N}\left(\left\{ X_{n}^{\left(1\right)}\right\} \right)\cap A_{N}\left(\left\{ X_{n}^{\left(2\right)}\right\} \right)}\right]\geq R_{2}\left(N\right)\right\} \right)=N^{-\xi\left(1\right)}.
\]
\end{lem}
For future use we denote
\begin{equation}
J\left(N\right)=\left\{ \omega\in\Omega\,:\, E_{\omega,\omega}^{z,z}\left[\left|\left[X^{\left(1\right)}\right]\cap\left[X^{\left(2\right)}\right]\cap\mathcal{P}\left(0,N\right)\right|{\mathbbm{1}}_{A_{N}\left(\left\{ X_{n}^{\left(1\right)}\right\} \right)\cap A_{N}\left(\left\{ X_{n}^{\left(2\right)}\right\} \right)}\right]\leq R_{2}\left(N\right)\,,\,\forall z\in\tilde{\mathcal{P}}\left(0,N\right)\right\} .\label{eq:intersection_event}
\end{equation}
Therefore, due to the last lemma, we have $P\left(J\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
\subsection{McDiarmid's inequality}
The following Azuma type inequality, proved by McDiarmid, see \cite{McD98},
is used in Section \ref{sec:Adding-time-estimation}.
\begin{thm}
\label{thm:Azuma_type_inequality}\cite[Theorem 3.14]{McD98} Let
$\left\{ M_{k}\right\} _{k=0}^{n}$ be a martingale with respect to
a probability measure $\mathtt{P}$, given by $M_{k}=E_{\mathtt{P}}\left[X|\mathscr{F}_{k}\right]$,
with $M_{0}=E_{\mathtt{P}}\left[X\right]$. For $1\leq k\leq n$ let
$U_{k}=\esup\left(\left|M_{k}-M_{k-1}\right|\Bigg|\mathscr{F}_{k-1}\right)$
and define $U=\sum_{k=1}^{n}U_{k}^{2}$. Then
\[
\mathtt{P}\left(\left|M_{n}\right|>\alpha\,,\, U\leq c\right)\leq2e^{-\frac{\alpha^{2}}{2c}}.
\]
\end{thm}
\subsection{Annealed estimation for a fixed time}
In this subsection we state some standard estimations on the annealed
measure of the random walk. The proof is a standard and straightforward
use of Fourier transform techniques and the regeneration structure
described in Subsection \ref{sub:Regeneration-times}. The first three
claims are proved in a very similar way to the proof of \cite[Lemma 4.2]{berger2008slowdown}
(see also Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time}
for another version). The formal statement is the following.
\begin{lem}
\label{lem:general_annealed_estimations} Assume that $P$ is uniformly
elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$. Then for
large enough $n\in\mathbb{N}$ and $x,y,z,w\in\mathbb{Z}^{d}$ such
that $\left\Vert x-y\right\Vert _{1}=1$, $\left\Vert z-w\right\Vert _{1}=1$
\begin{equation}
\mathbb{P}^{z}\left(X_{n}=x\right)\leq Cn^{-\frac{d}{2}},\label{eq:general_annealed_estimation_1}
\end{equation}
\begin{equation}
\left|\mathbb{P}^{z}\left(X_{n}=x\right)-\mathbb{P}^{z}\left(X_{n+1}=y\right)\right|\leq Cn^{-\frac{d+1}{2}},\label{eq:general_annealed_estimation_2}
\end{equation}
\begin{equation}
\left|\mathbb{P}^{z}\left(X_{n}=x\right)-\mathbb{P}^{w}\left(X_{n+1}=x\right)\right|\leq Cn^{-\frac{d+1}{2}}.\label{eq:general_annealed_estimation_3}
\end{equation}
In addition, for every $\varepsilon>0$ and every partition $\Pi_{n}^{\left(\varepsilon\right)}$
of $\mathbb{Z}^{d}$ into boxes of side length $n^{\varepsilon}$.
\begin{equation}
\sum_{\Delta\in\Pi_{n}^{\left(\varepsilon\right)}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\left[\max_{y\in\Delta}\mathbb{P}^{0}\left(X_{n}=y\right)-\mathbb{P}^{0}\left(X_{n}=x\right)\right]\leq Cn^{-\frac{1}{2}+3d\varepsilon}.\label{eq:general_annealed_estimation_4}
\end{equation}
\end{lem}
The proof of Lemma \ref{lem:general_annealed_estimations} can be
found in Appendix \ref{appendix_sub:Annealed-derivative-estimations}.
Before turning to the last estimation of this subsection we state
here a very simple claim to be used in several places.
\begin{claim}
\label{Clm:simple_claim_annealed_to_quenched}Let $A$ be an event
in the $\sigma$-algebra of $\left(\mathbb{Z}^{d}\right)^{\mathbb{N}}$and
assume that $\mathbb{P}\left(A\right)\leq\varepsilon$, then $P\left(\left\{ \omega\in\Omega\,:\, P_{\omega}\left(A\right)\geq\sqrt{\varepsilon}\right\} \right)\leq\sqrt{\varepsilon}$.
In particular if a sequence of events $\left\{ A_{N}\right\} $ satisfies
$\mathbb{P}\left(A_{N}\right)=1-N^{-\xi\left(1\right)}$, then $P\left(\left\{ \omega\in\Omega\,:\, P_{\omega}\left(A_{N}\right)=1-N^{-\xi\left(1\right)}\right\} \right)=1-N^{-\xi\left(1\right)}$. \end{claim}
\begin{proof}
Define the random variable $X:\Omega\to\left[0,1\right]$ by $X\left(\omega\right)=P_{\omega}\left(A\right)$.
By the Markov inequality $P\left(X\left(\omega\right)\geq\sqrt{\varepsilon}\right)\leq\frac{E\left[X\left(\omega\right)\right]}{\sqrt{\varepsilon}}=\frac{\mathbb{P}\left(A\right)}{\sqrt{\varepsilon}}\leq\frac{\varepsilon}{\sqrt{\varepsilon}}=\sqrt{\varepsilon}$.
\end{proof}
Next, we show that the location of the walk at time $n$ is concentrated
in a box which is a bit larger than $\sqrt{n}$. More formally we
have
\begin{lem}
\label{lem:good_estimation_for_the_location} Assume that $P$ is
uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Then \end{lem}
\begin{enumerate}
\item $\mathbb{P}^{z}\left(\left\Vert X_{n}-\mathbb{E}^{z}\left[X_{n}\right]\right\Vert _{\infty}>\sqrt{n}R_{5}\left(n\right)\right)\leq e^{-R_{5}\left(n\right)}=n^{-\xi\left(1\right)},$
\item $P\left(\left\{ \omega\in\Omega\,:\, P_{\omega}^{z}\left(\left\Vert X_{n}-\mathbb{E}^{z}\left[X_{n}\right]\right\Vert _{\infty}>\sqrt{n}R_{5}\left(n\right)\right)=n^{-\xi\left(1\right)}\right\} \right)=1-n^{-\xi\left(1\right)},$
\item For every $\delta>0$ there exists $C>0$ such that $\mathbb{P}^{z}\left(\left\Vert X_{n}-\mathbb{E}^{z}\left[X_{n}\right]\right\Vert _{\infty}>C\sqrt{n}\right)<\delta$.
\end{enumerate}
The proof of Lemma \ref{lem:good_estimation_for_the_location} can
be found in Appendix \ref{appendix_sub:Annealed-derivative-estimations}.
\section{\label{sec:Adding-time-estimation}Adding time estimation }
The goal of this section is to prove a generalized version of \cite[Proposition 4.5]{berger2008slowdown}.
The original lemma gives a bound on the difference between the probability
measures $\mathbb{P}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\cdot\right)$
and $P_{\omega}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\cdot\right)$
to hit any cube in a partition of $\partial^{+}{\mathcal{P}}\left(0,N\right)$
into cubes of side length $N^{\theta}$, for any $0<\theta\leq1$.
This estimation immediately implies that the total variation of the
two measures goes to zero as $N$ goes to infinity. Here, we show
that if an estimation on the hitting time $T_{\partial{\mathcal{P}}\left(0,N\right)}$
is added, then a similar estimation can be derived. More formally,
we have the following:
\begin{prop}
\label{prop:Time_(d-1)_box_difference}Let $d\geq4$ and assume $P$
is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
For every $0<\theta\leq1$, let $F\left(N\right)=F\left(N,\theta\right)$
be the event that for every $z\in\tilde{{\mathcal{P}}}\left(0,N\right)$, every
cube $\Delta$ of side length $N^{\theta}$ which is contained in
$\partial^{+}{\mathcal{P}}\left(0,N\right)$ and every interval $I$ of length
$N^{\theta}$
\[
\left|P_{{\omega}}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\Delta\,,\, T_{\partial{\mathcal{P}}\left(0,N\right)}\in I\right)-{\mathbb{P}}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\Delta\,,\, T_{\partial{\mathcal{P}}\left(0,N\right)}\in I\right)\right|\leq CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}
\]
Then $P\left(F\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
\end{prop}
The proof of Proposition \ref{prop:Time_(d-1)_box_difference} follows
the one of \cite[Proposition 4.5]{berger2008slowdown} (see also \cite[Section 4]{berger2008slowdown}
and in particular Lemma 4.15). Here are the main steps of the proof:
The proof starts with another version for annealed derivatives bounds
(see Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time}).
Next in Lemma \ref{lem:Distance_of_time_from_expectation} we prove
an annealed concentration inequality for the hitting time $T_{\partial\mathcal{P}\left(0,N\right)}$.
Lemma \ref{lem:First_difference_estimation} provides a first weak
estimation for the difference between the quenched and annealed hitting
probabilities for large enough boxes, i.e., $\theta>\frac{d}{d+1}$.
Using induction and the estimation from the last lemma we prove an
upper bound on the probability to hit a given box of side length $N^{\theta}$
in a time interval of length $N^{\theta}$ for every $0<\theta\leq1$
(see Lemma \ref{lem:Upper_estimation_on_the_quenched_exit}). In Lemma
\ref{lem:Estimation_for_far_hyperplane} we use the upper quenched
estimations in order to show that the difference between the quenched
and annealed hitting probabilities, in a slightly further hyperplane
are as required. Finally, in the proof of Proposition \ref{prop:Time_(d-1)_box_difference},
we show how to translate the estimations from the further hyperplane
back to the original hyperplane. The main tools used in the proof
are an environment exposure procedure, originally due to Bolthausen
and Sznitman \cite{bolthausen2002satic}, which defines a martingale
and allow the use of Azuma's and McDiarmid's inequalities, the intersection
estimate for two independent random walks from Lemma \ref{lem:Intersection_lemma}
and some general annealed estimation on the probability to hit boxes.
\begin{rem}
In Section \ref{sec:Lorentz_trans} we use Proposition \ref{prop:Time_(d-1)_box_difference}
for boxes whose side length is only asymptotic to $N^{\theta}$ (for
some $0<\theta<1$), i.e., the side length is $N^{\theta}+o\left(N^{\theta}\right)$.
One can verify that the same proof holds for such boxes as well.
\end{rem}
We start by stating another version for the estimation on the annealed
measure (see Lemma \ref{lem:general_annealed_estimations} and \cite[Lemma 4.2]{berger2008slowdown}).
\begin{lem}[Annealed derivative estimations]
\label{lem:Annealed_derivative_estimations_(d-1)+time} Assume $P$
is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Fix $z_{1}\in{\mathbb{Z}}^{d}$, $N\in\mathbb{N}$ and let $z\in\tilde{{\mathcal{P}}}\left(z_{1},N\right)$.
Let $\left\{ X_{n}\right\} $ be an RWRE starting at $z$. Then for
large enough $N$
\begin{enumerate}
\item For every $m\in\mathbb{N}$ and every $x\in\partial^{+}{\mathcal{P}}\left(z_{1},N\right)$
\begin{equation}
\mathbb{P}^{z}\left({T_{\partial\mathcal{P}(0,N)}}=m\,,\, X_{{T_{\partial\mathcal{P}(0,N)}}}=x\right)<CN^{-d}.\label{eq:annealed_derivative_estimation_1}
\end{equation}
\item For every $m\in\mathbb{N}$ and every $x,y\in\partial^{+}{\mathcal{P}}\left(z_{1},N\right)$
such that $\left\Vert x-y\right\Vert _{1}=1$
\begin{equation}
\left|\mathbb{P}^{z}\left({T_{\partial\mathcal{P}(0,N)}}=m\,,\, X_{{T_{\partial\mathcal{P}(0,N)}}}=x\right)-\mathbb{P}^{z}\left({T_{\partial\mathcal{P}(0,N)}}=m+1\,,\, X_{{T_{\partial\mathcal{P}(0,N)}}}=y\right)\right|<CN^{-d-1}.\label{eq:annealed_derivative_estimation_2}
\end{equation}
\item For every $m\in\mathbb{N}$, every $x\in\partial^{+}{\mathcal{P}}\left(z_{1},N\right)$
and every $1\leq j\leq d$
\begin{equation}
\left|\mathbb{P}^{z}\left({T_{\partial\mathcal{P}(0,N)}}=m\,,\, X_{{T_{\partial\mathcal{P}(0,N)}}}=x\right)-\mathbb{P}^{z+e_{j}}\left({T_{\partial\mathcal{P}(0,N)}}=m+1\,,\, X_{{T_{\partial\mathcal{P}(0,N)}}}=x\right)\right|<CN^{-d-1}.\label{eq:annealed_derivative_estimation_3}
\end{equation}
\end{enumerate}
\end{lem}
The proof of Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time}
can be found in Appendix \ref{appendix_sub:Annealed-derivative-estimations}.
Next, we prove an annealed concentration inequality for the hitting
time $T_{\partial{\mathcal{P}}}$.
\begin{lem}
\label{lem:Distance_of_time_from_expectation} Let $d\geq4$ and assume
$P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Then
\begin{equation}
\mathbb{P}^{z}\left(T_{\partial{\mathcal{P}}\left(0,N\right)}\neq T_{\partial^{+}{\mathcal{P}}\left(0,N\right)}\right)=N^{-\xi\left(1\right)}\label{eq:annealed_time_estimations_1}
\end{equation}
and for every $z\in\tilde{{\mathcal{P}}}\left(0,N\right)$
\begin{equation}
\mathbb{P}^{z}\left(\left|T_{\partial{\mathcal{P}}\left(0,N\right)}-\mathbb{E}^{z}\left[T_{\partial{\mathcal{P}}\left(0,N\right)}\right]\right|>NR_{2}\left(N\right)\right)=N^{-\xi\left(1\right)}.\label{eq:annealed_time_estimations_2}
\end{equation}
\end{lem}
\begin{proof}
The fact that (\ref{eq:annealed_time_estimations_1}) holds was proved
in \cite[Lemma 4.2(1)]{berger2008slowdown}. For (\ref{eq:annealed_time_estimations_2})
we first show that $\left|\mathbb{E}^{z}\left[\tau_{k}\right]-\mathbb{E}^{z}\left[\tau_{k}|B_{N}\right]\right|=N^{-\xi\left(1\right)}$
for every $1\leq k\leq N^{2}$, where $B_{N}$ is as defined in Corollary
\ref{cor:length_of_regenerations}. Indeed, using the notation $\tau_{0}=0$,
for every $1\leq k\leq N^{2}$
\begin{align*}
\mathbb{E}^{z}\left[\left|\tau_{k}-\tau_{k-1}\right|\cdot{\mathbbm{1}}_{B_{N}^{c}}\right] & \leq\mathbb{E}^{z}\left[\left|\tau_{k}-\tau_{k-1}\right|\cdot{\mathbbm{1}}_{\exists j\neq k\,\left|\tau_{j}-\tau_{j-1}\right|\geq R\left(N\right)}\right]+\mathbb{E}^{z}\left[\left|\tau_{k}-\tau_{k-1}\right|\cdot{\mathbbm{1}}_{\left|\tau_{k}-\tau_{k-1}\right|\geq R\left(N\right)}\right]\\
& \leq\mathbb{E}^{z}\left[\left|\tau_{k}-\tau_{k-1}\right|\right]\mathbb{P}^{z}\left(B_{N}^{c}\right)+\sum_{t>R\left(N\right)}t\cdot\mathbb{P}^{z}\left(\left|\tau_{k}-\tau_{k-1}\right|=t\right)\\
& \leq\mathbb{E}^{z}\left[\left|\tau_{k}-\tau_{k-1}\right|\right]\mathbb{P}^{z}\left(B_{N}^{c}\right)+\sum_{t>R\left(N\right)}t\cdot\exp\left(-\left(\log t\right)^{\alpha}\right)=N^{-\xi\left(1\right)},
\end{align*}
where for the last inequality we used Theorem \ref{thm:regeneration_times_thm}
and for the last equality we used Corollary (\ref{cor:length_of_regenerations}).
Therefore for every $1\leq k\leq N^{2}$
\begin{align*}
\left|\mathbb{E}^{z}\left[\tau_{k}-\tau_{k-1}\big|B_{N}\right]-\mathbb{E}^{z}\left[\tau_{k}-\tau_{k-1}\right]\right| & \leq\left|\mathbb{E}^{z}\left[\tau_{k}-\tau_{k-1}\big|B_{N}\right]-\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right){\mathbbm{1}}_{B_{N}}\right]\right|\\
& +\left|\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right){\mathbbm{1}}_{B_{N}}\right]-\mathbb{E}^{z}\left[\tau_{k}-\tau_{k-1}\right]\right|\\
& =\left|\left(1-\mathbb{P}^{z}\left(B_{N}\right)\right)\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right)\big|B_{N}\right]\right|+\left|\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right){\mathbbm{1}}_{B_{N}^{c}}\right]\right|\\
& =\mathbb{P}^{z}\left(B_{N}^{c}\right)\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right)\big|B_{N}\right]+\mathbb{E}^{z}\left[\left(\tau_{k}-\tau_{k-1}\right){\mathbbm{1}}_{B_{N}^{c}}\right]\\
& \leq R\left(N\right)\mathbb{P}^{z}\left(B_{N}^{c}\right)+N^{-\xi\left(1\right)}=N^{-\xi\left(1\right)}.
\end{align*}
Summing the differences $\left\{ \mathbb{E}^{z}\left[\tau_{j}-\tau_{j-1}\big|B_{N}\right]-\mathbb{E}^{z}\left[\tau_{j}-\tau_{j-1}\right]\right\} _{j=1}^{k}$
gives for every $1\leq k\leq N^{2}$
\[
\left|\mathbb{E}^{z}\left[\tau_{k}\big|B_{N}\right]-\mathbb{E}^{z}\left[\tau_{k}\right]\right|=N^{-\xi\left(1\right)}
\]
Since we know that $\mathbb{P}^{z}\left(B_{N}\right)=1-N^{-\xi\left(1\right)}$
(see Corollary \ref{cor:length_of_regenerations}) it is enough to
show that
\[
\mathbb{P}^{z}\left(\left|T_{\partial{\mathcal{P}}\left(0,N\right)}-\mathbb{E}^{z}\left[T_{\partial{\mathcal{P}}\left(0,N\right)}\right]\right|>NR_{2}\left(N\right)\Bigg|B_{N}\right)=N^{-\xi\left(1\right)}.
\]
Under the event $B_{N}$ there exist some $1\leq k\leq N^{2}$ such
that $\tau_{k}\leq T_{\partial P}\leq\tau_{k}+R\left(N\right)$ and
thus (using the first estimation)
\begin{align*}
\mathbb{P}^{z}\left(\left|T_{\partial{\mathcal{P}}\left(0,N\right)}-\mathbb{E}^{z}\left[T_{\partial{\mathcal{P}}\left(0,N\right)}\right]\right|>NR_{2}\left(N\right)\Bigg|B_{N}\right) & \leq\sum_{k=1}^{N^{2}}\mathbb{P}^{z}\left(\left|\tau_{k}-\mathbb{E}^{z}\left[\tau_{k}\right]\right|>\frac{1}{2}NR_{2}\left(N\right)\Bigg|B_{N}\right)\\
& \leq\sum_{k=1}^{N^{2}}\mathbb{P}^{z}\left(\left|\tau_{k}-\mathbb{E}^{z}\left[\tau_{k}\Big|B\left(N\right)\right]\right|>\frac{1}{4}NR_{2}\left(N\right)\Bigg|B_{N}\right)+N^{-\xi\left(1\right)}
\end{align*}
Note that conditioned on $B_{N}$ the first $N$ regenerations are
still i.i.d, so by Azuma's inequality this can be bounded by
\[
\sum_{k=1}^{N^{2}}2\exp\left(-\frac{N^{2}R_{2}^{2}\left(N\right)}{32kR^{2}\left(N\right)}\right)+N^{-\xi\left(1\right)}\leq e^{-R_{2}\left(N\right)}+N^{-\xi\left(1\right)}=N^{-\xi\left(1\right)}
\]
\end{proof}
\begin{lem}
\label{lem:First_difference_estimation} Let $d\geq4$ and assume
$P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Fix $0<\theta\leq1$. Let $L\left(N\right)=L\left(\theta,N\right)$
be the event that for every $\frac{2}{5}N^{2}\leq M\leq N^{2}$, every
$z\in\tilde{\mathcal{P}}\left(0,N\right)$, every $\left(d-1\right)$-dimensional
cube $\Delta$ of size $N^{\theta}$ which is contained in $H_{M}$
and every $1$-dimensional interval $I$ of length $N^{\theta}$
\[
\left|P_{{\omega}}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\,,\, B_{N}\right)-\mathbb{P}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\,,\, B_{N}\right)\right|\leq N^{d\left(\theta-1\right)}.
\]
Then for $\theta>\frac{d}{d+1}$, $P\left(L\left(\theta,N\right)\right)=1-N^{-\xi\left(1\right)}.$\end{lem}
\begin{proof}
Fix $\theta$, and let $\frac{d}{d+1}<\theta'<\theta$. Let $V=\left[N^{2\theta'}\right]$.
Fix $\frac{2}{5}N^{2}\leq M\leq N^{2}$, $v\in H_{M+V}$ and $m\in\mathbb{N}$.
Finally denote by ${\mathcal{G}}$ the ${\sigma}$-algebra determined by the configuration
on
\[
{\mathcal{P}}^{M}\left(0,N\right):={\mathcal{P}}\left(0,N\right)\cap\left\{ x\,:\,\left\langle x,e_{1}\right\rangle \leq M\right\} .
\]
We are interested in the quantity (see also Figure \ref{fig:2})
\[
J^{\left(M\right)}\left(v\right)=E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v,T_{M+V}=m\,,\, B_{N}\right)\bigg|{\mathcal{G}}\right].
\]
\begin{figure}[h]
\begin{centering}
\includegraphics[scale=0.4]{pic/proof_of_lemma_3. 5}\protect\caption{The quantity $J^{\left(M\right)}\left(v\right)$ is the probability
of hitting the point $v$, conditioned on the environments in the
gray area, and averaged (annealed) elsewhere. The small parallelogram
indicates the middle third $\tilde{\mathcal{P}}\left(0,N\right)$
in which the random walk starts.\label{fig:2}}
\par\end{centering}
\end{figure}
In order to estimate $J^{\left(M\right)}\left(v\right)$ we order
the vertices of $\mathcal{P}^{M}\left(0,N\right)$ lexicographically,
$x_{1},x_{2},\ldots$ with the first coordinate being the most significant
and let $\left\{ {\mathcal{F}}_{k}\right\} $ be the $\sigma$-algebra determined
by ${\omega}\left(x_{1},\cdot\right),\ldots,\omega\left(x_{k},\cdot\right)$,
so in particular for every $-N^{2}+1\leq l\leq M$ the vertices in
$H_{l}\cap\mathcal{P}^{M}\left(0,N\right)$ are exposed after those
in $H_{l-1}\cap\mathcal{P}^{M}\left(0,N\right)$.
Consider the martingale $M_{k}=E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\Big|B_{N}\right)\Big|\mathcal{F}_{k}\right]$.
In order to use McDiarmid's inequality we first bound $U_{k}:=\esup\left(\left|M_{k}-M_{k-1}\right|\Big|{\mathcal{F}}_{k-1}\right).$
We claim that for set of environments with $P$ probability $\geq1-N^{-\xi\left(1\right)}$
\[
U_{k}\leq CR\left(N\right)E\left[P_{{\omega}}^{z}\left(x_{k}\,\mbox{is visited}\Big|B_{N}\right)\Big|{\mathcal{F}}_{k-1}\right]V^{-\frac{d+1}{2}}.
\]
Indeed, let ${\omega}'$ be an environment that agrees with ${\omega}$ everywhere
except possibly in $x_{k}$. Let $\mathtt{P}$ be the distribution
of a walk that follows the law of ${\omega}$ on $\left\{ x_{i}\,:\, i\leq k\right\} $
and the annealed distribution on $\mathbb{Z}^{d}\backslash\left\{ x_{i}\,:\, i\leq k\right\} $
conditioned the event $B_{N}$. Equivalently, let $\mathtt{P}'$ be
the distribution of a walk that follows the law of ${\omega}'$ on $\left\{ x_{i}\,:\, i\leq k\right\} $
and the annealed on the complement conditioned the event $B_{N}$.
More precisely for an event $A\subset\left({\mathbb{Z}}^{d}\right)^{{\mathbb{N}}}$
we have $\mathtt{P}\left(A\right)=E\left[P_{\omega}^{z}\left(A\Big|B_{N}\right)\Big|{\mathcal{F}}_{k}\right]$
and equivalently for $\mathtt{P}'$. Then
\begin{align*}
U_{k} & \leq\sup_{{\omega}'}\left|\mathtt{P}'\left(X_{T_{M+V}}=v,T_{M+V}=m\right)-\mathtt{P}\left(X_{T_{M+V}}=v,T_{M+V}=m\right)\right|\\
& \leq\sup_{{\omega}'}\left|\mathtt{P}'\left(X_{T_{M+V}}=v,T_{M+V}=m\,,\,\left\{ x_{k}\,\mbox{is not visited}\right\} \right)-\mathtt{P}\left(X_{T_{M+V}}=v,T_{M+V}=m\,,\,\left\{ x_{k}\,\mbox{is not visited}\right\} \right)\right|\\
& +\sup_{{\omega}'}\left|\mathtt{P}'\left(X_{T_{M+V}}=v,T_{M+V}=m\,,\,\left\{ x_{k}\,\mbox{is visited}\right\} \right)-\mathtt{P}\left(X_{T_{M+V}}=v,T_{M+V}=m\,,\,\left\{ x_{k}\,\mbox{is visited}\right\} \right)\right|
\end{align*}
where the supremum is taken over all environments ${\omega}'$ that agree
with ${\omega}$ on ${\mathbb{Z}}^{d}\backslash\left\{ x_{k}\right\} $. Note that
on the event $\left\{ x_{k}\,\mbox{is not visited}\right\} $, the
distributions $\mathtt{P}$ and $\mathtt{P}'$ are the same and therefore
the difference of the probabilities equals zero. On the other hand
on the event $\left\{ x_{k}\,\mbox{is visited}\right\} $, we can
couple both walks together until the first hitting time of $x_{k}$
(which in particular implies that the hitting time of $x_{k}$ is
the same). Since we conditioned on the event $B_{N}$, the next regeneration
time after hitting $x_{k}$ is at most $R\left(N\right)$ steps later.
Therefore, from Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time}
it follows that
\[
\left|\mathtt{P}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\Big|x_{k}\,\mbox{is visited}\right)-\mathbb{P}^{x_{k}}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\right)\right|<CR\left(N\right)V^{-\frac{d+1}{2}}
\]
and
\[
\left|\mathtt{P}'^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\Big|x_{k}\,\mbox{is visited}\right)-\mathbb{P}^{x_{k}}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\right)\right|<CR\left(N\right)V^{-\frac{d+1}{2}}.
\]
Consequently we get
\[
U_{k}\leq CR\left(N\right)V^{-\frac{d+1}{2}}\mathtt{P}^{z}\left(x_{k}\,\mbox{is visited}\right)
\]
as required.
Next we show that $U:=\sum_{k}U_{k}^{2}$ is bounded by $CR_{2}^{2}\left(N\right)V^{-d-1}$
provided $\omega\in J\left(N\right)$, Indeed, noting that if $x$
is visited and $B_{N}$ holds, then the first visit to the layer $H_{\left\langle x,e_{1}\right\rangle -1}$
is in the box $B\left(x\right)=\left\{ y\,:\, y\in H_{\left\langle x,e_{1}\right\rangle -1},\,\left\Vert y-x\right\Vert _{\infty}\leq R\left(N\right)\right\} $
it follows that
\begin{align*}
U_{k} & \leq CR\left(N\right)V^{-\frac{d+1}{2}}\mathbf{P}^{z}\left(x_{k}\in\left[X\right]|{\mathcal{F}}_{k-1},B_{N}\right)\\
& \leq CR\left(N\right)V^{-\frac{d+1}{2}}\sum_{y\in B\left(x_{k}\right)}\mathbf{P}^{z}\left(T_{\left\langle x,e_{1}\right\rangle -1}=y\bigg|\mathcal{F}_{k-1}\right)\\
& =CR\left(N\right)V^{-\frac{d+1}{2}}\sum_{y\in B\left(x_{k}\right)}P_{\omega}^{z}\left(T_{\left\langle x,e_{1}\right\rangle -1}=y\right)\\
& \leq CR\left(N\right)V^{-\frac{d+1}{2}}\sum_{y\in B\left(x_{k}\right)}P_{\omega}^{z}\left(y\in\left[X\right]\right).
\end{align*}
Since $\left|B\left(x_{k}\right)\right|\leq C\cdot2^{d}\cdot R^{d}\left(N\right)$
and every $y$ is in $B\left(x\right)$ for at most $2^{d}R^{d}\left(N\right)$
points $x$, we get
\begin{align*}
U:=\sum_{k=1}^{n}U_{k}^{2} & \leq C\sum_{k=1}^{n}R^{2}\left(N\right)V^{-d-1}\cdot\left[\sum_{y\in B\left(x_{k}\right)}P_{{\omega}}^{z}\left(y\in\left[X\right]\right)\right]^{2}\\
& \leq R^{2}\left(N\right)\cdot2^{d}R^{d}\left(N\right)V^{-d-1}\sum_{k=1}^{n}\sum_{y\in B\left(x_{k}\right)}P_{{\omega}}^{z}\left(y\in\left[X\right]\right)^{2}\\
& \leq C\cdot2^{2d}\cdot R^{2d+2}\left(N\right)V^{-d-1}\sum_{y\in\mathcal{P}\left(0,N\right)}P_{{\omega}}^{z}\left(y\in\left[X\right]\,,\, B_{N}\left(\left\{ X_{n}\right\} \right)\right)^{2}\\
& \overset{\left(1\right)}{\leq}C\cdot2^{2d}\cdot R^{2d+2}\left(N\right)\cdot R_{2}\left(N\right)V^{-d-1}\leq C\cdot R_{2}^{2}\left(N\right)V^{-d-1},
\end{align*}
where for $\left(1\right)$ we used the assumption $\omega\in J\left(N\right)$.
Thus by McDiarmid's inequality (see Theorem \ref{thm:Azuma_type_inequality})
for every $\delta>0$
\begin{align*}
& P\left(\Biggl|E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, B_{N}\Big|{\mathcal{G}}\right)\right]-\mathbb{P}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, B_{N}\right)\Biggr|>\delta\right)\\
\leq & P\left(J\left(N\right)^{c}\right)+2\exp\left(-\frac{\delta^{2}}{2CR_{2}^{2}\left(N\right)V^{-d-1}}\right).
\end{align*}
In particular, for $\delta=\frac{1}{4}N^{-d}=\frac{1}{4}V^{-\frac{d+1}{2}}V^{\eta}$,
with $\eta=\frac{\left(d+1\right)\theta'-d}{2\theta'}>0$ we get
\begin{align*}
& P\left(\left|E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, B_{N}\Big|{\mathcal{G}}\right)\right]-\mathbb{P}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, B_{N}\right)\right|>\frac{1}{4}N^{-d}\right)\\
\leq & N^{-\xi\left(1\right)}+2\exp\left(-\frac{\left[N^{2\theta'}\right]^{2\eta}}{32CR_{2}^{2}\left(N\right)}\right)=N^{-\xi\left(1\right)}.
\end{align*}
Using Corollary \ref{cor:length_of_regenerations} this also gives
\[
P\left(\left|E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\Big|{\mathcal{G}}\right)\right]-\mathbb{P}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\right)\right|>\frac{1}{2}N^{-d}\right)=N^{-\xi\left(1\right)}.
\]
Let $W_{1}\left(N\right)\subset{\Omega}$ be the event
\[
\left|E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, A_{N}\,,\, B_{N}\Big|{\mathcal{G}}\right)\right]-\mathbb{P}^{z}\left(X_{T_{M+V}}=v\,,\, T_{M+V}=m\,,\, A_{N}\,,\, B_{N}\right)\right|\leq\frac{1}{2}N^{-d}
\]
for every $\frac{2}{5}N^{2}\leq M\leq N^{2}$, every $v\in H_{M+V}\cap{\mathcal{P}}\left(0,2N\right)$,
every $z\in\tilde{\mathcal{P}}\left(0,N\right)$ and every $m\in\left[0,N^{3}\right]$.
Then by the above argument $P\left(W_{1}\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
Consider now ${\omega}\in W_{1}\left(N\right)$, $\frac{2}{5}N^{2}\leq M\leq N^{2}$
a cube $\Delta$ of side length $N^{\theta}$ which is contained in
$H_{M}$ and an interval $I$ of length $N^{\theta}$. We wish to
estimate
\[
\left|P_{{\omega}}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\,,\, B_{N}\right)-\mathbb{P}^{z_{1}}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\,,\, B_{N}\right)\right|.
\]
Let $c\left(\Delta\right)$ and $c\left(I\right)$ be the centers
of the cube $\Delta$ and the interval $I$ respectively. Let $c'\left(\Delta\right)=c\left(\Delta\right)+V\frac{\vartheta}{\left\langle \vartheta,e_{1}\right\rangle }$,
$c'\left(I\right)=c\left(I\right)+V\frac{1}{\left\langle \mathbbm{v},e_{1}\right\rangle }$
and define (see also Figure \ref{fig:Bix_small_box})
\[
\Delta^{\left(1\right)}=\left\{ v\in H_{M+V}\,:\,\left\Vert v-c'\left(\Delta\right)\right\Vert _{\infty}<\frac{1}{2}\cdot\frac{9}{10}\cdot N^{\theta}\right\} ,
\]
\[
\Delta^{\left(2\right)}=\left\{ v\in H_{M+V}\,:\,\left\Vert v-c'\left(\Delta\right)\right\Vert _{\infty}<\frac{1}{2}\cdot\frac{11}{10}\cdot N^{\theta}\right\} ,
\]
\[
I^{\left(1\right)}=\left\{ t\in\mathbb{N}\,:\,\left|t-c'\left(I\right)\right|<\frac{1}{2}\cdot\frac{9}{10}\cdot N^{\theta}\right\} ,
\]
and
\[
I^{\left(2\right)}=\left\{ t\in\mathbb{N}\,:\,\left|t-c'\left(I\right)\right|<\frac{1}{2}\cdot\frac{11}{10}\cdot N^{\theta}\right\} .
\]
\begin{figure}[h]
\centering{}\includegraphics[scale=0.4]{pic/big_small_blocks}\protect\caption{\label{fig:Bix_small_box}Using the annealed walk in the area between
the hyperplane $H_{M}$ and $H_{M+V}$ we can turn the estimates on
the hitting probabilities of the small box $\Delta^{\left(1\right)}$
and big box $\Delta^{\left(2\right)}$ in the hyperplane $H_{M+V}$
into both quenched and annealed estiamtions for the hitting probability
of the box $\Delta$ in the hyperplane $H_{M}$. The probability to
hit $\Delta$ and not to hit $\Delta^{\left(2\right)}$ as well as
the probability not to hit $\Delta$ but to hit $\Delta^{\left(1\right)}$
are of order $N^{-\xi\left(1\right)}$. }
\end{figure}
Annealed estimations (for the proof see Appendix \ref{Appendix_sub:More-annealed-estimations})
yields
\begin{equation}
\mathbb{P}^{z}\left(X_{T_{M+V}}\in\Delta^{\left(1\right)}\,,\, T_{M+V}\in I^{\left(1\right)}\right)<\mathbb{P}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right)+N^{-\xi\left(1\right)}\label{eq:First_difference_estimation_annealed_estimation_1}
\end{equation}
\begin{equation}
\mathbb{P}^{z}\left(X_{T_{M+V}}\in\Delta^{\left(2\right)}\,,\, T_{M+V}\in I^{\left(2\right)}\right)>\mathbb{P}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right)-N^{-\xi\left(1\right)}\label{eq:First_difference_estimation_annealed_estimation_2}
\end{equation}
and also, due to Claim \ref{Clm:simple_claim_annealed_to_quenched},
for an event $W_{2}\left(N\right)$ of $P$ probability $\geq1-N^{-\xi\left(1\right)}$
\begin{equation}
E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}\in\Delta^{\left(1\right)}\,,\, T_{M+V}\in I^{\left(1\right)}\right)\bigg|\mathcal{G}\right]<P_{\omega}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right)+N^{-\xi\left(1\right)}\label{eq:First_difference_estimation_annealed_estimation_3}
\end{equation}
\begin{equation}
E\left[P_{{\omega}}^{z}\left(X_{T_{M+V}}\in\Delta^{\left(2\right)}\,,\, T_{M+V}\in I^{\left(2\right)}\right)\bigg|\mathcal{G}\right]>P_{\omega}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right)-N^{-\xi\left(1\right)}\label{eq:First_difference_estimation_annealed_estimation_4}
\end{equation}
Thus, from the definition of $W_{1}\left(N\right)$, $W_{2}\left(N\right)$
and the last $4$ estimations, it follows that $W_{1}\left(N\right)\cap W_{2}\left(N\right)\subset L\left(\theta,N\right)$
and the proof is complete. \end{proof}
\begin{lem}
\label{lem:Upper_estimation_on_the_quenched_exit} Let $d\geq4$ and
assume $P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
For every $0<\theta\leq1$ and $h\in\mathbb{N}$ let $D^{\left(\theta,h\right)}\left(N\right)$
be the event that for every $z\in\tilde{{\mathcal{P}}}\left(0,N\right)$, every
$\frac{1}{2}N^{2}\leq M\leq N^{2}$, every $\left(d-1\right)$-dimensional
cube of side length $N^{\theta}$ which is contained in $H_{M}$ and
every interval $I$ of length $N^{\theta}$
\begin{equation}
P_{{\omega}}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right)\leq R_{h}\left(N\right)N^{\left(\theta-1\right)d}\label{eq:Upper_estimation_on_the_quenched_exit_1}
\end{equation}
and
\begin{equation}
P_{\omega}^{z}\left(X_{T_{M}}\in\Delta\right)\leq R_{h}\left(N\right)N^{\left(\theta-1\right)\left(d-1\right)}.\label{eq:Upper_estimation_on_the_quenched_exit_2}
\end{equation}
Then for every $0<\theta\leq1$ there exists $h=h\left(\theta\right)$
such that $P\left(D^{\left(\theta,h\right)}\left(N\right)\right)=1-N^{-\xi\left(1\right)}$. \end{lem}
\begin{proof}
The proof of (\ref{eq:Upper_estimation_on_the_quenched_exit_2}) is
the content of \cite[Lemma 4.13]{berger2008slowdown} and therefore
we restrict attention to the proof of (\ref{eq:Upper_estimation_on_the_quenched_exit_1}).
We prove the lemma by descending induction on $\theta$. From Lemma
\ref{lem:First_difference_estimation} together with Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time}(1),
$P\left(D^{\left(\theta,1\right)}\right)\geq P\left(L\left(\theta,N\right)\right)=1-N^{-\xi\left(1\right)}$
for every $\frac{d}{d+1}<\theta\leq1$. For the induction step, fix
$\theta$ and assume that the statement of the lemma holds for some
$\theta'$ such that $\theta>\frac{d}{d+1}\theta'$. Define $h'=h\left(\theta'\right)$
and $\rho=\frac{\theta}{\theta'}>\frac{d}{d+1}$. Let
\[
S\left(N\right)=D^{\left(\rho,1\right)}\left(N\right)\cap\!\!\!\!\!\!\bigcap_{\footnotesize{\begin{array}{c}
z\in{\mathcal{P}}\left(0,2N\right)\\
s\in\left[-2NR_{5}\left(N\right),2NR_{5}\left(N\right)\right]
\end{array}}}\!\!\!\!\!\!\!\!\sigma_{z}\varrho_{s}\left(D^{\left(\theta',h'\right)}\left(\left[N^{\rho}\right]\right)\right)\cap T\left(N,\rho\right),
\]
where $\varrho$ is the time shift for the random walk, defined by
$\varrho\left(X_{1},X_{2},\ldots\right)=\left(X_{2},X_{3},\ldots\right)$
and
\[
T\left(N,\rho\right)=\left\{ {\omega}\in{\Omega}\,:\,\begin{array}{c}
\forall v\in{\mathcal{P}}\left(0,N\right),\quad P_{{\omega}}^{v}\left(X_{T_{\partial{\mathcal{P}}\left(v,\left[N^{\rho}\right]\right)}}\notin\partial^{+}{\mathcal{P}}\left(v,\left[N^{\rho}\right]\right)\right)<e^{-R\left(N\right)}\\
\mbox{and }P_{{\omega}}^{v}\left(\left|T_{\partial{\mathcal{P}}\left(v,\left[N^{\rho}\right]\right)}-\mathbb{E}^{v}\left[T_{\partial{\mathcal{P}}\left(v,\left[N^{\rho}\right]\right)}\right]\right|>NR_{2}\left(N\right)\right)=N^{-\xi\left(1\right)}
\end{array}\right\} .
\]
From the definition of $S\left(N\right)$, Lemma \ref{lem:Distance_of_time_from_expectation}
and the induction assumption we know that $P\left(S\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
Therefore we need to show that for some $h$ and all $N$ large enough,
we have $S\left(N\right)\subset D^{\left(\theta,h\right)}\left(N\right)$.
To this end fix ${\omega}\in S\left(N\right)$, $z\in\tilde{{\mathcal{P}}}\left(0,N\right)$,
$\frac{1}{2}N^{2}\leq M\leq N^{2}$, a $\left(d-1\right)$-dimensional
cube $\Delta$ of size length $N^{\theta}$ in ${\mathcal{P}}\left(0,N\right)\cap H_{M}$
and an interval $I\subset\left[\mathbb{E}^{z}\left[T_{M}\right]-NR_{2}\left(N\right),\mathbb{E}^{z}\left[T_{M}\right]+NR_{2}\left(N\right)\right]$
of length $N^{\theta}$. As before we denote by $c\left(\Delta\right)$
and $c\left(I\right)$ the centers of $\Delta$ and $I$ respectively.
Let $V=\left[N^{\rho}\right]^{2}$, $c\left(\Delta\right)'=c\left(\Delta\right)-V\frac{\vartheta}{\left\langle \vartheta,e_{1}\right\rangle }$
and $c\left(I\right)'=c\left(I\right)-V\frac{1}{\left\langle \mathbbm{v},e_{1}\right\rangle }$.
Since ${\omega}\in\bigcap_{z\in{\mathcal{P}}\left(0,2N\right)}\bigcap_{s\in\left[-2NR_{5}\left(N\right),2NR_{5}\left(N\right)\right]}\sigma_{z}\varrho_{s}\left(D^{\left(\theta',h'\right)}\left(\left[N^{\rho}\right]\right)\right)$
it follows that for every $v\in H_{M-V}$ and every $t\in\mathbb{N}$
\[
P_{{\omega}}^{v}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I-t\right)<R_{h'}\left(N\right)N^{\rho\left(\theta'-1\right)d}=R_{h'}\left(N\right)N^{\left(\theta-\rho\right)d}.
\]
In addition due to the Markov property of the quenched law
\begin{align*}
P_{{\omega}}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right) & =\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{{\footnotesize\begin{array}{c}
v\in H_{M-V}\cap{\mathcal{P}}\left(x',\left[N^{\rho}\right]\right)\\
\left|t-c'\left(I\right)\right|\leq N^{\rho}R_{5}\left(N^{\rho}\right)
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\! P_{{\omega}}^{z}\left(X_{T_{M-V}}=v\,,\, T_{M-V}=t\right)P_{{\omega}}^{v}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I-t\right)+N^{-\xi\left(1\right)}\\
& \leq\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{{\footnotesize\begin{array}{c}
v\in H_{M-V}\cap{\mathcal{P}}\left(x',\left[N^{\rho}\right]\right)\\
\left|t-c'\left(I\right)\right|\leq N^{\rho}R_{5}\left(N^{\rho}\right)
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\! P_{{\omega}}^{z}\left(X_{T_{M-V}}=v\,,\, T_{M-V}=t\right)R_{h'}\left(N\right)N^{\left(\theta-\rho\right)d}+N^{-\xi\left(1\right)}.
\end{align*}
Now, the last sum can be separated into the sum over $2^{d-1}R_{5}\left(N^{\rho}\right)^{d-1}$
$\left(d-1\right)$-dimensional cubes of side length $N^{\rho}$ and
$2R_{2}\left(N^{\rho}\right)$ intervals of length $N^{\rho}$. Since
${\omega}\in D^{\left(\rho,1\right)}\left(N\right)$ the probability to
hit each such cube in any time interval is bounded by $R_{1}\left(N\right)N^{\left(\rho-1\right)d}$.
Thus
\begin{align*}
P_{{\omega}}^{z}\left(X_{T_{M}}\in\Delta\,,\, T_{M}\in I\right) & <2^{d}R_{5}\left(N^{\rho}\right)^{d}R_{1}\left(N\right)N^{\left(\rho-1\right)d}R_{h'}\left(N\right)N^{\left(\theta-\rho\right)d}+N^{-\xi\left(1\right)}\\
& \leq R_{\max\left\{ 6,h'\right\} +1}\left(N\right)N^{\left(\theta-1\right)d}.
\end{align*}
and the proof is complete by taking $h=\max\left\{ 6,h'\right\} +1$.
\end{proof}
\begin{lem}
\label{lem:Estimation_for_far_hyperplane} Let $d\geq4$ and assume
$P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Let ${\mathcal{G}}$ be the ${\sigma}$-algebra generated by $\left\{ {\omega}\left(z\right)\,:\,\left\langle z,e_{1}\right\rangle \leq N^{2}\right\} $.
Let $\eta>0$, $V=\left[N^{\eta}\right]$ and define $R\left(N,V\right)$
to be the event that for every $z\in\tilde{{\mathcal{P}}}\left(0,N\right)$,
every $v\in H_{N^{2}+V}$ and every $m\in\mathbb{N}$
\[
\left|E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}=v\,,\, T_{N^{2}+V}=m\right)\bigg|{\mathcal{G}}\right]-\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}=v\,,\, T_{N^{2}+V}=m\right)\right|\leq N^{-d}V^{\frac{1-d}{6}}
\]
Then $P\left(R\left(N,V\right)\right)=1-N^{-\xi\left(1\right)}$. \end{lem}
\begin{proof}
Let $v\in H_{N^{2}+V}$, $m\in\mathbb{N}$ and let $\theta>0$ be
such that $\theta<\frac{1}{20}\eta$. Let $K$ be an integer such
that $2^{-K-1}N^{2}\leq V<2^{-K}N^{2}$, and for $0\leq k\leq K$
define (see also Figure \ref{fig:The bullet})
\[
{\mathcal{P}}^{\left(k\right)}={\mathcal{P}}\left(0,N\right)\cap\left\{ x\in\mathbb{Z}^{d}\,:\,0\leq N^{2}-\left\langle x,e_{1}\right\rangle \leq2^{-k}N^{2}\right\} ,\quad\forall1\leq k\leq K,
\]
\[
{\mathcal{P}}^{\left(0\right)}={\mathcal{P}}\left(0,N\right)\cap\left\{ x\in\mathbb{Z}^{d}\,:\,\frac{N^{2}}{2}\leq N^{2}-\left\langle x,e_{1}\right\rangle \right\} ,
\]
\[
F\left(v\right)=\left\{ x\in\mathcal{P}\left(0,N\right)\,:\,\Biggl\Vert x-v-\vartheta\frac{\left\langle x-v,e_{1}\right\rangle }{\left\langle \vartheta,e_{1}\right\rangle }\Biggr\Vert_{\infty}\leq\left|\left\langle v-x,e_{1}\right\rangle \right|^{\frac{1}{2}}R_{2}\left(N\right)\right\} ,
\]
\[
{\mathcal{P}}^{\left(k\right)}\left(v\right)=\mathcal{P}^{\left(k\right)}\cap F\left(v\right),
\]
and
\[
\widehat{{\mathcal{P}}}^{\left(k\right)}\left(v\right)=\left\{ y\in\mathbb{Z}^{d}\,:\,\exists x\in\mathcal{P}^{\left(k\right)}\left(v\right)\,\mbox{such that }\left\Vert x-y\right\Vert _{\infty}<R_{2}\left(N\right)\right\} .
\]
\begin{figure}[h]
\begin{centering}
\includegraphics[scale=0.4]{pic/parabola_pic}
\par\end{centering}
\protect\caption{The checkered areas are $\mathcal{P}^{\left(k\right)}\left(v\right)$
for different values of $k$. The environment in the filled area has
negligible influence on the probability of hitting $v$. (the picture
is not to scale)\label{fig:The bullet}}
\end{figure}
Condition on the event $D^{\left(\theta,h\right)}$ from Lemma \ref{lem:Upper_estimation_on_the_quenched_exit},
with $h$ such that $P\left(D^{\left(\theta,h\right)}\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
As in \cite[Lemma 4.14]{berger2008slowdown} for $0\leq k\leq K$
and ${\omega}\in D^{\left(\theta,h\right)}$, we have the estimation
\[
V\left(k\right)=E_{{\omega},{\omega}}^{z,z}\left[\left|\left[X^{\left(1\right)}\right]\cap\left[X^{\left(2\right)}\right]\cap{\mathcal{P}}^{\left(k\right)}\left(v\right)\right|\right]\leq\begin{cases}
R_{2}\left(N\right) & \quad,k=0\\
R_{h+1}\left(N\right)N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)}2^{-k\left\lfloor \frac{d+1}{2}\right\rfloor } & \quad,1\leq k\leq K
\end{cases}.
\]
Indeed, for $k=0$ this follows from Lemma \ref{lem:Intersection_lemma}
while for $k>0$
\begin{align*}
V\left(k\right) & =\sum_{x\in\mathcal{P}^{\left(k\right)}\left(v\right)}\left[P_{\omega}^{z}\left(x\,\mbox{is visited}\right)\right]^{2}\\
& \leq\sum_{x\in\mathcal{P}^{\left(k\right)}\left(v\right)}\left[\sum_{y\,:\,\left\Vert y-x\right\Vert _{\infty}<R\left(N\right)}P_{\omega}^{z}\left(X_{T_{\left\langle y,e_{1}\right\rangle }}=y\right)\right]^{2}+N^{-\xi\left(1\right)}\\
& \overset{\left(1\right)}{\leq}\sum_{x\in\mathcal{P}^{\left(k\right)}\left(v\right)}C\cdot R^{d}\left(N\right)\cdot\sum_{y\,:\,\left\Vert y-x\right\Vert _{\infty}<R\left(N\right)}\left[P_{\omega}^{z}\left(X_{T_{\left\langle y,e_{1}\right\rangle }}=y\right)\right]^{2}+N^{-\xi\left(1\right)}\\
& \overset{\left(2\right)}{\leq}C\cdot R^{2d}\left(N\right)\sum_{y\in\hat{\mathcal{P}}^{\left(k\right)}\left(v\right)}\left[P_{\omega}^{z}\left(X_{T_{\left\langle y,e_{1}\right\rangle }}=y\right)\right]^{2}+N^{-\xi\left(1\right)}\\
& \overset{\left(3\right)}{\leq}R_{2}\left(N\right)\sum_{y\in\hat{\mathcal{P}}^{\left(k\right)}\left(v\right)}R_{h}^{2}\left(N\right)N^{2\left(1-\theta\right)\left(1-d\right)}\\
& \leq R_{h+1}\left(N\right)N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)}2^{-k\left\lfloor \frac{d+1}{2}\right\rfloor },
\end{align*}
where for $\left(1\right)$ we used Cauchy-Schwarz inequality, for
$\left(2\right)$ we used the fact that each point is counted at most
$R^{d}\left(N\right)$ times, and for $\left(3\right)$ the assumption
$\omega\in D^{\left(\theta,h\right)}\left(N\right)$.
We now use again the filtration $\left\{ {\mathcal{F}}_{i}\right\} $ from Lemma
\ref{lem:First_difference_estimation}, and consider the martingale
\[
M_{i}=E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}=v\,,\, T_{N^{2}+V}=m\bigg|B_{N}\right)\Bigg|\mathcal{F}_{i}\right].
\]
In order to use McDiarmid's inequality we need to bound $U_{i}=\esup\left(\left|M_{i}-M_{i-1}\right|\Big|\mathcal{F}_{i-1}\right)$
under the assumption $\omega\in D^{\left(\theta,h\right)}\left(N\right)$.
Let $x$ be such that $\omega_{x}$ is measurable with respect to
${\mathcal{F}}_{i}$ but not with respect to ${\mathcal{F}}_{i-1}$. By a similar argument
as in the proof of Lemma \ref{lem:First_difference_estimation} we
have $U_{i}=N^{-\xi\left(1\right)}$ if $x\notin F\left(v\right)$,
while for $x\in F\left(v\right)$
\[
U_{i}\leq R\left(N\right)E\left[P_{{\omega}}^{z}\left(x\,\mbox{is hit}|B_{N}\right)\Big|\mathcal{F}_{i-1}\right]Der\left(N^{2}+V-\left\langle x,e_{1}\right\rangle \right)
\]
where $Der\left(N^{2}+V-\left\langle x,e_{1}\right\rangle \right)$
is the maximal first derivative of the annealed distribution with
respect to both place and time at distance $N^{2}+V-\left\langle x,e_{1}\right\rangle $.
By Lemma \ref{lem:Annealed_derivative_estimations_(d-1)+time} this
derivative is bounded by $CN^{-d-1}2^{k\frac{d}{2}}$. Therefore whenever
$\omega\in D^{\left(\theta,h\right)}$
\begin{align*}
U & =\sum_{i}U_{i}^{2}\leq C\sum_{k=0}^{K}V\left(k\right)N^{-2\left(d+1\right)}2^{kd}+N^{-\xi\left(1\right)}\\
& \leq CR_{h+1}\left(N\right)N^{-2\left(d+1\right)}\\
& +CR_{h+1}\left(N\right)N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)-2\left(d+1\right)}\sum_{k=1}^{K}2^{kd-k\frac{d+1}{2}}+N^{-\xi\left(1\right)}\\
& \leq CR_{h+1}\left(N\right)\left[N^{-2\left(d+1\right)}+N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)-2\left(d+1\right)}c2^{\frac{d-1}{2}K}\right]+N^{-\xi\left(1\right)}\\
& \leq CR_{h+1}\left(N\right)\left[N^{-2\left(d+1\right)}+N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)-2\left(d+1\right)}2^{\frac{d-1}{2}K}\right]+N^{-\xi\left(1\right)}
\end{align*}
Recalling that $K$ was chosen so that $2^{K}<N^{2}V^{-1}$ we can
bound the last sum term by
\begin{align*}
& CR_{h+1}\left(N\right)\left[N^{-2\left(d+1\right)}+N^{2\left(\frac{d+1}{2}+\left(1-\theta\right)\left(1-d\right)\right)-2\left(d+1\right)}N^{\left(d-1\right)}V^{-\frac{d-1}{2}}\right]+N^{-\xi\left(1\right)}\\
= & CR_{h+1}\left(N\right)\left[N^{-2\left(d+1\right)}+N^{-2d-2\theta\left(1-d\right)}V^{-\frac{d-1}{2}}\right]+N^{-\xi\left(1\right)}\\
\leq & CN^{-2d}N^{-\frac{d-1}{6}+\epsilon}
\end{align*}
for some small enough ${\varepsilon}>0$.
Using McDiarmid's inequality (see Theorem \ref{thm:Azuma_type_inequality})
{\footnotesize
\begin{align*}
& P\left(\left|E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\bigg|{\mathcal{G}}\right]-\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\right|>N^{-d}V^{\frac{1-d}{6}}\right)\\
\leq & P\left(\left|E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\bigg|{\mathcal{G}}\right]-\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\right|>N^{-d}V^{\frac{1-d}{6}},\, D^{\left(\theta,h\right)}\left(N\right)\right)+N^{-\xi\left(1\right)}\\
\leq & P\left(\left|E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\bigg|{\mathcal{G}}\right]-\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}=v,\, T_{N^{2}+V}=m,\, B_{N}\right)\right|>N^{-d}V^{\frac{1-d}{6}},\, U\leq CN^{-2d-\frac{d-1}{6}+\varepsilon}\right)+N^{-\xi\left(1\right)}\\
\leq & C\exp\left(-cN^{-\frac{d+1}{6}\left(1-2\eta\right)-\varepsilon}\right)+N^{-\xi\left(1\right)}=N^{-\xi\left(1\right)}.
\end{align*}
}Since $P\left(B_{N}\right)=1-N^{-\xi\left(1\right)}$ by Corollary
\ref{cor:length_of_regenerations}, this completes the proof.
\end{proof}
We are finally ready to prove Proposition \ref{prop:Time_(d-1)_box_difference}.
\begin{proof}[Proof of Proposition \ref{prop:Time_(d-1)_box_difference}]
Let $\eta>0$ and define $V=\left[N^{\eta}\right]$. By Lemma \ref{lem:Estimation_for_far_hyperplane}
we know that $P\left(R\left(N,V\right)\right)=1-N^{-\xi\left(1\right)}$.
As before, all we need to show is that $R\left(N,V\right)\cap S\left(N,V\right)\subseteq F\left(N,\theta\right)$
for an appropriate choice of $\eta>0$ and some event $S\left(N,V\right)$
satisfying $P\left(S\left(N,V\right)\right)=1-N^{-\xi\left(1\right)}$.
This is done identically as in the last step of the proof of Lemma
\ref{lem:First_difference_estimation}. Let ${\omega}\in R\left(N,V\right)$,
let $\Delta$ be a cube of side length $N^{\theta}$ which is contained
in $\partial^{+}{\mathcal{P}}\left(0,N\right)$ and let $I$ be an interval
of length $N^{\theta}$ in $\mathbb{N}$. As in Lemma \ref{lem:First_difference_estimation}
(see also Figure \ref{fig:Bix_small_box}) we denote by $c\left(\Delta\right)$
and $c\left(I\right)$ the center of $\Delta$ and $I$ respectively
and let $c'\left(\Delta\right)=c\left(\Delta\right)+V\frac{\vartheta}{\left\langle \vartheta,e_{1}\right\rangle }$,
$c'\left(I\right)=c\left(I\right)+\mathbb{E}^{0}\left[T_{V}\right]$.
Let $\Delta^{\left(1\right)}$ and $\Delta^{\left(2\right)}$ be $\left(d-1\right)$-dimensional
cubes that are contained in $H_{N^{2}+V}$ , centered at $c'\left(\Delta\right)$
and are of side lengths $N^{\theta}-R_{3}\left(N\right)\sqrt{V}$
and $N^{\theta}+R_{3}\left(N\right)\sqrt{V}$ respectively. In a similar
fashion let $I^{\left(1\right)}$ and $I^{\left(2\right)}$ be intervals
centered at $c'\left(I\right)$ which are of lengths $N^{\theta}-R_{3}\left(N\right)\sqrt{V}$
and $N^{\theta}+R_{3}\left(N\right)\sqrt{V}$ respectively.
As in the proof of Lemma \ref{lem:First_difference_estimation} (the
proof can be found in Appendix \ref{Appendix_sub:More-annealed-estimations})
we know that
\begin{equation}
\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(1\right)}\,,\, T_{N^{2}V}\in I^{\left(1\right)}\right)<\mathbb{P}^{z}\left(X_{T_{N^{2}}}\in\Delta\,,\, T_{N^{2}}\in I\right)+N^{-\xi\left(1\right)}\label{eq::Time_(d-1)_box_difference_annealed_estimation_1}
\end{equation}
\begin{equation}
\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(2\right)}\,,\, T_{N^{2}+V}\in I^{\left(2\right)}\right)>\mathbb{P}^{z}\left(X_{T_{N^{2}}}\in\Delta\,,\, T_{N^{2}}\in I\right)-N^{-\xi\left(1\right)}\label{eq::Time_(d-1)_box_difference_annealed_estimation_2}
\end{equation}
and using Claim \ref{Clm:simple_claim_annealed_to_quenched} for an
event $S\left(N,V\right)$ of probability $P\left(S\left(N\right)\right)=1-N^{-\xi\left(1\right)}$
\begin{equation}
E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(1\right)}\,,\, T_{N^{2}+V}\in I^{\left(1\right)}\right)\bigg|\mathcal{G}\right]<P_{\omega}^{z}\left(X_{T_{N^{2}}}\in\Delta\,,\, T_{N^{2}}\in I\right)+N^{-\xi\left(1\right)}\label{eq::Time_(d-1)_box_difference_annealed_estimation_3}
\end{equation}
\begin{equation}
E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(2\right)}\,,\, T_{N^{2}+V}\in I^{\left(2\right)}\right)\bigg|\mathcal{G}\right]>P_{\omega}^{z}\left(X_{T_{N^{2}}}\in\Delta\,,\, T_{N^{2}}\in I\right)-N^{-\xi\left(1\right)}.\label{eq::Time_(d-1)_box_difference_annealed_estimation_4}
\end{equation}
In addition on the event $R\left(N,V\right)$, for $i=1,2$
\[
\left|E\left[P_{{\omega}}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(i\right)}\,,\, T_{N^{2}+V}\in I^{\left(i\right)}\right)\bigg|{\mathcal{G}}\right]-\mathbb{P}^{z}\left(X_{T_{N^{2}+V}}\in\Delta^{\left(i\right)}\,,\, T_{N^{2}+V}\in I^{\left(i\right)}\right)\right|\leq\left|\Delta^{\left(i\right)}\right|\cdot\left|I^{\left(i\right)}\right|\cdot N^{-d}V^{\frac{1-d}{6}}.
\]
Therefore for ${\omega}\in R\left(N,V\right)\cap S\left(N,V\right)$ we
have
\begin{align*}
& \left|P_{{\omega}}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\Delta\,,\, T_{\partial{\mathcal{P}}\left(0,N\right)}\in I\right)-\mathbb{P}^{z}\left(X_{T_{\partial{\mathcal{P}}\left(0,N\right)}}\in\Delta\,,\, T_{\partial{\mathcal{P}}\left(0,N\right)}\in I\right)\right|\\
\leq & \left(\left|\Delta^{\left(1\right)}\right|\left|I^{\left(1\right)}\right|+\left|\Delta^{\left(2\right)}\right|\left|I^{\left(2\right)}\right|\right)N^{-d}V^{\frac{1-d}{6}}+\left(\left|\Delta^{\left(2\right)}\right|\left|I^{\left(2\right)}\right|-\left|\Delta^{\left(1\right)}\right|\left|I^{\left(1\right)}\right|\right)CN^{-d}+N^{-\xi\left(1\right)}\\
\leq & C\left[\left(N^{\theta}+R_{3}\left(N\right)\sqrt{V}\right)^{d}N^{-d}V^{\frac{1-d}{6}}+R_{3}\left(N\right)\sqrt{V}N^{\theta\left(d-1\right)-d}\right].
\end{align*}
Taking $\eta<2\theta$ we can bound the last term by
\[
C\left[N^{\theta d-d+\eta\frac{1-d}{6}}+R_{3}\left(N\right)N^{\theta\left(d-1\right)-d+\frac{\eta}{2}}\right].
\]
Notice that the exponents of the powers of $N$ are the same when
$\eta=\frac{6\theta}{d+2}<2\theta$, in which case the last bound
equals $C\left(1+R_{3}\left(N\right)\right)\cdot N^{-d\left(1-\theta\right)-\frac{d-1}{d+2}\theta}\leq CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}$
. Thus the proof is complete.
\end{proof}
\section{\label{sec:Lorentz_trans}From $\left(d-1\right)$-dimensional boxes
plus time to $d$-dimensional boxes in a fixed time}
The goal of this section is to use the estimation proved in Section
\ref{sec:Adding-time-estimation}, for the difference between the
quenched and annealed probabilities to hit boxes in a hyperplane within
a specific time interval, in order to achieve similar estimation for
the difference between the quenched and annealed probabilities to
hit a $d$-dimensional box in a specific time. Formally, we have the
following:
\begin{prop}
\label{Prop:Lorentz_transformation-1} Let $d\geq4$ and assume $P$
is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
For every $0<\theta\leq\frac{1}{2}$, let $H\left(N\right)=H\left(\theta,N\right)$
be the event that for every $z\in\tilde{\mathcal{P}}\left(0,N\right)$
every $d$-dimensional cube $\Delta$ of side length $N^{\theta}$
\[
\left|P_{\omega}^{z}\left(X_{N}\in\Delta\right)-\mathbb{P}^{z}\left(X_{N}\in\Delta\right)\right|\leq CN^{-d\left(1-\theta\right)-\frac{1}{3}\theta}
\]
Then $P\left(H\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.\end{prop}
\begin{rem}
The constant $\frac{1}{3}$ can in fact be replaced by any number
which is smaller than $\min\left\{ \frac{1}{2},\frac{d-2}{d+2}\right\} $.
\end{rem}
The idea of the proof is to exploit the estimation of Proposition
\ref{prop:Time_(d-1)_box_difference} and the fact that regeneration
times occur quite often. More precisely, we show that the event of
hitting a box $\Delta$ at time $N$ is bounded both from below and
from above by the event of hitting a certain hyperplane in a specific
$\left(d-1\right)$-dimensional box within a specific time interval.
This implies that the difference between the probabilities is roughly
the same as in Proposition \ref{prop:Time_(d-1)_box_difference} and
thus gives the required result.
\begin{proof}
Due to Lemma \ref{lem:good_estimation_for_the_location} we may restrict
ourselves to boxes $\Delta$ whose center $c\left(\Delta\right)$
satisfies $\left\Vert c\left(\Delta\right)-\mathbb{E}^{z}\left[X_{N}\right]\right\Vert _{\infty}<\sqrt{N}R_{5}\left(N\right)$.
Given a cube $\Delta$ of side length $N^{\theta}$ such that $c\left(\Delta\right)$
satisfies $\left\Vert c\left(\Delta\right)-\mathbb{E}^{z}\left[X_{N}\right]\right\Vert _{\infty}<\sqrt{N}R_{5}\left(N\right)$
let $\Delta^{\left(1\right)}$ and $\Delta^{\left(2\right)}$ be the
$\left(d-1\right)$-dimensional cubes in the hyperplane $H_{\left\langle c\left(\Delta\right),e_{1}\right\rangle -N^{\theta}}$
with center $c\left(\Delta\right)-\frac{N^{\theta}}{\left\langle \vartheta,e_{1}\right\rangle }\vartheta$
and side length $N^{\theta}-R_{5}\left(N\right)N^{\frac{1}{2}\theta}$
and $N^{\theta}+R_{5}\left(N\right)N^{\frac{1}{2}\theta}$ respectively.
Noting that
\[
L:=\left\langle c\left(\Delta\right),e_{1}\right\rangle -N^{\theta}\geq\left\langle \mathbb{E}^{z}\left[X_{N}\right],e_{1}\right\rangle -2\sqrt{N}R_{5}\left(N\right)\geq cN-2\sqrt{N}R_{5}\left(N\right),
\]
it follows from Proposition \ref{prop:Time_(d-1)_box_difference}
(for every $\omega\in F\left(N,\theta\right)$ and every $z\in\tilde{\mathcal{P}}\left(0,N\right)$)
that
\begin{equation}
\left|P_{\omega}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\right)-\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\right)\right|\leq CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta},\label{eq:Lorentz_trans_1}
\end{equation}
with $I^{\left(1\right)}=N-\frac{N^{\theta}}{\left\langle \mathbbm{v},e_{1}\right\rangle }+\left[-\frac{1}{2}\left(N^{\theta}-R_{5}\left(N\right)N^{\frac{1}{2}\theta}\right),\frac{1}{2}\left(N^{\theta}-R_{5}\left(N\right)N^{\frac{1}{2}\theta}\right)\right]$,
and
\begin{equation}
\left|P_{\omega}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)-\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)\right|\leq CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta},\label{eq:Lorentz_trans_2}
\end{equation}
with $I^{\left(2\right)}=N-\frac{N^{\theta}}{\left\langle \mathbbm{v},e_{1}\right\rangle }+\left[-\frac{1}{2}\left(N^{\theta}+R_{5}\left(N\right)N^{\frac{1}{2}\theta}\right),\frac{1}{2}\left(N^{\theta}+R_{5}\left(N\right)N^{\frac{1}{2}\theta}\right)\right]$.
In addition, by a standard CLT type arguments, as the one in Lemma
\ref{lem:First_difference_estimation} (see Appendix \ref{Appendix_sub:More-annealed-estimations}
for the proof) we have the following annealed estimations
\begin{equation}
\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\,,\, X_{N}\notin\Delta\right)\leq N^{-\xi\left(1\right)},\label{eq:Lorentz_trans_3}
\end{equation}
\begin{equation}
\mathbb{P}^{z}\left(\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)^{c}\,,\, X_{N}\in\Delta\right)\leq N^{-\xi\left(1\right)}.\label{eq:Lorentz_trans_4}
\end{equation}
Using Claim \ref{Clm:simple_claim_annealed_to_quenched} again this
also implies that
\begin{equation}
P\left(\left\{ \omega\in\Omega\,:\, P_{\omega}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\,,\, X_{N}\notin\Delta\right)\leq N^{-\xi\left(1\right)}\right\} \right)\geq1-N^{-\xi\left(1\right)}\label{eq:Lorentz_trans_5}
\end{equation}
and
\begin{equation}
P\left(\left\{ \omega\in\Omega\,:\, P_{\omega}^{z}\left(\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)^{c}\,,\, X_{N}\in\Delta\right)\leq N^{-\xi\left(1\right)}\right\} \right)\geq1-N^{-\xi\left(1\right)}.\label{eq:Lorentz_trans_6}
\end{equation}
Combining all of the above we get for an event with $P$ probability
$\geq1-N^{-\xi\left(1\right)}${\small
\begin{align*}
& P_{\omega}^{z}\left(X_{N}\in\Delta\right)\\
\overset{\eqref{eq:Lorentz_trans_6}}{\leq} & P_{\omega}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\,,\, X_{N}\in\Delta\right)+N^{-\xi\left(1\right)}\\
\leq & P_{\omega}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)+N^{-\xi\left(1\right)}\\
\overset{\eqref{eq:Lorentz_trans_2}}{\leq} & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)+CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}+N^{-\xi\left(1\right)}\\
\leq & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\right)+\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\backslash\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)\\
+ & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\backslash I^{\left(1\right)}\right)+CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}\\
\overset{\eqref{eq:Lorentz_trans_3}}{\leq} & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(1\right)}\,,\, X_{N}\in\Delta\right)+\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\backslash\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)\\
+ & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\backslash I^{\left(1\right)}\right)+CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}\\
\leq & \mathbb{P}^{z}\left(X_{N}\in\Delta\right)+\mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\backslash\Delta^{\left(1\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\right)\\
+ & \mathbb{P}^{z}\left(X_{T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}}\in\Delta^{\left(2\right)}\,,\, T_{\partial\mathcal{P}\left(0,\sqrt{L}\right)}\in I^{\left(2\right)}\backslash I^{\left(1\right)}\right)+CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}\\
\overset{(1)}{\leq} & \mathbb{P}^{z}\left(X_{N}\in\Delta\right)+CR_{5}\left(N\right)\cdot N^{\theta\left(d-\frac{1}{2}\right)}\cdot N^{-d}+CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}\\
= & \mathbb{P}^{z}\left(X_{N}\in\Delta\right)+CN^{-d\left(1-\theta\right)-\frac{1}{3}\theta}
\end{align*}
}Where for $\left(1\right)$ we used the annealed derivative estimation
proved in Lemma (\ref{lem:Annealed_derivative_estimations_(d-1)+time})
as well as the fact that the number of pairs of location and time
$\left(x,t\right)$ such that $\left(x,t\right)\in$$\left(\Delta^{\left(2\right)}\backslash\Delta^{\left(1\right)},I^{\left(2\right)}\right)$
or $\left(x,t\right)\in\left(\Delta^{\left(2\right)},I^{\left(2\right)}\backslash I^{\left(1\right)}\right)$
is bounded by $CR_{5}\left(N\right)N^{\theta\left(d-\frac{1}{2}\right)}$.
The other direction
\[
P_{\omega}^{z}\left(X_{N}\in\Delta\right)\geq\mathbb{P}^{z}\left(X_{N}\in\Delta\right)-CN^{-d\left(1-\theta\right)-\frac{1}{3}\theta}
\]
follows via the same argument except we use (\ref{eq:Lorentz_trans_1}),(\ref{eq:Lorentz_trans_4})
and (\ref{eq:Lorentz_trans_5}) instead of (\ref{eq:Lorentz_trans_2}),
(\ref{eq:Lorentz_trans_3}) and (\ref{eq:Lorentz_trans_6}).
\end{proof}
\section{Total variation bound for finite boxes}
In the previous section it was shown that for every $0<\theta\leq1$
the difference between the quenched and annealed probabilities to
hit a $d$-dimensional box of side length $N^{\theta}$ at time $N$
is bounded by $CN^{-d\left(1-\theta\right)-\frac{d-2}{d+2}\theta}$.
Since the particle must be inside a the box $\left[-N,N\right]^{d}$
at time $N$ this in particular implies that the total variation between
the quenched and annealed distribution over any partition of $\mathbb{Z}^{d}$
into $d$-dimensional boxes of side length $N^{\theta}$ goes to zero
as $N$ goes to infinity. The goal of this section is to strengthen
this result and prove that the same result hold for partitions of
$\mathbb{Z}^{d}$ into boxes whose side length is of constant size,
independent of $N$. More formally we have the following:
\begin{thm}
\label{thm:Total_variation_on_small_boxes} Let $d\geq4$ and assume
$P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
For $N,M\in\mathbb{N}$ denote by $G\left(N\right)=G\left(N,M\right)$
the set of environments $\omega\in\Omega$ such that for every $z\in\mathbb{Z}^{d}$
satisfying $\left\Vert z\right\Vert _{\infty}\leq N$
\[
\sum_{\Delta\in\Pi}\left|P_{\omega}^{z}\left(X_{N}\in\Delta\right)-\mathbb{P}^{z}\left(X_{N}\in\Delta\right)\right|\leq\frac{C_{2}}{M^{c_{1}}}+\frac{C_{2}}{N^{c_{1}}},
\]
where $\Pi$ is any partition of $\mathbb{Z}^{d}$ into boxes of side
length $M$. Then for an appropriate $0<c_{1},C_{2}<\infty$, $P\left(G\left(N\right)\right)=1-N^{-\xi\left(1\right)}$.
\end{thm}
The idea of the proof is to shrink the size of the boxes in steps,
each time by a constant factor from the previous step. This is done
as follows: First we fix some factor, say $\theta=\frac{1}{200}$.
Then, in the $k^{th}$ step of the process, we let the random walk
run for $N^{\frac{1}{2^{k}}}$ steps and ask for the difference between
the annealed and quenched measures hit the same box of side $N^{\frac{\theta}{2^{k}}}$.
Repeating the last procedure roughly $\log\log N$ times we get boxes
of constant side length $M$. The idea is to bound the total variation
of the $\left(k+1\right)^{th}$ step of this process by the one of
the $k^{th}$ step. Denoting by $\lambda_{k}$ the total variation
of the $k^{th}$ step we show that $\lambda_{k}\leq\lambda_{k-1}+CN_{k}^{-\alpha}$
for some $C,\alpha>0$. An additional short calculation then yields
the result.
\begin{proof}
We start with some notations to be used throughout the proof. Let
$\theta=\frac{1}{200}$. For $j\in\mathbb{N}$ denote $N_{j}=\left\lfloor N^{\frac{1}{2^{j}}}\right\rfloor $
and let $r\left(N\right)=\left\lceil \log_{2}\left(\frac{\log N}{\theta\log M}\right)\right\rceil $
(which is the minimal natural number such that $N_{r\left(N\right)}^{\theta}\leq M$).
Moreover, denote $n_{0}=n-\sum_{j=1}^{r\left(N\right)}N_{j}$ and
$n_{k}=\sum_{j=1}^{k}N_{j}$ for $1\leq k\leq r\left(N\right)$. For
$0\leq k\leq r\left(n\right)$ choose $\Pi_{k}$ to be a partition
of $\mathbb{Z}^{d}$ into boxes of side length $\left\lfloor N_{k}^{\theta}\right\rfloor $.
Finally, for $0\leq k\leq r\left(N\right)$ let
\[
\lambda_{k}=\sum_{\Delta\in\Pi_{k}}\left|P_{\omega}^{z}\left(X_{n_{k}}\in\Delta\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta\right)\right|.
\]
Note that in particular $\lambda_{r\left(N\right)}$ is the total
variation between the quenched and annealed measures on cubes of side
length $\leq M$ which is the term we wish to bound from above. In
fact, if one wish to be more precise, one should replace $N_{r\left(N\right)}$
by $M$ and thus get total variation for boxes of side length $M$,
this however does not influence the estimates to follow.
As stated before the main idea of the proof is to prove an inequality
of the form
\[
\lambda_{k}\leq\lambda_{k-1}+CN_{k}^{-\alpha}
\]
for some $\alpha>0$, which immediately implies $\lambda_{r\left(N\right)}\leq\lambda_{1}+C\sum_{k=1}^{r\left(N\right)}N_{k}^{-\alpha}$.
As it turns out the last term is bounded by $C_{2}M^{-c_{1}}$ for
some constants $0<c_{1},C_{2}<\infty$, while the first term, i.e.
$\lambda_{1}$, is bounded (due to Proposition \ref{Prop:Lorentz_transformation-1})
by $CN^{-\frac{d-2}{d+2}\theta}$, and the result follows.
We start estimating $\lambda_{k}$ for $k\geq2$. By the triangle
inequality and the Markov property of $P_{\omega}$ we have
\begin{align}
\lambda_{k} & =\sum_{\Delta\in\Pi_{k}}\left|P_{\omega}^{z}\left(X_{n_{k}}\in\Delta\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta\right)\right|\nonumber \\
& =\sum_{\Delta\in\Pi_{k}}\left|\sum_{\Delta'\in\Pi_{k-1}}\left[P_{\omega}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)\right]\right|\nonumber \\
& \leq\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|P_{\omega}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)\right|\nonumber \\
& \leq\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\left[P_{\omega}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\right]\right|\label{eq:lam_k_estimation_part_1}\\
& +\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\left[P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right]\right|\label{eq:lam_k+estimation_part_2}\\
& +\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\left[\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)\right]\right|\label{eq:lam_k+estimation_part_3}\\
& +\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)\right|.\label{eq:lam_k+estimation_part_4}
\end{align}
We turn to estimate each of the terms (\ref{eq:lam_k_estimation_part_1}-\ref{eq:lam_k+estimation_part_4})
separately.
For the first term (\ref{eq:lam_k_estimation_part_1}), we have
\begin{align*}
\eqref{eq:lam_k_estimation_part_1} & =\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\left[P_{\omega}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\right]\right|\\
& \leq\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\left|P_{\omega}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\right|\\
& =\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}\left|P_{\omega}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\right|\\
& =\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\left|P_{\omega}^{z}\left(X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\right|\\
& =\sum_{\Delta'\in\Pi_{k-1}}\left|P_{\omega}^{z}\left(X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\right|=\lambda_{k-1}.
\end{align*}
For the second term (\ref{eq:lam_k+estimation_part_2}), the triangle
inequality yields
\begin{align*}
\eqref{eq:lam_k+estimation_part_2} & =\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\left[P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right]\right|\\
& \leq\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\sum_{\Delta\in\Pi_{k}}\left|P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|.
\end{align*}
By Lemma \ref{lem:good_estimation_for_the_location} this can be bounded
by
\begin{align}
& \sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{\tiny{\begin{array}{c}
\Delta\in\Pi_{k}\\
\mbox{dist}\left(\Delta,u\right)\\
\leq\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\left|P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|\nonumber \\
+ & \left(n_{k}-n_{k-1}\right)^{-\xi\left(1\right)}.\label{eq:lam_k_estimation_part_2_2}
\end{align}
We say that a cube $\Delta'\in\Pi_{k-1}$ is good if for every $u\in\Delta'$
and every $\Delta\in\Pi_{k}$
\[
\left|P_{\omega}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|\leq CN_{k}^{\left(\theta-1\right)\left(d-1\right)-\frac{1}{3}\theta},
\]
otherwise we say that $\Delta'$ is bad. Note that the condition holds
trivially for all $\Delta$ such that $\mbox{dist}\left(u,\Delta\right)>n_{k}-n_{k-1}$.
Separating the sum in (\ref{eq:lam_k_estimation_part_2_2}) into the
sum over good and bad boxes, and noting that
\begin{itemize}
\item at time $n_{k}$ for every $u$ we only need to consider boxes $\Delta$
such that $\mbox{dist}\left(\Delta,u\right)\leq\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)=\sqrt{N_{k}}R_{5}\left(N_{k}\right)$
whose number is bounded by $\frac{N_{k}^{\frac{d}{2}}R_{5}^{d}\left(N_{k}\right)}{\left|\Delta\right|}$.
\item We only need to consider boxes $\Delta'$ such that $\mbox{dist}\left(z,\Delta'\right)\leq n_{k-1}$
\end{itemize}
we conclude that (\ref{eq:lam_k_estimation_part_2_2}) is bounded
by
\begin{align*}
& CN_{k}^{\left(\theta-1\right)\left(d-1\right)-\frac{1}{3}\theta}\cdot\frac{N_{k}^{\frac{d}{2}}R_{5}\left(N_{k}\right)}{\left\lfloor N_{k}^{\theta}\right\rfloor ^{d}}+2\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{\tiny{\begin{array}{c}
\mbox{dist}\left(z,\Delta'\right)\leq n_{k-1}\\
\Delta'\,\mbox{is bad}
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\\
\leq & CN_{k}^{1-\frac{4}{3}\theta-\frac{d}{2}}R_{5}^{d}\left(N_{k}\right)+2\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{\tiny{\begin{array}{c}
\mbox{dist}\left(z,\Delta'\right)\leq n_{k-1}\\
\Delta'\,\mbox{is bad}
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right).
\end{align*}
Using the annealed estimation from Lemma \ref{lem:general_annealed_estimations}
we can bound the last term by
\[
CN_{k}^{1-\frac{4}{3}\theta-\frac{d}{2}}R_{5}^{d}\left(N_{k}\right)+\#\left\{ \Delta'\in\Pi_{k-1}\,:\,\mbox{dist}\left(\Delta',z\right)\leq n_{k-1}\,,\,\Delta'\,\mbox{is bad}\right\} \cdot\frac{C\left\lfloor N_{k-1}^{\theta}\right\rfloor ^{d}}{n_{k-1}^{\frac{d}{2}}}.
\]
Notice however that Proposition \ref{Prop:Lorentz_transformation-1}
implies that the event $G=\left\{ \mbox{all boxes }\Delta'\,\mbox{such that }\mbox{dist}\left(\Delta',z\right)\leq n_{k-1}\,\mbox{are good}\right\} $
satisfies $P\left(G^{c}\right)\leq n_{k-1}^{d}\cdot N^{-\xi\left(1\right)}=N^{-\xi\left(1\right)}$
and therefore under the last event we have $\eqref{eq:lam_k+estimation_part_2}\leq CN_{k}^{1-\frac{4}{3}\theta-\frac{d}{2}}R_{5}^{d}\left(N_{k}\right)$.
Turning to deal with (\ref{eq:lam_k+estimation_part_3}), notice that
\begin{align*}
\eqref{eq:lam_k+estimation_part_3} & =\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\left[\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)P_{\omega}^{z}\left(X_{n_{k-1}}=u\bigg|X_{n_{k-1}}\in\Delta'\right)-\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)\right]\right|\\
& \leq\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\left|\max_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\min_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|\\
& =\!\!\!\!\sum_{\Delta'\in\Pi_{k-1}}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\sum_{\Delta\in\Pi_{k}}\left|\max_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\min_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|\\
& \overset{\left(1\right)}{\leq}\!\!\!\!\sum_{\Delta'\in\Pi_{k-1}}\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\sum_{\tiny{\begin{array}{c}
\Delta\in\Pi_{k}\,\,\mbox{s.t.}\,\,\exists u\in\Delta'\\
\mbox{dist}\left(\Delta,\mathbb{E}^{u}\left[X_{n_{k}-n_{k-1}}\right]\right)\\
\leq\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)
\end{array}}}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\left|\max_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\min_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\right|+\left(n_{k}-n_{k-1}\right)^{-\xi\left(1\right)},
\end{align*}
where for $\left(1\right)$ we used Lemma \ref{lem:good_estimation_for_the_location}.
Due to the annealed derivative estimation from Lemma \ref{lem:general_annealed_estimations}
we can bound the last term by
\begin{align*}
& \sum_{\Delta'\in\Pi_{k-1}}\!\!\!\!\mathbb{P}^{z}\left(X_{n_{k-1}}\in\Delta'\right)\underset{_{\mbox{num of relevant boxes}}}{\underbrace{\left(\frac{dN_{k-1}^{\theta}+\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)}{\left\lfloor N_{k}^{\theta}\right\rfloor }\right)^{d}}}\cdot\!\!\!\!\!\!\!\!\!\underset{_{\mbox{size of each box}}}{\underbrace{\vphantom{\frac{C}{\left(n_{k}\right)^{\frac{d+1}{2}}}}\left\lfloor N_{k}^{\theta}\right\rfloor ^{d}}}\!\!\!\!\cdot\underset{\,\,\,\,\,\,{}_{\mbox{derivative estimation}}}{\underbrace{\frac{C}{\left(n_{k}-n_{k-1}\right)^{\frac{d+1}{2}}}}}\!\!\!\!\!+\left(n_{k}-n_{k-1}\right)^{-\xi\left(1\right)}\\
= & \frac{C\left(dN_{k-1}^{\theta}+\sqrt{N_{k}}R_{5}\left(N_{k}\right)\right)^{d}}{N_{k}^{\frac{d+1}{2}}}+N_{k}^{-\xi\left(1\right)}\leq CR_{6}\left(N_{k}\right)N_{k}^{-\frac{1}{2}}.
\end{align*}
Finally, for (\ref{eq:lam_k+estimation_part_4}) we have
\begin{align*}
\eqref{eq:lam_k+estimation_part_4} & =\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta,\, X_{n_{k-1}}\in\Delta'\right)\right|\\
& =\sum_{\Delta\in\Pi_{k}}\sum_{\Delta'\in\Pi_{k-1}}\left|\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)\left[\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta\bigg|X_{n_{k-1}}=u\right)\right]\right|\\
& \leq\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)\sum_{\Delta\in\Pi_{k}}\left|\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta\bigg|X_{n_{k-1}}=u\right)\right|.
\end{align*}
Notice that under the event $B_{N}$, which by Corollary \ref{cor:length_of_regenerations}
satisfies $P\left(B_{N}\right)\geq1-N^{-\xi\left(1\right)}$, the
first regeneration time after hitting $u$ is after no more than $R\left(N\right)$
steps. Therefore, the distance between the regeneration times of both
annealed walks started in $u$ and started in $z$ conditioned to
hit $u$ is at most $2R\left(N\right)$ one from the other. Using
the annealed derivative estimation from Lemma \ref{lem:general_annealed_estimations}
for the annealed walks after the regeneration times we get
\[
\left|\mathbb{P}^{u}\left(X_{n_{k}-n_{k-1}}\in\Delta\right)-\mathbb{P}^{z}\left(X_{n_{k}}\in\Delta\bigg|X_{n_{k-1}}=u\right)\right|\leq\frac{CR\left(N\right)\cdot N_{k}^{d\theta}}{\left(n_{k}-n_{k-1}-R\left(N\right)\right)^{\frac{d+1}{2}}}\leq\frac{CR\left(N\right)\cdot N_{k}^{d\theta}}{N_{k}^{\frac{d+1}{2}}},
\]
recalling that due to Lemma \ref{lem:good_estimation_for_the_location}
we only need to consider boxes $\Delta$ at distance $\leq\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)$
from the annealed expectation $\mathbb{E}^{u}\left[X_{n_{k}-n_{k-1}}\right]$,
it follows that
\begin{align*}
\eqref{eq:lam_k+estimation_part_4} & \leq\sum_{\Delta'\in\Pi_{k-1}}\sum_{u\in\Delta'}\mathbb{P}^{z}\left(X_{n_{k-1}}=u\right)\cdot\left(\frac{dN_{k-1}^{\theta}+\sqrt{n_{k}-n_{k-1}}R_{5}\left(n_{k}-n_{k-1}\right)}{\left\lfloor N_{k}^{\theta}\right\rfloor }\right)^{d}\cdot\frac{CR\left(N\right)\cdot N_{k}^{d\theta}}{N_{k}^{\frac{d+1}{2}}}+N_{k}^{-\xi\left(1\right)}\\
& =\left(\frac{dN_{k-1}^{\theta}+\sqrt{N_{k}}R_{5}\left(N_{k}\right)}{\left\lfloor N_{k}^{\theta}\right\rfloor }\right)^{d}\cdot\frac{CR\left(N\right)\cdot N_{k}^{d\theta}}{N_{k}^{\frac{d+1}{2}}}+N_{k}^{-\xi\left(1\right)}\leq CR_{6}\left(N\right)N_{k}^{-\frac{1}{2}}.
\end{align*}
Combining all of the above we conclude that under the event $G\cap B_{N}$
(whose probability is $\geq1-N^{-\xi\left(1\right)}$) for every $k\geq1$
\[
\lambda_{k}\leq\lambda_{k-1}+CN_{k}^{1-\frac{4}{3}\theta-\frac{d}{2}}R_{5}^{d}\left(N_{k}\right)+CN_{k}^{-\frac{1}{2}}+CR\left(N\right)N_{k}^{-\frac{1}{2}}\leq\lambda_{k-1}+CN_{k}^{-\frac{1}{2}}R_{5}\left(N_{k}\right)\leq\lambda_{k-1}+CN_{k}^{-\frac{1}{3}}.
\]
Consequently
\begin{align*}
\lambda_{r\left(N\right)} & \leq\lambda_{1}+C\sum_{k=1}^{r\left(N\right)}N_{k}^{-\frac{1}{3}}=\lambda_{1}+C\sum_{k=1}^{r\left(N\right)}\frac{1}{\left\lfloor N^{\frac{1}{2^{k}}}\right\rfloor ^{\frac{1}{3}}}\\
& \leq\lambda_{1}+C\sum_{k=1}^{r\left(N\right)}N^{-\frac{1}{3\cdot2^{k}}}\leq\lambda_{1}+C\int_{1}^{r\left(N\right)+1}e^{-\frac{1}{3\cdot2^{t}}\cdot\log N}dt\\
& \overset{_{u=\frac{1}{3\cdot2^{t}}\cdot\log N}}{=}\lambda_{1}+C\int_{\alpha_{N}}^{\beta_{N}}\frac{e^{-u}}{-\ln3\cdot u}du,
\end{align*}
where $\alpha_{N}=\frac{1}{6}\cdot\log N$ and $\beta_{N}=\frac{1}{3\cdot2^{r\left(N\right)+1}}\cdot\log N$.
Since for large enough $N$ we have $\beta_{N}\geq\frac{\theta}{\sqrt[6]{M}}\log N\geq1$
we get
\[
\lambda_{k}\leq\lambda_{1}+C\int_{\alpha_{N}}^{\beta_{N}}-e^{-u}du=\lambda_{1}+C\left[e^{-u}\right]_{\alpha_{N}}^{\beta_{N}}\leq\lambda_{1}+Ce^{-\beta_{N}}=\lambda_{1}+\frac{C}{N^{\frac{1}{3\cdot2^{r\left(n\right)+1}}}}\leq\lambda_{1}+\frac{C}{M^{\frac{1}{6}}}.
\]
Finally, recalling the definition $\lambda_{1}$ and the fact that
$n_{0}\geq cN$ it follows from Proposition \ref{Prop:Lorentz_transformation-1},
that $\lambda_{1}\leq CN^{-\frac{1}{3}\theta}$, which completes the
proof.
\end{proof}
\section{Proof of Theorem \ref{thm:Absolutely_continuous_invariant_measure}}
In this section we prove our first main result, i.e., the existence
of a probability measure on the space of environments, which is equivalent
to the original i.i.d measure and is invariant with respect to the
point of view of the particle. The proof is divided into two parts.
In the first and main part of the proof the existence of an invariant
measure which is not singular with respect to the original i.i.d measure
is proved. In the second part we show that the existence of such a
measure guarantees the existence of an equivalent invariant measure.
In order to prove the existence of a non-singular invariant measure
we exploit the result from the last Section which allows us to construct
a coupling of the annealed and quenched law of the walk at time $N$
such that for most environments, i.e., with $P$ probability $\geq1-N^{-\xi\left(1\right)}$,
will keep them at distance at most $M$ one of the other with positive
probability independent of $N$. Using the uniform ellipticity, the
last coupling can be strengthen to guarantee the walks will coincide
at time $N$ with positive probability, which again is uniform in
$N$. Defining now, two random environments $\omega_{N},\omega_{N}^{'}$
which are the original environment shifted according to the location
of the annealed and quenched random walks at time $N$ respectively,
we get a coupling of the two such that $\omega_{N}=\omega_{N}^{'}$
with positive probability. Taking a Cesaro partial limit of the laws
of $\omega_{N}$ and $\omega_{N}^{'}$ we get two probability measures
on environments which are the original i.i.d measure and an invariant
measure with respect to the point of view of the particle respectively.
By taking the above coupling to the limit we can conclude that both
measure will give the same environment with positive probability,
and therefore in particular that they are not singular.
In the second part of the proof (see Lemma \ref{Lem:Either_Singular_or_absolutely_continuous})
we use general properties of any probability measure which is invariant
with respect to the point of view of the particle in order to show
that the existence of a non singular invariant probability measure
guarantees the existence of an equivalent invariant one.
In Subsection \ref{sub:Properties_of_the_Radon_Nikodym_derivative}
we discuss several properties of the Radon-Nikodym derivative of the
invariant measure with respect to the i.i.d measure. This includes
estimation on its average on a box as well as the existence of all
of its moments.
\subsection{\label{sub:Existence_of_an_equivalent_measure}Existence of an equivalent
measure}
\begin{lem}
\label{Lemma:there_exists_non_singular} Let $d\geq4$ and assume
$P$ is uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Then, there exists a measure $Q$ on on the space of environments
which is invariant with respect to the point of view of the particle
and is not singular with respect to the original i.i.d measure $P$.\end{lem}
\begin{proof}
Fix $0<{\varepsilon}<1$, a large $M\in\mathbb{N}$ and denote by $K\left(N\right)=K\left(N,M,\epsilon\right)$
the set of environments $\omega\in\Omega$ such that
\begin{equation}
\sum_{\Delta\in\Pi\left(M\right)}\left|P_{{\omega}}^{0}\left(X_{N}\in\Delta\right)-\mathbb{P}^{0}\left(X_{N}\in\Delta\right)\right|<\epsilon,\label{eq:Total_variation_box}
\end{equation}
where $\Pi\left(M\right)$ is a partition of $\mathbb{Z}^{d}$ into
$d$-dimensional boxes of side length $M$. By Theorem \ref{thm:Total_variation_on_small_boxes}
for every $\varepsilon>0$ there exists $M\in\mathbb{N}$ (independent
of $N$) such that $P\left(K\left(N\right)\right)\geq1-N^{-\xi\left(1\right)}$.
Equation (\ref{eq:Total_variation_box}) tells us that the total variation
distance of the respective distributions $\mathbb{P}^{0}\left(X_{N}\in\cdot\right)$
and $P_{{\omega}}^{0}\left(X_{N}\in\cdot\right)$ on $\Pi\left(M\right)$
is less than $\varepsilon$, and therefore there exists a coupling
$\tilde{\Theta}_{\omega,N,M}$ on $\Pi\left(M\right)\times\Pi\left(M\right)$
of both measures such that $\tilde{\Theta}_{\omega,N,M}\left(\Lambda_{\Pi}\right)>1-\varepsilon$,
where $\Lambda_{\Pi}=\left\{ \left(\Delta,\Delta\right)\,:\,\Delta\in\Pi\left(M\right)\right\} $.
Next, using the last coupling, we show how to construct a new coupling
of $\mathbb{P}^{0}\left(X_{N}=\cdot\right)$ and $P_{\omega}^{0}\left(X_{N}=\cdot\right)$
on $\mathbb{Z}^{d}\times\mathbb{Z}^{d}$ which gives a positive (independent
of $N$) probability to the event $\Lambda=\left\{ \left(x,x\right)\,:\, x\in\mathbb{Z}^{d}\right\} $.
Define $\Theta_{\omega,N}$ on $\mathbb{Z}^{d}\times\mathbb{Z}^{d}$
by
\[
\Theta_{\omega,N}\left(x,y\right):=\sum_{\Delta,\Delta'\in\Pi\left(M\right)}\tilde{\Theta}_{\omega,N-dM,M}\left(\Delta,\Delta'\right)\mathbb{P}^{0}\left(X_{N}=x\Big|X_{N-dM}\in\Delta\right)P_{\omega}^{0}\left(X_{N}=y\Big|X_{N-dM}\in\Delta'\right).
\]
Note that due to the law of total probability $\Theta_{\omega,N}$
is indeed a coupling of $\mathbb{P}^{0}\left(X_{N}=\cdot\right)$
and $P_{\omega}^{0}\left(X_{N}=\cdot\right)$.
For $x\in\mathbb{Z}^{d}$ let $\Delta_{x}$ be the unique cube that
contains $x$ in the partition $\Pi\left(M\right)$. Since the side
length of each box in the partition $\Pi\left(M\right)$ is $M$ it
follows that the the random walk can reach from each point in the
box $\Delta_{x}$ to $x$ in less than $dM$ steps. Recalling also
that the law of $P$ is uniformly elliptic with elliptic constant
$\eta$ (see (\ref{eq:elliptic_constant})) we conclude that
\begin{align*}
\Theta_{\omega,N}\left(x,x\right) & \geq\tilde{\Theta}_{\omega,N-dM,M}\left(\Delta_{x},\Delta_{x}\right)\mathbb{P}^{0}\left(X_{N}=x\Big|X_{N-dM}\in\Delta_{x}\right)P_{\omega}^{0}\left(X_{N}=x\Big|X_{N-dM}\in\Delta_{x}\right)\\
& \geq\tilde{\Theta}_{\omega,N-dM,M}\left(\Delta_{x},\Delta_{x}\right)\eta^{2dM}.
\end{align*}
Summing over $x$ we get
\[
\Theta_{\omega,N}\left(\Lambda\right)=\sum_{x\in\mathbb{Z}^{d}}\Theta_{\omega,N}\left(x,x\right)\geq\sum_{x\in\mathbb{Z}^{d}}\tilde{\Theta}_{\omega,N,M}\left(\Delta_{x},\Delta_{x}\right)\eta^{2dM}=\sum_{\Delta\in\Pi\left(M\right)}\tilde{\Theta}_{\omega,N,M}\left(\Delta,\Delta\right)M^{d}\eta^{2dM}>\left(1-\epsilon\right)M^{d}\eta^{2dM}.
\]
The last coupling allows us to construct for every $N$ two probability
measures on $\Omega$ that coincide with positive probability (independent
of $N$). Indeed, for $N\in\mathbb{N}$ let $Q_{N}$ and $P_{N}$
be defined by
\[
P_{N}\left(A\right)=E\left[\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{N}=x\right){\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]\quad\mbox{and}\quad Q_{N}\left(A\right)=E\left[\sum_{x\in\mathbb{Z}^{d}}P_{\omega}^{0}\left(X_{N}=x\right){\mathbbm{1}}_{\sigma_{x}\omega\in A}\right].
\]
Note that for every $N\in\mathbb{N}$ the measure $P_{N}$ is in fact
the i.i.d measure $P$ since the annealed walk is independent of the
environment distribution. Indeed, for every measurable event $A\subset\Omega$
\begin{align*}
P_{N}\left(A\right) & =E\left[\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{N}=x\right){\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]=\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{N}=x\right)E\left[{\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]\\
& =\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{N}=x\right)P\left(\sigma_{-x}A\right)=\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{T_{N}}=x\right)P\left(A\right)=P\left(A\right).
\end{align*}
Also note that using the coupling $\Theta_{\omega,N}$ we have for
every measurable event $A$
\begin{align*}
\left|Q_{N}\left(A\right)-P_{N}\left(A\right)\right| & =\left|E\left[\sum_{x\in\mathbb{Z}^{d}}\left[\mathbb{P}^{0}\left(X_{N}=x\right)-P_{\omega}^{0}\left(X_{N}=x\right)\right]{\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]\right|\\
& =\left|E\left[\sum_{x\in\mathbb{Z}^{d}}\left[\sum_{y\in\mathbb{Z}^{d}}\Theta_{\omega,N}\left(x,y\right)-\sum_{z\in\mathbb{Z}^{d}}\Theta_{\omega,N}\left(z,x\right)\right]{\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]\right|\\
& =\left|E\left[\sum_{x\in\mathbb{Z}^{d}}\left[\sum_{y\neq x}\Theta_{\omega,N}\left(x,y\right)-\sum_{z\neq x}\Theta_{\omega,N}\left(z,x\right)\right]{\mathbbm{1}}_{\sigma_{x}\omega\in A}\right]\right|\\
& \leq\max\left\{ \sum_{x\in\mathbb{Z}^{d}}\sum_{y\neq x}\Theta_{\omega,N}\left(x,y\right),\sum_{x\in\mathbb{Z}^{d}}\sum_{z\neq x}\Theta_{\omega,N}\left(z,x\right)\right\} <1-\left(1-\epsilon\right)M^{d}\eta^{2dM}.
\end{align*}
Let $\left\{ n_{k}\right\} $ be a subsequence such that the weak
limits of the Cesaro sequences $\left\{ \frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}Q_{N}\right\} _{k\geq1}$,
$\left\{ \frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}P_{N}\right\} _{k\geq1}$
and $\left\{ \frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\Theta_{\omega,N}\right\} _{k\geq1}$
exists. Since for every $N\in\mathbb{N}$ the measure $P_{N}$ equals
$P$ it follows that the limit of $\left\{ \frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}P_{N}\right\} _{k\geq1}$
is $P$ as well. Next, notice that the weak limit of $\left\{ \frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}Q_{N}\right\} _{k\geq1}$
which we denote by $Q$ is invariant with respect to the point of
view of the particle (see (\ref{eq:Transition kernel}) and (\ref{eq:invariant_probability_measure})
for the definition). Indeed, for every bounded continuous function
$f:\Omega\to\mathbb{R}$
\begin{align*}
\int_{\Omega}\mathfrak{R}f\left(\omega\right)dQ\left(\omega\right) & =\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\int_{\Omega}\mathfrak{R}f\left(\omega\right)dQ_{N}\left(\omega\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\int_{\Omega}\sum_{e\in\mathcal{E}_{d}}\omega\left(0,e\right)f\left(\sigma_{e}\omega\right)dQ_{N}\left(\omega\right)\\
& =\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\int_{\Omega}f\left(\omega\right)dQ_{N+1}\left(\omega\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=1}^{n_{k}}\int_{\Omega}f\left(\omega\right)dQ_{N}\left(\omega\right)=\int_{\Omega}f\left(\omega\right)dQ\left(\omega\right),
\end{align*}
where $\mathfrak{R}$ is as in (\ref{eq:Transition kernel}). Finally,
we show that $Q$ and $P$ are not singular. Using the coupling of
$P_{N}$ and $Q_{N}$, for every event $A\subset\Omega$ we have
\begin{align*}
\left|P\left(A\right)-Q\left(A\right)\right| & =\lim_{k\to\infty}\frac{1}{n_{k}}\left|\sum_{N=0}^{n_{k}-1}\left(P_{N}\left(A\right)-Q_{N}\left(A\right)\right)\right|\\
& \leq\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\left|P_{N}\left(A\right)-Q_{N}\left(A\right)\right|\leq1-\left(1-\epsilon\right)M^{d}\eta^{2dM}.
\end{align*}
Since this holds for all events, it follows that $\left\Vert P-Q\right\Vert _{TV}\leq1-\left(1-\epsilon\right)M^{d}\eta^{2dM}$
and thus $P$ and $Q$ are not singular. \end{proof}
\begin{lem}
\label{Lem:Either_Singular_or_absolutely_continuous} Assume $P$
is uniformly elliptic and i.i.d. If there exists a probability measure
$Q$ on the space of environments which is invariant with respect
to the point of view of the particle and is not singular with respect
to $P$, then there exists a probability measure $\tilde{Q}$ which
is invariant with respect to the point of view of the particle and
is also equivalent to $P$. \end{lem}
\begin{proof}
Denote by $Q=Q_{c}+Q_{s}$ the Lebesgue decomposition of $Q$ to an
absolutely continuous part $Q_{c}$ (w.r.t $P$) and a singular part
$Q_{s}$ (w.r.t $P$). Let $f=\frac{dQ_{c}}{dP}$ and define $A=\left\{ \omega\in\Omega\,:\, f\left(\omega\right)>0\right\} $.
From the invariance with respect to the point of view of the particle
and the uniform ellipticity we have
\[
Q=\sum_{e\in\mathcal{E}_{d}}\omega\left(e\right)\sigma_{e}\circ Q\geq\eta\sum_{e\in\mathcal{E}_{d}}\sigma_{e}\circ Q
\]
and therefore $\sigma_{e}\circ Q\ll Q$ for every $e\in\mathcal{E}_{d}$.
Since in addition we have $\left(\sigma_{e}\circ Q\right)_{s}=\sigma_{e}\circ Q_{s}$,
$\left(\sigma_{e}\circ Q\right)_{c}=\sigma_{e}\circ Q_{c}$ and $\frac{d\left(\sigma_{e}\circ Q\right)_{c}}{dP}\left(\cdot\right)=f\left(\sigma_{e}\left(\cdot\right)\right)$
we get that
\begin{equation}
Q_{c}=\sum_{e\in\mathcal{E}_{d}}\omega\left(e\right)\sigma_{e}\circ Q_{c}\geq\eta\sum_{e\in\mathcal{E}_{d}}\sigma_{e}\circ Q_{c}\label{eq:equivalent_measures}
\end{equation}
and thus
\[
f\left(\omega\right)\geq\eta\sum_{e\in\mathcal{E}_{d}}f\left(\sigma_{e}\left(\omega\right)\right).
\]
Consequently, $\omega\in A$ implies $\sigma_{e}\omega\in A$ for
every $e\in\mathcal{E}_{d}$, $P$ a.s.
In particular we get that $A$ is $\sigma_{e_{1}}$ invariant and
therefore by ergodicity that it is a $0-1$ event. This immediately
implies that if $Q$ is not singular with respect to $P$, i.e., $P\left(A\right)\neq0$,
then $P\left(A\right)=1$ and thus $P\ll Q_{c}$. Taking $\tilde{Q}=\frac{Q_{c}}{Q_{c}\left(\Omega\right)}$
we get that $\tilde{Q}$ is equivalent to the i.i.d measure and is
invariant with respect to the point of view of the particle (by the
first equality in (\ref{eq:equivalent_measures})). \end{proof}
\begin{rem}
\label{Rem:the_measure_is_in_fact_equivalent} Note that the sequence
of probability measures $\left\{ \sum_{N=0}^{n-1}Q_{N}\right\} _{n\geq1}$
equals to $\left\{ \sum_{N=0}^{n-1}\mathfrak{R}^{N}P\right\} _{n\geq1}$.
Recalling Theorem \ref{thm:Kozlov's_theorem} it follows that the
measure $\sum_{N=0}^{n-1}Q_{N}$ converges (without taking a subsequence)
to the equivalent measure $Q$ which is the \textbf{unique} probability
measure invariant with respect to the point of view of the particle.
In particular there is no need to restrict ourselves to the absolutely
continuous part as done in Lemma \ref{Lem:Either_Singular_or_absolutely_continuous}.
\end{rem}
\subsection{\label{sub:Properties_of_the_Radon_Nikodym_derivative}Some properties
of the Radon-Nikodym derivative}
In this subsection we discuss some properties of the equivalent probability
measure $Q$ and its Radon-Nikodym derivative. The next definition
will be useful in the statement of the lemmas:
\begin{defn}
Given two environments $\omega,\omega'\in\Omega$ define their distance
by
\[
\mbox{dist}\left(\omega,\omega'\right)=\min\left\{ \left\Vert x\right\Vert _{1}\,:\,\omega'=\sigma_{x}\omega\right\} ,
\]
where the minimum over an empty set is defined to be infinity.
\end{defn}
For future use we denote by $\Psi$ and $\Psi_{N}$ the couplings
of $P$ and $Q$ and of $P_{N}$ and $Q_{N}$ respectively on $\Omega\times\Omega$,
i.e.,
\begin{equation}
\Psi_{N}\left(A\right)=E\left[\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\omega,N}\left(x,y\right){\mathbbm{1}}_{\left(\sigma_{x}\omega,\sigma_{y}\omega\right)\in A}\right],\label{eq:Definition_of_Psi_N}
\end{equation}
and $\Psi$ is the weak limit of the Cesaro sequence $\left\{ \frac{1}{n}\sum_{N=0}^{n-1}\Psi_{N}\right\} _{n=1}^{\infty}$
along any converging sub-sequence which we denote from here on by
$\left\{ n_{k}\right\} _{k\geq1}$.
Our main goal is to prove the following concentration inequality for
the average of the Radon-Nikodym derivative on a box.
\begin{lem}
\label{lem:Radon_Nikodym_derivative_estimate} Let $M\in\mathbb{N}$
and denote by $\Delta_{0}$ a $d$-dimensional cube of side length
$M$ in $\mathbb{Z}^{d}$. Then for every $\varepsilon>0$,
\[
P\left(\left|\frac{1}{\left|\Delta_{0}\right|}\sum_{x\in\Delta_{0}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)-1\right|>\varepsilon\right)\leq M^{-\xi\left(1\right)}.
\]
\end{lem}
As a first step towards the proof of Lemma \ref{lem:Radon_Nikodym_derivative_estimate},
we prove the following:
\begin{lem}
\label{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}For
$M\in\mathbb{N}$ let $D_{M}^{\left(1\right)}:\Omega\to\left[0,\infty\right]$
and $D_{M}^{\left(2\right)}:\Omega\to\left[0,\infty\right]$ be defined
by
\[
D_{M}^{\left(1\right)}\left(\omega\right)=E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\bigg|\mathfrak{F}_{\omega_{1}}\right]\left(\omega\right)\quad\mbox{and}\quad D_{M}^{\left(2\right)}\left(\omega\right)=E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\bigg|\mathfrak{F}_{\omega_{2}}\right]\left(\omega\right),
\]
where $\mathfrak{F}_{\omega_{1}}$, $\mathfrak{F}_{\omega_{2}}$ are
the $\sigma$-algebras generated by the first respectively second
coordinate in $\Omega\times\Omega$ and $\Psi$ is as defined below
(\ref{eq:Definition_of_Psi_N}). For every $M\in\mathbb{N}$ there
exists an event $\mathbf{F}_{M}$ with the following properties:
\begin{enumerate}
\item $P\left(\mathbf{F}_{M}\right)=1-M^{-\xi\left(1\right)}$.
\item For every $\varepsilon>0$, if $M$ is large enough, then $D_{M}^{\left(1\right)}\left(\omega\right)\leq\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)$
and $\frac{dQ}{dP}\left(\omega\right)D_{M}^{\left(2\right)}\left(\omega\right)\leq\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)$
\end{enumerate}
\end{lem}
\begin{proof}
Let
\[
\mathbf{F}_{M}=\bigcap_{k=M}^{\infty}\left\{ \omega\in\Omega\,:\,\forall x\in\left[-k,k\right]^{d}\cap\mathbb{Z}^{d}\,,\,\sum_{\Delta\in\Pi_{M}}\left|\mathbb{P}^{x}\left(X_{k}\in\Delta\right)-P_{\omega}^{x}\left(X_{k}\in\Delta\right)\right|\leq\frac{C_{2}}{M^{c_{1}}}+\frac{C_{2}}{k^{c_{1}}}\right\} ,
\]
where $\Pi_{M}$ is a partition of $\mathbb{Z}^{d}$ into boxes of
side length $M$ and $0<c_{1},C_{2}<\infty$ are the constants from
Theorem \ref{thm:Total_variation_on_small_boxes}. Thus, by the same
theorem, we have $P\left(\mathbf{F}_{M}\right)=1-M^{-\xi\left(1\right)}$.
Fix some $\varepsilon>0$. The definition of $\mathbf{F}_{M}$ together
with the one of the couplings $\tilde{\Theta}_{\omega,k,M}$ constructed
in the proof of Lemma \ref{Lemma:there_exists_non_singular} implies
that for every $\omega\in\mathbf{F}_{M}$, every $k\geq M$ and every
$x\in\left[-k,k\right]^{d}\cap\mathbb{Z}^{d}$ we have $\tilde{\Theta}_{\sigma_{x}\omega,k,M}\left(\mathbf{\Lambda_{\Pi_{M}}}\right)>1-\frac{2C_{2}}{M^{c_{1}}}>1-\varepsilon$
for large enough $M$, where as before $\Lambda_{\Pi_{M}}=\left\{ \left(\Delta,\Delta\right)\,:\,\Delta\in\Pi_{M}\right\} $.
Before turning to prove the estimates for $\left\{ D_{M}^{\left(i\right)}\left(\omega\right)\right\} _{i\in\left\{ 1,2\right\} }$
we prove a similar results for the conditional expectations of $\Psi_{N}$.
For $N,M\in\mathbb{N}$ and $i\in\left\{ 1,2\right\} $ define $D_{M,N}^{\left(i\right)}:\Omega\to\left[0,\infty\right]$
by $D_{M,N}^{\left(i\right)}\left(\omega\right)=E_{\Psi_{N}}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\Bigg|\mathfrak{F}_{\omega_{i}}\right]\left(\omega\right)$.
Note that for $P$ almost every environment $\omega\in\Omega$ we
have
\begin{equation}
D_{M,N}^{\left(1\right)}\left(\omega\right)=\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\label{eq:definition_of_conditioned_probability_distance}
\end{equation}
and
\begin{equation}
D_{M,N}^{\left(2\right)}\left(\omega\right)=\left(\frac{dQ_{N}}{dP}\left(\omega\right)\right)^{-1}\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-y}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}.\label{eq:definition_of_conditioned_probability_distance_2}
\end{equation}
Indeed, using (\ref{eq:Definition_of_Psi_N}) we have for every measurable
event $A\subset\Omega$
\begin{align*}
E_{\Psi_{N}}\left[{\mathbbm{1}}_{A\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\right] & =\Psi_{N}\left(A\times\Omega\cap\left\{ \left(\omega_{1},\omega_{2}\right)\,:\,\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM\right\} \right)\\
& =E\left[\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\omega,N}\left(x,y\right){\mathbbm{1}}_{\left(\sigma_{x}\omega,\sigma_{y}\omega\right)\in A\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\sigma_{x}\omega,\sigma_{y}\omega\right)>dM}\right]\\
& =\sum_{x,y\in\mathbb{Z}^{d}}E\left[\Theta_{\omega,N}\left(x,y\right){\mathbbm{1}}_{\sigma_{x}\omega\in A}{\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\right],
\end{align*}
which by translation invariance of $P$ equals
\[
\sum_{x,y\in\mathbb{Z}^{d}}E\left[\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\omega\in A}{\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\right]=E\left[\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\omega\in A}{\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\right],
\]
and due to the fact that the first marginal of $\Psi_{N}$ is $P$
equals
\[
E_{\Psi_{N}}\left[{\mathbbm{1}}_{\left(\omega,\omega'\right)\in A\times\Omega}\cdot\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\right],
\]
Thus by the definition of conditional expectation (\ref{eq:definition_of_conditioned_probability_distance})
holds. A similar argument shows that
\begin{align*}
E_{\Psi_{N}}\left[{\mathbbm{1}}_{\Omega\times A}{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\right] & =E\left[\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega',N}\left(x,y\right){\mathbbm{1}}_{\omega'\in A}{\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\right]\\
& =E_{Q_{N}}\left[\left(\frac{dQ_{N}}{dP}\left(\omega'\right)\right)^{-1}\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega',N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}{\mathbbm{1}}_{\omega'\in A}\right]\\
& =E_{\Psi_{N}}\left[\left(\frac{dQ_{N}}{dP}\left(\omega_{2}\right)\right)^{-1}\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega_{2},N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\cdot{\mathbbm{1}}_{\Omega\times A}\left(\omega_{1},\omega_{2}\right)\right]
\end{align*}
and thus that (\ref{eq:definition_of_conditioned_probability_distance_2})
holds as well.
Since $\Theta_{\sigma_{-x}\omega,N}\left(x,y\right)>0$ implies $x\in\left[-N,N\right]^{d}\cap\mathbb{Z}^{d}$,
it follows that for large enough $M$, every $\omega\in\mathbf{F}_{M}$
and every $N\geq M$
\begin{align*}
\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM} & =1-\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}\leq dM}\\
& \leq1-\min_{z\in\left[-N,N\right]^{d}\cap\mathbb{Z}^{d}}\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-z}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}\leq dM}\\
& \leq1-\min_{z\in\left[-N,N\right]^{d}\cap\mathbb{Z}^{d}}\sum_{\Delta\in\Pi_{M}}\sum_{x,y\in\Delta}\Theta_{\sigma_{-z}\omega,N}\left(x,y\right)\\
& =1-\min_{z\in\left[-N,N\right]^{d}\cap\mathbb{Z}^{d}}\left(\sum_{\Delta\in\Pi_{M}}\tilde{\Theta}_{\sigma_{-z}\omega,N,M}\left(\Delta,\Delta\right)\right)\\
& =1-\min_{z\in\left[-N,N\right]^{d}\cap\mathbb{Z}^{d}}\tilde{\Theta}_{\sigma_{-z}\omega,N,M}\left(\Lambda_{\Pi_{M}}\right)<\varepsilon.
\end{align*}
Thus
\[
D_{M,N}^{\left(1\right)}\left(\omega\right)=\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-x}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\leq\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)
\]
and similarly,
\[
\frac{dQ_{N}}{dP}\left(\omega\right)D_{M,N}^{\left(2\right)}\left(\omega\right)=\sum_{x,y\in\mathbb{Z}^{d}}\Theta_{\sigma_{-y}\omega,N}\left(x,y\right){\mathbbm{1}}_{\left\Vert x-y\right\Vert _{1}>dM}\leq\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)
\]
Next we turn to prove the estimate for $\left\{ D_{M}^{\left(i\right)}\right\} _{i\in\left\{ 1,2\right\} }$.
It is enough to show that along some sub-sequence of $\left\{ n_{k}\right\} _{k\geq1}$
(which for simplicity we still denote by $\left\{ n_{k}\right\} _{k\geq1}$)
\begin{equation}
D_{M}^{\left(1\right)}\left(\omega\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}D_{M,N}^{\left(1\right)}\left(\omega\right)\quad\mbox{and}\quad D_{M}^{\left(2\right)}\left(\omega\right)=\left(\frac{dQ}{dP}\left(\omega\right)\right)^{-1}\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\frac{dQ_{N}}{dP}\left(\omega\right)D_{M,N}^{\left(2\right)}\left(\omega\right)\,,\, P\, a.s.\label{eq:Convergence_of_the_D_M}
\end{equation}
Indeed, if (\ref{eq:Convergence_of_the_D_M}) holds, then for $P$
almost every $\omega$ we have
\begin{align*}
D_{M}^{\left(1\right)}\left(\omega\right) & =\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}D_{M,N}^{\left(1\right)}\left(\omega\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\left[\sum_{N=0}^{M-1}D_{M,N}^{\left(1\right)}\left(\omega\right)+\sum_{N=M}^{n_{k}-1}D_{M,N}^{\left(1\right)}\left(\omega\right)\right]\\
& \leq\lim_{k\to\infty}\frac{1}{n_{k}}\left[M+\sum_{N=M}^{n_{k}-1}D_{M,N}^{\left(1\right)}\left(\omega\right)\right]\leq\lim_{k\to\infty}\frac{1}{n_{k}}\left[M+\sum_{N=M}^{n_{k}-1}\left(\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)\right)\right]\\
& =\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)
\end{align*}
and similarly
\begin{align*}
& \frac{dQ}{dP}\left(\omega\right)D_{M}^{\left(2\right)}\left(\omega\right)\\
= & \lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\frac{dQ_{N}}{dP}\left(\omega\right)D_{M,N}^{\left(2\right)}\left(\omega\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\left[\sum_{N=0}^{M-1}\frac{dQ_{N}}{dP}\left(\omega\right)D_{M,N}^{\left(2\right)}\left(\omega\right)+\sum_{N=M}^{n_{k}-1}\frac{dQ_{N}}{dP}\left(\omega\right)D_{M,N}^{\left(2\right)}\left(\omega\right)\right]\\
\leq & \lim_{k\to\infty}\frac{1}{n_{k}}\left[\sum_{N=0}^{M-1}\frac{dQ_{N}}{dP}\left(\omega\right)+\sum_{N=M}^{n_{k}-1}\frac{dQ_{N}}{dP}D_{M,N}^{\left(1\right)}\left(\omega\right)\right]\leq\lim_{k\to\infty}\frac{1}{n_{k}}\left[\sum_{N=0}^{M-1}\frac{dQ_{N}}{dP}\left(\omega\right)+\sum_{N=M}^{n_{k}-1}\left(\varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)\right)\right]\\
= & \varepsilon{\mathbbm{1}}_{\mathbf{F}_{M}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M}^{c}}\left(\omega\right)
\end{align*}
Turning to prove (\ref{eq:Convergence_of_the_D_M}), for every measurable
event $A\subset\Omega$ we have
\begin{align*}
& E\left[D_{M}^{\left(1\right)}\left(\omega\right){\mathbbm{1}}_{A}\left(\omega\right)\right]=E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\cdot{\mathbbm{1}}_{A\times\Omega}\left(\omega_{1},\omega_{2}\right)\right]\\
= & \Psi\left(\left\{ \left(\omega_{1},\omega_{2}\right)\,:\,\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM\right\} \cap A\times\Omega\right)\overset{\left(1\right)}{=}\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\Psi_{N}\left(\left\{ \left(\omega_{1},\omega_{2}\right)\,:\,\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM\right\} \cap A\times\Omega\right)\\
= & \lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{\Psi_{N}}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\cdot{\mathbbm{1}}_{A\times\Omega}\left(\omega_{1},\omega_{2}\right)\right]\overset{\left(2\right)}{=}\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{\Psi_{N}}\left[D_{M,N}^{\left(1\right)}\left(\omega_{1}\right)\cdot{\mathbbm{1}}_{A\times\Omega}\left(\omega_{1},\omega_{2}\right)\right]\\
= & \lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E\left[D_{M,N}^{\left(1\right)}\left(\omega_{1}\right)\cdot{\mathbbm{1}}_{A}\left(\omega_{1}\right)\right]=\lim_{k\to\infty}E\left[\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}D_{M,N}^{\left(1\right)}\left(\omega_{1}\right)\cdot{\mathbbm{1}}_{A}\left(\omega_{1}\right)\right]
\end{align*}
where $\left(1\right)$ is due to the definition of $\Psi$ below
(\ref{eq:Definition_of_Psi_N}) and $\left(2\right)$ uses the definition
of $D_{M,N}^{\left(1\right)}$ as the conditional expectation. This
implies that $\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}D_{M,N}^{\left(1\right)}$
converges in $L^{1}\left(P\right)$ to $D_{M}^{\left(1\right)}$ and
thus by standard arguments contains a subsequence that converges $P$
almost surely. Similraly, for $D_{M}^{\left(2\right)}$
\begin{align*}
& E_{Q}\left[D_{M}^{\left(2\right)}\left(\omega\right){\mathbbm{1}}_{A}\left(\omega\right)\right]=E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\cdot{\mathbbm{1}}_{\Omega\times A}\left(\omega_{1},\omega_{2}\right)\right]\\
= & \Psi\left(\left\{ \left(\omega_{1},\omega_{2}\right)\,:\,\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM\right\} \cap\Omega\times A\right)=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\Psi_{N}\left(\left\{ \left(\omega_{1},\omega_{2}\right)\,:\,\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM\right\} \cap\Omega\times A\right)\\
= & \lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{\Psi_{N}}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega_{1},\omega_{2}\right)>dM}\cdot{\mathbbm{1}}_{\Omega\times A}\left(\omega_{1},\omega_{2}\right)\right]=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{\Psi_{N}}\left[D_{M,N}^{\left(2\right)}\left(\omega_{2}\right)\cdot{\mathbbm{1}}_{\Omega\times A}\left(\omega_{1},\omega_{2}\right)\right]\\
= & \lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{Q_{N}}\left[D_{M,N}^{\left(2\right)}\left(\omega_{2}\right)\cdot{\mathbbm{1}}_{A}\left(\omega_{2}\right)\right]=\lim_{k\to\infty}\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}E_{Q}\left[\left(\frac{dQ}{dP}\left(\omega_{2}\right)\right)^{-1}\cdot\frac{dQ_{N}}{dP}\left(\omega_{2}\right)\cdot D_{M,N}^{\left(2\right)}\left(\omega_{2}\right)\cdot{\mathbbm{1}}_{A}\left(\omega_{2}\right)\right]\\
= & \lim_{k\to\infty}E_{Q}\left[\left(\frac{dQ}{dP}\left(\omega_{2}\right)\right)^{-1}\cdot\frac{1}{n_{k}}\sum_{N=0}^{n_{k}-1}\frac{dQ_{N}}{dP}\left(\omega_{2}\right)\cdot D_{M,N}^{\left(2\right)}\left(\omega_{2}\right)\cdot{\mathbbm{1}}_{A}\left(\omega_{2}\right)\right]
\end{align*}
proves the second quality in (\ref{eq:Convergence_of_the_D_M}), $Q$
(and thus $P$) almost surely for an appropriate sub-sequence.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:Radon_Nikodym_derivative_estimate}]
The proof deals separately with the events $B_{\varepsilon}^{-}=\left\{ \omega\in\Omega\,:\,\frac{1}{\left|\Delta_{0}\right|}\sum_{x\in\Delta_{0}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)<1-\varepsilon\right\} $
and $B_{\epsilon}^{+}=\left\{ \omega\in\Omega\,:\,\frac{1}{\left|\Delta_{0}\right|}\sum_{x\in\Delta_{0}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)>1+\varepsilon\right\} $.
We start with the event $B_{\varepsilon}^{-}$. The idea is to separate
the event $B_{\varepsilon}^{-}$ into two events the first with probability
$M^{-\xi\left(1\right)}$ and the second, denoted $S_{\varepsilon}^{-}$,
which will turn out to be of zero measure. To this end assume without
loss of generality that $\Delta_{0}$ is centered at the zero, denote
$M_{\varepsilon}=\frac{\epsilon}{6d^{2}}M$, define $\Delta_{0}^{-}=\left\{ x\in\mathbb{Z}^{d}\,:\,\left\Vert x\right\Vert _{\infty}<M-dM_{\varepsilon}\right\} $
and let
\[
S_{\varepsilon}^{-}=\left\{ \omega\in B_{\varepsilon}^{-}\,:\,\sigma_{x}\omega\in\mathbf{F}_{M_{\varepsilon}},\,\forall x\in\Delta_{0}\right\} ,
\]
where $\mathbf{F}_{M_{\varepsilon}}$ is the event from Lemma \ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}.
Due to property (1) of $\mathbf{F}_{M_{\varepsilon}}$ from Lemma
\ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}
\[
P\left(S_{\varepsilon}^{-}\right)\geq P\left(B_{\varepsilon}^{-}\right)-\left|\Delta_{0}\right|P\left(\mathbf{F}_{M_{\varepsilon}}^{c}\right)=P\left(B_{\varepsilon}^{-}\right)-M^{d}\cdot\left(M_{\varepsilon}\right)^{-\xi\left(1\right)}=P\left(B_{\varepsilon}^{-}\right)-M^{-\xi\left(1\right)},
\]
and therefore it is enough to show that $P\left(S_{\varepsilon}^{-}\right)=0$.
We claim that there exists an event $K\subset S_{\varepsilon}^{-}$
such that $\left(1\right)$ $P\left(K\right)\geq P\left(S_{\varepsilon}^{-}\right)\cdot\left(\left(4d\right)^{d}\left|\Delta_{0}\right|\right)^{-1}$
and $\left(2\right)$ if $\omega,\omega'\in K$ and $\omega\neq\omega'$,
then $\mbox{dist}\left(\omega,\omega'\right)>4dM$. Indeed, for every
$x\in\mathbb{Z}^{d}$ let $U_{x}$ be an independent (of everything
defined so far) random variable uniformly distributed on $\left[0,1\right]$,
and define%
\footnote{The event $K$ is not measurable in the $\sigma$-algebra of $\Omega$.
However, using Fubini's theorem we can find a section in $\Omega$
which is measurable having the desired properties.%
}
\[
K=\left\{ \omega\in S_{\varepsilon}^{-}\,:\,\forall x\in4d\Delta_{0}\,\mbox{if }\sigma_{x}\omega\in B_{\varepsilon}^{-}\,\mbox{then }U_{x}<U_{0}\right\} .
\]
Informally, from each family of environments whose distance is smaller
than $4dM$ we choose one uniformly. This immediately implies that
for two distinct points in $K$ property $(2)$ holds. Property $(1)$
on the other hand holds due to translation invariance of $P$.
Now, let
\[
H=\bigcup_{x\in\Delta_{0}}\sigma_{x}K\quad\mbox{and}\quad H^{-}=\bigcup_{x\in\Delta_{0}^{-}}\sigma_{x}K.
\]
By property (2) of $K$, in both cases this is a disjoint union, and
therefore, recalling once more the translation invariance of the measure
$P$, we have
\begin{align}
& P\left(H\right)=\left|\Delta_{0}\right|P\left(K\right)\nonumber \\
\mbox{and}\nonumber \\
& P\left(H^{-}\right)=\left|\Delta_{0}^{-}\right|P\left(K\right)=\left|\Delta_{0}\right|\left(1-\frac{\varepsilon}{6d^{2}}\right)^{d}P\left(K\right)>\left(1-\frac{\varepsilon}{6}\right)P\left(H\right).\label{eq:P_prob_estimation_for_H_and_H^-}
\end{align}
Going back to the definition of the event $B_{\epsilon}^{-}$ and
recalling that $K\subset S_{\varepsilon}^{-}\subset B_{\varepsilon}^{-}$
we get
\begin{align*}
Q\left(H\right) & =\int_{H}\frac{dQ}{dP}\left(\omega\right)dP\left(\omega\right)=\sum_{x\in\Delta_{0}}\int_{\sigma_{x}K}\frac{dQ}{dP}\left(\omega\right)dP\left(\omega\right)\\
& =\int_{K}\sum_{x\in\Delta_{0}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)dP\left(\omega\right)\\
& \leq\int_{K}\left(1-\epsilon\right)\left|\Delta_{0}\right|dP\left(\omega\right)=\left(1-\varepsilon\right)\left|\Delta_{0}\right|P\left(K\right)\\
& =\left(1-\varepsilon\right)P\left(H\right).
\end{align*}
Combining with (\ref{eq:P_prob_estimation_for_H_and_H^-}), for small
enough $\varepsilon>0$ this yields
\begin{equation}
Q\left(H\right)\leq\left(1-\varepsilon\right)P\left(H\right)=\frac{1-\varepsilon}{1-\frac{\varepsilon}{6}}\left(1-\frac{\varepsilon}{6}\right)P\left(H\right)<\frac{1-\varepsilon}{1-\frac{\varepsilon}{6}}P\left(H^{-}\right)<\left(1-\frac{\varepsilon}{3}\right)P\left(H^{-}\right).\label{eq:Q_prob_estimation_for_H_and_H^-}
\end{equation}
Let $A=\left\{ \left(\omega,\omega'\right)\,:\,\omega\in H^{-}\,,\,\omega'\notin H\right\} $.
Then by (\ref{eq:P_prob_estimation_for_H_and_H^-}) and (\ref{eq:Q_prob_estimation_for_H_and_H^-})
\begin{align}
\Psi\left(A\right) & \geq P\left(H^{-}\right)-Q\left(H\right)\geq P\left(H\right)-\left(1-\frac{\varepsilon}{3}\right)P\left(H^{-}\right)=\frac{\varepsilon}{3}P\left(H^{-}\right)>\frac{\varepsilon}{3}\left(1-\frac{\varepsilon}{6}\right)P\left(H\right)>\frac{\varepsilon}{4}P\left(H\right).\label{eq:Psi(A)_estimation}
\end{align}
By the construction of $K$, for every $\left(\omega,\omega'\right)\in A$
we have $\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}$
and therefore,
\begin{align}
\int_{H}D_{M_{\varepsilon}}^{\left(1\right)}\left(\omega\right)dP\left(\omega\right) & =\int_{H\times\Omega}D_{M_{\varepsilon}}^{\left(1\right)}\left(\omega\right)d\Psi\left(\omega,\omega'\right)\geq\int_{H^{-}\times\Omega}D_{M_{\varepsilon}}^{\left(1\right)}\left(\omega\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}\Big|\mathfrak{F}_{\omega}\right]\left(\omega\right){\mathbbm{1}}_{H^{-}\times\Omega}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{H^{-}\times\Omega}\left(\omega,\omega'\right)\Big|\mathfrak{F}_{\omega}\right]\left(\omega\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{H^{-}\times\Omega}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& \geq\int_{\Omega\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{A}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}{\mathbbm{1}}_{A}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\Psi\left(A\right)>\frac{\varepsilon}{4}P\left(H\right),\label{eq:first_int_D_M_H_estimation}
\end{align}
where for the last inequality we used (\ref{eq:Psi(A)_estimation}).
However, recalling that $H\subset\mathbf{F}_{M_{\varepsilon}}$ by
definition, and using Lemma \ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}
with $M_{\varepsilon}$ and $\frac{\varepsilon}{5}$ instead of $M$
and $\varepsilon$ we get
\begin{align}
\int_{H}D_{M_{\varepsilon}}^{\left(1\right)}\left(\omega\right)dP\left(\omega\right) & \leq\int_{H}\frac{\varepsilon}{5}{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}}\left(\omega\right)+{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}^{c}}\left(\omega\right)dP\left(\omega\right)\nonumber \\
& =\int_{H}\frac{\varepsilon}{5}dP\left(\omega\right)=\frac{\varepsilon}{5}P\left(H\right).\label{eq:second_int_D_M_H_estimation}
\end{align}
Combining (\ref{eq:first_int_D_M_H_estimation}) and (\ref{eq:second_int_D_M_H_estimation})
we must conclude that $P\left(H\right)=0$, and therefore $P\left(K\right)=0$.
This however, by property (1) of $K$, implies that $P\left(S_{\varepsilon}^{-}\right)=0$
and therefore finally that $P\left(B_{\varepsilon}^{-}\right)=M^{-\xi\left(1\right)}$.
Next we turn to deal with the event $B_{\varepsilon}^{+}$. As in
the proof for $B_{\varepsilon}^{-}$ for $\varepsilon>0$ we denote
$M_{\varepsilon}=\frac{\varepsilon}{6d^{2}}M$. Also assume without
loss of generality that $\Delta_{0}$ is centered in zero, define
$\Delta_{0}^{+}=\left\{ x\in\mathbb{Z}^{d}\,:\,\left\Vert x\right\Vert _{\infty}<M+dM_{\varepsilon}\right\} $
and let
\[
S_{\varepsilon}^{+}=\left\{ \omega\in B_{\varepsilon}^{+}\,:\,\sigma_{x}\omega\in\mathbf{F}_{M_{\varepsilon}},\,\forall x\in\Delta_{0}^{+}\right\} ,
\]
where $\mathbf{F}_{M_{\varepsilon}}$ is the event from Lemma \ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}.
Due to property (1) of $\mathbf{F}_{M_{\varepsilon}}$ from Lemma
\ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}
\[
P\left(S_{\varepsilon}^{+}\right)\geq P\left(B_{\varepsilon}^{+}\right)-\left|\Delta_{0}^{+}\right|P\left(\mathbf{F}_{M_{\varepsilon}}^{c}\right)=P\left(B_{\varepsilon}^{+}\right)-\left(1+\frac{\varepsilon}{6d}\right)^{d}M^{d}\cdot\left(M_{\varepsilon}\right)^{-\xi\left(1\right)}=P\left(B_{\varepsilon}^{+}\right)-M^{-\xi\left(1\right)}
\]
and therefore it is enough to show that $P\left(S_{\varepsilon}^{+}\right)=0$.
As for $S_{\varepsilon}^{-}$ we claim that there exists an event
$K\subset S_{\varepsilon}^{+}$ such that (1) $P\left(K\right)\geq P\left(S_{\varepsilon}^{+}\right)/\left(\left(4d\right)^{d}\left|\Delta_{0}^{+}\right|\right)^{-1}$
and (2) if $\omega,\omega'\in K$ and $\omega\neq\omega'$, then $\mbox{dist}\left(\omega,\omega'\right)>4d\left(M+M_{\varepsilon}\right)$.
Now, let
\[
H=\bigcup_{x\in\Delta_{0}}\sigma_{x}K\quad\mbox{and}\quad H^{+}=\bigcup_{x\in\Delta_{0}^{+}}\sigma_{x}K.
\]
By property (2) of $K$, in both cases this is a disjoint union, and
therefore, recalling once more the translation invariance of the measure
$P$, we have for small enough $\varepsilon>0$
\begin{align}
& P\left(H\right)=\left|\Delta_{0}\right|P\left(K\right)\nonumber \\
\mbox{and}\nonumber \\
& P\left(H^{+}\right)=\left|\Delta_{0}^{+}\right|P\left(K\right)=\left(1+\frac{\varepsilon}{6d^{2}}\right)^{d}\left|\Delta_{0}\right|P\left(K\right)<\left(1+\frac{\varepsilon}{6}\right)P\left(H\right).\label{eq:appendix_of_B_ep^+_1}
\end{align}
Going back to the definition of the event $B_{\varepsilon}^{+}$ and
recalling that $K\subset S_{\varepsilon}^{+}\subset B_{\varepsilon}^{+}$
we get
\begin{align}
Q\left(H\right) & =\int_{H}\frac{dQ}{dP}\left(\omega\right)dP\left(\omega\right)=\sum_{x\in\Delta_{0}}\int_{\sigma_{x}K}\frac{dQ}{dP}\left(\omega\right)dP\left(\omega\right)\nonumber \\
& =\int_{K}\sum_{x\in\Delta_{0}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)dP\left(\omega\right)\nonumber \\
& >\int_{K}\left(1+\varepsilon\right)\left|\Delta_{0}\right|dP\left(\omega\right)=\left(1+\varepsilon\right)\left|\Delta_{0}\right|P\left(K\right)\nonumber \\
& =\left(1+\varepsilon\right)P\left(H\right),\label{eq:appendix_of_B_ep^+_1.5}
\end{align}
and therefore, combining with (\ref{eq:appendix_of_B_ep^+_1}), for
small enough $\varepsilon>0$ this yields
\begin{equation}
Q\left(H\right)>\left(1+\varepsilon\right)P\left(H\right)=\frac{1+\varepsilon}{1+\frac{\varepsilon}{6}}\left(1+\frac{\varepsilon}{6}\right)P\left(H\right)>\frac{1+\varepsilon}{1+\frac{\varepsilon}{6}}P\left(H^{+}\right)>\left(1+\frac{\varepsilon}{3}\right)P\left(H^{+}\right).\label{eq:appendix_of_B_ep^+_2}
\end{equation}
Let $A=\left\{ \left(\omega,\omega'\right)\,:\,\omega\notin H^{+}\,,\,\omega'\in H\right\} $.
Then by (\ref{eq:appendix_of_B_ep^+_2})
\begin{equation}
\Psi\left(A\right)\geq Q\left(H\right)-P\left(H^{+}\right)>Q\left(H\right)-\frac{1}{1+\frac{\varepsilon}{3}}Q\left(H\right)=\frac{\frac{\varepsilon}{3}}{1+\frac{\varepsilon}{3}}Q\left(H\right)\geq\frac{\varepsilon}{4}Q\left(H\right)\label{eq:appendix_of_B_ep^+_3}
\end{equation}
By the construction of $K$, for every $\left(\omega,\omega'\right)\in A$
we have $\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}$
and therefore
\begin{align}
\int_{H}D_{M_{\varepsilon}}^{\left(2\right)}\left(\omega\right)dQ\left(\omega\right) & =\int_{\Omega\times H}D_{M_{\varepsilon}}^{\left(2\right)}\left(\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}\bigg|\mathfrak{F}_{\omega'}\right]\left(\omega'\right){\mathbbm{1}}_{\Omega\times H}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}E_{\Psi}\left[{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{\Omega\times H}\left(\omega,\omega'\right)\bigg|\mathfrak{F}_{\omega'}\right]\left(\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{\Omega\times H}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& \geq\int_{\Omega\times\Omega}{\mathbbm{1}}_{\mbox{dist}\left(\omega,\omega'\right)>dM_{\varepsilon}}{\mathbbm{1}}_{A}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\int_{\Omega\times\Omega}{\mathbbm{1}}_{A}\left(\omega,\omega'\right)d\Psi\left(\omega,\omega'\right)\nonumber \\
& =\Psi\left(A\right)\geq\frac{\varepsilon}{4}Q\left(H\right),\label{eq:appendix_of_B_ep^+_4}
\end{align}
where for the last inequality we used (\ref{eq:appendix_of_B_ep^+_3}).
However, since $H\subset\mathbf{F}_{M_{\varepsilon}}$ by definition,
and using Lemma \ref{lem:Perliminary_lemma_to_Radon_Nikodym_derivative_estimate}
with $M_{\varepsilon}$ and $\frac{\varepsilon}{5}$ instead of $M$
and $\varepsilon$ we get
\begin{align}
\int_{H}D_{M_{\varepsilon}}^{\left(2\right)}dQ\left(\omega\right) & \leq\int_{H}\frac{1}{f\left(\omega\right)}\left[\frac{\varepsilon}{5}{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}}+{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}^{c}}\right]dQ\left(\omega\right)=\int_{H}\left[\frac{\varepsilon}{5}{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}}+{\mathbbm{1}}_{\mathbf{F}_{M_{\varepsilon}}^{c}}\right]dP\left(\omega\right)\nonumber \\
& =\int_{H}\frac{\varepsilon}{5}dP\left(\omega\right)=\frac{\varepsilon}{5}P\left(H\right)\leq\frac{\varepsilon}{5}Q\left(H\right)\label{eq:appendix_of_B_ep^+_5}
\end{align}
Combining (\ref{eq:appendix_of_B_ep^+_4}) and (\ref{eq:appendix_of_B_ep^+_5})
we must conclude that $Q\left(H\right)=0$, therefore by (\ref{eq:appendix_of_B_ep^+_1.5})
that $P\left(H\right)=0$ and thus that $P\left(K\right)=0$. This
however, by property (1) of $K$, implies that $P\left(S_{\varepsilon}^{+}\right)=0$
and therefore finally that $P\left(B_{\varepsilon}^{+}\right)=M^{-\xi\left(1\right)}$.
\end{proof}
\begin{cor}
\label{cor:Existence_of_moments}Let $d\geq4$ and assume $P$ is
uniformly elliptic, i.i.d and satisfies $\left(\mathscr{P}\right)$.
Then $E\left[\left(\frac{dQ}{dP}\right)^{k}\right]<\infty$ for every
$k\in\mathbb{N}$.\end{cor}
\begin{proof}
For every $M\in\mathbb{N}$ large enough, Lemma \ref{lem:Radon_Nikodym_derivative_estimate}
implies
\[
P\left(\frac{dQ}{dP}\left(\omega\right)\geq2\left(2M+1\right)^{d}\right)\leq P\left(\frac{1}{\left(2M+1\right)^{d}}\sum_{x\in\left[-M.M\right]^{d}\cap\mathbb{Z}^{d}}\frac{dQ}{dP}\left(\sigma_{x}\omega\right)\geq2\right)=M^{-\xi\left(1\right)},
\]
Thus, $\frac{dQ}{dP}$ has super polynomial decay and the result follows.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:Absolutely_continuous_invariant_measure}]
The proof is the content of Lemma \ref{Lemma:there_exists_non_singular},
Lemma \ref{Lem:Either_Singular_or_absolutely_continuous} and Corollary
\ref{cor:Existence_of_moments}.
\end{proof}
\section{Proof of Theorem \ref{thm:Prefactor_thm}}
In this section we prove our second main result, the prefactor local
limit theorem. The uniqueness of the prefactor follows from its definition
quite easily and most of the work is concentrated into the proof of
existence. Our candidate for the prefactor is the Radon-Nikodym derivative
of the equivalent measure $Q$ constructed in the previous section.
The proof proceeds as follows: Instead of directly comparing the quenched
measure $P_{\omega}^{0}\left(X_{n}=\cdot\right)$ and the annealed
times the prefactor measure $\mathbb{P}^{0}\left(X_{n}=\cdot\right)f\left(\sigma_{\cdot}\omega\right)$,
we introduce two new measures (denoted temporarily by $\rho_{1}$
and $\rho_{2}$) and show that the total variation of the pairs $\left(\mathbb{P}^{0}\left(X_{n}=\cdot\right)f\left(\sigma_{\cdot}\omega\right),\rho_{1}\right)$,
$\left(\rho_{1},\rho_{2}\right)$, $\left(\rho_{2},P_{\omega}^{0}\left(X_{n}=\cdot\right)\right)$
goes to zero as $n$ goes to infinity for $P$ almost every environment.
Both measures $\rho_{1}$ and $\rho_{2}$ are constructed in a way
that allows us to exploit the previous results on the connection between
the the quenched and annealed measures in the total variations estimations.
More formally, we fix two parameters $0<\delta<\varepsilon<\frac{1}{2}$
and define the measures $\rho_{1}$ and $\rho_{2}$ as follows: For
$\rho_{1}$ we choose a point by first choosing a point according
to the annealed at time $n-n^{\varepsilon}$ times the prefactor and
then letting it ``evolve'' according to the quenched law for $n^{\varepsilon}$
steps. For $\rho_{2}$ we fix some partition of $\mathbb{Z}^{d}$
to boxes of side length $n^{\delta}$, choose a box according to the
quenched measure at time $n-n^{\varepsilon}$, choose a point inside
the box proportionally to its prefactor and then let it ``evolve''
into a new point according to the quenched law for $n^{\varepsilon}$
steps. For a more precise definition of the measures see Definition
\ref{Def:Lots_of_measures}.
\subsection{Uniqueness}
We start with a proof that the prefactor, if exists, is unique. Assume
both $f$ and $g$ satisfy (\ref{eq:Prefactor_thm}) and denote $h=f-g$.
By the triangle inequality for $P$ almost every $\omega\in\Omega$
\begin{equation}
\lim_{n\to\infty}\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=x\right)\left|h\left(\sigma_{x}\omega\right)\right|=0,\label{eq:pre_factor_uniqueness}
\end{equation}
i.e., $\lim_{n\to\infty}\mathbb{E}^{0}\left[\left|h\left(\sigma_{X_{n}}\omega\right)\right|\right]=0$,
$P$ a.s. If $h\neq0$ then there exists a measurable subset $A$
of $\Omega$ such that $P\left(A\right)>0$ and $\left|h\right|>c>0$
on $A$ . Thus for every $n\in\mathbb{N}$
\begin{align}
E\left[\mathbb{E}^{0}\left[\left|h\left(\sigma_{X_{n}}\omega\right)\right|\right]\right] & \geq E\left[\mathbb{E}^{0}\left[\left|h\left(\sigma_{X_{n}}\omega\right)\right|{\mathbbm{1}}_{\sigma_{X_{n}}\omega\in A}\right]\right]\geq cE\left[\mathbb{E}^{0}\left[{\mathbbm{1}}_{\sigma_{X_{n}}\omega\in A}\right]\right]\nonumber \\
& =cE\left[\mathbb{P}^{0}\left(\sigma_{X_{n}}\omega\in A\right)\right]=c\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=y\right)E\left[{\mathbbm{1}}_{\sigma_{y}\omega\in A}\right]\nonumber \\
& =c\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=y\right)P\left(\sigma_{y}\omega\in A\right)=cP\left(A\right)\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=y\right)\nonumber \\
& =cP\left(A\right)>0.\label{eq:uniquness_of_the_prefactor}
\end{align}
Since
\begin{align*}
E\left[\mathbb{E}^{0}\left[\left|h\left(\sigma_{X_{n}}\omega\right)\right|\right]\right] & =E\left[\sum_{y\in\mathbb{Z}^{d}}\left|h\left(\sigma_{y}\omega\right)\right|\mathbb{P}^{0}\left(X_{n}=y\right)\right]\\
& =\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=y\right)\cdot E\left[\left|h\left(\sigma_{y}\omega\right)\right|\right]\\
& =\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=y\right)\cdot E\left[\left|h\left(\omega\right)\right|\right]=E\left[\left|h\right|\right],
\end{align*}
the sequence $\left\{ \left|h\left(\sigma_{X_{n}}\omega\right)\right|\right\} _{n\in\mathbb{N}}$
is tight, and therefore by (\ref{eq:uniquness_of_the_prefactor})
it follows that $\lim_{n\to\infty}\mathbb{E}^{0}\left[\left|h\left(\sigma_{X_{n}}\omega\right)\right|\right]>0$
a contradiction to (\ref{eq:pre_factor_uniqueness}).
\subsection{Existence}
Let $f\left(\omega\right)$ be the Radon-Nikodym derivative of $Q$
(the equivalent measure w.r.t $P$, which is the invariant with respect
to the point of view of the particle). We will show that $f$ satisfies
Theorem \ref{thm:Prefactor_thm} starting with the following simple
proposition:
\begin{prop}
\label{prop:Invariance_of_the_prefactor}For $P$-almost every $\omega$
every $n\in\mathbb{N}$ and every $x\in\mathbb{Z}^{d}$
\[
f\left(\sigma_{x}\omega\right)=\sum_{y\in\mathbb{Z}^{d}}P_{\omega}^{y}\left(X_{n}=x\right)f\left(\sigma_{y}\omega\right).
\]
\end{prop}
\begin{proof}
For $n=1$ this follows from the definition of $f=\frac{dQ}{dP}$
as the Radon-Nikodym derivative of the measure $Q$ which is invariant
with respect to the point of view of the particle. Indeed, using (\ref{eq:Transition kernel})
and (\ref{eq:invariant_probability_measure}) and the translation
invariance of $P$ for every bounded measurable function $g:\Omega\to\mathbb{R}$
we have
\begin{align*}
\int_{\Omega}g\left(\omega\right)f\left(\omega\right)dP\left(\omega\right) & =\int_{\Omega}g\left(\omega\right)dQ\left(\omega\right)=\int_{\Omega}\mathfrak{R}g\left(\omega\right)dQ\left(\omega\right)\\
& =\int_{\Omega}\left(\mathfrak{R}g\left(\omega\right)\right)f\left(\omega\right)dP\left(\omega\right)=\int_{\Omega}\sum_{e\in\mathcal{E}_{d}}\omega\left(0,e\right)g\left(\sigma_{e}\omega\right)f\left(\omega\right)dP\left(\omega\right)\\
& =\int_{\Omega}\sum_{e\in\mathcal{E}_{d}}\omega\left(-e,0\right)g\left(\omega\right)f\left(\sigma_{-e}\omega\right)dP\left(\omega\right)\\
& =\int_{\Omega}g\left(\omega\right)\sum_{e\in\mathcal{E}_{d}}\omega\left(e,0\right)f\left(\sigma_{e}\omega\right)dP\left(\omega\right)
\end{align*}
and therefore
\[
f\left(\omega\right)=\sum_{e\in\mathcal{E}_{d}}\omega\left(e,0\right)f\left(\sigma_{e}\omega\right)=\sum_{e\in\mathcal{E}_{d}}P_{\omega}^{e}\left(X_{1}=0\right)f\left(\sigma_{e}\omega\right).
\]
Applying the last equality for $\sigma_{x}\omega$ gives the result
in the case $n=1$.
For $n>1$ the proof follows by induction. Indeed,
\begin{align*}
\sum_{y\in\mathbb{Z}^{d}}P_{\omega}^{y}\left(X_{n}=x\right)f\left(\sigma_{y}\omega\right) & =\sum_{y\in\mathbb{Z}^{d}}\sum_{z\in\mathbb{Z}^{d}}P_{\omega}^{y}\left(X_{n-1}=z\right)P_{\omega}^{z}\left(X_{1}=x\right)f\left(\sigma_{y}\omega\right)\\
& =\sum_{z\in\mathbb{Z}^{d}}P_{\omega}^{z}\left(X_{1}=x\right)\sum_{y\in\mathbb{Z}^{d}}P_{\omega}^{y}\left(X_{n-1}=z\right)f\left(\sigma_{y}\omega\right)\\
& \overset{\left(1\right)}{=}\sum_{z\in\mathbb{Z}^{d}}P_{\omega}^{z}\left(X_{1}=x\right)f\left(\sigma_{z}\omega\right)\overset{\left(2\right)}{=}f\left(\sigma_{x}\omega\right),
\end{align*}
where for $\left(1\right)$ we used the induction assumption and in
$\left(2\right)$ we used the case $n=1$.
\end{proof}
As stated at the beginning of the section the proof of Theorem \ref{thm:Prefactor_thm}
uses comparison with two additional probability measures which we
now define.
\begin{defn}
\label{Def:Lots_of_measures}For $n\in\mathbb{N}$ and $\omega\in\Omega$
define the following probability measures on $\mathbb{Z}^{d}$:
\begin{enumerate}
\item $\nu_{\omega}^{\mbox{ann}\times\mbox{pre},n}$ - the annealed at time
$n$ times the prefactor (normalized)
\[
\nu_{\omega}^{\mbox{ann}\times\mbox{pre},n}\left(x\right)=\frac{1}{Z_{\omega,n}}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right),
\]
where $Z_{\omega,n}=\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)$
is a normalizing constant. In Lemma \ref{lem:Evaluation_of_the_normalizing_constant}
we show that $\lim_{n\to\infty}Z_{\omega,n}=1$, $P$ almost surely.
\item $\nu_{\omega}^{\mbox{que},n}$ - the quenched measure at time $n$
\[
\nu_{\omega}^{\mbox{que},n}\left(x\right)=P_{\omega}^{0}\left(X_{n}=x\right).
\]
\item $\nu_{\omega}^{\mbox{box-que\ensuremath{\times}pre},n}=\nu_{\omega,\Pi}^{\mbox{box-que\ensuremath{\times}pre},n}$
- the quenched measure on boxes with a choice of a point in the box
proportional to the prefactor. Given a partition $\Pi$ of $\mathbb{Z}^{d}$
into boxes of side length $l$, we choose a box according to the quenched
measure at time $n$ and then choose a point inside of the box proportionally
to the value of the Radon-Nikodym derivative there.
\[
\nu_{\omega,\Pi}^{\mbox{box-que\ensuremath{\times}pre},n}\left(x\right)=\begin{cases}
P_{\omega}^{0}\left(X_{n}\in\Delta_{x}\right)\frac{f\left(\sigma_{x}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)} & \quad x\leftrightarrow n\\
0 & \quad\mbox{otherwise}
\end{cases},
\]
where $\Delta_{x}$ is the unique $d$-dimensional box that contains
$x$ in the partition $\Pi$.
\end{enumerate}
\end{defn}
Before turning to the proof of Theorem \ref{thm:Prefactor_thm} we
wish to study the normalization constant $Z_{\omega,n}$ of the measure
$\nu_{\omega}^{\mbox{ann}\times\mbox{pre},n}$
\begin{lem}
\label{lem:Evaluation_of_the_normalizing_constant}With the notations
as in Definition \ref{Def:Lots_of_measures} for $P$ almost every
$\omega$ we have $\lim_{n\to\infty}Z_{\omega,n}=1$.\end{lem}
\begin{proof}
Fix $\varepsilon>0$, $0<\delta<\frac{1}{6d}$ and let $\Pi$ be a
partition of $\mathbb{Z}^{d}$ into boxes of side length $n^{\delta}$.
If $x,y\in\Delta$ for some $\Delta\in\Pi$ then the annealed derivative
estimation, see Lemma \ref{lem:general_annealed_estimations}, gives
\begin{equation}
\left|\mathbb{P}^{0}\left(X_{n}=x\right)-\mathbb{P}^{0}\left(X_{n}=y\right)\right|\leq C\left\Vert x-y\right\Vert _{1}n^{-\frac{d+1}{2}}\leq Cn^{-\frac{d+1}{2}+\delta}.\label{eq:annealed_estimation for normalization constant}
\end{equation}
Denoting $\Pi_{n}=\left\{ \Delta\in\Pi\,:\,\Delta\cap\left[-n,n\right]^{d}\neq\emptyset\right\} $
we have
\[
\left|Z_{\omega,n}-1\right|=\left|\sum_{x\in\mathbb{Z}^{d}}\mathbb{P}\left(X_{n}=x\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|=\left|\sum_{\Delta\in\Pi_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|.
\]
By Lemma \ref{lem:good_estimation_for_the_location} There exists
$C_{\delta}>0$ such that $\mathbb{P}^{0}\left(\left\Vert X_{n}-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}>C_{\varepsilon}\sqrt{n}\right)<\varepsilon$.
Separating the sum into boxes in $\widehat{\Pi}_{n}=\left\{ \Delta\in\Pi_{n}\,:\,\Delta\cap\left\{ x\in\mathbb{Z}^{d}\,:\,\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert \leq C_{\varepsilon}\sqrt{n}\right\} \neq\emptyset\right\} $
and in $\Pi_{n}\backslash\widehat{\Pi}_{n}$ the last term is bounded
by
\begin{align}
& \leq\left|\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\label{eq:normalization_constant_lemma_0}\\
& +\left|\sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\left(\frac{1}{\left|\Delta\right|}\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\left[\mathbb{P}^{0}\left(X_{n}=y\right)-\mathbb{P}^{0}\left(X_{n}=x\right)\right]\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\label{eq:normalization_constant_lemma_1}\\
& +\left|\sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\frac{1}{\left|\Delta\right|}\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=y\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\label{eq:normalization_constnat_lemma_2}
\end{align}
We start by evaluating the term (\ref{eq:normalization_constant_lemma_0}).
By Lemma \ref{lem:Radon_Nikodym_derivative_estimate} there exists
some constant $C$, such that with probability $\geq1-n^{-\xi\left(1\right)}$
for every $\Delta\in\Pi_{n}$ (and in particular in $\Pi_{n}\backslash\widehat{\Pi}_{n}$)
we have $\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\left[f\left(\sigma_{y}\omega\right)+1\right]\leq C\left|\Delta\right|$. Therefore, under the above event, we can bound (\ref{eq:normalization_constant_lemma_0})
by
\begin{align*}
\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\left[f\left(\sigma_{x}\omega\right)+1\right] & \leq\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\max_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
x\leftrightarrow n
\end{array}}}\left[f\left(\sigma_{y}\omega\right)+1\right]\\
& \leq C\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\left|\Delta\right|\cdot\max_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right).
\end{align*}
Using Lemma \ref{lem:general_annealed_estimations}(\ref{eq:general_annealed_estimation_4})
and the definition of $\widehat{\Pi}_{n}$gives
\begin{align*}
\eqref{eq:normalization_constant_lemma_0} & \leq C\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\left|\Delta\right|\cdot\max_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\\
& \leq C\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\left[\max_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)-\mathbb{P}^{0}\left(X_{n}=y\right)\right]\\
& +C\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=y\right)\\
& \leq\frac{C}{n^{\frac{1}{2}-3d\delta}}+\varepsilon.
\end{align*}
Recalling that $\delta<\frac{1}{6d}$ and taking $n\to\infty$ this
gives (By an application of Borel Cantelli lemma)
\[
\limsup_{n\to\infty}\eqref{eq:normalization_constant_lemma_0}\leq\limsup_{n\to\infty}\sum_{\Delta\in\Pi_{n}\backslash\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\mathbb{P}^{0}\left(X_{n}=x\right)\left[f\left(\sigma_{x}\omega\right)+1\right]\leq\varepsilon,\quad P\, a.s.
\]
Next we deal with the term (\ref{eq:normalization_constant_lemma_1}).
Due to (\ref{eq:annealed_estimation for normalization constant}),
this is bounded by
\begin{align*}
& \sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\left|\frac{1}{\left|\Delta\right|}\sum_{\tiny{\begin{array}{c}
y\in\Delta\\
y\leftrightarrow n
\end{array}}}\left[\mathbb{P}\left(X_{n}=y\right)-\mathbb{P}\left(X_{n}=x\right)\right]\right|\left[f\left(\sigma_{x}\omega\right)+1\right]\\
\leq & \sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{\tiny{\begin{array}{c}
x\in\Delta\\
x\leftrightarrow n
\end{array}}}\frac{C}{n^{\frac{d+1}{2}-\delta}}\left[f\left(\sigma_{x}\omega\right)+1\right]\\
\leq\frac{C}{n^{\frac{1}{2}-\delta}} & \cdot\left(\frac{1}{n^{\frac{d}{2}}}\sum_{\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert \leq2C_{\varepsilon}\sqrt{n}}f\left(\sigma_{x}\omega\right)\right)+\frac{C}{n^{\frac{1}{2}-\delta}}.
\end{align*}
By Lemma \ref{lem:Radon_Nikodym_derivative_estimate} and an application
of Borel Cantelli for $P$ almost every $\omega$ and large enough
$n$ we have\\
$\frac{1}{n^{\frac{d}{2}}}\sum_{\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert \leq C_{\varepsilon}\sqrt{n}}f\left(\sigma_{x}\omega\right)\leq\left(4C_{\varepsilon}\right)^{d}$,
and thus the last term tends to zero as $n$ goes to infinity $P$
almost surely.
Finally for (\ref{eq:normalization_constnat_lemma_2}), we recall
that Lemma \ref{lem:general_annealed_estimations} also ensures $\mathbb{P}\left(X_{n}=x\right)\leq Cn^{-\frac{d}{2}}$
for every $x\in\mathbb{Z}^{d}$, and therefore
\begin{align*}
\eqref{eq:normalization_constnat_lemma_2} & =\left|\sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{x\in\Delta}\frac{1}{\left|\Delta\right|}\sum_{y\in\Delta}\mathbb{P}\left(X_{n}=y\right)\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\\
& \leq\sum_{\Delta\in\widehat{\Pi}_{n}}\frac{1}{\left|\Delta\right|}\sum_{y\in\Delta}\mathbb{P}\left(X_{n}=y\right)\left|\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\\
& \leq\frac{C}{n^{\frac{d}{2}}}\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\\
& =\frac{C}{n^{d\left(\frac{1}{2}-\delta\right)}}\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\frac{1}{\left|\Delta\right|}\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|
\end{align*}
Now, by Lemma \ref{lem:Radon_Nikodym_derivative_estimate}
\begin{align*}
& P\left(\frac{C}{n^{d\left(\frac{1}{2}-\delta\right)}}\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\frac{1}{\left|\Delta\right|}\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|>\varepsilon\right)\\
\leq & P\left(\exists\Delta\in\widehat{\Pi}_{n}\,:\,\left|\frac{1}{\left|\Delta\right|}\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|>\frac{\varepsilon}{C\cdot C_{\varepsilon}}\right)\\
\leq & n^{d\left(\frac{1}{2}-\delta\right)}P\left(\left|\frac{1}{\left|\Delta_{0}\right|}\sum_{x\in\Delta_{0}}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|>\frac{\varepsilon}{C\cdot C_{\varepsilon}}\right)=n^{d\left(\frac{1}{2}-\delta\right)}\cdot n^{-\xi\left(1\right)}=n^{-\xi\left(1\right)},
\end{align*}
where $\Delta_{0}$ is any choice for $\Delta_{0}\in\widehat{\Pi}_{n}$.
Therefore by Borell Cantelli we have
\[
\limsup_{n\to\infty}\frac{C}{n^{d\left(\frac{1}{2}-\delta\right)}}\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\frac{1}{\left|\Delta\right|}\sum_{x\in\Delta}\left[f\left(\sigma_{x}\omega\right)-1\right]\right|\leq\varepsilon.
\]
Combining all of the above we see that $P$ almost surely
\[
\limsup_{n\to\infty}\left|Z_{\omega,n}-1\right|\leq2\varepsilon.
\]
Since $\varepsilon>0$ was arbitrary, the result follows.
\end{proof}
Before turning to the main lemma in the proof of Theorem \ref{thm:Prefactor_thm}
we give two additional preliminary definitions needed in order to
construct the intermediate measures:
\begin{defn}
\label{Def:some_notations_for_measures_on_Z^d} Let $\nu_{\omega}^{1}$
and $\nu_{\omega}^{2}$ be two probability measures on $\mathbb{Z}^{d}$,
which may depend on $\omega\in\Omega$.
\begin{enumerate}
\item The $L^{1}$ distance of $\nu_{\omega}^{1}$ and $\nu_{\omega}^{2}$
is given by $\left\Vert \nu_{\omega}^{1}-\nu_{\omega}^{2}\right\Vert _{1}=\sum_{x\in\mathbb{Z}^{d}}\left|\nu_{\omega}^{1}\left(x\right)-\nu_{\omega}^{2}\left(x\right)\right|$
(note that this equals twice the total variation between $\nu_{\omega}^{1}$
and $\nu_{\omega}^{2}$).
\item The environment-convolution of $\nu_{\omega}^{1}$ and $\nu_{\omega}^{2}$
is a new probability measure on $\mathbb{Z}^{d}$, denoted $\left(\nu^{1}*\nu^{2}\right)_{\omega}$,
given by
\[
\left(\nu^{1}*\nu^{2}\right)_{\omega}\left(x\right)=\sum_{y\in\mathbb{Z}^{d}}\nu_{\omega}^{1}\left(y\right)\nu_{\sigma_{y}\omega}^{2}\left(x-y\right).
\]
\end{enumerate}
\end{defn}
We can now state the main Lemma in the proof of Theorem \ref{thm:Prefactor_thm}.
As already stated above, instead of comparing directly the $L^{1}$
distance of $\nu_{\omega}^{\mbox{ann}\times\mbox{pre},n}$ and $\nu_{\omega}^{\mbox{que},n}$,
i.e., the annealed times the prefactor and the quenched probability
measures, appearing in Theorem \ref{thm:Prefactor_thm}, we take a
more indirect approach and use two other measures as intermediaries.
This allows us to use previous results on the Radon-Nikodym derivative
and other relations between the quenched and annealed measures in
the evaluation of the $L^{1}$ distances.
\begin{lem}
\label{lem:Pre_factor_main_lemma} Fix $0<\delta<\varepsilon<\frac{1}{4}$,
and for $n\in\mathbb{N}$ abbreviate $k=\left\lceil n^{\epsilon}\right\rceil $
and $l=\left\lceil n^{\delta}\right\rceil $. Fix a partition $\Pi$
of $\mathbb{Z}^{d}$ into boxes of side length $l$. With the notations
as in Definitions \ref{Def:Lots_of_measures} and Definition \ref{Def:some_notations_for_measures_on_Z^d}
we have for $P$ almost every $\omega\in\Omega$
\begin{enumerate}
\item $\lim_{n\to\infty}\left\Vert \nu_{\omega}^{\mbox{ann}\times\mbox{pre},n}-\left(\nu^{\mbox{ann}\times\mbox{pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}\right\Vert _{1}=0.$
\item \textup{$\lim_{n\to\infty}\left\Vert \left(\nu^{\mbox{ann}\times\mbox{pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}-\left(\nu_{\Pi}^{\mbox{box-que\ensuremath{\times}pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}\right\Vert _{1}=0.$}
\item \textup{$\lim_{n\to\infty}\left\Vert \left(\nu_{\Pi}^{\mbox{box-que\ensuremath{\times}pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}-\left(\nu^{\mbox{que},n-k}*\nu^{\mbox{que},k}\right)_{\omega}\right\Vert _{1}=0.$}
\end{enumerate}
\end{lem}
\begin{rem}
$ $
\begin{enumerate}
\item In the temporary notation from the beginning of this Section we have
$\rho_{1}=\left(\nu^{\mbox{ann}\times\mbox{pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}$
and $\rho_{2}=\left(\nu^{\mbox{box-que\ensuremath{\times}pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}$.
\item Note that by the Markov property of the quenched walk $\nu_{\omega}^{\mbox{que},\cdot}$
we have $\left(\nu^{\mbox{que},n-k}*\nu^{\mbox{que},k}\right)_{\omega}=\nu_{\omega}^{\mbox{que},n}$.
\end{enumerate}
\end{rem}
\begin{proof}[Proof of Lemma \ref{lem:Pre_factor_main_lemma} part (1)]
We need to show that\textbf{
\[
\lim_{n\to\infty}\sum_{x\in\mathbb{Z}^{d}}\left|\frac{1}{Z_{n,\omega}}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-\frac{1}{Z_{n-k,\omega}}\sum_{y\in\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n-k}=y\right)f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|=0,
\]
}which by Lemma \ref{lem:Evaluation_of_the_normalizing_constant}
(and the fact that we can restrict attention to $x\in\mathbb{Z}^{d}$
such that $\left\Vert x\right\Vert _{1}\leq n$) is equivalent to
showing
\[
\lim_{n\to\infty}\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\left|\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n-k}=y\right)f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|=0.
\]
Denote $B_{n}=\left\{ x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\,:\,\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}\leq R_{5}\left(n\right)\sqrt{n}\right\} $.
By the triangle inequality
\begin{align}
& \sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\left|\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n-k}=y\right)f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|\nonumber \\
\leq & \sum_{x\in B_{n}}\left|\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\left[\mathbb{P}^{0}\left(X_{n}=x\right)-\mathbb{P}^{0}\left(X_{n-k}=y\right)\right]f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|\label{eq:Proof of_main_Lemma_part_1_1}\\
+ & \sum_{x\in B_{n}}\mathbb{P}^{0}\left(X_{n}=x\right)\left|f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|\label{eq:Proof of_main_Lemma_part_1_2}\\
+ & \sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\left|\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n-k}=y\right)f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|\label{eq:Proof of_main_Lemma_part_1_0}
\end{align}
Dealing with each of the terms separately (starting with (\ref{eq:Proof of_main_Lemma_part_1_1})),
by the annealed derivative estimation from Lemma \ref{lem:general_annealed_estimations}
\begin{align*}
\eqref{eq:Proof of_main_Lemma_part_1_1} & \leq\sum_{x\in B_{n}}\frac{Ck}{n^{\frac{d+1}{2}}}\left(\sum_{\mbox{dist}\left(y,B_{n}\right)\leq k}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right)\\
& \leq\frac{Ck}{n^{\frac{d+1}{2}}}\sum_{\mbox{dist}\left(y,B_{n}\right)\leq k}f\left(\sigma_{y}\omega\right)=\frac{Ck}{n^{\frac{1}{2}}}\cdot\frac{1}{n^{\frac{d}{2}}}\sum_{\mbox{dist}\left(y,B_{n}\right)\leq k}f\left(\sigma_{y}\omega\right).
\end{align*}
By Lemma \ref{lem:Radon_Nikodym_derivative_estimate} for $P$ almost
every $\omega$ and large enough $n$ we have $\frac{1}{n^{\frac{d}{2}}}\sum_{\mbox{dist}\left(y,B_{n}\right)\leq k}f\left(\sigma_{y}\omega\right)\leq2R_{6}\left(n\right)$.
Thus, using the fact that $k=n^{\varepsilon}\ll n^{\frac{1}{4}}$,
it follows that last term tends to zero $P$ almost surely as $n$
tends to $\infty$.
Turning to deal with (\ref{eq:Proof of_main_Lemma_part_1_2}), we
recall that by Proposition \ref{prop:Invariance_of_the_prefactor}
$f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)=0$
for every $x\in\mathbb{Z}^{d}$ such that $x+\left[-k,k\right]^{d}\cap\mathbb{Z}^{d}\subset\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}$.
In particular, denoting\\
$\tilde{B}_{n}=B_{n}\cap\left\{ x\in\mathbb{Z}^{d}\,:\, x+\left[-k,k\right]^{d}\cap\mathbb{Z}^{d}\subset\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\right\} $
and using the annealed estimations from Lemma \ref{lem:general_annealed_estimations}
\begin{align*}
\eqref{eq:Proof of_main_Lemma_part_1_2} & =\sum_{x\in\tilde{B}_{n}}\mathbb{P}^{0}\left(X_{n}=x\right)\left|f\left(\sigma_{x}\omega\right)-\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\right|\\
& \leq\sum_{x\in\tilde{B}_{n}}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)+\sum_{x\in\tilde{B}_{n}}\mathbb{P}^{0}\left(X_{n}=x\right)\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\\
& \leq\frac{C\left|\tilde{B}_{n}\right|}{n^{\frac{d}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}f\left(\sigma_{x}\omega\right)+\frac{C\left|\tilde{B}_{n}\right|}{n^{\frac{d}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}\sum_{\left\Vert y-x\right\Vert _{1}\leq k}f\left(\sigma_{y}\omega\right).
\end{align*}
where $\left|\tilde{B}_{n}\right|$ is the side of $\tilde{B}_{n}$.
By a worst case estimation we have $\left|\tilde{B}_{n}\right|\leq Ck\left(\sqrt{n}R_{5}\left(n\right)\right)^{d-1}$
and therefore
\begin{align*}
\eqref{eq:Proof of_main_Lemma_part_1_2} & =\frac{C\left|\tilde{B}_{n}\right|}{n^{\frac{d}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}f\left(\sigma_{x}\omega\right)+\frac{C\left|\tilde{B}_{n}\right|}{n^{\frac{d}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}\sum_{\left\Vert y-x\right\Vert _{1}\leq k}f\left(\sigma_{y}\omega\right).\\
& \leq\frac{CkR_{5}^{d-1}\left(n\right)}{n^{\frac{1}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}f\left(\sigma_{x}\omega\right)+\frac{CkR_{5}^{d-1}\left(n\right)}{n^{\frac{1}{2}}}\cdot\frac{1}{\left|\tilde{B}_{n}\right|}\sum_{x\in\tilde{B}_{n}}\sum_{\left\Vert y-x\right\Vert _{1}\leq k}f\left(\sigma_{y}\omega\right)
\end{align*}
Using again Lemma \ref{lem:Radon_Nikodym_derivative_estimate}, and
the choice $k=n^{\varepsilon}=o\left(n^{\frac{1}{4}}\right)$, it
follows that both sums tends to zero as $n$ goes to infinity, $P$
almost surely.
Finally we turn to deal with (\ref{eq:Proof of_main_Lemma_part_1_0}).
Using Lemma \ref{lem:good_estimation_for_the_location} we have $\mathbb{P}\left(X_{n}\notin B_{n}\right)=n^{-\xi\left(1\right)}$.
Recalling also that $k=\left\lceil n^{\varepsilon}\right\rceil \ll n$,
we note that if $P_{\omega}^{y}\left(X_{k}=x\right)>0$ and $x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}$
then $\left\Vert x-y\right\Vert _{1}\leq k$ and thus for large enough
$n$
\begin{align*}
\left\Vert y-\mathbb{E}^{0}\left[X_{n-k}\right]\right\Vert _{1} & \geq\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}-\left\Vert \mathbb{E}^{0}\left[X_{n}\right]-\mathbb{E}^{0}\left[X_{n-k}\right]\right\Vert _{1}-\left\Vert x-y\right\Vert _{1}\\
& \geq\sqrt{n}R_{5}\left(n\right)-2k\geq\frac{1}{2}\sqrt{n}R_{5}\left(n\right).
\end{align*}
This however, due to Lemma \ref{lem:good_estimation_for_the_location},
yields $\mathbb{P}^{0}\left(X_{n-k}=y\right)=n^{-\xi\left(1\right)}$,
and therefore
\begin{align*}
\eqref{eq:Proof of_main_Lemma_part_1_0} & \leq\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)+\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}\mathbb{P}^{0}\left(X_{n-k}=y\right)f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\\
& \leq n^{-\xi\left(1\right)}\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}f\left(\sigma_{x}\omega\right)+n^{-\xi\left(1\right)}\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\sum_{y\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{y}\omega\right)P_{\omega}^{y}\left(X_{k}=x\right)\\
& \leq2\cdot n^{-\xi\left(1\right)}\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{x}\omega\right).
\end{align*}
By Lemma \ref{lem:Radon_Nikodym_derivative_estimate} we have $P\left(\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}}f\left(\sigma_{x}\omega\right)\leq2n^{d}\right)>1-n^{-\xi\left(1\right)}$
and therefore by Borell-Cantelli lemma for large enough $n$
\[
\eqref{eq:Proof of_main_Lemma_part_1_0}\leq4n^{-\xi\left(1\right)}\cdot n^{d}=n^{-\xi\left(1\right)}\underset{_{n\to\infty}}{\longrightarrow}0.
\]
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:Pre_factor_main_lemma} part (2)]
Since
\[
\left\Vert \left(\nu^{\mbox{ann}\times\mbox{pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}-\left(\nu^{\mbox{box-que\ensuremath{\times}pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}\right\Vert _{1}\leq\left\Vert \nu_{\omega}^{\mbox{ann}\times\mbox{pre},n-k}-\nu_{\omega}^{\mbox{box-que\ensuremath{\times}pre},n-k}\right\Vert _{1},
\]
it is enough to deal with the right hand side and show that
\[
\lim_{n\to\infty}\sum_{\scriptsize{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n-k
\end{array}}}\left|\frac{1}{Z_{n-k,\omega}}\mathbb{P}^{0}\left(X_{n-k}=x\right)f\left(\sigma_{x}\omega\right)-\frac{f\left(\sigma_{x}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}P_{\omega}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|=0,\quad P\, a.s.
\]
Using Lemma \ref{lem:Evaluation_of_the_normalizing_constant} once
more , this is equivalent to showing
\begin{equation}
\lim_{n\to\infty}\sum_{\scriptsize{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|\mathbb{P}^{0}\left(X_{n-k}=x\right)-P_{\omega}^{0}\left(X_{n-k}\in\Delta_{x}\right)\frac{1}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\right|=0,\quad P\, a.s.\label{eq:Lemma_measures_part(2)_1}
\end{equation}
Denoting $B_{n}=\left\{ x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\,:\,\left\Vert x-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}\leq C_{\varepsilon}\sqrt{n}\right\} $
(with $C_{\varepsilon}$ such that $\mathbb{P}^{0}\left(\left\Vert X_{n}-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}>\frac{C_{\varepsilon}}{2}\sqrt{n}\right)<\varepsilon$
for large enough $n$) and using the triangle inequality the last
sum is bounded by
\begin{align}
& \sum_{\scriptsize{\begin{array}[t]{c}
x\in\left[-n,n\right]\cap\mathbb{Z}^{d}\backslash B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|\mathbb{P}^{0}\left(X_{n-k}=x\right)-\frac{1}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}P_{\omega}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|\label{eq:Lemma_measures_part(2)_0}\\
+ & \sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|\mathbb{P}^{0}\left(X_{n-k}=x\right)-\frac{2}{\left|\Delta_{x}\right|}\mathbb{P}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|\label{eq:Lemma_measures_part(2)_2}\\
+ & \sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|\frac{2}{\left|\Delta_{x}\right|}\mathbb{P}^{0}\left(X_{n-k}\in\Delta_{x}\right)-\frac{1}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\mathbb{P}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|\label{eq:Lemma_measures_part(2)_3}\\
+ & \sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|\frac{1}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\mathbb{P}^{0}\left(X_{n-k}\in\Delta_{x}\right)-\frac{1}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}P_{\omega}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|.\label{eq:Lemma_measures_part(2)_4}
\end{align}
Dealing with each of the terms separately, and starting with (\ref{eq:Lemma_measures_part(2)_0})
we have the following estimate:
\[
\eqref{eq:Lemma_measures_part(2)_0}\leq\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\mathbb{P}^{0}\left(X_{n-k}=x\right)f\left(\sigma_{x}\omega\right)+P_{\omega}^{0}\left(\left\Vert X_{n-k}-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}>C_{\varepsilon}\sqrt{n}\right).
\]
The term $\sum_{x\in\left[-n,n\right]^{d}\cap\mathbb{Z}^{d}\backslash B_{n}}\mathbb{P}^{0}\left(X_{n-k}=x\right)f\left(\sigma_{x}\omega\right)$
goes to zero as $n$ goes to infinity by the same argument used to
bound (\ref{eq:normalization_constant_lemma_0}) in Lemma \ref{lem:Evaluation_of_the_normalizing_constant}.
For the second term, Claim \ref{Clm:simple_claim_annealed_to_quenched}
implies that for a set of environments, with $P$ probability $>1-\sqrt{\varepsilon}$,
for large enough $n$
\[
P_{\omega}^{0}\left(\left\Vert X_{n-k}-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}>C_{\varepsilon}\sqrt{n}\right)\leq P_{\omega}^{0}\left(\left\Vert X_{n}-\mathbb{E}^{0}\left[X_{n}\right]\right\Vert _{1}>\frac{C_{\varepsilon}}{2}\sqrt{n}\right)\leq\sqrt{\varepsilon}.
\]
Since $\varepsilon>0$ was arbitrary this proves that the first term
goes to zero as $n$ goes to infinity.
Turning to (\ref{eq:Lemma_measures_part(2)_2}), the annealed derivative
estimations from Lemma \ref{lem:general_annealed_estimations}yields
\begin{align*}
\eqref{eq:Lemma_measures_part(2)_2} & \leq C\cdot\sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\frac{2}{\left|\Delta_{x}\right|}\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}\left|\mathbb{P}\left(X_{n-k}=x\right)-\mathbb{P}\left(X_{n-k}=y\right)\right|\\
& \leq C\cdot\sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\frac{2}{\left|\Delta_{x}\right|}\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}\frac{C}{\left(n-k\right)^{\frac{d+1}{2}}}\left\Vert x-y\right\Vert _{1}\\
& \overset{\left(1\right)}{\leq}C\cdot\sum_{x\in B_{n}}f\left(\sigma_{x}\omega\right)\frac{1}{\left|\Delta_{x}\right|}\sum_{y\in\Delta_{x}}\frac{C}{\left(n-k\right)^{\frac{d+1}{2}}}\cdot dn^{\delta}\\
& =\frac{C'\cdot n^{\frac{d}{2}+\delta}}{\left(n-k\right)^{\frac{d+1}{2}}}\cdot\left(\frac{1}{n^{\frac{d}{2}}}\sum_{x\in B_{n}}f\left(\sigma_{x}\omega\right)\right)\underset{_{n\to\infty}}{\longrightarrow}0,\quad P\, a.s.
\end{align*}
where for $\left(1\right)$ we used the fact that the side length
of every cube $\Delta$ is $n^{\delta}$ and for the limit we we used
Lemma \ref{lem:Radon_Nikodym_derivative_estimate}, the fact that
$k=n^{\epsilon}=o\left(n^{\frac{1}{4}}\right)$ and also that $\delta<\epsilon<\frac{1}{2}$.
Next we deal with (\ref{eq:Lemma_measures_part(2)_3}). Denoting $\widehat{\Pi}_{n}=\left\{ \Delta\in\Pi\,:\,\Delta\cap B_{n}\neq\emptyset\right\} $
and using the annealed derivative estimations from Lemma \ref{lem:Radon_Nikodym_derivative_estimate}
we conclude that
\begin{align*}
\eqref{eq:Lemma_measures_part(2)_3} & =\sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\frac{2}{\left|\Delta_{x}\right|}\mathbb{P}\left(X_{n-k}\in\Delta_{x}\right)\left|1-\frac{1}{\frac{2}{\left|\Delta_{x}\right|}\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\right|\\
& \leq\frac{C}{\left(n-k\right)^{\frac{d}{2}}}\sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)\left|1-\frac{1}{\frac{2}{\left|\Delta_{x}\right|}\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\right|\\
& \leq C\left(1-\frac{k}{n}\right)^{-\frac{d}{2}}\frac{1}{n^{\frac{d}{2}}}\sum_{\Delta\in\widehat{\Pi}_{n}}\sum_{x\in\Delta}f\left(\sigma_{x}\omega\right)\left|1-\frac{1}{\frac{2}{\left|\Delta\right|}\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\right|\\
& =C\left(1-\frac{k}{n}\right)^{-\frac{d}{2}}\frac{1}{n^{\frac{d}{2}\left(1-2\delta\right)}}\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\frac{1}{\left|\Delta\right|}\sum_{\scriptsize{\begin{array}[t]{c}
x\in\Delta\\
x\leftrightarrow n-k
\end{array}}}f\left(\sigma_{x}\omega\right)-1\right|.
\end{align*}
Using the same argument as the one at the end of the proof of Lemma
\ref{lem:Evaluation_of_the_normalizing_constant} (see the proof for
the bound on \ref{eq:normalization_constnat_lemma_2}) we get that
the last term goes to zero as $n$ goes to infinity $P$ a.s. Finally,
we estimate (\ref{eq:Lemma_measures_part(2)_4}).
\begin{align*}
\eqref{eq:Lemma_measures_part(2)_4} & \leq\sum_{\scriptsize{\begin{array}[t]{c}
x\in B_{n}\\
x\leftrightarrow n-k
\end{array}}}\frac{f\left(\sigma_{x}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
y\in\Delta_{x}\\
y\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{y}\omega\right)}\left|\mathbb{P}^{0}\left(X_{n-k}\in\Delta_{x}\right)-P_{\omega}^{0}\left(X_{n-k}\in\Delta_{x}\right)\right|\\
& =\sum_{\Delta\in\widehat{\Pi}_{n}}\left|\mathbb{P}^{0}\left(X_{n-k}\in\Delta\right)-P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\right|.
\end{align*}
The last term however is bounded by $Cn^{-\frac{1}{3}\delta}$ by
Proposition \ref{Prop:Lorentz_transformation-1} for $P$ almost every
$\omega$ and large enough $n$, and thus goes to zero as $n$ goes
to infinity.
\end{proof}
Part (3) of Lemma \ref{lem:Pre_factor_main_lemma} will follow from
the following more general lemma:
\begin{lem}
\label{lem:Lemma_for_part(3)} Let $x,y\in\mathbb{Z}^{d}$ satisfy
$\left\Vert x-y\right\Vert _{1}\leq n^{\theta}$ for some $\theta<\frac{1}{2}$.
Then the set of environments for which
\[
\left|P_{\omega}^{x}\left(X_{n}=z\right)-P_{\omega}^{y}\left(X_{n}=z\right)\right|=n^{-\xi\left(1\right)},\quad\forall z\in\mathbb{Z}^{d}
\]
has $P$ probability $\geq1-n^{-\xi\left(1\right)}$. \end{lem}
\begin{proof}
Fix $\theta<\theta'<1$ such that $\theta'<\frac{d+1}{2}\theta$,
$M\in\mathbb{N}$ and a partition $\Pi$ of $\mathbb{Z}^{d}$ into
boxes of side length $M$. By Theorem \ref{thm:Total_variation_on_small_boxes},
if $M$ is large enough, the set of environments $G\left(n,M\right)$
satisfying
\[
\sum_{\Delta\in\Pi}\left|P_{\omega}^{w}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)-\mathbb{P}^{w}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)\right|<\frac{1}{8}
\]
for every $w$ such that $\left\Vert w-x\right\Vert _{1}\leq n^{2}$
(and in particular for $w=y$) has probability $P\left(G\left(n,M\right)\right)\geq1-n^{-\xi\left(1\right)}$.
Consequently, using Lemma \ref{lem:general_annealed_estimations},
whenever $\left\Vert y-x\right\Vert _{1}\leq n^{\theta'}$, for large
enough $n$ we have
\begin{align*}
\left|P_{\omega}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)-P_{\omega}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)\right| & \leq\left|P_{\omega}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)-\mathbb{P}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)\right|+\left|\mathbb{P}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)-\mathbb{P}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)\right|\\
& +\left|P_{\omega}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)-\mathbb{P}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\Delta\right)\right|\\
& \leq\frac{1}{4}+\frac{Cn^{\theta'}}{n^{\frac{d+1}{2}\theta}}=\frac{1}{4}+Cn^{\theta'-\frac{d+1}{2}\theta}<\frac{1}{2}.
\end{align*}
Consequently, there exists a coupling of $P_{\omega}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\cdot\right)$
and $P_{\omega}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }\in\cdot\right)$
on $\Pi\times\Pi$ denoted $\tilde{\Xi}^{x,y}=\tilde{\Xi}_{n,\theta,\omega}^{x,y}$
such that $\tilde{\Xi}^{x,y}\left(\left\{ \left(\Delta,\Delta\right)\,:\,\Delta\in Q\right\} \right)>\frac{1}{2}$.
Using the uniform ellipticity, and the last coupling we can construct
a new coupling $\Xi_{1}^{x,y}$ of $P_{\omega}^{x}\left(X_{\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
and $P_{\omega}^{y}\left(X_{\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
on $\mathbb{Z}^{d}\times\mathbb{Z}^{d}$ such that $\Xi_{1}^{x,y}\left(\Lambda\right)\geq\frac{1}{2}\eta^{2dM}$,
where $\Lambda=\left\{ \left(z,z\right)\,:\, z\in\mathbb{Z}^{d}\right\} $
(for a more detailed explanation on the construction, see the proof
of Lemma \ref{Lemma:there_exists_non_singular}). Next, for $k\geq2$
we construct inductively a new coupling of $P_{\omega}^{x}\left(X_{k\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
and $P_{\omega}^{y}\left(X_{k\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
on $\mathbb{Z}^{d}\times\mathbb{Z}^{d}$ such that $\Xi_{\left\lceil \log^{2}n\right\rceil \cdot\left\lceil n^{\theta'}\right\rceil }^{x,y}\left(\Delta\right)=1-n^{-\xi\left(1\right)}$.
The construction goes as follows: First note that if $a,b\in\mathbb{Z}^{d}$
are any pair of points such that $\left\Vert a-x\right\Vert _{1},\left\Vert b-x\right\Vert _{1}\leq n^{2}$,
then, by the same reasoning, we can construct a coupling of $P_{\omega}^{a}\left(X_{\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
and $P_{\omega}^{b}\left(X_{\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
on $\mathbb{Z}^{d}\times\mathbb{Z}^{d}$, denoted $\Xi_{1}^{a,b}$,
such that $\Xi_{1}^{a,b}\left(\Lambda\right)>\frac{1}{2}\eta^{2dM}$.
Next, assuming the coupling $\Xi_{k-1}^{x,y}$ was constructed we
define $\Xi_{k}^{x,y}$ via the following procedure: Choose a pair
of points $\left(a,b\right)$ according to the previous coupling $\Xi_{k-1}^{x,y}$.
If $a=b$, couple the random walks together (to do exactly the same
path) for additional $\left\lceil n^{\theta}\right\rceil $ steps.
If $a\neq b$ and $\left\Vert a-b\right\Vert \leq n^{\theta'}$, couple
the random walks using the coupling $\Xi_{1}^{a,b}$. Finally, if
$\left\Vert a-b\right\Vert >n^{\theta'}$ we let the random walks
evolve independently. Formally, this is given by
\begin{align*}
\Xi_{k}^{x,y}\left(w_{1},w_{2}\right) & =\sum_{a,b\in\mathbb{Z}^{d}}\Xi_{k-1}^{x,y}\left(a,b\right)\Biggl[{\mathbbm{1}}_{a=b}{\mathbbm{1}}_{w_{1}=w_{2}}P_{\omega}^{a}\left(X_{\left\lceil n^{\theta}\right\rceil }=w_{1}\right)+{\mathbbm{1}}_{0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}}\cdot\Xi_{1}^{a,b}\left(w_{1},w_{2}\right)\\
& +{\mathbbm{1}}_{\left\Vert a-b\right\Vert _{1}>n^{\theta'}}\cdot P_{\omega}^{a}\left(X_{n^{\theta}}=w_{1}\right)P_{\omega}^{b}\left(X_{n^{\theta}}=w_{2}\right)\Biggr].
\end{align*}
It is not hard to verify that this indeed defines a coupling of $P_{\omega}^{x}\left(X_{k\left\lceil n^{\theta}\right\rceil }=\cdot\right)$
and $P_{\omega}^{y}\left(X_{k\left\lceil n^{\theta}\right\rceil }=\cdot\right)$,
and that in fact by the definition of $\theta'$ and the assumption
$\left\Vert x-y\right\Vert _{1}\leq n^{\theta}$ that $\Xi_{k}^{x,y}\left(a,b\right)=0$
whenever $\left\Vert a-b\right\Vert _{1}\geq n^{\theta'}$ and $n$
is large enough. Therefore,
\begin{align*}
\Xi_{k}^{x,y}\left(\Lambda\right) & =\sum_{w\in\mathbb{Z}^{d}}\sum_{a,b\in\mathbb{Z}^{d}}\Xi_{k-1}^{x,y}\left(a,b\right)\Biggl[{\mathbbm{1}}_{a=b}{\mathbbm{1}}_{w=w}P_{\omega}^{a}\left(X_{\left\lceil n^{\theta}\right\rceil }=w\right)+{\mathbbm{1}}_{0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}}\cdot\Xi_{1}^{a,b}\left(w,w\right)\\
& +{\mathbbm{1}}_{\left\Vert a-b\right\Vert >n^{\theta'}}\cdot P_{\omega}^{a}\left(X_{\left\lceil n^{\theta}\right\rceil }=w\right)P_{\omega}^{b}\left(X_{\left\lceil n^{\theta}\right\rceil }=w\right)\Biggr]\\
& \geq\sum_{w\in\mathbb{Z}^{d}}\sum_{a,b\in\mathbb{Z}^{d}}\Xi_{k-1}^{x,y}\left(a,b\right)\left[{\mathbbm{1}}_{a=b}P_{\omega}^{a}\left(X_{\left\lceil n^{\theta}\right\rceil }=w\right)+{\mathbbm{1}}_{0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}}\cdot\Xi_{1}^{a,b}\left(w,w\right)\right]\\
& =\Xi_{k-1}^{x,y}\left(\Lambda\right)+\sum_{a,b\in\mathbb{Z}^{d}}\Xi_{k-1}^{x,y}\left(a,b\right){\mathbbm{1}}_{0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}}\cdot\Xi_{1}^{a,b}\left(\Delta\right)\\
& \geq\Xi_{k-1}^{x,y}\left(\Lambda\right)+\sum_{a,b\in\mathbb{Z}^{d}}\Xi_{k-1}^{x,y}\left(a,b\right){\mathbbm{1}}_{0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}}\cdot\frac{1}{2}\eta^{2dM}\\
& =\Xi_{k-1}^{x,y}\left(\Lambda\right)+\frac{1}{2}\eta^{2dM}\Xi_{k-1}^{x,y}\left(\left\{ \left(a,b\right)\,:\,0<\left\Vert a-b\right\Vert _{1}\leq n^{\theta'}\right\} \right)\\
& =\Xi_{k-1}^{x,y}\left(\Lambda\right)+\frac{1}{2}\eta^{2dM}\left(1-\Xi_{k-1}^{x,y}\left(\Lambda\right)-\Xi_{k-1}^{x,y}\left(\left\{ \left(a,b\right)\,:\,\left\Vert a-b\right\Vert _{1}>n^{\theta'}\right\} \right)\right)\\
& \geq\Xi_{k-1}^{x,y}\left(\Lambda\right)+\frac{1}{2}\eta^{2dM}\left(1-\Xi_{k-1}^{x,y}\left(\Lambda\right)\right).
\end{align*}
Fixing $r>0$, as long as $\Xi_{j}^{x,y}\left(\Lambda\right)<1-n^{-r}$
for $j\leq k$ the last inequality gives
\[
\frac{\Xi_{k}^{x,y}\left(\Lambda\right)}{\Xi_{k-1}^{x,y}\left(\Lambda\right)}\geq1+\frac{1}{2}\eta^{2dM}\cdot\frac{n^{-r}}{1-n^{-r}}.
\]
which implies that $\Xi_{k}^{x,y}\left(\Lambda\right)$ grows exponentially
in this regime. Hence for some $C=C\left(r\right)<\infty$ we have
$\Xi_{\left\lceil C\log n\right\rceil }^{x,y}\left(\Lambda\right)>1-n^{-r}$
and in particular since $\left\{ \Xi_{k}^{x,y}\left(\Lambda\right)\right\} _{k\geq1}$
is non decreasing $\Xi_{\left\lceil \log^{2}n\right\rceil }^{x,y}\left(\Lambda\right)>\Xi_{\left\lceil C\log n\right\rceil }^{x,y}\left(\Lambda\right)>1-n^{-r}$
for every $r\in\mathbb{N}$ and large enough $n$, i.e., $\Xi_{\left\lceil \log^{2}n\right\rceil }^{x,y}\left(\Lambda\right)=1-n^{-\xi\left(1\right)}$.
We can now construct a coupling of $P_{\omega}^{x}\left(X_{n}=\cdot\right)$
and $P_{\omega}^{y}\left(X_{n}=\cdot\right)$ on $\mathbb{Z}^{d}$,
by using the coupling $\Xi_{\left\lceil \log^{2}n\right\rceil }^{x,y}$
until time $\left\lceil \log^{2}n\right\rceil \cdot\left\lceil n^{\theta}\right\rceil $.
Formally if they coincided until time $\left\lceil \log^{2}n\right\rceil \cdot\left\lceil n^{\theta}\right\rceil $
we couple them together (to go along the same path) until time $n$,
or if not to move independently until time $n$. This yields a coupling
such that $\sum_{z\in\mathbb{Z}}\left|P_{\omega}^{x}\left(X_{n}=z\right)-P_{\omega}^{y}\left(X_{n}=z\right)\right|<n^{-\xi\left(1\right)}$
as required.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:Pre_factor_main_lemma} part (3)]
Written explicitly (after some manipulations)
\begin{align}
& \left\Vert \left(\nu^{\mbox{box-que\ensuremath{\times}pre},n-k}*\nu^{\mbox{que},k}\right)_{\omega}-\left(\nu^{\mbox{que},n-k}*\nu^{\mbox{que},k}\right)_{\omega}\right\Vert \nonumber \\
= & \sum_{\scriptsize{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n
\end{array}}}\left|\sum_{\Delta\in\Pi}\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)P_{\omega}^{y}\left(X_{k}=x\right)\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\scriptsize{z\in\Delta}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]\right|\nonumber \\
\leq & \sum_{\scriptsize{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n
\end{array}}}\sum_{\Delta\in\Pi}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\left|\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}P_{\omega}^{y}\left(X_{k}=x\right)\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\scriptsize{z\in\Delta}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]\right|.\label{eq:Lemma_measures_part(3)_1}
\end{align}
Notice that for every $\Delta\in\Pi$ and $x\in\mathbb{Z}^{d}$ we
have
\[
\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}\left(\frac{1}{\left|\Delta\right|}\sum_{w\in\Delta}P_{\omega}^{w}\left(X_{k}=x\right)\right)\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\scriptsize{z\in\Delta}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]=0,
\]
and therefore (\ref{eq:Lemma_measures_part(3)_1}) equals{\footnotesize
\begin{align*}
& \sum_{\tiny{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n
\end{array}}}\sum_{\Delta\in\Pi}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\left|\!\!\!\sum_{\tiny{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}\!\!\!\left[P_{\omega}^{y}\left(X_{k}=x\right)-\left(\frac{2}{\left|\Delta\right|}\!\!\!\sum_{\tiny{\begin{array}[t]{c}
w\in\Delta\\
w\leftrightarrow n
\end{array}}}\!\!\! P_{\omega}^{w}\left(X_{k}=x\right)\right)\right]\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
z\in\Delta\\
z\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]\right|\\
= & \sum_{\tiny{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n
\end{array}}}\sum_{\Delta\in\Pi}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\left|\!\!\!\sum_{\tiny{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}\!\!\!\left[\frac{2}{\left|\Delta\right|}\!\!\!\sum_{\tiny{\begin{array}[t]{c}
w\in\Delta\\
w\leftrightarrow n
\end{array}}}\!\!\!\left(P_{\omega}^{y}\left(X_{k}=x\right)-P_{\omega}^{w}\left(X_{k}=x\right)\right)\right]\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
z\in\Delta\\
z\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]\right|\\
\leq & \sum_{\tiny{\begin{array}[t]{c}
x\in\mathbb{Z}^{d}\\
x\leftrightarrow n
\end{array}}}\sum_{\Delta\in\Pi}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\!\!\!\sum_{\tiny{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}\!\!\!\left|\frac{2}{\left|\Delta\right|}\!\!\!\sum_{\tiny{\begin{array}[t]{c}
w\in\Delta\\
w\leftrightarrow n
\end{array}}}\!\!\!\left(P_{\omega}^{y}\left(X_{k}=x\right)-P_{\omega}^{w}\left(X_{k}=x\right)\right)\right|\left|\frac{f\left(\sigma_{y}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
z\in\Delta\\
z\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right|\\
= & \sum_{\Delta\in\Pi}\sum_{\tiny{\begin{array}{c}
x\in\mathbb{Z}^{d}\\
\mbox{dist}\left(x,\Delta\right)\leq k\\
x\leftrightarrow n
\end{array}}}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\!\!\!\sum_{\tiny{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}\!\!\!\left|\frac{2}{\left|\Delta\right|}\!\!\!\sum_{\tiny{\begin{array}[t]{c}
w\in\Delta\\
w\leftrightarrow n
\end{array}}}\!\!\!\left(P_{\omega}^{y}\left(X_{k}=x\right)-P_{\omega}^{w}\left(X_{k}=x\right)\right)\right|\left|\frac{f\left(\sigma_{y}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
z\in\Delta\\
z\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right|.
\end{align*}
}By Lemma \ref{lem:Lemma_for_part(3)} applied with $k$ as $n$
and $\theta=\frac{\delta}{\epsilon}$ this is bounded by
\begin{align*}
& \sum_{\Delta\in\Pi}\sum_{\tiny{\begin{array}{c}
x\in\mathbb{Z}^{d}\\
\mbox{dist}\left(x,\Delta\right)\leq k\\
x\leftrightarrow n
\end{array}}}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)\sum_{\scriptsize{\begin{array}[t]{c}
y\in\Delta\\
y\leftrightarrow n-k
\end{array}}}k^{-\xi\left(1\right)}\left|\left[\frac{f\left(\sigma_{y}\omega\right)}{\underset{\tiny{\begin{array}[t]{c}
z\in\Delta\\
z\leftrightarrow n-k
\end{array}}}{\sum}f\left(\sigma_{z}\omega\right)}-P_{\omega}^{0}\left(X_{n-k}=y\Big|X_{n-k}\in\Delta\right)\right]\right|\\
\leq & \sum_{\Delta\in\Pi}\sum_{\tiny{\begin{array}{c}
x\in\mathbb{Z}^{d}\\
\mbox{dist}\left(x,\Delta\right)\leq k
\end{array}}}P_{\omega}^{0}\left(X_{n-k}\in\Delta\right)2k^{-\xi\left(1\right)}\\
\leq & 2k^{-\xi\left(1\right)}\cdot\left(2k+l\right)^{d}\underset{_{n\to\infty}}{\longrightarrow}0.
\end{align*}
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:Prefactor_thm}]
Combining all the claims of Lemma \ref{lem:Pre_factor_main_lemma}
and using the triangle inequality we conclude that
\[
\lim_{n\to\infty}\sum_{x\in\mathbb{Z}^{d}}\left|\frac{1}{Z}_{\omega,n}\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-\sum_{y\in\mathbb{Z}^{d}}P_{\omega}^{0}\left(X_{n-k}=y\right)P_{\sigma_{y}\omega}^{0}\left(X_{k}=x-y\right)\right|=0.
\]
Using the Markov property of the quenched law and Lemma \ref{lem:Evaluation_of_the_normalizing_constant},
this implies
\[
\lim_{n\to\infty}\sum_{x\in\mathbb{Z}^{d}}\left|\mathbb{P}^{0}\left(X_{n}=x\right)f\left(\sigma_{x}\omega\right)-P_{\omega}^{0}\left(X_{n}=x\right)\right|=0,
\]
and completes the proof. \end{proof}
\begin{acknowledgement}
The authors would like to thank Marek Biskup, Jean-Dominique Deuschel,
Tal Orenshtein, Jon Peterson, Pierre-François Rodriguez, Atilla Yilmaz
and Ofer Zeitouni for useful discussions.
\end{acknowledgement}
\bibliographystyle{alpha}
|
\section{Structure}
The static x-ray diffraction experiments were performed at the Material Science beamline of the Swiss Light Source, with a (2D+2S) surface diffractometer equipped with a PILATUS 100K two-dimensional detector \cite{Willmott2013}. The 2D detector allowed us to sample large three-dimensional reciprocal space volume data sets which were converted to I(h,k,l), intensity as a function of coordinates in reciprocal space \cite{Schleputz2011}. From these data reciprocal space maps could subsequently be extracted. The x-ray energy was 17~keV and the x-ray incidence angle was kept constant to ensure a homogeneous footprint. The sample temperature was controlled using a nitrogen cryojet, and the transition temperatures were in agreement with magnetization measurements. The sample was initially oriented in the AUS phase, and the resulting orientation matrix was used as reference when cooling to the MT phase.
In \figref{figrsm_T2} we show reciprocal space maps through the $(202)_{AUS}$ reflection as the sample is cooled from 408~K to 290~K. As the temperature is lowered we first (T=372~K and 356~K) see the appearance of the satellite reflections from the pre-martensite phase \cite{Zheludev1996,Mariager2014}, which has a modulation wavevector of $\mathbf{q} = [\xi\xi0]$ with $\xi \approx 1/3$. As the MT transition is reached (T = 356~K and 335~K) the regular lattice peak is split and several Bragg reflections corresponding to different MT twins can be seen, though none coincide exactly with the shown lattice plane.
\begin{figure}[ht]
\includegraphics[scale=.7]{Ni2MnGa_film_rsm_T_100_v2.eps}
\caption{(Color online) A series of reciprocal space maps taken at different temperatures during cooling. The premartensite satellite reflection at $\pm.42$ (which is .33 in r.l.u.) slowly appears from T = 408~K where it is hardly visible towards 356~K. At lower temperature the martensite transition has occurred and new satellite lattice peaks appear. The intensity of these peaks are independent of temperature (T = 315 and 290~K) and are clearly not located exactly on the [01$\overline{1}$]$_A$ rod.} \label{figrsm_T2}
\end{figure}
To identify the satellite reflections from the structural modulation of the MT phase, a cut has to be made between two MT peaks corresponding to the same twin. One example of such a reciprocal space map can be seen in \figref{figctr2}. It shows a cut including the $[202]_{MT}$ and $[224]_{MT}$ peaks corresponding to a single structural twin and can be used to identify the many satellite reflections located exactly between them.
\begin{figure*}[hbt]
\includegraphics[scale=.68]{Ni2MnGa_film_struct_ctrs3.eps}
\caption{(Color online) Reciprocal space map showing the streaks of scattering
extending from a [202]$_{MT}$ to a [224]$_{MT}$ Bragg peak. It shows the $c_{mt}$ plane and in AUS coordinates corresponds to a cut through [2.103 0.105 2.015] with surface normal [1.0568 0.0547 0].} \label{figctr2}
\end{figure*}
In \figref{figmod1} we show an example of a line scan along such a rod. We note that this scan is representative, that is, if there are any satellite reflections between two Bragg reflections corresponding to a single twin, we observe exactly these peaks. This structure is somewhat similar to scans published by \textcite{Eichhorn2011}, where they attributed it to a 7M structure. We observe the following concerning the location of the satellite reflections. First, all peaks appear modulo 2 r.l.u. and in pairs at $\pm q$ from the main reflections. In addition they are not placed at periodic intervals between the reflections with the main reflections occurring at $\Delta q = \pm 0.311(1), 0.51(1), 0.83(1)$. This is significantly different from the incommensurately modulated structure known from stochiometric Ni$_2$MnGa \cite{Righi2007} and it is not consistent with the typical 5M or 7M phases in any of their manifestations. The symmetric reflections indicate a periodic modulation along the \{110\} axes of the sample, but the different deviations $\Delta q$ indicate that more than a single modulation period exists. In fact, the observed reflections are consistent with two wavevectors with $\xi = 0.31$ and 0.51. Two such modulations will result in peaks at any combination of the modulation vectors, explaining the observed reflection at 0.83 (within the measurement precision). To test this we calculate the scattering from a structure with two such modulation periods \cite{Ustinov2009,Mariager2014} and as shown in \figref{figmod1} the result is a very good agreement between observed and calculated peak positions. The intensities are not reproduced as the calculation did not include the unit cell formfactor, allowed for different modulation amplitude for different atoms or included experimental factors such as polarization and the Lorentz factor.
We finally emphasize that the observed Bragg reflections are not compatible with an adaptive phase \cite{Kaufmann2010}. The satellite lattice reflections from such a structure would result in evenly spaced intermediate reflections, with only small possible deviations due to stacking faults \cite{Ustinov2009,Mariager2014}
\begin{figure}[ht]
\includegraphics[scale=.7]{Ni2MnGa_film_modulation1.eps}
\caption{(Color online) A line scan taken between two [0-22] and [2-24] peaks of a single MT twin. The rich structure in the immediate vicinity of the regular reflections (l = 2 and l = 4) arise from the other 11 twins. The red dotted line shows the calculated peak positions from a structure with two modulation periods.} \label{figmod1}
\end{figure}
\section{Optical data}
The photoinduced change in reflectivity was investigated in an optical pump-probe setup with a time resolution of $\sim 80$~fs. The 800~nm pump and probe pulse were generated with a 2~kHz Ti-Saphire laser system. The pump and probe pulses were cross polarized and incident along the [001] surface normal with the probe polarization aligned to the [100] crystal axis. The pump was focused to $500\times500~\mu m^2$ and the probe spot size was a factor of three smaller to ensure homogenous excitation of the probe region. The sample was mounted in a cryostat.
In \figref{figOpt}A we show the result of a time resolved optical pump-probe reflectivity measurements taken above and below $T_{MT}$ with a pump fluence of 1~mJ/cm$^2$. Below $T_{MT}$ the signal shows clear oscillations, with a beating indicating the presence of more than a single frequency. Above $T_{MT}$ no oscillations are present, beyond a single initial dip in reflectivity. In addition there is an initial electronic peak, whose amplitude as seen in \figref{figOpt}B scales with fluence. The phase transition is clearly visible as the active phonon modes change with the change in lattice symmetry. This behavior is qualitatively consistent with what we have observed in single crystal Ni$_2$MnGa \cite{Mariager2012a}. Both a Fourier analysis and a fit including two cosine functions reveal the existence of two distinct frequencies, around 1.3 and 0.9~THz just below T$_{MT}$ and hardening as the temperature is lowered. Given the two modulation wavevectors of the structure, the two coherent phonons can be attributed to two amplitudons, that is, an oscillation of the modulation amplitude. This is consistent with the existence of an amplitudon in stochiometric single crystal Ni$_2$MnGa \cite{Mariager2012a}. In \cite{Mariager2012a} we could not distinguish between an amplitudon and a zone folded acoustic mode, but with the structure now absolutely determined as incommensurately modulated \cite{Mariager2014} the observed phonon can be assigned to the amplitudon.
\begin{figure} [ht]
\includegraphics[scale=.8]{Ni2MnGa_dynPaper1_film_optical_v1.eps}
\caption{All optical pump-probe experiments. (A) Time resolved reflectivity as a function of temperature at a fluence of 1~mJ/cm$^2$. (B) Time resolved reflectivity as a function of pump fluence. The black lines shows the fits used to extract the frequencies presented in the main text.} \label{figOpt}
\end{figure}
\section{Amplitudon Model}
In the fit in FIG.1(e) of the main paper we assumed that the modulation amplitude varied linearly with the laser fluence. Several other models have been tested. An assumption that the modulation amplitude falls of as a power of f results in a significantly worse fits ($R^2 = 0.931$ for $x = x_0(1-\sqrt{f/f_c}$ )). On the other hand a model with a true threshold where $x = x_0$ for $f<f_c$ and 0 for $f\geq f_c$ cannot be ruled out ($R^2 = 0.990$) due to the low number of data points at low fluence and gives $f_c = 0.61(5)$~mJ/cm$^2$. This model is however complicated as it must be combined with the existence of a coherent phonon to account for the data in FIG.1(a) of the main text.
The critical fluence $f_c = 1.83(1)$~mJ/cm$^2$ can be converted to a corresponding temperature, though this number depends on the laser absorption depth. For $\delta = 20$~nm we find a critical energy density of 797~J/cm$^3$, which is about 1~eV per unit cell or a temperature rise of 250~K. While it then appears that the energy needed to quench the modulation on an ultrafast timescale is significantly higher than the thermal energy needed to reach $T_{MT}$, one explanation might be that this energy is distributed in various electronic channels from which only a fraction contributes to quenching the modulation.
\section{Time resolved rocking curves}
The data in FIG.2(b) of the main paper is composed from rocking curves (rotation of the sample around the surface normal) taken at different time delays. The rocking curves for four different pump fluences are shown in \figref{figAusRC}. The un-pumped rocking curve, as seen in \figref{figAusRC}A at t = -0.25~ps is, a combination of the rocking curves from 6 of the 12 MT twins, which are here probed in a single scan. The shape can be reproduced from the static x-ray data if one take into account the opening angles of the avalanche photo diode used for the time-resolved experiments. At a low fluence of 1.7~mJ/cm$^2$ the transition is only seen as a small increase in intensity at the location of the AUS peak, while the majority of the film remains in the MT phase. Only at the high fluence of 12~mJ/cm$^2$ in \figref{figAusRC}D is the high temperature AUS phase reached. Here the entire film is transformed and a single strong AUS (202) Bragg reflection occurs at $\Phi = 92.1^{\circ}$. At the intermediate fluences only parts of the film is transformed, resulting in a strained AUS structure.
The temperature rise due to a single laser pulse is given as:
\begin{equation}
\Delta T = (1-R)f_0(1-e^{-d/\delta})/C\rho d
\end{equation}
The specific heat of Ni$_2$MnGa is C = 0.35 - 0.4~J/gK and the density is $\rho = 8.14~g/cm^3$. The laser penetration depth at $\lambda=800$~nm is $\delta=20~nm$, the reflectivity is R = 0.15 and $d$ is the thickness of the layer in which the laser is absorbed. For $d = 1~\mu$m, the thickness of the film, the highest fluence of 12~mJ/cm$^2$ corresponds to a temperature rise of $\Delta T \approx 28$~K. This is just sufficient to raise the temperature of the film from 340~K to above $T_{MT}$, under the assumption that the heat is evenly distributed. This is in good agreement with the observed rocking curves where lower laser fluences only lead to a partial transformation.
\begin{figure*} [ht]
\includegraphics[scale=.65]{Ni2MnGa_film_rc_v1.eps}
\caption{(Color online) Rocking curves as a function of time delay at four different pump fluences. The legend shows the delay in units of ps, and is the same for all figures. (A) 1.7~mJ/cm$^2$, (B) 5.1~mJ/cm$^2$, (C) 8.5~mJ/cm$^2$, (D) 12~mJ/cm$^2$.} \label{figAusRC}
\end{figure*}
\putbib
\end{bibunit}
\end{document}
|
\section{Introduction}
\label{sec:intro}
The cosmological inflation at the early universe is an
attractive scenario which can solve the flatness and the
horizon problems, and simultaneously explains
the density perturbation of the initial universe.
Recent data from the Planck satellite~\cite{Ade:2013zuv}
show that the primordial non-Gaussianity in the Cosmic
Microwave Background (CMB) fluctuations is small and
the spectral index $n_s$ is less than $1$, and set an
upper bound on the tensor-to-scalar ratio $r$.
In addition to the Planck result, BICEP2
experiments reported the lower bound on the ratio $r$ \cite{Ade:2014xna}.
Because our universe is isotropic, most inflation models
assume a Lorentz scalar field called inflaton field
(which does not transform under the four-dimensional
Lorentz transformation of our universe) with very
specific forms of its potential terms (even of its
kinetic term) in order to be consistent with observations,
because parameters in the inflaton potential (as well as
the kinetic term) are severely constrained by the CMB data.
From the theoretical point of view, we may have to identify
some specific origin of such an inflaton scalar field itself,
otherwise it is difficult to restrict these parameters.
One of the origin could be a modulus field which
appears as a zero-mode of extra-dimensional components
in vector and/or tensor fields in higher-dimensional
spacetime with the compactified extra-dimensions.
Parameters in the modulus scalar potential are constrained
by the higher-dimensional Lorentz as well as gauge invariance.
Moduli fields are ubiquitous in the superstring/M theory,
one of the promising candidates for a unified description
of elementary particles and gravity, whose low energy
effective theories are described by supergravity.
The vacuum expectation values of closed (open) string
moduli fields determine, e.g., the size and the shape of
extra-dimensional space (the position of D-branes and
Wilson-lines of the gauge potential induced on them)
and so on, which accordingly determine phenomenological
aspects of the effective theory around the vacuum.
Therefore, it is important to study moduli inflation
scenarios based on the full supergravity framework, where
the local supersymmetry plays important roles to determine
the precise form of moduli kinetic and potential terms
as well as their couplings to matter fields.
The five-dimensional (5D) supergravity with the
compact fifth dimension, which has a full off-shell
formulation~\cite{Zucker:1999ej,Kugo:2000af}
with a local superconformal symmetry, provides a simple
but the attractive starting point for such a study.
Bacause a way of dimensional reduction keeping the off-shell
structure was proposed~\cite{Abe:2006eg} in 4D ${\cal N}=1$
superspace~\cite{Abe:2004ar,Paccetti:2004ri},
we can derive a four-dimensional (4D) effective action
for moduli and matter fields systematically, which has
a full 4D ${\cal N}=1$ local super(conformal) symmetry.
Then, we can easily write down the on-shell action (not
only in the Einstein frame but also in any other frame,
if necessary) for analyzing the moduli inflation and the
related particle cosmology.
In this paper, we study the moduli inflation starting from
the 5D off-shell supergravity compactified on orbifold
$S^1/Z_2$ with two fixed points. In addition to $Z_2$-even
vector and hypermultiplets which include multiplets of
supersymmetric standard model in their zero-modes, we
introduce $Z_2$-odd vector fields forming $Z_2$-odd
vector multiplets, whose fifth components yield multiple
moduli forming chiral multiplets in the 4D effective theory.
Numerous particle physics models were proposed so far in
such an orbifold framework, where the chirality of the
observed quarks and leptons arises as a consequence of
the orbifold structure, and there is a mechanism to
localize matter (and even gravity) fields exponentially
in the fifth dimension, which can be a source of the
observed hierarchical structure of quark and lepton
masses and mixings~\cite{ArkaniHamed:1999dc}
\footnote{See Ref.~\cite{Abe:2008an} for a realization
of the realistic flavor structure in the framework of off-sell
dimensional reduction.}
(and of the huge hierarchy between the weak and the
Planck scales~\cite{Randall:1999ee}).
A superpotential for such localized matter fields
is allowed at the fixed points where the supersymmetry
is reduced, which can be a source of moduli potential as well,
in addition to some nonperturbative effects such as
a gaugino condensation. We expect the exponential
form of the localized wavefunctions plays a certain
role to realize a successful moduli inflation
as well as their stabilization.
This paper is organized as follows.
In Sec.~\ref{sec:moduli}, we review the moduli fields
appearing in 5D supergravity on $S^1/Z_2$. Then,
we propose two simple models, one of them realizes
the small-field inflation and the other does the
large-field one in Secs.~\ref{sec:sinflation} and \ref{sec:linflation}, respectively, both are
triggered by the moduli dynamics.
We show their consistency with the recent
observations. Sec.~\ref{sec:conclusion}
is devoted to the conclusion. We show the
canonical normalization of fields for the large-field
model in Appendix~\ref{app:can}.
\section{Moduli effective action on orbifold $S^1/Z_2$}
\label{sec:moduli}
First in this section we review moduli effective action
appearing from a compactification of 5D (off-shell)
supergravity on orbifold $S^1/Z_2$.
The most general form of the 5D background metric
preserving a 4D flatness is given by
$ds^2=G_{MN}dx^Mdx^N
=e^{-2\sigma(y)}\eta_{\mu \nu}dx^\mu dx^\nu -dy^2$
where $M,N=0,1,2,3,4$ are 5D spacetime indices,
$\mu,\nu=0,1,2,3$ are 4D spacetime indices,
$y=x_{M=4}$ represents the fifth coordinate and
$\sigma(y)$ is an arbitrary function of $y$
(up to the following restriction).
Because the fifth direction is compactified on $S^1/Z_2$,
any field $f(x,y)$ (including gravity fields and then the
above function $\sigma(y)$) satisfies $f(x,y+L)=f(x,y)$
and $f(x,-y)=f(x,y)$ (for $Z_2$ even fields)
or $f(x,-y)=-f(x,y)$ (for $Z_2$ odd fields)
where $L$ is the length of orbifold segment, and then
there are two fixed points at $y=0$ and $y=L$.
The supersymmetry in 5D has eight supercharges.
For our purpose, relevant 5D supermultiplets
are vector multiplets
${\bm V}^I=\{ V^I, \Sigma^I \}$
with $I=1,2,\ldots,n_V$
and hypermultiplets
${\bm H}_\alpha=\{ {\cal H}_\alpha, {\cal H}^C_\alpha \}$
with $\alpha=1,2,\ldots,n_H+n_C$, where
$V^I$, $\Sigma^I$, ${\cal H}_\alpha$ and ${\cal H}^C_\alpha$
represent vector multiplets and three chiral multiplets,
respectively, under the 4D supersymmetry preserved
after the orbifolding which has four supercharges.
We introduce multiple $Z_2$-odd vector multiplets
${\bm V}^{I'}$ with $I'=1,2,\ldots,n_V'$ in which
the zero-modes of $Z_2$-even chiral multiplets
$\Sigma^{I'}$ become moduli chiral multiplets $T^{I'}$
in 4D, and a linear combination\footnote{
In the case $n_V'=1$, the single modulus $T^{I'=1}$
corresponds to a so-called radion (chiral multiplet)
satisfying $\langle {\rm Re}\,T^{I'=1} \rangle=L/\pi$,
while the radion for $n_V'>1$ is a linear combination
of $T^{I'}$s determined by the norm function
(\ref{eq:nf}).}
of these moduli becomes an inflaton
field in the moduli inflation scenario.
On the other hand, hypermultiplets ${\bm H}_\alpha$ are
introduced in order to generate a suitable moduli
potential for the inflation at the early universe
and for a moduli stabilization at the present universe.
The 5D off-shell (conformal) supergravity action
for vector and hypermultiplets is completely fixed
by identifying the numbers of multiplets,
$n_V \ge 1$ and $n_H \ge n_C \ge 1$ where $n_C$
is the number of compensator hypermultiplets,
and then determining a cubic polynomial of vector
multiplets
\begin{eqnarray}
{\cal N}(M) &=&
\sum_{I,J,K=1}^{n_V}C_{I,J,K}M^IM^JM^K,
\label{eq:nf}
\end{eqnarray}
with real coefficients $C_{I,J,K}$
for $I,J,K=1,2,\ldots,n_V$.
The manifold of vector multiplets is called
very special manifold governed by the norm function
${\cal N}({\bm V})$,
while that of hypermultiplets is dependent to $n_C$
(See Ref.~\cite{Zucker:1999ej,Kugo:2000af} and
references therein).
In this paper we choose $n_C=1$ for simplicity.
The hypermultiplet ${\bm H}_\alpha$ can be a non-trivial
representation of gauge symmetries in the 5D action
whose gauge fields are identified with the vector
fields $A_M^I$ in vector multiplets ${\bm V}^I$.
In this paper we identify the $Z_2$-odd vector
fields $A_M^{I'}$ in ${\bm V}^{I'}$ as gauge fields
of $U(1)_{I'}$ symmetries for simplicity,
and assign $U(1)_{I'}$ charges $c_{I'}^{(\alpha)}$
to the hypermultiplets ${\bm H_\alpha}$.
We introduce the same number of
(stabilizer) hypermultiplets $n_H=n_V'$ as that of
$Z_2$-odd vector multiplets in order to stabilize
the moduli $T^{I'}$ at a supersymmetric Minkowski
minimum.\footnote{This moduli stabilization
mechanism was proposed in Ref.~\cite{Maru:2003mq}
to stabilize a single modulus, i.e., the radion
for $n_V'=1$, which is extended here in this paper
to the case with multiple moduli for $n_V'>1$.}
So far we have set bulk configurations in the
5D supergravity compactified on $S^1/Z_2$.
In addition to these 5D data,
K\"ahler and superpotential terms are allowed at
the orbifold fixed points, where the supersymmetry
is reduced to the one with four supercharges.
For $Z_2$-even (stabilizer) chiral multiplets ${\cal H}_i$
($i=1,2,\ldots,n_H$) contained in the hypermultiplets
${\bm H}_\alpha$, we consider in our scenario
that the linear terms of ${\cal H}_i$,
\begin{eqnarray}
{\cal W} &=&
J_0^{(i)}\,{\cal H}_i\,\delta(y)
+J_L^{(i)}\,{\cal H}_i\,\delta(y-L),
\label{eq:w5d}
\end{eqnarray}
where $J_{0,L}^{(i)}$ are constants,
are dominant~\cite{Abe:2006eg}
in the superpotential ${\cal W}$ induced at the
fixed points $y=0$ and $y=L$, and the other terms
are forbidden or negligible due to some symmetries
or dynamics.\footnote{For $n_C=2$, a similar
moduli stabilization potential was proposed~\cite{Abe:2007zv}
in the framework of off-shell dimensional reduction.}
Furthermore, we assume that the terms
in the K\"ahler potential at the fixed points are
also negligible compared with the bulk contributions,
that can be a natural assumption if the radius $L/\pi$
of the compactified fifth dimension $y$ is larger
enough than the inverse of the mass scales
associated with these terms.
Now the (off-shell) supergravity action in 5D
spacetime is completely determined, and then
employing the off-shell dimensional reduction~\cite{Abe:2006eg},
we can integrate it over the fifth dimension $y$ and
find the following K\"ahler potential $K$
and the superpotential $W$,
\begin{eqnarray}
K &=& -\ln {\cal N}({\rm Re}\,T)
+Z_{i,\bar{i}}({\rm Re}\,T)\,|H_i|^2,
\nonumber \\
W &=& \left( J_0^{(i)}
+e^{-c_{I'}^{(i)}T^{I'}}J_L^{(i)} \right) H_i,
\label{eq:4dw}
\end{eqnarray}
in the 4D effective action, where
\begin{eqnarray}
Z_{i,\bar{j}}({\rm Re}\,T) &=&
\frac{1-e^{-2c_{I'}^{(i)}{\rm Re}\,T^{I'}}}{
c_{I'}^{(i)} {\rm Re}\,T^{I'}} \delta_{ij},
\label{eq:km}
\end{eqnarray}
is the K\"ahler metric of 4D zero-modes $H_{i}$
in a Kaluza-Klein expansion of $Z_2$-even components
${\cal H}_i$ of hypermultiplet ${\bm H}_{\alpha=i}$.
Here we remark that the exponential factors in $K$
and $W$ with the $U(1)_{I'}$ charges $c_{I'}^{(i)}$
in their exponents originate from the fact that the
wavefunctions of zero-modes $H_i$ are localized
exponentially in extra dimensions~\cite{Abe:2006eg},
which play important roles in this paper to realize
a successful moduli inflation at the early universe
as well as the moduli stabilization
at the late time.
We are ready to write down the effective 4D
scalar potential $V$ for zero-modes $T^{I'}$ and $H_i$,
where the former and the latter are called moduli and
stabilizer fields (both chiral multiplets) respectively,
using the standard formula of 4D supergravity
(with four supercharges) as
\begin{eqnarray}
V &=& e^K \left(
K^{m,\bar{n}}\,D_mW\,D_{\bar{n}} \bar{W}
-3|W|^2 \right),
\label{eq:4dsp}
\end{eqnarray}
where
$D_m W=W_m+K_mW$,
$W_m=\partial_m W$,
$K_m=\partial_m K$,
$m,n=\left\{ I', i \right\}$
and $K^{m,\bar{n}}$
is the inverse of K\"ahler metric
$K_{m,\bar{n}}=\partial_m \partial_{\bar{n}}K$.
The expectation values of moduli $T^{I'}$
and stabilizer fields $H_i$ at an extremum
of the scalar potential~(\ref{eq:4dsp})
are found as~\cite{Maru:2003mq}
\begin{eqnarray}
c_{I'}^{(i)} \langle T^{I'} \rangle
&=& \ln \frac{J_L^{(i)}}{J_0^{(i)}}, \qquad
\langle H_i \rangle \ = \ 0,
\label{eq:movev}
\end{eqnarray}
which satisfy
$\langle D_{I'}W \rangle
= \langle D_iW \rangle
= \langle W \rangle
= 0$
and then
$\langle V_{I'} \rangle
= \langle V_i \rangle
= \langle V \rangle
= 0$
where
$V_m=\partial_m V$.
Without moduli mixings in the K\"ahler metric,
$K_{I',\bar{J}'} = 0$ for $I' \ne J'$,
fields are stabilized at a supersymmetric
Minkowski minimum~(\ref{eq:movev}),
where both the modulus $T^{I'}$
and the stabilizer field $H_i$
obtain a supersymmetric mass
\begin{eqnarray}
m_{I'i}^2 &\simeq&
\frac{e^{\langle K \rangle}
\langle W_{I'i} \rangle^2}{
\langle K_{I',\bar{I}'} \rangle
\langle K_{i,\bar{i}} \rangle},
\label{eq:thmass}
\end{eqnarray}
where
$W_{mn}=\partial_m \partial_n W$
and then
$\langle W_{I'i} \rangle
=-c_{I'}^{(i)}e^{-c_{I'}^{(i)}
\langle T^{I'} \rangle}J_L^{(i)}$.
It is remarkable that
the mass square~(\ref{eq:thmass})
is exponentially suppressed with
its exponent proportional to the
$U(1)_{I'}$ charge $c_{I'}^{(i)}$
of the stabilizer field $H_i$,
that is one of the consequences of
wavefunction localization in the
extra dimension as mentioned
in Sec.~\ref{sec:intro}.
Note that , for
$\left| \langle W_{I'i} \rangle/\langle W_{J'j} \rangle \right|
\sim {\cal O}(1)$ ($\forall I',J',i,j$), this moduli stabilization
mechanism does not work with a sizable
moduli mixing in the K\"ahler metric,
$K_{I',\bar{J}'} \ne 0$ for $I' \ne J'$,
with which the above expectation
values~(\ref{eq:movev}) correspond
to a saddle point or a local maximum
of the scalar potential.
Therefore, the coefficients $C_{I',J',K'}$
in the norm function~(\ref{eq:nf})
are restricted\footnote{In the later
concrete example of our model
in Sec. \ref{sec:sinflation}, we will
take the norm function~(\ref{eq:nfinf})
and~(\ref{eq:qpnv2}) for $n_V'=2$,
that leads to such a diagonal metric.}
to those yielding an almost diagonal
moduli K\"ahler metric,
$K_{I',\bar{J}'} \approx 0$ for $I' \ne J'$
at least at the minimum~(\ref{eq:movev}).
On the other hand, for
$\left| \langle W_{I'i} \rangle/\langle W_{J'j} \rangle \right|
\ll 1$ ($\exists I',J',i,j$),
it is possible that a sizable K\"ahler
mixing does not spoil the stability of
the vaccum (\ref{eq:movev}) depending
on the hierarchy of $\langle W_{I'i} \rangle$.
The large-field model proposed in Sec.
\ref{sec:linflation}
utilizes this fact.
From a particle phenomenological point
of view, we have to introduce a supersymmetry
breaking sector, which in general affects
the moduli stabilization.
For the case presented in this paper,
the shift of position of
the minimum~(\ref{eq:movev})
is negligible if
the supersymmetric mass~(\ref{eq:thmass})
is larger enough than the supersymmetry
breaking scale. Even in this case,
the height of the minimum will be affected by
the supersymmetry breaking, and we just assume
$\langle V \rangle \approx 0$ even after the
breaking sector is incorporated.
We will discuss the validity
of this assumption in Sec. \ref{subsec:modpro}.
\section{A simple model for the small-field inflation}
\label{sec:sinflation}
In this section, we show that the moduli
potential discussed so far allows a scenario
of small-field moduli inflation that can realize the
observed WMAP and Planck data~\cite{Ade:2013zuv}.
We consider the case that
one pair of modulus and stabilizer fields,
e.g. $T^{I'=1}$ and $H_{i=1}$,
is decoupled from and lighter enough
than the other pairs
$T^{I' \ne 1}$ and $H_{i \ne 1}$,
that is,
$m_{I'=1,i=1}^2 \ll m_{I' \ne 1,i \ne 1}^2$.
Such a case can be naturally realized when
$c_{I'=1}^{(i \ne 1)}=c_{I' \ne 1}^{(i=1)}=0$
in Eq.~(\ref{eq:4dw}) and
$\left| c_{I' \ne 1}^{(i \ne 1)} \right|
<\left| c_{I'=1}^{(i=1)} \right|$
in Eq.~(\ref{eq:thmass}).
In this case, below the heavier mass scale
$m_{I' \ne 1,i \ne 1}$,
all the moduli $T^{I' \ne 1}$
and stabilizer fields $H_{i \ne 1}$
except the lighter pair
$T^{I'=1}$ and $H_{i=1}$
are strictly fixed at their supersymmetric
minimum with no fluctuations around it,
and they are replaced by their vacuum
expectation values~(\ref{eq:movev})
in the low energy effective action.
Then, the effective K\"ahler potential
and superpotential for the lighter fields
$T^1$ and $H_1$ are given by
\begin{eqnarray}
K_{\rm eff}(T^1,H_1)
&=& K(T^{I'},H_i)\,\Big|_0
\ = \ -\ln {\cal N}({\rm Re}\,T)\,\Big|_0
+Z_{1,\bar{1}}
({\rm Re}\,T)\,\Big|_0 \,|H_1|^2,
\nonumber \\
W_{\rm eff}(T^1,H_1)
&=& W(T^{I'},H_i)\,\Big|_0
\ = \ \left( J_0^{(1)}
+e^{-c_1^{(1)}T^1}J_L^{(1)} \right) H_1,
\label{eq:weff}
\end{eqnarray}
respectively, where
$f(T^{I'},H_i)\,\Big|_0 \equiv f(T^{I'},H_i)
\Big|_{\scriptsize \begin{array}{ll}
T^{I' \ne 1}=\langle T^{I' \ne 1} \rangle \\
H_{i \ne 1}=\langle H_{i \ne 1} \rangle
\end{array}}$
for an arbitrary function
$f(T^{I'},H_i)$, and then
\begin{eqnarray}
Z_{1,\bar{1}}({\rm Re}\,T)\,\Big|_0 &=&
\frac{1-e^{-2c_1^{(1)}{\rm Re}\,T^1}}{
c_1^{(1)} {\rm Re}\,T^1}.
\nonumber
\end{eqnarray}
The effective potential for the light fields,
\begin{eqnarray}
V_{\rm eff}(T^1,H_1) &=& e^{K_{\rm eff}}
\left( (K_{\rm eff})^{m,\bar{n}}\,
D_mW_{\rm eff}\,D_{\bar{n}} \bar{W}_{\rm eff}
-3|W_{\rm eff}|^2 \right),
\label{eq:4dspeff}
\end{eqnarray}
is obtained by using the effective K\"ahler potential
and superpotential~(\ref{eq:weff}),
where $m,n=\{I',i\}$ with $I'=1$ and $i=1$.
\subsection{The inflaton potential}
\label{ssec:potential}
From Eq.\,(\ref{eq:4dspeff}), we find the effective
potential for the modulus $T^1$ (whose real part
will be identified as the inflaton field later)
on the $H_1=0$ hypersurface,
\begin{eqnarray}
V_{\rm eff}(T^1,H_1=0)
&=& e^{K_{\rm eff}}
(K_{\rm eff})^{i=1,\bar{i}=\bar{1}}
\left| (W_{\rm eff})_{i=1} \right|^2
\,\Big|_{H_1=0}
\nonumber \\ &=&
\frac{c_1^{(1)} {\rm Re}\,T^1}{
{\cal N}({\rm Re}\,T)\,\Big|_0}
\times
\frac{\left| J_0^{(1)} \right|^2
\left| 1+\frac{J_L^{(1)}}{J_0^{(1)}}
e^{-c_1^{(1)}T^1} \right|^2}{
1-e^{-2c_1^{(1)}{\rm Re}\,T^1}},
\label{eq:veff}
\end{eqnarray}
where
$(K_{\rm eff})^{i,\bar{i}}\Big|_{H_1=0}
=1/Z_{i,\bar{i}}({\rm Re}\,T)\,\Big|_0$.
In the case
\begin{eqnarray}
{\cal N}({\rm Re}\,T)\Big|_0
&=& {\cal P}_0\,{\rm Re}\,T^1,
\label{eq:cdnf}
\end{eqnarray}
where ${\cal P}_0$ does not depend on $T^1$,
the first factor in Eq.~(\ref{eq:veff})
is independent to $T^1$, and then we find
\begin{eqnarray}
\lim_{{\rm Re}\,T^1 \to 0}
\left| V_{\rm eff}(T^1,H_1=0) \right|
&=& \infty, \nonumber \\
\lim_{{\rm Re}\,T^1 \to \infty}
V_{\rm eff}(T^1,H_1=0) &=&
c_1^{(1)} {\cal P}_0^{-1}
\left| J_0^{(1)} \right|^2
\ \equiv \ V_\infty,
\label{eq:limveff}
\end{eqnarray}
for $J_L^{(1)}/J_0^{(1)} \ne -1$ and $c_1^{(1)}>0$.
Fig.~\ref{fig:ip} shows
the ${\rm Re}\,T^1$-dependence of
$V_{\rm eff}(T^1,H_1)/V_\infty$
on the ${\rm Im}\,T^1=H_1=0$ hypersurface,
where the parameters are chosen as
\begin{eqnarray}
c_1^{(1)} &=& 1/10, \qquad
J_L^{(1)}/J_0^{(1)} \ = \ -3.9, \qquad
J_0^{(1)} \ = \ 10^{-4},
\label{eq:para}
\end{eqnarray}
in the Planck scale unit $M_{\rm Pl}=1$.
In Fig.~\ref{fig:ip}, we recognize the above
feature~(\ref{eq:limveff}) and expect that ${\rm Re}\,T^1$
can play a role of inflaton field, starting from its large
positive value on the flat region of the potential and slowly
rolling down to the minimum\footnote{We comment that this
shape of potential is essentially the same as the one in
Starobinski model~\cite{Starobinsky:1980te},
but the origin of the potential is quite different.}
given by Eq.~(\ref{eq:movev}) for $i=1$.
Furthermore, we find that the overshooting to negative region
${\rm Re}\,T^1<0$ is prohibited, which is also understood
from Eq~(\ref{eq:limveff}).
\begin{figure}[t]
\centering \leavevmode
\includegraphics[width=0.6\linewidth]{infpo.eps}
\caption{Scalar potential
$V_{\rm eff}(T^1,H_1)/V_\infty$
on the ${\rm Im}\,T^1=H_1=0$ hypersurface}
\label{fig:ip}
\end{figure}
Before analyzing the inflation dynamics, we should
recall the fact that the flatness of the potential
in the large ${\rm Re}\,T^1$ region is guaranteed
by the assumption~(\ref{eq:cdnf}).
The most general form of norm function~(\ref{eq:nf})
satisfying the condition~(\ref{eq:cdnf}) is found as
\begin{eqnarray}
{\cal N}(M) &=& {\cal P}(M)\,M^1+\cdots,
\label{eq:nfinf}
\end{eqnarray}
where
\begin{eqnarray}
{\cal P}(M) &=&
\sum_{J', K' \ne 1}^{n_V'} C_{1,J',K'}M^{J'}M^{K'},
\label{eq:qp}
\end{eqnarray}
is a quadratic polynomial of fields $M^{I' \ne 1}$
in $Z_2$-odd vector multiplets ${\bm V}^{I' \ne 1}$
other than ${\bm V}^{I'=1}$,
and the ellipsis stands for terms including fields
$M^{I''}$ in $Z_2$-even vector multiplets
${\bm V}^{I''}=\{V^{I''},\Sigma^{I''}\}$
with $I''=n_V'+1,n_V'+2,\ldots,n_V$
whose components $\Sigma^{I''}$ are $Z_2$-odd
chiral multiplets which do not carry any moduli.
The coefficient ${\cal P}_0$ of ${\rm Re}\,T^1$
in Eq.~(\ref{eq:cdnf}) is given by
${\cal P}_0 = {\cal P}({\rm Re}\,T)\Big|_0$,
which is a field independent constant
by definition~(\ref{eq:qp}).
Therefore, we find that the interesting flat region
is realized in a moduli stabilization potential
generated by a simple superpotential~(\ref{eq:w5d})
as a consequence of the peculiar form of norm
function~(\ref{eq:nfinf}) in 5D supergravity.
Note especially that, the condition~(\ref{eq:nfinf})
cannot be satisfied for $n_V'=1$, i.e., the single
modulus case where only the radion exists, because
the norm function ${\cal N}(M)$ is a cubic polynomial.
For $n_V'=2$, the quadratic polynomial ${\cal P}(M)$
is uniquely determined as
\begin{eqnarray}
{\cal P}(M) &=& C_{1,2,2}\,(M^2)^2.
\label{eq:qpnv2}
\end{eqnarray}
Finally we remark that,
although the norm function coefficients $C_{I,J,K}$
are free parameters in 5D supergravity, these are
closely related to the structure of the internal
manifold, if it is the 5D effective theory of a
more fundamental theory defined in more than five
dimensional spacetime with extra dimensions
compactified on some manifold.\footnote{
One of such examples is the 5D effective
theory of heterotic M-theory,
where the norm function coefficients correspond
to the intersection numbers of internal Calabi-Yau
three-fold~\cite{Lukas:1998yy}.}
In such a situation, the cosmological (as well as
phenomenological)
features of 5D supergravity are governed by the
internal manifold behind it.
\subsection{The inflation dynamics}
\label{ssec:inflation}
Based on the previous arguments, we identify one
of the moduli fields, the real part of the lightest
modulus, ${\rm Re}\,T^1$, as the inflaton field.
Although the inflation mechanism proposed in this
paper is applicable to any number $n_V' \ge 2$ of
$Z_2$-odd vector multiplets ${\bm V}^{I'}$,
in the following, we choose the minimal number
$n_V'=2$ just for simplicity and concreteness.
Then the norm function~(\ref{eq:nfinf})
is uniquely determined by the quadratic
monomial~(\ref{eq:qpnv2}), where we set
$C_{1,2,2}=1$ without loss of generality
which determines the normalization
of the field $M^2$.
By assuming that oscillations of
the other light fields
${\rm Im}\,T^1$,
${\rm Re}\,H_1$ and
${\rm Im}\,H_1$
than ${\rm Re}\,T^1$
around their expectation values
$\langle {\rm Im}\,T^1 \rangle
=\langle {\rm Re}\,H_1 \rangle
=\langle {\rm Im}\,H_1 \rangle=0$
are negligible during and after the inflation
(which will be confirmed in Sec. \ref{subsec:modpro}),
we solve the equation of motion for the single field
$\sigma \equiv {\rm Re}\,T^1$,
\begin{eqnarray}
\ddot\sigma +3H \dot\sigma
+\Gamma^\sigma_{\ \sigma \sigma} \dot\sigma^2
+g^{\sigma \sigma}
\frac{\partial V_{\rm eff}}{\partial \sigma}
&=& 0,
\label{eq:eominf}
\end{eqnarray}
where the dot denotes the derivative $\frac{d}{dt}$
with respect to a cosmic time $t$, $V_{\rm eff}$
is the effective potential~(\ref{eq:4dspeff}),
$g_{\sigma \sigma}=2(K_{\rm eff})_{I'=1,J'=1}$,
$g^{\sigma \sigma}=g_{\sigma \sigma}^{-1}$ and
$\Gamma^\sigma_{\ \sigma \sigma}$ is the
Christoffel symbol constructed by the metric
$g_{\sigma \sigma}$, all on the
${\rm Im}\,T^1={\rm Re}\,H_1={\rm Im}\,H_1=0$
hypersurface of the field space.
The Hubble parameter $H$ is given as
$H^2=\left( \frac{\dot{a}}{a} \right)^2
=\frac{1}{6} g_{\sigma \sigma} \dot\sigma^2
+\frac{V_{\rm eff}}{3}$ where
$a$ is the scale factor of 4D spacetime,
in which the 4D effective theory of
5D supergravity is defined.
Eq.~(\ref{eq:eominf}) is rewritten as
\begin{eqnarray}
\sigma''
&=& -\left( 1-
\frac{g_{\sigma \sigma}(\sigma')^2}{6} \right)
\left( 3\sigma'+6\sigma^2
\frac{V_{\rm eff}'}{V_{\rm eff}} \right)
+\frac{(\sigma')^2}{\sigma},
\label{eq:eominfe}
\end{eqnarray}
where the prime denotes the derivative
$\frac{d}{dN}=H^{-1}\frac{d}{dt}$
with respect to the number $N \equiv \ln a(t)$
of e-foldings, and we have used
$\Gamma^\sigma_{\ \sigma \sigma}=-1/\sigma$.
In the following analysis,
the numerical values of parameters
in the Planck scale unit $M_{\rm Pl}=1$
are chosen as
\begin{eqnarray}
c_2^{(2)} &=& 1/20, \qquad
J_L^{(2)}/J_0^{(2)} \ = \ -9, \qquad
J_0^{(2)} \ = \ 10^{-1},
\label{eq:parah}
\end{eqnarray}
for the heavy fields $T^2$ and $H^2$
as well as those~(\ref{eq:para})
for the light fields $T^1$ and $H^1$.
With these parameters,
the vacuum expectation values of fields
are given by Eq.~(\ref{eq:movev}),
and their numerical values are found as
\begin{eqnarray}
\langle T^1 \rangle \ \simeq \ 13.6, \qquad
\langle T^2 \rangle \ \simeq \ 43.9, \qquad
\langle H_1 \rangle \ = \
\langle H_2 \rangle \ = \ 0,
\nonumber
\end{eqnarray}
that determine
\begin{eqnarray}
{\cal P}_0
&=& {\cal P}({\rm Re}\,T)\Big|_0
\ = \ \langle {\rm Re}\,T^2 \rangle^2.
\nonumber
\end{eqnarray}
At this supersymmetric Minkowski minimum, the
supersymmetric mass squares~(\ref{eq:thmass})
of light ($I'=1$, $i=1$)
and heavy ($I'=2$, $i=2$) fields
are estimated respectively as
\begin{eqnarray}
m^2_{I'=1,i=1}
& \simeq & \left( 4.9 \times 10^{12} \
{\rm GeV} \right)^2, \qquad
m^2_{I'=2,i=2}
\ \simeq \ \left( 6.9 \times 10^{15} \
{\rm GeV} \right)^2,
\nonumber
\end{eqnarray}
while the inflation scale in our model is characterized
by the Hubble scale
\begin{eqnarray}
H_{\rm inf} &\equiv &\left( V_\infty /3M_{\rm Pl}^2\right)^{1/2}
\ \simeq \ 1.0 \times 10^{12} \ {\rm GeV},
\label{eq:sfhinf}
\end{eqnarray}
with $V_\infty$ given in Eq.~(\ref{eq:limveff}),
for $M_{\rm Pl}=2.4 \times 10^{18}$ GeV.
Because all of these scales $m_{I'=1,i=1}$,
$m_{I'=2,i=2}$ and $H_{\rm inf}$ are
below the compactification scale
\begin{eqnarray}
M_C &\equiv& \frac{\pi}{L}
\ \simeq \ \frac{\pi M_{\rm Pl}}{\langle
{\cal N}({\rm Re}\,T) \rangle^{1/2}}
\ \simeq \ 4.7 \times 10^{16} \ {\rm GeV},
\nonumber
\end{eqnarray}
we find the parameters chosen here ensure
the validity of 4D effective-theory
description during and after the inflation.
It is also confirmed that the heavy fields
$T^2$ and $H_2$ are decoupled from the
inflation dynamics due to $m_{I'=1,i=1}
\sim H_{\rm inf} \ll m_{I'=2,i=2}$,
and their oscillations can be neglected.
Now we consider a possibility of slow roll
inflation starting from a large value of
inflaton field $\sigma$ in the flat region of
the potential down to its VEV~(\ref{eq:movev})
at the minimum. To estimate the observed quantities,
we define the generalized slow roll parameters
for the
scalars having non-canonical kinetic term
\cite{Burgess:2004kv},
\begin{align}
\epsilon &\equiv \frac{M_{Pl}^2}{2}
\frac{\partial_\sigma V_{\rm eff}
g^{\sigma\sigma} \partial_\sigma V_{\rm eff}}{V_{\rm eff}^2}
\sim (2c_1^{(1)}\sigma)^2 \left(\frac{J_L^{(1)}}{J_0^{(1)}}e^{-c_1^{(1)}\sigma}\right)^2,
\nonumber\\
\eta &\equiv \frac{\nabla^\sigma\nabla_\sigma V_{\rm eff}}{V_{\rm eff}}
= \frac{g^{\sigma\sigma}\partial_\sigma^2 V_{\rm eff}
-g^{\sigma\sigma}\Gamma^\sigma_{\sigma\sigma} \partial_\sigma V_{\rm eff}}
{V_{\rm eff}}
\sim -(2c_1^{(1)} \sigma)^2 \frac{J_L^{(1)}}{J_0^{(1)}}e^{-c_1^{(1)}\sigma},
\end{align}
where $\nabla_\sigma$ is the covariant derivative
for the field $\sigma$. The observables such as the
power spectrum of scalar curvature perturbation,
its spectral index and the tensor-to-scalar ratio
are written in terms of these slow-roll parameters
as
\begin{align}
P_\xi (k)&=\frac{1}{24\pi^2} \frac{V}{\epsilon M_{\rm Pl}^4}, \nonumber\\
n_s &= 1+ \frac{d\,{\rm ln} P_{\xi}(k)}{d\,{\rm ln}\,k}
\simeq 1-6\,\epsilon +2\,\eta, \nonumber\\
r &=16\,\epsilon,
\label{eq:obss}
\end{align}
We numerically solve Eq.~(\ref{eq:eominfe})
with the initial conditions $\sigma=117$ and
$\sigma'=0$ at $N=0$. Fig.~\ref{fig:io} shows
the evolution of $\sigma$ as a function of $N$.
In this figure, we find the inflation ends
at about $N_{\rm end} \simeq 70.7$ where
the slow-roll condition
is violated (max $\{\epsilon, \eta\}\,=\,1$)
and then
the oscillation of inflaton starts.
\begin{figure}[h]
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{infosc.eps}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{infoscl.eps}
\end{center}
\end{minipage}
\caption{The behavior of inflaton field
$\sigma={\rm Re}\,T^1$ as a function
of the e-folding number $N$.}
\label{fig:io}
\end{figure}
First, we denote the field value
$\sigma=\sigma_\ast$
corresponding to the pivot scale
$k_0=0.05$ [Mpc$^{-1}$]
(at which the horizon exits) and
the scalar potential
$V_\ast^{1/4}
\equiv V^{1/4}(\sigma_\ast)$
at the pivot scale and
$V_{\rm end}^{1/4}
\equiv V^{1/4}(\sigma_{\rm end})$
at the end of inflation. In terms of
them, the e-folding number after
the pivot scale is given by~\cite{Liddle:1993fq},
\begin{eqnarray}
N_{\rm e} &\equiv& N_{\rm end}-N_\ast
\simeq \, 62+\ln
\frac{V_\ast^{1/4}}{10^{16}\ {\rm GeV}}
+\ln \frac{V_\ast^{1/4}}{V_{\rm end}^{1/4}}
-\frac{1}{3} \ln
\frac{V_{\rm end}^{1/4}}{\rho_R^{1/4}}
\ \simeq \ 56,
\label{eq:nend}
\end{eqnarray}
where $V_\ast^{1/4} \simeq V_{\rm end}^{1/4}
\simeq 2 \times 10^{15}$ GeV
and $\rho_R$ is the energy density
by which the universe is thermalized
with the reheating temperature
$T_R \simeq 1.05 \times 10^9$ GeV
whose numerical value will be
determined later in Sec.~\ref{subsec:rehs}.
Note that the energy of inflaton is assumed,
in Eq.~(\ref{eq:nend}), to be instantaneously
converted into radiation. On the other hand,
the same number $N_{\rm e}$ is estimated
based on a slow-roll approximation,
\begin{eqnarray}
N_{\rm e} \ = \ -\int_{t_{\rm end}}^{t_\ast}
d\tilde{t}\, H(\tilde{t})
\ \simeq \ \frac{1}{M_{\rm Pl}^2}
\int_{\sigma_{\rm end}}^{\sigma_\ast} d\sigma
\frac{V_{\rm eff}}{g^{\sigma \sigma}V'_{\rm eff}},
\label{eq:nast}
\end{eqnarray}
and then we find the numerical value
\begin{eqnarray}
\sigma_\ast &\simeq& 114,
\label{eq:sast}
\end{eqnarray}
is determined by equaling the Eq. (\ref{eq:nend})
and Eq. (\ref{eq:nast}).
Second, we check whether the WMAP and Planck
normalization on the power spectrum of scalar
curvature perturbation, $P_\xi(k_0)
=2.196_{-0.060}^{+0.051} \times 10^{-9}$
~\cite{Ade:2013zuv},
can be realized or not.
The slow-roll parameters $\epsilon$ and $\eta$
are obtained at the pivot scale $k_0$ by using the
numerical value (\ref{eq:sast}),
\begin{align}
\epsilon &\sim (2c_1^{(1)}\sigma)^2
\left(\frac{J_L^{(1)}}{J_0^{(1)}}e^{-c_1^{(1)}
\sigma}\right)^2\Bigl|_{\sigma =\sigma_\ast} \simeq
{\cal O}(10^{-6}),
\nonumber\\
\eta &\sim -(2c_1^{(1)} \sigma)^2
\frac{J_L^{(1)}}{J_0^{(1)}}e^{-c_1^{(1)}\sigma}\Bigl|_{\sigma =\sigma_\ast}
\simeq {\cal O}(-0.02),
\label{eq:slowpara}
\end{align}
which yield $P_\xi(k_0) \sim 2 \times 10^{-9}$ of
the correct order of the observed value.
Inversely speaking, the parameters $J_0^{(1)}$ and
$J_L^{(1)}$ are determined in Eq. (\ref{eq:para})
in such a way that the resultant $P_\xi(k_0)$ resides
in the observed region.
Also, the spectral index of the scalar curvature
perturbation, $n_s \,=0.9603\pm 0.0073$
~\cite{Ade:2013zuv}, at
the pivot scale is observed by the WMAP and
Planck collaborations. In our model, we can
realize the correct value of the spectral index,
$n_s \,\simeq\,0.96$ by using Eq. (\ref{eq:obss})
and Eq. (\ref{eq:slowpara}).
It implies that the $\eta$ problem is avoided by
the exponential factor and the large value of the
inflaton field, because the shift symmetry of
$\sigma$ is violated by its own superpotential
(\ref{eq:weff}).
We summarize the results of inflation dynamics
in Figs. \ref{fig:ps} and \ref{fig:si}.
From these figures drawn with the sample values
of parameters (\ref{eq:para}) and (\ref{eq:parah}),
we extract the numerical values of observables as
\begin{align}
N_e= 55.6 \;\text{e-folds},\qquad
P_\xi = 2.23\times 10^{-9},\qquad
n_s =0.959, \qquad
r=1.6\times 10^{-5}.
\label{eq:res}
\end{align}
So the simple model analyzed so far with
Re\,$T^1$ playing the role of inflation
is consistent with the WMAP and
Planck data~\cite{Ade:2013zuv}.
Note that this inflation mechanism
is categorized as the small-field model
of inflation due to the tiny
slow-roll parameter $\epsilon$ and the
field variable of the inflaton spends
\begin{align}
\Delta\hat{\sigma} \equiv \hat{\sigma}_\ast
-\hat{\sigma}_{\rm end}
\simeq 0.77 M_{\rm Pl},\;\;\; \hat{\sigma}=\frac{1}{2} {\rm log}\,\sigma,
\label{eq:infcans}
\end{align}
where we canonically normalize the field
$\sigma ={\rm Re}\,T^1$.
This small-field inflation leads to the small
tensor-to-scalar ratio as can be seen in
Eq. (\ref{eq:res}).
Although the results here are completely
consistent with WMAP and the current
Planck data, the tiny tensor-to-scalar ratio
$r\sim 10^{-5}$ in Eq. (\ref{eq:res})
contradicts with the most recent data from
the BICPE2 collaborations~\cite{Ade:2014xna}.
We discuss how to realize a successful large-field
inflation in Sec. \ref{sec:linflation}, which is one
of the few candidates to generate a sizable
tensor-to-scalar ratio within the framework
of single-field slow-roll inflation.
\begin{figure}[h]
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{ps.eps}
\caption{The power spectrum of scalar curvature
perturbation between 60 and 50
e-foldings before the end of inflation.}
\label{fig:ps}
\end{figure}
\begin{figure}
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{si.eps}
\caption{The spectral index of scalar curvature
perturbation between 60 and 50
e-foldings before the end of inflation.}
\label{fig:si}
\end{figure}
\vspace{4cm}
\subsection{Reheating temperature}
\label{subsec:rehs}
Before analyzing the large-field inflation, we comment on
the process to reheat the universe.
After the inflation, the energy of the inflaton is reduced
via inflaton decay into particles in the supersymmetric
standard model, although it depends on the concrete model.
Here we roughly estimate the reheating temperature via
the decay from the
inflaton into gauge boson pairs due to the dimensional
counting.
If the particles in the supersymmetric standard
model have the $U(1)_{I'=1}$ charge for the
vector multiplet $V^{I'=1}$ carrying inflaton, there are
terms like $Z_{Q,\bar{Q}}({\rm Re}\,T)|Q|^2$ in the
K\"ahler potential with $Q$ being the matter chiral
multiplet originated from the hypermultiplet
${\bm H}_\alpha$ and $Z_{Q,\bar{Q}}({\rm Re}\,T)$
is the K\"ahler metric of $Q$ given by Eq. (\ref{eq:km})
where $c_{I'}^{(i)}$ is replaced by the $U(1)_{I'}$ charge of
$Q$.
Although these couplings may enhance the inflaton
decay width into $Q$, we will not consider them in this
paper for simplicity just assuming the vanishing
$U(1)_{I'=1}$ charges for matter fields.
\footnote{If the $U(1)_{I'=1}$ charge of $Q$ is of
${\cal O}(1)$, the decay width into $Q$ is almost
the same order as those into the gauge boson
pairs, $\Gamma (\hat{\sigma} \rightarrow g^{(a)}+g^{(a)})$.}
The couplings between moduli and gauge fields
are conducted by the gauge kinetic function $f_a(T)$,
where $a=1,2,3$ represents the gauge groups in the
minimal supersymmetric standard model (MSSM),
$U(1)_Y$, $SU(2)_L$, $SU(3)_c$ respectively.
The relevant terms in the Lagrangian are
\begin{align}
{\cal L}&\supset -\cfrac{1}{4} {\rm Re}\,f_a
F^a_{\mu\nu}F^{a\mu\nu} \nonumber\\
&= -\cfrac{1}{4}\langle {\rm Re}\,f_a \rangle
F^a_{\mu\nu}F^{a\mu\nu}
-\cfrac{1}{4}\left\langle\frac{\partial\,{\rm Re}\,f_a}
{\partial \hat{\sigma}}\right\rangle
\delta \hat{\sigma} F^a_{\mu\nu}F^{a\mu\nu},
\end{align}
where $f_a =\sum_{I'=1}^2 \xi_a^{I'} T^{I'}$ and
$\xi_a^{I'} \equiv C_{I^{'},\,J^{''}=a,\,K^{''}=a}$.
Then the total decay width of the inflaton $\hat{\sigma}$ is
approximated as
\begin{align}
\Gamma\,\simeq\,\sum_{a=1}^3
\Gamma (\hat{\sigma} \rightarrow g^{(a)}+g^{(a)})&=
\sum_{a=1}^3\cfrac{N_G^a}{128\pi} \left\langle\cfrac{\xi_{a}^1}{\sqrt{(K_{\rm eff})_{T^1T^1}}
{\rm Re}\,f_a}\right\rangle^2
\cfrac{m_{\hat{\sigma}}^3}{M_{Pl}^2}, \nonumber\\
&\simeq \,2.32 \,{\rm GeV},
\end{align}
where $\sum_{a=1}^3\,N_G^a\,=\,12$ is the number of the
gauge boson in the MSSM and $\hat{\sigma}$ is the canonically
normalized inflaton field (\ref{eq:infcans}).
We choose the $\xi_1^1=\xi_2^1=\xi_3^1=0.27$,
otherwise zero to realize the gauge coupling unification
at the grand unification scale ($\simeq \,2.0\times 10^{16} {\rm GeV}$),
\begin{align}
\text{Re}\,f_a(\langle T \rangle) =
\left(\frac{1}{g_a}\right)^2 \simeq \,3.73.
\end{align}
Then the reheating temperature is roughly estimated
by equaling the expansion rate of the universe and the total decay width,
\begin{align}
\Gamma &\simeq \,H (T_R), \nonumber\\
\Leftrightarrow T_{R} &= \left( \cfrac{\pi^2 g_\ast}{90}\right)^{-1/4}
\sqrt{\Gamma \,M_{\rm Pl}} \simeq 1.05\times 10^9\,{\rm GeV},
\end{align}
where we use $g_\ast=915/4$ which is the
effective degrees of freedom of the radiation at
the reheating in the MSSM.
We restrict ourselves to the standard situation
that the coherent oscillation of inflaton field
dominates the energy density of the universe
after the inflation. The inflaton releases the
entropy and reheats the universe when it decay.
It is then assumed that the other field does
not dominate the energy density of the universe
which is verified in Sec. \ref{subsec:modpro}.
Finally in this section, we mention about the
one-loop correction to the moduli K\"ahler
potential. The modified K\"ahler potential in
the large volume limit is found as~\cite{Sakamura:2013wia},
\begin{align}
K =-\ln\,{\cal N}
+{\cal O}\left( \frac{1}{32\pi^2{\cal N}}\right) +\cdots,
\label{eq:oneloop}
\end{align}
where the leading contribution will depend on
the number of the charged fields under the
$Z_2$-odd vector multiplets $V^{I'}$.
Even if there are such contributions in the scalar potential,
our estimation in the previous section is not
changed due to the supersymmetry condition
(\ref{eq:movev}) at the vacuum.
Since Re\,$T^1$ rolls the potential from the
large field value, one-loop effect does not
affect the inflation mechanism, which is also
confirmed by the numerical analysis.
\section{A simple model for the large-field inflation}
\label{sec:linflation}
In this section, we discuss how to realize the
large-field inflation that would explain the WMAP,
Planck~\cite{Ade:2013zuv} and BICEP2 data~\cite{Ade:2014xna},
although there is a possible tension between these collaborations.
Unlike the previous section, we consider two light pair of modulus
and stabilizer fields, e.g. $T^{I'=1,2}$ and $H_{i=1,2}$
which are decoupled from the other heavy pairs $T^{I'\neq 1,2}$ and
$H_{i\neq 1,2}$. This scenario can be realized
when $c_{I'=1,2}^{(i \ne 1,2)}=c_{I' \ne 1,2}^{(i=1,2)}=0$
in Eq.~(\ref{eq:4dw}), $\left| c_{I' \ne 1,2}^{(i \ne 1,2)} \right|
<\left| c_{I'=1,2}^{(i=1,2)} \right|$
in Eq.~(\ref{eq:thmass}), $\left|J_{0}^{i= 1,2}\right|
< \left| J_{0}^{i\neq 1,2}\right|$ and
$\left|J_{L}^{i= 1,2}\right|
< \left| J_{L}^{i\neq 1,2}\right|$ in Eq.~(\ref{eq:thmass}).
Below the heavier mass scale $m_{I' \ne 1,2\,i\ne 1,2}$,
the effective K\"ahler potential and superpotential for
the light fields $T^{I'=1,2}$ and $H_{i=1,2}$ are given
by
\begin{eqnarray}
K_{\rm eff}(T^1,H_1,T^2,H_2)
&=& -\ln {\cal N}({\rm Re}\,T)\,\Big|_0
+Z_{1,\bar{1}}
({\rm Re}\,T)\,\Big|_0 \,|H_1|^2
+Z_{2,\bar{2}}
({\rm Re}\,T)\,\Big|_0 \,|H_2|^2,
\nonumber \\
W_{\rm eff}(T^1,H_1,T^2,H_2)
&=& \left( J_0^{(1)}
+e^{-c_{I'}^{(1)}T^{I'}}J_L^{(1)} \right) H_1
+\left( J_0^{(2)}
+e^{-c_{I'}^{(2)}T^{I'}}J_L^{(2)} \right) H_2,
\label{eq:weff2}
\end{eqnarray}
where
$f(T^{I'},H_i)\,\Big|_0 \equiv f(T^{I'},H_i)
\Big|_{\scriptsize \begin{array}{ll}
T^{I' \ne 1,2}=\langle T^{I' \ne 1,2} \rangle \\
H_{i \ne 1,2}=\langle H_{i \ne 1,2} \rangle
\end{array}}$
for an arbitrary function
$f(T^{I'},H_i)$, and then
\begin{eqnarray}
Z_{1,\bar{1}}({\rm Re}\,T)\,\Big|_0 =
\frac{1-e^{-2c_{I'}^{(1)}{\rm Re}\,T^{I'}}}{
c_{I'}^{(1)} {\rm Re}\,T^{I'}}, \,\,\,\,
Z_{2,\bar{2}}({\rm Re}\,T)\,\Big|_0 =
\frac{1-e^{-2c_{I'}^{(2)}{\rm Re}\,T^{I'}}}{
c_{I'}^{(2)} {\rm Re}\,T^{I'}}.
\label{eq:kml}
\end{eqnarray}
The effective potential for the light fields,
\begin{eqnarray}
V_{\rm eff}(T^1,H_1,T^2,H_2) &=& e^{K_{\rm eff}}
\left( (K_{\rm eff})^{m,\bar{n}}\,
D_mW_{\rm eff}\,D_{\bar{n}} \bar{W}_{\rm eff}
-3|W_{\rm eff}|^2 \right),
\label{eq:4dspeff2}
\end{eqnarray}
is obtained by using the effective K\"ahler potential
and superpotential~(\ref{eq:weff2}),
where $m,n=\{I',i\}$ with $I'=1,2$ and $i=1,2$.
The most general form of part of norm function
carrying the two light moduli $T^1$ and $T^2$ is written as
\begin{eqnarray}
{\cal N} ({\rm Re} T)\Big|_0 &=& C_{1,1,1}({\rm Re} T^1)^3 +
C_{1,1,2}({\rm Re} T^1)^2({\rm Re} T^2)
+C_{1,2,2} ({\rm Re} T^1)({\rm Re} T^2)^2
+C_{2,2,2} ({\rm Re} T^2)^3 \nonumber\\
&+& \cdots,
\label{eq:gnorm}
\end{eqnarray}
where the ellipsis stands for terms those do not contain
the two light moduli.\footnote{
We assume the couplings between the decoupled fields
$T^{I'}$ with $I^{'}=3,4,\cdots$ and the lighter fields
$T^1$ and $T^2$ are absent in the norm function for simplicity.}
To brighten the outlook for analyzing the
above scalar potential (\ref{eq:4dspeff2}),
we redefine the modulus field as
\begin{equation}
\;\hat{T^1} \equiv \frac{c_1^{(1)} T^1+c_2^{(1)} T^2}{c},\;\;\;
\;\hat{T^2} \equiv \frac{c_1^{(2)} T^1+c_2^{(2)} T^2}{d}.
\end{equation}
where $c$ and $d$ are the $U(1)$ charge of the stabilizer field
$H_1$ and $H_2$ for a linear combination of the $Z_2$-odd vector
fields $A_M^{I'}$ in ${\bf V}^{I'}$ with $I'=1,2$, respectively.
In this field base $\hat{T}^1$ and $\hat{T}^2$, the
mixing terms between $\hat{T}^1$ and $\hat{T}^2$ in the
superpotential are canceled and thus each of
$\hat{T}^1$ and $\hat{T}^2$ has the independent
superpotential to each other,
\begin{equation}
W_{\rm eff}(\hat{T}^1,H_1,\hat{T}^2,H_2)
\,=\, \left( J_0^{(1)}
+e^{-c\,\hat{T}^1}J_L^{(1)} \right) H_1
+\left( J_0^{(2)}
+e^{-d\,\hat{T}^2}J_L^{(2)} \right) H_2.
\label{eq:resup}
\end{equation}
The vacuum expectation values of moduli $\hat{T}^1,
\hat{T}^2$ and stabilizer fields $H_1, H_2$ are determined by
minimizing the scalar
potential~(\ref{eq:4dspeff2}) in a similar way to those of the
small-field inflation (\ref{eq:movev}) as
\begin{eqnarray}
\langle \hat{T}^{1} \rangle
&=& \frac{1}{c} \ln \frac{J_L^{(1)}}{J_0^{(1)}}, \qquad
\langle \hat{T}^{2} \rangle \ = \
\frac{1}{d} \ln \frac{J_L^{(2)}}{J_0^{(2)}}, \qquad
\langle H_1 \rangle \ = \ \langle H_2 \rangle \ = \ 0,
\label{eq:vacuum}
\end{eqnarray}
which satisfy
$\langle D_{\hat{I}'}W \rangle
= \langle D_iW \rangle
= \langle W \rangle
= 0$
and then
$\langle V_{\hat{I}'} \rangle
= \langle V_i \rangle
= \langle V \rangle
= 0$
for
$V_m=\partial_m V$.
When we construct the large-field inflation model in the
next subsection, we restrict ourselves to the case that
the coefficients $C_{I',J',K'}$ and the charges $c_{I'}^{(i)}$
for $I',J',K'=1,2$ and $i=1,2$ are chosen in such a way
that the norm function is written as
\begin{equation}
{\cal N} ({\rm Re} T)\Big|_0= a({\rm Re} \hat{T}^1)
({\rm Re} \hat{T}^2 -b\,{\rm Re} \hat{T}^1)^2 +\cdots,
\label{eq:norm}
\end{equation}
in the hatted field base, where $a$ and $b$ are positive
real numbers determined by the fixed values of $C_{I',J',K'}$
and $c_{I'}^{(i)}$ as, e.g.,
$a=c_1^{(1)}(d)^2/c\,(c_2^{(2)})^2,b=c\,c_1^{(2)}/d\,c_1^{(1)}$
for
$c_2^{(1)}=C_{1,1,1}=C_{1,1,2}=C_{2,2,2}=0$ and $C_{1,2,2}=1$.
The ellipsis in Eq.~(\ref{eq:norm}) has same meaning as that
in Eq.~(\ref{eq:gnorm}) and is irrelevant in the following
arguments.
The advantage of the norm function (\ref{eq:norm}) will be
explained in Sec.~\ref{subsec:linflation}, and here we
notice that it leads to the moduli
mixing in the K\"ahler metric, $K_{\hat{I}^{'},\bar{\hat{J}}^{'}}
\neq 0$ for $\hat{I}^{'} \neq \hat{J}^{'}$.
We have to check the positivity of the Hessian matrix with
the scalar potential (\ref{eq:4dspeff2}). The mass
matrix given by the potential (\ref{eq:4dspeff2})
is written in a block-diagonal form with two nonvanishing
blocks according to the absence
mixing terms between the moduli $\hat{T}^{I'=1,2}$
and the stabilizer fields $H_{i=1,2}$.
Since the following mixing terms are all vanishing at
the vacuum,
\begin{eqnarray}
&\langle V_{\hat{T}^1\bar{\hat{T}}^2} \rangle =
\langle V_{\hat{T}^1\bar{H}_1} \rangle =
\langle V_{\hat{T}^1\bar{H}_2} \rangle =
\langle V_{\hat{T}^2\bar{H}_1} \rangle =
\langle V_{\hat{T}^2\bar{H}_2}\rangle =0,\nonumber\\
&\langle (K_{\rm eff})^{\hat{T}^1\bar{H}_1} \rangle =
\langle (K_{\rm eff})^{\hat{T}^1\bar{H}_2} \rangle =
\langle (K_{\rm eff})^{\hat{T}^2\bar{H}_1} \rangle =
\langle (K_{\rm eff})^{\hat{T}^2\bar{H}_2} \rangle =0,
\end{eqnarray}
for $V_{mn}=\partial_n\partial_m V$,
there are no mixing between the moduli $\hat{T}^{I'}$ and the stabilizer
fields $H_i$ at the vacuum, that is, the mass (sub-)matrices of
the moduli $\hat{T}^{I'}$ and the stabilizer fields $H_i$ can be
analyzed independently.
First, we consider the mass-squared matrix $m_{t}^2$ of the
real parts of the moduli in the base of canonically normalized
field $t^{I'}$,
\begin{equation}
t^{I'} \,=\, \sum_{J'=1}^2 \sqrt{2(K_{\hat{T}})_{I'}} U_{I',J'}
{\rm Re}\,\hat{T}^{J'},
\label{eq:canbas}
\end{equation}
for $I'=1,2$, which is estimated as
\begin{equation}
m_t^2 =
\begin{pmatrix}
\sqrt{\frac{1}{(K_{\hat{T}})_1}} & 0 \\
0 & \sqrt{\frac{1}{(K_{\hat{T}})_2}}
\end{pmatrix}
U
\begin{pmatrix}
V_{\hat{T}^1\bar{\hat{T}}^1} & 0 \\
0 & V_{\hat{T}^2\bar{\hat{T}}^2}
\end{pmatrix}
U^{-1}
\begin{pmatrix}
\sqrt{\frac{1}{(K_{\hat{T}})_1}} & 0 \\
0 & \sqrt{\frac{1}{(K_{\hat{T}})_2}}
\end{pmatrix}
,
\label{eq:masst}
\end{equation}
where $V_{\hat{T}^{I'}\bar{\hat{T}}^{J'}} =
\langle
e^{K_{\rm eff}} (K_{\rm eff})^{H_i\bar{H}_j}
W_{\hat{T}^{I'}H_i}\overline{W_{\hat{T}^{J'}H_j}}
\rangle$, and
$(K_{\hat{T}})_1$, $(K_{\hat{T}})_2$ and $U$ are the
eigenvalues and diagonalizing matrix of the moduli
K\"ahler metric, respectively.
(Explicit form of them are shown in
Appendix \ref{app:can}.)
By contrast, the mass-squared matrix
of the imaginary part,
${\rm Im}\,\hat{T}^1$ and ${\rm Im}\,\hat{T}^2$, is
already diagonalized because K\"ahler potential
does not contain the imaginary parts of moduli, those
are prohibited by the $U(1)$ gauge symmetries. The
supersymmetric masses of canonically normalized moduli
$\phi^{I'}$,
\begin{equation}
\phi^{I'} \,=\, \sqrt{2(K_{\rm eff})_{\hat{T}^{I'}\bar{\hat{T}}^{I'}}}\,{\rm Im}\,\hat{T}^{I'},
\label{eq:canbasphi}
\end{equation}
for $I'=1,2$, are also same as those shown
in Eq. (\ref{eq:thmass}) where the K\"ahler potential and
superpotential are replaced by Eq. (\ref{eq:weff2}).
Second, the mass-squared matrix
of the stabilizer field is also evaluated
in the base of canonically normalized field $h_i$,
\begin{equation}
h_i \,=\, \sum_{j=1}^2 \sqrt{2(K_{H})_i}
\delta_{i,j} H_j,
\label{eq:canbash}
\end{equation}
for $i=1,2$, that is found as
\begin{equation}
m_h^2 =
\begin{pmatrix}
\sqrt{\frac{1}{(K_H)_1}} & 0 \\
0 & \sqrt{\frac{1}{(K_H)_2}}
\end{pmatrix}
\begin{pmatrix}
V_{H_1\bar{H}_1} & V_{H_1\bar{H}_2} \\
V_{H_2\bar{H}_1} & V_{H_2\bar{H}_2}
\end{pmatrix}
\begin{pmatrix}
\sqrt{\frac{1}{(K_H)_1}} & 0 \\
0 & \sqrt{\frac{1}{(K_H)_2}}
\end{pmatrix}
,
\label{eq:massh}
\end{equation}
where $V_{H_i\bar{H}_j} =\langle
e^{K_{\rm eff}}
(K_{\rm eff})^{\hat{T}^{I'}\bar{\hat{T}}^{J'}}
W_{\hat{T}^{I'}H_i}\overline{W_{\hat{T}^{J'}H_j}}
\rangle$, and
$(K_H)_1 =\langle (K_{\rm eff})_{H_1\bar{H}_1} \rangle$,
$(K_H)_2 =\langle (K_{\rm eff})_{H_2\bar{H}_2} \rangle$
are the eigenvalues of the K\"ahler metric
of the stabilizer fields, respectively.
From the mass-squared
matrices (\ref{eq:masst}) and
(\ref{eq:massh}), the supersymmetric masses
of the canonically normalized stabilizer fields
$h_i$ and the moduli $t^{I'}$ in the limit of
$\langle W_{\hat{T}^1H_1}\rangle \ll
\langle W_{\hat{T}^2H_2}\rangle$ are estimated as,
\begin{align}
&m_{t^1}^2 \simeq
\frac{e^{\langle K_{\rm eff} \rangle}
\langle (K_{\rm eff})^{H_1\bar{H}_1}\rangle
\langle W_{\hat{T}^1H_1}\rangle^2}
{\langle (K_{\rm eff})_{\hat{T}_1\bar{\hat{T}}_1}\rangle},
\qquad
m_{t^2}^2 \simeq
e^{\langle K_{\rm eff} \rangle}
\langle (K_{\rm eff})^{H_2H_2}\rangle
\langle (K_{\rm eff})^{\hat{T}^2\bar{\hat{T}}^2}\rangle
\langle W_{\hat{T}^2H_2}\rangle^2,
\nonumber\\
&m_{{\rm Re}\,h_1}^2=m_{{\rm Im}\,h_1}^2\simeq
\frac{e^{\langle K_{\rm eff} \rangle}
\langle (K_{\rm eff})^{\hat{T}^1\bar{\hat{T}}^1}\rangle
\langle W_{\hat{T}^1H_1}\rangle^2}
{\langle (K_{\rm eff})_{H_1\bar{H}_1}\rangle}, \nonumber\\
&m_{{\rm Re}\,h_2}^2=m_{{\rm Im}\,h_2}^2\simeq
\frac{e^{\langle K_{\rm eff} \rangle}
\langle (K_{\rm eff})^{\hat{T}^2\bar{\hat{T}}^2}\rangle
\langle W_{\hat{T}^2H_2}\rangle^2}
{\langle (K_{\rm eff})_{H_2\bar{H}_2}\rangle}.
\label{eq:massd}
\end{align}
These expressions show that the squared
masses of moduli and stabilizers are all positive
at the vacuum if there is a hierarchy
$\langle W_{\hat{T}^1H_1}\rangle \ll
\langle W_{\hat{T}^2H_2}\rangle$ as mentioned
in Sec. \ref{sec:moduli}, that confirms the stability
of the vacuum~(\ref{eq:vacuum}). \footnote{If the hierarchy
$\langle W_{\hat{T}^1H_1}\rangle \ll
\langle W_{\hat{T}^2H_2}\rangle$ does not
exist, sizable K\"ahler mixings may spoil the
stability of the vacuum.}
Because the two pair $(\hat{T}^1,H_1)$ and $(\hat{T}^2,H_2)$
of modulus and stabilizer have totally independent vacuum
expectation values to each other as shown in
Eq.~(\ref{eq:vacuum}), we can further consider the situation
that the first pair ($\hat{T}^1,H_1$) is lighter than the second
pair ($\hat{T}^2,H_2$) by assuming
$\left|J_0^1\right| < \left|J_0^2\right|$ and
$\left|J_L^1\right| < \left|J_L^2\right|$.
In this case, the second pair can also be integrated out, and
the effective potential for the first pair is given by
\begin{eqnarray}
V_{\rm eff}(\hat{T}^1,H_1) &=& e^{K_{\rm eff}}
\left( (K_{\rm eff})^{m,\bar{n}}\,
D_mW_{\rm eff}\,D_{\bar{n}} \bar{W}_{\rm eff}
-3|W_{\rm eff}|^2 \right),
\label{eq:veffl}
\end{eqnarray}
where $m,n=\{\hat{I}',i\}$ with $\hat{I}'=1$ and $i=1$
and the effective K\"ahler potential $K_{\rm eff}$
and superpotential $W_{\rm eff}$ are obtained as
\begin{eqnarray}
K_{\rm eff}(\hat{T}^1,H_1)
&=& -\ln {\cal N}({\rm Re}\,\hat{T})\,\Big|_0
+Z_{1,\bar{1}}
({\rm Re}\,\hat{T})\,\Big|_0 \,|H_1|^2,
\nonumber \\
W_{\rm eff}(\hat{T}^1,H_1)
&=& \left( J_0^{(1)}
+e^{-c\,\hat{T}^{1}}J_L^{(1)} \right) H_1.
\label{eq:weff3}
\end{eqnarray}
Here we adopt the notation
$f(T^{I'},H_i)\,\Big|_0 \equiv f(T^{I'},H_i)
\Big|_{\scriptsize \begin{array}{lll}
T^{I' \ne 1,2}=\langle T^{I' \ne 1,2} \rangle \\
\hat{T}^{I' \ne 2}=\langle \hat{T}^{I' \ne 2} \rangle \\
H_{i \ne 1}=\langle H_{i \ne 1} \rangle \\
\end{array}}$
for an arbitrary function
$f(T^{I'},H_i)$.
\subsection{The Inflation potential and dynamics}
\label{subsec:linflation}
From the above scalar potential (\ref{eq:veffl}),
we find the effective potential for the modulus
$\hat{T}^1$ on the $H_1 =0$
hypersurface,
\begin{eqnarray}
V_{\rm eff}(\hat{T}^1,H_1=0)
&=& e^{K_{\rm eff}}
(K_{\rm eff})^{i=1,\bar{i}=\bar{1}}
\left| (W_{\rm eff})_{i=1} \right|^2
\,\Big|_{H_1=0}
\nonumber \\ &=&
\Lambda^4 (1-\lambda\,\text{cos}(c\,\tau )),
\label{eq:veff2}
\end{eqnarray}
where
\begin{eqnarray}
\Lambda^4 &\equiv &
\frac{c}{(\langle {\rm Re}\,\hat{T}^2\rangle -b\, \sigma)^2}
\frac{J_{01}^2 + J_{L1}^2 e^{-2c\,\sigma}}{1-e^{-2c\,\sigma}},
\nonumber \\
\lambda &\equiv & 2\cfrac{J_{01}J_{L1}e^{-c\,\sigma}}{J_{01}^2
+ J_{L1}^2 e^{-2c\,\sigma}},
\label{eq:deflambda}
\end{eqnarray}
we adopted the norm function (\ref{eq:norm})
with $a=1$ and $\hat{T}^1 = \sigma + i\tau$.
Fig. \ref{fig:linfpo} shows the scalar potential on the
$(\sigma, \tau)$-plane, where the parameters
are chosen as
\begin{equation}
c=1/30,\qquad J_L^{(1)}\,=\,-4.4\times 10^{-3},
\qquad J_0^{(1)}\,=\,4.25\times 10^{-3},\qquad b=15,
\label{parameter}
\end{equation}
in the Planck unit $M_{{\rm Pl}} =1$.
The imaginary direction $\tau$ has a periodic
property as can be seen in Eq.~(\ref{eq:veff2}) and
the real direction $\sigma$ will be stabilized at the minimum
shown in Fig. \ref{fig:linfReT1} in which
the behaviors of potential on the hypersurfaces $\tau=10$
(dot dashed line), $\tau=5$ (dashed line) and
$\tau=0$ (thick line) are drawn.
As we can see from Fig. \ref{fig:linfReT1}, the negative
region of $\sigma<0$ is not allowed, because the
$\Lambda$ in Eq. (\ref{eq:veff2}) diverges in the
limit of $\sigma \rightarrow 0$, while the overshooting
to a large-field region, $\sigma >$
$\langle {\rm Re}\,\hat{T}^2 \rangle /b$,
is also prohibited by the structure of the
norm function (\ref{eq:norm}).
Since Re\,$\hat{T}^2$ is already stabilized by its own
minimum (\ref{eq:vacuum}), we find
\begin{equation}
\lim_{{\rm Re}\,\hat{T}^1 \to \langle {\rm Re}\,\hat{T}^2
\rangle /b}
\left| V_{\rm eff}(\hat{T}^1,H_1=0) \right|
= \infty.
\end{equation}
From these properties of the potential~(\ref{eq:veff2}),
we expect that the so-called natural inflation~\cite{Freese:1990rb}
would occur by identifying $\tau={\rm Im}\,\hat{T}^1$
as the inflaton field.
During the inflation, the real part $\sigma$ will take
a different field value from the one at the true minimum~(\ref{eq:vacuum})
and after the inflation, it rolls down to the minimum and oscillates around
the vacuum.
\begin{figure}[t]
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{linfpo.eps}
\caption{The scalar potential~(\ref{eq:veff2}) on the
$(\sigma, \tau)$-plane.}
\label{fig:linfpo}
\end{figure}
\begin{figure}[t]
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{ReT1potential.eps}
\caption{The scalar potential~(\ref{eq:veff2}) on the hypersurfaces $\tau=10$
(dotdashed line), $\tau=5$ (dashed line) and $\tau=0$
(thick line).}
\label{fig:linfReT1}
\end{figure}
To confirm the above statements, we solve the equations of
motion for two fields $\sigma$ and $\tau$ under the assumption
that the oscillations of the stabilizer fields, Re\,$H_1$ and Im\,$H_1$,
around their vacuum expectation values $\langle {\rm Re}\,H_1
\rangle =\langle {\rm Im}\,H_1 \rangle =0$ are negligible
during and after the inflation (which will be confirmed
in Sec. \ref{subsec:modpro}). The equations of motion for these
fields are written as
\begin{align}
\sigma^{''} &= -\left(1-{\cal L}_{\text{kin}} \right)
\left( 3\sigma^{'} +6\frac{\sigma^2 (\langle \hat{T}^2\rangle -b\sigma)^2}{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}\frac{\partial_{\sigma}V}{V}\right) +\frac{ (\sigma^{'})^2 -(\tau^{'})^2}{\sigma (\langle \hat{T}^2\rangle -b\sigma)} \left( \frac{(\langle \hat{T}^2\rangle -b\sigma)^3 -2b^3\sigma^3}{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}\right), \nonumber\\
\tau^{''} &= -\left(1-{\cal L}_{\text{kin}} \right)\left( 3\tau^{'} +6\frac{\sigma^2 (\langle \hat{T}^2\rangle -b\sigma)^2}{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}\frac{\partial_{\tau}V}{V}\right) +\frac{ 2\sigma^{'}\tau^{'}}{\sigma (\langle \hat{T}^2\rangle -b\sigma)} \left( \frac{(\langle \hat{T}^2\rangle -b\sigma)^3 -2b^3\sigma^3}{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}\right), \nonumber\\
{\cal L}_{\text{kin}} &\equiv\frac{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}{2\sigma^2 (\langle \hat{T}^2\rangle -b\sigma)^2} \left((\sigma^{'})^2 +(\tau^{'})^2\right),
\label{eq:eqm}
\end{align}
where the prime denotes the derivative $d/dN$ with respect
to the number $N$ of e-foldings as before, and we described
the Christoffel symbol for the target space in terms of the metric $g_{\sigma\sigma}=g_{\tau\tau}=\frac{(\langle \hat{T}^2\rangle -b\sigma)^2 +2b^2\sigma^2}{2\sigma^2 (\langle \hat{T}^2\rangle -b\sigma)^2}$.
In the following analysis, the numerical values of
parameters in the Planck unit $M_{{\rm Pl}} = 1$ are chosen
as
\begin{equation}
d\,=\,1/20, \qquad J_L^{(2)}/J_0^{(2)}\,=\,-9,\qquad J_0^{(0)}\,=\,10^{-1},
\end{equation}
for the heavier fields $\hat{T}^2$ and $H_2$ as well as
those (\ref{parameter}) for the light fields $\hat{T}^1$
and $H_1$. With these parameters, the vacuum
expectation values of fields are given by
\begin{equation}
\langle \hat{T}^1\rangle \simeq 1.04, \qquad
\langle \hat{T}^2\rangle \,=\, 43.94, \qquad
\langle H_1\rangle \,=\, \langle H_2\rangle \,=\,0.
\end{equation}
For the canonically normalized fields (\ref{eq:canbas}), (\ref{eq:canbasphi})
and (\ref{eq:canbash}), their vacuum expectation values are
\begin{equation}
\langle t^1\rangle \simeq 2.29, \qquad
\langle t^2\rangle \,\simeq\, 1.22, \qquad
\langle \phi^1\rangle \,=\, \langle \phi^2\rangle
\,=\,
\langle h_1\rangle \,=\, \langle h_2\rangle \,=\,0.
\end{equation}
The supersymmetric masses (\ref{eq:thmass})
and (\ref{eq:massd})
of these fields $t^{I'}$,$\phi^{I'}$ and $h_{i}$
are estimated as
\begin{align}
&(m_{t^1})^2 \simeq (m_{\phi^1})^2 \simeq
(m_{{\rm Re}\,h_1})^2 =(m_{{\rm Im}\,h_1})^2
\simeq
\left( 1.96 \times 10^{13}\,{\rm GeV}\right)^2,
\nonumber\\
&(m_{t^2})^2 \simeq
(m_{{\rm Re}\,h_2})^2 =(m_{{\rm Im}\,h_2})^2
\simeq
\left( 4.45 \times 10^{16}\,{\rm GeV}\right)^2,
\nonumber\\
&(m_{\phi^2})^2
\simeq
\left( 3.5 \times 10^{16}\,{\rm GeV}\right)^2.
\label{eq:massvac}
\end{align}
Note that the supersymmetric masses of $\phi^{I'}$ are
in general different from those of $t^{I'}$ due to the different
canonical normalization of Re$\,\hat{T}^{I'}$
and Im$\,\hat{T}^{I'}$ from each other. The Hubble scale is given by
\begin{equation}
H_{\rm inf}\,=\,\left( V_{\rm inf}/3M_{\rm Pl}^2\right)^{1/2}
\simeq 8.6 \times 10^{13} {\rm GeV},
\label{eq:hubl}
\end{equation}
where $V_{\rm inf} \sim \Lambda^4$ is estimated by Eq. (\ref{eq:veff2}).
We check that these masses (\ref{eq:massvac})
and Hubble scale (\ref{eq:hubl}) are below the
compactification scale
$M_C \simeq
\pi M_{\rm Pl}/\langle
{\cal N}({\rm Re}\,\hat{T}) \rangle^{1/2}
\simeq 2.6 \times 10^{17} {\rm GeV}$
to ensure the validity of 4D effective-theory
description.
The pair ($\hat{T}^2,H_2$)
is stabilized at $\hat{T}^1$- and $H_1$-independent
minimum and their masses are larger than the inflaton scale,
that is, $H_{\rm inf}^2\ll
(m_{t^2}^2),\,(m_{\phi^2}^2),\,(m_{{\rm Re}\,h^2}^2),\,(m_{{\rm Im}\,h^2}^2)$.
Then the heavier pair ($\hat{T}^2,H_2$) is decoupled
from the inflation dynamics.
Next we define the slow-roll
parameters for the multi-field case~\cite{Burgess:2004kv}
to estimate the observable quantities constrained by the
cosmological observations,
\begin{align}
\epsilon &
=\,\frac{g^{\sigma\sigma}}{2}\left(
\frac{\partial_{\sigma} V}{V}\right)^2
+\frac{g^{\tau\tau}}{2}\left(
\frac{\partial_{\tau} V}{V}\right)^2, \nonumber\\
\eta &=\,\text{minimum eigenvalue of} \;\left \{
\frac{1}{V}
\begin{pmatrix}
\nabla^{i} \nabla_{j} V & \nabla^{i} \nabla_{\bar{j}} V \\
\nabla^{\bar{i}} \nabla_{j} V & \nabla^{\bar{i}} \nabla_{\bar{j}} V
\end{pmatrix}
\right \}
,\nonumber\\
&=\frac{g^{\sigma\sigma}}{2}
\left(\frac{\partial_{\sigma} \partial_{\sigma} V}{V}
+\frac{\partial_\tau \partial_\tau V}{V}
-\sqrt{\left(\frac{\partial_{\sigma} \partial_{\sigma} V}{V}
-\frac{\partial_{\tau} \partial_{\tau} V}{V}
-2\Gamma^{\sigma}_{\sigma\sigma}
\frac{\partial_{\sigma} V}{V} \right)^2
+4\left( \frac{\partial_{\sigma} \partial_{\tau} V}{V} -
\Gamma^{\sigma}_{\sigma\sigma} \frac{\partial_{\tau} V}{V} \right)^2} \right),
\end{align}
where $i,j\,=\,\sigma,\tau$.
The observables such as the power spectrum $P_\xi (k)$
of scalar curvature perturbation, its spectral index $n_s$
and the tensor-to-scalar ratio $r$
are written in terms of these slow-roll parameters,
\begin{align}
P_\xi (k)&=\frac{1}{24\pi^2} \frac{V}{\epsilon\, M_{Pl}^4},
\nonumber\\
n_s &= 1+ \frac{d\,{\rm ln} P_{\xi}(k)}{d\,{\rm ln}\,k}
\simeq 1-6\epsilon +2\eta, \nonumber\\
r &=16\,\epsilon.
\label{eq:obs}
\end{align}
We numerically solve Eq. (\ref{eq:eqm}) with the initial
conditions $(\sigma, \tau)=(1,20)$ and
$(\sigma^{'}, \tau^{'})=(0,0)$ at $N=0$ and then
Fig. \ref{fig:linfosc}
shows the evolution of $\sigma$
and $\tau$ as a function of $N$. The time at
the end of inflation corresponds to an about
$N_{\rm end}\simeq 81.2$ e-folds, when the slow-roll
condition is violated (max\,$\{\epsilon, \eta\} =1$).
Fig. \ref{fig:linfosc} confirms a desired situation
that the real part $\sigma$ of the light modulus
is fixed to a certain field value different from its
vacuum expectation value during the inflation
and oscillates around the vacuum after the inflation.
Such a dynamics is explained as follows.
The mass square of $\sigma$ consists of those from
the Hubble-induced and the supersymmetric
contributions,
\begin{align}
\partial_\sigma\partial_\sigma V &\simeq \,
3f(\sigma) H_{\rm inf}^2 + m_{\rm SUSY}^2,
\label{eq:Ret1mass}
\end{align}
where $H_{\rm inf}$
is the inflation scale
defined by Eq. (\ref{eq:hubl}),
$m_{\rm SUSY} \sim m_{t^1}
\sim {\cal O}(10^{13}\,{\rm GeV})$
is the supersymmetric mass term originating
from the superpotential (\ref{eq:weff2}) and
$f(\sigma)$ is a function of $\sigma$
whose numerical value is of ${\cal O}(1)$ during and after the
inflation. By virtue of the Hubble-induced
contribution in Eq.~(\ref{eq:Ret1mass}), the
real part $\sigma$ is ``stabilized" (at a different
point from the minimum of potential) with
its field value estimated below during the
inflation caused by the slowly rolling imaginary
part $\tau$ playing a role of inflaton field.
The ``stabilized" value of $\sigma$ during the inflation
can be estimated analytically from the approximated
equation of motion for $\sigma$ under
the slow-roll regime, $\sigma^{'}\ll 1$ and
$\tau^{'}\ll 1$,
\begin{align}
\sigma^{'} &=-g^{\sigma\sigma}\frac{V_{\sigma}}{V}
\nonumber \\
&=-g^{\sigma\sigma}\left( \frac{2}{
\langle \hat{T}^2 \rangle/b -\,\sigma} -
\frac{2c\,e^{-2c\,\sigma}}{1-e^{-2c\,\sigma}}
-c \right)
+\frac{V_{\rm vac}(\sigma)}{V},
\label{eq:trajsigma}
\end{align}
where $V_{\rm vac}(\sigma) =e^{K_{\rm eff}}
(K_{\rm eff})^{H_1\bar{H}_1}c
\left(\left|J_0^{(1)}\right|^2
-\left|J_L^{(1)}\right|^2 e^{-2c\,\sigma} \right)
/{\cal N}$
and we dropped the mixing term proportional
to $\sigma^{'} \tau^{'}$.
The field value $\sigma=\sigma_{\rm inf}$ during the inflation
is given by equaling the first parenthesis of Eq.
(\ref{eq:trajsigma}) to $0$,
\begin{align}
&\frac{2}{
\langle \hat{T}^2 \rangle/b -\,\sigma_{\rm inf}} -
\frac{2c\,e^{-2c\,\sigma_{\rm inf}}}{1-e^{-2c\,\sigma_{\rm inf}}}
-c \,=\,0, \nonumber\\
\Leftrightarrow
&\frac{\langle \hat{T}^2 \rangle}{b} \,=\,
\frac{-2+c\,\sigma_{\rm inf}
+e^{2c\,\sigma_{\rm inf}}(2+c\,\sigma_{\rm inf})}
{c\,(1+e^{2c\,\sigma_{\rm inf}})}.
\end{align}
One of the advantages of the current setup in
our model building is that we can choose
the value of $\sigma_{\rm inf}$ close to the vacuum
expectation value $\langle \sigma \rangle$ at the minimum
of potential given by Eq. (\ref{eq:vacuum}), if we employ
the parameters of the heavier modulus $\hat{T}^2$
in such a way that the following relation holds,
\begin{equation}
\frac{\langle \hat{T}^2 \rangle}{b} \,\simeq\,
\frac{-2+c\,\langle \sigma \rangle
+e^{2c\,\langle \sigma \rangle}(2+c\,\langle \sigma \rangle)}
{c\,(1+e^{2c\,\langle \sigma \rangle})},
\label{eq:parab}
\end{equation}
which is already adopted in the above numerical
analysis.
Therefore, the inflaton dynamics caused by the light
modulus field $\hat{T}^1=\sigma+i\tau$ can be
dominated by its imaginary part $\tau$, and then it
is classified as a single-field inflation which can avoid
sizable magnitudes of the isocurvature
perturbations possibly caused by the dynamics of the
other fields (most likely $\sigma$) than the inflaton.
As the inflaton $\tau$ rolls down toward the minimum
(\ref{eq:vacuum}), the real part $\sigma$ also tends
to go there,
because the value of $V$ approaches $V_{\rm vac}$
shown in Eq. (\ref{eq:trajsigma}).
The discussion here is
confirmed in Fig. \ref{fig:linftra}, where
the black dotted curve is the inflationary trajectory
on the ($\tau$, $\sigma$)-plane evaluated under
the slow-roll approximation
(\ref{eq:trajsigma}) and the red solid curve represents
the same trajectory by solving the full equations of
motion (\ref{eq:eqm}) numerically.
\begin{figure}[t]
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{linfosc.eps}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{linfoscl.eps}
\end{center}
\end{minipage}
\caption{The behavior of the $\sigma={\rm Re}\,\hat{T}^1$
(black curves) and $\tau={\rm Im}\,\hat{T}^1$ (red curves)
as a function of the e-folding number $N$.}
\label{fig:linfosc}
\end{figure}
\begin{figure}[t]
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{linftra.eps}
\end{center}
\end{minipage}
\begin{minipage}{0.5\hsize}
\begin{center}
\includegraphics[width=0.8\linewidth]{linftral.eps}
\end{center}
\end{minipage}
\caption{The inflaton trajectory
on the ($\tau-\sigma$)-plane.
(The black dotted curve is evaluated under
the slow-roll approximation (\ref{eq:trajsigma}),
while the red solid curve is drawn by solving the
full equations of motion (\ref{eq:eqm}) numerically.)}
\label{fig:linftra}
\end{figure}
From the observational point of view, the above
inflationary dynamics can be considered as a
single-field inflation if the scalar
density perturbation is successfully produced and
in this case the inflation mechanism is essentially
categorized into the so-called
natural inflation~\cite{Freese:1990rb}. (With our
parameter settings, the value of $\lambda$ defined
in Eq.~(\ref{eq:deflambda}) is almost equal to $1$ due to
$\sigma_{\rm inf} \sim \langle \sigma \rangle$
caused by the parameter choice~(\ref{parameter})).
Therefore the effective potential for the
canonically normalized field $\phi^1 \equiv k\tau$
is given by
\begin{equation}
V_{\rm eff} = \Lambda^4 (1-\lambda \,(\hat{c}\,\phi^1)),
\label{eq:infpoeff}
\end{equation}
where $\hat{c} \equiv c/k$ and
$k\equiv \sqrt{2(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}} =
\sqrt{\frac{(\langle T^2\rangle -b\,\sigma)^2 +2b^2\,\sigma^2}
{2\sigma^2 (\langle T^2\rangle -b\,\sigma)^2}}$
and the slow-roll parameters are explicitly
shown in terms of $\phi^1$ as
\begin{align}
\epsilon &= \frac{M_{\rm Pl}^2}{2}
\left(\frac{V^{'}}{V}\right)^2
=\frac{(\hat{c}\,M_{\rm Pl})^2}{2} \lambda^2
\frac{1-{\rm cos}^2 (\hat{c}\,\phi^1)}
{\left( 1-\lambda\,{\rm cos} (\hat{c}\,\phi^1)
\right)^2}, \nonumber\\
\eta &=M_{\rm Pl}^2
\frac{V^{''}}{V}
=(\hat{c}\,M_{\rm Pl})^2 \lambda
\frac{{\rm cos} (\hat{c}\,\phi^1)}
{1-\lambda\,{\rm cos} (\hat{c}\,\phi^1)}
\end{align}
those yield
\begin{align}
r&=16\,\epsilon, \nonumber\\
\xi^2 &=M_{\rm Pl}^4
\frac{V^{'} V^{'''}}{V^2}
=-2(\hat{c}\,M_{\rm Pl})^2 \epsilon,
\end{align}
where the prime denotes the derivative $d/d\phi^1$
with respect to the canonically normalized inflaton
field $\phi^1$.
The axion identified as the inflaton in the
terminology of the natural inflation here
corresponds to the zero mode of fifth component of
the $U(1)_{I'=1}$ gauge field, $A_y^{I'=1}$,
in our framework of 5D supergravity models,
and here the axion decay constant $f_{\phi^1}$ is
given by $\hat{c}^{-1}$. Although we need
the large axion decay constant $f_{\phi^1} \geq M_{\rm Pl}$
in order to get the large tensor-to-scalar ratio
$r\sim{\cal O}(0.1)$ in the natural inflation,
this large axion decay constant is obtained from the
small $U(1)_{I'=1}$ charge $c$ shown in Eq.~(\ref{parameter})
in our framework.
In addition to the natural realization of the large axion decay
constant, the $\eta$ problem peculiar to the general
four-dimensional supergravity models is avoided here,
because the K\"ahler potential does not
include the axion field $\tau$ whose appearance is prohibited
by the $U(1)_{I'=1}$ symmetry.
In the same way as the case of small-field
inflation, we denote the field values
$(\sigma,\tau) =(\sigma_\ast,\tau_\ast)$
corresponding to the pivot scale,
the number of e-foldings $N=N_\ast$ and
the height of scalar potential $V=V_\ast
\equiv V(\sigma_\ast,\tau_\ast)$ at the pivot scale
as well as $V=V_{\rm end} \equiv V(\sigma_{\rm end}, \tau_{\rm end})$
at the end of inflation. In terms of them,
the following e-foldings number
$N_{\rm e} \equiv N_{\rm end}-N_\ast$ can be
estimated as~\cite{Liddle:1993fq}
\begin{align}
N_{\rm e} \simeq 62 +
\text{ln} \frac{V_{\ast}^{1/4}}{10^{16}\,{\rm GeV}}
+\text{ln} \frac{V_\ast^{1/4}}{V_{\rm end}^{1/4}}
-\frac{1}{3} \text{ln} \frac{V_{\rm end}^{1/4}}{\rho_R^{1/4}},
\label{eq:efold2real}
\end{align}
where we used $V_{\rm end}^{1/4}
\simeq 1.9 \times 10^{16}\,{\rm GeV}$,
$\rho_R^{1/4} =(\pi^2 g_\ast/30)T_R \simeq
2.0 \times 10^{10}\,{\rm GeV}$.
The effective degrees of freedom
of the radiation $g_\ast=915/4$
at the reheating temperature $T_R$
can be fixed by assuming the MSSM with
$T_R \simeq 6.8 \times 10^{9}\,{\rm GeV}$
whose numerical value will be determined
later in Sec. \ref{subsec:rehl}. On the other hand,
the same e-folding number $N_{\rm e}$ is also
evaluated by
\begin{align}
N_{\rm e} = -\int^{t_\ast}_{t_{\text{end}}} H(\tilde{t})\,d\tilde{t},
\label{eq:efold2ast}
\end{align}
therefore we find the numerical values $\sigma_{\ast}$,
$ \tau_\ast$ and $V_{\ast}^{1/4}$ by equaling
Eq. (\ref{eq:efold2real}) to Eq. (\ref{eq:efold2ast}),
\begin{equation}
\sigma_\ast \,\simeq\, 0.98,\qquad \tau_\ast \,\simeq\, 17.9, \qquad
V_\ast^{1/4} \,\simeq\, 1.9\times 10^{16}\,{\rm GeV},
\qquad N_{\ast}\,=\,16.6,
\qquad N_{\rm e} \,=\, 64.6.
\end{equation}
Next, we check whether the power spectrum $P_\xi$ of scalar
curvature perturbation, its spectral index $n_s$, the running
of its spectral index $dn_s/d\,\ln k$
and the tensor-to-scalar ratio $r$, all at
the pivot scale $k_0$, can be consistent with the
recent observations or not.
Especially, the BICEP2 collaboration~\cite{Ade:2014xna} has
reported that a large value of
the tensor-to-scalar ratio,
\begin{equation}
r\,=\,0.16^{+0.06}_{-0.05},
\end{equation}
after considering the foreground dust. We extract
the numerical values of these observables from our
model as follows,
\begin{equation}
P_\xi \,=\,2.18 \times 10^{-9},\,\,\,
n_s \,=\, 0.967, \,\,\,
dn_s/d\ln k \,=\, -5.3\times 10^{-4}, \,\,\,
r\,=\,0.12,
\end{equation}
where the running of the spectral
index is
defined by ${\rm d}n_s/{\rm d}\,{\rm ln}\,k =-24\epsilon^2
+16\epsilon\,\eta -2\xi^2$.
These results of inflaton dynamics are summarized in
Figs. \ref{fig:psl}, \ref{fig:sil} and \ref{fig:tts}.
Note that our estimations are consistent with recent studies
~\cite{Freese:2014nla} reporting the consistency of the natural
inflation with recent observations.
These predictions are similar to those of the chaotic inflation,
because the scalar potential (\ref{eq:infpoeff})
is similar to that of the chaotic inflation~\cite{Linde:1983gd}
in the parameter region of the large axion decay constant
$f_{\phi^1}=\hat{c}^{-1}\ge M_{\rm Pl}$.
Note that this inflation mechanism is classified as the so-called
large-field inflation whose change of the canonically normalized inflaton
field $\phi^1$ is given by
\begin{equation}
\Delta \phi^1 \,= \,\phi^1_\ast -\phi^1_{\rm end}
\simeq 14.6\,M_{\rm Pl}.
\end{equation}
This kind of natural inflation scenario with the large
axion decay constant is also discussed in Ref.~\cite{Kim:2004rp},
where the large axion decay constant is effectively generated
from sub-Planckian decay constants.
In the following section \ref{subsec:modpro}
and \ref{subsec:rehl}, we discuss the field oscillation
during inflationary era and the moduli-induced
gravitino problem via the inflaton decay and
the reheating process.
\begin{figure}[h]
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{psl.eps}
\caption{The power spectrum of scalar curvature
perturbation between 70 and 60
e-foldings before the end of inflation.}
\label{fig:psl}
\end{figure}
\begin{figure}[h]
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{sil.eps}
\caption{The spectral index of scalar curvature
perturbation between 70 and 60
e-foldings before the end of inflation.}
\label{fig:sil}
\end{figure}
\begin{figure}
\centering \leavevmode
\includegraphics[width=0.5\linewidth]{tts.eps}
\caption{The tensor-to-scalar ratio
between 70 and 60
e-foldings before the end of inflation.}
\label{fig:tts}
\end{figure}
\vspace{2cm}
\subsection{Moduli problem}
\label{subsec:modpro}
In this section, we consider the cosmological
moduli problem~\cite{Lyth:1995ka}
such as moduli-induced gravitino problem~\cite{Endo:2006zj}
and the effect of field oscillation after the inflation.
In our both inflation models,
the moduli do not induce the SUSY breaking,
therefore they do not decay into the gravitino
which means that there is no moduli-induced
gravitino problem.
Even if there is a source of the SUSY breaking in
the superpotential, the moduli will not get the F-term
because they have large supersymmetric masses.
In addition to the above issues, we have to check the field oscillation after
inflationary era, because if the fields other than the inflaton
oscillate after the inflation and dominate
the universe, they affect the particle cosmology.
Since both $\hat{T}^2$ and $H_2$ have a supersymmetric
mass larger than the inflaton mass, we expect that these fields
do not oscillate.
By contrast, the stabilzer field $H_1$ and
the inflaton get the same order of the supersymmetric
mass as each other at the vacuum, and then $H_1$ could
have been stabilized at a point different from its true
minimum during the inflation
and oscillated around the minimum after the inflation.
In the inflationary era however,
the mass of $H_1$ is given by the Hubble-induced mass
proportional to $H_{\rm inf}$ shown in Eqs.~(\ref{eq:sfhinf})
and (\ref{eq:hubl}) in the small- and the large-field inflation
scenarios proposed in the previous Sec.~\ref{sec:sinflation}
and here in this Sec.~\ref{sec:linflation}, respectively.
Therefore, in each scenario, $H_1$ is fixed strictly
at the origin during and after inflation and does not oscillate
and dominate the universe.
We further remark that, in the case of small-field inflation
discussed in Sec.~\ref{sec:sinflation},
the imaginary part of modulus Im\,$T^1$ does not oscillate
as well if the initial position of Im\,$T^1$ is located at the
origin. This is because
the K\"ahler potential has a shift symmetry for the
imaginary part and the inflationary dynamics does not
involve the imaginary direction.
In order to estimate the effects of supersymmetry
breaking on the inflation dynamics,
we consider the following superpotential,
\begin{align}
W &=W_{\rm eff} +\Delta W (\hat{T}^1),
\label{eq:supbrea}
\end{align}
where $W_{\rm eff}$ represents the superpotential
terms responsible for the inflation given in
Eq.~(\ref{eq:weff3}), and
$\Delta W(\hat{T}^1)$ describes the supersymmetry
breaking sector, which may involve the inflaton
multiplet $\hat{T}^1$ in general.
We assume that the other
fields such as $\hat{T}^2$ and $H_2$
are stabilized at their supersymmetric minimum.
The following analysis can be applied to
both the inflation scenarios by replacing
the original effective superpotential $W_{\rm eff}$
with the modified one (\ref{eq:supbrea}) in each
scenario.
The position of the supersymmetry breaking minimum
will be determined by estimating the deviation from the
supersymmetric Minkowski minimum (\ref{eq:movev})
by assuming
$\langle \Delta W \rangle \sim
\langle \partial_{\hat{T}^1}(\Delta W \rangle) \ll 1$
(in the unit $M_{\rm Pl}=1$)
and employing the reference point method~\cite{Abe:2006xp}.
As the reference point which should be
selected as close to the true minimum as
possible,
we set it in such a way that the following conditions,
\begin{align}
&D_{H_1} W|_{\rm ref} = W_{H_1} +(K_{\rm eff})_{H_1} W =0, \nonumber\\
& \hspace{2cm} \leftrightarrow c\,\hat{T}^1|_{\rm ref}
=\text{ln} \frac{J_L^{(1)}}{J_0^{(1)}} \,\,{\rm and}
\;\;H_1|_{\rm ref}=0, \nonumber\\
&D_{\hat{T}^1} W|_{\rm ref} = (K_{\rm eff})_{\hat{T}^1} \Delta W,
\end{align}
are satisfied at the point, where the effective K\"ahler
potential $K_{\rm eff}$ is given in Eqs.~(\ref{eq:weff}) or
(\ref{eq:weff3}) for each scenario. Then we expand the field
$\varphi=\varphi |_{\rm ref} +\delta \varphi$ for
$\varphi=\hat{T}^1, H_1$
and evaluate
the deviations $\delta \varphi$ from the reference point
$\varphi |_{\rm ref}$.
We find the following variations,
\begin{align}
\delta \hat{T}^1 ={\cal O}\left( \frac{|\Delta W|^2}{W_{\hat{T}^1H_1}} \right),
\;\;\; \delta H_1= -\frac{(K_{\rm eff})_{\hat{T}^1} \Delta W}{W_{\hat{T}^1H_1}} +{\cal O}(|\Delta W|^2),
\end{align}
minimize the scalar potential at their first order,
which implies our reference point method is
valid if the supersymmetry breaking scale is smaller than
the supersymmetric masses of the moduli and stabilizers, that is,
$\langle \Delta W \rangle \ll \langle W_{\hat{T}^1H_1}\rangle$
in the unit $M_{\rm Pl}=1$. In the same way,
the F-terms of $\hat{T}^1$ and $H_1$ at the supersymmetry
breaking minimum are estimated as
\begin{align}
\sqrt{(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}} F^{\hat{T}^1} &=
-e^{K_{\rm eff}/2} \sqrt{(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}}
(K_{\rm eff})^{\hat{T}^1\bar{J}}\,\overline{D_JW}
\simeq {\cal O}\left(\frac{(m_{3/2})^3}{(m_{t^1})^2}\right),\nonumber\\
\sqrt{(K_{\rm eff})_{H_1\bar{H}_1}} F^{H_1} &=
-e^{K_{\rm eff}/2} \sqrt{(K_{\rm eff})_{H_1\bar{H}_1}}
(K_{\rm eff})^{H_1\bar{J}}\,\overline{D_JW}
\simeq {\cal O}\left(\frac{(m_{3/2})^3}{(m_{h_1})^2}\right),
\end{align}
where $(m_{t^1})^2$ and $(m_{h_1})^2$
are given in Eq.~(\ref{eq:massd}).
We conclude that
if the size of supersymmetry breaking is much smaller
than the inflation scale which we assume in this paper,
they do not affect the
inflation mechanism and the related cosmology after
the inflation. In fact, the field $\hat{T}^1$ and $H_1$
have almost vanishing F-terms which means that
the decay channels from $\hat{T}^1$ and $H_1$
into the gravitino are suppressed and they do not
induce the moduli-induced gravitino problem.
The coherent oscillation of $H_1$ after the inflation
is also suppressed, because the amplitude of the
oscillation of $H_1$,
\begin{align}
&\Delta H_1\,\simeq \, \delta H_1|_{\rm inf} -
\delta H_1|_{\rm vac}\,\simeq\,
{\cal O}\left( \frac{\Delta W}{H_{\rm inf}} \right)
-{\cal O}\left( \frac{\Delta W}{m_{h_1}} \right),
\end{align}
is small enough, where $\delta H_1|_{\rm inf}$ and
$\delta H_1|_{\rm vac}$ are the deviations of $H_1$
from the supersymmetric Minkowski minimum~(\ref{eq:movev})
during the inflation and at the true minimum where the
supersymmetry is broken, respectively.
\subsection{Reheating temperature}
\label{subsec:rehl}
Finally, we show the decay channel and the reheating
process after the end of inflation.
As shown in Fig. \ref{fig:linftra}, after the inflation,
both $\sigma$ and $\tau$ oscillate around the
minimum and they decay into particles in the MSSM
at the time
$t_{\rm dec}^{t^1}$ and $t_{\rm dec}^{\phi^1}$
respectively where $t^1$ and $\phi^1$ are the
canonically normalized field,
$t^1\,=\,\sqrt{2(K_{\hat{T}})_1}U_{1,1}\sigma +
\sqrt{2(K_{\hat{T}})_1}U_{1,2}({\rm Re}\,\hat{T}^2)$,
and $\phi^1 =\sqrt{2(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}}\tau$ given in Eqs.~(\ref{eq:canbas})
and (\ref{eq:canbasphi}), respectively.
Note that the eigenvalue $(K_{\hat{T}})_{I'}$
and diagonalizing matrix $U_{I',J'}$ ($I'=1,2$)
of the K\"ahler metric
are explicitly shown in Appendix
\ref{app:can}. In the following analysis,
we neglect the oscillation of Re\,$\hat{T}^2$
and use the sudden-decay approximation.
The decay time $t_{\rm dec}^{t^1}\,=\,1/\Gamma^{t^1}$
is the inverse of the decay width of $t^1$ which depends
on the concrete model of the particle physics.
We assume that the modulus $t^1$ mainly
decay into the gauge boson pairs, $t^1 \to g^{(a)}+g^{(a)}$, for simplicity.
If some matter chiral multiplets $Q$
originating in the hypermultiplet ${\bm H}_\alpha$
have $U(1)_{I'=1,2}$ charges under the $Z_2$-odd
vector multiplets $V^{I'=1,2}$ carrying the inflaton,
they induce couplings like $Z(t^1)|Q|^2$ in
the K\"ahler potential where
$Z(t^1)$ is the K\"ahler metric of $Q$ given by
Eq. (\ref{eq:km}) where the $U(1)_{I'=1,2}$ charges are
replaced by those of $Q$.
These couplings will enhance the inflaton decay width into $Q$
depending on the $U(1)_{I'=1,2}$ charge of $Q$.
\footnote{If the $U(1)_{I'=1,2}$ charge of $Q$ is of
${\cal O}(1)$,
the decay width into $Q$ is of almost the same order as that
of $\Gamma^{t^1} (t^1 \rightarrow g^{(a)}+g^{(a)})$.}
In our model, the decay channel via the
F-term of $t^1$ is kinematically forbidden,
because its vacuum expectation value is negligibly small
as mentioned previously.
After all, $t^1$ mostly decays into the gauge boson pairs
with the following decay width,
\begin{align}
\sum_{a=1}^3
\Gamma^{t^1} (t^1 \rightarrow g^{(a)}+g^{(a)})&\simeq
\sum_{a=1}^3\cfrac{N_G^a}{64\pi }
\left\langle\cfrac{\xi_a^1}{{\rm Re}\,f_a}\right\rangle^2
\left\langle
\frac{U_{2,2}}{\sqrt{2(K_{\hat{T}})_1}(U_{1,1}U_{2,2}-
U_{1,2}U_{2,1} )} \right\rangle^2
\cfrac{m_{t^1}^3}{M_{\rm Pl}^2}
\nonumber\\
&\simeq 0.86\,\,{\rm GeV},
\label{eq:t1dec}
\end{align}
where $N_G^a$ is the number of the gauge bosons
for the gauge group $G^a$ with $a=1,2,3$ representing
the three gauge groups in the MSSM, $a=\,U(1)_Y$, $SU(2)_L$,
$SU(3)_c$, respectively. We are adopting the numerical values
of input parameters, those yield
$\langle \sqrt{2(K_{\hat{T}})_1}\rangle \simeq 0.86$,
$\langle U_{1,1}\rangle \simeq -39.68$,
$\langle U_{2,1}\rangle \simeq 0.025$,
$\langle U_{1,2}\rangle \,=\,
\langle U_{2,2}\rangle \,=\,1$ and
$m_{t^1} \simeq 1.96\times 10^{13}\,{\rm GeV}$.
Especially we set $\xi_a^1=3.58$ and $\xi_a^{I' \ne 1}=0$
to realize the correct gauge coupling
$\langle f_a\rangle \,=\,1/(g_a)^2 \simeq 3.73$
at the grand unification scale
($\simeq\,2.0\times10^{16} [\text{GeV}]$).
Because
$\Gamma^{t^1} (t^1 \rightarrow g^{(a)}+g^{(a)})$
with $a=1,2,3$ are assumed to be dominant, the total decay width is
given by
\begin{align}
\Gamma^{t^1} \simeq
\sum_{a=1}^3
\Gamma^{t^1} (t^1 \rightarrow g^{(a)}+g^{(a)}).
\end{align}
On the other hand, the decay time $t_{\rm dec}^{\phi^1}
\,=\,1/\Gamma^{\phi^1}$
is estimated from the following terms in the Lagrangian,
\begin{align}
{\cal L}&\supset -\cfrac{1}{8}
{\rm Im}\,f_a\epsilon^{\mu\nu\rho\sigma} F^a_{\mu\nu}F^a_{\rho\sigma}
\nonumber\\
&=-\cfrac{1}{8}
\langle {\rm Im}\,f_a\rangle\epsilon^{\mu\nu\rho\sigma}
F^a_{\mu\nu}F^a_{\rho\sigma}
-\cfrac{1}{8}\left\langle\cfrac{\partial\,{\rm Im}\,f_a}{\partial \phi^1}\right\rangle
\delta \phi^1 \,\epsilon^{\mu\nu\rho\sigma} F^a_{\mu\nu}F^a_{\rho\sigma}.
\end{align}
Then the total decay width of the field $\phi$ is computed as follows,
\begin{align}
\Gamma^{\phi^1}\simeq
\sum_{a=1}^3
\Gamma^{\phi^1} (\phi^1 \rightarrow g^{(a)}+g^{(a)})&\simeq
\sum_{a=1}^3\cfrac{N_G^a}{128\pi }
\left\langle\cfrac{\xi_a^1}{\sqrt{(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}}{\rm Re}\,f_a}\right\rangle^2
\cfrac{m_{\phi^1}^3}{M_{\rm Pl}^2}
\nonumber\\
&\simeq 1359\,\,{\rm GeV},
\label{eq:phidec}
\end{align}
where the given input parameters lead to
$\sqrt{(K_{\rm eff})_{\hat{T}^1\bar{\hat{T}}^1}} \simeq 0.61$ and
$m_{\phi^1} \simeq 1.96\times 10^{13}\,{\rm GeV}$.
Since both the fields $t^1$ and $\phi^1$ have the almost degenerate
supersymmetric masses (\ref{eq:massvac}),
the differences
between $\Gamma^{t^1}$ shown in Eq.~(\ref{eq:t1dec})
and $\Gamma^{\phi^1}$ in Eq.~(\ref{eq:phidec}) come from the
K\"ahler metric when the fields $\sigma$ and $\tau$ are
diagonalized.
From the expressions (\ref{eq:t1dec}) and (\ref{eq:phidec}),
we find the decay time $t^{\phi^1}$ is much smaller than $t^{t^1}$,
i.e., $t^{\phi^1} \ll t^{t^1}$. It indicates that the inflaton $\phi^1$
decay into the radiation faster than the decay of
the real part of the modulus $t^1$ into the radiation.
Then the reheating temperature is estimated by
equaling the expansion rate of the universe and the total decay width,
\begin{align}
\Gamma^{\phi^1}&\simeq H (T_R), \nonumber\\
\Leftrightarrow T_{R} &=
\left( \cfrac{\pi^2 g_\ast}{90}\right)^{-1/4}
\sqrt{\Gamma^{\phi^1} M_{\rm Pl}} \simeq 6.8\times 10^9\,{\rm GeV},
\end{align}
where $g_\ast =915/4$ is the effective degrees of
freedom of the radiation at the reheating in the MSSM.
Since $t^1$ behaves as the non-relativistic
particle after the inflation, its energy density decreases as $a^{-3}$
compared to that of the radiation $a^{-4}$, where $a$ is the scale factor.
Thus whether there is a second reheating or not after $t^1$
decays depends on the following condition. If the following
condition is satisfied, $t^1$ dominates the universe and
it induces the second reheating,
\begin{equation}
1 \leq \frac{\rho_{t^1}}{\rho_R}\biggl|_{T=T_{t^1}} \simeq
\frac{\rho_{t^1}}{\rho_R}\biggl|_{T=T_R}
\left( \frac{T_{R}}{T_{t^1}} \right),
\label{eq:dom}
\end{equation}
where $\rho_{t^1}$ and $\rho_R$ are the energy densities
of $t^1$ and the radiation, respectively, and
$T_{t^1}$ is the decay temperature of $t^1$ given by
\begin{align}
T_{t^1} &=
\left( \cfrac{\pi^2 g_\ast}{90}\right)^{-1/4}
\sqrt{\Gamma^{t^1} M_{\rm Pl}} \simeq 1.7\times 10^8\,
{\rm GeV}.
\end{align}
After the inflation, the field $t^1$ and the inflaton $\phi^1$
oscillate at the same time and the difference between
them is only the size of the decay
width. Thus we expect that the amplitude of
$\Delta t^1$ is small enough at the time $t^{\phi^1}$
and the energy density $\rho_{t^1}\simeq
m_{t^1}^2(\Delta t^1)^2$ is neglected compared
to that of the radiation
$\rho_R^{1/4}\,=\,2\times 10^{10}\,{\rm GeV}$.
It follows that the above condition~(\ref{eq:dom})
is not satisfied, and then the second reheating does
not occur.
Finally we comment on the one-loop corrections to the
moduli K\"ahler potential given by Eq. (\ref{eq:oneloop}).
Although the loop correction to the effective K\"ahler
potential depends on the moduli Re\,$\hat{T}^{I'}$ with $I'=1,2$
via the Norm function ${\cal N}$ shown in Eq~(\ref{eq:norm}),
the Re\,$\hat{T}^1$-dependence of the potential is similar to that
of the tree-level K\"ahler potential. Since the scalar potential
also diverges in the limit $b\,{\rm Re}\,\hat{T}^1
\rightarrow {\rm Re}\,\hat{T}^2$ due to the behavior of the
following factor in this limit,
\begin{equation}
e^{K_{\rm eff}} \simeq \frac{e^{1/(32\pi^2 {\cal N})}}{{\cal N}}
\rightarrow \infty,
\end{equation}
the modulus ${\rm Re}\,\hat{T}^1$ is not destabilized during
and after the inflation.
Such a behavior implies that the field Re\,$\hat{T}^1$
remains stabilized during the inflation, which is
considered as the single-field inflation
with the imaginary part of modulus identified as the inflaton.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we proposed the effective mechanism
to realize successful inflation to explain the
cosmological observations based on the 5D
supergravity models on $S^1/Z_2$.
In our framework, we can realize both the small-
and the large-field inflation scenarios, where
the role of inflaton is played by a linear combination
of the moduli appearing after
compactifying the fifth direction.
These two inflation scenarios would be compatible
with numerous particle physics models constructed in
5D.
In the case of the small-field inflation,
the real part of the light modulus, Re\,$T^1$,
is considered as the inflaton and the inflaton potential is induced
by the superpotential of the stabilizer field, $H_1$,
which has a localized wavefunction in the fifth
dimension.
This small-field inflaton potential is consistent with
WMAP and Planck data~\cite{Ade:2013zuv},
although they cannot explain the large tensor-to-scalar
ratio reported by BICEP2~\cite{Ade:2014xna}.
We have also studied the particle cosmology in this
inflationary scenario and find that there is no moduli
and gravitino over production.
We further presented a different setup within the
same 5D supergravity framework, realizing the
large-field inflation which produces the sizable
tensor-to-scalar ratio consistent with the results
from BICEP2.
In this scenario, the two light pairs of moduli and
stabilizer fields ($\hat{T}^{I'},H_i$) with $I',i=1,2$ are introduced
and the inflaton is identified
as the imaginary part of the lightest modulus Im\,$\hat{T}^1$.
The moduli potential is induced by the superpotential of the
stabilizer fields as in the small-field scenario.
The inflaton potential is similar to the one of
natural inflation~\cite{Freese:1990rb},
but in our framework, the axion decay constant
is given by the $U(1)$ charges originated
from the $Z_2$-odd vector multiplets carrying the inflaton field.
Both the inflation scenarios proposed in this paper are
free from the $\eta$ problem which is peculiar to the
inflationary dynamics in the four-dimensional
${\cal N}=1$ supergravity models.
In the large-field scenario, when the imaginary part of the modulus
rolls down in its potential, the real part of the modulus
will be destabilized because there is a runaway
direction in its potential in general.
However, in our model, the real part of the modulus,
Re\,$\hat{T}^1$, can be stabilized during the inflation, because of the
potential barrier is produced by the real
part of the heavier modulus, Re\,$\hat{T}^2$, in the K\"ahler potential.
After the inflation,
both the imaginary and the real part of the
modulus, Im$\,\hat{T}^1$ and Re$\,\hat{T}^1$
oscillate but only the inflaton Im\,$\hat{T}^1$ reheats the universe.
The reheating temperature is estimated
from the decay width of the inflaton
into the gauge boson pairs.
The stabilizer fields are also fixed at the origin
during (-and after-) the inflation by the
Hubble-induced and their own supersymmetric
masses. Therefore there is no cosmological moduli
problem also in the large-field scenario.
Both the proposed inflation scenarios are insensitive
to the supersymmetry breaking required by the particle
phenomenology, if the breaking scale is lower than the
inflation scale.
This is because the large supersymmetric
masses are provided from the superpotential
of charged stabilizer fields, those are controlled
by the $U(1)$ charges under the 5D vector
multiplets carrying the moduli.
The branching ratio of the moduli decaying
into gravitino is suppressed due to such the
supersymmetric masses of moduli.
The field in the supersymmetry breaking sector
may oscillate after the inflation if the size
of supersymmetry breaking is smaller than
the inflation scale.
The further model building of the particle
cosmology including the concrete matter
sectors remains as a future work.
The moduli potential as well as their kinetic terms
are strictly constrained by the symmetries in
higher-dimensional spacetime, although the moduli
behave as Lorentz scalars in the four-dimensional spacetime
with the extra-dimensions compactified.
For the inflationary dynamics proposed in this paper,
the $U(1)_{I'}$ symmetries played essential roles,
those generate the localized wavefunctions of charged
stabilizer zero-modes and then yield the suitable moduli
potential. It would be possible that the 5D supergravity
studied in this paper is derived as the 5D effective theory
of supergravities in more-than-five dimensional spacetime,
superstrings in ten-dimensions and the M-theory
in eleven-dimensions~\cite{Lukas:1998yy}.
In such cases, the coefficients $C_{I,J,K}$ in the norm function
will be related to the geometric structure of the internal space
(e.g., the intersection numbers of Calabi-Yau manifold) and the
above $U(1)_{I'}$ symmetries might originate from certain local
symmetries with the gauge fields in the higher-dimensional spacetime.
Our 5D models not only work well observationally, but also would
be theoretically instructive and extensible from the above points of view.
\subsection*{Acknowledgement}
The authors would like to thank T.~Higaki, Y.~Sakamura
and Y.~Yamada for useful discussions and comments.
The work of H.~A. was supported in part by the Grant-in-Aid for
Scientific Research No.~25800158 from the Ministry of Education,
Culture, Sports, Science and Technology (MEXT) in Japan.
H.~O. was supported in part by a Grant-in-Aid for JSPS Fellows
No. 26-7296 and a Grant for
Excellent Graduate Schools from the MEXT in Japan.
|
\section{Introduction\label{INTRO}}
Imagine a cup of coffee into which cream has just been poured. \ At first, the
coffee and cream are separated. \ Over time, the two liquids diffuse until
they are completely mixed. \ If we consider the coffee cup a closed system, we
can say that its \textit{entropy} is increasing over time, in accordance with
the second law of thermodynamics. \ At the beginning, when the liquids are
completely separated, the system is in a highly ordered, low-entropy state.
\ After time has passed and the liquids have completely mixed, all of the
initial structure is lost; the system has high entropy.
Just as we can reason about the disorder of the coffee cup system, we can also
consider its \textquotedblleft complexity.\textquotedblright\ \ Informally, by
complexity we mean the amount of information needed to describe everything
\textquotedblleft interesting\textquotedblright\ about the system. \ At first,
when the cream has just been poured into the coffee, it is easy to describe
the state of the cup: it contains a layer of cream on top of a layer of
coffee. \ Similarly, it is easy to describe the state of the cup after the
liquids have mixed: it contains a uniform mixture of cream and coffee.
\ However, when the cup is in an intermediate state---where the liquids are
mixed in some areas but not in others---it seems more difficult to describe
what the contents of the cup look like.
Thus, it appears that the coffee cup system starts out at a state of low
complexity, and that the complexity first increases and then decreases over
time. \ In fact, this rising-falling pattern of complexity seems to hold true
for many closed systems. \ One example is the universe itself. \ The universe
began near the Big Bang in a low-entropy, low-complexity state, characterized
macroscopically as a smooth, hot, rapidly expanding plasma. It is predicted to
end in the high-entropy, low-complexity state of heat death, after black holes
have evaporated and the acceleration of the universe has dispersed all of the
particles (about $10^{100}$ years from now). \ But in between, complex
structures such as planets, stars, and galaxies have developed. \ There is no
general principle that quantifies and explains the existence of
high-complexity states at intermediate times in closed systems. \ It is the
aim of this work to explore such a principle, both by developing a more formal
definition of \textquotedblleft complexity,\textquotedblright\ and by running
numerical experiments to measure the complexity of a simulated coffee cup
system. \ The idea that complexity first increases and then decreases in as
entropy increases in closed system has been suggested informally
\cite{gell-mann,carroll}, but as far as we know this is the first quantitative
exploration of the phenomenon.
\section{Background\label{BG}}
Before discussing how to define \textquotedblleft
complexity,\textquotedblright\ let's start with the simpler question of how to
define entropy in a discrete dynamical system. \ There are various definitions
of entropy that are useful in different contexts. \ Physicists distinguish
between the Boltzmann and Gibbs entropies of physical systems. \ (There is
also the phenomenological thermodynamic entropy and the quantum-mechanical
von~Neumann entropy, neither of which are relevant here.) \ The Boltzmann
entropy is an objective feature of a microstate, but depends on a choice of
coarse-graining. \ We imagine coarse-graining the space of microstates into
equivalence classes, so that each microstate $x_{a}$ is an element of a unique
macrostate $X_{A}$. \ The volume $W_{A}$ of the macrostate is just the number
of associated microstates $x_{a}\in X_{A}$. \ Then the Boltzmann entropy of a
microstate $x_{a}$ is the normalized logarithm of the volume of the associated
macrostate:
\begin{equation}
S_{\mathrm{Boltzmann}}(x_{a}):=k_{B}\log W_{A},
\end{equation}
where $k_{B}$ is Boltzmann's constant (which we can set equal to $1$). \ The
Boltzmann entropy is independent of our knowledge of the system; in
particular, it can be nonzero even when we know the exact microstate. \ The
Gibbs entropy (which was also studied by Boltzmann), in contrast, refers to a
distribution function $\rho(x)$ over the space of microstates, which can be
thought of as characterizing our ignorance of the exact state of the system.
\ It is given by
\begin{equation}
S_{\mathrm{Gibbs}}[\rho]:=-\sum_{x}\rho(x)\log\rho(x).
\end{equation}
In probability theory, communications, information theory, and other areas,
the Shannon entropy of a probability distribution $D=(p_{x})_{x}$ is the
expected number of random bits needed to output a sample from the
distribution:
\begin{equation}
H(D):=-\sum_{x}p_x\log p_{x}.
\end{equation}
We see that this is essentially equivalent to the Gibbs entropy, with a slight
change of notation and vocabulary.
Finally, in computability theory, the entropy of an $n$-bit string $x$ is
often identified with its \emph{Kolmogorov complexity} $K(x)$: the length of
the shortest computer program that outputs $x$.\footnote{A crucial fact
justifying this definition is that switching from one (Turing-universal)
programming language to another changes $K(x)$ by at most an additive
constant, independent of $x$. \ The reason is that in one language, we can
always just write a compiler or interpreter for another language, then specify
$x$ using the second language. \ Also, throughout this paper, we will assume
for convenience that the program receives $x$'s length $n$ as input. \ This
assumption can change $K(x)$ by at most an additive $O(\log n)$ term.}
\ Strings that are highly patterned---meaning low in disorder---can be
described by a short program that takes advantage of those patterns. \ For
example, a string consisting of $n$ ones could be output by a short program
which simply loops $n$ times, printing `1' each time. \ Conversely, strings
which have little regularity cannot be compressed in this way. \ For such
strings, the shortest program to output them might simply be one that
hard-codes the entire string.
Fortunately, these notions of entropy are closely related to each other, so
that one can often switch between them depending on convenience. \ The Gibbs
and Shannon entropies are clearly equivalent. \ The Boltzmann entropy is
equivalent to the Gibbs entropy under the assumption that the distribution
function is flat over microstates within the given macrostate, and zero
elsewhere--\textit{i.e.}, given the knowledge of the system we would actually
obtain via macroscopic observation. \ For a computable distribution $D$ over
$n$-bit strings, the Kolmogorov complexity of a string sampled from $D$ tends
to the entropy of $D$ \cite{livitanyi}. \ (Thus, the Kolmogorov complexity of
a sequence of random numbers will be very high, even though there is no
\textquotedblleft interesting structure\textquotedblright in it.)
Despite these formal connections, the three kinds of entropy are calculated in
very different ways. \ The Boltzmann entropy is well-defined once a specific
coarse-graining is chosen. \ To estimate the Shannon entropy $H (D)$ of a
distribution $D$ (which we will henceforth treat as identical to the
corresponding Gibbs entropy), one in general requires knowledge of the entire
distribution $D$, which could potentially require exponentially many samples
from $D$. \ At first glance, the Kolmogorov complexity $K (x)$ seems even
worse: it is well-known to be \emph{uncomputable} (in fact, computing $K (x)$
is equivalent to solving the halting problem). \ On the other hand, in
practice one can often estimate $K (x)$ reasonably well by the compressed file
size, when $x$ is fed to a standard compression program such as \texttt{gzip}.
\ And crucially, unlike Shannon entropy, Kolmogorov complexity is well-defined
even for an individual string $x$. \ For these reasons, we chose to use $K
(x)$ (or rather, a computable approximation to it) as our estimate of entropy.
Of course, \textit{none} of the three measures of entropy capture
\textquotedblleft complexity,\textquotedblright\ in the sense discussed in
Section \ref{INTRO}. \ Boltzmann entropy, Shannon entropy, and Kolmogorov
complexity are all maximized by \textquotedblleft random\textquotedblright\ or
\textquotedblleft generic\textquotedblright\ objects and distributions,
whereas a complexity measure should be low both for \textquotedblleft
simple\textquotedblright\ objects \textit{and} for \textquotedblleft
random\textquotedblright\ objects, and large only for \textquotedblleft
interesting\textquotedblright\ objects that are neither simple nor random.
This issue has been extensively discussed in the complex systems and
algorithmic information theory communities since the 1980s. \ We are aware of
four quantitative ideas for how to define \textquotedblleft
complexity\textquotedblright\ or \textquotedblleft
interestingness\textquotedblright\ as distinct from entropy. \ While the
definitions look extremely different, it will turn out happily that they are
all related to one another, much like with the different definitions of
entropy. \ Note that our primary interest here is in the complexity of a
configuration defined at a single moment in time. One may also associate
measures of complexity to dynamical processes, which for the most part we
won't discuss.
\subsection{Apparent Complexity\label{APPCOMP}}
The first notion, and arguably the one that matches our intuition most
directly, we call \textit{apparent complexity}.\footnote{Here we are using
\textquotedblleft apparent\textquotedblright\ in the sense of
\textquotedblleft directly perceivable,\textquotedblright\ without meaning to
imply any connotation of \textquotedblleft illusory.\textquotedblright} \ By
the apparent complexity of an object $x$, we mean $H\left( f\left( x\right)
\right) $, where $H$ is any of the entropy measures discussed previously, and
$f$ is some \textquotedblleft denoising\textquotedblright\ or
\textquotedblleft smoothing\textquotedblright\ function---that is, a function
that attempts to remove the \textquotedblleft incidental\textquotedblright\ or
\textquotedblleft random\textquotedblright\ information in $x$, leaving only
the \textquotedblleft interesting, non-random\textquotedblright\ information.
\ For example, if $x$\ is a bitmap image, then $f\left( x\right) $\ might
simply be a blurred version of $x$.
Apparent complexity has two immense advantages. \ First, it is simple: it
directly captures the intuition that we want something \textit{like} entropy,
but that leaves out \textquotedblleft incidental\textquotedblright%
\ information. \ For example, while the Kolmogorov complexity of a random
sequence would be very large, the apparent complexity of the same sequence
would typically be quite small, since the smoothing procedure would average
out the random fluctuations. \ Second, we can plausibly hope to
\textit{compute} (or at least, approximate) apparent complexity: we need
\textquotedblleft merely\textquotedblright\ solve\ the problems of computing
$H$ and $f$. \ It's because of these advantages that the complexity measure we
ultimately adopt for our experiments will be an approximate variant of
apparent complexity.
On the other hand, apparent complexity also has a large disadvantage: namely,
the apparent arbitrariness in the choice of the denoising function $f$. \ Who
decides which information about $x$ is \textquotedblleft
interesting,\textquotedblright\ and which is \textquotedblleft
incidental\textquotedblright? \ Won't $f$ depend, not only on the type of
object under study (bitmap images, audio recordings, etc.), but even more
worryingly, on the prejudices of the investigator? \ For example, suppose we
choose $f$\ to blur out details of an image that are barely noticeable to the
human eye. \ Then will studying the time-evolution of $H\left( f\left(
x\right) \right) $ tell us anything about $x$ itself, or only about various
quirks of the human visual system?
Fortunately, the apparent arbitrariness of the smoothing procedure is less of
a problem than might initially be imagined. \ It is very much like the need
for a coarse-graining on phase space when one defines the Boltzmann entropy.
\ In either case, these apparently-arbitrary choices are in fact
well-motivated on physical grounds. \ While one \textit{could} choose bizarre
non-local ways to coarse-grain or smooth a distribution, natural choices are
typically suggested by our physical ability to actually observe systems, as
well as knowledge of their dynamical properties (see for example
\cite{brun1999classical}). \ When deriving the equations of fluid dynamics
from kinetic theory, in principle one could choose to average over cells of
momentum space rather than in position space; but there is no physical reason
to do so, since interactions are local in position rather than momentum.
\ Likewise, when we observe configurations (whether with our eyes, or with
telescopes or microscopes), large-scale features are more easily discerned
than small-scale ones. \ (In field theory this feature is formalized by the
renormalization group.) \ It therefore makes sense to smooth configurations
over local regions in space.
Nevertheless, we would ideally like our complexity measure to \textit{tell us}
what the distinction between \textquotedblleft random\textquotedblright\ and
\textquotedblleft non-random\textquotedblright\ information consists of,
rather than having to decide ourselves on a case-by-case basis. \ This
motivates an examination of some alternative complexity measures.
\subsection{Sophistication\label{SOPH}}
The second notion---one that originates in work of Kolmogorov himself---is
\textit{sophistication}. \ Roughly speaking, sophistication seeks to
generalize Kolmogorov complexity to capture only the non-random information in
a string---while using Kolmogorov complexity itself to define what is meant by
\textquotedblleft non-random.\textquotedblright\ \ Given an $n$-bit string
$x$, let a \textit{model} for $x$ be a set $S\subseteq\left\{ 0,1\right\}
^{n}$\ such that $x\in S$. \ Let $K\left( S\right) $ be the length of the
shortest program that enumerates the elements of $S$, in any order (crucially,
the program must halt when it is done enumerating the elements). \ Also, let
$K\left( x|S\right) $\ be the length of the shortest program that outputs
$x$ given as input a description of $x$. \ Then we can consider $x$ to be a
\textquotedblleft generic\textquotedblright\ element of $S$ if $K\left(
x|S\right) \geq\log_{2}\left\vert S\right\vert -c$ for some small constant
$c$. \ This means intuitively that $S$ is a \textquotedblleft
maximal\textquotedblright\ model for $x$: one can summarize all the
interesting, non-random properties of $x$\ by simply saying that $x\in S$.
Now the $c$-\textit{sophistication} of $x$ or $\operatorname*{soph}_{c}\left(
x\right) $, defined by Koppel \cite{koppel}, is the minimum of $K\left(
S\right) $\ over all models $S$ for $x$ such that $K\left( S\right)
+\log_{2}\left\vert S\right\vert \leq K\left( x\right) +c$. \ (The optimal
such $S$ is said to \textquotedblleft witness\textquotedblright%
\ $\operatorname*{soph}_{c}\left( x\right) $.) \ In words,
$\operatorname*{soph}_{c}\left( x\right) $\ is the smallest possible amount
of \textquotedblleft non-random\textquotedblright\ information in a program
for $x$ that consists of two parts---a \textquotedblleft
non-random\textquotedblright\ part (specifying $S$)\ and a \textquotedblleft
random\textquotedblright\ part (specifying $x$ within $S$)---assuming the
program is also near-minimal. \ We observe the following:
\begin{itemize}
\item[(i)] $\operatorname*{soph}_{c}\left( x\right) \leq K\left( x\right)
+O\left( 1\right) $, since we can always just take $S=\left\{ x\right\} $
as our model for $x$.
\item[(ii)] \textit{Most} strings $x$ satisfy $\operatorname*{soph}_{c}\left(
x\right) =O\left( 1\right) $, since we can take $S=\left\{ 0,1\right\}
^{n}$\ as our model for $x$.
\item[(iii)] If $S$ witnesses $\operatorname*{soph}_{c}\left( x\right) $,
then $\log_{2}\left\vert S\right\vert \leq K\left( x\right) -K\left(
S\right) +c\leq K\left( x|S\right) +c$, meaning that $x$\ must be a
\textquotedblleft generic\textquotedblright\ element of $S$.
\end{itemize}
It can be shown (see G\'{a}cs, Tromp, and Vit\'{a}nyi \cite{gtv}\ or
Antunes\ and Fortnow \cite{soph}) that there do exist highly \textquotedblleft
sophisticated\textquotedblright\ strings $x$, which satisfy
$\operatorname*{soph}_{c}\left( x\right) \geq n-c-O\left( \log n\right) $.
\ Interestingly, the proof of that result makes essential use of the
assumption that the program for $S$ halts, after it has finished listing $S$'s
elements. \ If we dropped that assumption, then we could always achieve
$K\left( S\right) =O\left( \log n\right) $, by simply taking $S$ to be the
set of all $y\in\left\{ 0,1\right\} ^{n}$ such that $K\left( y\right) \leq
K\left( x\right) $, and enumerating those $y$'s in a dovetailing fashion.
Recently, Mota et al.\ \cite{maas}\ studied a natural variant of
sophistication, in which one only demands that $S$\ be a maximal model for $x$
(i.e., that $K\left( x|S\right) \geq\log_{2}\left\vert S\right\vert -c$),
and not that $S$ also lead to a near-optimal two-part program for $x$ (i.e.,
that $K\left( S\right) +\log_{2}\left\vert S\right\vert \leq K\left(
x\right) +c$). \ More formally, Mota et al.\ define the \textit{na\"{\i}ve
}$c$\textit{-sophistication} of $x$, or $\operatorname*{nsoph}_{c}\left(
x\right) $, to be the minimum of $K\left( S\right) $\ over all models $S$
for $x$ such that $K\left( x|S\right) \geq\log_{2}\left\vert S\right\vert
-c$. \ By point (iii) above, it is clear that $\operatorname*{nsoph}%
_{c}\left( x\right) \leq\operatorname*{soph}_{c}\left( x\right) $.
\ \textit{A priori}, $\operatorname*{nsoph}_{c}\left( x\right) $\ could be
much smaller $\operatorname*{soph}_{c}\left( x\right) $, thereby leading to
two different sophistication notions. \ However, it follows from an important
2004 result of Vereshchagin and Vit\'{a}nyi \cite{vereshvitanyi} that
$\operatorname*{soph}_{c+O\left( \log n\right) }\left( x\right)
\leq\operatorname*{nsoph}_{c}\left( x\right) $\ for all $x$, and hence the
two notions are basically equivalent.
Sophistication is sometimes criticized for being \textquotedblleft
brittle\textquotedblright: it is known that increasing the parameter $c$\ only
slightly can cause $\operatorname*{soph}_{c}\left( x\right) $\ and
$\operatorname*{nsoph}_{c}\left( x\right) $\ to fall drastically, say from
$n-O\left( \log n\right) $\ to $O\left( 1\right) $. \ However, a simple
fix to that problem is to consider the quantities $\min_{c}\left\{
c+\operatorname*{soph}_{c}\left( x\right) \right\} $\ and $\min_{c}\left\{
c+\operatorname*{nsoph}_{c}\left( x\right) \right\} $. \ Those are known,
respectively, as the \textit{coarse sophistication} $\operatorname*{csoph}%
\left( x\right) $\ and \textit{na\"{\i}ve coarse sophistication}
$\operatorname*{ncsoph}\left( x\right) $, and they satisfy
$\operatorname*{ncsoph}\left( x\right) \leq\operatorname*{csoph}\left(
x\right) \leq\operatorname*{ncsoph}\left( x\right) +O\left( \log n\right)
$.
The advantage of sophistication is that it captures, more cleanly than any
other measure, what exactly we mean by \textquotedblleft
interesting\textquotedblright\ versus \textquotedblleft
random\textquotedblright\ information. \ Unlike with apparent complexity, with
sophistication there's no need to specify a smoothing function $f$, with the
arbitrariness that seems to entail. \ Instead, if one likes, the definition of
sophistication picks out a smoothing function for us: namely, whatever
function maps $x$\ to its corresponding model $S$.
Unfortunately, this conceptual benefit\ comes at a huge computational price.
\ Just as $K\left( x\right) $\ is uncomputable, so one can show that the
sophistication measures are uncomputable as well. \ But with $K\left(
x\right) $, at least we can get better and better upper bounds, by finding
smaller and smaller compressed representations for $x$. \ By contrast, even to
\textit{approximate} sophistication requires solving two coupled optimization
problems: firstly over possible models $S$, and secondly over possible ways to
specify $x$ given $S$.
A second disadvantage of sophistication is that, while there \textit{are}
highly-sophisticated strings, the only known way to produce such a string
(even probabilistically) is via a somewhat-exotic diagonalization argument.
\ (By contrast, for \textquotedblleft reasonable\textquotedblright\ choices of
smoothing function $f$, one can easily generate $x$ for which the apparent
complexity $H\left( f\left( x\right) \right) $\ is large.) \ Furthermore,
this is not an accident, but an unavoidable consequence of sophistication's
generality. \ To see this, consider any short probabilistic program $P$: for
example, the coffee automaton that we will study in this paper, which has a
simple initial state and a simple probabilistic evolution rule. \ Then we
claim that \textit{with overwhelming probability, }$P$\textit{'s output }%
$x$\textit{\ must have low sophistication.} \ For as the model $S$, one can
take the set of all possible outputs $y$\ of $P$ such that $\Pr\left[
y\right] \approx\Pr\left[ x\right] $. \ This $S$\ takes only $O\left( \log
n\right) $\ bits to describe (plus $O\left( 1\right) $\ bits for $P$
itself), and clearly $K\left( x|S\right) \geq\log_{2}\left\vert S\right\vert
-c$\ with high probability over $x$.
For this reason, sophistication as defined above seems irrelevant to the
coffee cup or other physical systems:\ it simply never becomes large for such
systems! \ On the other hand, note that the two drawbacks of sophistication
might \textquotedblleft cancel each other out\textquotedblright\ if we
consider \textit{resource-bounded} versions of sophistication: that is,
versions where we impose constraints (possibly severe constraints) on both the
program for generating $S$, and the program for generating $x$ given $S$.
\ Not only does the above argument fail for resource-bounded versions of
sophistication, but those versions are the only ones we can hope to compute
anyway! \ With Kolmogorov complexity, we're forced to consider proxies (such
as \texttt{gzip} file size) mostly just because $K\left( x\right) $\ itself
is uncomputable. \ By contrast, even if we \textit{could} compute
$\operatorname*{soph}_{c}\left( x\right) $\ perfectly, it would never become
large for the systems that interest us here.
\subsection{Logical Depth\label{DEPTH}}
A third notion, introduced by Bennett \cite{bennett:depth}, is \textit{logical
depth}. \ Roughly speaking, the logical depth of a string $x$ is the amount of
time taken by the shortest program that outputs $x$. \ (Actually, to avoid the
problem of \textquotedblleft brittleness,\textquotedblright\ one typically
considers something like the minimum amount of time taken by any program that
outputs $x$ and whose length is at most $K\left( x\right) +c$, for some
constant \textquotedblleft fudge factor\textquotedblright\ $c$. \ This is
closely analogous to what is done for sophistication.)
The basic idea here is that, both for simple strings and for random ones, the
shortest program will \textit{also} probably run in nearly linear time. \ By
contrast, one can show that there exist \textquotedblleft
deep\textquotedblright\ strings, which can be generated by short programs but
only after large amounts of time.
Like sophistication, logical depth tries to probe the internal structure of a
minimal program for $x$---and in particular, to distinguish between the
\textquotedblleft interesting code\textquotedblright\ in that program\ and the
\textquotedblleft boring data\textquotedblright\ on which the code acts. \ The
difference is that, rather than trying to measure the \textit{size} of the
\textquotedblleft interesting code,\textquotedblright\ one examines how long
it takes to run.
Bennett \cite{bennett:depth} has advocated logical depth as a complexity
measure, on the grounds that logical depth encodes the \textquotedblleft
amount of computational effort\textquotedblright\ used to produce $x$,
according to the \textquotedblleft most probable\textquotedblright\ (i.e.,
lowest Kolmogorov complexity) hypothesis about how $x$ was generated. \ On the
other hand, an obvious disadvantage of logical depth is that it's even less
clear how to estimate it in practice than was the case for sophistication.
A second objection to logical depth is that even short, fast programs can be
extremely \textquotedblleft complicated\textquotedblright\ in their behavior
(as evidenced, for example, by cellular automata such as Conway's Game of
Life). \ Generating what many people would regard as a visually complex
pattern---and what \textit{we} would regard as a complex, milk-tendril-filled
state, in the coffee-cup system---simply need not take a long time! \ For this
reason, one might be uneasy with the use of running time as a proxy for complexity.
\subsection{Light-Cone Complexity\label{LCC}}
The final complexity measure we consider was proposed by Shalizi, Shalizi, and
Haslinger \cite{ssh}; we call it \textit{light-cone complexity}. \ In contrast
to the previous measures, light-cone complexity does not even try to define
the \textquotedblleft complexity\textquotedblright\ of a string $x$, given
only $x$ itself. \ Instead, the definition of light-cone complexity assumes a
\textit{causal structure}: that is, a collection of spacetime points
$\mathcal{A}$ (assumed to be fixed), together with a transitive, cycle-free
binary relation indicating which points $a\in\mathcal{A}$\ are to the
\textquotedblleft future\textquotedblright\ of which other points in
$\mathcal{A}$. \ The set of all points $b\in\mathcal{A}$\ to $a$'s future is
called $a$'s \textit{future light-cone}, and is denoted $F\left( a\right) $.
\ The set of all points $b\in\mathcal{A}$\ to $a$'s past (that is, such that
$a$ is to $b$'s future)\ is called $a$'s \textit{past light-cone}, and is
denoted $P\left( a\right) $. \ For example, if we were studying the
evolution of a $1$-dimensional cellular automaton, then $\mathcal{A}$\ would
consist of all ordered pairs $\left( x,t\right) $\ (where $x$ is position
and $t$ is time), and we would have%
\begin{align}
F\left( x,t\right) & =\left\{ \left( y,u\right) :u>t,~~\left\vert
x-y\right\vert \leq u-t\right\} ,\\
P\left( x,t\right) & =\left\{ \left( y,u\right) :u<t,~~\left\vert
x-y\right\vert \leq t-u\right\} .
\end{align}
Now given a spacetime point $a\in\mathcal{A}$, let $V_{a}$\ be the actual
value assumed by the finite automaton at $a$\ (for example, \textquotedblleft
alive\textquotedblright\ or \textquotedblleft dead,\textquotedblright\ were we
discussing Conway's Game of Life or some other $2$-state system). \ In
general, the finite automaton might be probabilistic, in which case $V_{a}%
$\ is a random variable, with a Shannon entropy $H\left( V_{a}\right) $\ and
so forth. \ Also, given a set $S\subseteq\mathcal{A}$, let $V_{S}:=\left(
V_{a}\right) _{a\in S}$ be a complete description of the values at
\textit{all} points in $S$. \ Then the light-cone complexity at a point
$a\in\mathcal{A}$, or $\operatorname*{LCC}\left( a\right) $, can be defined
as follows:%
\begin{align}
\operatorname*{LCC}\left( a\right) & =I\left( V_{P\left( a\right)
}:V_{F\left( a\right) }\right) \\
& =H\left( V_{P\left( a\right) }\right) +H\left( V_{F\left( a\right)
}\right) -H\left( V_{P\left( a\right) },V_{F\left( a\right) }\right) .
\end{align}
In other words, $\operatorname*{LCC}\left( a\right) $\ is the \textit{mutual
information} between $a$'s past and future light-cones: the number of bits
about $a$'s future that are encoded by its past. \ If we want the light-cone
complexity of (say) an entire spatial slice, we could then take the sum of
$\operatorname*{LCC}\left( a\right) $\ over all $a$ in that slice, or some
other combination.
The intuition here is that, if the cellular automaton dynamics are
\textquotedblleft too simple,\textquotedblright\ then $\operatorname*{LCC}%
\left( a\right) $\ will be small simply because $H\left( V_{P\left(
a\right) }\right) $\ and $H\left( V_{F\left( a\right) }\right) $\ are
both small. \ Conversely, if the dynamics are \textquotedblleft too
random,\textquotedblright\ then $\operatorname*{LCC}\left( a\right) $\ will
be small because $H\left( V_{P\left( a\right) },V_{F\left( a\right)
}\right) \approx H\left( V_{P\left( a\right) }\right) +H\left(
V_{F\left( a\right) }\right) $: although the past and future light-cones
both have plenty of entropy, they are uncorrelated, so that knowledge of the
past is of barely any use in predicting the future. \ Only in an intermediate
regime, where there are interesting \textit{non}-random dynamics, should there
be substantial uncertainty about $V_{F\left( a\right) }$\ that can be
reduced by knowing $V_{P\left( a\right) }$.
As Shalizi et al.\ \cite{ssh}\ point out, a major advantage of light-cone
complexity, compared to sophistication, logical depth, and so on, is that
light-cone complexity has a clear \textquotedblleft operational
meaning\textquotedblright: it is easy to state the question that light-cone
complexity is answering. \ That question is the following: \textquotedblleft
how much could I possibly predict about the configurations in $a$'s future,
given complete information about $a$'s past?\textquotedblright\ \ The reason
to focus on light-cones, rather than other sets of points, is that the
light-cones are automatically determined once we know the causal structure:
there seems to be little arbitrariness about them.
On the other hand, depending on the application, an obvious drawback of
light-cone complexity is that it can't tell us the \textquotedblleft
inherent\textquotedblright\ complexity of an object $x$, without knowing about
$x$'s past and future. \ If we wanted to use a complexity measure to make
\textit{inferences} about $x$'s past and future, this might be seen as
question-begging. \ A less obvious drawback arises if we consider a dynamical
system that changes slowly with time: for example, a version of the coffee
automaton where just a single cream particle is randomly moved at each time
step. \ Consider such a system in its \textquotedblleft late\textquotedblright%
\ stages: that is, after the coffee and cream have fully mixed. \ Even then,
Shalizi et al.'s $\operatorname*{LCC}\left( a\right) $\ measure will remain
large, but not for any \textquotedblleft interesting\textquotedblright%
\ reason: only because $a$'s past light-cone will contain almost the same
(random) information as its future light-cone, out to a very large distance!
\ Thus, $\operatorname*{LCC}$\ seems to give an intuitively wrong answer in
these cases (though no doubt one could address the problem by redefining
$\operatorname*{LCC}$\ in some suitable way).
The computational situation for $\operatorname*{LCC}$\ seems neither better
nor worse to us than that for (say) apparent complexity or resource-bounded
sophistication. \ Since the light-cones $P\left( a\right) $\ and $V\left(
a\right) $\ are formally infinite, a first step in estimating
$\operatorname*{LCC}\left( a\right) $---as Shalizi et al.\ point out---is to
impose some finite cutoff $t$ on the number of steps into $a$'s past and
future one is willing to look. \ Even then, one needs to estimate the mutual
information $I\left( V_{P_{t}\left( a\right) }:V_{F_{t}\left( a\right)
}\right) $\ between the truncated light-cones $P_{t}\left( a\right) $\ and
$F_{t}\left( a\right) $, a problem that na\"{\i}vely requires a number of
samples exponential in $t$. \ One could address this problem by simply taking
$t$ extremely small (Shalizi et al.\ set $t=1$). \ Alternatively, if a large
$t$ was needed, one could use the same Kolmogorov-complexity-based approach
that we adopt in this paper for apparent complexity. \ That is, one first
replaces the mutual information by the mutual \textit{algorithmic} information%
\begin{equation}
K\left( V_{P_{t}\left( a\right) }:V_{F_{t}\left( a\right) }\right)
=K\left( V_{P_{t}\left( a\right) }\right) +K\left( V_{F_{t}\left(
a\right) }\right) -K\left( V_{P_{t}\left( a\right) },V_{F_{t}\left(
a\right) }\right) ,
\end{equation}
and then estimates $K\left( x\right) $\ using some computable proxy such as
\texttt{gzip}\ file size.
\subsection{Synthesis\label{SYNTHESIS}}
It seems like we have a bestiary of different complexity notions.
\ Fortunately, the four notions discussed above can all be related to each
other; let us discuss how.
First, one can view apparent complexity as a kind of \textquotedblleft
resource-bounded\textquotedblright\ sophistication. \ To see this, let $f$ be
any smoothing function. \ Then $K\left( f\left( x\right) \right) $, the
Kolmogorov complexity of $f\left( x\right) $, is essentially equal to
$K\left( S_{f,x}\right) $, where%
\begin{equation}
S_{f,x}:=\left\{ y:f\left( y\right) =f\left( x\right) \right\}
.\label{sfx}%
\end{equation}
Thus, if instead of minimizing over \textit{all} models $S$ for $x$ that
satisfy some condition, we consider only the particular model $S_{f,x}$ above,
then sophistication reduces to apparent complexity. \ Note that this argument
establishes neither that apparent complexity is an upper bound on
sophistication, nor that it's a lower bound. \ Apparent complexity could be
larger, if the minimization found some model $S$\ for $x$ with $K\left(
S\right) \ll K\left( S_{f,x}\right) $. \ But conversely, sophistication
could also be larger, if the model $S_{f,x}$\ happened to satisfy $K\left(
x|S_{f,x}\right) \ll\log_{2}\left\vert S_{f,x}\right\vert $ (that is, $x$ was
a highly \textquotedblleft non-generic\textquotedblright\ element of $S_{f,x}$).
Second, Antunes\ and Fortnow \cite{soph} proved a close relation between
coarse sophistication and a version of logical depth. \ Specifically, the Busy
Beaver function, $\operatorname*{BB}\left( k\right) $, is defined as the
maximum number of steps for which a $k$-bit program can run before halting
when given a blank input. \ Then given a string $x$, Antunes and Fortnow
\cite{soph}\ define the \textit{Busy Beaver computational depth}
$\operatorname*{depth}_{BB}\left( x\right) $ to be the minimum, over all
programs $p$ that output a model $S$ for $x$ in $BB\left( k\right) $\ steps
or fewer, of $\left\vert p\right\vert +k-K\left( x\right) $. They then prove
the striking result that $\operatorname*{csoph}$ and $\operatorname*{depth}%
_{BB}$ are essentially equivalent: for all $x\in\left\{ 0,1\right\} ^{n}$,%
\begin{equation}
\left\vert \operatorname*{csoph}\left( x\right) -\operatorname*{depth}%
\nolimits_{BB}\left( x\right) \right\vert =O\left( \log n\right) .
\end{equation}
Third, while light-cone complexity is rather different from the other three
measures (due to its taking as input an entire causal history), it can be
loosely related to apparent complexity as follows. \ If $\operatorname*{LCC}%
\left( a\right) $\ is large, then the region around $a$ must contain large
\textquotedblleft contingent structures\textquotedblright: structures that are
useful for predicting future evolution, but that might have been different in
a different run of the automaton. \ And one might expect those structures to
lead to a large apparent complexity in $a$'s vicinity. \ Conversely, if the
apparent complexity is large, then one expects contingent structures (such as
milk tendrils, in the coffee automaton), which could then lead to nontrivial
mutual information between $a$'s past and future light-cones.
Having described four complexity measures, their advantages and disadvantages,
and their relationships to each other, we now face the question of which
measure to use for our experiment. \ While it would be interesting to study
the rise and fall of light-cone complexity in future work, here we decided to
restrict ourselves to complexity measures that are functions of the current
state. \ That leaves apparent complexity, sophistication, and logical depth
(and various approximations, resource-bounded versions, and hybrids thereof).
Ultimately, we decided on a type of apparent complexity. \ Our reason was
simple: because even after allowing resource bounds, \textit{we did not know
of any efficient way to approximate sophistication or logical depth}. \ In
more detail, given a bitmap image $x$ of a coffee cup, our approach first
\textquotedblleft smears $x$ out\textquotedblright\ using a smoothing function
$f$, then uses the \texttt{gzip}\ file size of $f\left( x\right) $\ as an
upper bound on the Kolmogorov complexity $K\left( f\left( x\right) \right)
$\ (which, in turn, is a proxy for the Shannon entropy $H\left( f\left(
x\right) \right) $\ of $f\left( x\right) $\ considered as a random
variable). \ There are a few technical problems that arise when implementing
this approach (notably, the problem of \textquotedblleft border pixel
artifacts\textquotedblright). \ We discuss those problems and our solutions to
them in Section \ref{ALGS}.
Happily, as discussed earlier in this section, our apparent complexity measure
can be related to the other measures. \ For example, apparent complexity can
be seen as an \textit{extremely} resource-bounded variant of sophistication,
with the set $S_{f,x}$\ of equation (\ref{sfx}) playing the role of the model
$S$. \ As discussed in Section \ref{APPCOMP}, one might object to our apparent
complexity measure on the grounds that our smoothing function $f$ is
\textquotedblleft arbitrary,\textquotedblright\ that we had no principled
reason to choose it rather than some other function. \ Interestingly, though,
one can answer that objection by taking inspiration from light-cone
complexity. \ Our smoothing function $f$ will \textit{not} be completely
arbitrary, for the simple reason that the regions over which we
coarse-grain---namely, squares of contiguous cells---will correspond to the
coffee automaton's causal structure.\footnote{Technically, if we wanted to
follow the causal structure, then we should have used \textit{diamonds} of
continguous cells rather than squares. \ But this difference is presumably
insignificant.}
\section{The Coffee Automaton\label{SETUP}}
The coffee cup system that we use as our model is a simple stochastic cellular
automaton. \ A two-dimensional array of bits describes the system's state,
with ones representing particles of cream, and zeros representing particles of
coffee. \ The cellular automaton implementation used for this project is
written in Python; source code is available for download.\footnote{At
\texttt{www.scottaaronson.com/coffee\_automaton.zip}}
The automaton begins in a state in which the top half of the cells are filled
with ones, and the bottom half is filled with zeros. \ At each time step, the
values in the cells change according to a particular transition rule. \ We
consider two different models of the coffee cup system, each having its own
transition rule.
\subsection{Interacting Model\label{INT}}
In the interacting model of the coffee cup system, only one particle may
occupy each cell in the state array. \ The transition rule for this model is
as follows: at each time step, one pair of horizontally or vertically
adjacent, differing particles is selected, and the particles' positions are
swapped. \ This model is \textit{interacting} in the sense that the presence
of a particle in a cell prevents another particle from entering that cell.
\ The movements of particles in this model are not independent of one another.
This model reflects the physical principle that two pieces of matter may not
occupy the same space at the same time. \ However, the interactions between
particles that make this model more realistic also make it harder to reason
about theoretically.
\subsection{Non-Interacting Model\label{NONINT}}
In the non-interacting model of the coffee cup system, any number of cream
particles may occupy a single cell in the state array. \ Coffee particles are
not considered important in this model; they are simply considered a
background through which the cream particles move. \ The transition rule for
this model is as follows: at each time step, each cream particle in the system
moves one step in a randomly chosen direction. \ This model is
\textit{non-interacting} in that the location of each cream particle is
independent of all the others. \ The presence of a cream particle in a
particular cell does not prevent another cream particle from also moving into
that cell.
We consider this model because it is easier to understand theoretically.
\ Since the particles in the system do not interact, each particle can be
considered to be taking an independent random walk. \ The dynamics of random
walks are well-understood, so it is easy to make theoretical predictions about
this model (see Appendix \ref{ANALYSIS}) and compare them to the experimental results.
\section{Approximating Apparent Complexity\label{ALGS}}
While Kolmogorov complexity and sophistication are useful theoretical notions
to model our ideas of entropy and complexity, they cannot be directly applied
in numerical simulations, because they are both uncomputable. \ As such, while
we use these concepts as a theoretical foundation, we need to develop
algorithms that attempt to approximate them.
Evans et al.\ \cite{oscr} propose an algorithm, called the optimal symbol
compression ratio (OSCR) algorithm, which directly estimates Kolmogorov
complexity and sophistication. \ Given an input string $x$, the OSCR algorithm
produces a two-part code. \ The first part is a codebook, which maps symbols
chosen from the original input string to new symbols in the encoded string.
\ The second part of the code is the input string, encoded using the symbols
in this codebook. \ The goal of OSCR is to select which symbols to put in the
codebook such that the total size of the output---codebook size plus encoded
string size---is minimized. \ The optimal codebook size for $x$ is an estimate
of $K\left( S\right) $, the sophistication of $x$. \ The optimal total size
of the output for $x$ is called the minimum description length (MDL) of the
string, and is an estimate of $K\left( x\right) $.
The OSCR approach seems promising because of its direct relationship to the
functions we are interested in approximating. \ However, we implemented a
version of this algorithm, and we found that our implementation does not
perform well in compressing the automaton data. \ The output of the algorithm
is noisy, and there is no obvious trend in either the entropy or complexity
estimates. \ We conjecture that the noise is present because this compression
method, unlike others we consider, does not take into account the
two-dimensionality of the automaton state.
An alternative metric adopts the idea of coarse-graining. \ Here we aim to
describe a system's state on a macroscopic scale---for example, the coffee cup
as it would be seen by a human observer from a few feet away---by smoothing
the state, averaging nearby values together. \ Conceptually, for an automaton
state represented by a string $x$, its coarse-grained version is analogous to
a typical set $S$ which contains $x$. \ The coarse-grained state describes the
high-level, \textquotedblleft non-random\textquotedblright\ features of
$x$---features which it has in common with all other states from which the
same coarse-grained representation could be derived. \ Thus, the descriptive
size of the coarse-grained state can be used as an estimate for the state's
sophistication, $K\left( S\right) $. \ To estimate the descriptive size of
the coarse-grained state, we compress it using a general file compression
program, such as \texttt{gzip} or \texttt{bzip}. \ Shalizi \cite{shalizi}
objects to the use of such compression programs, claiming that they do not
provide consistently accurate entropy estimates and that they are too slow.
\ In our experiments, we have not seen either of these problems; our
simulations run in a reasonable amount of time and produce quite consistent
entropy estimates (see, for instance, Figure \ref{multicompress}). \ We
therefore use such compression programs throughout, though we consider
alternative approaches in Section \ref{CONC}.
Having defined the notion of coarse-graining, we can then define a two-part
code based on it. \ If the first part of the code---the typical set---is the
coarse-grained state, then the second part is $K\left( x|S\right) $, the
information needed to reconstruct the fine-grained state given the
coarse-grained version. \ The total compressed size of both parts of the code
is an estimate of the Kolmogorov complexity of the state, $K\left( x\right)
$.
We attempted to implement such a two-part code, in which the second part was a
diff between the fine-grained and coarse-grained states. \ The fine-grained
state, $x$, could be uniquely reconstructed from the coarse-grained array and
the diff. \ In our implementation of this two-part code, our estimate of
$K\left( x|S\right) $ suffered from artifacts due to the way the diff was
represented. \ However, defining a two-part code based on coarse-graining is
possible in general.
In light of the artifacts produced by our implementation of the two-part code,
we chose to pursue a more direct approach using coarse-graining. \ We
continued to use the compressed size of the coarse-grained state as an
approximation of $K\left( S\right) $. \ However, instead of approximating
$K\left( x|S\right) $ and using $K\left( S\right) +K\left( x|S\right) $
as an estimate of $K\left( x\right) $, we approximated $K\left( x\right) $
directly, by measuring the compressed size of the fine-grained array. \ This
approach avoided the artifacts of the diff-based code, and was used to
generate the results reported here.
\section{Coarse-Graining Experiment\label{CG}}
\subsection{Method\label{CGMETHOD}}
To derive a coarse-grained version of the automaton state from its original,
fine-grained version, we construct a new array in which the value of each cell
is the average of the values of the nearby cells in the fine-grained array.
\ We define \textquotedblleft nearby\textquotedblright\ cells as those within
a $g\times g$ square centered at the cell in question. The value of $g$ is
called the grain size, and here is selected experimentally. \ This procedure
is illustrated in Figure \ref{finecoarsefig}.
\begin{figure}\begin{center}%
\begin{tabular}
[c]{cc}%
\textbf{Fine-Grained} & \textbf{Coarse-Grained}\\%
\begin{tabular}
[c]{|c|c|c|c|c|}\hline
0 & 1 & 0 & 0 & 1\\\hline
0 & \cellcolor{silver}1 & \cellcolor{silver}1 & \cellcolor{silver}0 &
1\\\hline
0 & \cellcolor{silver}1 & \cellcolor{silver}0 & \cellcolor{silver}1 &
0\\\hline
1 & \cellcolor{silver}0 & \cellcolor{silver}1 & \cellcolor{silver}1 &
1\\\hline
1 & 0 & 1 & 0 & 0\\\hline
\end{tabular}
&
\begin{tabular}
[c]{|c|c|c|c|c|}\hline
0.5 & 0.5 & 0.5 & 0.5 & 0.5\\\hline
0.5 & $0.\overline{4}$ & $0.\overline{4}$ & $0.\overline{4}$ & 0.5\\\hline
0.5 & $0.\overline{5}$ & \cellcolor{silver}$0.\overline{6}$ & $0.\overline{6}$
& $0.\overline{3}$\\\hline
0.5 & $0.\overline{5}$ & $0.\overline{5}$ & $0.\overline{5}$ & 0.5\\\hline
0.5 & $0.\overline{6}$ & 0.5 & $0.\overline{6}$ & 0.5\\\hline
\end{tabular}
\\
&
\end{tabular}
\end{center}
\caption{Illustration of the construction of the coarse-grained array, using
an example grain size of $3$. \ The values of the shaded cells at left are
averaged to produce the value of the shaded cell at right.}%
\label{finecoarsefig}%
\end{figure}
Given this array of averages, we then threshold its floating-point values into
three buckets. \ Visually, these buckets represent areas which are mostly
coffee (values close to $0$), mostly cream (values close to $1$), or mixed
(values close to $0.5$). \ The estimated complexity of the state, $K\left(
S\right) $, is the file size of the thresholded, coarse-grained array after
compression. \ Analogously, the estimated entropy of the automaton state is
the compressed file size of the fine-grained array.
\subsection{Results and Analysis\label{CGRESULTS}}
Results from simulation of the automaton using the coarse-graining metric are
shown in Figure \ref{initialgraph}.
\begin{figure}[h]
\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig2_int}}
\fbox{\includegraphics[height=150px, width=200px]{fig2_nonint}}%
\caption{The estimated entropy and complexity of an automaton using the
coarse-graining metric. Results for the interacting model are shown at left,
and results for the non-interacting model are at right.}%
\label{initialgraph}%
\end{center}
\end{figure}
Both the interacting and non-interacting models show the predicted increasing,
then decreasing pattern of complexity. \ Both models also have an increasing
entropy pattern, which is expected due to the second law of thermodynamics.
\ The initial spike in entropy for the non-interacting automaton can be
explained by the fact that all of the particles can move simultaneously after
the first time step. \ Thus, the number of bits needed to represent the state
of the non-interacting automaton jumps after the first time step. \ With the
interacting automaton, by contrast, particles far from the coffee-cream border
cannot move until particles closer to the border have moved, so there is less
change in the automaton at each time step. \ Therefore, the estimated entropy
of this model is predictably more continuous throughout.
A visualization of the automaton's changing state over time is provided in
Figures \ref{interact} and \ref{noninteract}. \ This visualization is
generated by converting each cell's value to a grayscale color value; lighter
colors correspond to larger values. \ Visually, the fine-grained
representation of the state continues to grow more complicated with time,
while the coarse-grained representation first becomes first more and then less complicated.
\begin{figure}[h]\begin{center}%
\begin{tabular}
[c]{ccc}%
$\mathbf{t=0}$ & $\mathbf{t = 8 \times10^{6}}$ & $\mathbf{t = 2 \times10^{7}}%
$\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_00}} &
\fbox{\includegraphics[height=75px, width=75px]{fig3_01}} &
\fbox{\includegraphics[height=75px, width=75px]{fig3_02}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig3_11}} &
\fbox{\includegraphics[height=75px, width=75px]{fig3_12}}\\
& &
\end{tabular}
\caption{Visualization of the state of the interacting automaton of size $100$
over time. \ The top row of images is the fine-grained array, used to estimate
entropy. The bottom row is the coarse-grained array, used to estimate
complexity. \ From left to right, the images represent the automaton state at
the beginning of the simulation, at the complexity maximum, and at the end of
the simulation.}%
\label{interact}%
\end{center}\end{figure}
\begin{figure}[h]\begin{center}%
\begin{tabular}
[c]{ccc}%
$\mathbf{t=0}$ & $\mathbf{t = 5000}$ & $\mathbf{t = 10000}$\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_00}} &
\fbox{\includegraphics[height=75px, width=75px]{fig4_01}} &
\fbox{\includegraphics[height=75px, width=75px]{fig4_02}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig4_11}} &
\fbox{\includegraphics[height=75px, width=75px]{fig4_12}}\\
& &
\end{tabular}
\caption{Visualization of the state of the non-interacting automaton of size
$100$ over time.}%
\label{noninteract}%
\end{center}\end{figure}
The \texttt{gzip} compression algorithm was used to generate the results in
Figure \ref{initialgraph}, and is used throughout when a general file
compression program is needed. \ The results achieved using the
coarse-graining metric are qualitatively similar when different compression
programs are used, as shown in Figure \ref{multicompress}.
\begin{figure}[h]\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig5}}%
\caption{Coarse-grained complexity estimates for a single simulation of the
interacting automaton, using multiple file compression programs.}%
\label{multicompress}%
\end{center}\end{figure}
Given these results, it is informative to examine how complexity varies with
$n$, the size of the automaton.
\begin{figure}[h]
\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig6_int}}
\fbox{\includegraphics[height=150px, width=200px]{fig6_nonint}}%
\caption{Graphs of automaton size versus entropy maximum value. \ Quadratic
curve fits are shown, with $r^{2}$ values of $0.9999$ for both the interacting
and non-interacting automaton.}%
\label{nentropy}%
\end{center}\end{figure}\begin{figure}\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig7_int}}
\fbox{\includegraphics[height=150px, width=200px]{fig7_nonint}}%
\caption{Graphs of automaton size versus complexity maximum value. \ Linear
curve fits are shown, with $r^{2}$ values of $0.9798$ for the interacting
automaton and $0.9729$ for the non-interacting automaton.}%
\label{ncomplexity}%
\end{center}\end{figure}\begin{figure}\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig8_int}}
\fbox{\includegraphics[height=150px, width=200px]{fig8_nonint}}%
\caption{Graphs of automaton size versus time to complexity maximum. Quadratic
curve fits are shown, with $r^{2}$ values of 0.9878 for the interacting
automaton and 0.9927 for the non-interacting automaton.}%
\label{timetomax}%
\end{center}\end{figure}
The well-fit quadratic curve for the maximum values of entropy (Figure
\ref{nentropy}) is expected. \ The maximum entropy of an automaton is
proportional to the number of particles in the automaton. \ This is because,
if the state of the automaton is completely random, then the compressed size
of the state is equal to the uncompressed size--the number of particles. \ As
the automaton size, $n$, increases, the number of particles increases to
$n^{2}$.
The maximum values of complexity appear to increase linearly as the automaton
size increases (Figure \ref{ncomplexity}). \ That is, maximum complexity is
proportional to the side length of the two-dimensional state array. \ This
result is expected, since the automaton begins in a state which is symmetric
along its vertical axis, and complexity presumably develops along a single
dimension of the automaton. \ The time that it takes for the automaton to
reach its complexity maximum appears to increase quadratically with the
automaton size, or proportionally to the number of particles in the automaton
(Figure \ref{timetomax}). \ This result is also expected, since the time for
$n^{2}$ particles to reach a particular configuration is proportional to
$n^{2}$.
\section{Adujsted Coarse-Graining Experiment\label{ACG}}
\subsection{Method\label{ACGMETHOD}}
Though the original coarse-graining approach produces the hypothesized
complexity pattern, the method of thresholding used in the previous
experiment---dividing the floating-point values into three buckets---has the
potential to introduce artificial complexity. \ Consider, for example, an
automaton state for which the coarse-grained array is a smooth gradient from
$0$ to $1$. \ By definition, there will be some row of the array which lies on
the border between two threshold values. \ Tiny fluctuations in the values of
the coarse-grained array may cause the cells in this row to fluctuate between
two threshold values. In such a case, the small fluctuations in this border
row would artificially increase the measured complexity of the coarse-grained
array. \ This case is illustrated in Figure \ref{coarsegraining}.
\begin{figure}\begin{center}
\fbox{\includegraphics[height=75px, width=75px]{fig9}}
\fbox{\includegraphics[height=75px, width=75px]{fig9_thresh}}%
\caption{A coarse-grained array consisting of a smooth gradient from $0$ to
$1$ is shown at left. \ At right is the same array after a small amount of
simulated noise has been added and the values have been thresholded.}%
\label{coarsegraining}%
\end{center}\end{figure}
We propose an adjustment to the coarse-graining algorithm that helps to
minimize these artifacts. \ First, we use a larger number of
thresholds---seven, in contrast to the three used in the original experiment.
\ Additionally, we allow each cell in the array to be optionally,
independently adjusted up or down by one threshold, in whatever manner
achieves the smallest possible file size for the coarse-grained array.
This adjustment helps to compensate for the thresholding artifacts--such
random fluctuations could be removed by adjusting the fluctuating pixels.
\ However, since each pixel can be adjusted independently, there are
$2^{n^{2}}$ possible ways to adjust a given coarse-grained array.
Because we cannot search through this exponential number of possible
adjustments to find the optimal one, we develop an approximation algorithm to
produce an adjustment that specifically targets pixels on the border between
two thresholds. \ Given the properties of the automaton---it begins with rows
of dark cells on top, and light cells on the bottom---it is likely that each
row of the coarse-grained array will contain similar values. \ Thus, we adjust
the coarse-grained array by using a majority algorithm. \ If a cell is within
one threshold value of the majority value in its row, it is adjusted to the
majority value.
The hope is that
\begin{enumerate}
\item[(1)] this adjustment will reduce artificial border complexity by
\textquotedblleft flattening\textquotedblright\ fluctuating border rows to a
single color,
\item[(2)] the adjustment will not eliminate actual complexity, since
complicated structures will create value differences in the coarse-grained
array that are large enough to span multiple thresholds.
\end{enumerate}
\subsection{Results and Analysis\label{ACGRESULTS}}
Results from simulation of the automaton using the adjusted coarse-graining
metric are shown in Figure \ref{bettergraph}. \ Visualizations of the
automaton state are shown in Figures \ref{size100} and
\ref{size100noninteract}.
\begin{figure}\begin{center}
\fbox{\includegraphics[height=150px, width=200px]{fig10_int}}
\fbox{\includegraphics[height=150px, width=200px]{fig10_nonint}}%
\caption{The estimated entropy and complexity of an automaton using the
adjusted coarse-graining metric.}%
\label{bettergraph}%
\end{center}\end{figure}\begin{figure}\begin{center}%
\begin{tabular}
[c]{ccc}%
$\mathbf{t=0}$ & $\mathbf{t = 1.4 \times10^{7}}$ & $\mathbf{t = 4 \times
10^{7}}$\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_00}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_01}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_02}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig11_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_11}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_12}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig11_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_21}} &
\fbox{\includegraphics[height=75px, width=75px]{fig11_22}}\\
& &
\end{tabular}
\caption{Visualization of the state of the interacting automaton of size $100$
over time. \ The rows of images represent the fine-grained state, the original
coarse-grained state, and the coarse-grained state after adjustment,
respectively.}%
\label{size100}%
\end{center}\end{figure}\begin{figure}\begin{center}%
\begin{tabular}
[c]{ccc}%
$\mathbf{t=0}$ & $\mathbf{t = 10000}$ & $\mathbf{t = 20000}$\\
\fbox{\includegraphics[height=75px, width=75px]{fig3_00}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_01}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_02}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig12_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_11}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_12}}\\
\fbox{\includegraphics[height=75px, width=75px]{fig12_10}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_21}} &
\fbox{\includegraphics[height=75px, width=75px]{fig12_22}}\\
& &
\end{tabular}
\caption{Visualization of the state of the non-interacting automaton of size
$100$ over time. Note that the coarse-grained images are darker than for the
previous coarse-graining metric, because a larger number of thresholds were
used.}%
\label{size100noninteract}%
\end{center}\end{figure}
While this metric is somewhat noisier than the original coarse-graining
method, it results in a similarly-shaped complexity curve for the interacting
automaton. \ For the non-interacting automaton, however, the complexity curve
is flattened to a lower value.
This result for the non-interacting automaton is actually borne out by
theoretical predictions. \ The basic story is as follows; for details, see
Appendix \ref{ANALYSIS}. \ If we consider the automaton state to
\textquotedblleft wrap around\textquotedblright\ from right to left, then by
symmetry, the expected number of cream particles at a particular location in
the automaton depends solely on the vertical position of that location. \ The
expectations of all cells in a particular row will be the same, allowing the
two-dimensional automaton state to be specified using a single dimension.
\ Modeling each particle of cream as taking a random walk from its initial
position, it is possible to calculate the expected number of particles at a
given position as a function of time. \ Further, Chernoff bounds can be used
to demonstrate that the actual number of particles in each grain of the
coarse-grained state is likely to be close to the expectation, provided that
the grain size is large enough. \ Since it is possible to specify the expected
distribution of particles in the non-interacting automaton at all times using
such a function, the complexity of the non-interacting automaton state is
always low.
We believe thresholding artifacts caused the apparent increase in complexity
for the non-interacting automaton when regular coarse-graining was used. \ Our
adjustment removes all of this estimated complexity from the non-interacting
automaton, but preserves it in the interacting automaton. \ This evidence
suggests that the interacting automaton model may actually have intermediate
states of high complexity, even if the non-interacting model never becomes complex.
\section{Conclusions and Further Work\label{CONC}}
Of the metrics considered in this project, the coarse-graining approaches such
as apparent complexity provide the most effective estimate of complexity that
produces results which mirror human intuition. \ However, this metric suffers
from the disadvantage that it is based on human intuition and perceptions of
complexity. \ Ideally, a complexity metric would be found which produces
similar results without relying on such assumptions. \ The OSCR approach seems
promising for its independence from these assumptions and for its theoretical
foundations. \ It is possible that a different implementation of this
algorithm could produce better results than the one we used for this project.
It would also be worthwhile to investigate other complexity metrics, beyond
those already explored in this paper. \ Shalizi et al.\ \cite{ssh} propose a
metric based on the concept of light cones. \ They define $C\left( x\right)
$, the complexity of a point $x$ in the spacetime history, as the mutual
information between descriptions of its past and future light cones. \ Letting
$P\left( x\right) $ be the past light cone and $F\left( x\right) $ the
future light cone, $C\left( x\right) =H\left( P\left( x\right) \right)
+H\left( F\left( x\right) \right) -H\left( P\left( x\right) ,F\left(
x\right) \right) $. \ This metric is of particular interest because it
avoids the problem of artifacts created by coarse-graining; it can also be
approximated in a way that avoids the use of \texttt{gzip}. \ Running
experiments with the automaton using the light cone metric, and comparing the
results to those generated using coarse-graining, could provide more
information about both metrics.
Ultimately, numerical simulation is of limited use in reasoning about the
problem of complexity. \ Approximation algorithms can provide only an upper
bound, not a lower bound, on Kolmogorov complexity and sophistication. \ To
show that a system really does become complex at intermediate points in time,
it is necessary to find a lower bound for the system's complexity. \ Future
theoretical work could help provide such a lower bound, and could also
generate further insight into the origins of complexity in closed systems.
\section{Acknowledgments}
We thank Alex Arkhipov, Charles Bennett, Ian Durham, Dietrich Leibfried, Aldo
Pacchiano, and Luca Trevisan for helpful discussions.
\section{Appendix: The Non-Interacting Case\label{ANALYSIS}}
Let's consider the non-interacting coffee automaton on an $n\times n$ grid
with periodic boundary conditions. \ At each time step, each cream particle
moves to one of the $4$ neighboring pixels uniformly at random. \ Let
$a_{t}\left( x,y\right) $ be the number of cream particles at point $\left(
x,y\right) $ after $t$ steps.
\begin{claim}
\label{eclaim}For all $x,y,t$, we have $\operatorname*{E}\left[ a_{t}\left(
x,y\right) \right] \leq1$.
\end{claim}
\begin{proof}
By induction on $t$. \ If $t=0$, then $a_{0}\left( x,y\right) \in\left\{
0,1\right\} $. \ Furthermore, by linearity of expectation,%
\[
\operatorname*{E}\left[ a_{t+1}\left( x,y\right) \right] =\frac
{\operatorname*{E}\left[ a_{t}\left( x-1,y\right) \right]
+\operatorname*{E}\left[ a_{t}\left( x+1,y\right) \right]
+\operatorname*{E}\left[ a_{t}\left( x,y-1\right) \right]
+\operatorname*{E}\left[ a_{t}\left( x,y+1\right) \right] }{4}.
\]
\end{proof}
Now let $B$ be an $L\times L$\ square of pixels, located anywhere on the
$n\times n$\ grid. \ Let $a_{t}\left( B\right) $\ be the number of cream
particles in $B$ after $t$ steps. \ Clearly%
\begin{equation}
a_{t}\left( B\right) =\sum_{\left( x,y\right) \in B}a_{t}\left(
x,y\right) .
\end{equation}
So it follows from Claim \ref{eclaim}\ that $\operatorname*{E}\left[
a_{t}\left( B\right) \right] \leq L^{2}$.
Fix some constant $G$, say $10$. \ Then call $B$ \textquotedblleft
bad\textquotedblright\ if $a_{t}\left( B\right) $\ differs from
$\operatorname*{E}\left[ a_{t}\left( B\right) \right] $ by more than
$L^{2}/G$. \ Suppose that at some time step $t$, no $B$ is bad. \ Also,
suppose we form a coarse-grained image by coloring each $B$ one of $G$ shades
of gray, depending on the value of%
\begin{equation}
\left\lfloor \frac{a_{t}\left( B\right) G}{L^{2}}\right\rfloor
\end{equation}
(or we color $B$ white if $a_{t}\left( B\right) >L^{2}$). \ Then it's clear
that the resulting image will be correctable, by adjusting each color by
$\pm1$, to one where all the $B$'s within the same row are assigned the same
color---and furthermore, that color is simply%
\begin{equation}
\left\lfloor \frac{\operatorname*{E}\left[ a_{t}\left( B\right) \right]
G}{L^{2}}\right\rfloor .
\end{equation}
If this happens, though, then the Kolmogorov complexity of the coarse-grained
image can be at most $\log_{2}\left( n\right) +\log_{2}\left( t\right)
+O\left( 1\right) $. \ For once we've specified $n$ and $t$, we can simply
\textit{calculate} the expected color for each $B$, and no color ever deviates
from its expectation.
So our task reduces to upper-bounding the probability that $B$ is bad. \ By a
Chernoff bound, since $a_{t}\left( B\right) $\ is just a sum of independent,
$0/1$ random variables,%
\begin{equation}
\Pr\left[ \left\vert a_{t}\left( B\right) -\operatorname*{E}\left[
a_{t}\left( B\right) \right] \right\vert >\delta\operatorname*{E}\left[
a_{t}\left( B\right) \right] \right] <2\exp\left( -\frac
{\operatorname*{E}\left[ a_{t}\left( B\right) \right] \delta^{2}}%
{3}\right) .
\end{equation}
Plugging in $L^{2}/G=\delta\operatorname*{E}\left[ a_{t}\left( B\right)
\right] $, we get%
\begin{equation}
\Pr\left[ \left\vert a_{t}\left( B\right) -\operatorname*{E}\left[
a_{t}\left( B\right) \right] \right\vert >\frac{L^{2}}{G}\right]
<2\exp\left( -\frac{L^{4}}{3G^{2}\operatorname*{E}\left[ a_{t}\left(
B\right) \right] }\right) .
\end{equation}
Since $\operatorname*{E}\left[ a_{t}\left( B\right) \right] \leq L^{2}$
from above, this in turn is at most%
\begin{equation}
2\exp\left( -\frac{L^{2}}{3G^{2}}\right) .
\end{equation}
Now, provided we choose a coarse-grain size%
\begin{equation}
L\gg G\sqrt{3\ln\left( 2n^{2}\right) }=\Theta\left( G\sqrt{\log n}\right)
,
\end{equation}
the above will be much less than $1/n^{2}$. \ In that case, it follows by the
union bound that, at each time step $t$, with high probability \textit{none}
of the $L\times L$\ squares $B$ are bad (since there at most $n^{2}$ such
squares). \ This is what we wanted to show.
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:int}
In recent years, the amount of available extragalactic data has helped to establish a comprehensive picture of our Universe
and its evolution \cite[e.g.][]{Percival2010, Riess2011, Hinshaw2013, Planck2013}. These data, by and large, have enforced
the standard cosmological paradigm where initial perturbations in the mass density field grow via gravitational instability
and eventually form the cosmic structure we observe today. The clustering process is inevitably associated with peculiar
motions of matter, namely deviations from a pure Hubble flow. On large scales, these motions exhibit a coherent pattern, with
matter generally flowing from underdense to overdense regions. If galaxies indeed move much like test particles, they should
appropriately reflect the underlying peculiar velocity field which contains valuable information and, in principle, could be
used to constrain and discriminate between different cosmological models.
Usually relying on galaxy peculiar velocities estimated from measured redshifts and distance indicators, most approaches in
the literature have focused on extracting this information within local volumes of up to $100h^{-1}$ Mpc and larger centered
on the Milky Way \cite[e.g.,][]{Riess1995, Dekel1999, Zaroubi2001, Hudson2004, Sarkar2007, Lavaux2010, feldwh10, ND11,
Turnbull2012, Feindt2013}. Common distance indicators are based on well-established relations between observable intrinsic
properties of a given astronomical object, where one of them depends on the object's distance. A typical example is the
Tully-Fisher relation \cite{TF77} between rotational velocities of spiral galaxies and their absolute magnitudes. Due to
observational challenges, the number of galaxies in distance catalogs is relatively small compared to that of redshift catalogs,
limiting the possibility of exploring the cosmological peculiar velocity field to low redshifts $z\sim$ 0.02--0.03. Moreover,
all known distance indicators are potentially plagued by systematic errors \cite{lyn88, Strauss1995} which could give rise to
unwanted biases in the inferred velocities and thus renders their use for cosmological purposes less desirable.
To probe the flow of galaxies at deeper redshifts, one needs to resort to non-traditional distance indicators. One method,
for instance, exploits the kinetic Sunyaev-Zel'dovich effect to measure the cosmic bulk flow, i.e. the volume average of
the peculiar velocity field, out to depths of around 100--500$h^{-1}$ Mpc \cite[e.g,][]{Haehnelt1996, Osborne2011, Lavaux2013,
planck_bf}. Another strategy is based on the apparent anisotropic clustering of galaxies in redshift space which is commonly
described as redshift-space distortions. This effect is a direct consequence of the additional displacement from distances to
redshifts due to the peculiar motions of galaxies, and it yields reliable constraints on the amplitude of large-scale coherent
motions and the growth rate of density perturbations \cite[e.g.,][]{Hamilton1998, Peacock2001, Scoccimarro2004, Guz08}.
Galaxy peculiar motions also affect luminosity estimates based on measured redshifts, providing another way of tackling the
problem. Since the luminosity of a galaxy is independent of its velocity, systematic biases in the estimated luminosities
of galaxies can be used to explore the peculiar velocity field. The idea has a long history. It was first adopted to constrain
the velocity of the Virgo cluster relative to the Local Group by correlating the magnitudes of nearby galaxies with their
redshifts \cite{TYS}. Although in need of very large galaxy numbers to be effective, methods based on this idea use only
measured galaxy luminosities and their redshifts to derive bounds on the large-scale peculiar velocity field. Therefore, these
methods do not require the use of traditional distance indicators and they are also independent of galaxy bias. Using the
nearly full-sky 2MASS Redshift Survey (2MRS) \cite{Huchra2012}, for example, this approach has recently been adopted to
constrain bulk flows in the local Universe within $z\sim 0.01$ \cite{Nusser2011, Branchini2012}. Furthermore, it has been used
to determine the current growth rate of density fluctuations by reconstructing the full linear velocity field from the
clustering of galaxies \cite{Nusser1994, Nusser2012}.
Here we seek to apply this luminosity-based approach to obtain peculiar velocity information from galaxy redshifts and
apparent magnitudes of the Sloan Digital Sky Survey (SDSS) \cite{York2000}. The goals of our analysis are:
\begin{itemize}
\item A demonstration of the method's applicability to datasets with large galaxy numbers.
\item An updated estimate of the $r$-band luminosity function of SDSS galaxies at $z\sim 0.1$, accounting for evolution
in galaxy luminosities.
\item Novel bounds on bulk flows and higher-order moments of the peculiar velocity field at redshifts $z\sim 0.1$.
\item First constraints on the angular peculiar velocity power spectrum and cosmological parameters without additional
input such as galaxy clustering information.
\end{itemize}
The paper is organized as follows: we begin with introducing the luminosity method and its basic equations in section \ref{section2}.
In section \ref{section3}, we then describe the SDSS galaxy sample used in our analysis, together with a suite of mock catalogs
which will allow us to assess uncertainties and known systematics inherent to the data. After a first test of the method, we attempt
to constrain peculiar motions in section \ref{section4}, assuming a redshift-binned model of the velocity field. Because of the mixing
between different velocity moments arising from the SDSS footprint, bulk flow measurements are interpreted with the help of
galaxy mocks. Including higher-order velocity moments, we proceed with discussing constraints on the angular velocity power in
different redshift bins and their implications. As an example of cosmological parameter estimation, we further infer the quantity
$\sigma_{8}$, i.e. the amplitude of the linear matter power spectrum on a scale of $8h^{-1}$ Mpc, and compare the result to the
findings from the corresponding mock analysis. Other potential issues and caveats related to our investigation are addressed at the
section's end. In section \ref{section5}, we finally summarize our conclusions and the method's prospects in the context of
next-generation surveys. For clarity, some of the technical material is separately given in an appendix. Throughout the paper, we
adopt the standard notation, and all redshifts are expressed in the rest frame of the cosmic microwave background (CMB) using the
dipole from ref. \cite{Fixsen1996}.
\section{Methodology}
\label{section2}
\subsection{Variation of observed galaxy luminosities}
\label{section2a}
In an inhomogeneous universe, the observed redshift $z$ of an object (a galaxy) is generally different from its cosmological
redshift $z_{c}$ defined for the unperturbed background. To linear order in perturbation theory, one finds the well-known
expression \cite{SW}
\begin{equation}
\begin{split}
\frac{z-z_{c}}{1+z} &= \frac{V(t,r)}{c} - \frac{\Phi(t,r)}{c^2}\\
&{ } - \frac{2}{c^2}\int_{t(r)}^{t_0}{\rm d} t \frac{\partial\Phi\left\lbrack\hat{\vr} r(t),t\right\rbrack}{\partial t}\approx \frac{V(t,r)}{c},
\end{split}
\label{eq:sw}
\end{equation}
where $V$ is the physical radial peculiar velocity of of the object, $\Phi$ denotes the gravitational potential and $\hat{\vr}$
is a unit vector along the line of sight to the object. The last step explicitly assumes low redshifts where the velocity
$V$ makes the dominant contribution.\footnote{The first two terms on the right-hand side of eq. \eqref{eq:sw} describe
the Doppler effect and the gravitational redshift, respectively. The last one reflects the energy change of a photon
passing through a time-dependent potential well and is equivalent to the late-time integrated Sachs-Wolfe effect.} Note
that all fields are considered relative to their present-day values at a comoving radius of $r(t=t_{0})$ and that we have
substituted $z$ for $z_{c}$ in the denominator on the left-hand side of eq. \eqref{eq:sw}, which simplifies part of the
analysis presented below and is consistent at the linear level.
The observed absolute magnitude $M$, computed using the galaxy redshift $z$, rather than the (unknown) cosmological
redshift $z_{c}$, differs from the true value $M^{(t)}$ because of the shift ${\rm DM}(z)-{\rm DM}(z_c)$ in the distance
modulus ${\rm DM}=25+5\log_{\rm 10}\lbrack D_{L}/{\rm Mpc}\rbrack$, where $D_{L}$ is the luminosity distance. Hence,
\begin{equation}
\begin{split}
M &= m - {\rm DM}(z) - K(z) + Q(z)\\
&= M^{(t)} + 5\log_{10}\dfrac{D_{L}(z_{c})}{D_{L}(z)},
\end{split}
\label{eq:magvar}
\end{equation}
where $m$ is the apparent magnitude, $K(z)$ is the $K$-correction \cite[e.g.,][]{Blanton2007}, and the function $Q(z)$
accounts for luminosity evolution. Since the variation $M-M^{(t)}$ of magnitudes distributed over the sky is systematic,
it can be used to gain information on the peculiar velocity field. In the following, we will discuss how this may be
achieved with the help of maximum-likelihood techniques.
\subsection{Statistical description
\label{section2c}
\subsubsection{Inference of bulk flows and other velocity moments}
Before introducing our methodology, we need to specify a suitable model of the velocity field. Although a popular option
is to characterize peculiar velocities in terms of bulk flows, one could aim at a more complete description of the
peculiar velocity field. Given the current data, however, a full three-dimensional estimate of the velocity field would
be entirely dominated by the noise. A more promising approach is the following: first, we subdivide the galaxy data into
suitable redshift bins and consider the bin-averaged velocity $\tilde{V}$. Supposing for the moment that we are dealing
with a single bin, we then proceed to decompose $\tilde{V}(\hat{\vr} )$ (evaluated at the galaxy position $\hat{\vr}$) into
spherical harmonics, i.e.
\begin{equation}
\begin{split}
a_{lm} &= \int{\rm d}\Omega\tilde{V}(\hat{\vr} )Y_{lm}(\hat{\vr} ),\\
\tilde{V}(\hat{\vr} ) &= \sum\limits_{l,m}a_{lm}Y_{lm}^{*}(\hat{\vr} ),\quad l>0,
\end{split}
\label{eq:2c1}
\end{equation}
where the sum over $l$ is cut at some maximum value $l_{\rm max}$. A bulk flow of the entire volume, denoted as $\boldsymbol{v}_{B}$,
corresponds to the dipole term ($l=1$) in eq. \eqref{eq:2c1}. Building on the pioneering work of \cite{TYS}, for example,
the analysis presented in \cite{Nusser2011} has initially been restricted to a model with $l_{\rm max}=1$ when considering
galaxies from the 2MRS \cite{Huchra2012}.
Assuming that redshift errors can be neglected \cite{Nusser2011}, we write the probability of observing a
galaxy with magnitude $M$, given only its redshift and angular position $\hat{\vr}$ on the sky, as
\begin{equation}
P\left (M\vert z,a_{lm}\right ) = P\left (M\vert z,\tilde{V}(\hat{\vr} )\right ) = \frac{\phi(M)}{\eta\left (M^{+},M^{-}\right )},
\label{eq:2c2}
\end{equation}
where $\phi(M)$ is the galaxy luminosity function (LF) and $\eta\left (M^{+},M^{-}\right )$ is defined as
\begin{equation}
\eta\left (M^{+},M^{-}\right ) = \int_{M^{+}}^{M^{-}}\phi(M){\rm d} M.
\label{eq:2b2}
\end{equation}
The corresponding limiting magnitudes $M^{\pm}$ are given by
\begin{equation}
\begin{split}
M^{+} &= \max\left\lbrack M_{\rm min}, m^{+} - {\rm DM}(z_{c}) - K(z) + Q(z)\right\rbrack,\\
M^{-} &= \min\left\lbrack M_{\rm max}, m^{-} - {\rm DM}(z_{c}) - K(z) + Q(z)\right\rbrack,
\end{split}
\label{eq:app1c}
\end{equation}
where $m^{\pm}$ are the sample's limiting apparent magnitudes and the cosmological redshift $z_{c}$ depends on
the velocity $\tilde{V}$ and the observed redshift $z$ because of eq. \eqref{eq:sw}. The velocity model enters the
expression for the limiting magnitudes $M^{\pm}$ since it induces a shift in the distance modulus.
The coefficients
$a_{lm}$ of the flow modes can, therefore, be inferred by maximizing the total log-likelihood obtained from the sum
over all galaxies in a sample, i.e. $\log P_{\rm tot} = \sum\log P_{i}$. The rational for this is to find the set of
$a_{lm}$ which minimizes the spread in the observed magnitudes \cite{Nusser2011}. The spherical harmonics provide
an orthogonal basis only in the case of an all-sky survey, and the partial sky coverage of the SDSS implies that
the inferred moments will not be statistically independent. For example, a quadrupole velocity mode ($l=2$) would
contaminate the estimate of a bulk flow $\boldsymbol{v}_{B}$, which must be taken into account when interpreting any results.
The monopole term ($l=0$) is completely degenerate with an overall shift of the magnitudes, and hence it is not
included.
If the number of available galaxies is large enough, the central limit theorem implies that $P_{\rm tot}$ becomes
approximately normal, and we have
\begin{equation}
\log P_{\rm tot}\left (\boldsymbol{d}\vert\boldsymbol{x}\right ) = -\frac{1}{2}\left (\boldsymbol{x}-\overline{\boldsymbol{x}}\right )^{\rm T}\bm{\Sigma}^{-1}
\left (\boldsymbol{x}-\overline{\boldsymbol{x}}\right ),
\label{eq:2c3}
\end{equation}
where $\boldsymbol{x}$ is a vector of all model parameters, $\overline{\boldsymbol{x}}$ is the corresponding mean, $\boldsymbol{d}$ denotes the data
(or Bayesian evidence), and $\bm{\Sigma}$ is the covariance matrix describing the expected error of our estimate. We
have numerically verified that this approximation is extremely accurate for the SDSS which comprises several hundred
thousands of galaxies. The distribution's mean in eq. \eqref{eq:2c3} simply corresponds to the maximum-likelihood
estimate $\hat{\boldsymbol{x}}^{\rm ML}$ of the vector $\boldsymbol{x}$, and $\bm{\Sigma}$ can be estimated either by inverting the
observed Fisher matrix $\mathbf{F}$ which is defined as
$\mathbf{F}_{\alpha\beta}=-\partial\log P_{\rm tot}/(\partial x_{\alpha}\partial x_{\beta})$ evaluated at the
maximum value $\hat{\boldsymbol{x}}^{\rm ML}$ or from a realistic set of mock galaxy catalogs. The increasing number of
parameters associated with the higher-order moments of $\tilde{V}$ typically renders a full numerical evaluation of
$\log P_{\rm tot}$ unfeasible. A solution to this problem is based on approximating the total log-likelihood
function to second order (see section \ref{sectionnum} and appendix \ref{app1} for details). In the realistic
application to SDSS data, the model parameters $\boldsymbol{x}$ include the coefficients $a_{lm}$ (for each redshift bin) as
well as the LF parameters. They will be determined simultaneously by maximizing $\log P_{\rm tot}$.
\subsubsection{Inference of angular velocity power spectra}
Let us now focus on large scales where linear theory is applicable. Assuming Gaussian initial conditions, the
cosmological peculiar velocity field on these scales is then fully characterized by its power spectrum. The relevant
quantity here is the angular velocity power spectrum $C_{l}=\langle\lvert a_{lm}^{2}\rvert\rangle$. Under these
preliminaries, the problem of inferring the $C_{l}$ becomes equivalent to the more familiar estimation of the CMB
anisotropy power spectrum, and may thus be tackled with the same general techniques \cite{Tegmark1997, Bond1998}. To
estimate the power spectrum, one simply maximizes the probability of observing the data given the $C_{l}$, i.e.
\begin{equation}
P(C_{l})\equiv P\left (\boldsymbol{d}\vert C_{l}\right )\propto\int{\rm d} a_{lm}P\left (\boldsymbol{d}\vert a_{lm}\right )P\left (a_{lm}\vert C_{l}\right ),
\label{eq:2c4}
\end{equation}
which is obtained by constructing the posterior likelihood according to Bayes' theorem and marginalizing over the
$a_{lm}$. Here the individual $a_{lm}$ are uncorrelated and taken to be normally distributed, i.e. one has
\begin{equation}
P\left (a_{lm}\vert C_{l}\right ) = \prod\limits_{l,m}\left (2\pi C_{l}\right )^{-1/2}\exp
\left (-\frac{\lvert a_{lm}\rvert^{2}}{2C_{l}}\right ),
\label{eq:2c5}
\end{equation}
and $P\left (\boldsymbol{d}\vert a_{lm}\right )$ is derived from marginalizing $P_{\rm tot}\left (\boldsymbol{d}\vert\boldsymbol{x}\right )$ over the
remaining parameters in $\boldsymbol{x}$. Within the Gaussian approximation, carrying out the integration in eq. \eqref{eq:2c4}
is straightforward and the resulting expressions are presented in appendix \ref{app3}.
Considering a particular model like, for example, the standard $\Lambda$CDM cosmology, the $C_{l}$ are fully specified
by a set of cosmological parameters $\zeta_{k}$. Therefore, accounting for this dependency in the prior probability,
the above technique may further be used to constrain cosmological key quantities such as $\sigma_{8}$ from the observed
peculiar velocity field alone. Given the characteristics of current galaxy redshift surveys, it is clear that these
constraints will be much less tight than those obtained by other means such as CMB analysis, but still valuable as a
complementary probe and consistency check.
A successful application of the method requires a large number of galaxies to beat the statistical (Poissonian)
errors. The method does not require accurate redshifts and can be used with photometric redshifts to recover signals
on scales larger than the spread of the redshift error. Other related maximum-likelihood approaches based on reduced
input (photometric redshifts or just magnitudes) \cite{Itoh2010,Abate2012} consider integrated quantities such as
number densities, resulting in less sensitive measurements of bulk flows and higher-order moments of the peculiar
velocity field.
\subsection{Estimating the galaxy luminosity function}
\label{section2d}
A reliable measurement of the galaxy LF represents a key step in our approach. A corresponding
estimator should be flexible enough to capture real features in the luminosity distribution, but also physical in
the sense of returning a smooth function over the range of interest. To meet these requirements, we shall adopt the
spline-based estimator introduced in \cite{Branchini2012} for our analysis.\footnote{Since the underlying principle
is a reduction in the spread of observed magnitudes, even unrealistic models of the luminosity distribution should
yield unbiased measurements of the velocity information \cite{Nusser2012}, albeit with larger statistical errors.}
In this case, the unknown LF is written as a piecewise-defined function, i.e.
\begin{equation}
\phi(M) = \varphi_{i}(M),\qquad M_{i-1}\leq M<M_{i},
\label{eq:2d1}
\end{equation}
where $\varphi_{i}$ is a third-order polynomial defined such that the second derivative of $\phi$ with respect to $M$ is
continuous on the interval $[M_{0},M_{N-1}]$ and vanishes at the boundaries.
The cubic spline in eq. \eqref{eq:2d1} may be regarded as a generalization of the stepwise estimator originally proposed
in \cite{efs88}, and the actual spline coefficients are determined employing the standard techniques summarized in
\cite{Press2002}. Since there occur only polynomial expressions, derivatives and integrals of $\phi(M)$ are of particularly
simple form, allowing quite an efficient evaluation of the previously defined likelihood functions. LFs which are obtained
according to this procedure might exhibit spurious wiggles, especially at the corresponding bright and faint ends.
As is already discussed in \cite{Branchini2012}, however, these wiggles can be sufficiently suppressed by adding an appropriate
penalty term to the total likelihood function or by enforcing (log-)linear behavior of $\phi(M)$ beyond suitable bright and
faint magnitude thresholds. Alternatively, it is also possible to simply choose magnitude cuts and the total number of spline
points in such a way that the number of galaxies in each magnitude interval is large enough to avoid this problem for all
practical purposes. In the present analysis, we will follow the latter approach when maximizing the total log-likelihood.
In addition to the spline-based estimator, which is most relevant when considering real observations, we shall also use a
parametric estimator that assumes a widely used Schechter form of the LF \cite{Sandage1979,schechter}, i.e.
\begin{equation}
\phi(M) \propto 10^{0.4(1+\alpha^{\star})(M^{\star}-M)}\exp{\left (-10^{0.4(M^{\star}-M)}\right )},
\label{eq:2d2}
\end{equation}
where $M^{\star}$ and $\alpha^{\star}$ are the usual Schechter parameters. The normalization of $\phi(M)$ cancels in the
likelihood function and does not concern us here. Although it does not provide a good fit to all datasets, the Schechter form
and its corresponding estimator turn out very useful for the analysis of both mock catalogs and the real galaxy sample
presented in section \ref{section4}.
\section{Datasets}
\label{section3}
\subsection{NYU Value-Added Galaxy Catalog}
\label{section3a}
We will use the SDSS galaxies from the latest publicly available NYU Value-Added Galaxy Catalog (NYU-VAGC)
\cite{Blanton2005}.\footnote{\url{http://sdss.physics.nyu.edu/vagc/}}. This catalog is based on the SDSS Data Release $7$
(DR7) \cite{abaz}, and contains galaxies with a median redshift of $z\approx 0.1$, observed in five different photometric
bands with magnitudes corrected for Galactic extinction according to \cite{Schlegel1998}. Using Petrosian magnitudes, we
decide to work with the $r$-band, mainly because it gives the largest spectroscopically complete galaxy sample
\cite{Blanton2001, Strauss2002}, which is an important factor for the statistical method we have introduced in section
\ref{section2}. To minimize incompleteness and to exclude galaxies with questionable photometry and redshifts, we choose
the subsample NYU-VAGC {\tt safe} which contains only galaxies whose apparent $r$-band magnitudes satisfy
$14.5 < m_{r} < 17.6$.\footnote{The SDSS photometry is known to exhibit small offsets from the AB magnitude system
\cite{Oke1983}. For the SDSS $r$-band, this amounts to a shift of around $0.01$ \cite{Eisenstein2006} which we will
take into account when calculating absolute magnitudes below. } The subsample accounts for fiber collisions following
the correction scheme {\tt nearest}, but this is expected to be of little relevance in our analysis which should be
insensitive to galaxy clustering. Also, since we are interested in minimizing
systematics due to uncertainties in $K$-corrections and luminosity evolution (see section \ref{section4}), we shall adopt
the $^{0.1}r$-bandpass when dealing with absolute magnitudes \cite{Blanton2003B}, and further impose cuts on redshifts
(expressed in the CMB frame) and observed absolute magnitudes $M_{r}$ such that only galaxies with $0.02 < z < 0.22$ and
$-22.5 < M_{r} - 5\log_{10}h < -17.0$ are selected. The number of galaxies contained in our final working sample is
approximately $5.4\times 10^{5}$ and may slightly vary, depending on the assumed background cosmology which enters the
calculation of $M_{r}$ through the luminosity distance. For realistic flat cosmologies with a total matter density
$\Omega_{m}\approx 0.3$, however, these variations are typically on the order of a few hundred galaxies and thus not very
significant. Since we are concerned with relatively low redshifts $z\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.2$, we assume a linear dependence of the
luminosity evolution on redshift for simplicity, i.e.
\begin{equation}
\label{eq:qz}
Q(z)= Q_{0}(z-z_{0}),
\end{equation}
where we set the pivotal redshift $z_{0}=0.1$. Furthermore, $K$-corrections for individual galaxies are taken from the
NYU-VAGC and have been calculated with the software package {\tt kcorrect} {\tt v4{\_}1{\_}4} \cite{Blanton2007}. To
calculate the limiting absolute magnitudes $M^{\pm}$ in the $^{0.1}r$-bandpass at a given redshift $z$, however, we
resort to a mean $K$-correction of the form
\begin{equation}
\overline{K}(z) = -2.5\log_{10}(1.1) + \sum_{i=1}^{3}\gamma_{i}(z - 0.1)^{i},
\label{eq:4b}
\end{equation}
where $\gamma_{1}\approx 0.924$, $\gamma_{2}\approx 2.095$, and $\gamma_{3}\approx -0.184$ are determined by directly
fitting the individual $K$-corrections listed in the NYU-VAGC. When calculating the total likelihood function introduced
in section \ref{section2c}, all galaxies are weighted according to the angular (redshift) completeness. All remaining
details relevant to the analysis of the NYU-VAGC galaxy redshift data will be separately discussed in section \ref{sectionnum}.
\subsection{Mock galaxy catalogs}
\label{section3b}
To test the performance of our approach, we resort to two different suites of galaxy mock catalogs. The first set of mocks
is based on the LasDamas simulations \cite{McBride2009} while the second one is obtained from the real NYU-VAGC dataset that
we analyse in this work.
\subsubsection{LasDamas mock catalogs}
These mock galaxy catalogs are obtained by populating the LasDamas simulations \cite{McBride2009} with artificial galaxies,
using a halo occupation distribution model \cite[e.g.,][]{pesm,Seljak2000,Berlind2002} to match the observed clustering of
SDSS galaxies in a wide luminosity range. The goal of these catalogs is to benchmark our method and validate its implementation,
using a sample with overall characteristics (number density of objects, sky coverage, etc.) similar to that of the real catalog,
ignoring all sources of systematic biases.
Here we will consider a total of $60$ mocks from the public {\tt gamma} release, modeled after a volume-limited subsample of
SDSS DR7 cut at $M_{r}<-20$, which cover the full SDSS footprint (``North and South'') and a redshift range $0.02<z<0.106$ with
a median of $z\approx 0.08$.\footnote{\url{http://lss.phy.vanderbilt.edu/lasdamas/mocks.html}} The typical galaxy number in
these mocks is around $1-1.5\times 10^{5}$, and we shall use them as a basic test of bulk flow measurements. To this end, an
observed redshift is assigned to each galaxy according to
\begin{equation}
cz = cz_{c} + V + c\epsilon_{z},
\label{eq:3b1}
\end{equation}
where $z_{c}$ corresponds to the redshift entry in the mock catalog, the radial velocity $V=\hat{\vr}\cdot\boldsymbol{v}_{B}$ is the
line-of-sight component of the bulk flow $\boldsymbol{v}_{B}$, and $\epsilon_{z}$ is a random measurement error drawn from a
Gaussian distribution with $c\sigma_{z}=15$ km s$^{-1}$.\footnote{Although the redshifts listed in the LasDamas {\tt gamma}
mocks include distortions from peculiar velocities, we interpret them as cosmological redshifts $z_{c}$ for simplicity.
This has no adverse effect on testing our method's performance.} Similarly, observed $r$-band magnitudes are assigned
with the help of eq. \eqref{eq:magvar}, but without including the $K$-correction term. Assuming the linear luminosity
evolution in \eqref{eq:qz} with $Q_{0}=1.6$, the true galaxy magnitudes $M^{(t)}$ are randomly extracted from the
Schechter distribution given by eq. \eqref{eq:2d2} with the parameters $M^{\star}=-20.44+5\log_{10}h$ and
$\alpha^{\star}=-1.05$ \cite{Blanton2003}. Although it is irrelevant for the present purposes, this procedure ignores
the masses of dark matter halos, meaning that very massive halos may host very faint galaxies and vice versa. We also
add a Gaussian random error to $M_{r}$ with $\sigma_{M}=0.03$, and further trim the resulting mock catalogs by requiring
$M_{r}<-20.25$ to prevent problems related to Malmquist bias. Finally, our choice of the bulk flow $\boldsymbol{v}_{B}$ used in the
benchmark runs will be described in section \ref{section4}. As the LasDamas simulations assume a flat $\Lambda$CDM model
with $\Omega_{m}=0.25$ and $h=0.7$, we adopt the same cosmology for the mocks.
\begin{table*}
\caption{Summary of the $\Lambda$CDM cosmologies described in the text.}
\begin{tabular*}{0.95\linewidth}{@{\extracolsep{\fill}}lcccccc}
\noalign{\medskip}
\hline
\hline
\noalign{\smallskip}
Parameter set & $\Omega_{b}$ & $\Omega_{m}$ & $\Omega_{\Lambda}$ & $h$ & $n_{s}$ & $\sigma_{8}$
\tabularnewline
\noalign{\smallskip}
\hline
\noalign{\smallskip}
{\tt param{\_}mock} & 0.0455 & 0.272 & 0.728 & 0.702 & 0.961 & 0.8
\tabularnewline
{\tt param{\_}wmap} & 0.0442 & 0.2643 & 0.7357 & 0.714 & 0.969 & 0.814
\tabularnewline
{\tt param{\_}planck} & 0.049 & 0.3175 & 0.6825 & 0.671 & 0.962 & 0.834
\tabularnewline
\noalign{\smallskip}
\hline
\hline
\end{tabular*}
\label{table1}
\end{table*}
\subsubsection{NYU-VAGC mock catalogs}
Starting directly from the previously described NYU-VAGC dataset, we generate a second set of mock catalogs built from the
angular positions and spectroscopic redshifts of the observed galaxies. The goal of these mocks is to investigate the impact
of known observational biases, incompleteness, and cosmic variance while preserving the spatial distributions of the galaxies
in the real SDSS DR7 catalog.
Just as in the case of the LasDamas mock catalogs, we interpret the observed spectroscopic redshifts as the cosmological ones
and obtain the corresponding measured redshifts from eq. \eqref{eq:3b1}, where $V$ is now determined from the full linear
velocity field evaluated at redshift $z=0$. The velocity field is obtained from a random realization sampled on a cubic grid
with $1024^{3}$ points and a comoving mesh size of $4h^{-1}$ Gpc, assuming the linear power spectrum $P_{v}(k)$ of a flat
$\Lambda$CDM cosmology with total matter and baryonic density parameters $\Omega_{m}=0.272$ and $\Omega_{b}=0.0455$, respectively,
scalar spectral index $n_{s}=0.961$, $h=0.702$, and $\sigma_{8}=0.8$ (corresponding to the parameter set {\tt param{\_}mock}
which is listed in table \ref{table1}). To ensure a high level of (statistical) independence between the final mocks, we
perform appropriate translations and rotations of the survey data reference frame relative to the sampling grid before each
galaxy is assigned a velocity equal to that of the nearest grid point. Because of small-scale nonlinearities and the finite
grid sampling, we further add uncertainties to the line-of-sight components of these velocities which are generated from a
normal distribution with $\sigma_{V} = 250$ km s$^{-1}$.
The luminosities are assigned exactly as for the LasDamas mocks using the appropriate cuts in apparent and absolute magnitudes
specified in section \ref{section3a}. In addition, we simulate two known systematic errors in the photometric calibration
of SDSS data \cite{Pad2008} that have a potential impact on our analysis. The first one arises from various magnitude offsets
between the individual SDSS stripes, and is modeled by considering another random error with $\sigma_{\rm stripe} = 0.01$. The
second, more serious error results from unmodeled atmospheric variations during the time of observation, ultimately causing an
overall zero-point photometric tilt of roughly $0.01$ in magnitudes over the survey region. To mimic this tilt, we include in
each mock a magnitude offset in the form of a randomly oriented dipole normalized such that its associated root mean square
(rms) over all galaxies is $\delta m_{\rm dipole}=0.01$ \cite{Pad2008}.
With this procedure we obtain a total of $269$ galaxy mocks (both flux- and volume-limited), mimicking the characteristics of the
real NYU-VAGC sample. These mocks will be used to explore the distribution of measured bulk flow vectors and to study constraints
on the power spectrum $C_{l}$ or cosmological parameters for a realistic choice of the large-scale peculiar velocity field.
\section{Data analysis}
\label{section4}
We now proceed to apply our method to the SDSS data. To achieve our goals outlined in section \ref{sec:int}, we will begin with
a short description of some additional preliminaries and present the general line of action in section \ref{sectionnum}. After
this, we will estimate the $r$-band LF of SDSS galaxies at $z\sim 0.1$ in section \ref{sectionlfestimate}, which provides the
basis for our investigations. The results of our velocity analysis of the NYU-VAGC galaxy sample are then presented and discussed
in sections \ref{sec:bf} and \ref{sectionvcosmo}.
\subsection{General line of action}
\label{sectionnum}
A major obstacle in constraining the velocity field from the SDSS is the partial coverage (only about $20$\%) of the sky. Since
the $Y_{lm}$ no longer form an orthogonal basis on this limited mask, the maximum-likelihood approach yields a statistical mixing
between the estimated velocity moments, effectively probing a combination of different multipoles. In the case of the bulk flow,
for instance, this would correspond to a superposition of several terms up to even the hexadecapole of the peculiar velocity
field \cite{Tegmark2004}. Of course, one may resort to an orthogonal basis set for $l_{\rm max}$ in pixel space. Because we are
going through the full maximum-likelihood procedure, however, there is no gain in doing so, i.e. all the information is already
contained in the measured $a_{lm}$ and their covariance matrix. Also, the results expressed in such orthogonal bases typically
have a less obvious physical interpretation.
Additional difficulties arise from a too flexible LF model, i.e. oversampling issues related to the spline-based estimator, and
the linear evolution term $Q(z)$ which actually mimics the formally ignored monopole contribution in eq. \eqref{eq:2c1} over the
redshift range of interest. Both may contribute to the mode mixing and further complicate the interpretation of the corresponding
results. Similarly, the presence of systematic errors in the SDSS photometry (see section \ref{section3b}) can lead to spurious
flows which contaminate the velocity measurements and bias possible estimates of velocity power and cosmological parameters.
Despite these limitations, however, we show below that such measurements can still provide meaningful constraints if one
interprets them with the help of suitable mock catalogs sharing the same angular mask (see section \ref{section3b}). For
instance, estimates of different quantities can be directly compared to the corresponding distributions obtained from the
mocks where systematic effects are under control. As for the data (and mock) analysis presented below, we shall thus employ
the following basic strategy:
\begin{enumerate}
\item Assume a set of parameters that describe the background cosmology and select the galaxy sample according to the absolute
magnitudes and luminosity distances computed, respectively, from apparent magnitudes and redshifts (see section \ref{section3a}).
\item Assuming the linear luminosity evolution model specified in eq. \eqref{eq:qz}, determine the $^{0.1}r$-band LF parameters
including $Q_{0}$ for the case of a vanishing peculiar velocity field, i.e. $a_{lm} = 0$. The value of $Q_{0}$ is kept fixed in
the following steps while the LF parameters are free to vary, except when using the fixed LF estimator explored in section
\ref{sec:bf}.
\item Compute the maximum-likelihood estimate $\hat{\boldsymbol{x}}^{\rm ML}$ of the parameter vector $\boldsymbol{x}$ introduced in section \ref{section2c}
for a suitable $l_{\rm max}$. These parameters specify both the velocity model and the LF. Approximating $P_{\rm tot}$
locally by a Gaussian distribution and taking the previously found $\phi$ with $a_{lm} = 0$ as an initial guess, this is achieved
by iteratively solving for $\hat{\boldsymbol{x}}^{\rm ML}$ until the (exact) likelihood peak is reached. The required derivatives of
$\log{P_{\rm tot}}$ can be calculated analytically and are summarized in appendix \ref{app1}. Convergence is reached after $3$--$5$
iterations for a relative accuracy $10^{-6}$--$10^{-10}$. The CPU time depends on the value of $l_{\rm max}$, but is typically
around a few tens of minutes for about half a million objects. The results are potentially prone to mask-induced degeneracies
related to the spline point separation $\Delta M$ when estimating the LF. We will describe below the various approaches used to
investigate this issue.
\item Estimate the random errors of $\hat{\boldsymbol{x}}^{\rm ML}$ from the covariance matrix $\bm{\Sigma}$ which is computed by directly
inverting the observed Fisher matrix $\mathbf{F}$. The Fisher matrix
$\mathbf{F}_{\alpha\beta}=-\partial\log P_{\rm tot}/(\partial x_{\alpha}\partial x_{\beta})$ is evaluated at the maximum value
$\hat{\boldsymbol{x}}^{\rm ML}$.\footnote{An immediate worry is that $\mathbf{F}$ could turn out singular or ill-conditioned.
Except for issues, which are related to the normalization of the spline-based LF and can be easily overcome with the
help of standard techniques \cite{James2006}, we have not encountered such problems in our study which considers only
the first few multipoles, i.e. $l_{\rm max}\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 5$.}
\item Marginalize the resulting distribution $P_{\rm tot}(\boldsymbol{d}\vert\boldsymbol{x} )$ over all LF parameters that are unrelated to the velocity
field and construct the posterior probability for the $a_{lm}$ and, subsequently, the probability $P(C_{l})$ according to the
prescription given in section \ref{section2c}. Then maximize the latter with respect to $C_{l}$ to estimate the angular power.
Given the characteristics of the SDSS data, such estimates are expected to be quite uncertain, and thus we will limit ourselves
to a proof of concept.
\item Alternatively, consider a spatially flat $\Lambda$CDM model where the $C_{l}$ are not free and independent, but
fully determined by the cosmological parameters $\zeta_{k}$, i.e. $C_{l}=C_{l}(\zeta_{k})$. Constraints are obtained
by sampling the probability $P[C_{l}(\zeta_{k})]$ as a function of $\zeta_{k}$ on a discrete grid. Although other
parameter choices are briefly discussed, we will focus on the quantity $\sigma_{8}$ which corresponds to the amplitude
of the linear matter power spectrum on a scale of $8h^{-1}$ Mpc. To ensure that linear theory remains a valid
description on the physical scales probed in the analysis, we further have to set $l_{\rm max}$ accordingly (see section
\ref{sectionvcosmo} for details).
\end{enumerate}
Except in the case of examining the LasDamas mocks, which contain substantially less galaxies than the other samples
(see section \ref{section3b} above) and cover a smaller redshift range, we will consider the peculiar velocity field
in two redshift bins with $0.02 < z_{1} < 0.07 < z_{2} < 0.22$, comprising about $N_{1}\sim 1.5\times 10^{5}$ and
$N_{2}\sim 3.5\times 10^{5}$ galaxies, respectively. This specific choice is mainly driven by the accuracy of the bulk
flow estimates presented in section \ref{sec:bf}. For two redshift bins, the uncertainties are typically around $100$ km
s$^{-1}$ which is also larger than the expected variation of the flow amplitude within the respective bins, yielding a good
compromise between accuracy and evolution. Further, the actual bin widths are determined by requiring comparable
signal-to-noise ratios which we roughly estimate from the expected variance of the velocity field within the corresponding
bin volumes and the Poisson noise due the finite number of objects. As for the study of the observed data sample, we adopt
the latest cosmological parameters based on the Wilkinson Microwave Anisotropy Probe (WMAP) combined with ground-based
experiments \cite{Calabrese2013} and the recent measurements by the Planck satellite \cite{Planck2013} which are summarized
in table \ref{table1} and denoted by {\tt param{\_}wmap} and {\tt param{\_}planck}, respectively.
\begin{figure}
\includegraphics[width=0.95\linewidth]{fig1.eps}
\caption{The $^{0.1}r$-band LF as obtained from the NYU-VAGC sample: shown are the maximum-likelihood result adopting
the spline-based estimator with $\Delta M=0.5$ (solid line), and two fits based on the Schechter form (dashed line)
and its extension (dotted line; zoomed panel only) which is defined by eq. \eqref{eq:4a}.}
\label{fig2}
\end{figure}
\subsection{The \boldmath{$^{0.1}r$}-band luminosity function of NYU-VAGC galaxies}
\label{sectionlfestimate}
As described in section \ref{sectionnum}, we begin with estimating $\phi(M)$ in the $^{0.1}r$-band from the NYU-VAGC
data for a vanishing velocity field. Adopting the spline-based estimator with a separation of $\Delta M = 0.5$ between
individual spline points, the resulting $\phi(M)$ is shown as a solid line in figure \ref{fig2}, where the normalization
is chosen such that the integral of $\phi(M)$ over the considered absolute magnitude range becomes unity, i.e.
$\eta(-22.5+5\log_{10}h,-17+5\log_{10}h)=1$, and error bars are computed from the ``constrained'' covariance matrix
obtained by enforcing the LF normalization to guarantee a non-singular Fisher matrix. The shape of $\phi(M)$ and the
found evolution parameter, $Q_{0} = 1.6\pm 0.11$, are in good agreement with previous studies based on earlier data
releases \cite{Blanton2003,Montero2009}. While the simple Schechter form with $M^{\star}-5\log_{10}h=-20.52\pm 0.04$
and $\alpha^{\star}=-1.10\pm 0.03$ (dashed line) describes the estimated $\phi$ reasonably well, it does not capture
the visible feature at the faint end.\footnote{We emphasize that the variation in the faint-end slope is robust with
respect to different choices of $\Delta M$ and not an artifact caused by the spline estimator.} Therefore, we consider
an extension to eq. \eqref{eq:2d2} which, after using the relation $L/L^{\star}=10^{0.4(M^{\star}-M)}$, takes the form
\begin{equation}
\phi \propto \left (\frac{L}{L^{\star}}\right )^{\beta^{\star}_{1}}\left\lbrack 1 + 10^{-2.5}
\left (\frac{L}{L^{\star}}\right )^{\beta^{\star}_{2}}\right\rbrack\exp{\left (-\frac{L}{L^{\star}}\right )}
\label{eq:4a}
\end{equation}
and is equivalent to the sum of two Schechter functions with different choices of normalization and $\alpha^{\star}$.
Fitting the above to the spline estimate yields the parameters $M^{\star} = -20.46\pm 0.03$,
$\beta^{\star}_{1} = -1.01\pm 0.03$, and $\beta^{\star}_{2} = -1.64\pm 0.11$, giving a much better representation of
the observed trend. This is illustrated in the zoomed panel of figure \ref{fig2}, where the new result (dotted line) is
compared to both the spline (solid line) and the previous Schechter fit (dashed line).
To further assess our result, we
also calculate the predicted redshift distribution ${\rm d} N/{\rm d} z$ of galaxies which is directly proportional to the radial
selection function $S(z)$, i.e. the fraction of galaxies included in the sample at redshift $z$. The selection function
is easily obtained as an integral of the LF over the magnitude range defined by the redshift-dependent limiting
absolute magnitudes.
\begin{figure}
\includegraphics[width=0.95\linewidth]{fig2.eps}
\caption{Redshift distribution of SDSS galaxies from the NYU-VAGC sample: the histogram (solid line) represents the
observed distribution normalized to unity over the considered redshift range for bins with $\Delta z = 2.5\times 10^{-3}$.
The predicted distribution (dashed line) assumes the spline-based estimate of the LF. }
\label{fig3}
\end{figure}
Figure \ref{fig3} shows that the predicted and observed redshift distributions match quite well, except for a slight
disagreement on the order of a few percent near the high-$z$ cut. This small discrepancy is most likely caused by a
combination of both the limited linear evolution model and the use of different $K$-corrections (individual and mean)
when estimating $\phi(M)$ and the selection function (see section \ref{sectioncaveats} for a discussion of how
luminosity evolution and $K$-corrections impact our peculiar velocity results). Note that all of the above assumes
the set {\tt param{\_}wmap}. Repeating the analysis with the parameters from {\tt param{\_}planck}, however, does not
yield any significant changes.
\begin{figure}
\includegraphics[width=0.95\linewidth]{fig3.eps}
\caption{Histograms of bulk flow measurements obtained from the customized LasDamas mocks: shown are
the recovered distributions for both a known (dashed line) and unknown (solid line) flow direction.}
\label{fig1}
\end{figure}
\subsection{Constraining bulk flows}
\label{sec:bf}
As a first application of our method, we address how it may be used to constrain the bulk flow, $\boldsymbol{v}_{B}$, in the
NYU-VAGC data. We begin with the LasDamas mocks for testing the ability of the method to detect large, anomalous
bulk flows in a SDSS-like catalog. Then we apply the method to the real NYU-VAGC and discuss whether our results
are consistent with the standard $\Lambda$CDM cosmology.
\subsubsection{LasDamas benchmark}
\label{benchmark}
We make use of the LasDamas mocks introduced in section \ref{section3b} and assume
a constant $\boldsymbol{v}_{B}$ of $1000$ km s$^{-1}$ pointing toward the direction $(l, b) \approx (266^{\circ}, 33^{\circ})$
expressed in Galactic coordinates. Note that both the magnitude and the direction of $\boldsymbol{v}_{B}$ are chosen in accordance
with the recent controversial claim of a ``dark flow'' out to depths of around 300--600$h^{-1}$ Mpc \cite{Kash2008,
Keisler2009, Kash2010, Osborne2011}.
Using the Schechter estimator for the LF and setting $l_{\rm max}=1$, we follow
the procedure outlined in section \ref{sectionnum} to recover the flow $\boldsymbol{v}_{B}$ from the customized LasDamas mocks.
The histogram in figure \ref{fig1} shows the resulting component along the input direction for the cases that it is
known (dashed line) and unknown (solid line), i.e. the direction is allowed to vary freely. Clearly, the magnitude of
$\boldsymbol{v}_{B}$ is successfully extracted in both cases, and the corresponding rms values of $111$ and $125$ km s$^{-1}$ are
fully consistent with each other as is expected from Gaussian statistics. Although not presented here, the found
distributions along the other (perpendicular) directions for a freely varying $\boldsymbol{v}_{B}$ are consistent with a zero
velocity and exhibit a similar scatter. Of course, the current setup neglects any contamination due to leakage from
other multipoles or systematic errors in the data. If these effects remain subdominant in the sense that their
combination leads to changes comparable to or less than the estimated random errors, our result suggests that the
method is capable of constraining large coherent bulk flows using the available galaxy data from the NYU-VAGC. As we
will show below, this condition seems reasonably satisfied, at least for the results in the low-redshift bin with
$0.02<z<0.07$.
\begin{table*}
\caption{Summary of ``bulk flow'' measurements ($l_{\rm max}=1$) in two redshift bins for the parameter set
{\tt param{\_}wmap} and the different models of the LF described in the text.}
\centering
\begin{tabular*}{0.95\linewidth}{@{\extracolsep{\fill}}lcccccc}
\noalign{\medskip}
\hline
\hline
\noalign{\smallskip}
& \multicolumn{3}{c}{$0.02<z<0.07$} & \multicolumn{3}{c}{$0.07<z<0.22$}
\tabularnewline
\noalign{\smallskip}
$\phi(M)$ & $v_{x}$ [km/s] & $v_{y}$ [km/s] & $v_{z}$ [km/s] & $v_{x}$ [km/s] & $v_{y}$ [km/s] & $v_{z}$ [km/s]
\tabularnewline
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Hybrid & $-227\pm 128$ & $-326\pm 113$ & $-239\pm 73$ & $-367\pm 92$ & $-439\pm 85$ & $-25\pm 71$
\tabularnewline
Fixed & $-175\pm 126$ & $-278\pm 111$ & $-147\pm 58$ & $-340\pm 90$ & $-409\pm 81$ & $-45\pm 43$
\tabularnewline
Schechter & $-151\pm 130$ & $-277\pm 116$ & $-102\pm 78$ & $-422\pm 93$ & $-492\pm 86$ & $-150\pm 74$
\tabularnewline
\noalign{\smallskip}
\hline
\hline
\end{tabular*}
\label{table2}
\end{table*}
\subsubsection{Constraints from the NYU-VAGC}
In the next step, we seek constraints on the velocity field for the case $l_{\rm max}=1$, now using the real NYU-VAGC
galaxy data. As we have argued above, the angular mask of our sample causes such ``bulk flow'' estimates to suffer
from multipole mixing and their interpretation requires the use of mock catalogs. Another mask-related problem arises
from additional degeneracies between the velocity multipoles and the LF, depending on the assumed spline point
separation. For $\Delta M=0.5$, this already becomes an issue, and the straightforward remedy is to increase the
separation to an adequate value. To account for alternative solutions and to further judge our method's robustness,
however, we will consider the following three representative approaches in our analysis:
\begin{enumerate}
\item Fix the LF to its estimate for a vanishing velocity field, i.e. use a predetermined shape of $\phi(M)$ for the
analysis. The rational of this fixed estimator is to evaluate the impact of adding degrees of freedom in the LF model.
\item Adopt a hybrid model by fitting a Schechter form to the spline-based LF estimate for a vanishing velocity field
and expressing the LF as the sum of a Schechter function and the corresponding (fixed) residual.
\item Work exclusively with the Schechter parameterization of the LF.
\end{enumerate}
Featuring the highest flexibility among the above, we expect estimates based on the hybrid LF model to be the most
reliable ones. The corresponding flow measurements will be expressed in a specific Cartesian coordinate system defined
by its $x$-, $y$-, and $z$-axes pointing toward Galactic coordinates $(l,b)\approx (81^{\circ},-7^{\circ})$,
$(172^{\circ},-1^{\circ})$, and $(90^{\circ},83^{\circ})$, respectively. In particular, the system's $z$-axis is
chosen such that it approximately penetrates the central patch of galaxies observed in the northern hemisphere, and thus
it is expected to give the tightest constraints. Using these coordinates, for example, the anomalous bulk flow incorporated
in the LasDamas mocks (see the first paragraph of section \ref{benchmark}) can be written as
$\boldsymbol{v}_{B}^{\rm T}\approx (-894,-80,442)$ in units of km s$^{-1}$. To aid the following discussion of our results, let us
further introduce $v_{K}$ as the component along the direction of this anomalous flow which points toward
$(l, b)\approx (266^{\circ}, 33^{\circ})$.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig4.eps}
\caption{Histograms of ``bulk flow'' measurements obtained from the simple NYU-VAGC mocks: shown are the recovered
distributions (black lines) and corresponding Gaussian estimates (red lines) for the two redshift bins with
$0.02<z<0.07$ (solid lines) and $0.07<z<0.22$ (dashed lines) along the direction of the anomalous flow assumed in
the first paragraph of section \ref{benchmark}.}
\label{fig4}
\end{figure}
Regarding the real NYU-VAGC galaxy sample, the inferred ``bulk flows'' for the cosmology defined by {\tt param{\_}wmap}
are summarized in table \ref{table2}. A comparison between the estimated flow components shows that the results based on
the various LF models are different, but consistent within their $1\sigma$-values, where the quoted errors are derived
from the observed Fisher information. All measurements are in very good agreement with $v_{K}\approx 120\pm 115$ and
$355\pm 80$ km s$^{-1}$ for the first and second redshift bin, respectively.
To make sense of these numbers, we compare them to the distribution of $v_{K}$ found with the help of the simple NYU-VAGC
mocks (see section \ref{section3b}) which is presented in figure \ref{fig4}. Note that the mock analysis leading to this
distribution has been performed using a pure Schechter estimator of the LF. Employing the other LF models listed above,
however, gives very similar results and will leave our conclusions unchanged.\footnote{Despite having less degrees of
freedom, this is also true in the case of the fixed LF estimator, where only the spread in the $v_{z}$-component is
significantly reduced.} As can be seen from the figure, the observed distributions in both redshift bins are well described
by Gaussian profiles with (nearly) zero mean and standard deviations of approximately $170$ and $200$ km s$^{-1}$, respectively.
In contrast to the Fisher errors, the dispersion found from the mocks includes both the cosmic signal and contributions due
to the magnitude dipole introduced in section \ref{section3b}. Repeating the mock analysis after removing the latter leads to
a decrease in the dispersions of around $9$\% and $62$\% for the first and second bin, respectively. Systematic errors induced
by the magnitude dipole are expected to increase with the redshift since a bulk flow with amplitude $v_{B}$ is expected to
induce a magnitude offset around $\delta{m}=5\log_{10}(1-v_{B}/cz)$ \cite{Nusser2011}. On the contrary, cosmic variance is
expected to decrease with the volume, and thus with the redshift. The increase of the dispersion with the redshift indicates
that the errors induced by the magnitude dipole obliterate the contribution due to cosmic variance.
We find our measurements of $v_{K}$ to be fully compatible with the distribution obtained from the mocks and consistent with
zero at a $1\sigma$ (first redshift bin) and $2\sigma$ (second redshift bin) confidence level. Given that the estimated flow
components are not necessarily uncorrelated, however, it is more appropriate to consider the joint distribution of the bulk
flow components which is adequately characterized by a multivariate Gaussian. From our set of mocks and the actual data, we
find that correlations between the different components are relatively mild and the corresponding (linear) correlation
coefficients typically take values around $0.1$--$0.3$. Choosing the hybrid model of the LF, for instance, the bulk flow
measured in both bins is consistent with zero within $1.5\sigma$ if we adopt the usual confidence levels of multivariate
normally distributed data with three degrees of freedom. Surprisingly, the bulk flow amplitudes associated with the estimated
components in this case are $v_{B}\approx 490\pm 100$ and $580\pm 80$ expressed in units of km s$^{-1}$, where the errors are
purely statistical and relatively small. Despite these rather large values, the recovered flows correspond to detections at
the $1.5\sigma$ level. The reason is that the distribution of amplitudes does not follow a Gaussian, but has a long tail and
is closely related to the $\chi^{2}$-distribution. The estimated flows are pointing toward
$(l,b)\approx (310^{\circ},-25^{\circ})\pm (30^{\circ},10^{\circ})$ and $(310^{\circ},5^{\circ})\pm (10^{\circ},15^{\circ})$
in Galactic coordinates for the first and second redshift bin, respectively. Note that changing the cosmological parameters
to {\tt param{\_}planck} or using the other LF models yields basically the same results.
\begin{figure*}
\vspace{10pt}
\includegraphics[width=0.88\linewidth]{fig5.eps}
\caption{Scatter plots of ``bulk flow'' components versus the linear evolution parameter $Q_{0}$ obtained from the
simple NYU-VAGC mocks: the panels illustrate the resulting distributions of $v_{x}$ (black squares), $v_{y}$ (red
circles), and $v_{z}$ (blue triangles) for the two redshift bins with $0.02<z<0.07$ (left) and $0.07<z<0.22$ (right).}
\label{fig5}
\end{figure*}
As for the comparison of our flow measurements with the mock catalog results, we point out that the simple NYU-VAGC
mocks are not only built and analyzed with a slightly different cosmological model, but also ignore any redshift
dependence of the peculiar velocity field which is assumed at $z=0$ (see section \ref{section3b}). For typical choices
of cosmological parameters and the redshift range of interest, this amounts to small differences $\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 3$\% and can,
therefore, be ignored in our analysis. Another concern is that fixing the linear evolution as described in section
\ref{sectionnum} causes a bias in the flow components since the monopole-like term $Q(z)$ might leak in through the mask.
To ensure that this is not the case, we plot the inferred components for both redshift bins against the estimate of
the parameter $Q_{0}$ in figure \ref{fig5}. A brief visual inspection of the scatter already indicates that
there is no evidence for a correlation between these quantities. This is confirmed by calculating the linear correlation
coefficients which turn out smaller than $0.1$ in all the cases. Together with the above findings, we thus conclude that
the SDSS galaxy data exhibit no hint toward anomalously large flows. Accounting for the known magnitude tilt in the
photometric calibration, our velocity measurements further appear fully consistent with the expectations of a
$\Lambda$CDM cosmology.
\subsection{Higher-order multipoles: constraining angular power and \boldmath{$\sigma_{8}$}}
\label{sectionvcosmo}
As we have outlined in section \ref{section2c}, the luminosity-based approach considered in this work is analogous to
the analysis of CMB anisotropies and, in principle, it is straightforward to constrain the angular velocity power spectrum
using basically the same techniques. Given the characteristics of the SDSS data and our previous findings from section
\ref{sec:bf}, however, we already expect such constraints to be rather weak and potentially biased because of the
systematic magnitude tilt described in section \ref{section3b}. Nevertheless, we shall explore the potential of this
approach and illustrate some examples involving simple velocity models.
\begin{figure*}
\vspace{10pt}
\includegraphics[width=0.87\linewidth]{fig6.eps}
\caption{Constraints on angular velocity power in a randomly selected mock for a model with $l_{\rm max}=2$: shown
are the joint $1\sigma$ and $2\sigma$ confidence regions of $\tilde{C}_{1}$ and $\tilde{C}_{2}$ (see text) for the
first (solid lines) and second (dashed lines) redshift bin, respectively, estimated with (right panel) and without
(left panel) a systematic dipole in the galaxy magnitudes. The hexagons indicate the corresponding results obtained
from directly using the known galaxy velocities.}
\label{fig6}
\end{figure*}
\subsubsection{Constraints with no cosmology priors}
Let us assume a velocity model with $l_{\rm max}=2$ and assess the impact of a tilt in the zero-point photometry
with the help of a mock galaxy catalog randomly chosen form the NYU-VAGC set. To facilitate a direct interpretation
in terms of velocities, we additionally define the dimensional quantity
\begin{equation}
\tilde{C}_{l}\equiv \sqrt{\frac{2l+1}{4\pi}C_{l}}
\label{eq:4d}
\end{equation}
which will be used in what follows below. Again, we assume the latest pre-Planck $\Lambda$CDM cosmology determined
by {\tt param{\_}wmap} and also work with the hybrid estimator of the LF (see section \ref{sec:bf}). Figure \ref{fig6}
shows the joint $1\sigma$ and $2\sigma$ confidence regions of $\tilde{C}_{1}$ and $\tilde{C}_{2}$, estimated after
maximizing the likelihood $P(C_{l})$ in eq. \eqref{eq:2c4} for the same mock, with (right panel) and without (left
panel) mimicking the systematic magnitude dipole offset (see section \ref{section3b}). Here the posterior likelihood
is constructed separately for each redshift bin after marginalizing $P\left (\boldsymbol{d}\vert a_{lm}\right )$ over the $a_{lm}$
of the respective other one, and the resulting contour lines are derived using the quadratic estimator presented in
ref. \cite{Bond1998}. The effect of a spurious magnitude dipole mostly affects the probability contour along the
$\tilde{C}_{1}$-axis, i.e. the power in the dipole, and as expected, the amplitude of the effect increases with the
redshift. To quantify the smearing introduced when one estimates velocities through luminosity variations, we further
compare the contours with the values of $\tilde{C}_{1}$ and $\tilde{C}_{2}$ inferred directly from the galaxy peculiar
velocities in the mocks (hexagons in the figure). The estimated constraints are consistent with these values within
the (large) $1$--2$\sigma$ bounds.
\begin{figure*}
\vspace{10pt}
\includegraphics[width=0.45\linewidth]{fig7.eps}
\hfill
\includegraphics[width=0.45\linewidth]{fig8.eps}
\caption{Left panel: same as figure \ref{fig6}, but now for the real NYU-VAGC galaxy data which is missing direct
velocity estimates. Right panel: adopting a model with $l_{\rm max}=3$ and fixed LF to analyze the real galaxy
data, the plot shows the marginalized joint $1\sigma$ and $2\sigma$ confidence regions of $\tilde{C}_{1}$ and
$\tilde{C}_{2}$ (solid line), $\tilde{C}_{1}$ and $\tilde{C}_{3}$ (dashed line), and $\tilde{C}_{2}$ and
$\tilde{C}_{3}$ (dotted line) for the first redshift bin.}
\label{fig7}
\end{figure*}
Repeating the analysis for the real SDSS galaxies, we end up with the confidence regions depicted in the left panel
of figure \ref{fig7}. Although it is not very constraining in the present case, our analysis restricts the $\tilde{C}_{l}$
for $l_{\rm max}=2$ to several hundred km s$^{-1}$ and is consistent with zero power. This fully agrees with the
predictions of the $\Lambda$CDM model and does not suggest any anomalous properties. We also note a striking resemblance
in contour trends with the mock result in figure \ref{fig6} from which it is tempting to deduce the existence of a
formidable dipole contamination in the real data. Since, among other uncertainties, there is still leakage due to the
survey geometry, however, strong statements like that cannot be made.
Including higher velocity multipoles with $l_{\rm max}\geq 3$, the constraints become even weaker as the level of
degeneracy increases. To give a final example, we assume another model with fixed LF and set $l_{\rm max}=3$. The
corresponding confidence regions of the different $\tilde{C}_{l}$ are shown in right panel of figure \ref{fig7}.
Note that the tighter bounds are only a consequence of reducing the available degrees of freedom.
\begin{figure*}
\vspace{5pt}
\includegraphics[width=0.9\linewidth]{fig9.eps}
\caption{Distribution of $\sigma_{8}$ estimated from the simple NYU-VAGC mocks: shown are the recovered
histograms (black lines) with (solid lines) and without (dashed lines) the inclusion of a systematic dipole
in the galaxy magnitudes, using the information in both redshift bins (left) and the first redshift bin
with $0.02<z<0.07$ only (right). Note that the Gaussian curves (red lines) are obtained by fitting the
observed cumulative distribution, excluding the cases where no large-scale power is detected.}
\label{fig8}
\end{figure*}
\subsubsection{Constraints with cosmology priors}
Next, we shall consider constraints on cosmological parameters by imposing a $\Lambda$CDM prior on the angular power
spectrum as detailed above in section \ref{section2c}. In doing so, it is convenient to divide the parameters that define
the cosmological models into two categories. The ones that characterize the background cosmology and that are used to
estimate absolute magnitudes and compute distances for sample selection, and those that characterize the density fluctuations.
Here we focus on $\sigma_{8}$ which belongs to the latter category, and assume that all other parameters are fixed to their
values in {\tt param{\_}wmap}. At the end of this section, we shall briefly discuss other choices and comment on the
possibility of constraining background parameters such as $\Omega_{m}$.
\begin{figure*}
\vspace{10pt}
\includegraphics[width=0.91\linewidth]{fig10.eps}
\caption{Raw estimates of $\sigma_{8}$ obtained from the real NYU-VAGC galaxy data: shown is the derived
$\Delta\chi^{2}$ as a function of $\sigma_{8}$ for both redshift bins (left panel) and the first redshift bin
with $0.02<z<0.07$ only (right panel), adopting the hybrid (solid line), fixed (dashed line), and Schechter
(dotted line) estimator of the LF.}
\label{fig9}
\end{figure*}
Like in the previous section, we assess the impact of a spurious tilt in the estimated magnitudes. To guarantee the validity
of linear theory, we set $l_{\rm max}=5$ which corresponds to considering physical scales above $\sim 100h^{-1}$ Mpc.
Concerning the calculation of the theoretical $C_{l}$ which is required for the prior probability and summarized in appendix
\ref{app2}, we adopt the parameterized form of the matter power spectrum $P(k)$ given in ref. \cite{EH98}. Moreover, the
galaxy redshift distribution $p(z)$ used to compute the bin-averaged velocity field given by eq. \eqref{eq:app2a} is taken to
be of the form
\begin{equation}
p(z) \propto z^{a}\exp{\left\lbrack -\left (z/\overline{z}\right )^{b}\right\rbrack },
\label{eq:4e}
\end{equation}
where the parameters $a=1.31$, $b=1.94$, and $\overline{z}=0.1$ are found by directly fitting eq. \eqref{eq:4e} to the
observed distribution. As is customary, $\sigma_{8}$ is inferred from discretely sampling the posterior probability and
interpolating the corresponding result. In our calculations, we will choose a step size of $0.05$.
Applying this procedure to the full suite of mock catalogs with and without the inclusion of a systematic magnitude
dipole, we obtain the histograms shown in figure \ref{fig8}. While the results for the combination of both redshift
bins stem from maximizing $P(C_{l})$ in eq. \eqref{eq:2c4} using the full probability $P\left (\boldsymbol{d}\vert a_{lm}\right )$,
those for the low-$z$ bin are computed by constructing the posterior probability after marginalizing
$P\left (\boldsymbol{d}\vert a_{lm}\right )$ over all $a_{lm}$ in the high-$z$ bin. Note that the former approach is actually
inconsistent as it incorrectly assumes that the $a_{lm}$ are uncorrelated between different bins. Accounting for these
missing correlations, however, leads only to differences of a few percent in the estimated values of $\sigma_{8}$ and
its error, suggesting that they may be safely neglected. As described in section \ref{section3b}, the NYU-VAGC mocks
are based on the parameter set {\tt param{\_}mock} and assume an input value of $\sigma_{8}=0.8$. In addition, we assume
a Schechter LF for the mock analysis (just like in section \ref{sec:bf}), but using the other LF models does not significantly
change the results. The spikes in the histograms at $\sigma_8=0$ correspond to the cases in which we do not detect
any power. Once we exclude those, the histograms are reasonably well represented by Gaussian distributions with
standard deviations of $\sim 0.3$--$0.4$ (solid and dashed, red lines). As is readily seen from comparing the left
and right panels, the presence of a systematic magnitude dipole (solid lines) causes a bias in the estimate of
$\sigma_{8}$ which is rather severe for higher redshifts, i.e. $0.07<z<0.22$ (left panel). As expected, removing the
dipole (dashed lines) also eliminates the bias, thus leading to the same mean value of $\sigma_{8}$ in both cases.
Expressing the bias in numbers, the dipole contribution to galaxy magnitudes amounts to a systematic shift of
$\Delta\sigma_{8}\approx 0.13$ and $\Delta\sigma_{8}\approx 0.52$ for the low-$z$ bin and the combination of both
redshift bins, respectively.
Considering now the real SDSS galaxy sample, we perform exactly the same analysis to obtain measurements of $\sigma_{8}$
for the different LF estimators introduced in section \ref{sec:bf}. Our results are presented in figure \ref{fig9} which
shows the derived $\Delta\chi^{2}$ as a function of $\sigma_{8}$, obtained using the information from both redshift bins
(left panel) and that of the first redshift bin only (right panel). Similar to what we have discovered in
our investigation of ``bulk flows'', the values based on different LF models agree very well within their corresponding
$1\sigma$ errors, and we get $\sigma_{8}\sim 1.0$--$1.1$ in the low-$z$ bin and $\sigma_{8}\sim 1.5$--$1.6$ over the full
$z$-range. Remarkably, the measured values and uncertainties closely match the inferred biased distributions of the previous
mock analysis depicted in figure \ref{fig8}. If the magnitude tilt in the SDSS data is the only relevant source of
systematic errors and sufficiently characterized by a dipole-like modulation, we can use the bias estimated from the
mocks to correct our measurements. Taking the result of the hybrid LF estimator (solid line in figure \ref{fig9}), for
example, we obtain corrected values of $\sigma_{8}=1.09\pm 0.38$ (both bins) and $\sigma_{8}=0.95\pm 0.53$ (low-$z$ bin)
which are fully compatible with each other and also consistent with the expectation of the $\Lambda$CDM model. The quoted
errors are the statistical errors inferred from the NYU-VAGC data. Note that changing the cosmology to {\tt param{\_}planck}
or choosing a different LF estimator has only a minor impact on the results.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{fig11.eps}
\caption{Scatter plots of $\sigma_{8}$ versus the evolution parameter $Q_{0}$ for the simple NYU-VAGC mocks:
using the information from both redshift bins, the plot illustrates the resulting distributions obtained with
(black squares) and without (red circles) a systematic dipole in the galaxy magnitudes.}
\label{fig10}
\end{figure}
Again, one may ask whether fixing the linear evolution as described in section \ref{sectionnum} causes an additional bias
in our measurements of $\sigma_{8}$. To answer this question, we plot the derived values of $\sigma_{8}$ for both redshift
bins against the estimate of the evolution parameter $Q_{0}$ in figure \ref{fig10}, using the simple NYU-VAGC mocks with
(black squares) and without (red circles) the magnitude dipole. As before (see section \ref{sec:bf}), the linear correlation
coefficients turn out $\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.1$, and there is no indication for a correlation between these quantities.
Of course, one is not restricted to $\sigma_{8}$, but also free to look at other cosmological parameters or various
combinations thereof. Considering the two parameters $h$ and $\Omega_{b}$ which, together, determine the baryonic matter
density, for instance, we find that the respective constraints turn out weaker than before, and are also highly
degenerate with $\sigma_{8}$. Similar statements should hold for the parameter $\Omega_{m}$. However, we do not explicitly
check this because changing $\Omega_{m}$ alters the background cosmology and implies a modification of the survey volume and
the selection of the galaxy sample. Taking this into account would substantially increase the workload of the analysis.
\subsection{Caveats}
\label{sectioncaveats}
\subsubsection{Coherent photometric errors and spurious signals}
Although our analysis of the SDSS galaxy data has tried to account for known systematics such as the zero-point photometric
tilt (see the mock description in section \ref{section3b}), there could exist additional errors in the photometric calibration.
It is important to note that the impact of such errors increases with the redshift and significantly affects the results
obtained at $z\lower.5ex\hbox{$\; \buildrel > \over \sim \;$} 0.1$ in our analysis. An example is the possibility of zero-point offsets in magnitudes between the
whole northern and southern hemispheres due to observing the galaxies in disconnected regions. If this offset was around
$0.01$--$0.02$, it should basically contribute a spurious flow of $\sim 100$--$250$ km s$^{-1}$ to the actual bulk motion
along the connecting axis, assuming a redshift of $z=0.1$ for all galaxies. Since the SDSS footprint gives much more weight
to the northern sample, however, the effect is much less pronounced. As we have verified with the help of our galaxy mocks,
it has just a mild impact on velocity measurements along the previously defined $z$-axis (see section \ref{sec:bf}), thus
leaving our conclusions unchanged. In fact, the missing evidence for any large-scale flow anomaly found in sections \ref{sec:bf}
and \ref{sectionvcosmo} already indicates that there are no other relevant systematics which would otherwise require serious
fine-tuning.
As for the photometric tilt, which constitutes the main source of systematic errors in the present analysis, it
is worth pointing out that, although difficult in practice, such a photometric bias could in principle be characterized and
corrected for by using additional information from star or galaxy counts and clustering from the very same SDSS DR7 dataset.
Another possibility is to recalibrate the SDSS photometry with the help of independent datasets. For example, this could be
accomplished using observations of the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS) which covers $3/4$
of the sky visible from Hawaii \cite{Kaiser2002, Kaiser2010} or, indeed, any future, wide galaxy survey with good photometric
stability. With this respect, a major step forward in control on the photometric calibration is expected from galaxy surveys
carried out from space like, for instance, the planned Euclid mission \cite{euclid2011}.
Imperfect corrections for Galactic extinction might also cause a systematic large-scale offset in the estimated magnitudes.
In the NYU-VAGC, this correction is based on the maps given in ref. \cite{Schlegel1998}. The recent comparison between these
and the reddening maps obtained from the Pan-STARRS1 stellar photometry \cite{Schlafly2014} does not hint at any large-scale
coherent residual with an amplitude comparable to that of the known magnitude tilt. The new dust maps that will be obtained
from the Planck data will settle the issue.
\subsubsection{Environmental dependence of the luminosity function}
There are several studies which strongly hint toward a dependence of the LF on the large-scale
environment of galaxies \cite[e.g.,][]{Balogh2001, Mo2004, Croton2005, Park2007, Merluzzi2010, Faltenbacher2010}.
However, investigations trying to shed light on the connection between luminosity and galaxy density are typically
limited to scales of a few Mpc, thus not probing the large scales relevant to our work. As a matter of fact, an
analysis addressing these environmental dependencies of galaxy luminosities on scales $\sim 100h^{-1}$ Mpc and
above is still unavailable. Nonetheless, we may get an idea from extrapolating the observed dependence into the
large-scale domain. Using that the rms of density fluctuations averaged over scales $\lower.5ex\hbox{$\; \buildrel > \over \sim \;$} 100h^{-1}$ is less
than about $0.07$, this gives rise to a particularly small effect if the overdensity of the large-scale
environment turns out to be the most important factor. As we have already suggested in ref. \cite{Nusser2013},
it should further be feasible to take such effects into account by performing measurements over independent
volumes which are classified in terms of their average density.
\subsubsection{Luminosity evolution and $K$-corrections}
The analysis conducted in this paper assumes that the evolution of galaxy luminosity can be effectively described with
a linear model of the form $Q(z)= Q_{0}(z-z_{0})$ where $z_{0}=0.1$ and $Q_{0}$ is determined according to section
\ref{sectionnum}. To assess the robustness of our results with respect to this specific model, we have carried out
a few simple tests. In a first run, we have examined the influence of varying $Q_{0}$ within its estimated $3\sigma$
confidence interval. This typically leads to changes in the measured ``bulk flow'' components of several tens of
km s$^{-1}$ and causes deviations less than $10$\% in the found values of $\sigma_{8}$. Moreover, we have explored
nonlinear evolution models with second- and third-order terms. In this case, the corresponding changes in our
velocity measurements turn out even smaller and can be safely neglected. Similarly, we have studied the impact of
different $K$-corrections using the mean correction given by eq. \eqref{eq:4b} and two-dimensional
polynomial fits as a function of redshift and $g-r$ color \cite{Chilin2010}.\footnote{As advised by the SDSS
collaboration, we use model magnitudes to estimate the $g-r$ colors of galaxies.} Again, the resulting differences
are marginal and do not affect any of our conclusions.
\section{Conclusions}
\label{section5}
We have exploited the well-known fact that peculiar motion induces spatially-coherent variations in the observed
luminosity distribution of galaxies to probe the cosmic velocity field at $z\sim 0.1$ from the luminosity distribution
of SDSS galaxies in the NYU-VAGC. The method adopted here extends the maximum-likelihood approach proposed in ref.
\cite{Nusser2011} to constrain the peculiar velocity field beyond the bulk flow component. Considering the bin-averaged
peculiar velocity field in two different redshift bins, $0.02<z<0.07$ and $0.07<z<0.22$, we have demonstrated how
the method permits bounds on the corresponding angular velocity power spectrum and cosmological parameters. The main
results of our analysis can be summarized as follows:
\begin{itemize}
\item
To assess the robustness of our analysis against potential systematic errors, we have used a suite of mock galaxy catalogs
obtained both from numerical simulations and from the NYU-VAGC dataset itself to match the real data as close as possible.
We have identified three main obstacles which potentially hamper the analysis of the SDSS data: the survey geometry, which
causes mixing between different moments, the possible degeneracies between the velocity multipoles and the estimator of the
LF, and the presence of a coherent photometric tilt of about $0.01$ magnitude across the survey region. While the impact of
mode-to-mode mixing can be readily quantified by modeling the sky coverage and the influence of the LF has been evaluated
by applying different estimators to the mocks, the latter effect is less trivial to account for. Here we have modeled the
zero-point photometry offset by adding a randomly oriented dipole normalized such that the corresponding rms over all
galaxies is $\delta m_{\rm dipole}=0.01$ for each individual mock galaxy catalog. Our results suggest that the systematic
tilt in the observed galaxy magnitudes is sufficiently described by this dipole contribution.
\item
Accounting for the known systematics in the SDSS photometry, the estimated ``bulk flows'' are consistent with the predictions
of the standard $\Lambda$CDM model at $\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 1$--$2\sigma$ confidence in both redshift bins. The combined analysis of the
corresponding three Cartesian components further corroborates this result. Using an independent estimator, this confirms the
findings of the CMB studies in refs. \cite{Osborne2011, planck_bf} which provide an upper bulk flow limit of a few hundred
km s$^{-1}$ at a $95$\% confidence limit on similar scales.
\item
Our analysis yields direct constraints on the angular velocity power spectrum $C_{l}$ (considering terms up to the octupole)
defined in section \ref{section2c}, independent of a prior on the cosmological model. All of the estimated $C_{l}$ are
consistent with the corresponding theoretical power spectra of the $\Lambda$CDM cosmology.
\item
Assuming a prior on the $C_{l}$ as dictated by the $\Lambda$CDM model with fixed density parameters and a Hubble constant,
we have used the method to infer the parameter $\sigma_{8}$ which determines the amplitude of the velocity field. After
correcting for known systematics, we obtain $\sigma_{8}\approx 1.1\pm 0.4$ for the combination of both redshift bins and
$\sigma_{8}\approx 1.0\pm 0.5$ for the low-redshift bin only. As anticipated, the found constraints on velocity moments and
$\sigma_{8}$ are not very tight. However, they show the validity of our approach in view of future analyses with different
datasets.
\item
As for the encountered data-inherent issues, current and next-generation spectroscopic surveys are designed to alleviate
most of them, thanks to their large sky coverage (e.g., eBOSS\footnote{ http://www.sdss3.org/future/}, DESI \cite{bigboss2011,
Levi2013}) and improved photometric calibration in ground-based surveys (e.g., PAN-STARRS \cite{Kaiser2002, Kaiser2010}) and
especially in space-borne experiments like Euclid \cite{euclid2011}. Note that since uncertainties in the measured redshifts
play a little role in our error budgets, the method is also suitable for application to wide photometric redshift surveys
such as the 2MASS Photometric Redshift catalog (2MPZ) \citep{Bilicki2014} and, again, Euclid.
\item
These excellent observational perspectives give us confidence that the method considered here will become a full-fledged
cosmological probe, independent and alternative to the more traditional ones based on galaxy clustering, gravitational
lensing and redshift space distortions. We expect that combining all these approaches will result in superior control
over potential systematic errors that might affect the estimate of cosmological quantities, chief among them the growth
rate $f(\Omega)$ of density fluctuations \cite{Nusser2012}.
\item The main interest here is the methodological aspect since the novel approach to estimate the angular velocity power
spectrum or cosmological parameters, developed in analogy to the statistical treatment of CMB anisotropies, is to be regarded
as a proof of concept guiding future analyses. As a final remark, we point out that it should be conceivable to reverse the
ansatz taken in this work, allowing one to constrain luminosity evolution and to improve the photometric calibration of a
galaxy sample in a given cosmological framework.
\end{itemize}
\begin{acknowledgments}
This research was supported by the I-CORE Program of the Planning and Budgeting Committee, THE ISRAEL SCIENCE FOUNDATION (grants No. 1829/12
and No. 203/09), the German-Israeli Foundation for Research and Development, the Asher Space Research Institute, and in part by the Lady
Davis Foundation. M.F. is supported by a fellowship from the Minerva Foundation. E.B. is supported by INFN-PD51 INDARK, MIUR PRIN 2011
``The dark Universe and the cosmic evolution of baryons: from current surveys to Euclid'', and the Agenzia Spaziale Italiana from the
agreement ASI/INAF/I/023/12/0.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
Quantum key distribution (QKD) \cite{Gisin02,Scarani09} is a method by which two parties, Alice and Bob, may generate a shared secret key over an insecure quantum channel monitored by an eavesdropper, Eve. Any QKD protocol relies on several assumptions, namely, any eavesdropper must obey the laws of quantum mechanics; Alice and Bob have the freedom to choose at least one of two measurement settings; and there is no classical information leaking from Alice or Bob's laboratories. Most conventional QKD protocols further assume that Alice and Bob have near perfect control of their measurement devices as well as their state preparation. Device-independant QKD \cite{Mayers04,Acin07,Ekert14} is a protocol that, remarkably, is free from making these additional assumptions; Alice and Bob need no knowledge of the inner workings of their devices nor even the dimension of the space their quantum states reside in.
In this paper, we use the novel approach of combining the encoding scheme of \cite{GKP} with the results of \cite{Mayers04,Acin07} to create a device-independent quantum cryptography protocol for continuous variables (CVs). CV quantum information offers higher efficiency detectors, cheap off-the-shelf components and the experimentally accessible Gaussian resources. Furthermore, by encoding the CV space of a harmonic oscillator into a finite-dimensional code space we are able to take advantage of results which have previously only been applied to discrete-variable (DV) QKD.
The first proposals for continuous-variable QKD \cite{Weedbrook12} relied on `non-classical' states of light such as squeezed states \cite{Hillary00,Gottesman01}. In fact, one of these protocols was proven unconditionally secure \cite{Gottesman01}. As the field matured it was recognized that such non-classical states were not required and that the more experimentally available class of coherent states were sufficient \cite{Grosshans02,Weedbrook12}.
Device-independent QKD provides a way by which two parties may share a private key despite having no knowledge of the inner workings of their respective devices. Conversely, in conventional QKD protocols it is regularly assumed that both parties have a high degree of control over both state preparation as well as measurement. Although, recently relaxing the condition of trusting the measurement device was achieved \cite{Lo12,Braunstein12}. The security in this device-independent approach comes instead from the fact that the two parties are able to violate a Bell inequality \cite{Brunner13}, which can remarkably be used to put a bound on the amount of information that a potential eavesdropper could, in principle, obtain.
Here we introduce a CV version of device-independent QKD. Our protocol goes as follows. Alice first generates a Bell state which has been created using an encoding based on the Gottesman-Kitaev-Preskill scheme where a qubit is encoded into an infinite dimensional space of a harmonic oscillator. After this the protocol continues in a similar fashion where she keeps one encoded qubit and sends the other qubit to Bob over an insecure quantum channel. Hence, the results of DV device-independent QKD can then be applied to the system yielding the first implementation of device-independent QKD for CVs.
It is known in the field of CV quantum information that all Gaussian resources are insufficient for violating a Bell inequality \cite{Bell87,Paternostro09}. This means that one should already expect non-Gaussian states or measurements as being a requirement \cite{Wenger03,Garcia05,Cavalcanti07,Brask12,Qian14}, despite the fact that they are typically more difficult to produce in a lab. This highlights the challenges faced when attempting to create a CV version of device-independent QKD because most current CV-QKD protocols use Gaussian states. Fortunately, if we use for example, a single mode of the electromagnetic field as our harmonic oscillator, we are able to use CV resources, including high efficiency detectors and off-the-shelf components. The major drawback of DV device-independent QKD is that in order to close the detector loophole one needs high efficiency detectors \cite{Rowe01}. The detector loophole issue is often overcome by CV quantum information where we can take advantage of such high detection efficiencies~\cite{Wenger03,Garcia05,Cavalcanti07,Brask12,Qian14,Reid13}.
This paper is structured as follows. In \sec{background} we discuss separately the necessary encoding scheme as well as the results of DV device-independent QKD. In \sec{cvdiqkd} we relate these concepts to CV quantum information and discuss formally the kind of measurements that are necessary. Following this we investigate the resources required in order to implement the protocol in \sec{resources}. Since we are only capable of making approximations of the desired encoding in the real world, we consider the effects of such approximations on the encoding and resulting key rate in \sec{finitesq}. Finally, \sec{conclusions} presents some discussions and concluding comments as well as some interesting open questions.
\section{Background}
\label{sec:background}
The premise of this paper is to propose a method of implementing device-independent QKD with CV states. This is accomplished by embedding a two-level Hilbert space into the full infinite-dimensional space and then using results from DV-QKD. Here we discuss the encoding scheme proposed by Gottesman, Kitaev, and Preskill (GKP) in \cite{GKP} as well as the DV version of device-independent QKD.
\subsection{GKP encoding}
\label{sec:encoding}
The GKP encoding \cite{GKP} provides a method to encode a qubit in the infinite-dimensional space of an oscillator in such a way that one can protect against arbitrary, but small, shifts in the canonical variables $q$ and $p$ as well as carrying out fault-tolerant universal quantum computation on the encoded space \cite{GKP,Menicucci14}. The stabilizer generators of a two-dimensional Hilbert space in an infinite-dimensional Hilbert space with canonical variables $q,p$ are given by~ \cite{Gottesman01}
\begin{align}
S_q=\exp(2iq\sqrt\pi), \hspace{1mm} S_p=\exp(-2ip\sqrt\pi).
\end{align}
The stabilizers are simply shift operators for $q,p$, and if the eigenvalues are $S_q=S_p=1$ then the allowed values of $q$ and $p$ are integer multiples of $\sqrt \pi$. Since the codewords are invariant under shifts by integer multiples of $2\sqrt \pi$ we can define a basis for the encoded qubit as
\begin{align}
\ket{\bar j_L}\propto \sum_{s\in\mathbb Z} \ket{(2s+j)\sqrt\pi}_q,
\end{align}
for $j=0,1$, and where the subscript $q$ indicates the $q$ (`position')-basis. These states can be approximated optically using Schr\"odinger cat states \cite{Vasconcelos10}, or by a variety of other methods \cite{Travaglione02,Pirandola04,Pirandola05,Pirandola06}. Encoded Pauli gates are defined as $\bar Z\equiv \exp(iq\sqrt\pi)$ and $\bar X\equiv \exp(-ip\sqrt\pi)$; since these operators commute with the stabilizers they also preserve the code subspace.
The set of Clifford operations on the encoded subspace correspond to symplectic (or Gaussian) transformations on the CV space of the oscillator; these operations can be implemented in a fault tolerant way \cite{GKP}. To achieve universal quantum computation we must be able to implement a non-Clifford gate on the encoded subspace \cite{Lloyd99}, for example, the addition of a $\pi/8$-gate (T-gate) to the Clifford group will make for a universal set of gates. The T-gate can be implemented with a non-symplectic transformation on the oscillator; this is more experimentally difficult than symplectic transformations and requires a non-Gaussian resource such as photon counting. The physical resources required to implement these gates are discussed in more detail in \sec{resources}.
\subsection{Device-independent quantum key distribution}
\label{sec:diqkd}
The DV-QKD protocol \cite{Mayers04,Acin07} begins with Alice and Bob sharing a quantum channel that emits pairs of entangled particles. To consider the worse case scenario, we allow Eve full control over the source \cite{Lo99} which, if she is honest, emits the state $\ket{\psi_{AB}}=1/\sqrt 2(\ket{00}+\ket{11})$. But in general she is free to create any arbitrary state $\rho_{ABE}$ which may be entangled between not only Alice and Bob but herself as well. To generate a secret key, Alice chooses a basis to measure in from $\{A_0,A_1,A_2\}$ while Bob chooses a basis from $\{B_1,B_2\}$ and they get outcomes of $a_i,b_j\in \{+1,-1\}$, respectively~\cite{Pironio09}. After all measurements are performed, if Alice had chosen measurement $A_0$ and Bob chosen measurement $B_1$ they extract a single bit of raw key corresponding to their measurement outcome. Instead, if they had measurement settings corresponding to $\{A_0,B_2\}$, their outcomes are completely uncorrelated and so this case is discarded. For all other measurement settings Alice and Bob use their results to violate the CHSH inequality \cite{Brunner13}
\begin{align}
\mathcal S&=\langle a_1b_1 \rangle +\langle a_1 b_2 \rangle + \langle a_2b_1 \rangle - \langle a_2b_2 \rangle\leq 2.
\end{align}
The CHSH inequality puts a bound on the values of $\mathcal S$ consistent with local hidden-variable theories in accordance with Bell's theorem \cite{Bell,CHSH}. Violation of this inequality by quantum mechanics arises due to the fact that entanglement can provide nonlocal correlations that cannot be produced by shared randomness. If Alice and Bob share a nonlocal correlation then, regardless of how this correlation came to exist, Eve cannot have full knowledge of the correlation or else she would be in possession of a local variable capable of reproducing the correlations \cite{Acin07}.
A set of measurements which give the desired behaviour in the above protocol and which maximize the violation of the CHSH inequality are given by \cite{Pironio09}
\begin{align}
\begin{aligned}
\label{eq:bases}
A_0=B_1&= Z, & A_1&= 1/\sqrt 2 ( Z + X),\\
B_2&= X, & A_2&=1/\sqrt 2( Z - X).
\end{aligned}
\end{align}
For the moment, $Z$ and $X$ in the above expression \eq{bases} have no relation to the encoded Pauli gates $\bar Z$ and $\bar X$; although we will make this connection in \sec{cvdiqkd}. The main result shown by Ac\'in et al. \cite{Acin07} is that the Holevo quantity between Eve and Bob, after Alice and Bob have symmetrized their marginals, is bounded as
\begin{align}
\chi(B_1{:}E)&\leq h\left(\frac{1+\sqrt{(\mathcal S/2)^2-1}}{2}\right),
\end{align}
where $\chi(B_1{:} E)=S(\rho_E)-\frac{1}{2}\sum_{b_1=\pm 1}S(\rho_{E|b1})$ is the Holevo quantity and $h=-p\log_2p-(1-p)\log_2(1-p)$ is the binary entropy. This provides a method which Alice and Bob can use to keep Eve honest and bound her knowledge using only their violation of the CHSH inequality.
\section{Continuous-Variable Device Independence Protocol}
\label{sec:cvdiqkd}
The CV version of device-independent QKD begins with Alice creating an encoded Bell state. This encoding is based on the GKP encoding as given in \sec{encoding}. Once this Bell state is created she keeps one qubit for herself and sends the other entangled qubit to Bob over an insecure and lossy quantum channel. Apart form this initial encoding, the protocol follows the same steps as in typical DV-QKD protocols \cite{Mayers04,Acin07}.
A set of measurements which maximize the violation of the CHSH inequality, for the encoded state $\ket{\Phi^+}=1/\sqrt 2(\ket{\bar 0\bar 0}+\ket{\bar 1 \bar1})$, consist of measurements $A_1,A_2,B_1,B_2$, as defined in \sec{diqkd}, which act on the encoded subspace. We can destructively measure the observables $\bar Z$ and $\bar X$ by performing a suitable homodyne measurement of the $\hat q$ or $\hat p$ quadrature, respectively. By measuring the $\hat q$ quadrature we expect that the only outcomes possible will be integer multiples of $\sqrt\pi$; even multiples corresponding to a $\ket{\bar 0}$ state and odd multiples corresponding to $\ket{\bar 1}$. Imperfections in the measurement and the encoded state will result in other measured values, but we can apply classical error correction and adjust the value to the nearest $k\sqrt\pi$ for an integer $k$. The outcome of the measurement $\bar Z$ is then given by $(-1)^k$. We can measure the other three observables by first applying a change of basis gate which takes us to the $\bar Z$ basis, and in this way we need only consider homodyne measurements of the $\hat q$ quadrature.
We assume in this section that we are able to implement Clifford gates as well as $\pi/8$-gates on our encoded space and also that we can carry out homodyne measurements on the CV space; the resources required to do this are discussed in \sec{resources}. From here onwards we drop the over-bar notation to denote encoded operations; all gates are to be understood as acting on the encoded space while symplectic transformations are understood to be in relation to the oscillator.
It is readily seen that we can measure in the $X$ basis by using the change of basis gate $H$, and one can easily verify that $ H^\dagger X H= Z$. Since the Hadamard gate $H$ is in the Clifford group we can implement an encoded $H$ by carrying out symplectic transformations on the full CV space. Unfortunately, it is not possible to change from the $A_1$ or $A_2$ basis to the $Z$ basis by using only Clifford gates, which means that we will need to go beyond symplectic transformations in the CV space. This can be readily seen by recognizing that $A_1= H$, suppose there existed a Clifford gate $C$ such that $C^\dagger H C= Z$. This would imply that $C Z C^\dagger= H$ and so $C$ is not a Clifford gate by definition.
The required change of basis gates can be calculated as: $\mathcal I B_1 \mathcal I = Z,~H^\dagger B_2 H=Z,~\alpha^\dagger A_1 \alpha=Z$, and $\beta^\dagger A_2\beta =Z$, where $\alpha=PHTHP$, $\beta=ZPHTHP$ and $T$ is an encoded $\pi/8$-gate. It is important to note that while the latter two gates are not Clifford gates, they can be decomposed exactly as a composition of Clifford gates with only one non-Clifford $T$-gate. Furthermore, it is possible to shift the problem of implementing a $T$-gate to a state preparation problem, and since preparation can be done `offline' we require only Gaussian operations and one auxiliary state to carry out our CV device-independent QKD.
\section{Required resources}
\label{sec:resources}
In order for Alice and Bob to implement the necessary measurements they must be able to perform gates on the encoded states as well as homodyne detection on one quadrature. The necessary set of gates include $ H, P, T$ (no need for $Z$, since $ Z= P^2$). The first two gates correspond to Clifford operations while the last one is a non-Clifford gate. The set of Clifford gates on the encoded states correspond to symplectic transformations on the CV space, given as \cite{GKP}: $H: (q,p)\rightarrow (p,-q)$, $P: (q,p)\rightarrow (q,p-q)$, and $C_{NOT}: (q_1,p_1,q_2,p_2)\rightarrow (q_1,p_1-p_2,q_1+q_2,p_2)$. The encoded $C_{NOT}$ gate is used not by Alice or Bob but in the preparation of the encoded Bell state by Eve.
In order to implement an encoded $\pi/8$-gate we need a non-symplectic transformation which requires a non-Gaussian resource. The addition of photon counting to Gaussian resources is sufficient to carry out non-symplectic transformations. In particular, one is able to create either a $\pi/8$ state or a cubic phase state which can then be used to implement a $T$-gate on the encoded space \cite{GKP,Gu09}. Fortunately, one can generate these states offline and use them as required throughout the protocol, effectively shifting the issues of non-Gaussian operations to state preparation. In this way, one needs only to have a supply of non-Gaussian states and be capable of performing symplectic transformations (including homodyne detection) in order to implement the QKD protocol. In the case of an optical mode, the set of symplectic transformations can be achieved with linear optics (phase shifters and beam splitters) and squeezing operations (non-linear crystals).
Fortunately, Alice and Bob do not need to choose a measurement basis, which is used to check for a CHSH violation, very often; the probability to choose between the possible options need not be uniform, although this would work as well. If we suppose that Alice and Bob share $N$ quantum states, it is enough to use $\sim \sqrt N$ pairs to check for a CHSH violation, so long as the measurements are causally independent \cite{Masanes11}. This condition would be satisfied for memoryless devices, or devices which could have internal memory reliably cleared after every run. One protocol \cite{Masanes11}, also provides security against coherent attacks, which is the most general form attack. Since we have chosen the measurement basis corresponding to generating a key as $Z$, this means that in the limit of large $N$ almost all of the time we need only perform Gaussian operations. Hence, we need only perform non-Clifford operations a small fraction of the time, in order to estimate the CHSH violation and thus keep the eavesdropper honest. Many other such DV device-independent QKD protocols exist and offer different key rates with different underlying assumptions \cite{Barret05,Lim13,Vazirani12}, but typically one still requires the ability to make measurements in a set of four bases which violates the CHSH inequality.
\section{Gaussian finite-squeezing effects}
\label{sec:finitesq}
In practice, the encoded GKP states will not consist of delta peaks at $\sqrt\pi$ intervals, but instead the peaks will have some finite width and they will be modulated by a larger envelope to ensure the state is of finite energy. One way to produce an ideal GKP state is to prepare a momentum eigenstate $\ket{p=0}$, and then measure the value of $q~ \text{(mod 2}\sqrt\pi)$. Since the position is completely undetermined for a momentum eigenstate all values of $q$ are equally likely, and this measurement will project out a state that differs from a $ Z$ eigenstate by a shift of $q$ which can then be corrected.
If instead of an unphysical momentum eigenstate, which corresponds to infinite squeezing, we can consider a finitely-squeezed state given by $\psi_{sq}(p)=\pi^{-1/4}\kappa^{-1/2}\exp\left(-\frac{1}{2}p^2/\kappa^2\right)$ \cite{Braunstein05}, where $\kappa=e^{-r}$ for squeezing parameter $r\in\left[0,\infty\right)$. In the position representation this state is given by $\psi_{sq}(q)=\pi^{-1/4}\kappa^{1/2}\exp(-\frac{1}{2}q^2\kappa^2)$.
An ideal homodyne measurement of $q$ is a projection-valued measure (PVM) with projectors corresponding to position eigenstates $P_x=\ket x\bra x$, or infinitely squeezed states in position. If we allow the homodyne measurement to have a Gaussian acceptance of width $\Delta$, we replace the PVM with a positive-operator valued measure (POVM) which consists of an ideal homodyne measurement convolved with a Gaussian window. This leads to POVM elements given by
\begin{align}
\label{eq:povm}
\Pi_x&=(2\pi\Delta^2)^{-1/2}\int_{-\infty}^\infty dy e^{-\frac{1}{2}(x-y)^2/\Delta^2}\ket y\bra y.
\end{align}
An ideal measurement of $q~ \text{(mod 2}\sqrt\pi)$ is described by the PVM with elements $P'_x=\sum_{s=-\infty}^\infty P_{x-2s\sqrt\pi}$ for $x\in[0,2\sqrt\pi)$, and if we let $P_x\rightarrow \Pi_x$ we obtain the result for a homodyne detector with a Gaussian acceptance. Without loss of generality suppose we obtain a result corresponding to $\Pi_0$, then the state will be transformed to $\psi_{\bar 0}(q)\propto \sum_{s=\mathbb Z} \exp(-\frac{1}{2}q^2\kappa^2) \psi'_{sq}(q+2s\sqrt\pi)$,
where $\psi'_{sq}$ is a squeezed vacuum state in position with width $\Delta$. If we obtain a result other than $\Pi_0$ we can simply apply a shift to correct the state. This is of the same type of approximate codeword proposed in the GKP paper \cite{GKP}. Notice that our initial squeezing determines the size of the overall envelope, width $\kappa^{-1}$, while the precision of our homodyne measurement determines the width of the individual peaks.
If we further approximate by replacing $\exp(-\frac{1}{2}q^2\kappa^2)\rightarrow \exp(-\frac{1}{2}(2s\sqrt\pi)^2\kappa^2)$ in the summation above, which corresponds to scaling each peak by a constant factor, we find
\begin{align}
|\psi_{\bar 0}(q)|^2&=\frac{2\kappa}{\Delta\sqrt{\pi}}\sum_{s=-\infty}^\infty e^{-4\pi\kappa^2 s^2}e^{-(q-2s\sqrt\pi)^2/\Delta^2}.
\end{align}
We can correct for shifts in the position which are less than $\sqrt\pi/2$, and thus bound the error by adding up the contribution from all of the tails further than $\sqrt\pi/2$ from their respective peak. Assuming that $\kappa\sqrt\pi\ll 1$ the probability of error is bounded as $P_e<2\Delta^2/(\kappa\pi)\exp{(-\frac{1}{4}\pi/\Delta^2)}$ \cite{Gottesman01}.
The errors from incorrectly identifying an encoded state will determine the amount by which one is able to violate the CHSH inequality. Consider one term in the CHSH quantity $\mathcal S$. The correlator is defined as $\langle a_ib_j\rangle=P(a=b|ij)-P(a\neq b|ij)$ for outcomes $a,b$ and measurement choices $i,j$. If we assume that our gates are perfect then all errors will come from incorrectly identifying an encoded state. We can calculate the value of $\mathcal S$ after error correction by computing the expectation values of the various measurements. This value is plotted in \fig{keyrate}, and it can be seen that we start to violate the CHSH inequality for parameters $\Delta=\kappa$ corresponding to squeezing greater than 5 dB. This shows that the value of the CHSH quantity is scaled according to the error rate, assuming perfect gates.
The quantum bit error rate (QBER) \cite{Scarani09} is defined as $Q=P(a\neq b|01)=2P_e(1-P_e)$ since we are only extracting a key for the cases where Alice does measurement $A_0$ and Bob does measurement $B_1$. This corresponds to either Alice or Bob incorrectly identifying the state while the other party does not make an error. The secret-key rate $r$, under collective attacks, with one-way classical post-processing from Bob to Alice, is lower bounded by the Devetak-Winter rate \cite{Acin07,Devetak05}
\begin{align}
r\geq r_{DW}=I(A_0{:} B_1)-\chi(B_1{:} E),
\end{align}
where $I(A_0{:}B_1)=1-h(Q)$ is the mutual information between Alice and Bob ($h$ being the binary entropy), and $\chi(B_1{:} E)$ is the Holevo quantity.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.84]{keyrate.pdf}
\caption{(color online) The extractable secret-key rate is plotted as a function of the squeezing for the symmetric case $\Delta=\kappa$, where $\Delta$ is the width of the individual peaks and $\kappa^{-1}$ is the width of the Gaussian envelope in the GKP encoding. Note that currently the maximal amount of single-mode squeezing achieved is $12.7$ dB \cite{Eberle10,Mehmet11}. The shaded region indicates a violation of the CHSH inequality.}
\label{fig:keyrate}
\end{figure}
In \fig{keyrate}, we plot both the QBER and the key rate $r$. Notice that the extractable key rate remains zero even for values of $\mathcal S$ slightly larger than two. The key rate grows rapidly for squeezing beyond 6 dB, for example, a squeezing of 10 dB yields a key rate of $\approx 98\%$. Note that the well known critical QBER of $11\%$ for BB84 \cite{Shor00} as well as $7.1\%$ for DV device-independent QKD \cite{Acin07} are higher than the $\approx 3.5\%$ critical QBER for this proposal. This is due to the fact that one requires a suitable enough approximation to a GKP encoded state in order to have a high enough violation of the CHSH inequality, and by doing so one immediately achieves a corresponding low probability of error $P_e$. Typically one desires a high critical QBER as it generally tolerates more imperfections in the protocol. However, in this case the difficulty arises from the need to violate the CHSH inequality and if one is able to do so then one already obtains a small QBER. Intuitively, as the width of the individual peaks $\Delta$ in the encoded state become larger, and equivalently the QBER, the overall state resembles a Gaussian state and thus cannot violate the CHSH inequality.
\section{Discussion of loss, comparison to discrete-variables and conclusion}
\label{sec:conclusions}
By harnessing the results of discrete-variable QKD, with a qubit encoding in a harmonic oscillator, we provided the first device-independent QKD protocol for continuous variables. This protocol derived its security from the ability to violate a Bell inequality and, remarkably, does not require Alice or Bob to know the inner workings of their devices. We showed how both the CHSH violation and the resulting extractable key-rate depended on the quality of the approximate codewords. We also showed that, in terms non-Gaussian resources, we required a modest one $T$-gate for each of $\sim \sqrt N$ of $N$ total Bell pairs. Thankfully, from an experimental point of view, what this means is that only Gaussian operations (e.g., homodyne detection) are needed most of the time.
It should be noted that our encoding scheme is experimentally challenging. However, it is still practical, with many proposals already existing \cite{Vasconcelos10,Travaglione02,Pirandola04,Pirandola05,Pirandola06}. It is hoped that our paper will further motivate experimental advances using such encodings. Given the technological challenges, distances in our scheme will be limited (although not fundamentally). However, it should be noted that such limitations are also faced by the discrete-variable version of device-independent QKD, which is currently limited to a few kilometers. This is because it requires a detection efficiency of approximately $95\%$ to achieve a key rate on the order of $10^{-10}$ per pulse \cite{Marcos11}.
Interestingly, one can also consider the distances over which our continuous-variable protocol will perform well. We do this by calculating the Wigner function of an approximate encoded state and then send it through an amplitude damping channel. We numerically find that, for example, at 2.3km, with 0.2dB/km loss, we get a key rate of 0.35 bits/state. This is comparable to discrete-variable device-independent QKD where such schemes are limited to only a few kilometers \cite{Marcos11}. Furthermore, we note that the distance of our protocol can also be improved by using distillers as was shown for discrete-variable states \cite{Gisin10} or by the application of heralded amplifiers~\cite{Lutkenhaus11}.
It is an interesting open question whether one can devise a device-independent continuous-variable QKD protocol with more readily accessible states. One possible avenue to explore is lifting the requirement of a CHSH inequality violation by considering a protocol where only one party trusts their device. This one-sided device-independent QKD requires one to violate only an EPR-steering inequality \cite{Reid13}, which amounts to Alice and Bob checking that they have entanglement and ruling out local hidden state models \cite{Branciard12}.
C.W. acknowledges support from NSERC. We are grateful to Hoi-Kwong Lo and Norbert L\"utkenhaus for fruitful discussions.
\twocolumngrid
|
\section{Introduction}
The OPERA experiment \cite{1}, located at the Gran Sasso laboratory (LNGS), aims at observing the $\nu_{\mu}\rightarrow\nu_{\tau}$ neutrino oscillation in the direct appearance mode in the CERN neutrino to Gran Sasso (CNGS) \cite{3a,3b} beam by detecting the decay of the $\tau$ produced in charged current (CC) interactions. A detailed description of the detector can be found in \cite{1,2a,2b,2c,2d,2e,2f}. The OPERA detector consists of two identical Super Modules (SM), each of them consisting of a target area and a muon spectrometer, as shown in Fig. \ref{strauss-fig1}. The target area consists of alternating layers of scintillator strip planes and target walls. The muon spectrometer is used to reconstruct and identify muons from $\nu_{\mu}$-CC interactions and estimate their momentum and charge.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig01-eps-converted-to.pdf}}}
\caption{Picture of the OPERA detector, with a view of a reconstructed neutrino interaction occurring in the 2nd Super Module.}
\label{strauss-fig1}
\end{myfigure}
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig14-eps-converted-to.pdf}}}
\caption{The CNGS neutrino beamline. Figure from \cite{6}.}
\label{author-fig4}
\end{myfigure}
The target walls are trays in which target units of $10\times12.5\mathrm{ cm^2}$ and a depth of 10\,$X_0$ in lead (7.9\,cm) with a mass of around 10~kg each are stored: they are also refered to as bricks. A brick is formed by alternating layers of lead plates and emulsion films (2 emulsion layers separated by a plastic base) building an emulsion cloud chamber (ECC). This provides high granularity and high mass, which is ideal for $\nu_{\tau}$ interaction detection. Fig. \ref{strauss-fig3} left shows an image of the unwrapped ECC brick. On the right, the arrangement of the scintillator strip planes (Target Tracker, TT) and the ECC is shown. Note an extra pair of emulsion films in a removable box called a changeable sheet (CS), shown in blue. The total mass of each target area is about 625\,tons, leading to a target mass of 1.25\,ktons for 145'000 bricks.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig03-eps-converted-to.pdf}\includegraphics{thomasstrauss_2012_0n_fig05-eps-converted-to.pdf}}}
\caption{$\tau$ detection principle in OPERA.}
\label{strauss-fig3}
\end{myfigure}
\section{DAQ and analysis}
The information from the reconstructed event, recorded by the electronic detectors, is used to predict the most probable ECC for the neutrino interaction vertex \cite{4}. A display of a reconstructed event is indicated in Fig. \ref{strauss-fig1}. Fig. \ref{strauss-fig2} shows the procedure for localizing the event vertex in the ECC, by extrapolating the reconstructed tracks from the electronic detector to the CS emulsion films.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig02-eps-converted-to.pdf}}}
\caption{Event detection principle in the OPERA experiment. The candidate interaction brick is determined from the prediction of the electronic detector (blue). The changeable sheet (CS) is used to confirm the prediction. From the CS result, the tracks are followed up to the interaction point inside the ECC.}
\label{strauss-fig2}
\end{myfigure}
The signal recorded in the CS films will confirm the prediction from the electronic detector, or will act as a veto and trigger a search in neighboring bricks to find the correct ECC in which the neutrino interaction is contained. After a positive CS result, the ECC will be unpacked and the emulsion films are developed and sent to one of the various scanning stations in Japan and Europe. Dedicated automatic scanning systems allow us to follow the tracks from the CS prediction up to their stopping point. Around these stopping points, a volume of 1\,cm$^2$ times 15\,emulsion films will be scanned to find the interaction vertex. As illustrated in Fig. \ref{strauss-fig2}, only track segments in the active emulsion volume are visible and a reconstruction of the event is needed to find tracks and vertices. A dedicated procedure, called a ``decay search" is used to search for possibly interesting topologies, like the $\tau$-decay pictured in Fig. \ref{strauss-fig2}. The accuracy of the track reconstruction goes from cm in the electronic detector, down to mm for the CS analysis and to micrometric precision in the final vertex reconstruction (after aligning the ECC emulsion plates with passing-through cosmic-ray tracks).
\newline
\subsection{Tau detection} $\tau$ detection is only possible due to the micrometric resolution of the emulsion films, as it allows us to separate the primary neutrino interaction vertex from the decay vertex of the $\tau$ particle. The most prominent background for $\tau$ decay is either hadron scattering or charged charm decays. The background from hadron scattering can be controlled by cuts applied to the event kinematics. The background due to charm can be reduced by identifying the muon at the primary vertex, as charm will occur primarily in $\nu_{\mu}$-CC interactions (for further details see \cite{2e,4}). After topological and kinematical cuts are applied, the number of background events in the nominal events sample is anticipated to be 0.7 at the end of the experiment.
In 2010 the first $\nu_{\tau}$ candidate was reconstructed inside the OPERA emulsion. It was recorded in an event classified as a neutral current, as no muon was identified in the electronic detector. To crosscheck the $\tau$ hypothesis, all tracks were followed downstream of the vertex until their stopping or their re-interaction point. They were all attributed to be hadrons, and no soft muon ($E<2$\,GeV) was found. In 2010, the expected total number of $\tau$ candidates was 0.9, while the expected background was less than 0.1 events. More details on the analysis are presented in \cite{2e}. In Fig. \ref{strauss-fig8} shows a picture of this event. Track number 4, labeled as the parent, is the $\tau$, decaying into one charged daughter track. The two showers are most likely connected to the decay vertex of the $\tau$, rather than activity from the primary interaction. Thus the decay is compatible with: $\tau\rightarrow\eta^-(\pi^-+\pi_0)\nu_{\tau}$.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig08-eps-converted-to.pdf}}}
\caption{Display of the 2010 $\tau^-$ candidate event. Top left: view transverse to the neutrino
direction. Top right: same view zoomed on the vertices. Bottom: longitudinal view. Figure from \cite{2e}.}
\label{strauss-fig8}
\end{myfigure}
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig09-eps-converted-to.pdf}}}
\caption{MC distribution of: a) the kink angle for $\tau$ decays, b) the path length of the $\tau$, c) the momentum of the decay daughter, d) the total transverse momentum $P_T$ of the detected daughter particles of $\tau$ decays with respect to the parent track. The red band shows the 68.3\,\% of values allowed for the candidate event, and the dark red line indicates the
most probable value. The dark shaded area represents the excluded region corresponding to the a priori $\tau$ selection cuts. Figure from \cite{2e}.}
\label{strauss-fig9}
\end{myfigure}
Fig. \ref{strauss-fig9} shows the cuts used for the selection criteria defined at the time of the proposal and the kinematic variables of the $\tau$ decay observed by the OPERA experiment. At the time of this conference, the number of expected $\tau$ candidates was 1.7, with 0.5 events expected in the single-prong channel. The expected background for the analyzed event sample corresponding to $4.9\times10^{19}$\,protons on target (pot) was $0.16\pm0.05$ events.
At the time of writing these proceedings, a second $\tau$ candidate appeared \cite{7a}.
\subsection{Physics run performance and data analysis status}
Since 2007, the OPERA experiment has collected a total of $18.5\times10^{19}$\,pot. This corresponds to about 15'000 interactions in the target areas of the experiment. Fig. \ref{strauss-fig6} shows from top to bottom, as a function of time, the integrated number of events occurring in the target (showing the CNGS shutdown periods), with their vertex reconstructed by the electronic detectors, for which at least one brick has been extracted, for which at least one CS has been analysed, for which this analysis has been positive (track stubs corresponding to the event have been found), for which the brick has been analysed, for which the vertex has been located, for which the decay search has been completed.
The efficiency of the analysis of the most probable CS is rather low, about 65\,\%, and is significantly lower for NC events, among which $\nu_{\tau}$ interactions are the most likely to be found. To recover this loss, multiple brick extraction is performed; this brings the final efficiency of observing tracks of the event in the CS to about 74\,\%. The efficiency for locating an event seen on the CS in the brick is about 70\,\%. At the time of this conference, a total of 4611 events has been localised in the bricks and the decay search has been completed for 4126 of them.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig07-eps-converted-to.pdf}}}
\caption{Events recorded and analyzed in the OPERA experiment from 2008 until 2012.}
\label{strauss-fig6}
\end{myfigure}
After an ECC has been identified to be the most likely interaction brick, the ECC is developed and sent to one of the scanning laboratories, where it is scanned within a short time. The efficiency for locating an event within the ECC is about 70\,\% with respect to the number of positive CS results. After the event has been located, a dedicated decay search is performed to obtain a data sample which can be compared to MC and which provides uniform data quality from all laboratories. This decay search includes the search for decay daughters and a reconstruction of the kinematics of the particles at the vertex. At the time of this conference, 4611 events have been localized in the ECC, with a total of 4126 CC and NC events having completed the decay search.
\subsection{Charm decay topologies in neutrino interactions}
In about 5\,\% of the $\nu_{\mu}$-CC interactions, the production of a charmed particle at the primary vertex takes place. Charmed particles have lifetimes similar to the $\tau$ and similar decay channels. Thus charm events provide a subsample of decay topologies similar to $\tau$ decay, for which the detection efficiency can be estimated based on MC simulation. A study of a high purity selection of charm events in 2008 and 2009 shows agreement with the data \cite{5}. One-prong charm decay candidates are retained if the charged daughter particle has a momentum larger than 1 GeV/c. This leads to an efficiency of $\epsilon_{\mbox{short}} = 0.31 \pm 0.02 (\mbox{stat.})\pm 0.03 (\mbox{syst.})$ for short and $\epsilon_{\mbox{long}} = 0.61 \pm 0.05 (\mbox{stat.}) \pm 0.06 (\mbox{syst.})$ for long charm decays, wherein `long' means the production and the decay vertices are not located in the same lead plate. The number of events for which the charm search is complete is 2167 CC interactions. In these we expect $51\pm7.5$ charm candidates, with a background of $5.3\pm2.3$ events. The number of observed candidates is 49, which is in agreement with expectations.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig10-eps-converted-to.pdf}}}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig11-eps-converted-to.pdf}}}
\caption{Display of one of the charm events. Top: View of the reconstructed event in the emulsion. Bottom: Zoom in the vertex region of the primary and secondary vertex. }
\label{strauss-fig10}
\end{myfigure}
Fig. \ref{strauss-fig10} shows a charm decay detected in the OPERA experiment. In the electronic detector reconstruction, two muons were observed, one charged positively, the other negatively. The $\mu^-$ is attached to the primary vertex, while the $\mu^+$ is connected to the decay vertex. This topology corresponds with a charged charm decaying into a muon and the measured kinematic parameters are a flight length of 1330\,$\mu$m and a kink angle of 209\,mrad. The impact parameter (IP) of the $\mu^+$ with respect to the primary vertex is 262\,$\mu$m, and its momentum is measured as 2.2\,GeV/c. This accounts for a transverse momentum ($P_T$) of 0.46\,GeV/c.
\section{$\nu$-velocity measurement}
Due to the time structure of the CERN SPS beam, the OPERA experiment is able to trigger on the proton spill hitting the CNGS target. As a result, the electronic detector provides a time signal of the recorded events, which can be used to measure the neutrino velocity in the CNGS beam. One needs to measure with precision of some ns the flight time (time of flight - TOF) between CERN and LNGS, and the distance between reference points in the detector and the CNGS. The concept of the neutrino time of flight measurement is illustrated in Fig. \ref{strauss-fig15}.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig15-eps-converted-to.pdf}}}
\caption{Scheme of the time of flight measurement. Figure from \cite{6}.}
\label{strauss-fig15}
\end{myfigure}
The procedures are explained in great detail in \cite{6}. Since the time of the conference, an instrumental mistake has been identified that makes the results presented at this conference obsolete. Updated results taken from \cite{7b} are presented below.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig16-eps-converted-to.pdf}}}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig17-eps-converted-to.pdf}}}
\caption{Top: Scheme of the timing sytem at CERN. Bottom: Scheme of the timing system at LNGS. Figures from \cite{6}.}
\label{strauss-fig16}
\end{myfigure}
Fig. \ref{strauss-fig16} shows the timing systems both at CERN and LNGS, which allowed time calibration between both sites within accuracy of $\pm4$\,ns.
The distance between CERN and the OPERA detector was measured via GPS geodesy and extrapolation down to the location of both the CNGS target and the OPERA detector with terrestrial traverse methods. The effective baseline is measured as $731278.0\pm0.2$\,m.
The proton wave form for each SPS extraction was measured with a beam current transformer (BCT). The sum of the wave forms restricted to those associated to a neutrino interaction in OPERA was used as PDF for the time distribution of the events within the extraction. The maximum likelihood method was used to extract the time shift between the two distributions, i.e. the neutrino time of flight. Internal NC and CC interactions in the OPERA target and external CC interactions occurring in the upstream rock from the 2009, 2010 and 2011 CNG runs were used for this analysis. As shown in Fig. \ref{strauss-fig20}, it is measured to be: $$\delta t = \mbox{TOF}_c - \mbox{TOF}_\nu = (6.5\pm7.4(\mbox{stat.})\pm ^{+8.3}_{-8.0}(\mbox{syst.}))\,\mbox{ns.}$$ Modifying the analysis by using each neutrino interaction waveform as PDF instead of their sum gives a comparable result of
$$\delta t = (3.5\pm5.6(\mbox{stat.})\pm ^{+9.4}_{-9.1}(\mbox{syst.}))\,\mbox{ns.}$$ No energy dependence was observed.
\begin{myfigure}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig19-eps-converted-to.pdf}}}
\centerline{\resizebox{70mm}{!}{\includegraphics{thomasstrauss_2012_0n_fig20-eps-converted-to.pdf}}}
\caption{Top: Comparison of the measured neutrino interaction time distributions (data points) and the proton PDF (red and blue line) for the two SPS extractions resulting from the maximum likelihood analysis. Bottom: Blow-up of the leading edge (left plot) and the trailing edge (right plot) of the measured neutrino interaction time distributions (data points) and the proton PDF (red line) for the first SPS extraction after correcting for $\delta t=6.5$\,ns. Within errors, this second extraction is equal to the first one. Figures from \cite{6}.}
\label{strauss-fig20}
\end{myfigure}
To cross-check for systematic effects, a dedicated bunched beam run was performed, where the SPS proton delivery is split into 3\,ns long spills, separated by 524\,ns in time during autumn 2011, and a similar mode of 3\,ns with 100\,ns separation in spring 2012.
The value of $\delta t$ obtained in 2011 by using timing information provided by the target tracker is $1.9\pm3.7$\,ns; it is $0.8\pm3.5$\,ns when based on the spectrometer data \cite{6}. For the 2012 run, the corresponding values of $\delta t$ are $\delta t =(-1.6\pm1.1(stat.)^{+6.1}_{-3.7})$\,ns \cite{7b}. All results are in agreement with the measurement from standard CNGS beam operation.
\section{Conclusions}
The OPERA experiment detected two $\tau$ neutrino events appearing in the CNGS beam. Further, we measured the neutrino velocity to be in agreement with the speed of light in a vacuum to the O($10^{-6}$). Other short decay topologies like $\nu_e$ or charm decays can also be detected and are in agreement with MC expectations, thus providing a benchmark for validating the $\tau$ efficiency expectations.
\thanks
Firstly, I thank the organizers of the workshop and the OPERA PTB for the possibility to join this workshop. The OPERA collaboration thanks CERN, INFN and LNGS for their support and work. In addition OPERA is grateful for funding from the following national agencies:
Fonds de la Recherche Scientique - FNRS and Institut Interuniversitaire des Sciences Nucl\'eaires for Belgium; MoSES for Croatia; CNRS and IN2P3 for France; BMBF for Germany; INFN for Italy; JSPS (Japan Society for the Promotion of Science), MEXT (Ministry of Education, Culture, Sports, Science and Technology), QFPU (Global COE program of Nagoya University, ``Quest for Fundamental Principles in the Universe", supported by JSPS and MEXT) and Promotion and Mutual Aid Corporation for Private Schools of Japan for Japan; The Swiss National Science Foundation (SNF), the University of Bern and ETH Zurich for Switzerland; the Russian Foundation for Basic Research (grant 09-02-00300 a), the Programs of the Presidium of the Russian Academy of Sciences ``Neutrino Physics" and ``Experimental and theoretical researches of fundamental interactions connected with work on the accelerator of CERN", the support programs of leading schools (grant 3517.2010.2), and the Ministry of Education and Science of the Russian Federation for Russia; the Korea Research Foundation Grant (KRF-2008-313-C00201) for Korea; and TUBITAK ``Scientific and Technological Research Council of Turkey", for Turkey. In addition the OPERA collaboration thanks the technical collaborators and the IN2P3 Computing Centre (CC-IN2P3).
|
\section{Introduction} \label{Introduction}
Since 2009, \textit{Twitter} has become a popular online social network (OSN) among millions of users. As of 2013, Twitter has 200 million users creating more than 400 million tweets per day, making it the most populous micro-blogging service.~\footnote{https://blog.twitter.com/2013/celebrating-twitter7} Users join Twitter via an easy sign-up process. During the sign up process, users set some profile attributes as part of their account, where some are mandatory (`username') and some optional (e.g. name, location). \textit{Username} attribute is an important attribute, with which a user can be referred, searched and tagged \emph{uniquely} in a tweet by any other user on Twitter. After the sign up process is complete, Twitter assigns a \emph{unique}, numeric, and constant \emph{ID} to the user, which can only be known to other users via Twitter's API request, and not via Twitter web / mobile interface. A user is therefore associated with two unique attributes on Twitter -- username and user ID. \\
\indent User ID is not changeable however, changes to username are allowed according to user convenience and requirements.~\footnote{https://support.twitter.com/articles/14609-changing-your-username} Figure~\ref{fig:change} shows a Twitter user who changes her username from `alone\_trix!' to `JOX\_4!'. Few OSNs such as Facebook~\footnote{https://www.facebook.com/help/105399436216001\#What-are-the-guidelines-around-creating-a-custom-username?} put a sealing on the number of times a user can change her username, while Twitter does not. Allowing changes to a username is beneficial because it may help a user to accommodate her changing attributes, likes, dislikes, and her changing interests over time. However, such username changes may lead to unwanted consequences -- a user search with her past username and with no information of her numeric user ID, may lead to non-searchability (no results) or unreachability (broken link~\footnote{Sorry, the page doesn't exist page on Twitter}) to the user's profile. Further, Figure~\ref{fig:change} shows a scenario where search for a Twitter user, who had username `alone\_trix!' at an earlier timestamp, redirects to a different user who picked the same username at a later timestamp. If changing usernames may lead to loss of connectivity to a user profile, broken links, or redirection to a different user profile, no user may intend to do it. However, we observe that around 10\% of observed 8.7 million Twitter users change their usernames over time (as discussed in Section~\ref{Methodology}). \\
\begin{figure*}[t]
\centering
\subfigure[User-ID `12x6917x09' recorded on November 8, 2013, holding the username `alone\_trix!' ]{
\includegraphics[scale=0.27]{snapshot1_an.png}\label{fig:change1}
}
\quad
\subfigure[User-ID `12x6917x09' recorded on January 15, 2014, changed her username to `JOX\_4!' ]{
\includegraphics[scale=0.27]{snapshot2_an.png}\label{fig:change2}
}
\quad
\subfigure[User-ID `55x814x82' recorded on January 15, 2014, holding the username `alone\_trix!']{
\includegraphics[scale=0.27]{snapshot3_an.png}\label{fig:change3}
}
\caption{ shows a Twitter user who changes her username over time (a, b). If the user ID is not known, her new username cannot be found. Two different users with two different unique IDs pick the same username `alone\_trix!' at two different timestamps (a, c). A Twitter search for user ID `12x6917x09' with her old username `alone\_trix!' redirects to a different user with user ID `55x814x82'. }
\label{fig:change}
\end{figure*}
\indent To the best of our knowledge, there is very little work to understand why users temporally change profile attributes on OSNs, specifically username. We make the first attempt to explore this behavior of changing usernames temporally and term it as \emph{username changing behavior}, in this paper. We try to understand properties of users who temporally change usernames and explore a set of reasons which explain why users change usernames over time and what usernames they switch to. Our contributions are:
\begin{itemize}
\item
\vspace{-3mm}
We present the first longitudinal study to understand username changing behavior on Twitter. We find that around 10\% of 8.7 million Twitter users change their usernames temporally.
\item
We analyze a focused set of users exhibiting username changing behavior and observe that few users change their usernames frequently, and within 24 hours of the earlier username change. Users choose less similar new usernames while a few favor one of their past usernames as their new username.
\item
We show that frequency of changing username is weakly correlated with users' in-degree, activity and year of account creation. Users who change usernames are more popular, and active on Twitter than users who do not change.
\item
We learn that users change their usernames for reasons such as to gain space in a tweet, to gain more followers, to promote usernames, to suit a trending event, to gain / lose anonymity or to avoid boredom.
\end{itemize}
\vspace{-3mm}
\indent We believe that a comprehension of username changing behavior can help Twitter to develop tailored username suggestion algorithms for its users. Currently, Twitter offers username suggestions only during sign up process and not later. We think that Twitter can suggest suitable usernames to a user, even after her account creation, based on her activity with her past usernames and associated profits / risks. Further, we think that a record of past usernames can be helpful for planning promotional content for users on Twitter, for correlating multiple Twitter profiles to a single real-world user, and for judging usernames of users on other OSNs (explored in Section~\ref{Applications}). \\
\indent The paper is organized as follows. We first discuss the detailed methodology in Section~\ref{Methodology}. We then discuss username changing behavior characteristics in Section~\ref{In-Depth}, present properties of users who change usernames in Section~\ref{user} and then explore some plausible reasons for such behavior in Section~\ref{reasons}. We present some applications of our work in Section~\ref{Applications}. We then discuss the applicability and generalizability of our observations in Section~\ref{Discussion}, present related work and conclude the paper with some future directions.
\section{Methodology} \label{Methodology}
To qualitatively and quantitatively study behavior of changing username on Twitter, we collect 8.7 million users on Twitter and monitor them at regular intervals from October 1, 2013 - November 26, 2013. We also create a smaller set of 10,000 users who are selected to be monitored every 15 minutes for any username change. We describe our data collection framework now.
\subsection{Data collection} \label{DataCollection}
Our data collection methodology is divided into two stages -- Seed collection, and Seed monitoring. Seed collection stage collects a set of users who are monitored in Seed monitoring stage. Details of each stage is as follows:
\subsubsection{Seed collection}
We collect a seed set of 8,767,576 users recorded by an event monitoring tool, MultiOSN~\cite{Dewan:2013:MRM:2528228.2528235}. Researchers shared the profiles of users who tweeted at least once about any of the 17 events (see Table~\ref{list_events}) monitored by MultiOSN, during April 1, 2013 - September 3, 2013. We refer to the seed set of 8,767,576 users as 8.7M users in the rest of the paper.
\begin{table}[ht]
\begin{center}
\small
\begin{tabular}{|p{4.5cm}|p{3.1cm}|}
\hline
{\bf{Global Events}} & {\bf \raggedright{Local Events} } \\ \hline
Indian Premier League (IPL) & Bangalore Blasts \\ \hline
Boston Blasts & Uttrakhand Floods \\ \hline
Texas Fertilizer Plant Blast & Bodhgaya Blasts \\ \hline
Oklahoma Tornado & Telangana State \\ \hline
Champions Trophy Cricket & FoodBill \\ \hline
Nelson Mandela in ICU & Onion Crisis \\ \hline
Royal Baby & AsaramBapu Conviction \\ \hline
\raggedright{Earthquake\_Pakistan ($ \rm{16^{th}}$ April)} & \\ \hline
\raggedright{Earthquake\_MiddleEast ($ \rm{1^{st}}$ May)} & \\ \hline
Mother's Day Parade shooting & \\ \hline
\end{tabular}
\caption{Events monitored to collect seed users.}
\label{list_events}
\end{center}
\vspace{-8mm}
\end{table}
\subsubsection{Seed monitoring}
In this stage, we monitor temporal changes in profile attributes of 8.7M user profiles, collected in the earlier stage. We query 8.7M users four times after definite intervals during October 1, 2013 - November 26, 2013, via Twitter Search API~\footnote{https://dev.twitter.com} and record their observed profile attributes in a database along with a timestamp. The process of querying 8.7M users is termed as \textit{scan}, in this paper. Table~\ref{collection} describes the time duration of each scan. Due to varying Internet speeds at servers and use of less number of authentication tokens during initial querying, each scan took different number of days to record 8.7M user profiles. Further, not all user accounts are activated during each scan, some deactivate / delete their accounts, while some are suspended by Twitter. We, therefore, record different number of user profiles in each scan. \\
\indent We analyze four scans of 8.7M users collected over a period of 2 months (Oct 1 - Nov 26). A user is marked to have changed her username, if her username values in two consecutively timed scans are different. By comparing two consecutive scans, old and new usernames of a user are recorded. Note that, Twitter usernames are case-insensitive, therefore any case changes are not counted as username changes. We find that 853,827 users of 8.7M users (10\%) change their usernames at least once during our observation period (Oct 1 - Nov 26). Further, 853,827 users constitute 904,518 username change instances implying that a few users change their usernames multiple times. Therefore, we believe that it is significant to explore the characteristics of username changing behavior on Twitter. However, four scans of 8.7M users lack in necessary data. \\ \indent Due to huge number of users to query and limited Twitter API calls, each scan took long time to query each user in 8.7M dataset. Due to long scan stretches and intermediate intervals, we could neither record the exact date and time when users actually changed usernames nor all username changes a user went through. To capture the actual time and date when users changed their usernames as well as capture most username change instances triggered by users, we needed to scan 8.7M users at short intervals. Given Twitter allows 60 API calls per 15 minutes for users/lookup,~\footnote{https://dev.twitter.com/docs/rate-limiting/1.1/limits} scanning 8.7M users at short intervals (every 15 minutes) would require 1,462 application authentication tokens. We, therefore, use limited authentication tokens to respect the Twitter API resources utilization and initiate a fifteen-minute scan.
\subsection{Fifteen-minute scan}
Our focus in this research is on the users who change their usernames over time. We, therefore, first select 711,609 users who change their usernames at least once, during October 1, 2013 - November 15, 2013, as recorded during our Scan-I, Scan-II and Scan-III. Out of these 711,609 users, we \emph{randomly} sample 10,000 users~\footnote{We are continuously collecting data for the 711,609 users and hope to have a larger dataset soon.} and start to monitor them at short intervals. We query 10K users via Twitter API every 15 minutes. We term the faster scan of 10K users as \emph{Fifteen-minute scan}. Fifteen-minute scan starts on November 22, 2013; we bookmark the scan till March 19, 2014 and use this 117 days scan dataset for our analysis.~\footnote{We continue to scan 10K users after Mar 19, 2014 and record any username change.} If an observed username of a user profile exhibits a change in comparison to the most recent username recorded either in Scan-III or in fifteen-minute dataset which recorded her past username changes, fifteen-minute scan records the user profile in the fifteen-minute dataset and timestamp it. With fifteen-minute scan, we could record the exact timestamp when user changed her username, with an error limit of 15 minutes. Further, we observe that our regular Scan-IV took only one snapshot while fifteen-minute scan took 794 snapshots of 10K users, during Nov 22 - Nov 26. Scan-IV misses 712 username change instances triggered by 607 users, well captured by fifteen-minute scan. Therefore, fifteen-minute scan is successful in capturing most username changes made by the monitored users.
\begin{table}[h]
\begin{center}
\small
\begin{tabular}{|p{1.5cm}|p{2.3cm}|p{1.5cm}|p{1.5cm}|}
\hline
{\bf{Name of scan}} & {\bf \raggedright{Period of scan} } & \raggedright{\textbf{\# users queried}} & {\textbf{\# users recorded} } \\ \hline
Seed set & Apr 1 - Sep 3 & 8,767,576 & 8,767,576 \\ \hline
Scan-I & Oct 1 - Oct 16 & 8,767,576 & 8,380,827 \\ \hline
Scan-II & Oct 25 - Oct 30 & 8,767,576 & 8,396,594 \\ \hline
Scan-III & Nov 8 - Nov 15 & 8,767,576 & 7,271,129 \\ \hline
Scan-IV & Nov 22 - Nov 26 & 8,767,576 & 8,388,010 \\ \hline
Fifteen-minute Scan & Nov 22 - Mar 19 & 10,000 & 2,698 \\ \hline
\end{tabular}
\caption{describes our four scans of 8.7M users and fifteen minute scan of 10K users. }
\label{collection}
\end{center}
\end{table}
\indent In the rest of the paper, we use the following definitions and notations. Each time a user changes her username, an event is counted and is termed as \emph{username change event / instance}. Old (dropped) username of a user $i$ at time $t_{1}$ is denoted by $u_{io}$ and new username at time $t_{2}$ is denoted by $u_{in}$, where $t_{1} < t_{2}$. A user $i$ may use different usernames at different times, thereby creating multiple username change instances and constituting a \emph{username sequence}, denoted by $u_{i1}, u_{i2}, u_{i3}, \dots, u_{in}$. A user $i$ may choose her old usernames again, if still available. Such usernames are termed as \emph{revisited-usernames} and are denoted by $u_{ir}$.
\subsection{Representativeness of the dataset} \label{Representativeness}
We examine geographical locations of 10K users to understand if they span across diverse locations.
Geographical location of a user can be estimated by user's two profile attributes -- `location' and `timezone'. Location attribute of a user is an unformatted text, where user can provide any set of characters describing her location, such as `Moon'. Timezone attribute is selected from a drop down list of timezones and user can decide to use any timezone. Both, location and timezone attribute may hold incorrect values, which may bias our understanding on representativeness of the dataset. We therefore use geo-tagged tweets by the users, to record their location. We map 1,849 unique latitude, longitude pairs from where 926 users (9\% of 10K) have posted their tweets (see Figure~\ref{fig:map}). We observe that users in our dataset, tweet from different locations around the world and not biased to only a few locations. Therefore, our analysis and results can be generalized to Twitter population from various global locations, who opt to change their usernames over time.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.47]{Geo_10K_FT.png}
\caption{shows geographical distribution of 926 users who geo-tagged their tweets. We see that users in our dataset span diverse geographical locations. }
\label{fig:map}
\end{figure}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\includegraphics[scale=0.15]{how_many_times_f_d1.eps}\label{fig:frequency1}
}
\quad
\subfigure[]{
\includegraphics[scale=0.23]{cumulative_dist_general.eps}\label{fig:freq}
}
\quad
\subfigure[]{
\includegraphics[scale=0.23]{cumulative_dist2.eps}\label{fig:distance1}
}
\caption{(a) Username change frequency distribution (b) Cumulative distribution of username change instances vs \# of days after which users change usernames (c) Cumulative distribution of username change instances vs the longest common subsequence length between old and new usernames. We observe that few users change usernames frequently and most username changes (61\%) happen within 24 hours of earlier change (0 day). For 75\% of username change instances, new usernames are less similar to old ones.}
\label{fig:how}
\end{figure*}
\section{Username changing behavior} \label{In-Depth}
We now explore the characteristics of username changing behavior on Twitter. We study frequency and patterns of username change and consequences to users' old usernames.
\subsection{Frequency of username change}
We intend to understand if users change their usernames once or repeatedly. Out of 10K users, fifteen-minute scan records 2,698 users who change their usernames at least once after Nov 22, 2014. We plot distribution of users v/s number of times they change their usernames (see Figure~\ref{fig:frequency1}). Most users (1,474) change their usernames once during our observation period (Nov 22 - Mar 19), however few users change about 30 times. One user changes her username 96 times in 117 days. On manual inspection, the user~\footnote{https://twitter.com/intent/user?user\_id=796101102} seems to have malicious intentions with half completed tweets, tweets with same text, and frequent posts in short duration. We further investigate the time after which users change their usernames. We plot cumulative distribution of username change events with the number of days after which users trigger a username change event (see Figure~\ref{fig:freq}). We observe that around 61\% of username change events are triggered within 24 hours (zero day) of the previous username change. We conclude that a few users frequently change usernames within short period of time. \\
\indent We now inquire what usernames users switch to. Do they change to random usernames or to usernames similar to their dropped username? We calculate longest common subsequence length between consecutive usernames picked by the user in a username sequence. For instance, if a user $i$ chooses username $u_{i1}$ at time $t_{1}$, username $u_{i2}$ at time $t_{2}$ and username $u_{i3}$ at time $t_{3}$ (where $t_{1}$ $\leq$ $t_{2}$ $\leq$ $t_{3}$), then longest common subsequence lengths are calculated for \{$u_{i1}$, $u_{i2}$\} and \{$u_{i2}$, $u_{i3}$\}. We then normalize the length scores for each username pair with the length of the earlier username used by the user. We choose longest common subsequence matching in comparison to edit distance because the prior metric captures maximal alignments of characters between two usernames while later counts misalignments. We intend to penalize less for addition / re-arrangement of characters during a new username selection. Figure~\ref{fig:distance1} shows the cumulative distribution of username change events versus the longest subsequence matching length between the usernames (old and new) associated with the event. Around 75\% of username change instances demonstrate that the new username is less similar to the old username (length $\leq$ 0.5) and around 10\% instances have new usernames highly similar and derived from the old usernames (length $\geq$ 0.8). We therefore conclude that most users tend to pick a username which is less similar to the earlier username.
\vspace{-3mm}
\subsection{Patterns of username selection} \label{patterns}
Researchers have observed that most users choose same or similar usernames across multiple OSNs, owing to the human memory limitation to remember username for each OSN, they register to~\cite{ZafaraniL13,perito}. However, on Twitter, they need not remember their old usernames and therefore, username selection might be completely random and un-related to their earlier username. We observe the same in earlier section where 75\% of username change instances have less similar usernames. With this observation, we speculate that when a user chooses a username, she favors a username different from the earlier one. \\
\indent We analyze 2,698 users of fifteen-minute dataset and found that out of 1,224 users who change username twice or more, 779 users (64\%) choose new usernames whenever they change. However, 445 users (36\%) choose to pick at least one of their dropped usernames again. For instance, a user $i$ picks a username $u_{i1}$ at time $t_{1}$, change to username $u_{i2}$ at $t_{2}$ and then switches back to username $u_{i1}$ at time $t_{3}$. We term such instances as \emph{username switching-back instances / events}. For such instances, $u_{i1}$, the username to which user $i$ switches back to, is termed as \emph{revisited-username} and is denoted by $u_{ir}$. Username(s) picked by the user in the meantime, collectively are termed as \emph{transit usernames}. Username $u_{i2}$ is an instance of transit username. We now extract patterns in which users choose to pick a revisited-username via the following methodology. \\
\indent We analyze a user's username sequence, for instance, \{$u_{i1}$, $u_{i2}$, $u_{i3}$, $u_{i2}$, $u_{i4}$, $u_{i3}$\}. Such a username sequence can be re-written as \{$u_{i1}$, $u_{ir1}$, $u_{ir2}$, $u_{ir1}$, $u_{i4}$, $u_{ir2}$\}. The user chooses to revisit two of her past usernames ($u_{i2}$, $u_{i3}$) denoted by ($u_{ir1}$, $u_{ir2}$). We extract three patterns from the given sequence -- \{$u_{ir1}$, $u_{ir2}$, $u_{ir1}$\}, \{$u_{ir2}$, $u_{ir1}$, $u_{i4}$, $u_{ir2}$\}, and \{$u_{i1}$, $u_{ir1}$, $u_{ir2}$, $u_{ir1}$\}. Pattern \{$u_{ir1}$, $u_{ir2}$, $u_{ir1}$\} represents a pattern of length 3 where user chooses an immediate dropped username. Pattern \{$u_{i1}$, $u_{ir1}$, $u_{ir2}$, $u_{ir1}$\} represents a pattern of length 4 where user chooses an immediate dropped username and contains a sub-pattern \{$u_{ir1}$, $u_{ir2}$, $u_{ir1}$\}. Pattern \{$u_{ir2}$, $u_{ir1}$, $u_{i4}$, $u_{ir2}$\} of length 4 which represents that user chooses to revisit a username used by her, two username change events earlier. We extract all patterns and sub-patterns of different lengths where user chooses to reuse her past username, from 445 user sequences (see Table~\ref{fig:pattern}). \\
\indent We observe that (sub)pattern \{$u_{irj}$, $u_{in}$, $u_{irj}$\} is most frequent among username switching-back instances than others. Users favor most-immediate dropped username when they wish to choose from their past usernames. We think that users hop to most recently dropped username again, either because their other past usernames are taken, they may not remember their rest past usernames or they remember the benefits of their most recently dropped username. We believe that such a characteristic can help Twitter to design customized username recommendation feature. Twitter can recommend its users the most beneficial username among their past usernames based on followers count gain (Section~\ref{gainloss}), activity and other metrics, rather than they choosing recently dropped usernames again, owing to lack of information about suitability of their other past usernames.
\begin{table}[htbp]
\vspace{-3mm}
\begin{center}
\small
\begin{tabular}{|p{4cm}|p{2.5cm}|}
\hline
\raggedright{{\bf Pattern / Sub-pattern}} & {\bf{\# of instances }} \\ \hline
$u_{ir1}$ - $u_{i2}$ - $u_{ir1}$ & 228 \\ \hline
$u_{i1}$ - $u_{ir1}$ - $u_{i2}$ - $u_{ir1}$ & 104 \\ \hline
$u_{ir1}$ - $u_{i2}$ - $u_{i3}$ - $u_{ir1}$ & 54 \\ \hline
$u_{ir1}$ - $u_{ir2}$ - $u_{ir1}$ - $u_{ir2}$ & 18 \\ \hline
$u_{ir1}$ - $u_{i2}$ - $u_{i3}$ - $u_{i4}$ - $u_{ir1}$ & 12\\ \hline
$u_{i1}$ - $u_{ir1}$ - $u_{i2}$ - $u_{i3}$ - $u_{ir1}$ & 8 \\ \hline
$u_{ir1}$ - $u_{i2}$ - $u_{ir1}$ - $u_{i3}$ - $u_{ir1}$ & 9 \\ \hline
$u_{ir1}$ - $u_{ir2}$ - $u_{i3}$ - $u_{ir2}$ - $u_{ir1}$ & 3\\ \hline
$u_{i1}$ - $u_{ir1}$ - $u_{ir2}$ - $u_{ir1}$ - $u_{ir2}$ & 9 \\ \hline
$u_{ir1}$ - $u_{ir2}$ - $u_{i3}$ - $u_{ir1}$ - $u_{ir2}$ & 2 \\ \hline
\end{tabular}
\caption{Popular patterns of revisited-username selection. Most users switch back to recently dropped username.}
\label{fig:pattern}
\end{center}
\end{table}
\vspace{-6mm}
\subsection{Squatted usernames}
Till now, we explore the properties of new usernames picked by users. We are also interested in examining the properties of old dropped usernames i.e. the usernames which users vacate to pick new usernames. On Twitter, there are four pools to which an old username can belong to -- \emph{free-username pool}, \emph{taken-username pool},~\footnote{https://support.twitter.com/groups/51-me/topics/205-account-settings/articles/14609-changing-your-username} \emph{suspended / deactivated-username pool}~\footnote{https://support.twitter.com/articles/15348-my-account-information-is-already-taken\#deactivatedaccount} and \emph{squatted-username pool.}~\footnote{https://support.twitter.com/articles/18370-username-squatting-policy} A username $u_{io}$ belongs to a free-username pool if no one else uses it on Twitter. If another user selects username $u_{io}$ as her own, the username moves to taken-username pool. If the user with username $u_{io}$ deactivates her account or is suspended by Twitter, the username is blocked forever (for now) and is not available to anyone, which thereby moves the username in suspended / deactivated-username pool. If an inactive user profile registers her account using $u_{io}$, in order to block or preserve that username, and not to allow others to use it, the username is said to belong to squatted-username pool. \\
\indent Squatted usernames on OSNs have been investigated as a challenge in literature by law researchers~\cite{ramsey2010brandjacking,curtin2010name}. Researchers analyzed trademark infringement cases on social media like Twitter. We were curious to know if cybersquatting exists on Twitter in the form of username squatting. We could not consider cases of trademark infringements, because of the lack of ground truth, where companies filed a trademark infringement report to Twitter. Therefore, we check for generic users, if users' past / dropped usernames are reserved by inactive Twitter user profiles. \\
\indent For our fifteen-minute dataset, we observe that for around 7\% of 2,698 users, at least one of their vacated usernames are blocked by inactive Twitter profiles (either created by themselves or others) who either show no activity (i.e. no tweets) or have zero followers. We think that inactive profiles may have been created to avoid slip of past usernames in Twitter's free-username pool. Twitter considers such blocked usernames as `squatted' usernames. However, according to Twitter username squatting policy, squatted or inactive usernames are not released, unless the username causes trademark infringement. In such scenarios, common Twitter users are suggested to choose a modified version of their wished username. We suggest that Twitter should implement a username request feature, where a user can put her request for a username, if the username is not available and as soon as the username is vacated, the user can be notified. This way, Twitter can avoid squatted usernames and users can use their requested username. Further, investigating if a user is blocking her own past usernames via inactive profiles, can help Twitter understand her malicious intentions.
\section{User characteristics} \label{user}
We now attempt to understand the characteristics of users who opt to change their usernames. Do users with high popularity or activity change their usernames frequently? Do new users change their usernames more frequently than old users? Do they differ from users who never change their usernames? We answer each of these questions now.
\subsection{In-Degree centrality}
On Twitter, users tweet, reply or indulge in conversations with their username. Changing usernames by a popular user may lead to confusion among her followers or may lead to loss of tweets in case someone else picks the username. In such a scenario, we speculate that users with high centrality / in-degree would like to avoid any username change. We use fifteen-minute dataset of 2,698 users and measure the in-degree of the users (see Figure~\ref{centrality}). We correlate in-degree of a user with the frequency of username changes. We remove an outlier user who changes her username 96 times (in-degree - 14.3K). We observe that number of times a user changes her username is weakly and positively correlated to the in-degree of the user (Pearson correlation: 0.0083). We therefore conclude that irrespective of the popularity, a user may still wish to change the username multiple number of times. We suspect that popular users change interests, and put opinions about recent events frequently and therefore, to reflect the same, change their username multiple times.
\subsection{Activity}
An active user on Twitter, who engages herself in conversations and group chats, may change her username less frequently to avoid confusion during tagging / replying in a tweet. We therefore speculate that active users change their usernames less frequently. We analyze 2,698 users and measure their activity with the number of tweets they create. We remove an outlier user who changes her username 96 times (tweet count - 8,741). Figure~\ref{statuses} shows the tweet distribution of users with the number of times they change their usernames. We observe a weak and positive correlation between the two (Pearson correlation: 0.0322). We infer that users who actively post on Twitter change their usernames frequently for possible reasons such as to gain traction or change behavior during trending topics (explored in Section~\ref{reasons}).
\begin{figure}[htbp]
\centering
\subfigure[]{
\includegraphics[width=0.5\linewidth]{In-degree_dist2.eps} \label{centrality}
}
\subfigure[]{
\includegraphics[width=0.5\linewidth]{activity2.eps} \label{statuses}}
\caption{ (a) user's in-degree distribution, (b) user activity distribution versus frequency of username changes. We find that in-degree centrality and activity of a user are weakly correlated with the number of times she changes her username.}
\label{fig:multiple}
\end{figure}
\subsection{Age of the account}
Users who registered themselves on Twitter, long time ago might have chosen most stable and beneficial username for themselves than users who have registered recently and are still in exploratory stage. We examine if old user accounts engage themselves in username changing behavior or only new users change their usernames multiple times. Figure~\ref{age} shows the age of the account distribution for 2,698 users, with the number of times users change their usernames. Inset graph shows distribution of monitored 10K users across year of their account creation. We observe negative and weak correlation between the age of the Twitter account and the frequency with which the account changes username (Pearson correlation: -0.0798). Negative correlation implies that older accounts change their usernames less frequently, however, is not necessarily true for all old accounts. A 2009 account changes her username 19 times, while a 2010 account changes her username 31 times. We therefore infer that irrespective of the age of their account on Twitter, users change their usernames.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\linewidth]{Age_dist_f.eps}
\caption{shows age of the user profile v/s number of username changes and distribution of 10K users monitored. We observe a weak correlation between the account age with the frequency of username change. }
\label{age}
\end{figure}
\subsection{Normal v/s Changing users}
As we observe for 8.7M users, 10\% of users change their usernames over time. A large proportion (90\%) still does not engage in this behavior. We now explore the similarities and differences between the properties of users who do not change their usernames with users who do. We compare in-degree, out-degree and activity of 2,698 users with characteristics of two random samples of size 2,698 users,~\footnote{We take two random samples to justify the generalizability of observations. } extracted from 8.7M users who have not changed their usernames during Oct 1 - Nov 26 (see Figure~\ref{comparison}). We observe that users who change their usernames demonstrate higher popularity and activity and actively follow other Twitter users as compared to users who do not bother to change their username. Note that, both random samples do not differ in their properties but differ from the users who change their usernames over time.
\begin{figure*}[htbp]
\centering
\subfigure[In-degree distribution of users]{
\includegraphics[scale=0.24]{Followers_comp.eps} \label{in-deg}
}
\subfigure[Out-degree distribution of users]{
\includegraphics[scale=0.24]{Friends_comp.eps} \label{out-deg}}
\subfigure[Activity distribution of users]{
\includegraphics[scale=0.24]{statuses_comp.eps} \label{status}}
\caption{shows a comparison between 2,698 users who changed their usernames with two random samples of 2,698 users who never changed their usernames. Users who change usernames demonstrate superiority in terms of popularity, and activity and users being followed as compared to users who don't. }
\label{comparison}
\end{figure*}
\section{Plausible reasons} \label{reasons}
We now examine the reasons why users opt for a username change by analyzing fifteen-minute dataset. We validate the reasons with the interactions we have with the users we monitor, as discussed in this section.
\subsection{Space gain}
Twitter allows its users to post 140 characters in a tweet. Long usernames might allow less space to convey content, while short usernames might give the liberty to add more content, hashtags or URLs~\cite{chhabra:phi.sh/ocial:-the-phishin:2011:yuqfj}. Therefore, users with long old usernames may change to short new usernames to benefit from space gain. We calculate the length difference between consecutive usernames in a username sequence i.e. between old and new username of a user and plot the distribution (see Figure~\ref{fig:length}). We observe that out of 6,132 instances of username change by 2,698 users, 2,687 instances exist where the new username is shorter than old username (44\%), 2,378 instances exist where new username is longer than the old username (39\%), 1,067 instances witness no length change between old and new username (17\%). Therefore, around half the population changes usernames to benefit from space gain, however the other half choose same or longer length username. A female responder from the user survey (discussed in Section~\ref{Discussion}) mentions that she changes her username to gain space. We conclude that space gain is a possible (but not only) reason for users to change their usernames.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.23]{Space_gain_2.eps}
\caption{Length difference between new and old username. We observe 44\% username change instances where new username is shorter than the old one. Space gain can be one of the reasons (but not only) for changing username. }
\label{fig:length}
\end{figure}
\subsection{Gain followers or avoid loss of followers}\label{gainloss}
As observed in Section~\ref{patterns}, 445 users choose to switch back to any of their earlier usernames. We examine the possible reasons for users to change and switch back to any of their past usernames (revisited-username). In real-world, researchers found that users keep the identity which causes maximum benefits, in terms of friends and reputation~\cite{mcfarland2005motives}. We validate the finding in our online scenario. We examine fifteen-minute dataset and observe three plausible reasons of why users switch to an earlier used username -- loss due to transit usernames, gain due to revisited-username, or both. We discuss each of the reasons now.
\vspace{-2mm}
\subsubsection{Loss or no gain via transit-usernames}
We quantify loss in terms of loss of followers for the user. We explore if loss or no gain of followers due to the transit usernames, collectively, prompts users to switch back to the revisited-username. We calculate the follower count difference between the time user first uses the revisited username to the next time the user uses the revisited username. We observe that for 158 users (36\%), transit-usernames collectively cause a loss or no gain of followers (follower count change $\leq$0). For instance, a user ID 76xx33242 with 3,288 followers, might have experienced loss of 674 followers due to her transit-username, and therefore might have switched back to her earlier username after 11 days. Maximum observed loss of followers due to transit usernames is -6,118 for a user with 215,084 followers earlier. Further, 26\% of 158 users switch back to their earlier username within 24 hours. We also talk to some of the users by tweeting them to know why they switch back to earlier username. A user responds that he gains no followers and receives a lot of irrelevant tweets with the transit-username and therefore he changes back to the earlier username (see Figure~\ref{fig:spam}).
\vspace{-2mm}
\subsubsection{Gain via revisited-username}
We quantify gain in terms of the gain of followers for the user. We explore whether the user wishes for revisited-username again because it has caused her a gain in follower count in the past, and therefore has helped her to increase her popularity. We calculate the follower count difference between the time user picks the revisited username to the time when she drops it and initiate another username change event. We observe that for 211 users (47\%), revisited-username has gained followers. For instance, user ID 56xx15159 gains 51 more followers by choosing a revisited-username for the first time, gains 1 follower due to transit username and gains 63 followers when she switches back to the revisited-username. We, therefore, infer that the experience of the user with her past usernames governs her decision on selection of revisited-username. \vspace{-2mm}
\subsubsection{Loss via transit \& gain via revisited username}
A mix of both reasons explained earlier is also observed in our dataset. 41 users (9\%) experience both loss of followers due to transit-usernames and gain of followers due to the revisited-username. A male responder from the user survey (discussed in Section~\ref{Discussion}) mentions that he gains followers due to revisited-username and lose followers due to transit usernames, hence switches back.
\indent We therefore infer that, most users do not stake their popularity for a new username and will change immediately if they observe a fall in their followers. To choose a revisited-username among the past usernames, users prefer most recently dropped username as we observed earlier (see Table~\ref{fig:pattern}). We reason this behavior with the limited memory and gauging capability of users to understand which usernames in their past caused them maximum benefit, further with no such help from Twitter. Twitter does not provide any support to help users to measure gain or loss of followers as they change usernames over time. We suggest that Twitter can develop a better understanding on which usernames a user should choose, based on the associated benefits and losses observed in the past, thereby can suggest better and tailored usernames to its users.\\
\indent On the other side, there are instances where users have gained followers due to transit-usernames or have lost followers due to revisited-username. For 287 users (64\%), transit usernames cause a gain of followers ($\geq$ 1). A user with 1 follower earlier, gains 106,970 followers within 3 days with her transit username, and yet changes back to her earlier username.Further, only 6\% of such users who gain via transit-username, switch back to the earlier username within 24 hours. We think that users who gain via transit usernames intend to wait till they achieve maximum benefit of followers and then switch back to earlier username due to other reasons explained in the section.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.38]{spam.png}
\caption{User changes username to pick a brand's old username (transit-username), receives no followers gain, and irrelevant tweets, and hence, switches back to his earlier username.}
\label{fig:spam}
\end{figure}
\begin{table*}[htbp]
\begin{center}
\small
\begin{tabular}{|l|l|l|l|l|p{3cm}|}
\hline
\raggedright{{\bf ID}} & {Scan - I} & {Scan - II } & {Scan - III} & {Group} & {Date of observation when user holds the tracked username} \\ \hline
12xx62463x & \textbf{CollaGe\_InFo} & Dictionary\_ID & Dictionary\_ID & Sajan Group & 2013-04-01 \\ \hline
11xx79686x & DaiLy\_GK & \textbf{CollaGe\_InFo} & Geo\_Account & Sajan Group & 2013-10-02 \\ \hline
95xx1822x & Geonewspak9 & DictioNary\_GK & \textbf{CollaGe\_InFo} & Sajan Group & 2013-10-25 \\ \hline
19xx56472x & - & - & \textbf{CollaGe\_InFo} & Sajan Group & 2013-12-04 \\
\hline
\hline
60xx2762x & \textbf{Peshawar\_sMs} & MoBile\_TricKes & BBC\_PAK\_NEWS & Sajan Group & 2013-04-08 \\ \hline
11xx37099x & Vip\_Wife & \textbf{Peshawar\_sMs} & UBL\_Cricket & Sajan Group & 2013-10-25 \\ \hline
28xx1645x & NFS002cric & NaKaaM\_LiFe& \textbf{Peshawar\_sMs} & Sajan Group & 2013-11-08 \\ \hline
\hline
70xx9502x & \textbf{FuNNy\_SardaR} & MaST\_DuLHaN & KaiNaT\_LipS & Khan Group & 2013-04-01 \\ \hline
99xx9356x & SaIrA\_JoX & \textbf{FuNNy\_SardaR} & MaST\_DuLHaN & Khan Group & 2013-10-02 \\ \hline
12xx73970x & - & - & \textbf{FuNNy\_SardaR} & Khan Group & 2013-12-04 \\ \hline
\end{tabular}
\caption{Rotational use of a username among the members of a group in Twitter. We observe that different users pick the same username at different times, which might intend towards promotion of the username.}
\label{tab:rotation}
\end{center}
\vspace{-3mm}
\end{table*}
\subsection{Rotational use of a username in a group} \label{rotation}
Owing to limited number of users in fifteen minute dataset who changed their usernames, we use our four scans and seed set of 8.7M users for this analysis. We observe that a few user profiles, who change their usernames, belong to or follow a group and their past usernames are picked by other user profiles following / belonging to the same group. Table~\ref{tab:rotation} shows two such groups and the rotation of a username among the profiles, as observed in four scans. Username `Collage\_InFo' is used by a user with user id 12xx6246xx as recorded in Scan-I. The same username is later used by user with user id 11xx7968xx as observed in Scan-II, followed by user 95xx182xx in Scan-III and by different user later. All users who use a particular username at different timestamps, claim that they belong to a group named as `Sajan Group', either in their name attribute or in their bio attribute. Similar behavior is observed with username `Peshawar\_sMs' and `FuNNy\_SardaR'. We observe 70 such usernames in four scans and seed set, which are picked by different users at different timestamps. We speculate that either user profiles who keep the same username belong to the same real-world user, or the intention is to popularize the username itself. We think that users change usernames in order to let other users in the group use the username.~\footnote{We do not tweet such users to avoid alarming them in case they have malicious intentions.}
\subsection{Other possible reasons}
Other reasons we either inspect manually from our dataset or learn by tweeting and asking the users we monitor are --
\begin{itemize}
\item
\vspace{-3mm}
\textbf{Change of username identifiability} -- Few users in our dataset change usernames to reverse the identifiability of the usernames i.e. either to make them personal or anonymous. For instance, a user named `loried ligarreto' changes her username from `loriedligarreto' to `sienteteotravez' (feel again in English) implying that user intends to make her username anonymous. In other instances, we observe users who previously pick less identifiable usernames, make them personal later. For example, a user named `rodrigo' changes her username from `unosojosverdes' (green eyes in English) to `rodrigothomas\_', thereby implicating that user wishes to associate her real identity to her username.
\item
\textbf{Change of Events} -- A user responds that she represents Sahara India FabClub. She has supported Sahara's Pune Warriors team in IPL event with username `pwifanclub' and then Sahara F1 team with username `ForceIndia@!' and therefore has changed her username (see Figure~\ref{fig:changes1}).
\item
\textbf{No specific reason} -- Few users respond that they change their usernames without any specific reasons, that they get bored of the earlier one (see Figure~\ref{fig:changes2}).
\end{itemize}
\begin{figure}[htbp]
\centering
\subfigure[]{
\includegraphics[scale=0.35]{event_change_an.png}\label{fig:changes1}
}
\quad
\subfigure[]{
\includegraphics[scale=0.35]{tired_old_an.png}\label{fig:changes2}
}
\caption{Username change (a) due to change of events over time, (b) due to boredom. }
\label{fig:how}
\end{figure}
\section{Applications} \label{Applications}
We now discuss the application domains where recorded past usernames can be helpful. We show that past usernames of a user can help Twitter to come up with customized username suggestions and can help to create a unified social footprint of a user by locating her usernames on other OSNs. Tracking a username and user profiles who choose a particular username can help in correlating user profiles to a single real-world person.
\subsection{Username suggestion}
We now know that users change usernames frequently on Twitter. They change to non similar usernames as compared to earlier ones, which help them to increase their popularity, gain more space, maintain their anonymity (identity), or avoid boredom. We suggest that the understanding of such facts can help Twitter to suggest usernames to its users, they may wish to change to, by recording and analyzing their past usernames and assessing their suitability to the users. Suggestions can further be customized based on users' username selection patterns. We think that suggesting usernames to users may improve users' experience with Twitter, in terms of better visibility, newer connections and higher relevant interactions with newly suggested usernames.
\subsection{Identity resolution}
With a user registered on multiple OSNs with varying characteristics, it is difficult to find her identity on multiple OSNs. The problem of finding and resolving a user's multiple identities across different OSNs is termed as ``Identity resolution in OSNs" \cite{jain2013seek}. Till now, researchers have suggested to use most recent profile attributes of users to help in identity resolution~\cite{ZafaraniL13, perito,Irani,Malhotra, goga2013exploiting}, however little research explores the potential of past profile attribute changes, specifically usernames, in solving identity resolution in OSNs.\\
\indent We hypothesize that a user might use any of her past usernames from Twitter as her username on other OSNs, to remember selected chosen usernames for all OSNs. We test this hypothesis with users who change their username at least once as recorded in fifteen-minute dataset and the four scans. To test the hypothesis, we need users' other OSNs usernames, to compare them with users' past usernames. However, we don't have direct access to such information. Therefore, we use self-identification method~\cite{jain2013seek} to search and find users' usernames on other OSNs. We extract their self-claimed account (username) on other OSNs, via their URL attribute value. For instance, a user mentions `$http://www.facebook.com/xyzabc$' as her URL attribute thereby exposes and self-identifies her Facebook username as `$xyzabc$'. With this methodology, we find 214,636 out of 853,827 users and 993 out of 2,698 users, mention their other OSN accounts via URL attribute on Twitter. We then compare extracted usernames on other OSNs with past usernames of the user on Twitter.~\footnote{Few users change URL temporally to direct to their other OSN accounts or new account on an earlier referred OSN.}\\ \indent We find 9\% of 214,636 users and 8\% of 993 users choose at least one of their past usernames to register on other OSNs. Therefore, an exact string search with user's past usernames can help locating her other OSNs accounts. Note that, we might have missed users for whom the hypothesis is true, because of no information on their other OSN accounts. Further, we note that a few users change their usernames on other OSNs in the same pattern as they do on Twitter. For instance, a Twitter user changes username from ``happygul@!!" to``gulben!!!" as well as changes her Instagram~\footnote{www.instagram.com} username from ``happygul@!!" to ``gulben!!!", as observed in the URL attribute of the user. Therefore, historic usernames of a user may help to suggest her username on other OSNs and therefore, may address identity resolution in OSNs.
\subsection{Multiple identities correlation}
As observed in Section~\ref{rotation}, few users pick same username at different timestamps. We observe 70 such usernames in four scans and seed set, which are picked by different users at different timestamps. We inquire about few instances and send tweets to users, asking what are the possible reasons for such a behavior. Two users reply as `` Both the user profiles belong to me." (see Figure~\ref{fig:multiple_rotation}). In this way, multiple identities of a real-world user may be correlated and linked together via monitoring the user profiles and their username changes, if they use the same set of usernames over time, thus helping multiple identities correlation on Twitter.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.37]{multiple_rotation_an.png}
\caption{User mentions that two different profiles had same username (@mrk!!!) at different times, because two profiles (@mrk!!! and @wagle!!!!) belong to the same person. }
\label{fig:multiple_rotation}
\end{figure}
\section{Discussion} \label{Discussion}
In this work, we attempt to characterize, an unexplored username changing behavior on an OSN. We observe 8.7 million users at long intervals while 10,000 users at short intervals. We derive that around 10\% of users change their username owing to reasons such as space gain, gain more followers or switch with other user's username. Few users undergo a username change multiple times (96 times in 117 days) while many users change their usernames once. In some scenarios, users switch back to their past usernames. Number of times a user changes her username is weakly correlated to her popularity, activity and age of her account. We further observe that users choose their username on other OSNs out of the usernames they used on Twitter in the past. Further, monitoring which users pick the same username at different timestamps might help in understanding if different online users reflect the single real-world entity.\\
\indent We discuss quantifiable analysis, draw inferences and attempt to validate the inferences with user feedback. We wrote tweets to few Twitter users we monitored, asking them the reasons, and benefits of their username change. Only few users responded and most ignored the tweet. We then planned a survey of 10 questions customized for each user with their past usernames.~\footnote{http://nemo.iiitd.edu.in/findingnemo/user1/} We designed the survey to comprehend the reasons, benefits or losses for which users changed their usernames. We tweeted the survey to users with a message ``Changed ur username $<$\#$>$times in $\#$months? \#Help us know why. Fill $<$survey link$>$" over a week using two Twitter accounts. We tweeted the survey to only 500 users out of 2,698 users, with @tag. The reason was that our accounts got suspended multiple times because we used automated scripts to send tweets.~\footnote{We even randomized our tweeting time, but Twitter still suspended our accounts.} Three users filled the survey, two males and one female. One user responded that she changed her username to gain space, another user responded that he changed because earlier username was inappropriate and the third user changed for no specific reason. Further one male user switched back to his past username, because he gained followers due to revisited-username and lost followers due to transit usernames. Due to limited responses, we do not list other reasons for changing usernames. \\
\indent In this work, we strictly focus on the username attribute changes over time. However, we observe that apart from username, other profile attributes change over time as well. Users change their name, description, location, URL and profile picture more frequently than their username. By monitoring identity changes of a user, a user's true identity attributes, missing attributes and recency / validity of their attributes on OSNs can be estimated. We sense an immense potential in the historical data of an OSN user.
\section{Related Work}
Researchers have examined the temporal nature of two important user attributes on an OSN namely content and network, however little has been explored about temporal changes to profile attributes of the user. Content attribute of users were studied on Twitter to understand their posting behavior. Authors suggested that at any time, users' post characteristics depends on three factors -- breaking news, posts from social friends and the user's interests~\cite{Xu:2012:MUP:2348283.2348358}. A user's interests in terms of topics she posts about, were studied to capture temporal changes in her topical interests and therefore her posting behavior~\cite{abel2011analyzing}. Depending on the user's pace of keeping unto new topics, authors built a framework to model users' temporally changing interests and used the model to build a personalized recommender system. Network attributes of a user were studied on OSNs, in terms of evolution of her friend connections~\cite{fan2011incremental,mislove2008growth}, and involvement in groups~\cite{motamedicharacterizing}. \\
\indent Profile attributes were studied temporally by \cite{5231867}, where authors crawled 2 million Myspace profiles twice over a period of a year. Authors compared the evolved honesty and accountability of the users, derived from the user's profile, content and network attributes. Authors however did not present any insights on why users preferred an identity change at the first place. Further, reasons of an identity change were explored and reasoned in the real-world by sociology and psychology researchers~\cite{mcfarland2005motives,burke2006identity}. Researchers observed 6,000 high school adolescents after a year, found identity changes of a user in terms of affiliations of the user in the group, description of herself \cite{mcfarland2005motives}. They found that network attributes of a user such as betweenness play a major role in understanding if a user is likely to change identity. Other reasons such as maturity of the adolescent, also plays a role in the likelihood of an identity change. To the best of our knowledge, we could not find any research which focused and explored changes to a unique attribute of a user on OSN i.e. username. We, in this work, addressed the research gaps, by conducting a study on Twitter users who change their profile attributes over time, and focused on a user's unique identifiable attribute -- username.
\section{Limitations and Future Work}
We make the first attempt to study temporal changes in username attribute of a user and therefore acknowledge limitations of our work. Firstly, we monitor a small set of 10K users who have demonstrated the act of username change earlier in the past. We plan to expand this dataset, and generalize our findings. Secondly, we understand that our observation periods could have been longer. But with the frequency of username changes we observe, we think that inferences derived with our observation period of 4 months, can be generalized to a longer observation duration. Lastly, we receive limited user feedback to support our reasons and inferences. In future, we plan to perform an extensive user study and build automated systems to prove quantifiably that past identity change patterns could help in username suggestion, identity resolution and multiple identities correlation. We plan to investigate temporal changes to other profile attributes of Twitter users in our future work.
{
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{intro}
Half a century has passed since Alexander Andreev reported the curious retro-reflection of electrons at the interface between a normal metal and a superconductor \cite{And64}. One reason why Andreev reflection is still very much a topic of active research, is the recent interest in Majorana zero-modes \cite{Sil14}: Nondegenerate bound states at the Fermi level ($E=0$) consisting of a coherent superposition of electrons and holes, coupled via Andreev reflection. These are observed in the differential conductance as a resonant peak around zero bias voltage $V$ that does not split upon variation of a magnetic field $B$ \cite{Ali12,Lei12,Sta13,Bee13b}. In the $B,V$ plane the conductance peaks trace out an unusual Y-shaped profile, distinct from the more common X-shaped profile of peaks that meet and immediately split again. (See Fig.\ \ref{fig_XY}.)
\begin{figure}[tb]
\centerline{\includegraphics[width=0.8\linewidth]{XY}}
\caption{Left panel: Magnetic field $B$-dependence of peaks in the differential conductance $G=dI/dV$. The peak positions trace out an X-shaped or Y-shaped profile in the $B$-$V$ plane. Right panel: Location of the poles of the scattering matrix $S(\varepsilon)$ in the complex energy plane $\varepsilon=E-i\gamma$. The arrows indicate how the poles moves with increasing magnetic field.
}
\label{fig_XY}
\end{figure}
It is tempting to think that the absence of a splitting of the zero-bias conductance peak demonstrates that the quasi-bound state is nondegenerate, hence Majorana. This is mistaken. As shown in a computer simulation \cite{Pik12}, the Y-shaped conductance profile is generic for superconductors with broken spin-rotation and broken time-reversal symmetry, irrespective of the presence or absence of Majorana zero-modes. The theoretical analysis of Ref.\ \onlinecite{Pik12} focused on the ensemble-averaged conductance peak, in the context of the weak antilocalization effect \cite{Bro95,Alt96,Ios12,Bag12}. Here we analyse the sample-specific conductance profile, by relating the X-shape and Y-shape to different configurations of poles of the scattering matrix in the complex energy plane \cite{Pik11}.
\section{Andreev billiard}
\label{Andreevbilliard}
\begin{figure}[tb]
\centerline{\includegraphics[width=0.7\columnwidth]{SystemSetup}}
\caption{Schematic illustration of an Andreev billiard.}
\label{fig:SystemSetup}
\end{figure}
\subsection{Scattering resonances}
\label{scatres}
We study the Andreev billiard geometry of Fig.\ \ref{fig:SystemSetup}: A semiconductor quantum dot strongly coupled to a superconductor and weakly coupled to a normal metal. In the presence of time-reversal symmetry an excitation gap is induced in the quantum dot by the proximity effect \cite{Bee04}. We assume that the gap is closed by a sufficiently strong magnetic field. Quasi-bound states can then appear near the Fermi level ($E=0$), described by the Hamiltonian
\begin{equation}
{\cal H} = \sum_{\mu, \nu} |\mu\rangle H_{\mu \nu} \langle\nu| + \sum_{\mu, a} \bigl( |\mu\rangle W_{\mu a} \langle a| + |a\rangle W^{\ast}_{\mu a} \langle\mu| \bigr). \label{Eq01}
\end{equation}
The bound states in the closed quantum dot are eigenvalues of the $M\times M$ Hermitian matrix $H=H^{\dagger}$. The $M\times N$ matrix $W$ couples the basis states $|\mu\rangle$ in the quantum dot to the normal metal, via $N$ propagating modes $|a\rangle$ through a point contact. In principle we should take the limit $M\rightarrow\infty$, but in practice $M\gg N$ suffices.
The amplitudes of incoming and outgoing modes in the point contact at energy $E$ (relative to the Fermi level) are related by the $N\times N$ scattering matrix \cite{Guh98,Bee97}
\begin{equation}
S(E) = 1 + 2 \pi i W^{\dagger} \left(H - i \pi WW^{\dagger}-E\right)^{-1}W.\label{Eq03}
\end{equation}
This is a unitary matrix, $S(E)S^{\dagger}(E)=1$.
A scattering resonance corresponds to a pole $\varepsilon=E-i\gamma$ of the scattering matrix in the complex energy plane, which is an eigenvalue of the non-Hermitian matrix
\begin{equation}
H_{\rm eff}=H - i \pi WW^{\dagger}.\label{Heffdef}
\end{equation}
The positive definiteness of $WW^{\dagger}$ ensures that the poles all lie in the lower half of the complex plane, $\gamma\geq 0$, as required by causality. Particle-hole symmetry implies that $\varepsilon$ and $-\varepsilon^{\ast}$ are both eigenvalues of $H_{\rm eff}$, so the poles are symmetrically arranged around the imaginary axis.
The differential conductance $G(V)=dI/dV$ of the quantum dot, measured by grounding the superconductor and applying a bias voltage to the normal metal, is obtained from the scattering matrix via \cite{Pik12}
\begin{align}
G(V)=\frac{e^{2}}{h}\left[\frac{N}{2}-\frac{1}{2}\,{\rm Tr}\,S(eV)\tau_z S^\dagger(eV)\tau_z\right],\label{Gehbasis}
\end{align}
in the electron-hole basis, and
\begin{align}
G(V)=\frac{e^{2}}{h}\left[\frac{N}{2}-\frac{1}{2}\,{\rm Tr}\,S(eV)\tau_y S^\dagger(eV)\tau_y\right],\label{GMbasis}
\end{align}
in the Majorana basis. The Pauli matrices $\tau_y$, $\tau_z$ act on the electron-hole degree of freedom. The two bases are related by the unitary transformation
\begin{equation}
S\mapsto USU^{\dagger},\;\;U=\sqrt{\tfrac{1}{2}}\begin{pmatrix}
1&1\\
i&-i
\end{pmatrix}.\label{Udef}
\end{equation}
\subsection{Gaussian ensembles}
\label{gaussens}
For a random-matrix description we assume that the scattering in the quantum dot is chaotic, and that this applies to normal scattering from the electrostatic potential as well as to Andreev scattering from the pair potential. In the large-$M$ limit we may then take a Gaussian distribution for $H$,
\begin{equation}
P(H)\propto \exp\left(-\frac{c}{M} \,{\rm Tr}\,H^2\right).\label{Eq02}
\end{equation}
By taking the matrix elements of $H$ to be real, complex, or quaternion numbers (in an appropriate basis), one obtains the Wigner-Dyson ensembles of non-superconducting chaotic billiards \cite{Mehta,Forrester,handbook}. Particle-hole symmetry then plays no role, because normal scattering does not couple electrons and holes.
Altland and Zirnbauer introduced the particle-hole symmetric ensembles appropriate for an Andreev billiard \cite{Alt97}. The two ensembles without time-reversal symmetry are obtained by taking the matrix elements of $i\times H$ (instead of $H$ itself) to be real or quaternion. When $iH$ is real there is only particle-hole symmetry (class D), while when $iH$ is quaternion there is particle-hole and spin-rotation symmetry (class C).
Both the Wigner-Dyson (WD) and the Altland-Zirnbauer (AZ) ensembles are characterized by a parameter $\beta\in\{1,2,4\}$ that describes the strength of the level repulsion factor in the probability distribution of distinct eigenvalues $E_i$ of $H$: a factor $\prod_{i<j}|E_i-E_j|^\beta$ in the WD ensembles and a factor $\prod'_{i<j}|E_i^2-E_j^2|^\beta$ in the AZ ensembles. (The prime indicates that the product includes only the positive eigenvalues.)
In the WD ensembles the parameter $\beta$ also counts the number of degrees of freedom of the matrix elements of $H$: $\beta=1$, 2 or 4 when $H$ is real, complex, or quaternion, respectively. In the AZ ensembles this connection is lost: $\beta=2$ in the class C ensemble ($iH$ real) as well as in the class D ensemble ($iH$ quaternion).
The coefficient $c$ can be related to the average spacing $\delta_0$ of distinct eigenvalues of $H$ in the bulk of the spectrum,
\begin{equation}
c=\frac{\beta\pi^{2}}{8\delta_0^{2}}\times\begin{cases}
2& \text{in the WD ensembles},\\
1& \text{in the AZ ensembles.}
\end{cases}\label{cdef}
\end{equation}
The coefficient \eqref{cdef} for the AZ ensembles is twice as small as it is in the WD ensembles with the same $\beta$, on account of the $\pm E$ symmetry of the spectrum, see App.\ \ref{cpmEsym}.
Because the distribution of $H$ is basis independent, we may without loss of generality choose a basis such that the coupling matrix $W$ is diagonal,
\begin{equation}
W_{mn}=w_n \delta_{mn},\;\;1\leq m\leq M,\;\;1\leq n\leq N.\label{Wdef}
\end{equation}
The coupling strength $w_n$ is related to the tunnel probability $\Gamma_n\in (0,1)$ of mode $n$ into the quantum dot by \cite{Guh98,Bee97}
\begin{equation}
|w_n|^2 = \frac{M\delta_0}{\pi^2 \Gamma_n}\bigl( 2 - \Gamma_n - 2 \sqrt{1-\Gamma_n} \bigr).\label{Eq04}
\end{equation}
\subsection{Class C and D ensembles}
\label{symclassCD}
We summarize the properties of the $\beta=2$ Altland-Zirnbauer ensembles, symmetry class C and D, that we will need for our study of the Andreev resonances. (See App.\ \ref{CIDIIIensembles} for the corresponding $\beta=1,4$ formulas in symmetry class CI and DIII.) Similar formulas can be found in Ref.\ \onlinecite{Iva02}.
When Andreev scattering operates together with spin-orbit coupling, one can combine electron and hole degrees of freedom from the same spin band into a real basis of Majorana fermions. [This change of basis amounts to the unitary transformation \eqref{Udef}.] In the Majorana basis the constraint of particle-hole symmetry reads simply
\begin{equation}
H=-H^{\ast},\label{Dconstraint}
\end{equation}
so we can take $H=iA$ with $A$ a real antisymmetric matrix. In the Gaussian ensemble the upper-diagonal matrix elements $A_{nm}$ ($n<m$) all have identical and independent distributions,
\begin{equation}
P(\{A_{nm}\})\propto\prod_{1=n<m}^{M}\exp\left(-\frac{\pi^{2}A_{nm}^{2}}{2M\delta_0^{2}}\right),\label{GaussEns}
\end{equation}
see Eqs.\ \eqref{Eq02} and \eqref{cdef}. This is the $\beta=2$ class-D ensemble, without spin-rotation symmetry.
The $\beta=2$ class-C ensemble applies in the absence of spin-orbit coupling, when spin-rotation symmetry is preserved. Andreev reflection from a spin-singlet superconductor couples only electrons and holes from different spin bands, which cannot be combined into a real basis state. It is then more convenient stay in the electron-hole basis and to eliminate the spin degree of freedom by considering a single spin band for the electron and the opposite spin band for the hole. (The matrix dimensionality $M$ and the mean level spacing $\delta_0$ then refer to a single spin.) In this basis the particle-hole symmetry requires
\begin{equation}
H=-\tau_y H^\ast \tau_y,\label{Cconstraint}
\end{equation}
where the Pauli matrix $\tau_y$ operates on the electron and hole degrees of freedom.
The constraint \eqref{Cconstraint} implies that $H=iQ$ with $Q$ a quaternion anti-Hermitian matrix. Its matrix elements are of the form
\begin{equation}
\begin{split}
&Q_{nm}=a_{nm}\tau_0+ib_{nm}\tau_x+ic_{nm}\tau_y+id_{nm}\tau_z,\\
&n,m=1,2,\ldots M/2,
\end{split}
\label{quaterniondef}
\end{equation}
with real coefficients $a,b,c,d$ (to ensure that $Q_{nm}=\tau_y Q_{nm}^{\ast}\tau_y$). Anti-Hermiticity of $Q$ requires that the off-diagonal elements are related by $a_{nm}=-a_{mn}$ and $x_{nm}=x_{mn}$ for $x\in\{b,c,d\}$. On the diagonal $a_{nn}=0$. In the Gaussian ensemble the independent matrix elements have the distribution
\begin{align}
&P(\{Q_{nm}\})\propto\prod_{n=1}^{M/2}\exp\left(-\frac{\pi^{2}}{2M\delta_0^{2}}(b_{nn}^2+c_{nn}^2+d_{nn}^2)\right) \nonumber\\
&\quad\times \prod_{1=n<m}^{M/2}\exp\left(-\frac{\pi^{2}}{M\delta_0^{2}}(a_{nm}^2+b_{nm}^2+c_{nm}^2+d_{nm}^2)\right),\label{GaussEnsC}
\end{align}
\section{Andreev resonances}
\label{Aresonances}
\subsection{Accumulation on the imaginary axis}
\label{accumulation}
\begin{figure*}[tb]
\centerline{\includegraphics[width=0.8\linewidth]{scatterplots}}
\caption{Scatter plot of the poles $\varepsilon=E-i\Gamma$ of 5000 scattering matrices $S(\varepsilon)$, in the Gaussian ensembles of class D, C, and A (first, second, and third column), for ballistic coupling ($\Gamma=1$, first row) and for tunnel coupling ($\Gamma=0.2$, second row). In each case the Hamiltonian has dimension $M\times M=500\times 500$ and the scattering matrix $N\times N=50\times 50$. Only a narrow energy range near $E=0$ is shown, to contrast the accumulation of the poles on the imaginary axis in class D and the repulsion in class C. The blue horizontal lines indicate the expected boundaries \eqref{rhoEgammab} of the class-A scatter plot in the limit $N,M/N\rightarrow\infty$.}
\label{fig_poles}
\end{figure*}
In Fig.\ \ref{fig_poles} we show the location of the poles of the scattering matrix in the complex energy plane, for the $\beta=2$ Altland-Zirnbauer ensembles with and without spin-rotation symmetry (class C and D, respectively). The $\beta=2$ Wigner-Dyson ensemble (class A, complex $H$) is included for comparison. The poles are eigenvalues $\varepsilon$ of the non-Hermitian effective Hamiltonian \eqref{Eq01}, with $H$ distributed according to the Gaussian distribution \eqref{Eq02}--\eqref{cdef}, $\beta=2$, and coupling matrix $W$ given by Eqs.\ \eqref{Wdef}--\eqref{Eq04}. For simplicity we took identical tunnel probabilities $\Gamma_n\equiv\Gamma$ for each of the $N$ modes connecting the quantum dot to the normal metal.
The number $M$ of basis states in the quantum dot is taken much larger than $N$, to reach the random-matrix regime. In class C this number is necessarily even, as demanded by the particle-hole symmetry relation \eqref{Cconstraint}. The symmetry relation \eqref{Dconstraint} in class D imposes no such constraint, and when $M$ is odd there is an unpaired Majorana zero-mode in the spectrum \cite{Iva02,note3}. The class-D superconductor with a Majorana zero-mode is called topologically nontrivial, while class C or class D without a zero-mode is called topologically trivial \cite{Ryu10,Has10,Qi11}. For a more direct comparison of class C and class D we take $M$ even in both cases, so both superconductors are topologically trivial.
In the absence of particle-hole symmetry (class A), the poles $\varepsilon=E-i\gamma$ of the scattering matrix have a density \cite{Fyo97}
\begin{align}
&\rho(E,\gamma)=\frac{N}{4\pi\gamma^2},\;\;\gamma_{\rm min}<\gamma<\gamma_{\rm max},\label{rhoEgammaa}\\
&\gamma_{\rm min}=N\Gamma\delta_0/4\pi,\;\;\gamma_{\rm max}=\gamma_{\rm min}/(1-\Gamma),
\label{rhoEgammab}
\end{align}
for $|E|\ll M\delta_0$ and asymptotically in the limit $N,M/N\rightarrow\infty$. For $|E|\gtrsim\delta_0$ all three $\beta=2$ ensembles A, C, D have a similar density of poles, but for smaller $|E|$ the densities are strikingly different, see Fig.\ \ref{fig_poles}. While in class C the poles are repelled from the imaginary axis, in class D they accumulate on that axis.
As pointed out in Ref.\ \onlinecite{Pik11}, a nondegenerate pole $\varepsilon=-i\gamma$ on the imaginary axis has a certain stability, it cannot acquire a nonzero real part $E$ without breaking the $\varepsilon\leftrightarrow-\varepsilon^{\ast}$ symmetry imposed by particle-hole conjugation. To see why this stability is not operative in class C, we note that on the imaginary axis $\gamma$ is a real eigenvalue of the matrix
\begin{align}
&\Omega=-Q+\pi WW^{\dagger}\;\;\text{in class C},\label{OmegadefC}\\
&\Omega=-A+\pi WW^{\dagger}\;\;\text{in class D}.\label{OmegadefD}
\end{align}
In both classes the matrix $\Omega$ commutes with an anti-unitary operator, ${\cal C}\Omega=\Omega{\cal C}$, with ${\cal C}=i\tau_{y}{\cal K}$ in class C and ${\cal C}={\cal K}$ in class D. (The operator ${\cal K}$ performs a complex conjugation.) In class C this operator ${\cal C}$ squares to $-1$, so a real eigenvalue $\gamma$ of $\Omega$ has a Kramers degeneracy \cite{note2} and hence nondegenerate poles $\varepsilon=-i\gamma$ on the imaginary axis are forbidden. In class D, in contrast, the operator ${\cal C}$ squares to $+1$, Kramers degeneracy is inoperative and nondegenerate poles are allowed and in fact generic.
\subsection{Square-root law}
\label{squareroot}
As we explain in App.\ \ref{orthogonalmapping}, for ballistic coupling ($\Gamma=1$) the statistics of poles on the imaginary axis can be mapped onto the statistics of the real eigenvalues of an $M\times M$ random orthogonal matrix with $N$ rows and columns deleted --- which is a solved problem \cite{Kho10,For10}. The linear density profile $\rho_0(\gamma)$ on the imaginary axis is
\begin{equation}
\rho_0(\gamma)=\sqrt{\frac{N\Gamma}{8\pi}}\,\frac{1}{\gamma},\;\;\gamma_{\rm min}<\gamma<\gamma_{\rm max},\label{rho0result}
\end{equation}
for $1\ll N\Gamma\ll M$ and $\gamma_{\rm min}$, $\gamma_{\rm max}$ given by Eq.\ \eqref{rhoEgammab}. We {\em conjecture} that this density profile, derived \cite{Kho10} for $\Gamma=1$, holds also for $\Gamma<1$. In Fig.\ \ref{fig_densities} we give numerical evidence in support of this conjecture.
\begin{figure}[tb]
\centerline{\includegraphics[width=0.8\linewidth]{densities_fit}}
\caption{Double-logarithmic plot of the probability distribution $\rho(\gamma)$, normalized to unity, of the imaginary part $\gamma$ of the poles of the scattering matrix. The curves are calculated by averaging over some 2000 realizations of the class-D ensemble, with $N=10$, $M=500$, $\Gamma=0.9$. The red dashed curve includes all poles, while the blue solid curve includes only the poles on the imaginary axis ($E=0$). The black dashed lines are the predicted slopes from Eq.\ \eqref{rhoEgammaa} and \eqref{rho0result}.
}
\label{fig_densities}
\end{figure}
\begin{figure}[tb]
\centerline{\includegraphics[width=0.8\linewidth]{G_sqrt_M}}
\caption{Average of the number $N_{\rm Y}$ of poles on the imaginary axis for an $N\times N$ scattering matrix $S(\varepsilon)$ in symmetry class D. Colors distinguish different tunnel couplings $\Gamma<1$, and $N$ is increased together with $M=80\,N$. The slope of the dashed black line is the large-$N$ asymptote \eqref{squarerootlaw}.}
\label{fig_polesgammacount}
\end{figure}
In Fig.\ \ref{fig_polesgammacount} we show how the average number $\langle N_{\rm Y}\rangle$ of class-D poles on the imaginary axis depends on the dimensionality $N$ of the scattering matrix and on the tunnel probability $\Gamma$. We compare with the square-root law \cite{note4}
\begin{equation}
\langle N_{\rm Y}\rangle=-\sqrt{\frac{N\Gamma}{8\pi}}\ln(1-\Gamma),\label{squarerootlaw}
\end{equation}
implied by integration of our conjectured density profile \eqref{rho0result}. This $\sqrt{N}$ scaling is generic for random-matrix ensembles that exhibit accumulation of eigenvalues on the real or imaginary axis, such as the Ginibre ensemble \cite{Gin65,Leh91,Ede94} (real Gaussian matrices without any symmetry) and the Hamilton ensemble \cite{Bee13} (matrices of the form ${\cal M}=HJ$ with $H$ a symmetric real Gaussian matrix and $J={\begin{pmatrix}
0&1\\
-1&0
\end{pmatrix}}$ a fixed anti-symmetric matrix). Fig.\ \ref{fig_polesgammacount} shows that the Andreev resonances follow the same square-root law.
\section{X-shaped and Y-shaped conductance profiles}
\begin{figure}[tb]
\centerline{\includegraphics[width=0.75\columnwidth]{Pole_dynamics}}
\caption{Parametric evolution of the differential conductance $G(V,\alpha)$ (color scale) and the real part $E$ of the poles of the scattering matrix $S_{\alpha}(\varepsilon)$. These are results for a single realization of the class D ensemble with $M=120$, $N=6$, and $\Gamma=0.3$.
}
\label{fig:Pole_dynamics}
\end{figure}
\begin{figure}[tb]
\centerline{\includegraphics[width=0.8\columnwidth]{Cuts}}
\caption{Four cuts through the parametric evolution of Fig.\ \ref{fig:Pole_dynamics}, showing the differential conductance $G=dI/dV$ (top row) and scattering matrix poles $\varepsilon=E-i\gamma$ (bottom row).
}
\label{fig:Cuts}
\end{figure}
In Ref.\ \onlinecite{Pik12} it was found in a computer simulation of a superconducting InSb nanowire that the conductance resonances trace out two distinct profiles in the voltage-magnetic field plane: an X-shape or a Y-shape. In the X-shaped profile a pair of conductance resonances merges and immediately splits again upon variation of voltage $V$ or magnetic field $B$. In the Y-shaped profile a pair of peaks merges at $V=0$ and then stays pinned to zero voltage over a range of magnetic field values. Here we wish to relate this phenomenology to the parametric evolution of poles of the scattering matrix in the complex energy plane \cite{Pik11}.
For that purpose we introduce a parameter dependence in the Hamiltonian $H$ of the Andreev billiard,
\begin{equation}
H_\alpha = (1-\alpha) H_0 + \alpha H_1, \label{Eq05}
\end{equation}
and calculate the differential conductance as a function of $V$ and $\alpha$. We work in symmetry class D (broken time-reversal and broken spin-rotation symmetry), so $H_0$ and $H_1$ are purely imaginary antisymmetric matrices (in the Majorana basis). We draw them from the Gaussian distribution \eqref{GaussEns}. The scattering matrix $S_\alpha$, obtained from $H_\alpha$ via Eq.\ \eqref{Eq03}, gives the differential conductance $G(V,\alpha)$ via Eq.\ \eqref{GMbasis}. For each $\alpha$ we also compute the poles $\varepsilon=E-i\gamma$ of $S(\varepsilon)$ in the complex energy plane.
Fig.\ \ref{fig:Pole_dynamics} shows a typical realization where the number $N_{\rm Y}$ of conductance poles on the imaginary axis switches between zero and two when $\alpha$ varies in the interval $[0,1]$. The color-scale plot shows $G(V,\alpha)$, while the dots trace out the projection of the poles of $S_{\alpha}(\varepsilon)$ on the real axis. Labels X and Y indicate the two types of profiles, and Fig.\ \ref{fig:Cuts} shows the corresponding conductance peaks and scattering matrix poles.
Inspection of the figures shows that the X-shaped profile appears when two scattering matrix poles cross when projected onto the real axis. (They do not cross in the complex energy plane.) The Y-shaped profile appears when $N_{\rm Y}$ jumps by two.
\section{Conclusion}
\label{conclude}
For a closed superconducting quantum dot, the distinction between topologically trivial and nontrivial is the absence or presence of a level pinned to the middle of the gap (a Majorana zero-mode). When the quantum dot is connected to a metallic reservoir, the bound states become quasi-bound, $E\mapsto E-i\gamma$, with a finite life time $\hbar/2\gamma$. The distinction between topologically trivial and nontrivial then becomes whether the number $N_{\rm Y}$ of quasi-bound states with $E=0$ is even or odd.
One can now distinguish two types of transitions \cite{Pik11}: At a topological phase transition $N_{\rm Y}$ changes by $\pm 1$ \cite{note3}. At a ``pole transition'' $N_{\rm Y}$ changes by $\pm 2$. Both types of transitions produce the same Y-shaped conductance profile of two peaks that merge and stick together for a range of parameter values --- distinct from the X-shaped profile that happens without a change in $N_{\rm Y}$.
There is a variety of methods to distinguish the pole transition from the topological phase transition \cite{Pik12}: Since $N_{\rm Y}\simeq \Gamma^{3/2}\sqrt{N}$ for $\Gamma\ll 1$, one way to suppress the pole transitions is to couple the metal to the superconductor via a small number of modes $N$ with a small transmission probability $\Gamma$. The pole transitions are a sample-specific effect, while the topological phase transition is expected to be less sensitive to microscopic details of the disorder. One would therefore not expect the pole transitions to reproduce in the same sample upon thermal cycling. If one can measure from both ends of a nanowire, one might search for correlations between the conductance peaks at the two ends. The Majorana zero-modes come in pairs, one at each end, so there should be a correlation in the conductance peaks measured at the two ends, which we would not expect to be there for the peaks due to the pole transition.
\acknowledgments
This research was supported by the Foundation for Fundamental Research on Matter (FOM), the Netherlands Organization for Scientific Research (NWO/OCW), an ERC Synergy Grant, and the China Scholarship Council.
|
\section{Introduction}
Recall that for a profinite group $G$ and a prime number $p$, the Zassenhaus ($p$-)filtration $(G_{(n)})$ of $G$ is defined inductively by
\[
G_{(1)}=G, \quad G_{(n)}=G_{(\lceil n/p\rceil)}^p\prod_{i+j=n}[G_{(i)},G_{(j)}],
\]
where $\lceil n/p \rceil$ is the least integer which is greater than or equal to $n/p$. (Here for two closed subgroups $H$ and $K$ of $G$, $[H,K]$ means the smallest closed subgroup of $G$ containing the commutators $[x,y]=x^{-1}y^{-1}xy$, $x\in H, y\in K$. Similarly, $H^p$ means the smallest closed subgroup of $G$ containing the $p$-th powers $x^p$, $x\in H$.) Zassenhaus filtrations of groups were introduced in \cite{Zas} and are now recognized as being of fundamental importance in determining the structure and properties of various types of groups. For example, in the case of absolute Galois groups, these filtrations and their subquotients have recently been investigated in \cite{CEM, Ef1,Ef2, EM,MT, EMT}. In the case of arbitrary groups, this filtration has also been referred to as the dimension series, with the subgroups $G_{(n)}$ being called the dimension subgroups in characteristic $p$ (see \cite[Chapters 11, 12]{DDMS}). Our goal is to develop a method for determining the ${\mathbb F}_p$-dimension of subquotients of the Zassenhaus filtration in the case of finitely generated pro-$p$ groups.
Let $G$ be a finitely generated pro-$p$-group. For each $n\geq 1$, we set
\[
c_n(G) =\dim_{{\mathbb F}_p}(G_{(n)}/G_{(n+1)}).
\]
Note that since $G$ is finitely generated, $c_n(G)$ is finite for every $n\geq 1$ (see Section 2 for more details).
We will proceed to derive an explicit formula for $c_n(G)$ for various families of groups $G$, including finitely generated free pro-$p$-groups, Demushkin groups, and free pro-2 products of finitely many copies of the cyclic group $C_2$ of order 2. Galois theory provides much of the underlying motivation, as many of these groups are realizable as Galois groups of maximal $p$-extensions of local fields (see \cite{De1}, \cite{Sha}) and other fields (see \cite[Proposition 1.3]{EH}).
Shafarevich \cite{Sha} demonstrated that for certain fields $F$ not containing primitive $p$-th roots of unity, one could show that the Galois group of the maximal $p$-extension of $F$ was a free pro-$p$-group simply by determining the cardinality of some of its filtration quotients.
In Remarks~\ref{rmks:characterization} (1) we show that the numbers $c_n(G)$, $n=1,2,\ldots$, are sufficient to determine finitely generated free pro-$p$-groups in the family of all finitely generated pro-$p$-groups.
In Remarks~\ref{rmks:characterization} (2) we are able to determine finitely generated free pro-p groups in the family of all Galois groups of the maximal $p$-extensions of fields containing a primitive $p$-th root of unity by just two numbers, $c_1(G)$ and $c_2(G)$.
In Remarks~\ref{rmks:SAP} we show that $c_1(G)$ and $c_2(G)$ are sufficient to determine the Galois groups of the maximal $2$-extensions of Pythagorean fields in two significant cases.
In Subsection 4.2 we study groups $G$ which are the free products of several copies of the cyclic group of order 2 in the category of pro-$2$-groups. These groups can be realized as the Galois groups of the maximal 2-extensions of Pythagorean SAP fields, and therefore they are significant in Galois theory. Each such group G contains a free pro-2-subgroup $H$ of index 2. In Corollary~\ref{cor:quotient} we are able to use knowledge of the numbers $c_n(G)$ and $c_n(H)$,
to obtain the interesting relation $H_{(n)} = H\cap G_{(n)}$ for each $n\geq 2$. This is yet another example illustrating that the numbers $c_n(G)$ can be very useful in group theory and Galois theory.
In this paper we provide a unifying principle for deriving the dimensions $c_n(G)$ in a number of interesting cases. We observe that the formulas obtained look simple, elegant, and potentially useful.
We would also like to note that when $S$ is a finitely generated free pro-$p$ group, a formula for $c_n(S)$ is implicitly given in \cite{Ga}, where an ${\mathbb F}_p$-basis for $S_{(n)}/S_{(n+1)}$ is provided.
When we interpret the groups $G$ considered in this paper as Galois groups, our formulas lead to formulas for the order of related Galois groups. For example, if $G$ is isomorphic to the maximal pro-$p$-quotient $G_F(p)$ of the absolute Galois group $G_F$ of a field $F$, and if we denote by $F_{(n)}$ the fixed field of $G_F(p)_{(n)}$, then $|{\rm Gal}(F_{(n)}/F)|=p^{\sum_{i=1}^{n-1} c_i(G)}$. As indicated above, these Galois groups play a fundamental role in current Galois theory. Furthermore, we observe in Sections 3 and 5 that our formulas also lead to the determination of the minimal number of topological generators of $G_{(n)}$, for $G$ a free pro-$p$-group or a Demushkin pro-$p$-group.
In fact the orders of Galois groups ${\rm Gal}(F_{(n)}/F)$ are of considerable interest in current Galois theory research. In particular, in \cite{Ef1,MT, EMT}, based partially on the special cases in \cite{EM,MS2}, the Kernel Unipotent Conjecture was formulated. If this conjecture is true, we would obtain a characterization of $G_F(p)_{(n)}$, where $n\geq 3,$ as the intersection of the kernels of all Galois representations $\rho\colon G_F(p)\to {\mathbb U}_n({\mathbb F}_p)$. In order to prove the Kernel Unipotent Conjecture in the case when $G_F(p)$ is finitely generated, one may try to produce enough such representations. However, in order to check whether the intersection of the kernels of given representations is in fact $G_F(p)_{(n)}$, it would be useful to know $|{\rm Gal}(F_{(n)}/F)|$. This strategy resembles the previous successful strategy of Shafarevich, mentioned above. Another very interesting project in current Galois theory is to study the possible Koszul duality relating the Galois cohomology algebra $H^*(G,{\mathbb F}_p)$ to the Lie algebra $\bigoplus_{n=1}^\infty G_{(n)}/G_{(n+1)}$ and its universal enveloping algebra. In order to check some corollaries of this possible Koszul duality, determination of the numbers $c_n(G)$ could play an important role.
The structure of our paper is as follows: In Section 2 we discuss Hilbert-Poincar\'e series and provide a general formula for $c_n(G)$ (see Theorem~\ref{thm:general}). In Section 3 we provide an explicit formula for $c_n(G)$ when $G$ is a free pro-$p$-group of finite rank. In Section 4 we treat the case when $G$ is a free pro-2 product of finite copies of $C_2$.
We also show that in some significant special cases, knowledge of just $c_1(G)$ and $c_2(G)$ is sufficient to determine certain Galois groups within large families of
pro-$2$-groups.
In Section 5 we treat the case in which $G$ is a Demushkin group. In the last section we discuss some other groups.
\\
\\
{\bf Acknowledgements: }
The first author gratefully acknowledges discussions with I. Efrat, J. Labute and A. Topaz; the latter two having provided some extra motivation for the work in this paper. All of the authors would like to thank the referee for valuable suggestions related to the exposition.
\section{Hilbert-Poincar\'e series}
Let $F$ be a unital commutative ring. A graded free $F$-module $V=\bigoplus_{i=0}^\infty V_n$ is called {\it locally finite} if ${\rm rank}_F(V_n)<\infty$ for all $n\geq 0$. For such a graded free $F$-module $V$, the {\it Hilbert-Poincar\'e series} $P_V(t)\in {\mathbb Z}[[t]]$ of $V$ is the formal power series
\[
P_V(t)=\sum_{n=0}^\infty {\rm rank}_F(V_n) t^n.
\]
Let $G$ be a finitely generated pro-$p$-group. It is convenient to use the completed group algebra ${\mathbb F}_p[[G]]$ of $G$ over ${\mathbb F}_p$
\[
{\mathbb F}_p[[G]]:= \varprojlim_{N} {\mathbb F}_p[G/N].
\]
Thus ${\mathbb F}_p[[G]]$ is the topological inverse limit of the usual group rings ${\mathbb F}_p[G/N]$, where $N$ runs through open normal subgroups of $G$. A standard reference for completed group rings is \cite[Chapter 5]{NSW}. We also use the convenient references \cite[Chapter 7]{Ko} and \cite[Chapters 7 and 12]{DDMS}. We recall that $I(G)\subset {\mathbb F}_p[[G]]$ denotes the augmentation ideal of ${\mathbb F}_p[[G]]$ which is the closed two-sided ideal of ${\mathbb F}_p[[G]]$ generated by the elements $g-1$, for $g\in G$. Thus if $\epsilon\colon {\mathbb F}_p[[G]]\to {\mathbb F}_p$ is the homomorphism such that $\epsilon(g)=1$ for all $g\in G$, then $I(G)=\ker\epsilon$.
We denote by $I^n(G)$ the $n$-th power of the augmentation ideal $I(G)$.
There are two graded ${\mathbb F}_p$-algebras associated to $G$ and ${\mathbb F}_p[[G]]$ respectively, which are defined by
\[
{\rm gr}(G)= \bigoplus_{n\geq 1} G_{(n)}/G_{(n+1)} \quad \text{ and } \quad {\rm gr}({\mathbb F}_p[[G]]) =\bigoplus_{n\geq 0} I^n(G)/I^{n+1}(G).
\]
Then ${\rm gr}(G)$ is a restricted Lie algebra. (See \cite[Chapter 12]{DDMS}.) Furthermore since $G$ is finitely generated, the two graded algebras ${\rm gr}(G)$ and ${\rm gr}({\mathbb F}_p[[G]])$ are locally finite (see \cite[Section 7.4]{Ko}). We recall that $c_n(G)=\dim_{{\mathbb F}_p} G_{(n)}/G_{(n+1)}$ and we let $a_n(G):= \dim_{{\mathbb F}_p} I^n(G)/I^{n+1}(G)$.
As pointed out in \cite[page 312]{DDMS}, $a_n(G)=\dim_{{\mathbb F}_p} I_0^n(G)/I_0^{n+1}(G)$, where $I_0(G)$ is the augmentation ideal of ${\mathbb F}_p[G]$ - the usual group ring of $G$. Thus our results below apply to this case as well. In several places we use results from discrete groups which extend in a straightforward way to pro-$p$-groups. We usually mention this, but in some cases we omit explicitly mentioning such a standard extension.
The following theorem, Theorem~\ref{thm:JL}, is a consequence of a beautiful theory of Jennings and Lazard \cite[Chapters 11 and 12]{DDMS} viewing the Zassenhaus filtration subgroups $G_{(n)}$ as dimension subgroups. (See also \cite{Qu}.)
\begin{thm}[Jennings-Lazard]
\label{thm:JL}
Let the notation be as above.
\begin{enumerate}
\item The graded algebra ${\rm gr}({\mathbb F}_p[[G]])$ is a restricted universal enveloping algebra of ${\rm gr}(G)$.
\item We have
\begin{equation}
\label{eq:fundamental} P_{{\rm gr}({\mathbb F}_p[[G]])}(t) =
\sum_{n=0}^\infty a_n(G) t^n=\prod_{n=1}^\infty \left(\frac{1-t^{np}}{1-t^n}\right)^{c_n(G)}.
\end{equation}
\end{enumerate}
\end{thm}
\begin{proof} (1) See \cite[Theorem 12.8]{DDMS}.
(2) See \cite[Theorem 12.16]{DDMS} (see also \cite[Proposition 2.3]{Er}).
\end{proof}
The following lemma is an important technical tool which allows us to derive our results in a concise way. It relies on one fundamental result of Lichtman and also on a simple, but quite remarkable formula which can be traced back to the work of Lemaire in \cite[Chapter 5]{Le}.
\begin{lem}
\label{lem:free product}
Let $G_1$ and $G_2$ be two finitely generated pro-$p$-groups. Let $G=G_1*G_2$ be the free product of $G_1$ and $G_2$ in the category of pro-$p$-groups. Then
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t) = (P_{{\rm gr}({\mathbb F}_p[[G_1]])}^{-1}(t)+P_{{\rm gr}({\mathbb F}_p[[G_2]])}^{-1}(t)-1)^{-1}.
\]
\end{lem}
\begin{proof} By \cite[Theorem 1]{Li}, the graded ${\mathbb F}_p$-algebra ${\rm gr}({\mathbb F}_p[[G]])$ is a free product (i.e., a categorical coproduct)
of ${\rm gr}({\mathbb F}_p[[G_1]])$ and ${\rm gr}({\mathbb F}_p[[G_2]])$. The statement then follows from \cite[Equation (1.2), page 56]{PP}.
\end{proof}
\begin{rmk}
\label{rmk:direct product}
Let $G=G_1\times G_2$ be the direct product of two finitely generated pro-$p$-groups $G_1$ and $G_2$. We first observe that every commutator in $G$ is the product of a commutator in $G_1$ and a commutator in $G_2$, and every $p$-power in $G$ is the product of a $p$-power in $G_1$ and a $p$-power in $G_2$. Then we can check that $G_{(n)}=(G_1)_{(n)}\times (G_2)_{(n)}$, and that
\[\frac{G_{(n)}}{G_{(n+1)}}\simeq\frac{(G_1)_{(n)}}{(G_1)_{(n+1)}}\times \frac{(G_2)_{(n)}}{(G_2)_{(n+1)}}.\]
This implies that $c_{n}(G)=c_{n}(G_1)+c_{n}(G_2)$, and that
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t) = P_{{\rm gr}({\mathbb F}_p[[G_1]])}(t)\cdot P_{{\rm gr}({\mathbb F}_p[[G_2]])}(t).
\]
In fact, since ${\rm gr}(G)\simeq {\rm gr}(G_1) \oplus {\rm gr}(G_2)$, one can show that
\[{\rm gr}({\mathbb F}_p[[G]])\simeq {\rm gr}({\mathbb F}_p[[G_1]]) \otimes {\rm gr}({\mathbb F}_p[[G_2]]).\]
\end{rmk}
\begin{examples}
\mbox{}
In our examples below, $d$ can be any natural number, and in (3), $d=0$ is also meaningful.
\begin{enumerate}
\item If $G$ is a free pro-$p$-group of rank $d$, then (see Section 3)
\[ P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \frac{1}{1-dt}.
\]
\item If $G=C_p$ is a cyclic group of order $p$, then
\[ P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= 1+t+\cdots+t^{p-1}.
\]
Indeed, since ${\rm gr}(C_p)=C_p$, the graded algebra ${\rm gr}({\mathbb F}_p[[G]])$, which is a universal enveloping algebra of ${\rm gr}(C_p)$ by Theorem~\ref{thm:JL}, is isomorphic to ${\mathbb F}_p[X]/(X^p)$, the preceding statement follows.
\\
\item If $G=C_p*\cdots*C_p$ is a free product of $d+1$ copies of $C_p$ the cyclic group of order $p$, then
\[ P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \frac{1+t+\cdots + t^{p-1},}{1-dt-\cdots - dt^{p-1}}.
\]
This follows by induction on $d$, and by using (2) above, and Lemma~\ref{lem:free product}.
\\
\item If $G={\mathbb Z}_p^d$, then
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\frac{1}{(1-t)^d}.
\]
This follows from Remark~\ref{rmk:direct product} and (1) above.
\\
\item If $G:={\mathbb Z}_2^d \rtimes C_2$, where $C_2=\langle x \rangle$ and the action of $C_2$ on ${\mathbb Z}_2^d$ is given by $xyx=y^{-1}$, for all $y\in {\mathbb Z}_2^d$, then (see Corollary \ref{cor:superPy})
\[
P_{{\rm gr}({\mathbb F}_2[[G]])}(t)=\frac{1+t}{(1-t)^d}\prod_{i=1}^{\infty}\frac{1}{1-t^{2i+1}}.
\]
\item If $G$ is a Demushkin pro-$p$-group of rank $d$, then (see Section 5)
\[ P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \frac{1}{1-dt+t^2}.
\]
\item If $G$ is a free product of a cyclic group of order 2 and a free pro-$2$-group of rank $d$, then
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \frac{1+t}{1-dt-dt^2}.
\]
This follows by using Lemma~\ref{lem:free product} and (1)-(2) above.
\qed
\end{enumerate}
\end{examples}
Below we shall describe a general method for deriving a formula for $c_n(G)$ if we know the Hilbert-Poincar\'e series $P_{{\rm gr}({\mathbb F}_p[[G]])}(t)$.
So we assume that we are given a power series $P(t)=1+\sum_{n\geq 1}a_nt^n\in {\mathbb Z}[[t]]$. We define $c_n,n=1,2,\ldots$ by
\[
P (t)=1+\sum_{n\geq 1}a_n t^n=\prod_{n=1}^\infty \left(\frac{1-t^{np}}{1-t^n}\right)^{c_n}.
\]
We write $\log P(t)= \sum_{n\geq 1} b_n t^n$.
We shall derive a formula for $c_n$ using the values $b_1,\ldots,b_n$. To do this, it is convenient to introduce a new auxiliary sequence $w_1,w_2,\ldots$ defined below.
Taking logarithms and using $\log(\dfrac{1}{1-t})=\sum \limits_{\nu=1}^\infty \dfrac{1}{\nu}t^\nu$, we obtain
\[
\sum_{n=1}^\infty b_n t^n =\sum_{m=1}^\infty c_m \sum_{\nu=1}^\infty \frac{1}{\nu} (t^{m\nu}-t^{mp\nu}).
\]
Equating the coefficients of $t^n$, we obtain
\[
b_n=\sum_{m\nu=n} \frac{1}{\nu}c_m - \sum_{mp\nu=n}\frac{1}{\nu}c_m.
\]
Hence
\[
nb_n= \sum_{m\mid n} m c_m -\sum_{mp\mid n} mp c_m.
\]
Recall that for two integers $a$ and $b$, the symbol $a\mid b$ means that $a$ divides $b$. Now we define the sequence $w_n, n=1,2,\ldots$ by
\[
w_n=\frac{1}{n}\sum_{m\mid n} \mu(n/m) mb_m.
\]
Then by the M{\"o}bius inversion formula,
\[
nb_n= \sum_{m\mid n} m w_m.
\]
Here $\mu$ is the M\"obius function: for a positive integer $d$,
\[
\mu(d)=
\begin{cases}
(-1)^r & \text{ if $d$ is a product of $r$ distinct prime numbers},\\
0 &\text{ otherwise}.
\end{cases}
\]
\begin{rmk}
\label{rmk:wn}
From the definition of $w_n$ we see that
\[
P (t)=1+\sum_{n\geq 1}a_n t^n=\prod_{n=1}^\infty \frac{1}{(1-t^n)^{w_n}}.
\]
\end{rmk}
\begin{lem}
\label{lem:coprime}
If $(n,p)=1$ then $c_n=w_n$.
\end{lem}
\begin{proof}
Assume that $(n,p)=1$. Then we have
\[nb_n=\sum_{m\mid n} mc_m.\]
Hence by the M{\"o}bius inversion formula, we have
\[
c_n=\frac{1}{n}\sum_{m\mid n} \mu(n/m) mb_m=w_n.
\qedhere
\]
\end{proof}
\begin{lem}
\label{lem:not coprime}
If $p$ divides $n$, then we have
\[
c_n = c_{n/p}+w_n.
\]
\end{lem}
\begin{proof}
We proceed by induction on $n$. Clearly $c_p-c_1=\dfrac{pb_p-b_1}{p}=w_p$, hence the statement is true for $n=p$. Therefore we assume now that $n>p$ and $p\mid n$. We assume that the statement is true for every $m$ such that $p\mid m\mid n$, $m\not=n$. We are going to show that the statement is also true for $n$.
We have
\[
\begin{aligned}
nb_n &=\sum_{m\mid n} m c_m -\sum_{pm\mid n} pm c_m\\
&=\sum_{m\mid n} m c_m -\sum_{p\mid m\mid n} m c_{m/p}\\
&= \sum_{m\mid n, (m,p)=1} m c_m +\sum_{p\mid m\mid n} m (c_m-c_{m/p})\\
&= \sum_{m\mid n, (m,p)=1} m w_m +\sum_{p\mid m\mid n,m\not= n} m w_m + n(c_n-c_{n/p})\\
&=\sum_{m\mid n, m\not= n} m w_m +n(c_n-c_{n/p}).
\end{aligned}
\]
Combining with
\[
nb_n=\sum_{m\mid n} mw_m,
\]
we obtain $c_n-c_{n/p}=w_n$.
\end{proof}
\begin{prop}
\label{prop:key}
If $n=p^k m $ with $(m,p)=1$, then
\[
c_n =w_m +w_{pm}+\cdots + w_{p^km}.
\]
\end{prop}
\begin{proof}
This follows from the above two lemmas.
\end{proof}
\begin{thm}
\label{thm:general}
Let $G$ be a finitely generated pro-$p$-group. We write
\[ \log P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\sum_{n\geq 1}b_nt^n\in {\mathbb Q}[[t]],\] and we define $w_n(G)$ by
\[
w_n(G):=\frac{1}{n}\sum_{m\mid n} \mu(n/m) mb_m.
\]
Let $n=p^k m $ with $(m,p)=1$. Then
\[
c_n(G) =w_m(G) +w_{pm}(G)+\cdots + w_{p^km}(G).
\]
\end{thm}
\begin{proof} This follows from Theorem~\ref{thm:JL} and Proposition~\ref{prop:key}.
\end{proof}
Let $G$ be a finitely generated pro-$p$-group. We write $\log P_{{\rm gr}({\mathbb F}_p[[G]])}(t) = \sum_{n\geq 1} b_n t^n$ and recall that we have defined $w_n(G)$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m) mb_m.
\]
At first glance the definition of $w_n$ may look a bit artificial. One may ask whether $w_n$ appears more naturally as the rank or dimension of some free finitely generated abelian group. Below we shall give a partial answer to this question.
Recall that for a profinite group $G$, the descending central series $(G_n)$ is defined inductively by
\[
G_1=G,\quad G_{n+1}=[G_n,G].
\]
Let $J(G)$ be the augmentation ideal of the completed group ring ${\mathbb Z}_p[[G]]$. (Here ${\mathbb Z}_p[[G]]$ and $J(G)$ are defined similarly to ${\mathbb F}_p[[G]]$ and $I(G)$.)
Then we have two graded ${\mathbb Z}_p$-algebras associated to $G$ and ${\mathbb Z}_p[[G]]$ respectively which are defined by
\[
{\rm gr}_\gamma(G)= \bigoplus_{n\geq 1} G_{n}/G_{n+1} \quad \text{ and } \quad {\rm gr}({\mathbb Z}_p[[G]]) =\bigoplus_{n\geq 0} J^n(G)/J^{n+1}(G).
\]
\begin{lem}
\label{lem:integral version}
Let $G$ be a finitely generated pro-$p$-group. Assume that the graded algebra ${\rm gr}_\gamma(G)=\bigoplus_{n\geq1} G_n/G_{n+1}$ is torsion free. Let $e_n(G)={{\rm rank}}_{{\mathbb Z}_p} G_n/G_{n+1}$.
\begin{enumerate}
\item[(a)] The graded algebra ${\rm gr}({\mathbb Z}_p[[G]])$ is a universal enveloping algebra of ${\rm gr}_\gamma(G)$.
\item[(b)] $J^n(G)/J^{n+1}(G)$ is a free module over ${\mathbb Z}_p$ of finite rank $d_n(G)$, and
\[
P_{{\rm gr}({\mathbb Z}_p[[G]])}(t) =
\sum_{n=0}^\infty d_n(G) t^n=\prod_{n=1}^\infty \frac{1}{(1-t^n)^{e_n(G)}}.
\]
\end{enumerate}
\end{lem}
\begin{proof} (a) This follows from \cite[Theorem 1.3]{Har}. In \cite[Theorem 1.3]{Har}, a discrete group $G$ is considered, but an adaptation of this proof to the profinite case is straightforward.
(b) This follows from (a) and \cite[Proposition 2.5]{La3}.
\end{proof}
\begin{prop}
\label{prop:wn}
Let $G$ be a finitely generated pro-$p$-group. Assume that the graded algebra ${\rm gr}_\gamma(G)=\bigoplus_{n\geq1} G_n/G_{n+1}$ is torsion free. The following are equivalent.
\begin{enumerate}
\item[(a)] ${\rm rank}_{{\mathbb Z}_p} J^n(G)/J^{n+1}(G)=\dim_{{\mathbb F}_p} I^n(G)/I^{n+1}(G)$ for all $n\geq 1$.
\item[(b)] $w_n(G)={\rm rank}_{{\mathbb Z}_p} G_n/G_{n+1}$ for all $n\geq 1$.
\end{enumerate}
\end{prop}
\begin{proof} We keep the existing notation as in Lemma~\ref{lem:integral version}.
(a) $\Rightarrow$ (b): Assume that ${\rm rank}_{{\mathbb Z}_p} J^n(G)/J^{n+1}(G)=\dim_{{\mathbb F}_p} I^n(G)/I^{n+1}(G)$ for all $n$. Then by Theorem~\ref{thm:JL}, Remark~\ref{rmk:wn} and Lemma~\ref{lem:integral version}, we have
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\prod_{n=1}^\infty \frac{1}{(1-t^n)^{w_n(G)}}= P_{{\rm gr}({\mathbb Z}_p[[G]])}(t)=\prod_{n=1}^\infty \frac{1}{(1-t^n)^{e_n(G)}}.
\]
Therefore $w_n(G)=e_n(G)$ for all $n\geq 1$.
(b) $\Rightarrow$ (a): Assume that $w_n(G)=e_n(G)$ for all $n\geq 1$. Then by Theorem~\ref{thm:JL}, Remark~\ref{rmk:wn} and Lemma~\ref{lem:integral version}, we have
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=P_{{\rm gr}({\mathbb Z}_p[[G]])}(t).
\]
Therefore ${\rm rank}_{{\mathbb Z}_p} J^n(G)/J^{n+1}(G)=\dim_{{\mathbb F}_p} I^n(G)/I^{n+1}(G)$ for all $n\geq 1$.
\end{proof}
\begin{rmk}
We shall see in Sections 3 and 5, Remark~\ref{rmk:wn free} and Lemma~\ref{lem:wn Demushkin}, that both a free finitely generated pro-$p$-group and a Demushkin group with a relation of the form $r=[x_1,x_2]\cdots [x_{d-1},x_d]$ satisfy the equivalent statements in Proposition~\ref{prop:wn}.
\end{rmk}
\begin{question}
Let $G$ be a finitely generated pro-$p$-group. We assume that the graded algebra $\bigoplus_{n\geq1} G_n/G_{n+1}$ is torsion free. Is this true that
\[{\rm rank}_{{\mathbb Z}_p}(G_n/G_{n+1})=w_n(G)?\]
\end{question}
\section{Free pro-$p$-groups}
Throughout this section we assume that $S$ is a free pro-$p$-group on a finite set of generators $x_1,\ldots,x_d$.
We recall the Magnus homomorphism from the completed group algebra ${\mathbb F}_p[[S]]$ to the ${\mathbb F}_p$-algebra ${\mathbb F}_p\langle\langle X_1,\ldots,X_d\rangle\rangle$ of the formal power series in $d$ non-commuting variables $X_1,\ldots,X_d$ over ${\mathbb F}_p$.
The homomorphism is given by
\[
\psi\colon {\mathbb F}_p[[S]] \to {\mathbb F}_p \langle\langle X_1,\ldots,X_d\rangle\rangle, x_i\mapsto 1+X_i.
\]
The ${\mathbb F}_p$-algebra ${\mathbb F}_p\langle\langle X_1,\ldots,X_d\rangle\rangle$ is equipped with a natural valuation $v$ given by
\[
v(\sum a_{i_1,\ldots,i_k}X_{i_1}\cdots X_{i_k})=\inf\{k\mid a_{i_1,\ldots,i_k}\not=0\}\in {\mathbb Z}_{\geq 0}\cup \{\infty\},
\]
making it a compact topological ${\mathbb F}_p$-algebra. One basic result is the following.
\begin{lem}
\label{lem:0a}
The Magnus homomorphism $\psi$ is a (topological) isomorphism.
\end{lem}
\begin{proof} See for example, \cite[Chapter I, Proposition 7]{Se} or \cite[Chapter 6]{Laz}.
\end{proof}
\begin{cor} The Hilbert-Poincar\'e series
\[
P_{{\rm gr}({\mathbb F}_p[[S]])}(t)=\frac{1}{1-dt}.
\]
\end{cor}
\begin{proof}
Via the Magnus homomorphism, the augmentation ideal $I(S)$ is mapped to the ideal $I=(X_1,\ldots,X_d)$ of ${\mathbb F}_p\langle\langle X_1,\ldots,X_d\rangle\rangle$. Hence
\[
a_n(S):= \dim_{{\mathbb F}_p}(I^n(S)/I^{n+1}(S))=\dim_{{\mathbb F}_p}(I^n/I^{n+1}),
\]
which is equal to the number of non-commutative monomials of degree $n$ in $d$ variables $X_1,\ldots,X_n$. Hence $a_n(S)=d^n$. The statement then follows.
\end{proof}
We define $w_n(S)$ by
\[w_n(S)=\frac{1}{n}\sum_{m\mid n} \mu(m) d^{n/m}.
\]
\begin{rmk}
\label{rmk:wn free}
Let $(S_n)$ be the lower central series of $S$.
Then by Witt's result, $S_{n}/S_{n+1}$ is a free ${\mathbb Z}_p$-module of finite rank $w_n(S)$.
\end{rmk}
Theorem~\ref{thm:general} immediately implies the following result.
\begin{prop}
\label{prop:cn free}
If $n=p^k m $ with $(m,p)=1$, then
\[
c_n(S) =w_m(S) +w_{pm}(S)+\cdots + w_{p^km}(S).
\qedhere
\]
\end{prop}
\begin{rmks}
\label{rmks:characterization}
(1) If a finitely generated pro-$p$-group $G$ has Hilbert-Poincar\'e series of a finitely generated free pro-$p$-group, then $G$ is itself free. In other words, if
$ P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\displaystyle \frac{1}{1-dt}, $
then $G$ is free of rank $d$. Indeed, we first have $c_1(G)=w_1(G)=d$, which is equal to the minimal number of topological generators of $G$. Hence there exists a minimal presentation of $G$:
\[
1 \to R\to S\to G\to 1,
\]
where $S$ is a free pro-$p$-group of rank $d$. We then have $c_n(G)=c_n(S)$ for all $n\geq 1$. Hence $|S/S_{(n)}|=|G/G_{(n)}|$ for all $n\geq 1$. Thus the natural epimorphism
\[
S/S_{(n)} \twoheadrightarrow G/G_{(n)}
\]
is in fact an isomorphism. This implies that $R\subseteq S_{(n)}$ for all $n\geq 1$. Therefore by \cite[Theorem 7.11]{Ko}, $R=1$ and hence $S\simeq G$.
(2) Let $G$ be a finitely generated pro-$p$-group. In the case in which $G$ is realizable as the Galois group of a maximal $p$-extension of a field $F$ containing a primitive $p$-th root of unity, it is noteworthy to point out that if $c_1(G)=c_1(S)$ and $c_2(G)=c_2(S)$ for some finitely generated free pro-$p$-group $S$, then $G$ is in fact isomorphic to $S$. Indeed, as $c_1(S)=c_1(G)$ we have a short exact sequence
\[
1 \to R\to S\stackrel{\pi}{\to} G\to 1.
\]
Since $c_1(S)=c_1(G)$ and $c_2(S)=c_2(G)$, we see that $|S/S_{(3)}|=|G/G_{(3)}|$. Thus the natural epimorphism
\[
S/S_{(3)} \twoheadrightarrow G/G_{(3)}
\]
is in fact an isomorphism. Hence by \cite[Theorem C]{EM} (see also \cite[Theorem D]{CEM} for the case $p=2$) we see that $\pi\colon S\to G$ is an isomorphism.
\end{rmks}
In Corollary~\ref{cor:unipotent} relying on Lemma~\ref{lem:coefficient} below, we obtain an interesting purely group-theoretical corollary of our formula for $c_n(S)$. For each positive integer $n$, let ${\mathbb U}_{n+1}({\mathbb F}_p)$ be the group of all upper-triangular unipotent $(n+1)\times(n+1)$-matrices with entries in ${\mathbb F}_p$.
\begin{lem}
\label{lem:coefficient}
Let $n$ be a positive integer. If ${\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}=1$, then $c_n(S)=0$ for every free pro-$p$-group $S$.
\end{lem}
\begin{proof} Let $S$ be a free pro-$p$-group.
Assume that ${\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}=1$. Then for any (continuous) representation $\rho: S\to {\mathbb U}_{n+1}({\mathbb F}_p)$, we have $\rho(S_{(n)})\subseteq {\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}=1$. Hence
\[
S_{(n+1)}\subseteq S_{(n)}\subseteq \bigcap \ker(\rho\colon S\to {\mathbb U}_{n+1}({\mathbb F}_p)),
\]
where $\rho$ runs over the set of all representations (continuous homomorphisms) $S \to {\mathbb U}_{n+1}({\mathbb F}_p)$. On the other hand, we know that the Kernel Unipotent Conjecture is true for $S$ (see \cite{Ef1}, and also \cite{Ef2}, \cite{MT}). This means that we have
\[
S_{(n+1)}=\bigcap \ker(\rho\colon S\to {\mathbb U}_{n+1}({\mathbb F}_p)).
\]
Therefore, $S_{(n+1)}=S_{(n)}$, i.e., $c_n(S)=0$, as desired.
\end{proof}
\begin{cor}
\label{cor:unipotent}
Let $n$ be a positive integer. Then
$ {\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}\simeq {\mathbb F}_p$ and
\[
n=\max\{h\mid {\mathbb U}_{n+1}({\mathbb F}_p)_{(h)}\not=1\}.
\]
\end{cor}
\begin{proof} We first observe that if $S$ is a free pro-$p$-group of rank $d>1$, then all numbers $w_n(S)$, $n=1,2,\ldots$, are positive. Therefore from Proposition~\ref{prop:cn free} we see that $c_n(S)\not=0$ for all $n\in {\mathbb N}$. Hence by Lemma~\ref{lem:coefficient}, ${\mathbb U}_{(n+1)}({\mathbb F}_p)_{(n)}\not=1$.
On the other hand, it is well-known that ${\mathbb U}_{n+1}({\mathbb F}_p)_{(n+1)}=1$. Hence ${\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}\subseteq Z({\mathbb U}_{n+1}({\mathbb F}_p))\simeq {\mathbb F}_p$, where $Z({\mathbb U}_{n+1}({\mathbb F}_p))$ is the center of ${\mathbb U}_{n+1}({\mathbb F}_p)$. Therefore
\[
{\mathbb U}_{n+1}({\mathbb F}_p)_{(n)}=Z({\mathbb U}_{n+1}({\mathbb F}_p))\simeq {\mathbb F}_p,
\]
and the second assertion is also clear.
\end{proof}
\begin{ex} Let $S$ be a free pro-$p$-group of finite rank $d$. We have
\[
\begin{aligned}
c_1(S)&=d,\\
c_2(S)&= \begin{cases}
\frac{d^2-d}{2} \text{ if } p\not=2,\\
\frac{d^2+d}{2} \text{ if } p=2,
\end{cases}\\
c_3(S) &=\begin{cases}
\frac{d^3-d}{3} \text{ if } p\not=3,\\
\frac{d^3+2d}{3} \text{ if } p=3,
\end{cases}\\
c_4(S)&= \begin{cases}
\frac{d^4-d^2}{4} \text{ if } p\not=2,\\
\frac{d^4+d^2+2d}{4} \text{ if } p=2,
\end{cases}\\
c_5(S)&=\begin{cases}
\frac{d^5-d}{5} \text{ if } p\not=5,\\
\frac{d^5+4d}{5} \text{ if } p=5.
\end{cases}
\end{aligned}
\]
Observe that our numbers $c_n(S)$, $n=1,2,\ldots$, also detect the minimal numbers of generators of $S_{(n)}$. Indeed by the pro-$p$-version of Schreier's formula for each open subgroup $T$ of $S$, we have the following expression for the minimal number of generators $d(T)$ of $T$:
\[
d(T)=[S:T](d(S)-1)+1.
\]
Therefore
\[
d_n(S):=d(S_{(n)})= p^{\sum_{i=1}^{n-1}c_i(S)}(d-1)+1.
\]
\end{ex}
Next we are going to give an explicit ${\mathbb F}_p$-basis for $S_{(n)}/S_{(n+1)}$, for each $n$. We shall first recall a definition of Hall commutators of weight $n$ and their linear order. This was originally introduced by M. Hall in \cite[Section 4]{Ha} (see also \cite[Definition 2.3]{Vo}).
\begin{defn} The set $C_n$ of {\it Hall commutators of weight $n$} together with a total order $<$ is inductively defined as follows:
\begin{enumerate}
\item $C_1=\{x_1,\ldots,x_d\}$ with the ordering $x_1>\cdots >x_d$.
\item Assume $n>1$ and that we have defined Hall commutators and their ordering for all weights $<n$.
Then $C_n$ is the set of all commutators $[c_1,c_2]$ where $c_1\in C_{n_1},c_2\in C_{n_2}$ such that $n_1+n_2=n$, $c_1>c_2$ and if $c_1=[c_3,c_4]$ then we also require that $c_2\geq c_4$. The set $C_n$ is ordered lexicographically, i.e., $[c_1,c_2]<[c_1^\prime,c_2^\prime]$ if and only if $c_1<c_1^\prime$, or $c_1=c_1^\prime$ and $c_2<c_2^\prime$. Finally commutators of weight $n$ are greater than all commutators of smaller weight.
\end{enumerate}
\end{defn}
The following theorem was proved by M. Hall in the discrete case. The extension of his theorem to the pro-$p$ case is straightforward.
\begin{thm}[{\cite[Theorem 4.1]{Ha}}] The Hall commutators of weight $n$ represent a basis of $S_n/S_{n+1}$ as a free ${\mathbb Z}_p$-module.
In particular $w_n(S)= |C_n|$.
\end{thm}
The following theorem is due to Lazard (see \cite[Theorem 11.2]{DDMS}).
\begin{thm}[Lazard]
For each $n$, one has
\[
G_{(n)}=\prod_{ip^j\geq n}G_i^{p^j}.
\]
\end{thm}
\begin{cor}
Let us write $n=p^km$ with $(m,p)=1$. Then a basis of the ${\mathbb F}_p$-vector space $S_{(n)}/S_{(n+1)}$ can be represented by the following set
\[
C_m^{p^k} \bigsqcup C_{pm}^{p^{k-1}}\bigsqcup \cdots \bigsqcup C_{p^{k-1}m}^p\bigsqcup C_n.
\]
\end{cor}
\begin{proof}
By Lazard's theorem, we can check that the above set defines a set of generators for the ${\mathbb F}_p$-vector space $S_{(n)}/S_{(n+1)}$. Now by Proposition~\ref{prop:cn free} and by a counting argument, we see that this set defines a basis for the ${\mathbb F}_p$-vector space $S_{(n)}/S_{(n+1)}$.
\end{proof}
\section{Free products of a finite number of cyclic groups of order 2}
\subsection{Free products of finitely many cyclic groups of order 2} Let $d$ be a non-negative integer. Let $G=C_2*\cdots * C_2$ be a free product in the category of pro-$2$-groups of $d+1$ copies of $C_2$, where $C_2$ is the group of order 2.
In this section we shall consider Pythagorean fields. A field $F$ is said to be Pythagorean if each finite sum of squares in $F$ is again a square in $F$. A Pythagorean field is called a formally real field if $-1$ is not a square. Pythagorean fields play an important role in Galois theory, real algebraic geometry and the algebraic theory of quadratic forms. We refer a reader to a beautiful exposition of related topics in \cite{Lam}.
Let us recall that a formally real Pythagorean field $K$ with $|K^\times/(K^\times)^2|=2^{d+1}$ is called an SAP field if $K$ admits exactly $d+1$ orderings. These SAP fields form an interesting and well investigated family of fields. (See \cite[Chapter 17]{Lam}.)
\begin{thm}
\label{thm:SAP}
Let $F$ be a field with $|F^\times/(F^\times)^2|=2^{d+1}$. Then $F$ is an SAP field if and only if $G_F(2)$ is isomorphic to $G=C_2*\cdots *C_2$, the free product of $d+1$ copies of $C_2$.
\end{thm}
\begin{proof}
The "only if" part follows from \cite{Mi}.
We now prove the "if" part. Suppose that $G_F(2)$ is isomorphic to $C_2*\cdots*C_2$ ($d+1$ copies of $C_2$). Then $F$ is formally real Pythagorean, and
$|F^\times/(F^\times)^2|=2^{d+1}$. Now we pick any SAP field $K$ which has exactly $d+1$ orderings. From the "only if" part, we see that $G_F(2)\simeq G_K(2)$. In particular, $G_F(2)/G_F(2)_{(3)}\simeq G_K(2)/G_K(2)_{(3)}$. Then \cite[Theorem 3.8]{MS2} implies that the Witt ring $WF$ of $F$ is isomorphic to the Witt ring $WK$ of $K$. We know that for a Pythagorean field $L$, the Witt ring $WL$ of $L$ determines the space of orderings $X_L$ of $L$. Hence the space of orderings of $F$ is isomorphic to the space of orderings of $K$. In particular $F$ admits exactly $d+1$ orderings. Therefore $F$ is an SAP field.
\end{proof}
\begin{cor}
\label{cor:SAP}
Let $F$ be any Pythagorean field, and let $K$ be an SAP field. Assume that $|F^\times/(F^\times)^2|=|K^\times/(K^\times)^2|=2^{d+1}$ . Then there exists an epimorphism $G_K(2)\simeq C_2*\cdots*C_2 \twoheadrightarrow G_F(2)$.
\end{cor}
\begin{proof}
By \cite[Remark 17.5]{Lam}, $F$ has at least $d+1$ orderings, and we can choose $d+1$ involutions $\sigma_1,\ldots,\sigma_{d+1}$ in $G=G_F(2)$ such that $\sigma_1,\ldots,\sigma_{d+1}$ minimally generate $G_F(2)$. The statement then follows from the previous theorem.
\end{proof}
Our treatment below is purely group-theoretical. However the group $G$ plays an important role as the maximal pro-$2$ quotient of the absolute Galois group of SAP fields.
(We refer the interested reader to \cite{Haran}, \cite{Mi}.) Also it is interesting to observe that if $G=C_2*\cdots *C_2$ is the Galois group as above, then $G$ is already determined by its quotient $G/G_{(3)}$. (See also Remarks~\ref{rmks:SAP} for closely related observations.)
More precisely, assume that $H$ is another pro-2-group which is realizable as the Galois group of the maximal 2-extension of a field $F$, and that $H/H_{(3)}\simeq G/G_{(3)}$, then $H\simeq G$. (See \cite{MS1,MS2,Mi}.)
The Hilbert-Poincar\'e series of ${\rm gr}({\mathbb F}_2[[G]])$ is
\[
P_{{\rm gr}({\mathbb F}_2[[G]])}(t)= \frac{1+t}{1-dt}.
\]
\begin{rmk} Since the cohomology algebra $H^*(C_2,{\mathbb F}_2)$ is isomorphic to ${\mathbb F}_2[X]$, the polynomial algebra in one variable $X$ over ${\mathbb F}_2$, we have
\[
P_{H^*(G,{\mathbb F}_2)}(t)=\sum_{n=0}^\infty \dim_{{\mathbb F}_2}(H^n(G,{\mathbb F}_2)) t^n= 1+(d+1)t+(d+1)t^2+\cdots= \frac{1+dt}{1-t}
\]
by \cite[Theorems 4.1.4-4.1.5]{NSW}. Therefore, we have
\[
P_{H^*(G,{\mathbb F}_2)}(t) P_{{\rm gr}({\mathbb F}_2[[G]])}(-t)=1.
\]
This is not a coincidence. One can show that the cohomology algebra $H^*(G,{\mathbb F}_2)$ is Koszul and that ${\rm gr}({\mathbb F}_2[[G]])$ is its Koszul dual. The above equality is just a special case of the well-known relation between the two Hilbert-Poincar\'e series of a Koszul algebra and its dual (see \cite[Corollary 2.2]{PP}).
\qed
\end{rmk}
We have
\[
\log P_{{\rm gr}({\mathbb F}_2[[G]])}(t)= \log(\frac{1}{1-dt})-\log(\frac{1}{1+t})=\sum_{n\geq 1} \frac{1}{n} (d^n-(-1)^n)t^n.
\]
Now we define the sequence $w_n(G), n=1,2,\ldots$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m) (d^m-(-1)^m).
\]
\begin{prop}
\label{prop:cn free product of C2}
If $n=2^k m $ with $(m,2)=1$, then
\[
c_n(G) =w_m(G) +w_{2m}(G)+\cdots + w_{2^km}(G).
\qed
\]
\end{prop}
\subsection{Free products of the cyclic group of order 2 as semidirect products} Let $G=C_2*\cdots * C_2$ be a free product in the category of pro-$2$-groups of $d+1$ copies of $C_2$ In this subsection we shall show that $G$ is isomorphic to a semidirect product $H\rtimes C_2$ of a free pro-$2$-group $H$ and $C_2$. We also provide provide a relation between $G_{(n)}$ and $H_{(n)}$.
We define the numbers $\epsilon_n$, $n=1,2,\ldots$ by
\[
\epsilon_n=\frac{1}{n}\sum_{m\mid n} \mu(n/m)(-1)^m,
\]
i.e., by the M{\"o}bius inversion formula,
\[
\label{eq:epsilon}
\tag{*}(-1)^n= \sum_{m\mid n} m \epsilon_m.
\]
\begin{lem}
\label{lem:epsilon}
We have $\epsilon_1=-1$, $\epsilon_2=1$ and $\epsilon_n=0$ for $n\geq 3$.
\end{lem}
\begin{proof} The equation (\ref{eq:epsilon}) determines $\epsilon_n$, $n\in {\mathbb N}$, uniquely. But $\epsilon_1=-1$, $\epsilon_2=1$ and $\epsilon_n=0$ for $n\geq 3$ work as for these numbers
\[
\sum_{m\mid n} m \epsilon_m
=\begin{cases}
-1 &\;\text{if $n$ is odd},\\
-1+2=1 &\; \text{if $n$ is even}.
\qedhere
\end{cases}
\]
\end{proof}
Let us write
\[G=C_2*C_2*\cdots *C_2=\langle x_0\mid x_0^2\rangle*\langle x_1\mid x_1^2\rangle *\cdots * \langle x_d\mid x_d^2\rangle.\]
For ease of notation, we consider $x_0,x_1,\ldots, x_d$ as elements of $G$. We consider a continuous homomorphism $\varphi:G\to C_2= \langle x\mid x^2\rangle$ defined by $x_i\mapsto x$ for all $i=0,1,\ldots,d$. For each $i=1,\ldots,d$, we set $y_i=x_0x_i\in G$ and let $H$ be the closed subgroup of $G$ generated by $y_1,\ldots,y_d$.
\begin{lem} Let the notation be as above.
\begin{enumerate}
\item $\ker\varphi=H$.
\item $H$ is a free pro-$2$-group of rank $d$.
\item We have $G\simeq H\rtimes C_2$, where the action of $C_2$ on $H$ is given by $xy_ix=y_i^{-1}$.
\end{enumerate}
\end{lem}
\begin{proof}
(1) Clearly $y_i\in \ker\varphi$, hence $H\subseteq \ker\varphi$. Now consider any element $\gamma\in \ker\varphi$. Then for each open neighborhood $U$ of $\gamma$ in $G$, there exists an element $g=x_{i_1}\cdots x_{i_r}\in U$, $i_1,\ldots,i_r\in \{1,\ldots,d\}$ such that $1=\varphi(g)=x^r$. Hence $r=2s$ is even. Since $x_0 y_ix_0= y_i^{-1}$, we obtain
\[
g= x_0 y_{i_1}\cdots x_0 y_{i_r}=y_{i_1}^{-1}y_{i_2}\cdots y_{i_{r-1}}^{-1}y_{i_r}.
\]
Thus $g\in H$. Therefore $\gamma\in H$ and $H=\ker\varphi$.
(2) By profinite analogue of the well known Kurosch's subgroup theorem in the theory of free products of discrete groups due to E. Binz, J. Neukirch and G. Wenzel explained in \cite[Theorem 4.2.1 and Remarks below this Theorem]{NSW}, we see that $H$ is indeed a free pro-$2$-group of rank $d$.
(3) This follows by observing that $\psi\colon C_2=\langle x\mid x^2 \rangle \to G$ which maps $\bar{x}$ to $x_1$, is a section of $\varphi$.
\end{proof}
The following proposition and corollary are remarkable properties of the pair $\{H,G\}$.
\begin{prop}
\label{prop:comparison}
We have $c_1(H)=d=c_1(G)-1$ and $c_n(H)=c_n(G)$ for all $n\geq 2$.
\end{prop}
\begin{proof}
It is clear that $c_1(H)=w_1(H)=d$ and $c_1(G)=w_1(G)=d+1$. Hence $c_1(H)=d=c_1(G)-1$. We shall show that $c_n(H)=c_n(G)$ for any $n\geq 2$.
We note that
\[ w_n(H)-w_n(G)=\frac{1}{n}\sum_{m\mid d}\mu(n/m)(-1)^m =\epsilon_n.
\]
By Lemma~\ref{lem:epsilon}, one has $w_2(H)=w_2(G)+1$ and $w_n(H)=w_n(G)$ for every $n\geq 3$.
If $n>1$ is odd, then
\[
c_n(H)=w_n(H)=w_n(G)=c_n(G).
\]
If $n$ is even, then by writing $n=2^km$ with $m$ odd, we have
\[
\begin{aligned}
c_n(H)&= w_m(H)+w_{2m}(H) +w_{4m}(H)+\cdots+w_{2^km}(H)\\
&=w_m(G)+w_{2m}(G) +w_{4m}(G)+\cdots+w_{2^km}(G)=c_n(G).
\end{aligned}
\]
(Note that we always have $w_m(H)+w_{2m}(H)=w_m(G)+w_{2m}(G)$ for every $m\geq 1$ odd.)
\end{proof}
\begin{cor}
\label{cor:quotient}
Let $n\geq 2$ be an integer.
\begin{enumerate}
\item $H_{(n)}=H\cap G_{(n)}$.
\item $G/G_{(n)}\simeq H/H_{(n)}\rtimes C_2$, where the action of $C_2$ on $H/H_{(n)}$ is given by $\bar{x}\bar{y}_i\bar{x}=\bar{y}_i^{-1}$.
\end{enumerate}
\end{cor}
\begin{proof}
(1) Clearly $H_{(n)}\subseteq H\cap G_{(n)}$. We proceed by induction on $n$ to show that $H_{(n)}=H\cap G_{(n)}$. First consider the case $n=2$. We have an exact sequence
\[
1\to H/H\cap G_{(2)} \to G/G_{(2)}\to C_2\to 1.
\]
This implies that $[H: H\cap G_{(2)}]= [G:G_{(2)}]/2=2^d=[H:H_{(2)}]$. Hence $H_{(2)}=H\cap G_{(2)}$. Assume that $H_{(n)}=H\cap G_{(n)}$ for some $n\geq 2$. Then from the exact sequence
\[
1\to H/H\cap G_{(n)} \to G/G_{(n)}\to C_2\to 1,
\]
we obtain $[H:H_{(n)}]=[H:H\cap G_{(n)}]=[G:G_{(n)}]/2$. From a similar exact sequence we obtain
\[
\begin{aligned}
{[H:H\cap G_{(n+1)}]}&=\frac{1}{2}[G:G_{(n+1)}] = \frac{1}{2}[G:G_{n}] [G_{(n)}:G_{(n+1)}]\\
&= [H:H_{(n)}] [H_{(n)}:H_{(n+1)}]=[H:H_{(n+1)}].
\end{aligned}
\]
Here the equality $[G_{(n)}:G_{(n+1)}]=[H_{(n)}:H_{(n+1)}]$ follows from Proposition~\ref{prop:comparison}.
Therefore $H_{(n+1)}=H\cap G_{(n+1)}$.
(2) This follows from (1).
\end{proof}
\subsection{Another semidirect product}
In this subsection we consider an example in which $G$ is the semidirect product $G:={\mathbb Z}_2^d \rtimes C_2= H \rtimes \langle x \rangle$, where the action of $C_2$ on $H:={\mathbb Z}_2^d$ is given by $xyx=y^{-1}$, for all $y\in H$. This group $G$ is realizable as the maximal pro-$2$-quotient of the absolute Galois group of a superpythagorean field.
Recall that a formally real Pythagorean field $F$ with $|F^\times/(F^\times)^2|=2^{d+1}<\infty$ is called a superpythagorean field if $F$ admits exactly $2^d$ orderings.
\begin{prop}
Let $F$ be a Pythagorean field with $|F^\times/(F^\times)^2|=2^{d+1}$. Then there exists an epimorphism $G_F(2)\twoheadrightarrow G={\mathbb Z}_2^d\rtimes C_2$.
\end{prop}
\begin{proof}
We choose any ordering $P$ in $F$ and an ${\mathbb F}_2$-basis $[a_1],\ldots, [a_d]$ of $P/(F^\times)^2$. By \cite{Be1} we know that there exists a field $E$, the Euclidean closure of $F$ with respect to $P$ such that $F(2)=E(\sqrt{-1})$, $E$ is a formally real field and $(E^\times)^2\cap F^\times=P$. We can pick for each $a_i$ as above, a sequence
\[
\sqrt{a_i}, \sqrt[4]{a_i},\ldots, \sqrt[2^n]{a_i},\ldots,
\]
such that all $\sqrt[2^n]{a_i}$ are in $E^\times$. Indeed, by induction on $n$ we may assume that $\sqrt[2^n]{a_i}$ is in $E^\times$. Then we can pick $\sqrt[2^{n+1}]{a_i}$ in $(E^\times)^2$ because $E^\times=(E^\times)^2\cup -(E^\times)^2$. We set
\[
\tilde{M}:=\bigcup_{n=1}^\infty F(\sqrt[2^n]{a_1},\ldots,\sqrt[2^n]{a_d}).
\]
Then $\tilde{M}$ is formally real since $\tilde{M}$ is a subfield of $E$. We recall that for each $n\in {\mathbb N}$, $F(\sqrt{-1})$ contains a primitive $2^n$-th root of unity $\zeta_{2^n}$. (See \cite[Chapter II, Theorem 8]{Be}.) We may also assume that $\zeta_{2^{n+1}}^2=\zeta_{2^n}$.
We set $M:=\tilde{M}(\sqrt{-1})$. Then $M/F$ is a Galois extension.
We show that ${\rm Gal}(M/F{\sqrt{-1}})$ is isomorphic to ${\mathbb Z}_2^d$. This follows from Kummer theory. In fact, let $\tau_1, \ldots, \tau_d$ be elements in ${\rm Gal}(M/F(\sqrt{-1})$ such that for each $i=1,\ldots,d$, one has
\[
\tau_i(\sqrt[2^n]{a_i})=\zeta_{2^n} \sqrt[2^n]{a_i}, \; \text { and }\tau_i(\sqrt[2^n]{a_j})=\sqrt[2^n]{a_j}\; \forall j\not=i.
\]
Then ${\rm Gal}(M/F(\sqrt{-1}))=(\prod_{i=1}^d\langle \tau_i\rangle) \simeq {\mathbb Z}_2^d$.
We observe that the restriction of a nontrivial element of ${\rm Gal}(E(\sqrt{-1})/E)$ to $M$ gives us a nontrivial element $\sigma\in {\rm Gal}(M/\tilde{M})$. Thus we have a spliting
\[
{\rm Gal}(M/F)\simeq {\rm Gal}(M/F{\sqrt{-1}})\rtimes \langle \sigma \rangle,
\]
where $ \langle \sigma \rangle\simeq C_2$, and the action of $C_2$ on ${\rm Gal}(M/F{\sqrt{-1}})$ is by involution.
The natural projection
\[
G_F(2)={\rm Gal}(F(2)/F)\to {\rm Gal}(M/F)\simeq {\mathbb Z}_2^d\rtimes C_2
\]
gives the desired epimorphism.
\end{proof}
\begin{cor}
\label{cor:superPytha}
Let $F$ be a a field with $|F^\times/(F^\times)^2|=2^{d+1}$. Then $F$ is a superpythagorean field if and only if $G_F(2)$ is isomorphic to the group $G={\mathbb Z}_2^d \rtimes C_2$.
\end{cor}
\begin{proof}
Assume that $F$ is a superpythagorean field with $|F^\times/(F^\times)^2|=2^{d+1}$. Let the notation be as in the previous proposition. Then ${\rm Gal}(M/F)\simeq G={\mathbb Z}_2^d \rtimes C_2$. On the other hand, from \cite[Example 3.8, (ii)]{Wa} (see also \cite[Chapter III, Theorem 1]{Be}), we know that ${\rm Gal}(M/F)$ is equal to $G_F(2)$. Hence $G_F(2)\simeq {\mathbb Z}_2^d \rtimes C_2$.
The converse direction is proved in a similar fashion to the proof of the "if" part in Theorem~\ref{thm:SAP}.
\end{proof}
\begin{cor}
\label{cor:superPy}
Let $F$ be any Pythagorean field, and let $K$ be a superpythagorean field. Assume that $|F^\times/(F^\times)^2|=|K^\times/(K^\times)^2|=2^{d+1}$ . Then there exists an epimorphism $G_F(2) \twoheadrightarrow G_K(2)\simeq {\mathbb Z}_2^d\rtimes C_2$.
\end{cor}
\begin{proof}
This follows from the previous theorem and corollary.
\end{proof}
\begin{lem} Let $G=H\rtimes \langle x \rangle={\mathbb Z}_2^d\rtimes C_2$ be as above. Let $n\geq 2$ be an integer, and let $s=\lceil \log_2n \rceil$. Then $G_{(n)}=H^{2^s}$.
\end{lem}
\begin{proof} We proceed by induction on $n$. We first observe that $[y,x]=y^{-1}x^{-1} y x=(y^{-1})^2$ and $(yx)^2=y^2$, for every $y\in H$. Hence
\[
G_{(2)}=G^2[G,G]=G^2=H^2.
\]
The lemma is true for $n=2$. We assume that the lemma is true for $j$ with $2\leq j<n$. Then
\[
\begin{aligned}
G_{(n)}&=G_{(\lceil n/2 \rceil)}^2\prod_{i+j=n}[G_{(i)},G_{(j)}]\\
&=G_{(\lceil n/2\rceil)}^2[G,G_{(n-1)}]\\
&=(H^{2^{s-1}})^2=H^{2^s}.
\end{aligned}
\]
Here we use that $G_{n-1}\subseteq H^{2^{s-1}}$, and hence $[G,G_{(n-1)}]\subseteq H^{2^s}$.
\end{proof}
An immediate consequence of the above lemma is the following result.
\begin{cor}
\label{cor:cn superPy}
Let $n\geq 2$ be an integer. We have
\[
c_n(G)=
\begin{cases}
d+1 &\text{ if $n=1$},\\
d &\text{ if $n=2^s$ for some $1\leq s\in {\mathbb Z}$},\\
1 & \text{ if $n$ is not a power of 2}.
\end{cases}
\]
\end{cor}
\begin{cor}
\label{cor:superPy}
We have
\[
P_{{\rm gr}({\mathbb F}_2[[G]])}(t)=\frac{1+t}{(1-t)^d}\prod_{i=1}^{\infty}\frac{1}{1-t^{2i+1}}.
\]
\end{cor}
\begin{proof}
We write $\log P_{{\rm gr}({\mathbb F}_2[[G]])}= \sum_{n\geq 1} b_n(G) t^n $, and let
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m) mb_m(G).
\]
By Lemma~\ref{lem:coprime}, if $n$ is odd then $w_n(G)=c_n(G)$. In particular, $w_1(G)=c_1(G)=d+1$, and $w_{2i+1}(G)=c_{2i+1}(G)=1$ for $i\geq 1$.
By Lemma~\ref{lem:not coprime}, $w_2(G)=c_2(G)-c_1(G)=d-(d+1)=-1$.
We claim that $w_n(G)=0$ if $ n$ is even and $n\geq 4$. Indeed, if $n=2^s$ with $s\geq 2$, then by Lemma~\ref{lem:not coprime},
\[w_{2^s}(G)=c_{2^s}(G)-c_{2^{s-1}}(G)=d-d=0.\]
Now if $n=2m$, where $m$ is not a power of $2$, then also by Lemma~\ref{lem:not coprime},
\[w_{2m}(G)=c_{2m}(G)-c_{m}(G)=1-1=0.\]
The corollary then follows from Remark~\ref{rmk:wn}.
\end{proof}
\begin{rmks}
\label{rmks:SAP}
It is an interesting fact that $c_1(G)$ and $c_2(G)$ can be sufficient in determining $G$ itself within some large families of pro-$p$-groups. We mentioned an example in Remarks~\ref{rmks:characterization} (2). Here are two other instances.
(1) Suppose that $K$ is an SAP field with $|K^\times/(K^\times)^2|=2^{d+1}$. Then $G_K(2)=C_2*\cdots *C_2$, the free product of $d+1$ copies of $C_2$.
By Proposition~\ref{prop:cn free product of C2}, one has
\[
c_1(G_K(2))=d+1 \text { and } c_2(G_K(2))=\frac{d(d+1)}{2}.
\]
Now let $F$ be a formally real Pythagorean field $F$ with $|F^\times/(F^\times)^2|< \infty$. We assume that $c_1(G_F(2))=d+1$ and that $c_2(G_F(2))=d(d+1)/2$ for some integer $d\geq 0$. Then we claim that $F$ is an SAP field with exactly $d+1$ orderings. So, quite remarkably, within the family of formally real Pythagorean fields with finitely many square classes, the numbers $c_1(G_F(2))$ and $c_2(G_F(2))$ above suffice to characterize SAP fields $F$.
We shall now prove this claim. Because $c_1(G_F(2))=d+1$, we see that $G_F(2)$ has $d+1$ minimal generators, and therefore $|F^\times/(F^\times)^2|=2^{d+1}$.
We pick any SAP field $K$ with $|K^\times/(K^\times)^2|=2^{d+1}$.
By Corollary~\ref{cor:SAP}, there exists an epimorphism $\varphi\colon G_K(2)\twoheadrightarrow G_F(2)$.
We have
\[
\begin{aligned}
|G_K(2)/{G_K(2)}_{(3)}| &=c_1(G_K(2))+c_2(G_K(2))\\
&=d+\frac{d(d+1)}{2}\\
&=c_1(G_F(2))+c_2(G_F(2))\\
&=|G_F(2)/{G_F(2)}_{(3)}|.
\end{aligned}
\]
This implies that the induced epimorphism $G_K(2)/{G_K(2)}_{(3)}\twoheadrightarrow G_F(2)/{G_F(2)}_{(3)}$ is an isomorphism. By \cite[Theorem D]{CEM}, $\varphi \colon G_K(2)\to G_F(2)$ is an isomorphism. This implies that $F$ is a SAP field by Theorem~\ref{thm:SAP}.
(2) Suppose that $K$ is a superpythagorean field with $|K^\times/(K^\times)^2|=2^{d+1}<\infty$. By Corollary~\ref{cor:cn superPy}, one has
\[
c_1(G_K(2))=d+1 \; \text { and } c_2(G_K(2))=d.
\]
Now let $F$ be a formally real Pythagorean field $F$ with $|F^\times/(F^\times)^2|< \infty$. We assume that $c_1(G_F(2))=d+1$, $c_2(G_F(2))=d$ for some integer $d\geq 0$. Then we claim that $F$ is a superpythagorean field. So within the family of formally real Pythagorean fields with finitely many square classes, the numbers $c_1(G_F(2))$ and $c_2(G_F(2))$ above, also suffice to characterize superpythagorean fields $F$.
We shall now prove this claim. We pick any superpythagorean field $K$ with $|K^\times/(K^\times)^2|=2^{d+1}$. By Corollary~\ref{cor:superPy}, we have an epimorphism $\varphi\colon G_F(2)\twoheadrightarrow G_K(2)$.
We have
\[
\begin{aligned}
|G_F(2)/{G_F(2)}_{(3)}| &=c_1(G_F(2))+c_2(G_F(2))\\
&=d+1+d\\
&=c_1(G_K(2))+c_2(G_K(2))\\
&=|G_K(2)/{G_K(2)}_{(3)}|.
\end{aligned}
\]
This implies that the induced epimorphism $G_F(2)/{G_F(2)}_{(3)}\twoheadrightarrow G_K(2)/{G_K(2)}_{(3)}$ is an isomorphism. By \cite[Theorem D]{CEM}, $\varphi \colon G_F(2)\to G_K(2)$ is an isomorphism. This implies that $F$ is a superpythagorean field by Corollary~\ref{cor:superPytha}.
\qed
\end{rmks}
\section{Demushkin groups}
Recall that a pro-$p$-group $G$ is said to be a Demushkin group if
\begin{enumerate}
\item $\dim_{{\mathbb F}_p} H^1(G,{\mathbb F}_p)<\infty,$
\item $\dim_{{\mathbb F}_p} H^2(G,{\mathbb F}_p)=1,$
\item the cup product $H^1(G,{\mathbb F}_p)\times H^1(G,{\mathbb F}_p)\to H^2(G,{\mathbb F}_p)$ is a non-degenerate \mbox{bilinear} form.
\end{enumerate}
By the work of \cite{De1,De2}, \cite{Se1} and \cite{La1}, we now have a complete classification of Demushkin groups.
Let $G$ be a Demushkin group of rank $d=\dim_{{\mathbb F}_p} H^1(G,{\mathbb F}_p)$. Let $c_n=c_n(G)$. Then by \cite[Theorem 5.1 (g)]{La3} (see also \cite{Fo,Ga,LM}), we have the Hilbert-Poincar\'e series
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \frac{1}{1-dt+t^2}.
\]
We write $1-dt+t^2=(1-at)(1-bt)$ so that $a+b=d$ and $ab=1$. Then
\[
\log P_{{\rm gr}({\mathbb F}_p[[G]])}(t)= \log(\frac{1}{1-at})+\log(\frac{1}{1-bt})=\sum_{n\geq 1} \frac{1}{n}(a^n+b^n).
\]
We define the sequence $w_n(G), n=1,2,\ldots$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(m) (a^{n/m}+b^{n/m})=\frac{1}{n}\sum_{m\mid n} \mu(n/m) (a^m+b^m).
\]
\begin{rmk} The numbers $w_n(G)$ are given by the formula
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m) \left[ \sum_{0\leq i\leq [m/2]} (-1)^i \frac{m}{m-i} {m-i \choose i} d^{m-2i} \right].
\]
(See \cite[Proof of Proposition 4]{La2}.)
\end{rmk}
\begin{prop} If $n=p^k m $ with $(m,p)=1$, then
\[
c_n(G) =w_m(G) +w_{pm}(G)+\cdots + w_{p^km}(G).
\qedhere
\]
\end{prop}
\begin{ex} Let $G$ be a Demushkin pro-$p$-group of finite rank $d$. We have
\[
\begin{aligned}
c_1(G)&=d,\\
c_2(G)&= \begin{cases}
\frac{d^2-d-2}{2} \text{ if } p\not=2,\\
\frac{d^2+d-2}{2} \text{ if } p=2,
\end{cases}\\
c_3(G) &=\begin{cases}
\frac{d^3-4d}{3} \text{ if } p\not=3,\\
\frac{d^3-d}{3} \text{ if } p=3,
\end{cases}\\
c_4(G)&= \begin{cases}
\frac{d^4-5d^2+4}{4} \text{ if } p\not=2,\\
\frac{d^4-3d^2+2d}{4} \text{ if } p=2,
\end{cases}\\
c_5(G)&=\begin{cases}
\frac{d^5-5d^3+4d}{5} \text{ if } p\not=5,\\
\frac{d^5-5d^3+9d}{5} \text{ if } p=5.
\end{cases}
\end{aligned}
\]
Observe that our numbers $c_n(G)$, $n=1,2,\ldots$, also detect the minimal numbers of generators of $G_{(n)}$. Indeed by the remarkable result of I. V. Ando\v{z}skii and independently by J. Dummit and J. Labute for each open subgroup $T$ of the Demushkin group $G$, we have the following expression for the minimal number of generators $d(T)$ of $T$:
\[
d(T)=[G:T](d(G)-2)+2.
\]
(See \cite[Theorem 3.9.15]{NSW}.)
Therefore
\[
d_n(G):=d(G_{(n)})= p^{\sum_{i=1}^{n-1}c_i(G)}(d-2)+2.
\]
\end{ex}
From now on we assume that $G=F/\langle r\rangle$, where $F$ is a free pro-$p$-group on generators $x_1,x_2,\ldots,x_d$, and
\[
r=[x_1,x_2][x_3,x_4]\cdots [x_{d-1},x_d].
\]
Then we extract from \cite{La2} the following fact.
\begin{lem}
\label{lem:wn Demushkin} For every $n$, $w_n(G)={\rm rank}_{{\mathbb Z}_p} G_n/G_{n+1}$.
\end{lem}
\begin{proof} This follows from \cite[Theorem and proof of Proposition 4]{La2}. (Although \cite{La2} only treats abstract discrete groups, his results are also true for pro-$p$-groups with virtually the same proofs; one only has to replace ${\mathbb Z}$ by ${\mathbb Z}_p$, subgroups by closed subgroups, and group rings by completed group rings.)
\end{proof}
\begin{cor} Assume that for each $n$, $B_n$ represents a ${\mathbb Z}_p$-basis of $G_n/G_{n+1}$.
Let us write $n=p^km$ with $(m,p)=1$. Then a basis of the ${\mathbb F}_p$-vector space $G_{(n)}/G_{(n+1)}$ can be represented by the following set
\[
B_m^{p^k} \bigsqcup B_{pm}^{p^{k-1}}\bigsqcup \cdots \bigsqcup B_{p^{k-1}m}^p\bigsqcup B_n.
\]
\end{cor}
\section{Some other groups}
\subsection{Free products of a finite number of Demushkin groups and free pro-$p$-groups}
Let $G$ be a free pro-$p$ product of $r$ Demushkin groups of ranks $d_1,\ldots, d_r$, and of a free pro-$p$-group of rank $e$.
The Hilbert-Poincar\'e series of ${\rm gr}({\mathbb F}_p[[G]])$ is
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\frac{1}{1-(d_1+\cdots +d_r+e)t+ rt^2}=: \frac{1}{(1-at)(1-bt)}.
\]
We define the sequence $w_n(G), n=1,2,\ldots$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(m) (a^{n/m}+b^{n/m})=\frac{1}{n}\sum_{m\mid n} \mu(n/m) (a^m+b^m).
\]
\begin{prop} If $n=p^k m $ with $(m,p)=1$, then
\[
c_n(G) =w_m(G) +w_{pm}(G)+\cdots + w_{p^km}(G).
\qedhere
\]
\end{prop}
\subsection{A free product of a cyclic group of order 2 and a free pro-$2$-group}
We first consider the case of $p=2$ because this is the case of interest in Galois theory of $2$-extensions, and because this case is a bit simpler than the general case of any prime $p$. This latter case will be covered in the next subsection.
Let $G=C_2* S$ be a free pro-$2$ product of the cyclic group $C_2$ of order 2 and a free pro-$2$-group of rank $d$.
The Hilbert-Poincar\'e series of ${\rm gr}({\mathbb F}_2[[G]])$ is
\[
P_{{\rm gr}({\mathbb F}_2[[G]])}(t)=\left(\frac{1}{1+t}-dt\right)^{-1}=\frac{1+t}{1-dt- dt^2}=: \frac{1+t}{(1-at)(1-bt)}.
\]
We define the sequence $w_n(G), n=1,2,\ldots$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m) (a^m+b^m-(-1)^m).
\]
\begin{prop} If $n=2^k m $ with $(m,2)=1$, then
\[
c_n(G) =w_m(G) +w_{pm}(G)+\cdots + w_{2^km}(G).
\qedhere
\]
\end{prop}
\subsection{A free product of a cyclic group of order $p$ and a free pro-$p$-group}
Let $G=C_p* S$ be a free pro-$p$ product of the cyclic group $C_p$ of order $p$, and a free pro-$p$-group of rank $d$. We shall find a formula for $c_n(G)$.
The Hilbert-Poincar\'e series of ${\rm gr}({\mathbb F}_p[[G]])$ is
\[
P_{{\rm gr}({\mathbb F}_p[[G]])}(t)=\frac{1+t+\cdots+t^{p-1}}{1-dt- dt^2-\cdots-dt^p}=: \frac{(1-\xi_1 t)\cdots(1-\xi_{p-1}t)}{(1-a_1t)\cdots (1-a_pt)}.
\]
We define the sequence $w_n(G), n=1,2,\ldots$ by
\[
w_n(G)=\frac{1}{n}\sum_{m\mid n} \mu(n/m)( a_1^m+\cdots+a_p^m-(\xi_1^m+\cdots+\xi_{p-1}^m)).
\]
\begin{prop} If $n=p^k m $ with $(m,p)=1$, then
\[
c_n(G) =w_m(G) +w_{pm}(G)+\cdots + w_{p^km}(G).
\qedhere
\]
\end{prop}
\begin{rmk}
Note that
\[
\xi_1^n+\cdots+\xi_{p-1}^n
=\begin{cases}
-1 \text{ if } (n,p)=1,\\
p-1 \text{ if } p\mid n.
\end{cases}
\]
We shall compute $a_1^n+\cdots+a_{p}^n$.
From
\[
\frac{1}{(1-a_1t)\cdots(1-a_pt)}=\frac{1}{1-(dt+dt^2+\cdots+dt^p)},
\]
taking the logarithms of both sides, we obtain
\[
\begin{aligned}
& \sum_{n\geq 1}\frac{1}{n}(a_1^n+\cdots+a_p^n)t^n= \sum_{n\geq 1} \frac{1}{n} (dt+dt^2+\cdots +dt^p)^n\\
&= \sum_{n\geq 1} \frac{1}{n}\sum_{\substack{k_1+\cdots+k_p=n,\\k_i \geq 0}} \binom{n}{k_1,\ldots,k_p} (dt)^{k_1} (dt^2)^{k_2}\cdots (dt^p)^{k_p}\\
&= \sum_{M}\sum_{\substack{k_1+2k_2+\cdots+pk_p=M,\\k_i \geq 0}}\\
&\hspace*{60pt}
\left[\frac{1}{M-k_2-\cdots-(p-1)k_p} \binom{M-k_2-\cdots-(p-1)k_p}{k_1,\ldots,k_p} d^{M-k_2-\cdots-(p-1)k_p}\right] t^M.
\end{aligned}
\]
Finally comparing the coefficients of $t^n$ gives us the required formula for $a_1^n+\cdots+a_{p}^n$,
\[
\begin{aligned}
&a_1^n+\cdots+a_p^n\\
&=\sum_{\substack{k_1+2k_2+\cdots+pk_p=n,\\k_i \geq 0}}\frac{n}{n-k_2-\cdots-(p-1)k_p} \binom{n-k_2-\cdots-(p-1)k_p}{k_1,\ldots,k_p} d^{n-k_2-\cdots-(p-1)k_p}.
\end{aligned}
\]
\end{rmk}
|
\section{Introduction} \label{intro}
Quantum computation is believed to be more powerful than classical computation.
There are quantum algorithms that efficiently solve a problem for which no efficient classical algorithm is known so far,
such as the Shor's algorithm~\cite{S1997}, an algorithm for Pell's equation and the principal ideal problems~\cite{H2002,S2009}, and an algorithm for
approximate solutions of knot invariants~\cite{FLW2002,AJL2006,WY2008,AAEL2007}.
However, it has not been yet fully understood what makes quantum computation more powerful than classical one.
One approach to address this question is to investigate whether or not quantum computation described by circuits comprising a restricted class of quantum gates
still outperform classical computation.
From this perspective, it has been shown that, if the gates are restricted to those generating a little amount of entanglement~\cite{JL2003} or so-called matchgates acting on
nearest neighbor qubits~\cite{V2002,JM2008}, quantum computation is efficiently simulated by classical computers.
Quantum computation by diagonal quantum circuits in the computational basis with a separable initial state, which are often called
{\it instantaneous quantum polynomial-time (IQP)} circuits, is also attracting much attention~\cite{SB2009,BJS2011,NV2012,JWAURB2013,FM2013}.
A study of diagonal quantum circuits is motivated by a fact that they are rather classical in the sense that all gates commute each other.
Hence, they are suited for investigating the boarder of quantum and classical computational power.
Since noncommutativity of operators is one of the significant features of classical theory,
one may expect that diagonal quantum circuits would not outperform classical computers.
This natural expectation is however likely to be incorrect.
In Ref.~\cite{BJS2011}, it has been shown that IQP circuits are not classically simulatable under the assumption that the polynomial hierarchy (PH) does not collapse at the third level,
which is a highly plausible assumption in computational complexity theory.
Such computational power of IQP circuits is intuitively related to
a fact that a quantum circuit diagonal in the computational basis
can generate highly entangled state by choosing an appropriate separable initial state.
In fact, almost all states generated by IQP circuits are extremely entangled~\cite{NTM2012}.
These results clarify the distinction between quantum and classical computational power,
and give an insight that quantum theory is not necessarily reduced to classical one by imposing only commutativity of operators.
Diagonal quantum circuits also have a practical importance
since they are experimentally much simpler to realize and less sensitive to decoherence than non-diagonal quantum circuits.
Diagonal gates are fault-tolerantly realizable even by current technology, for instance, in superconducting and semiconducting systems~\cite{ABDPST2009}.
Moreover, the commutativity of all gates in a diagonal quantum circuit allows us to implement the circuit by a single time-independent commuting Hamiltonian.
In such an implementation by a Hamiltonian dynamics, a control of the order of interactions is not necessary since all interactions can be simultaneously applied due to their commutativity,
so that it significantly reduces practical time to perform computation and makes the implementation more robust.
For these reasons, it is worth studying what diagonal quantum circuits can perform beyond classical information processing,
which directly provides experimentally realizable quantum tasks by currently available technology.
In spite of their computational power and practical merits of diagonal quantum circuits, little is known about concrete applications of diagonal quantum circuits so far.
One of the applications of quantum diagonal circuits is a generation of {\it random states}, which are an ensemble of pure states uniformly distributed
in a Hilbert space with respect to the unitarily invariant measure.
Random states have many utilities in a wide range of applications in quantum information processing such as
quantum communicational tasks~\cite{L1997}, efficient measurements~\cite{RBSC2004}, an algorithmic use~\cite{RRS2005,S2006} and an estimation of gate fidelities~\cite{DCEL2009}.
However, generating random states requires exponential resource since the number of parameters in random states scales exponentially with the number of qubits.
Hence, efficient generations of approximate random states, called a {\it state $t$-design}, have been intensely studied~\cite{DCEL2009,EWSLC2003,DLT2002,ODP2007,DOP2007,Z2008,HL2009,DJ2011,HL2009TPE,BHH2012,CHMPS2013}.
A $t$-design of an ensemble is another ensemble that simulates up to the $t$th-order statistical moments of the original one~\cite{RBSC2004,AE2007}.
A state $t$-design can be approximately but efficiently generated by a quantum circuit~\cite{BHH2012} and can be used instead of random states in
many applications~\cite{L2009}.
In Refs.~\cite{NM2013,NKM2013}, it was shown that non-diagonal gates are not necessary for generating a state design
by presenting a way of approximately generating a state $t$-design by diagonal quantum circuits. Although the degree of approximation is small but constant and
cannot be improved by applying additional diagonal gates in general,
it was also shown that a state $2$-design can be exactly generated by combining a diagonal quantum circuit with a classical random procedure~\cite{NM2013}.
These results are practically useful since they provide a concrete application of diagonal quantum circuits in
quantum informational tasks exploiting a state $t$-design.
This allows us to experimentally demonstrate quantum advantages.
This paper aims to review the results on diagonal quantum circuits in terms of their computational power and applications.
It is organized as follows.
The computational power of diagonal quantum circuits is summarized in Sect.~\ref{sec:CompPower}.
The rest of the paper is devoted to overview the results obtained in our previous papers~\cite{NM2013,NKM2013} about applications of diagonal quantum circuits,
which are given in Sect.~\ref{sec:App}.
We summarize the paper with concluding remarks in Sect.~\ref{sec:Sum}.
Before leaving the introduction, we explain our notation in this paper.
We consider an $n$-qubit system, where its Hilbert space is denoted by $\mathcal{H}=(\mathbb{C}^2)^{\otimes n}$.
The eigenstates of the Pauli-$Z$ ($X$) operator are denoted by $\ket{0}$ and $\ket{1}$ ($\ket{+}$ and $\ket{-}$), which corresponds to the eigenvalues $1$ and $-1$, respectively.
They are related by $\ket{\pm} = (\ket{0} \pm \ket{1})/\sqrt{2}$.
We define the computational basis by tensor products of $\ket{i}$ ($i=0,1$), and denote it by $\{ \ket{\bar{m}} \}_{m=0,\cdots, 2^n-1}$, where $\bar{m}$ is a binary representation of $m$.
Throughout the paper, we study diagonal quantum circuits in the computational basis, so that we do not refer to the basis in the following.
\section{Computational power of diagonal quantum circuits} \label{sec:CompPower}
We review results on the computational power of diagonal quantum circuits mainly obtained in Refs.~\cite{BJS2011,FM2013}.
In Sect.~\ref{ssec:ComClass}, we provide a very brief introduction of some complexity classes, which will be used in the following subsections.
The results of computational complexity on IQP circuits are summarized in Sect.~\ref{ssec:IQP}.
\subsection{Computational complexity classes} \label{ssec:ComClass}
We briefly explain computational complexity classes for decision problems, which are the problems with yes-or-no answers.
We overview only the complexity classes related to an investigation of IQP circuits. For more details, see, e.g., Refs.~\cite{AB2009,P1994}.
We first explain complexity classes called {\it polynomial-time (P)} and {\it nondeterministic polynomial-time (NP)}.
The class P is a class of the problems solvable in polynomial time by classical computers.
The class NP is problems with the following properties:
if the answer is yes, there exists a proof of polynomial length to confirm this fact that can be verified in P,
and if the answer is no, all proofs purported to that the answer is yes are rejected in P.
The class P is clearly included in NP, but it has not been shown whether the inclusion is strict or not.
From the complexity classes P and NP, a {\it polynomial hierarchy (PH)} is defined based on an idea of oracle machines.
A class $\Delta_k$ ($k=1,2,\cdots$) in the hierarchy are recursively defined by $\Delta_1=P$ and $\Delta_{k+1}=P^{N\Delta_k}$,
where $N\Delta_k$ is the non-deterministic class associated with $\Delta_k$, e.g., $N\Delta_1=NP$.
The ${\rm P}^{N\Delta_k}$ is a set of problems that are in $P$ if an oracle belonging to a complexity class $N\Delta_k$ is allowed to use.
Since one use of an oracle is counted as one step, ${\rm P}^{N\Delta_k}$ is expected to to be much harder than $N\Delta_k$ although it has not been proven.
A complexity class PH is then defined by the union of all classes $\Delta_k$ (see e.g.~\cite{AB2009}).
In contrast to a deterministic feature of P, NP and PH in the sense that the computation is deterministically performed,
there are probabilistic complexity classes.
In particular, two classes {\it bounded-error probabilistic polynomial-time (BPP)} and {\it probabilistic polynomial-time (PP)} are important for investigating IQP circuits.
The problems in BPP can be solved probabilistically in polynomial time, where the probability of obtaining a correct answer should be greater than or equal to $2/3$.
If the probability of obtaining a correct answer is relaxed to be strictly greater than $1/2$, the class of the problems is called PP.
The distinction between BPP and PP is whether or not the lower bound of the probability that the algorithm provides a correct answer,
which should be strictly greater than $1/2$, can depend on the input size.
For BPP, the probability $2/3$ can be replaced with an arbitrary $p>1/2$ as long as it is constant, which enables us to amplify the probability arbitrary close to 1 by running the algorithm polynomially many times.
On the other hand, the lower bound of the probability is possibly dependent on the input size for PP, e.g., $p>1/2 + 1/2^n$. Hence, BPP is a subclass of PP.
For quantum computation, we explain only {\it bounded-error quantum polynomial-time (BQP)}~\cite{BV1993}, which is a quantum version of BPP.
Formally, BQP is a class of problems solvable by a quantum computer in polynomial time, where the probability to obtain a correct answer is greater than or equal to $2/3$.
The class BQP includes BPP and P, and is included in PP~\cite{ADH1997}, so that P $\subseteq$ BPP $\subseteq$ BQP $\subseteq$ PP.
Although the relation between BQP and NP is not exactly known,
there is an evidence that BQP is not included in PH~\cite{A2009}, implying that BQP is likely to be not included in NP. See also Fig.~\ref{Fig:CC}
Finally, we introduce computational classes with a post-selection.
For every probabilistic computational class $A$,
its post-selected version can be defined by allowing a post-selection, which is simply denoted by post-A,
e.g., post-BPP, post-PP, and post-BQP.
Although a post-selection may not be realistic,
it helps an investigation of relations between different complexity classes for which a post-selection is not allowed.
\begin{figure}[tb]
\centering
\includegraphics[width=60mm, clip]{CC.eps}
\caption{A diagram for the relations between complexity classes P, NP, PH, BPP, and BQP. Note that it is not clear whether or not the inclusions are strict.}
\label{Fig:CC}
\end{figure}
\subsection{IQP circuits and their computational power} \label{ssec:IQP}
Instantaneous quantum polynomial-time (IQP) circuits are quantum circuits comprising only diagonal gates in the computational basis with a separable initial state and
measurement in a separable basis.
In spite of the commutativity of diagonal gates, they have quantum features
if the initial state and the measurement basis are appropriately chosen.
The definition of IQP circuits is given by the following.
\begin{Definition}[{\bf IQP circuits~\cite{SB2009,BJS2011}}]
{\it An IQP circuit for $n$ qubits is a quantum circuit with the following structure:
each gate in the circuit is diagonal in the Pauli-$Z$ basis, the input state is $\ket{+}^{\otimes n}$,
and the output is the result of a measurement in the Pauli-$X$ basis on a specified set of output qubits.}
\end{Definition}
A complexity class related to IQP circuits is simply denoted by IQP.
Its classical simulatability has been studied in Refs.~\cite{BJS2011,FM2013},
where the simulatability of quantum circuits is defined in two ways, a {\it strong} and {\it weak} simulation~\cite{BJS2011}.
\begin{Definition}[{\bf Simulatability of quantum circuits}] \label{Def:SimQC}
{\it Let $\mathcal{C}$ and $P(\mathcal{C})$ be a quantum circuit of a uniform family and the probability distribution obtained by a
measurement in a separable basis, respectively.
A circuit family $\mathcal{C}$ is strongly simulatable if the output probability distribution $P(\mathcal{C})$ can be computed in polynomial time.
A circuit family $\mathcal{C}$ is weakly simulatable if there exists a method to sample the output according to
the probability distribution $P(\mathcal{C})$ in polynomial time.}
\end{Definition}
Although it is clear that strong simulatability implies weak one, their difference is large
since there exist quantum circuit families that are classically weakly simulatable but its exact strong simulation is in a complexity class $\sharp P$-complete~\cite{V2010}, which is strongly believed to be a much harder than P.
As the exact simulation of quantum circuits is too restrictive, an approximate simulation is also introduced in terms of a multiplicative error
for a weak simulation of quantum circuits, e.g., in Ref.~\cite{BJS2011}.
A circuit family $\mathcal{C}$ is weakly simulatable with a multiplicative error $c \geq 1$
if there is a sampling method according to the probability distribution $\tilde{P}(\mathcal{C})$, where
\begin{equation}
\frac{1}{c} P(\mathcal{C}) \leq \tilde{P}(\mathcal{C}) \leq c P(\mathcal{C}).
\end{equation}
The following theorem is about classical simulatability of IQP circuits.
\begin{Theorem}[{\bf Hardness of classical simulations of IQP circuits by Bremner et al.~\cite{BJS2011}}] \label{Thm:IQPhard}
{\it If the output probability distributions generated by IQP circuits could be classically weakly simulated to within multiplicative error
$1 \leq c < \sqrt{2}$ then PH $= \Delta_3$. This is the case even for IQP circuits with two-qubit gates.}
\end{Theorem}
The proof of Theorem~\ref{Thm:IQPhard} relies on an investigation of a post-selected version of IQP circuits.
It was first shown that if a post-selection is allowed, computational power of IQP circuits is equal to that of BQP.
This is interesting in itself since it implies that a post-selection closes the gap between diagonal and non-diagonal circuits.
The computational power of post-BQP is also known to be equal to PP~\cite{A2005}, resulting in a relation that
\begin{equation}
{\rm post {\text -} IQP} = {\rm PP}. \label{Eq:ccc}
\end{equation}
Moreover, under the assumption that IQP circuits are weakly classically simulatable as stated in Theorem~\ref{Thm:IQPhard},
it was shown that post-IQP $\subseteq$ post-BPP.
This relation and Eq.~\eqref{Eq:ccc} imply, together with a fact post-BPP $\subseteq$ post-BQP $=$ PP~\cite{A2005}, that post-BPP $=$ PP.
By using relations that PH $\subseteq$ P$^{\rm PP}$~\cite{T1991} and P$^{\rm post {\text -} BPP} \subseteq \Delta_3$~\cite{HHT1997},
it is obtained that PH $\subseteq \Delta_3$, meaning a collapse of the polynomial hierarchy at its third level (see Ref.~\cite{BJS2011} for the full detail).
Since it is highly implausible that PH $= \Delta_3$,
Theorem~\ref{Thm:IQPhard} implies that IQP circuits are highly unlikely to be weakly simulatable by classical computers.
One intuitive understanding of this computational power of IQP circuits
is obtained by studying the properties of states that appear during the computation by IQP circuits.
Although IQP circuits are capable to use only a restricted ensemble of states due to the commutativity of all gates and a fixed separable initial state,
it was shown that such an ensemble of states covers the whole Hilbert space fairly uniformly in the sense that
the ensemble is hard to be distinguished from the uniformly distributed states in the Hilbert space~\cite{NKM2013}.
Consequently, a state generated by an IQP circuit is typically highly entangled~\cite{NTM2012}.
These results are shown in the formulation of a $t$-design of random states, which we will explain in more detail in Sects.~\ref{ssec:RMRS} and~\ref{sssec:Approx}.
A fact that IQP circuits can exploit such a uniformly distributied states in a Hilbert space indicates that the outputs of the circuits
should typically have quantum features in spite of the commutativity of all gates,
and is expected to result in the computational power exceeding the classical one.
Although Theorem~\ref{Thm:IQPhard} shows that there is no universal classical method to simulate an output distribution of {\it any} IQP circuits,
it is also interesting to address a question {\it what types of IQP circuits are classically simulatable}.
This question is partially answered by mapping the output probability distribution to the partition function of the Ising models with imaginary coupling constants
on an associated graph~\cite{FM2013}.
The following Theorem is obtained by investigating whether or not the partition function of the mapped Ising model is exactly computed:
\begin{Theorem}[{\bf Classically simulatable subclass of IQP circuit by Fujii et al.~\cite{FM2013}}] \label{Thm:CSIQP}
{\it If IQP circuits are sufficiently sparse,
or if IQP circuits contain only two-qubit gates of the form $\exp[ i Z_l \otimes Z_m ]$ acting on nearest-neighbor qubits $(l,m)$ on a two-dimensional graph,
then the IQP circuits are classically simulatable in a strong sense.}
\end{Theorem}
The condition of the sparse property in Theorem~\ref{Thm:CSIQP} is given by an associated graph of IQP circuits, which is obtained by identifying qubits and diagonal gates in the circuit with vertices and hyperedges on a graph. If the associated graph is an {\it independent and full rank bipartite (IFRB)} graph, the sparse condition is satisfied
(see Ref.~\cite{FM2013} for details).
Theorem~\ref{Thm;CSIQP} implies that there exist IQP circuits that are easy to classically simulate.
Although only two specific cases were found in Ref.~\cite{FM2013},
the method developed in the paper, mapping an output probability distribution of IQP circuits to the partition function of the Ising models,
works in general and will provide a good tool for studying classical simulatability of IQP circuits.\\
Theorems~\ref{Thm:IQPhard} and~\ref{Thm:CSIQP} imply that IQP circuits show both classical and quantum computational power depending on the detailed structure of the circuits and that their computational power is on the border of classical and quantum ones.
Thus, IQP circuits will provide a good framework to address a question what property of quantum circuits distinguishes quantum from classical in terms of computational power.
Although this question can be addressed by investigating non-diagonal quantum circuits, IQP circuits are probably more suitable since they have
a much simpler structure than non-diagonal ones.
\section{Applications of diagonal quantum circuits} \label{sec:App}
In this section, we provide two applications of diagonal quantum circuits.
Since the applications are closely related to random unitary matrices and random states, we explain them in Sect.~\ref{sec:DUM}.
We also introduce random diagonal-unitary matrices and phase-random states, which are a restricted version of random unitary matrices and random states, respectively.
We then show how to implement a $t$-design of random diagonal-unitary matrices by diagonal quantum circuits in Sect.~\ref{sec:DUbyIQP}.
This guarantees an efficient realization of the applications of random diagonal-unitary matrices by diagonal quantum circuits.
In Sects~\ref{ssec:Gt} and~\ref{ssec:Therm}, we provide two applications of random diagonal-unitary matrices, generating a state $t$-design~\cite{NKM2013,NM2013} and a thermalizing algorithm for classical Hamiltonian.
\subsection{Random matrices, random states and $t$-designs} \label{sec:DUM}
We overview definitions and applications of random unitary matrices, random states and their $t$-designs.
In Sect.~\ref{ssec:RMRS}, we explain random unitary matrices and random states and
briefly summarize their properties and applications.
A study of $t$-designs of random unitary matrices and random states is summarized in Sect.~\ref{ssec:t}.
\subsubsection{Random unitary matrices and random states} \label{ssec:RMRS}
Random unitary matrices are originally introduced in the field of random matrix theory~\cite{M1990} and defined by an ensemble of unitary matrices uniformly distributed in a
unitary group with respect to the Haar measure.
Similarly, random states are an ensemble of pure states distributed uniformly in a Hilbert space,
which are obtained by applying random unitary matrices on a fixed initial state.
Their definitions are given by the following.
\begin{Definition}[{\bf Random unitary matrices~\cite{M1990} and Random states}]
{\it Let $\mathcal{U}(d)$ be the unitary group of degree $d$. Random unitary matrices $\mathcal{U}_{\rm Haar}$ are the ensemble of unitary matrices uniformly distributed in $\mathcal{U}(d)$ with respect to the Haar measure $d\mu_{\rm Haar}$.
Random states are the ensemble of states uniformly distributed in a Hilbert space, which are given by
$\{ U_{\mu} \ket{\Psi} \}_{U_\mu \in \mathcal{U}_{\rm Haar}}$ for any fixed state $\ket{\Psi}$ in a Hilbert space with dimension $d$.}
\end{Definition}
Note that the distribution of $\{ U_{\mu} \ket{\Psi} \}_{U_\mu \in \mathcal{U}_{\rm Haar}}$ is independent of
the choice of $\ket{\Psi}$ due to the unitary invariance of the Haar measure, i.e., $d\mu_{\rm Haar}(U)=d\mu_{\rm Haar}(UV)=d\mu_{\rm Haar}(V)$ for any $U,V \in \mathcal{U}(d)$.
The unitarily invariant property of random unitary matrices and random states leads to many uses of them in the field of quantum information.
Random unitary matrices are used for cryptographic use~\cite{BR2003,AMTW2000,AS2004,DN2006,HLSW2004,Au2009}, quantum communication tasks~\cite{L2000,HHL2004}, a quantum data hiding~\cite{DLT2002,TDL2001}, and as a mathematical tool to construct a counterexample of additivity conjecture~\cite{H2009}.
Random states similarly have many utilities such as for saturating a classical communication capacity of noisy quantum channels~\cite{L1997},
for efficient measurements~\cite{RBSC2004,RRS2005,S2006} and for the estimation of gate fidelities~\cite{DCEL2009}.
Random states have been also often used to reveal generic properties of quantum states.
Since random states are uniformly distributed in a Hilbert space, their properties are supposed to show generic properties of quantum states.
For instance, entanglement of random states, called {\it generic entanglement}, has been intensively investigated from this point of view~\cite{L1978,P1993,FK1994,R1995,ZS2001,HLW2006,G2007,FMPPS2008,PFPFS2010,NMV2010,NMV2011}, and
it has been shown that almost all random states are almost maximally entangled.\\
In analogy with random unitary matrices and random states, {\it random diagonal-unitary matrices} and {\it phase-random states} have been proposed in Refs.~\cite{NTM2012,NM2013}.
They are originally introduced to investigate a typical behavior of time-independent Hamiltonian dynamics, but
are turned out to be useful for investigating typical properties of diagonal quantum circuits and of the states generated by them.
Random diagonal-unitary matrices also have applications in quantum informational tasks as we will see in Sects.~\ref{ssec:Gt} and~\ref{ssec:Therm}.
\begin{Definition}[{\bf Random diagonal-unitary matrices~\cite{NM2013} and Phase-random states~\cite{NTM2012}}]
{\it Random diagonal-unitary matrices in an orthonormal basis $\{ \ket{u_n} \}$, denoted by $\mathcal{U}_{\rm diag}(\{ \ket{u_n} \})$, are an ensemble of diagonal unitary matrices of the form $U_{\varphi} = \sum_{n=1}^{d} e^{i \varphi_n} \ketbra{u_n}{u_n}$, where the phases $\varphi_n$ are uniformly distributed according to the normalized Lebesgue measure d$\varphi$ = d$\varphi_1 \cdots $d$\varphi_{d} / (2\pi)^{d} $ on $[0,2\pi)^d$.
Phase-random states are an ensemble of states $\{ U \ket{\Psi} \}_{U \in \mathcal{U}_{\rm diag}(\{ \ket{u_n} \})}$, which depends on the choice of the initial state $\ket{\Psi}$.}
\end{Definition}
As mentioned, we consider only random diagonal-unitary matrices and phase-random states in the computational basis in this paper.
Note that phase-random states depend on the choice of the initial state since random diagonal-unitary matrices do not have unitary invariance.
By choosing the initial state to be $\ket{+}^{\otimes n}$, the corresponding phase-random states are identified with an ensemble of all states that can be generated by IQP circuits.
Hence, generic properties of the states during the computation by IQP circuits can be revealed by investigating such phase-random states.
It has been shown that almost all phase-random states of which initial state is $\ket{+}^{\otimes n}$ are almost maximally entangled and they have even higher entanglement than generic entanglement of random states~\cite{NTM2012}. This implies that computation by IQP circuits possibly utilizes highly entangled states during the computation.
\subsubsection{$t$-designs} \label{ssec:t}
Although random (diagonal-)unitary matrices and (phase-)random states have been studied in quantum information from many perspectives,
they cannot be efficiently generated since the number of parameters scales exponentially with the number of qubits.
Hence, it is important to introduce approximate ones, which are called $t$-designs.
In the following, we denote by $\mathbb{E}$ expectations over a probability distribution for simplicity.
If necessary, we specify the probability space taken over for the expectation.
A $t$-design of an ensemble is defined by an ensemble that simulates up to the $t$th-order statistical moments of the original ensemble on average,
and an $\epsilon$-approximate $t$-design is an ensemble that approximates the $t$-design, where $\epsilon$ is a degree of approximation.
In the case of approximate designs of matrices, the degree of approximation is often evaluated by the diamond norm~\cite{KSV2002}.
For a superoperator $\mathcal{E}$ acting on the bounded operators on a Hilbert space $\mathcal{H}$, the diamond norm is defined by
\begin{equation}
|\!| \mathcal{E} |\!|_{\diamond} := \sup_d \sup_{X \neq 0} \frac{|\!| (\mathcal{E} \otimes {\rm id}_d) X |\!|_1 }{ |\!|X|\!|_1}, \notag
\end{equation}
where ${\rm id}_d$ is the identity operator on another $d$-dimensional Hilbert space $\mathcal{H}^\prime$ and $X$ is any positive operator on $\mathcal{H} \otimes \mathcal{H^\prime}$.
To define an $\epsilon$-approximate $t$-design, let $\mathcal{V}$ be an ensemble of unitary matrices and
$\mathcal{E}_{\mathcal{V}}(\rho)$ be a superoperator such that
\begin{align}
\mathcal{E}_{\mathcal{V}}^{(t)}(\rho) &:= \mathbb{E}_{U \in \mathcal{V}} [ U^{\otimes t} \rho (U^{\dagger})^{\otimes t}], \label{Eq:superop}
\end{align}
for any states $\rho$.
Then, an $\epsilon$-approximate unitary $t$-design is defined as follows (see, e.g., Ref.~\cite{HL2009}):
\begin{Definition}[{\bf $\epsilon$-approximate unitary $t$-designs~\cite{DCEL2009,TG2007}}] \label{Def:Ut}
{\it Let $\mathcal{U}$ be random unitary or diagonal-unitary matrices.
An $\epsilon$-approximate $t$-design of $\mathcal{U}$, denoted by $\mathcal{U}^{(t,\epsilon)}$, is an ensemble of unitary matrices such that
\begin{equation*}
|\!| \mathcal{E}^{(t)}_{\mathcal{U}} - \mathcal{E}^{(t)}_{\mathcal{U}^{(t,\epsilon)}} |\! |_{\diamond} \leq \epsilon.
\end{equation*}
The $t$-designs for random unitary and diagonal-unitary matrices are called {\it unitary} and {\it diagonal-unitary} $t$-designs, respectively.
When $\epsilon=0$, the design is called exact.}
\end{Definition}
From a property of the diamond norm, an operational meaning of a $t$-design is given by that a unitary $t$-design cannot be distinguished from random unitary matrices even if we have $t$-copies of the system and apply any operations on them.
We note that
there are several definitions of an $\epsilon$-approximate unitary designs in terms of different measures of the distance\footnote{In several definitions of a $t$-design, it is required that the ensemble is finite, but we do not impose such a condition in this paper to be more general.}.
However, they are all equivalent in the sense that,
if $\mathcal{V}$ is an $\epsilon$-approximate unitary $t$-design in one of the definitions,
then it is also an $\epsilon'$-approximate unitary $t$-design in other definitions, where $\epsilon' = {\rm poly} (2^{tn}) \epsilon$ (see Ref.~\cite{L2010} for details).
Similarly, an $\epsilon$-approximate state $t$-design is defined as follows.
\begin{Definition}[{\bf $\epsilon$-approximate state $t$-designs~\cite{RBSC2004,AE2007}}] \label{Def:appstate}
{\it Let $\Upsilon$ be random states or phase-random states.
An $\epsilon$-approximate $t$-design of $\Upsilon$, denoted by $\Upsilon^{(t,\epsilon)}$, is an ensemble of states such that
\begin{equation}
\biggl|\!\biggl|
\mathbb{E}_{\ket{\psi} \in \Upsilon^{(t,\epsilon)}} [ \ketbra{\psi}{\psi}^{\otimes t} ]
-
\mathbb{E}_{\ket{\psi} \in \Upsilon} [\ketbra{\psi}{\psi}^{\otimes t}]
\biggr|\! \biggr|_{1} \leq \epsilon,
\end{equation}
where $|\!| \cdot |\!|_{1} = \tr | \cdot |$ is the trace norm.
In particular, we refer to a $t$-design of random states and phase-random states as a state $t$-design and a toric $t$-design, respectively.
When $\epsilon=0$, the design is called exact.}
\end{Definition}
In most applications of random unitary matrices and random states, their $t$-designs for small $t$ can be exploited~\cite{L2009}.
Hence, an efficient implementation of a unitary $t$-design, which also provides an efficient generation of a state $t$-design, is important for their applications.
In particular, an implementation of a unitary $t$-design by a {\it random circuit} has been intensely studied~\cite{ODP2007,DOP2007,Z2008,HL2009,DJ2011,BHH2012},
where a random circuit is a quantum circuit comprising random two-qubit gates that act on randomly chosen pairs of qubits.
In Ref.~\cite{BHH2012}, it has been shown that random circuits with a constraint that each two-qubit gate acts on nearest neighbor qubits
achieves an $\epsilon$-approximate $t$-design by applying $O(Nt^4(N+\log1/\epsilon))$ gates.
In contrast to the implementation of a unitary $t$-design,
diagonal-unitary $t$-designs for general $t$ cannot be achieved by only one- and two-qubit {\it diagonal} gates.
This is because of the abelian property of the corresponding group
and a multi-qubit diagonal gate cannot be generally expressed by a product of diagonal matrices acting on smaller number of qubits.
For instance, a diagonal matrix ${\rm diag} (1,1,1,1,1,1,1,-1 )$ acting on three qubits cannot be decomposed into a product of diagonal two-qubit matrices.
Hence, implementing a diagonal unitary $t$-design by diagonal quantum circuits is a non-trivial task.
We review the results on the implementation of a diagonal-unitary $t$-design in the next section.
\subsection{Achieving a diagonal-unitary $t$-design by diagonal quantum circuits} \label{sec:DUbyIQP}
In this Section, we provide an efficient implementation of a diagonal-unitary $t$-design by diagonal quantum circuits.
This guarantees that the applications of a diagonal-unitary $t$-design, which will be given in Sects.~\ref{ssec:Gt} and~\ref{ssec:Therm}, are efficiently realizable by diagonal quantum circuits.
In particular, we consider to implement a diagonal-unitary $t$-design by a certain type of diagonal quantum circuits, called an {\it $r$-qubit phase-random circuit with a gate set $\mathcal{G}$} that generally containes multi-qubit gates. We introduce it in Sect.~\ref{ssec:PRC}.
In Sect.~\ref{ssec:ExactCons}, we provide a necessary and sufficient condition for an $r$-qubit phase-random circuit to achieve an exact diagonal-unitary $t$-design obtained in Ref.~\cite{NKM2013}.
In Sect.~\ref{ssec:ApproxCons},
we also show an approximate implementation of a diagonal-unitary $2$-design by using a simpler gate set composed only of the controlled-$Z$ gate and
single-qubit random diagonal gates~\cite{NM2013}.
Since experimental manipulations of multi-qubit gates are not easy, this may help experimental implementations of the design.
\subsubsection{Phase-random circuit} \label{ssec:PRC}
\begin{figure}[tb!]
\centering
\begin{tabular}{ccc}
\begin{minipage}{0.33\hsize}
\centering
\includegraphics[width=40mm, clip]{rPRC.eps}
\end{minipage}
\begin{minipage}{0.33\hsize}
\centering
\includegraphics[width=50mm, clip]{PRC.eps}
\end{minipage}
\begin{minipage}{0.33\hsize}
\centering
\includegraphics[width=50mm, clip]{PRCCZ.eps}
\end{minipage}
\end{tabular}
\caption{
Panel (a) shows an $r$-qubit phase-random circuit with a diagonal gate set $\mathcal{G}_r$ in an $n$-qubit system, where $r=3$.
The black circles indicate the places of qubits $I_r$ on which the $r$-qubit gate acts on, and each $r$-qubit gate is randomly drawn from a diagonal gate set $\mathcal{G}_r$.
Panel (b) shows a $3$-qubit phase-random circuit used in Sect.~\ref{ssec:ExactCons} in a $5$-qubit system.
In this case, the gates are applied on all combinations of $r$ qubits out of the $n$ qubits, so that $I_r$ is deterministically chosen.
Note that the order of the application is arbitrary since all gates are commutable.
Panel (c) shows the phase-random circuit with a gate set $\mathcal{G}_{\rm CZ}$ in Sect.~\ref{ssec:ApproxCons}.
The circle represents a single-qubit random diagonal gates followed by the controlled-$Z$ gate. The places of qubits $I_2$ are randomly chosen from $\{1,\cdots, n\}$.
}
\label{Fig:rPRC}
\end{figure}
An {\it $r$-qubit phase-random circuit with a gate set $\mathcal{G}_r$} consists of $T$ diagonal unitary gates. Each of the gates act on $r$ qubits.
For the $p$th gate, we choose $r$ different numbers $I_r^{(p)} \subset \{1,2,\cdots, n \}$ and apply an $r$-qubit diagonal gate $W_{I_r^{(p)}}$ on qubits
at sites $I_r^{(p)}$, where the gate $W_{I_r^{(p)}}$ is randomly chosen from a gate set $\mathcal{G}_r$.
An instance of the circuit is then specified by a set of parameters, $\mathcal{C}_T:=\{I_r^{(p)},W_{I_r^{(p)}}\}_{p=1}^T$, and the unitary operation corresponding to the circuit is given by $U_T = W_{I_r^{(T)}} W_{I_r^{(T-1)}} \cdots W_{I_r^{(1)}}$.
Thus, an $r$-qubit phase-random circuit with length $T$ is described by a set of the unitary operations $\{U_T\}_{\mathcal{C}_T}$.
See also Fig.~\ref{Fig:rPRC} (a).
\subsubsection{Exact implementation} \label{ssec:ExactCons}
Consider an $r$-qubit phase-random circuit with a diagonal gate set $\mathcal{G}_r$, where the gate set is given by
a set of diagonal gates with random phases in the computational basis,
\begin{equation}
\mathcal{G}_r = \{ {\rm diag} (e^{i \varphi_1}, e^{i \varphi_2}, \cdots, e^{i \varphi_{2^r}})\}_{\varphi_k \in [0,2 \pi)}. \label{Eq:G_r}
\end{equation}
We apply an $r$-qubit diagonal gates randomly drawn from the gate set $\mathcal{G}_r$ on all combinations of $r$ qubits out of $n$ qubits.
In this circuit, the choice of the place of qubits acted by the $p$th gate, $I_r^{(p)}$, is deterministic (see also Fig.~\ref{Fig:rPRC} (b)).
The random parameters in the circuit are then only phases in the diagonal gates,
leading to the probability measure on the phase-random circuit given by $\prod_{t=1}^T \prod_{k=1}^{2^r} d\varphi_k^{(t)}/2 \pi$.
The following Theorem provides a relation between $r$ and $t$ for this phase-random circuit to achieve a diagonal-unitary $t$-design.
\begin{Theorem}[{\bf Exact implementation of a diagonal-unitary $t$-design by Y. Nakata et al~\cite{NKM2013}}] \label{Thm:NecSuf_r}
{\it The $r$-qubit phase-random circuit with the gate set $\mathcal{G}_r$ defined above is an exact diagonal-unitary $t$-design if and only if
$r > \log_2 t $ for $t \leq 2^{n}-1$ and $r=n$ for $t \geq 2^N$.}
\end{Theorem}
Theorem~\ref{Thm:NecSuf_r} is obtained based on another equivalent definition of a $t$-design in terms of $\mathbb{E} [U^{\otimes t} \otimes (U^{\dagger})^{\otimes t}]$
(see e.g. Ref.~\cite{L2010}).
Then, we use a fact that
a matrix element in $U^{\otimes t} \otimes (U^{\dagger})^{\otimes t}$ containing a term $e^{i \varphi}$ becomes zero by averaging it over all $U \in \mathcal{C}_T$
since $\varphi$ is randomly chosen from $[0, 2 \pi)$.
By comparing the place of the constant terms in $U^{\otimes t} \otimes (U^{\dagger})^{\otimes t}$ for the phase-random circuit $\{U_T\}_{\mathcal{C}_T}$ and that for
random diagonal-unitary matrices $\mathcal{U}_{\rm diag}$, the statement of Theorem is reduced to a set identification problem.
The set identification problem can be solved in a combinatorial manner, and we obtain Theorem~\ref{Thm:NecSuf_r}.
The number of gates in the circuit is $\binom{n}{r}$, which should be $poly(n)$ for the circuit to have an efficient classical description,
implying that $r$ should be constant.
Hence, Theorem~\ref{Thm:NecSuf_r} implies that a diagonal-unitary $t$-design can be efficiently implemented by the phase-random circuit when
$t$ is constant with respect to the number of qubits.
If we restrict the circuit to use only two-qubit gates, we obtain an exact $t$-design only for $t\leq 3$.
This is contrasted to an implementation of an approximate unitary $t$-design for any $t$ by using two-qubit gates~\cite{BHH2012}.
This difference comes from a fact that there exist universal gate sets for non-diagonal quantum circuits,
but does not exist a counterpart for diagonal quantum circuits, except a trivial one, if the gate set is restricted to be diagonal ones.
We also note that the gate set $\mathcal{G}_r$ can be replaced with a finite set of simpler multi-qubit gates. For the details, see Ref.~\cite{NKM2013}.
\subsubsection{Approximate implementation by a two-qubit gate set} \label{ssec:ApproxCons}
In Theorem~\ref{Thm:NecSuf_r}, the gate set $\mathcal{G}_r$ is chosen to be $r$-qubit diagonal gates with random phases.
Since it may not be experimentally feasible to manipulate multi-qubit gates with random parameters,
it will be helpful to investigate what can be achieved by a fixed multi-qubit gate and random single-qubit diagonal gates.
Motivated by this, it was shown that a simpler gate set containing the controlled-$Z$ gate can achieve an approximate diagonal-unitary $2$-design~\cite{NM2013}.
Let us consider
a $2$-qubit phase-random circuit with a diagonal gate set $\mathcal{G}_{\rm CZ}$ given by
\begin{equation}
\mathcal{G}_{\rm CZ} = \biggl\{\begin{pmatrix} 1 & 0 \\ 0 & e^{i \alpha} \end{pmatrix} \otimes \begin{pmatrix} 1 & 0 \\ 0 & e^{i \beta} \end{pmatrix}
\begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & -1 \end{pmatrix} \biggr\}_{ \alpha, \beta \in [ 0, 2 \pi)}.
\end{equation}
In this case, we choose a pair of two qubits $I_2^{(p)}$ for the $p$-th gate randomly from $\{1,\cdots, n\}$ (see Fig.~\ref{Fig:rPRC} (c)).
Thus, the probability measure of the circuit is given by $\prod_{p=1}^T \frac{2}{N(N-1)} d \alpha_p d\beta_p/(2\pi)^2$.
In this phase-random circuit, there exist terms in
$\mathbb{E}_{\mathcal{C}_T}[U^{\otimes t} \otimes (U^{\dagger})^{\otimes t}]$ that are equal to $-1$ due to the use of the controlled-$Z$ gate,
while all elements in $\mathbb{E}_{\mathcal{U}_{\rm diag}}[U^{\otimes t} \otimes (U^{\dagger})^{\otimes t}]$ are either $0$ or $1$.
Consequently, the phase-random circuit does not achieve an exact diagonal-unitary $t$-design.
Nevertheless, it achieves an $\epsilon$-approximate diagonal-unitary $2$-design by applying at most
$O(N^2(N+\log \epsilon^{-1}))$ gates randomly drawn from the gate set $\mathcal{G}_{\rm CZ}$ as stated in the following Theorem.
\begin{Theorem}[{\bf Approximate implementation of a diagonal 2-design by Y. Nakata et al~\cite{NM2013}}] \label{Thm:phase-random circuitCZ}
{\it The $2$-qubit phase-random circuit with a gate set $\mathcal{G}_{\rm CZ}$ of a length $T$ is an $\epsilon$-approximate diagonal-unitary $2$-design
if $T \geq T_{conv}(\epsilon)$, where
\begin{equation}
\frac{N}{2} + \biggl(\frac{N^2}{4} + O(N) \biggr)\log (2\epsilon)^{-1} \leq T_{conv}(\epsilon)
\leq 7 N^3 \log2 + N^2 \log\epsilon^{-1} + O(N^2).
\end{equation}
Therefore, the phase-random circuit is an $\epsilon$-approximate diagonal-unitary $2$-design after applying at most
$O(N^2(N+\log \epsilon^{-1}))$ two-qubit gates.}
\end{Theorem}
Theorem~\ref{Thm:phase-random circuitCZ} is proven by a method developed in Ref.~\cite{ODP2007,DOP2007}, which maps the transformation of a state by the circuit
to a Markov chain on a certain graph.
By investigating the mixing time of the Markov chain, we obtain Theorem~\ref{Thm:phase-random circuitCZ}.
Note that a random choice of a pair of two qubits in the phase-random circuit is crucial in Theorem~\ref{Thm:phase-random circuitCZ}.
If the qubits are deterministically chosen, the commutativity of gates makes the circuit stationary after the two-qubit gates are applied on all pairs of qubits.
It is easy to check that such a stationary circuit is not a diagonal-unitary $2$-design.
When the two qubits are randomly chosen, the classical randomness prevents the circuit from being stationary even after
the gates are applied on all pairs of qubits.
This is also clear from the probability measure of the phase-random circuit.
As a result, the degree of approximation is reduced to be arbitrarily small and obtain
an $\epsilon$-approximate diagonal-unitary $2$-design.
\subsection{Generating a state $t$-design by a diagonal-unitary $t$-design} \label{ssec:Gt}
In Ref.~\cite{NKM2013}, it was shown that applying a diagonal-unitary $t$-design on a specific initial state achieves a good approximate {\it state} $t$-design
for any $t$, although the degree of approximation is constant.
The degree of approximation can be improved by combining a diagonal-unitary $t$-design with a classical random procedure,
and particularly, an exact state $2$-design are obtained~\cite{NM2013}.
Since a state $t$-design for small $t$ is used in many quantum informational tasks and
diagonal-unitary $t$-design can be implemented by diagonal quantum circuits, these results provides an application of diagonal quantum circuits.
We review a generation of an approximate state $t$-design in Sect.~\ref{sssec:Approx}, and an exact one for $t=2$ in Sect.~\ref{sssec:Exact}.
\subsubsection{Approximate state $t$-design} \label{sssec:Approx}
The following proposition states that a diagonal-unitary $t$-design with an appropriate separable initial state
generates an approximate state $t$-design.
\begin{Proposition}[{\bf Generating an approximate state $t$-design by Y. Nakata et al~\cite{NKM2013}}] \label{Prop:Protocol1}
{\it An ensemble of states obtained
by applying a diagonal-unitary $t$-design to an initial state $\ket{+}^{\otimes n}$
is an $\eta(n,t)$-approximate state $t$-design, where
\begin{equation}
\eta(N,t)=\frac{t(t-1)}{2^n} + O(\frac{1}{2^{2n}}).
\end{equation}}
\end{Proposition}
Although the degree of approximation $\eta(n,t)$ is constant and cannot be improved by applying additional diagonal gates, it is already a good approximation for $t$
independent of $n$.
This good approximation of random states by the phase-random states is a consequence of the {\it concentration of measure} of random states~\cite{L2001}.
The concentration of measure in this case means that
almost all random states are equal-amplitudes states in the sense that
$\ket{\Psi} = \sum_{k=1}^{2^n} c_k \ket{u_k}$, where $|c_k| \sim 2^{-n/2}$ and $\{ \ket{u_k} \}$ is some orthonormal basis.
Hence, random states can be well-approximated by an ensemble of states with equal-amplitudes in a fixed basis.
However, the distribution of amplitudes of states in an ensemble generally depends on the basis,
which differs from the distribution of coefficients of random states independent of the basis due to the unitary invariance.
This makes the ensemble in Proposition~\ref{Prop:Protocol1} not an exact but approximate state design.
Proposition~\ref{Prop:Protocol1} implies that a diagonal-unitary $t$-design is capable to generate an ensemble of states of which distribution is
hard to distinguish from the uniform one as long as looking at lower order statistical moments.
Since a diagonal-unitary $t$-design is obtained by a phase-random circuit with an initial state $\ket{+}^{\otimes n}$, which is an IQP circuit,
computation by IQP circuits is typically exploiting uniformly distributing states in the Hilbert space, as mentioned in Sect.~\ref{ssec:IQP}.
\subsubsection{Exact state $2$-design by virtue of classical randomness} \label{sssec:Exact}
Although a diagonal quantum circuit achieves only an approximate state $t$-design with a fixed degree of approximation,
the degree of approximation can be improved by combining them with a classical random procedure.
In particular, the resulting ensemble becomes an exact design when $t=2$ as stated below.
\begin{Proposition}[{\bf Generating an exact state $2$-design by Y. Nakata et al~\cite{NM2013}}] \label{Prop:2exact}
{\it Consider the following protocol.
\begin{enumerate}
\item With probability $\frac{1}{2^n+1}$, choose a random $n$-bit sequence $\bar{m}$ and generate a state $\ket{\bar{m}}$.
\item With probability $\frac{2^n}{2^n+1}$, apply a diagonal-unitary $2$-design on an initial state $\ket{\vec{+}}=\ket{++ \cdots +}$.
\end{enumerate}
Then, the resulting ensemble is an exact state $2$-design.}
\end{Proposition}
Proposition~\ref{Prop:2exact} is simply obtained by looking at the difference between
$\mathbb{E}_{\ket{\psi} \in \Upsilon^{(2)}_{\rm Haar}} [\ketbra{\psi}{\psi}^{\otimes 2}]$, where $ \Upsilon^{(2)}_{\rm Haar}$ represents a state $2$-design, and
$\mathbb{E}_{U \in \mathcal{U}^{(2)}_{\rm diag}} [( U \ketbra{\vec{+}}{\vec{+}} U^{\dagger})^{\otimes 2}]$, which is given by
\begin{align}
\mathbb{E}_{\ket{\psi} \in \Upsilon^{(2)}_{\rm Haar}} [\ketbra{\psi}{\psi}^{\otimes 2}]
=
\frac{2^n}{2^n +1}
\mathbb{E}_{U \in \mathcal{U}^{(2)}_{\rm diag}} [( U \ketbra{\vec{+}}{\vec{+}} U^{\dagger})^{\otimes 2}]
+
\frac{1}{2^n + 1}\sum_{m} \ketbra{\bar{m}}{\bar{m}}^{\otimes 2}. \label{Eq:protexact}
\end{align}
The equation~\eqref{Eq:protexact} implies that an exact state $2$-design is obtained by a probabilistic mixture of the ensemble generated by a diagonal-unitary $2$-design
$\{ U \ketbra{\vec{+}}{\vec{+}} U^{\dagger}\}_{U \in \mathcal{U}^{(2)}_{\rm diag}}$ and
product states $\{ \ket{\bar{m}} \}$.
This protocol of generating an exact state $2$-design has experimental merits compared to previously known protocols listed below:
\begin{itemize}
\item A generation of an exact unitary 2-design using Clifford operations is known~\cite{DLT2002}, which requires $O(n^8)$ bits and $O(n^2)$ quantum gates.
In the protocol, unitary matrices of generating an exact design are classically calculated and are decomposed into one- and two-qubit unitary gates.
Thus, for obtaining a state $2$-design,
it needs to repeat calculating a gate decomposition and constructing the corresponding quantum circuit.
\item An $\epsilon$-approximate unitary 2-design is obtained by a quantum circuit composed of one- and two-qubit Clifford gates, where
some gates are applied probabilistically~\cite{DCEL2009}. The number of gates is $O(n(n+ \log 1/\epsilon))$~\cite{HL2009} in Definition~\ref{Def:appstate}.
\item
A random circuit~\cite{ODP2007,DOP2007} with length $O(n(n+ \log 1/\epsilon))$ is an $\epsilon$-approximate unitary 2-design~\cite{HL2009,DJ2011},
where all gates are randomly chosen from a gate set called a {\it 2-copy gapped} gate set, e.g., a set of the controlled-NOT gate and random single-qubit gates.
\item A local random circuit with length $O(n t^4( n + \log 1/\epsilon))$ gates is an $\epsilon$-approximate unitary $t$-design~\cite{BHH2012} for any $t$.
The circuit is composed of random $SU(4)$ gates acting on nearest neighbor qubits.
\end{itemize}
The protocol in Proposition~\ref{Prop:2exact} is experimentally preferable because of the following reasons.
First, a diagonal-unitary $2$-design is achieved by diagonal quantum circuits comprising only two-qubit diagonal gates.
Moreover,
the circuits are implementable by a single time-independent commuting Hamiltonian due to the commutativity of all the gates.
Since such a Hamiltonian can be simultaneously applied on all qubits, the practical time of an implementation is significantly reduced compared to non-diagonal ones.
Hence, the implementation of a quantum part in the protocol is easier and more robust than other protocols.
Finally, most of the previous protocols except the first one achieve only an {\it approximate} $2$-design,
while the protocol in Proposition~\ref{Prop:2exact} achieves an {\it exact} one.
We also emphasize that there is no drawback in the protocol of Proposition~\ref{Prop:2exact} since the same number of gates as that of previous ones are used in the protocol.
One may wonder how much the degree of approximation is improved by adding classical randomness in the case of general $t$.
It was shown that the improvement is limited to be $O(2^{(1-t)n})$ for general $t$~\cite{NKM2013}.
Since the degree of approximation without classical randomness is $\eta(n,t) = O(1/2^n)$,
the protocol does not result in a significant improvement except for $t=2$.
\subsection{Verifying the principle of apparently equal a priori probability in quantum statistical mechanics} \label{ssec:Therm}
Diagonal quantum circuits can be also used to verify the foundation of quantum statistical mechanics.
In quantum statistical mechanics, a derivation of a canonical thermal state $e^{-\beta H}/\tr e^{-\beta H}$ for a given Hamiltonian $H$ with an inverse temperature $\beta$
from natural assumptions is one of the fundamental problems~\cite{N1955,S1989,H1998,GLTZ2006}.
Recently, a new development on the problem based on {\it the principle of apparently equal a priori probability} has been made~\cite{PSW2006}.
We introduce a quantum algorithm of generating a thermal state based on the principle by using a diagonal quantum circuit, which we call a {\it thermalizing algorithm}.
We overview the principle in Sect.~\ref{sssec:can} and provide the thermalizing algorithm for a certain class of Hamiltonians in Sect.~\ref{sssec:themal}.
\subsubsection{Appearance of a canonical thermal state} \label{sssec:can}
In Ref.~\cite{PSW2006}, it was shown that a standard assumption for the derivation of canonical thermal states, i.e., {\it the equal a priori probability postulate} which is also known as a {\it microcanonical assumption}, can be replaced by a weaker assumption.
To clarify the situation, let us consider a composite system $S+B$, where $S$ and $B$ represent a system and a thermal bath, respectively,
and denote their Hilbert spaces by $\mathcal{H}_S$ and $\mathcal{H}_B$, respectively.
Let $H_{\rm tot}=H_S+H_B+H_{\rm int}$ be a Hamiltonian acting on $\mathcal{H}_S \otimes \mathcal{H}_B$, where $H_S$ ($H_B$) acts on only $\mathcal{H}_S$ ($\mathcal{H}_B$) and
$H_{\rm int}$ acts on both.
We denote the eigen decomposition of $H_{\rm tot}$ by $\sum_i e_i \ketbra{e_i}{e_i}$ and define a subspace restricted by total energy $E$,
$\mathcal{H}_E = {\rm span}\{ \ket{e_i} | E - \delta < e_i< E\}$, where $\delta$ is supposed to be sufficiently small.
Then, the equal a priori probability postulate states that
it is always the case in the systems describable by thermodynamics that the equiprobable state $\mathbb{I}_{\mathcal{H}_E}/\tr \mathbb{I}_{\mathcal{H}_E}$ is realized, where $\mathbb{I}_{\mathcal{H}_E}$ is a projector onto $\mathcal{H}_E$.
This postulate leads to a thermal state in the system $S$ in the thermodynamic limit, i.e., $\tr_B \mathbb{I}_{\mathcal{H}_E}/\tr \mathbb{I}_{\mathcal{H}_E} =
e^{-\beta H_S}/\tr e^{-\beta H_S}$ where the inverse temperature $\beta$ is determined by the energy $E$ in the total system $S+B$.
However, the assumption is very strong since it is a statement about one single initial state.
What has shown in Ref.~\cite{PSW2006} is that the assumption of the equiprobable state can be relaxed to random states in $\mathcal{H}_E$.
This is the principle of apparently equal a priori probability.
More precisely, it has been shown that, for almost all random states $\ket{\Psi}$ in $\mathcal{H}_E$, the reduced density state in the system $S$ is very close to
that of the equiprobable state, $\tr_{B} \ketbra{\Psi}{\Psi} \sim \tr_B \mathbb{I}_{\mathcal{H}_E}/\tr \mathbb{I}_{\mathcal{H}_E} = e^{-\beta H_S}/\tr e^{-\beta H_S}$.
Since this implies that a state randomly drawn from the Hilbert space $\mathcal{H}_E$ results in a thermal state in a system $S$ with high probability,
it is a stronger statement than the equal a priori probability postulate.
Similarly, it has been shown that this is also the case for certain types of phase-random states~\cite{NTM2012},
that is, almost all phase-random states in the subspace $\mathcal{H}_E$ locally equilibrate to a thermal state in the above sense if the basis and the initial state of the phase-random states satisfy certain conditions.
In particular, if the initial state is an equal superposition of all eigenstates in $\mathcal{H}_E$, the condition is satisfied.
\subsubsection{A thermalizing algorithm} \label{sssec:themal}
\begin{figure}[tb]
\centering
\includegraphics[width=70mm, clip]{THM.eps}
\caption{A quantum circuit for implementing thermalizing algorithm of classical Hamiltonians by using random diagonal-unitary matrices. The upper half of the circuit represents ancilla qubits and the lower half represents the system and the artificial thermal bath.
RDU and QPE (QPE$^{-1}$) denote random diagonal-unitary matrices and the (inverse of) quantum phase estimation, respectively. }
\label{Fig:THM}
\end{figure}
The principle of apparently equal a priori probability offers a possibility of a quantum algorithm that realizes a thermal state by using random and phase-random states in $\mathcal{H}_E$.
Here, we particularly provide a quantum algorithm using a diagonal quantum circuit, which efficiently realizes a thermal state of classical Hamiltonians.
By classical Hamiltonians, we mean those with separable eigenstates.
Such algorithms are important from the viewpoint not only of an experimental verification of the foundation of quantum statistical mechanics,
but also of obtaining a thermal state of arbitrary classical Hamiltonian in experiments.
The latter has utilities in condensed matter physics, since it enables us to measure expectation values of any observables easily in experiments,
and also has a possible application in quantum information processing that exploits a thermal state~\cite{FNOM2013}.
Note that there are several algorithms that realize a thermal state of a given Hamiltonian without knowing its eigenenergies and eigenstates, e.g.,
quantum Metropolis algorithm~\cite{TOVPV2011} and an artificial thermalization circuit~\cite{RGE2012}, which are based on different mechanisms from the principle of apparently equal a priori probability.
As shown in Ref.~\cite{NTM2012}, it is sufficient for realizing a thermal state to prepare phase-random states with equal-amplitudes in the computational basis in the subspace $\mathcal{H}_E$.
To see this more clearly, let us consider $n=n_S + n_B$ qubits, where $n_S$ and $n_B$ are the number of qubits in the system $S$ and in the thermal bath $B$, respectively.
We denote by $\{ \ket{e_l^E} \}_{l=1,\cdots, d_E}$ the eigenstates of $H_{\rm tot}$ in the restricted subspace $\mathcal{H}_E$.
Note that they are a subset of the computational basis by the assumption that the Hamiltonian is classical.
The principle of apparently equal a priori probability based on phase-random states in $\mathcal{H}_E$~\cite{NTM2012} states that
almost all phase-random states in the form of
\begin{equation}
\frac{1}{\sqrt{d_E}} \sum_{l=1}^{d_E} e^{i \varphi_l} \ket{e_l^E},
\end{equation}
are locally close to a thermal state.
This state is approximately achieved by the quantum circuit presented in Fig.~\ref{Fig:THM}.
The circuit uses $r$ ancilla qubits initially prepared in $\ket{+}^{\otimes r}$, random diagonal-unitary matrices, the quantum phase estimation (QPE)~\cite{K1995,NC2000} and the projective measurement on the ancilla qubits in the computational basis.
Random diagonal-unitary matrices can be replaced by a diagonal-unitary $t$-design by a similar argument in Ref.~\cite{L2009},
and the most parts of QPE except the quantum Fourier transformation can be also implemented by a diagonal quantum circuit since the Hamiltonians are classical.
At each part of the circuit, the state changes as follows:
\begin{enumerate}
\item The random diagonal-unitary matrices acting on the system and the thermal bath add random phases $\{\varphi_l\}$ to $\ket{+}^{\otimes n}$,
and generates a phase-random state $2^{-n/2} \sum_{l=1}^{2^n} e^{i \varphi_l} \ket{\bar{l}}$ in $\mathcal{H}_S+\mathcal{H}_B$.
\item After QPE, the state is approximately $\ket{\Psi_{e}}=2^{-n/2} \sum e^{i \varphi_l} \ket{\bar{l}} \otimes \ket{\bar{e}_l}$ where $\bar{e}_l$ is a binary representation of the eigenenergy $e_l$ of the total Hamiltonian $H_{\rm tot}$.
\item By performing the projective measurement $P:=\{ P_E, P_{\neg E} \}$, where $P_E$ is a projection operator onto $\mathcal{H}_E$ and $P_{\neg E} = I-P_E$, on the ancilla qubits, the state is probabilistically changed into $\frac{1}{\sqrt{d_E}} \sum_{l=1}^{d_E} e^{i \varphi_l} \ket{e_l^E} \otimes \ket{\bar{e}_l}$.
Note that this measurement is done in the computational basis.
The success probability is given by a probability to obtain the outcome corresponding to $P_E$.
Otherwise, the algorithm fails.
\item The inverse of QPE changes the state back to $\ket{\Psi_f}= (\frac{1}{\sqrt{d_E}} \sum_{l=1}^{d_E} e^{i \varphi_l} \ket{e_l^E}) \otimes \ket{+}^r$,
so that we obtain a phase-random state in the subspace $\mathcal{H}_E$ by tracing out the ancilla qubits.
\end{enumerate}
There are two error factors in the algorithm, which originates from QPE.
First, the eigenenergy is approximated by binary numbers within a precision of $2^{-r}$. This approximation results in a round-off error~\cite{TOVPV2011}.
Second, QPE does not exactly transform the state to $\ket{\Psi_{e}}$. This is inherited in the final state, resulting in $\sqrt{(1-\epsilon)}\ket{\Psi_f} + \sqrt{\epsilon} \ket{\Psi_{Error}}$. However, these errors can be sufficiently suppressed by preparing a large number of ancilla qubits such that $2^{-r} \ll \Delta e$, where $\Delta e$ is the minimum energy gap of $H_{\rm tot}$. It is often the case for local Hamiltonians that $\Delta e$ scales at worst exponentially with the number of particles $n$.
Hence, the error of the algorithm is sufficiently small if $r$ is chosen to be $poly(n)$.
Note that obtaining a thermal state at low temperatures is generally difficult since the probability to obtain $P_E$ in the projective measurement depends on the temperature as shown in Ref.~\cite{RGE2012}.
Although the main aim of our algorithm is an experimental verification of the principle of apparently equal a priori probability,
it has an advantage to some extent for the purpose of obtaining a thermal state even compared to other thermalizing algorithms~\cite{TOVPV2011,RGE2012}
since most parts of the algorithm are implementable
by diagonal quantum circuits.
On the other hand, it has certain drawbacks.
One is that our algorithm needs to use the Hadamard gate in the quantum Fourier transformation in QPE.
The number of the Hadamard gates is $r$, which is the number of ancilla qubits and determines the error of the algorithm.
Thus, there is a trade-off relation between the number of the Hadamard gates and the precision of the algorithm.
The other is that our algorithm works only for classical Hamiltonians. This drawback can be solved if random diagonal-unitary matrices in our algorithm are replaced
by random unitary matrices. However, such an algorithm necessarily requires non-diagonal quantum circuits, which is probably not feasible by current experimental technology.
\section{Summary and concluding remarks} \label{sec:Sum}
In this paper, we have reviewed a study of diagonal quantum circuits in the computational basis
motivated by theoretical and practical interests.
From the theoretical point of view, diagonal quantum circuits are a good framework to investigate the origin of quantum speed-up
since quantum computation described by diagonal quantum circuits is supposed to be on the boarder of classical and quantum computation.
Based on this idea, we have reviewed the computational power of IQP circuits in terms of classical simulatability.
It has been shown that the output probability distribution of IQP circuits is in general highly implausible to be classically simulated.
This is the case even if the circuits are composed only of $2$-qubit diagonal gates.
However, there also exist IQP circuits that are classically simulatable.
On the other hand, diagonal quantum circuits have a practical importance since realizations of diagonal gates in experiments are more feasible than non-diagonal gates.
They are implementable in a fault-tolerant manner by current technology, so that
any quantum tasks using diagonal quantum circuits are likely to be already realizable in experiments.
We have reviewed two applications of diagonal quantum circuits, generating an approximate state $t$-design and a thermalizing algorithm.
Such applications enable us to experimentally demonstrate quantum advantages in informational tasks and quantum nature in statistical mechanics.
It is worth investigating the diagonal quantum circuits further from both theoretical and practical point of view.
On classical simulatability of IQP circuits, it is important to study the complete classification of IQP circuits
since
clarifying computational power of IQP circuits will lead a better understanding of the origin of quantum speed-up.
Since IQP circuits have a simple structure,
this approach is more suitable than studying a distinction between quantum computation by non-diagonal quantum circuits and classical one.
It is also interesting to consider a difficulty of average instances of IQP circuits.
The results about hardness of classical simulation reviewed in this paper are obtained by deriving a highly implausible statement from
an assumption that {\it all} IQP circuits can be classical simulatable.
This methodology, however, does not answer to the question about hardness of classical simulatability of a {\it specific} IQP circuit.
One way to address this question is to study classical simulatability of randomly chosen IQP circuits, which is an idea of the study of average instances in computational complexity theory.
Investigating a difficulty of average instances of IQP circuits is important to fully understand classical simulatability of IQP circuits.
It is also practically desirable to consider what diagonal quantum circuits can perform beyond classical information processing.
This will be an important step toward a realization of a quantum computer since such tasks are experimentally realizable to
demonstrate quantum advantages and will accelerate an experimental challenge of making a quantum computer.
One possible application of diagonal quantum circuits is a {\it decoupling} of two systems~\cite{DBWR2010,BF2013}, which is often used in quantum informational tasks.
The decoupling is originally proposed by using random unitary matrices, but it is not necessary to use them.
Since random diagonal-unitary matrices have similar properties to random unitary matrices in some aspects,
an approximate decoupling may be achievable by using diagonal-unitary designs realized by diagonal quantum circuits.\\
This work is supported by Project for Developing Innovation Systems of the Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan.
Y. N. acknowledges JSPS Postdoctoral Fellowships for Research Abroad.
M.~M acknowledge support from JSPS by KAKENHI, Grant No. 23540463.
M. M also acknowledge to the ELC project (Grand No. 24106009) for encouraging the research project in this paper.
|
\section{Introduction}
Let $\Omega$ be a convex polygonal domain of $\mathbb{R}^2$ with boundary $\partial \Omega$.
We are interested in a mixed finite element method for the nonlinear elliptic Monge-Amp\`ere equation: find a smooth
convex function $u$ such that
\begin{align} \label{m1}
\begin{split}
\det (D^2 u) & = f \, \text{in} \, \Omega\\
u & = g \, \text{on} \, \partial \Omega.
\end{split}
\end{align}
For $u \in C^2(\Omega)$, $D^2 u=\bigg( (\partial^2 u) / (\partial x_i \partial x_j)\bigg)_{i,j=1,\ldots, 2} $ denotes the Hessian matrix of $u$ and $\det D^2 u$ denotes its determinant. The function $f$ defined on $\Omega$ is assumed to satisfy $f \geq c_0 >0$ for a constant $c_0 >0$ and we assume that $g \in C(\partial \Omega)$ can be extended to a function $\tilde{g} \in C(\tir{\Omega})$ which is convex in $\Omega$.
We consider a mixed formulation with unknowns the scalar variable $u$ and the Hessian $D^2 u$. The scalar variable and the components of the Hessian are approximated by Lagrange elements of degree $k \geq 2$. The method considered in this paper was analyzed from different point of views in \cite{Neilan2013} and \cite{AwanouLiMixed1} for smooth solutions of \eqref{m1}. In both \cite{Neilan2013} and \cite{AwanouLiMixed1} the convergence of the method for Lagrange elements of degree $k=1$ and $k=2$ was left unresolved. In this paper we resolve this issue for quadratic elements.
The ingredients of our approach consist in a fixed point argument, which yields the convergence of a time marching method, a ''rescaling argument'', i.e. the solution of a rescaled version of the equation, and
the continuity of the eigenvalues of a matrix as a function of its entries. This is the same approach we took in the case of the standard finite element discretization of the Monge-Amp\`ere equation \cite{Awanou-Std01}.
With the mixed methods, one can apply directly Newton's method to the discrete nonlinear problem and still have numerical evidence of convergence to a larger class of non smooth solutions than what is possible with the standard finite element discretization. We refer to \cite{Neilan2013,Lakkis11b} for the numerical results. Moreover with the standard finite element discretization \cite{Awanou-Std01}, convexity must be enforced weakly through appropriate iterative methods. Although the number of unknowns in the mixed methods is higher, in \cite{Neilan2013,Lakkis11b} the discrete Hessian was eliminated from the discrete equations in the implementation. However, as observed in \cite{AwanouLiMixed1} this prevents numerical convergence for smooth solutions when linear elements are used to approximate all the unknowns. We note that in \cite{Neilan2013} a stabilized method was proposed which works numerically for non smooth solutions in two dimension. It consists in using piecewise constants for the discrete Hessian and linear elements for the scalar variable. The analysis for smooth solutions of the lowest order methods discussed in \cite{AwanouLiMixed1,Neilan2013} cannot be done with the approach of this paper. The techniques used in this paper generalize to the three-dimensional problem but only for $k \geq 3$.
It should be possible to extend the approach taken in this paper to the formulation where discontinuous elements are used to approximate the unknowns \cite{Neilan2013}. Numerical results reported in \cite{Neilan2013} indicate the latter approach could lead to a less accurate approximation of the Hessian. For simplicity, and to focus on the methodology we present, we do not consider such an extension in this paper.
We organize the paper as follows. In the second section we introduce some notation and preliminaries. The error analysis of the mixed method is done in section \ref{error}.
\section{Notation and Preliminaries} \label{notation}
We use the usual notation $L^p(\Omega), 2 \leq p \leq \infty$ for the Lebesgue spaces and $H^s(\Omega), 1 \leq s < \infty$ for the Sobolev spaces of elements of $L^2(\Omega)$ with weak derivatives of order less than or equal to $s$ in $L^2(\Omega)$. We recall that $H_0^1(\Omega)$ is the subset of $H^1(\Omega)$
of elements with vanishing trace on $\partial \Omega$. We also recall that $W^{s,\infty}(\Omega)$ is the Sobolev space of functions with weak derivatives of order less than or equal to $s$ in $L^{\infty}(\Omega)$.
For a given normed space $X$, we denote by $X^{2}$ the space of vector fields with components in $X$ and by $X^{2 \times 2}$ the space of matrix fields with each component in $X$.
The norm in $X$ is denoted by $|| . ||_X$ and we omit the subscript $\Omega$ and superscripts $2$ and $2 \times 2$ when it is clear from the context.
The inner product in $L^2(\Omega), L^2(\Omega)^2$, and $L^2(\Omega)^{2 \times 2}$ is denoted by $(,)$ and we use $\< , \>$ for the inner product on $L^2(\partial \Omega)$ and $L^2(\partial \Omega)^2$. For inner products on subsets of $\Omega$, we will simply append the subset notation.
We denote by $n$ the unit outward normal vector to $\partial \Omega$. We recall that for a matrix $A$, $A_{ij}$ denote its entries and the cofactor matrix of
$A$, denoted $\cof A$, is the matrix with entries $(\cof A)_{ij}=(-1)^{i+j} \det(A)_i^j$ where $\det(A)_i^j$ is the determinant of the matrix obtained from $A$ by deleting its $i$th row and its $j$th column. For two matrices $A=(A_{ij})$ and $B=(B_{ij})$, $A: B=\sum_{i,j=1}^2 A_{ij} B_{ij}$ denotes their Frobenius inner product. A quantity which is constant is simply denoted by $C$.
For a scalar function $v$ we denote by $D v$ its gradient vector and recall that $D^2 v$ denotes the Hessian matrix of second order derivatives.
The divergence of a matrix field is understood as the vector obtained by taking the divergence of each row.
In this section and section \ref{error} we assume that \eqref{m1} has a solution which is sufficiently smooth. Put $\sigma=D^2 u$. Then the unique convex solution $u \in H^3(\Omega)$ of \eqref{m1} satisfies the following mixed problem:
Find $(u,\sigma) \in H^2(\Omega) \times H^1(\Omega)^{2 \times 2}$ such that
\begin{align} \label{m11}
\begin{split}
(\sigma,\tau) + (\div \tau, D u) - \< D u, \tau n \> & = 0, \forall \tau \in H^1(\Omega)^{2 \times 2} \\
(\det \sigma,v) & = (f,v), \forall v \in H_0^1(\Omega)\\
u & = g \, \text{on} \, \partial \Omega.
\end{split}
\end{align}
It is proved in \cite{AwanouLiMixed1} that the above variational problem is well defined.
\subsection{Discrete variational problem}
We denote by $\mathcal{T}_h$ a triangulation of $\Omega$ into simplices $K$ and assume that $\mathcal{T}_h$ is quasi-uniform. We denote by $V_h$ the standard Lagrange finite element space of degree $k \geq 2$ and denote by $\Sigma_h$ the space of symmetric matrix fields with components in the Lagrange finite element space of degree $k \geq 2$. Let $I_h$ denote the standard Lagrange interpolation operator from $H^s(\Omega), s \geq k+1$ into the space $V_h$. We use as well the notation $I_h$ for the matrix version of the Lagrange interpolation operator mapping $H^s(\Omega)^{2 \times 2}$, for $s \geq k+1$, into $\Sigma_h$.
We consider the problem:
find $(u_h, \sigma_h) \in V_h \times \Sigma_h$ such that
\begin{align} \label{m11h}
\begin{split}
(\sigma_h,\tau) + (\div \tau, D u_h) - \< D u_h, \tau n \> & = 0, \forall \tau \in \Sigma_h\\
(\det \sigma_h, v) & = ( f, v), \, \forall v \in V_h \cap H_0^1(\Omega)\\
u_h & = g_h \, \text{on} \, \partial \Omega,
\end{split}
\end{align}
where $g_h=I_h \tilde{g}$.
It follows from the analysis in \cite{Neilan2013,AwanouLiMixed1} that \eqref{m11h} is well-posed for $k \geq 3$ and error estimates were given. In section \ref{error} we give an error analysis valid for $k \geq 2$
For $v_h \in V_h$, we will make the abuse of notation of using $D^2 v_h$ to denote the Hessian of $v_h$ computed element by element. We will need the broken Sobolev norm
$$
||v||_{H^k(\mathcal{T}_h)} = \bigg( \sum_{K \in \mathcal{T}_h} ||v||^2_{H^k(K)}
\bigg)^{\frac{1}{2}}.
$$
\subsection{Properties of the Lagrange finite element spaces}
We recall some properties of the Lagrange finite element space of degree $k \geq 1$ that will be used in this paper. They can be found in \cite{Brenner02,Bramble86}. We have
Interpolation error estimates.
\begin{align} \label{interpol}
\begin{split}
||v - I_h v||_{H^j} & \leq C h^{k+1-j} ||v||_{H^{k+1}}, \forall v \in H^{s}(\Omega), j=0,1, \, \\
||v - I_h v||_{L^{\infty}} & \leq C h^{k} |v|_{H^{k+1}}, \forall v \in H^{s}(\Omega).
\end{split}
\end{align}
Inverse inequalities
\begin{align}
||v||_{L^{\infty}} & \leq C h^{-1} ||v||_{L^2}, \forall v \in V_h \label{inverse0} \\
||v||_{H^1} & \leq C h^{-1} ||v||_{L^2}, \forall v \in V_h \label{inverse1} \\
||v||_{H^{k+1}(\mathcal{T}_h)} & \leq C h^{-k-1} ||v||_{L^{2}}, \forall v \in V_h. \label{inverse2}
\end{align}
Scaled trace inequality
\begin{align}
||v||_{L^2(\partial \Omega)} &\leq C h^{-\frac{1}{2}} ||v||_{L^2},\ \forall v \in V_h. \label{trace-inverse}
\end{align}
\subsection{Algebra with matrix fields}
We collect in the following lemma some properties of matrix fields, the proof of which can be found in \cite{AwanouLiMixed1,AwanouPseudo10}.
\begin{lemma}
For $K \in \mathcal{T}_h$ and $u, v \in C^2(K)$ we have
\begin{align} \label{mean-v}
\det D^2 u - \det D^2 v = \cof(t D^2 u + (1-t) D^2 v): (D^2 u - D^2 v),
\end{align}
for some $t \in [0,1]$. It can be shown that $t=1/2$, \cite{Brenner2010b}.
For two $2 \times 2$ matrix fields $\eta$ and $\tau$
\begin{align}
||\cof (\eta):\tau||_{L^2} & \leq C ||\eta||_{L^{\infty}}^{} ||\tau||_{L^2}, \label{cof-est} \\
\cof (\eta) - \cof (\tau) & = \cof(\eta-\tau). \label{cof-mv}
\end{align}
\end{lemma}
\subsection{Continuity of the eigenvalues of a matrix as a function of its entries}
Let $\lambda_1(A)$ and $\lambda_2(A)$ denote the smallest and largest eigenvalues of the symmetric matrix $A$. We have
\begin{lemma} [\cite{Awanou-Std01}, Lemma 3.1] \label{lem-1}
There exists constants $m, M >0$ independent of $h$ and a constant $C_{conv} > 0$ independent of $h$ such that for all $v_h \in V_h$ with
$v_h=g_h$ on $\partial \Omega$ and
$$
||v_h-I_h u||_{H^1} < C_{conv} h^{2},
$$
we have
$$
m \leq \lambda_1(\cof D^2 v_h(x)) \leq \lambda_2(\cof D^2 v_h(x)) \leq M, \forall x \in K, K \in \mathcal{T}_h.
$$
\end{lemma}
The following lemma was used implicitly in \cite{AwanouPseudo10,Awanou-Std01,Awanou-Std-fd}.
\begin{lemma} \label{time-trick}
Assume $0 < \alpha < 1$ and $\alpha^{} \leq (m+M)/(2m)$ for constants $m, M >0$.
Let $B$ be a symmetric matrix field such that
$$
0 < m \alpha^{} \leq \lambda_1(B(x)) \leq \lambda_2 (B(x)) \leq M \alpha^{ }, \forall x \in \Omega.
$$
Then for $\nu=(m+M)/2$
$$
\gamma \equiv \sup_{v, w \in V_h \atop |v|_{H^1}=1, |w|_{H^1}=1 } \bigg| (D v , D w) - \frac{1}{\nu} (B D v, D w)\bigg|,
$$
satisfies $0 < \gamma < 1$.
\end{lemma}
\begin{proof}
Since $\lambda_1(B)$ and $\lambda_2(B)$ are the minimum
and maximum respectively of the Rayleigh quotient $((B z) \cdot z)/||z||^2$, where $||z||$ denotes the Euclidean norm of $\mathbb{R}^2$, we have
for $x \in \Omega$
$$
m \alpha^{} ||z||^2 \leq (B(x) z)\cdot z \leq M \alpha ||z||^2, z \in \mathbb{R}^2.
$$
This implies
\begin{equation*}
m \alpha^{} |w|_{H^1}^2 \leq \int_{\Omega} [ B(x) D w(x)] \cdot D w(x) \, \ud x \leq M \alpha^{} |w|_{H^1}^2, w \in V_h.
\end{equation*}
If we assume in addition that $|w|_{H^1}=1$, we get
\begin{equation*}
m \alpha^{} \leq \int_{\Omega} [ B(x) D w(x)] \cdot D w(x) \, \ud x \leq M \alpha^{}, w \in V_h.
\end{equation*}
It follows that
\begin{equation*}
(1-\frac{M \alpha^{}}{\nu}) \leq \int_{\Omega} [I - \frac{1}{\nu} B(x) D w(x)] \cdot D w(x) \, \ud x \leq (1-\frac{m \alpha^{}}{\nu}) , w \in V_h.
\end{equation*}
Since $\nu= (m+M)/2$, we have
\begin{align*}
1- \frac{\alpha^{} M}{ \nu} & = \frac{m+M - 2 M \alpha^{}}{m+M} < 1 \\
1- \frac{\alpha^{} m}{ \nu} & = \frac{m+M - 2 m \alpha^{}}{m+M} < 1.
\end{align*}
If we define
$$
\beta \equiv \sup_{v \in V_h, |v|_{H^1}=1 } \bigg|(D v , D v) - \frac{1}{\nu} (B D v, D v)\bigg|,
$$
by the assumptions on $\alpha$, we have
$$
0 < \beta < 1.
$$
We can define a bilinear form on $V_h$ by the formula
\begin{align*}
(p,q) & = \int_{\Omega} [(I -\frac{1}{\nu}(B(x)) D p(x)] \cdot D q(x) \ud x.
\end{align*}
Then because
$$
(p,q) = \frac{1}{4} ((p+q,p+q) - (p-q,p-q)),
$$
and using the definition of $\beta$, we get assuming that $|p|_{H^1}=|q|_{H^1}=1$,
\begin{align*}
|(p,q)| & \leq \frac{\beta}{4} (p+q,p+q) + \frac{\beta}{4} (p-q,p-q) \\
& \leq \frac{\beta}{4} |p+q|_{H^1}^2 + \frac{\beta}{4} |p-q|_{H^1}^2 = \beta.
\end{align*}
This completes the proof.
\end{proof}
\section{Error analysis of the mixed method for smooth solutions} \label{error}
We will assume without loss of generality that $h \leq 1$.
The goal of this section is to prove the local solvability of \eqref{m11h} for Lagrange elements of degree $k \geq 2$. We define for $\rho >0$,
$$\bar B_h(\rho)=\{(w_h, \eta_h) \in V_h\times \Sigma_h,\ \| w_h-I_hu\|_{H^1 }\leq \rho,\ \|\eta_h-I_h\sigma\|_{L^{2}}\leq h^{-1}\rho\}.$$
We are interested in elements $(w_h, \eta_h) \in V_h\times \Sigma_h$ satisfying
\begin{equation} \label{discrete-H}
(\eta_h, \tau)+(\div \tau, Dw_h)-\<Dw_h, \tau n\>=0, \forall \tau\in \Sigma_h.
\end{equation}
We define
\begin{align*}
\begin{split}
Z_h & =\{ \, (w_h, \eta_h) \in V_h\times \Sigma_h, w_h =g_h \, \text{on} \, \partial \Omega, (w_h, \eta_h) \ \text{solves} \ \eqref{discrete-H}
\, \} \, \text{and}
\end{split}
\end{align*}
\begin{equation*}
B_h(\rho)=\bar B_h(\rho)\cap Z_h.
\end{equation*}
In \cite{AwanouLiMixed1} the local solvability of \eqref{m11h} was obtained by a fixed point argument which consists in a linearization at the exact solution of \eqref{m1}. To be able to obtain results for quadratic elements we use a time marching method combined with a rescaling argument.
This is the point of view we took in \cite{Awanou-Std01,Awanou-Std-fd}. We first describe the time marching method at the continuous level.
Let $\nu >0$. We consider the sequence of problems
\begin{align*}
-\nu \Delta u^{r+1} &= -\nu \Delta u^{r} + \det D^2 u^r -f \, \text{in} \, \Omega \\
u^{r+1} &= g \, \text{on} \, \partial \Omega.
\end{align*}
Put $\sigma^{r+1}=D^2 u^{r+1}$. We obtain the equivalent problems
\begin{align*}
\sigma^{r+1}& =D^2 u^{r+1} \, \text{in} \, \Omega \\
-\nu \tr \sigma^{r+1} &= -\nu \tr \sigma^{r} + \det \sigma^r -f, \, \text{in} \, \Omega \\
u^{r+1} &= g \, \text{on} \, \partial \Omega,
\end{align*}
where $\tr A$ denotes the trace of the matrix $A$.
We are thus lead to consider the sequence of discrete problems:
find $(u_h^{r+1}, \sigma_h^{r+1}) \in V_h \times \Sigma_h$ such that $u_h^{r+1} = g_h \, \text{on} \, \partial \Omega$ and
\begin{align}
(\sigma_h^{r+1},\tau) & + (\div \tau, D u_h^{r+1}) - \< D u_h^{r+1}, \tau n \> = 0, \forall \tau \in \Sigma_h \label{m11h-iter01}\\
-\nu (\tr \sigma^{r+1},v) &= -\nu (\tr \sigma^{m},v) +(\det \sigma_h^{r}-f, v) , \, \forall v \in V_h \cap H_0^1(\Omega), \label{m11h-iter02}
\end{align}
given an initial guess $(u^0_h,\sigma^0_h)$. We prove below the convergence of $(u_h^{r+1}, \sigma_h^{r+1}) $ to a local solution $(u_h,\sigma_h)$ of the discrete problem \eqref{m11h}. Although \eqref{m11h-iter01}--\eqref{m11h-iter02} may be used in the computations, it is better to use in practice Newton's method.
Let $\alpha >0$. We define a mapping $T: V_h\times \Sigma_h\rightarrow V_h\times\Sigma_h$ by
\begin{eqnarray*}
T(w_h, \eta_h)=(T_1(w_h, \eta_h), T_2(w_h, \eta_h)),
\end{eqnarray*}
where $T_1(w_h, \eta_h)$ and $T_2(w_h, \eta_h)$ satisfy
\begin{align}
\begin{split}
(\eta_h-T_2(w_h, \eta_h), \tau)& +(\div\tau, D(w_h-T_1(w_h, \eta_h))) \\
& \ -\<D(w_h-T_1(w_h, \eta_h)), \tau n\> =(\eta_h, \tau) \\
& \qquad \qquad \qquad +(\div \tau, Dw_h)-\<Dw_h, \tau n\>, \quad \forall\ \tau\in \Sigma_h \label{eqn.1}
\end{split}
\end{align}
\begin{align}
-\nu (\tr T_2(w_h, \eta_h),v) & = - \nu (\tr \eta_h,v) + (\det \eta_h - \alpha^2 f, v), \ \forall\ v\in V_h\cap H^1_0(\Omega) \label{eqn.2}\\
T_1(w_h, \eta_h)&=w_h \quad \text{on}\quad \partial\Omega. \label{eqn.3}
\end{align}
Note that \eqref{eqn.1} is equivalent to
\begin{align}
(T_2(w_h, \eta_h), \tau) +(\div\tau, D T_1(w_h, \eta_h) ) -\<D T_1(w_h, \eta_h), \tau n\> = 0 \ \forall \ \tau\in \Sigma_h. \label{eqn.11}
\end{align}
Let $I$ denote the $2 \times 2$ identity matrix. We first make the following important observation.
For $v\in V_h\cap H^1_0(\Omega)$ and $\tau=v I$, we have $\div \tau= D v$ and since $v=0$ on $\partial \Omega$, we have in addition $\tau n=0$ on $\partial \Omega$. Thus using \eqref{eqn.11}
we obtain
\begin{equation} \label{w-eta-comp}
-\nu (\tr T_2(w_h, \eta_h),v) = - \nu (T_2(w_h, \eta_h), v I) = \nu (D T_1(w_h, \eta_h), D v).
\end{equation}
Similarly, we obtain that if $(w_h,\eta_h)$ solves \eqref{discrete-H}, then
\begin{equation} \label{w-eta-comp2}
(\tr \eta_h ,v) = -(D w_h, D v), \forall v\in V_h\cap H^1_0(\Omega).
\end{equation}
\begin{lemma} \label{final-lem2}
The mapping $T$ is well defined
and if $( \alpha w_h, \alpha \eta_h)$ is a fixed point of \eqref{eqn.1}--\eqref{eqn.3} with $w_h=g_h$ on $\partial \Omega$, then $( w_h, \eta_h)$ solves the nonlinear problem \eqref{m11h}.
\end{lemma}
\begin{proof}
To prove the first assertion, it is enough to prove that if $(w_h,\eta_h) \in V_h \times \Sigma_h$ is such that $w_h = 0 \, \text{on} \, \partial \Omega$ and
\begin{align*}
(\eta_h,\tau) + (\div \tau, D w_h) - \< D w_h, \tau n \> &= 0, \forall \tau \in \Sigma_h \\
-\nu (\tr \eta_h,v) &= 0 , \, \forall v \in V_h \cap H_0^1(\Omega),
\end{align*}
then $w_h=0$ and $\eta_h=0$.
Using \eqref{w-eta-comp2}, we obtain $0= - (\tr \eta_h,v) = (D w_h ,D v)$, for all $v \in V_h\cap H^1_0(\Omega)$. Thus $|w_h|^2_{H^1} =0$. This proves that $w_h=0$ by Poincar\'e's inequality. Using $\tau=\eta_h$ we obtain as well $\eta_h=0$.
The proof of the second assertion is immediate.
\end{proof}
We recall from \cite[Remark 3.6]{AwanouLiMixed1}, see also \cite{Neilan2013,Lakkis11b}, that for $v_h \in V_h$, there exists a unique $\eta_h \in \Sigma_h$ denoted $H(v_h)$, such that
\begin{equation} \label{disc-H}
(H(v_h), \tau)+(\div \tau, D v_h)-\<D v_h, \tau n\>=0, \forall \tau\in \Sigma_h,
\end{equation}
holds. To see this
consider the problem: find $\eta_h \in \Sigma_h$ such that
\begin{eqnarray} \label{disc-H2}
(\eta_h, \tau)=-(\div\tau, D v_h )+\<D v_h, \tau n\>, \quad \forall \tau\in \Sigma_h.
\end{eqnarray}
For $\tau\in \Sigma_h$, we define $F(\tau) = -(\div\tau, D v_h )+\<D v_h, \tau n\>$. Clearly $F$ is linear. By the Schwarz inequality, \eqref{inverse1} and \eqref{trace-inverse}
\begin{align*}
|-(\div\tau, D v_h)+\<D v_h, \tau\cdot n\> |& \leq C ||\tau||_{H^1} || v_h||_{H^1} + C || v_h||_{H^1(\partial \Omega)} ||\tau||_{L^2(\partial \Omega)} \\
& \leq C (h^{-1} || v_h||_{H^1} + h^{-\frac{1}{2}} || v_h||_{H^1(\partial \Omega)} ) ||\tau||_{L^2}.
\end{align*}
Thus a unique solution $\eta_h=H(v_h)$ exists by the Lax-Milgram Lemma.
\begin{rem} \label{H-vh-rem}
From the definition of $H(v_h)$ \eqref{disc-H} and \eqref{disc-H2}, we have for $v_h \in V_h$,
$$
H(\alpha v_h) = \alpha H(v_h).
$$
\end{rem}
\begin{lemma} \label{est-disc-H}
Let $v_h \in V_h$ such that $||v_h -I_h u||_{H^1} \leq \mu$. Then
$$
||H(v_h) - I_h \sigma||_{L^2} \leq C h^{-1} \mu +Ch^{k-1}.
$$
\end{lemma}
\begin{proof}
For $\tau \in \Sigma_h$, by \eqref{m11} and \eqref{disc-H} we have
\begin{align*}
(H(v_h) -I_h\sigma, \tau) & = (H(v_h) - \sigma, \tau) + (\sigma -I_h\sigma, \tau) \\
& = (\sigma-I_h\sigma, \tau)-(\div\tau, D(v_h-u))+\<D( v_h-u), \tau n\> \\
& = (\sigma-I_h\sigma, \tau)-(\div\tau, D(v_h-I_h u))+\<D( v_h- I_h u), \tau n\> \\
& \qquad \qquad -(\div\tau, D(I_h u -u))+\<D( I_h u -u), \tau n\>.
\end{align*}
Let $\tau=H(v_h) -I_h\sigma$. By the Schwarz inequality, \eqref{inverse1} and \eqref{trace-inverse}
\begin{align*}
\| \tau \|_{L^2}^2 & \leq \|\sigma-I_h\sigma\|_{L^2} \| \tau \|_{L^2}+ C \|\tau\|_{H^1}\|D(v_h - I_hu)\|_{L^2}\\
& \qquad +C \|D(v_h - I_hu)\|_{L^2(\partial\Omega)}\|\tau \|_{L^2(\partial\Omega)} + C \|\tau\|_{H^1}\|D( I_hu -u )\|_{L^2}\\
& \qquad \qquad \qquad+C \|D(I_h u -u)\|_{L^2(\partial\Omega)}\|\tau \|_{L^2(\partial\Omega)} \\
& \leq \|\sigma-I_h\sigma\|_{L^2} \|\tau \|_{L^2}+ C h^{-1} \mu \|\tau\|_{L^2} +C h^{-1} \|D(v_h - I_hu)\|_{L^2(\Omega)}\|\tau \|_{L^2(\Omega)} \\
& \qquad +C h^{-1} \|\tau\|_{L^2}\| I_hu -u \|_{H^1}\ + C h^{-\frac{1}{2}} \|D(I_hu-u)\|_{L^2(\partial\Omega)}\|\tau\|_{L^2}.
\end{align*}
Therefore
\begin{align*}
\|\tau \|_{L^2} & \leq Ch^{k+1}+ C h^{-1} \mu +Ch^{k-1}+Ch^{k-\frac{1}{2}} \\
& \leq C h^{-1} \mu +Ch^{k-1}.
\end{align*}
This proves the result.
\end{proof}
It follows from Lemma \ref{est-disc-H}, with $\mu=0$, that $(I_h u, H(I_h u)) \in B_h(\rho)$, i.e.
the ball $B_h(\rho)\neq \emptyset$ for
$\rho = C_0 h^{k}$ for a constant $C_0 >0$. See also \cite[Lemma 3.5]{AwanouLiMixed1}. As a consequence, see also \cite{Neilan2013},
\begin{equation} \label{disc-H-Ihu}
|| H(I_h u) - I_h \sigma ||_{L^2} \leq C_0 h^{k-1}.
\end{equation}
Let
$$
\tilde{B}_h(\rho) = \{ \, v_h \in V_h, v_h=g_h \, \text{on} \, \partial \Omega, ||v_h-I_hu||_{H^1} \leq \rho \, \},
$$
and consider the mapping
\begin{align*}
\tilde{T}_1: V_h \to V_h, \, \text{defined by} \, \tilde{T}_1(v_h) = T_1(v_h, H(v_h)).
\end{align*}
The motivation to introduce a discrete Hessian $H(v_h)$ in this paper, as opposed to the approach in \cite{AwanouLiMixed1}, is given by Lemma \ref{final-lem0} below.
\begin{lemma} \label{final-lem0}
If $w_h$ is a fixed point of $\tilde{T}_1$, then $(w_h,H(w_h))$ is a fixed point of $T$ and equivalently, if $(w_h,\eta_h)$ is a fixed point of $T$, then $w_h$ is a fixed point of $\tilde{T}_1$.
\end{lemma}
\begin{proof}
The result was given as \cite[Remark 3.6 ]{AwanouLiMixed1}. Let $w_h$ be a fixed point of $\tilde{T}_1$. We have $T_1(w_h,H(w_h)) = w_h$ and by \eqref{eqn.11} and \eqref{disc-H}, $T_2(w_h,H(w_h)) = H(T_1(w_h,H(w_h))) = H(w_h)$. This proves that $(w_h,H(w_h))$ is a fixed point of $T$.
Conversely if $(w_h,\eta_h)$ is a fixed point of $T$, then $\tilde{T}_1(w_h) = T_1(w_h,H(w_h)) = T_1(w_h,\eta_h) = w_h$. This completes the proof.
\end{proof}
\begin{lemma} \label{move-ball}
We have for $0 \leq \alpha \leq 1$
\begin{align}
|| \alpha I_h u - T_1( \alpha I_h u, H( \alpha I_h u) )||_{H^1} & \leq \frac{C_1}{\nu} \alpha^2 h^{k-1}, \label{u-ball}
\end{align}
for a positive constant $C_1$.
\end{lemma}
\begin{proof}
Since $T_1( \alpha I_h u, H(\alpha I_h u))- \alpha I_h u=0$ on $\partial \Omega$, by \eqref{w-eta-comp} and \eqref{eqn.2} we have using $w_h= \alpha I_h u$, $\eta_h= H(\alpha I_h u)$ and
$v=T_1(w_h,\eta_h)-w_h$
\begin{align*}
\nu (D T_1(w_h, \eta_h), D v) & = -\nu (\tr T_2(w_h, \eta_h),v)
= - \nu (\tr \eta_h,v) + (\det \eta_h - \alpha^2 f, v).
\end{align*}
It follows that
\begin{align*}
\begin{split}
\nu | D v|_{L^2}^2 & = -\nu (D w_h, D v) - \nu (\tr \eta_h,v) + (\det \eta_h - \alpha^2 f, v).
\end{split}
\end{align*}
Therefore, using \eqref{w-eta-comp2}, we get
\begin{align} \label{u1}
\begin{split}
\nu | D v|_{L^2}^2 & = (\det \eta_h - \alpha^2 f, v).
\end{split}
\end{align}
On the other hand since $f=\det D^2 u = \det \sigma$, by \eqref{mean-v} and Remark \ref{H-vh-rem}, on each element $K$
\begin{align}
\begin{split} \label{partial-est1}
\det \eta_h - \alpha^2 f & = \det H(\alpha I_h u) - \alpha^2 \det \sigma = \det \alpha H( I_h u) - \alpha^2 \det \sigma \\
& = \alpha^2 ( \det H( I_h u) - \det \sigma)\\
& = \alpha^2 (\cof(t H( I_h u) +(1-t) \sigma):(H( I_h u) - \sigma)),
\end{split}
\end{align}
for some $t \in [0,1]$.
By \eqref{interpol} we have $\|I_h\sigma\|_{L^{\infty}}\leq C\|\sigma\|_{L^{\infty}}$. Thus by \eqref{disc-H-Ihu} and \eqref{inverse0}
\begin{align*}
|| H(I_h u) ||_{L^{\infty}} & \leq || H(I_h u) - I_h \sigma ||_{L^{\infty}} + \|I_h\sigma\|_{L^{\infty}} \leq C h^{-1} || H(I_h u) - I_h \sigma ||_{L^{2}} + \|I_h\sigma\|_{L^{\infty}} \\
& \leq C h^{k-2} + C\|\sigma\|_{L^{\infty}} \leq C, \, \text{since} \, k \geq 2.
\end{align*}
Thus by \eqref{cof-est} and \eqref{disc-H-Ihu}
\begin{align*}
\|\det(H(I_h u))-\det\sigma\|_{L^2(K)} & \leq C\|t H(I_h u)+(1-t)\sigma\|_{L^{\infty}(K)}^{} \| H(I_h u)-\sigma\|_{L^2(K)} \\
& \leq C \| H(I_h u) -\sigma\|_{L^2(K)} \\
& \leq C \| H(I_h u) -I_h \sigma\|_{L^2(K)} + C \| I_h \sigma -\sigma\|_{L^2(K)} \\
& \leq C h^{k-1}.
\end{align*}
Therefore by \eqref{interpol} and \eqref{partial-est1}
\begin{align} \label{u3}
\begin{split}
\| \det \eta_h - \alpha^2 f \|_{L^2} = \alpha^2 \| \det(H(I_h u))-\det \sigma \|_{L^2}& \leq C \alpha^2 h^{k-1}.
\end{split}
\end{align}
And so combining \eqref{u1}--\eqref{u3}, \eqref{disc-H-Ihu}, Cauchy-Schwarz inequality, the interpolation error estimate \eqref{interpol} and Poincare's inequality, we get
\begin{align*}
|v|_{H^1}^2 & \leq \frac{C}{\nu} \alpha^2 h^{k-1} ||v||_{L^2} \leq \frac{C}{\nu} \alpha^2 h^{k-1} ||v||_{H^1},
\end{align*}
from which \eqref{u-ball} follows.
\end{proof}
We will need the following lemma
\begin{lemma} \label{zh-lem}
Let $(w_h, \eta_h) \in Z_h$. Then for a piecewise smooth symmetric matrix field $P$
\begin{align}
\begin{split}
((\cof P): \eta_h,v) + ((\cof P) D w_h , D v ) &
\leq C h ||v||_{H^1} ||w_h||_{H^1},
\end{split}
\end{align}
for all $v \in V_h \cap H_0^1(\Omega)$ and for a constant $C$ which depends on $||\cof P||_{H^{k+1}(\mathcal{T}_h)}$.
\end{lemma}
\begin{proof}
The proof is the same as the proof of \cite[Lemma 3.7]{AwanouLiMixed1}. There the proof was given for $P=D^2 u$, but it carries over to the general case of this lemma line by line.
The dependence of the constant $C$ on $||\cof P||_{H^{k+1}(\mathcal{T}_h)}$ arises from the use in the proof of the approximation property $||P_{\Sigma_h} (v \cof P) - v \cof P||_{H^m(\mathcal{T}_h)} \leq C h^{k+1-m} ||v \cof P ||_{H^{k+1}(\mathcal{T}_h)}$. Here $P_{\Sigma_h}$ denotes the $L^2$ projection operator into $\Sigma_h$.
\end{proof}
\begin{lemma} \label{l-inf-est-dis-H}
For $(w_h,\eta_h) \in B_h(\rho), \rho = C_0 h^k$, we have
\begin{align*}
||\eta_h-D^2 w_h||_{L^{\infty}} & \leq C h^{k-2}.
\end{align*}
\end{lemma}
\begin{proof}
Recall that for $(w_h,\eta_h) \in B_h(\rho)$, we have $\eta_h=H(w_h)$. We have by \eqref{inverse0}, \eqref{disc-H-Ihu}
\begin{align*}
||\eta_h-D^2 w_h||_{L^{\infty}} & \leq ||H(w_h)-D^2 w_h||_{L^{\infty}} \\
& \leq ||H(w_h)-I_h \sigma||_{L^{\infty}} + || I_h \sigma -D^2 w_h||_{L^{\infty}} \\
& \leq C h^{-1} ||H(w_h)-I_h \sigma||_{L^{2}} + || I_h \sigma -D^2 u||_{L^{\infty}} + || D^2 u -D^2 w_h||_{L^{\infty}} \\
& \leq C h^{k-2} + C h^{k+1} + || D^2 u -D^2 I_h u||_{L^{\infty}} + || D^2 I_h u -D^2 w_h||_{L^{\infty}} \\
& \leq C h^{k-2} + C h^{-1} ||I_h u - w_h||_{H^1} \\
& \leq C h^{k-2}.
\end{align*}
\end{proof}
The next lemma states a crucial contraction property of the mapping $T_1$ in $\alpha B_h(\rho)$.
\begin{lemma}
Let $(w_1,\eta_1),(w_2,\eta_2) \in B_h(\rho)$ with $\rho \leq \min(C_0,C_{conv}) h^k$. We have
\begin{align} \label{T-contra}
\begin{split}
|T_1(\alpha w_1, \alpha \eta_1)-T_1(\alpha w_2, \alpha \eta_2)|_{H^1} & \leq a |\alpha w_1- \alpha w_2 |_{H^1},
\end{split}
\end{align}
for $0 < a < 1$, $h$ sufficiently small, $\alpha=h^{k+2}$ and $\nu=(m+M)/2$.
\end{lemma}
\begin{proof}
Put $v=T_1(\alpha w_1, \alpha \eta_1)-T_1(\alpha w_2, \alpha \eta_2)$. By assumption $v \in V_h \cap H_0^1(\Omega)$.
Using \eqref{w-eta-comp} and \eqref{eqn.2} we obtain
\begin{align*}
\nu (D T_1(\alpha w_1, \alpha \eta_1) - D T_1(\alpha w_2, \alpha \eta_2), D v) & = - \nu( \tr T_2(\alpha w_1, \alpha \eta_1) - \tr T_2(\alpha w_2, \alpha \eta_2),v ) \\
& = -\nu (\tr \alpha \eta_1 - \tr \alpha \eta_2,v) + (\det \alpha \eta_1 - \det \alpha \eta_2,v).
\end{align*}
Therefore, using \eqref{mean-v}, we have for some $t \in [0,1]$ and with the notation
$$
Q=t \eta_1 + (1-t) \eta_2 \ \text{and} \ \tir{Q} = t D^2 w_1 + (1-t) D^2 w_2,
$$
\begin{align} \label{contraction-1}
\begin{split}
|v|_{H^1}^2 & = - (\tr \alpha \eta_1 - \tr \alpha \eta_2,v) \\
& \qquad \qquad + \frac{1}{\nu} ((\cof \alpha (t \eta_1 + (1-t) \eta_2)):\alpha (\eta_1-\eta_2),v) \\
& = \big((-I + \frac{1}{\nu} \cof \alpha Q): \alpha (\eta_1-\eta_2),v\big) \\
& = -(I: \alpha (\eta_1-\eta_2),v) - (D \alpha (w_1-w_2),D v) \\
& \ + \frac{1}{\nu} ( (\cof \alpha Q): \alpha (\eta_1-\eta_2),v) + \frac{1}{\nu} ((\cof \alpha Q)D \alpha (w_1-w_2),D v) \\
& \quad + (D\alpha (w_1-w_2),D v) - \frac{1}{\nu} ((\cof \alpha \tir{Q})D \alpha (w_1-w_2),D v) \\
& \qquad + \frac{1}{\nu} ((\cof \alpha \tir{Q})D \alpha (w_1-w_2),D v) - \frac{1}{\nu} ((\cof \alpha Q)D \alpha (w_1-w_2),D v).
\end{split}
\end{align}
For $(w_1,\eta_1),(w_2,\eta_2) \in B_h(\rho)$, $ t (w_1,\eta_1)+ (1-t) (w_2,\eta_2) \in B_h(\rho)$ and thus for $h$ sufficiently small, by Lemmas \ref{lem-1} and \ref{time-trick}
we get
\begin{equation} \label{contraction-12}
| (D(w_1-w_2),D v) - \frac{1}{\nu} ((\cof \alpha \tir{Q} )D(w_1-w_2),D v)| \leq \gamma |w_1-w_2|_{H^1} |v|_{H^1},
\end{equation}
for $0 < \gamma < 1$.
On the other hand, by Lemma \ref{zh-lem}, with $P=I$, we have
\begin{equation} \label{contraction-13}
|-(I:(\eta_1-\eta_2),v) - (D(w_1-w_2),D v)| \leq C h |w_1-w_2|_{H^1} |v|_{H^1}.
\end{equation}
Applying Lemma \ref{zh-lem}, with $P=Q$, we get
\begin{align} \label{contraction-14}
\begin{split}
|( (\cof Q):(\eta_1-\eta_2),v) + ((\cof Q)D(w_1-w_2),D v) | & \leq C h ||\cof Q||_{H^{k+1}(\mathcal{T}_h)} \\
& \quad \quad |w_1-w_2|_{H^1} |v|_{H^1}.
\end{split}
\end{align}
Finally, since by \eqref{cof-mv}
$$
\cof Q - \cof \tir{Q} = \cof(Q-\tir{Q} ) = \cof\bigg(t (\eta_1 - D^2 w_1) + (1-t) (\eta_2 - D^2 w_2) \bigg),
$$
we get using Lemma \ref{l-inf-est-dis-H}
$$
||\cof Q - \cof \tir{Q} ||_{L^{\infty}} \leq C h^{k-2} \leq C, \, \text{since} \, k \geq 2.
$$
Thus
\begin{align} \label{contraction-15}
\begin{split}
\bigg|\frac{1}{\nu} ((\cof \tir{Q})D(w_1-w_2),D v) - \frac{1}{\nu} ((\cof Q)D(w_1-w_2),D v)\bigg| & \leq C |w_1-w_2|_{H^1} \\
& \quad \qquad |v|_{H^1}.
\end{split}
\end{align}
We conclude from \eqref{contraction-1}--\eqref{contraction-15} that
\begin{align} \label{contraction-2}
|v|_{H^1} \leq (\gamma + C h + C \alpha h ||\cof Q||_{H^{k+1}(\mathcal{T}_h)} +C \alpha ) |\alpha w_1- \alpha w_2|_{H^1}.
\end{align}
Using the inverse estimate \eqref{inverse2} and noting that $\rho \leq h^2$
\begin{align*}
||\cof Q||_{H^{k+1}(\mathcal{T}_h)} & \leq C h^{-k-1} ||\cof Q||_{L^{2}} \leq C h^{-k-1} ||Q||_{L^{2}}^{} \\
& \leq C h^{-k-1} || t \eta_1 + (1-t) \eta_2 ||_{L^{2}}^{} \\
& \leq C h^{-k-1} (||\eta_1||_{L^{2}} + ||\eta_2||_{L^{2}} )^{} \\
& \leq C h^{-k-1} (||\eta_1 - I_h \sigma||_{L^2} + ||\eta_2 - I_h \sigma||_{L^2} + 2 ||I_h \sigma||_{L^2})^{} \\
& \leq C h^{-k-1} (h^{-1} \rho + || \sigma||_{L^2})^{} \leq C h^{-k-1} (C h + || \sigma||_{L^2}) \leq C h^{-k-1}.
\end{align*}
Since $\gamma < 1$, and $\alpha=h^{k+2}$, for $h$ sufficiently small, $C h + C \alpha h ||\cof Q||_{H^{k+1}(\mathcal{T}_h)} +C \alpha < 1-\gamma$. We conclude from
\eqref{contraction-2} that \eqref{T-contra} holds.
\end{proof}
\begin{lemma} \label{final-lem1}
For $\rho = \min(C_0,C_{conv}) h^k$, the mapping $\tilde{T}_1$ has a unique fixed point in $\alpha \tilde{B}_h(\rho)$ for $\alpha=h^{k+2}$.
\end{lemma}
\begin{proof}
Note that by \eqref{T-contra}, $\tilde{T}_1$ is a strict contraction in $\alpha \tilde{B}_h(\rho)$ for $\rho \leq$ $ \min(C_0,C_{conv}) h^k$. We now show that $\tilde{T}_1$ maps $ \alpha \tilde{B}_h(\rho)$ into itself.
Let $v_h \in \tilde{B}_h(\rho)$. We have by \eqref{T-contra} and \eqref{u-ball}
\begin{align*}
||\tilde{T}_1(\alpha v_h)- \alpha I_hu||_{H^1} & \leq ||\tilde{T}_1(\alpha v_h)-\tilde{T}_1(\alpha I_h u) ||_{H^1} + ||\tilde{T}_1(\alpha I_h u)- \alpha I_h u ||_{H^1} \\
& \leq a ||\alpha v_h - \alpha I_h u||_{H^1} + C_1 \alpha^2 h^{k-1} \\
& \leq a \alpha \rho + C_1 \alpha h^{2 k+ 1} = a \alpha \rho + C_1 h^{k+1} \alpha h^{ k}.
\end{align*}
Therefore for $h$ sufficiently small, $C_1 h^{k+1} \leq \min(C_0,C_{conv}) (1-a) $ and so
$$
||\tilde{T}_1(\alpha v_h)- \alpha I_hu||_{H^1} \leq a \alpha \rho + (1-a) \alpha \rho.
$$
The result then follows from the Banach fixed point theorem.
\end{proof}
We can now state the main result of this paper
\begin{thm}
Problem \eqref{m11h} has a unique local solution $(u_h,\sigma_h)$ for $k \geq 2$ and $h$ sufficiently small. We have
\begin{align*}
||u_h - I_h u||_{H^1} & \leq C h^k \\
||\sigma_h - I_h \sigma||_{H^1} & \leq C h^{k-1}.
\end{align*}
\end{thm}
\begin{proof}
Recall that for $(u_h,\sigma_h) \in B_h(\rho)$, we have $\sigma_h=H(u_h)$.
The result follows from Lemmas \ref{final-lem0}, \ref{final-lem1} and \ref{final-lem2}, the definition of $B_h(\rho)$ and \eqref{disc-H-Ihu}.
The local solution $u_h$ given by Lemma \ref{final-lem1} satisfies $||u_h - I_h u||_{H^1} \leq C h^k $. Since by Lemma \ref{final-lem0}, $(u_h,H(u_h))$ is a fixed point of $T$, by Lemma \ref{final-lem2}, $(u_h,H(u_h))$ solves \eqref{m11h}. By the definition of $B_h(\rho)$ $\sigma_h=H(u_h)$ and by \eqref{disc-H-Ihu}, we have $||\sigma_h - I_h \sigma||_{H^1} \leq C h^{k-1}$.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
We present here the first realistic \ac{GW} search
pipeline for coalescing compact binaries containing a \ac{NS}
and a \ac{BH} with spin aligned with the orbital angular momentum.
Our pipeline includes a physical template bank, signal-based vetoes and
coincidence between multiple detectors. We show that a
simple extension of traditional search methods to include the effects
of aligned spin can lead to an appreciable improvement in detection
efficiency, even during the early observational runs of advanced
\ac{GW} detectors, before they reach full sensitivity.
See \cite{Harry:2010zz,Accadia:2012zzb} for descriptions
of the advanced LIGO and Virgo detectors. In addition, the KAGRA detector
is currently under construction in Japan \cite{Somiya:2011np} and an
advanced detector has also been proposed in India \cite{LIGOIndia}.
In this paper we shall focus on neutron star--black hole (NS-BH)
binary systems which are promising sources for the advanced detectors
and pose a computational challenge. Based on our current understanding
of the population and evolution of binary systems, it is expected that
the coalescence rate for NS-BH systems within the sensitive volume of
the advanced detectors is in the range 0.2-300/year
\cite{Abadie:2010cf}. To achieve this detection rate, we must be able
to distinguish signals from noise at a matched filter \ac{SNR} of $8$
or above in the LIGO detectors. Thus we require accurate models of the
signal waveforms for matched filtering, as well as effective methods to
exclude false alarms due to non-Gaussian artifacts in the data.
With one exception \cite{Abbott:2007ai}, previous searches of initial LIGO data
did not incorporate the effect of the compact objects' angular momentum (spin)
in the waveforms used for filtering the data\footnote{However, some other LIGO
searches (see e.g.~\cite{oai:arXiv.org:1209.6533}) have quantified how well
they could detect spinning systems, even though they conducted the primary
search using waveform models without spin.}. This was because the search
methods and detector sensitivity at the time did not warrant the inclusion of
spin effects \cite{Abbott:2007ai,VanDenBroeck:2009gd}. In general, including
extra parameters such as spin in the search increases the size of the template
bank, making the search computationally more demanding and increasing the
false-alarm rate. An important question therefore is whether more accurate
waveform models can offset this increase in false-alarm rate. It has been
recently demonstrated that, for the case of \ac{BH}-\ac{BH} binaries, including
a single effective aligned-spin parameter in the search space does improve the
detection rate \cite{Privitera:2013xza}, but the question remains open for
NS-BH systems. The initial LIGO detectors had relatively low sensitivity at
higher frequencies meaning that the modified phase in the expected signal due
to spin was less visible. The situation will be different in the advanced
detector era. The advanced LIGO detectors will be able to discern the extra
features in the waveform due to the effects of spin for a significant number of
events \cite{Brown:2012qf,Harry:2013tca}. Furthermore, the increased
computational requirements for spinning searches can be met by improving the
analysis software used to process the data and exploiting modern computational
platforms such as \acp{GPU}.
The aim of this paper is twofold. First, we show
that it is indeed important to incorporate spin in searches for NS-BH \ac{CBC}
events, even in the early advanced detector era. Second, we describe a new
software package for \ac{CBC} searches known as PyCBC, which is designed to
meet the computational challenges of the advanced detector era.
Since the NS-BH merger rate is uncertain by about three orders of
magnitude, it is clear that much remains unknown about the population
of compact binary systems. A measurement of this rate would constrain
models of the formation and evolution of stellar binaries
\cite{Mandel:2009nx,O'Shaughnessy:2005qc}. NS-BH systems are also of
interest astrophysically because they (along with double \ac{NS}
systems) are expected to be progenitors of short-hard gamma-ray bursts
\cite{1989Natur.340..126E,1992ApJ...395L..83N}. A detection of NS-BH
coalescences would allow us to explore the behavior of compact objects
in the strong field regime and observation of the merger phase would
provide important information about the tidal disruption of \ac{NS}
and their equation of state
\cite{PhysRevD.81.064026,Pannarale:2011pk}. NS-BH systems have thus
been of significant interest for numerical relativity simulations
\cite{PhysRevD.77.084002,PhysRevD.79.044024,PhysRevD.78.104015,
PhysRevD.74.121503,PhysRevD.74.104018}.
The spin angular momentum of binary objects affects the intrinsic evolution
of their orbits due to spin-orbit and spin-spin couplings in the post-Newtonian
orbital energy and \ac{GW} flux. If the spin of the objects in the
binary system is not aligned with
the orbital angular momentum, then the orbits will also precess
\cite{Apostolatos:1994mx}. Searches for such precessing signals are
computationally demanding as such signals are described by many
independent
parameters.
Previous investigations of the effect of spin on
\ac{GW} searches largely focused on the precessing case,
for which a number of phenomenological search templates have been
proposed \cite{Apostolatos:1995pj,Apostolatos:1996rf,Buonanno:2002fy,Grandclement:2002dv}.
However, none of these attempts were successful when applied to real data.
Here we focus instead on the simpler problem of ``aligned-spin''
systems, where the spin angular momenta are aligned with the orbital
angular momentum.
Including the effect of aligned spins still increases the size of the
template bank. The larger number of templates
increases the number of false alarms in pure noise. False alarms from
non-Gaussian transients (glitches) triggering spinning templates had
an adverse affect on previous attempts to include spin effects in
searches \cite{VanDenBroeck:2009gd}. To counter this problem we also
include here a signal-based veto, the $\chi^2$-test
\cite{Allen:2004gu,Babak:2012zx} used in previous LIGO searches
\cite{Abbott:2009tt, oai:arXiv.org:1209.6533, Colaboration:2011np,
Abadie:2011kd, Abadie:2010yb, Abbott:2009qj, PRD.73.102002,
PRD.69.122001, PRD.72.082001, PRD.72.082002, 0004-637X-715-2-1453}.
This veto reduces the significance of glitch-induced triggers in the
search, and thus greatly reduces the threshold on signal SNR that must
be applied to achieve a desired false-alarm rate. In order to
simulate the behavior of real advanced LIGO data, we analyze two
months of real data from the two 4\,km initial LIGO detectors at
Hanford (H1) and Livingston (L1), recolored to a spectrum typical of
the sensitivity that advanced detectors are expected to achieve in
2016-2017 \cite{Aasi:2013wya}.
A comparison of searches with and without including spin effects
depends critically on the expected distribution of spin magnitudes and
orientations in the target astrophysical population. The maximum
theoretical spin for an isolated Kerr \ac{BH} is given by $\chi =
1$ where $\chi$ is the dimensionless ratio $cJ/Gm^2$ between the spin
angular momentum $J$ and the mass $m$. The maximum
value of $\chi$ to which a \ac{BH} can spin up due to accretion of
matter from a thin accretion disk is thought to be very close to this
limit \cite{Thorne:1974ve}. A number of stellar mass \acp{BH} have
been discovered using X-ray techniques. Observations suggest that many
of them have quite large spins, even close to this maximum limit \cite{McClintock:2013vwa}.
This is especially true of the \acp{BH} in high-mass X-ray binaries
whose measured spins are all above $0.85$. These high-mass systems are
the most likely to form NS-BH binaries and it is likely that the
\acp{BH} were born with these high spins since they have had insufficient
time to spin up due to accretion \cite{McClintock:2013vwa}.
Binary systems that are potential sources for advanced LIGO
are expected to have undergone a hypercritical common envelope (HCE) phase
\cite{Dominik:2012kk}. The available modeling of this phase suggests
that hypercritical accretion onto the \ac{BH} will further spin up
the \acp{BH} \cite{O'Shaughnessy:2005qc} from their spin values
before HCE. Taken in conjunction with the X-ray data, this suggests
that many of the \acp{BH} observable to the LIGO detectors will
have large spins. We test our analysis with the full range of spin values
from $-1$ to $1$ (where negative values indicate spins anti-aligned with
the orbital angular momentum) but we also display results for restricted
ranges, including a high spin range $0.7$ to $1$.
The maximum possible spin for a \ac{NS} is set by the break-up velocity,
which for expected equations of state corresponds to $\chi \sim 0.7$
\cite{Lo:2010bj}; realistic \ac{NS} values are thought to lie below this,
braked by r-mode instabilities and perhaps by \ac{GW} emission
\cite{Wagoner:1984pv,Bildsten:1998ey,Chakrabarty:2003kt}. The maximal
spin observed for accreting millisecond pulsars corresponds to $\chi \sim 0.4$,
but the maximum value observed in a binary of two compact objects is only
$\chi \sim 0.03$ \cite{Damour:2012yf}.
In NS-BH binary systems the \ac{BH} is likely to form first due to its larger
mass and is therefore unable to contribute matter to spin up the \ac{NS}.
We therefore largely ignore the spin of the \ac{NS} in NS-BH binaries
and concentrate on a single spin, that of the \ac{BH}.
While some studies suggest that an appreciable fraction of NS-BH systems may have
significant spin misalignments \cite{Belczynski:2007xg}, others suggest a
small misalignment for most systems \cite{Kalogera:1999tq}. For small
misalignments, the aligned-spin search would be close to optimal.
A closed form for the waveform of a single spin precessing system
has recently been provided \cite{Lundgren:2013jla}, but a full search based
on this method has not yet been implemented and may, as an initial step,
require an efficient single-spin aligned search similar to that presented
here. We may therefore view this investigation as the first step towards a
fully precessing search.
A complete gravitational waveform includes the merger and post-merger ringdown
signal as well as the inspiral signal. For simplicity, we ignore the merger and
ringdown part of the waveform and ignore the possibility that the \ac{NS} may
be tidally disrupted and destroyed during the inspiral phase
\cite{Foucart:2012vn}. A particular recent model for a complete waveform is
EOBNRv2, based on the Effective One-Body (EOB) framework calibrated by
numerical relativity waveforms \cite{Pan:2011gk}. Standard inspiral-only
waveforms were found to match EOBNRv2 waveforms for total masses below 11.4
solar masses for the advanced LIGO detectors \cite{Brown:2012nn}. Most of our
simulated signals have total masses below this limit, although it is not yet
known what the effect of spin is on this limit. The effect of merger and
ringdown will be studied in detail elsewhere.
The aligned-spin search pipeline employed in this paper is based on the PyCBC
software package \cite{pycbc}. PyCBC is a newly-developed toolkit for \ac{CBC} searches
in the advanced detector era written in the Python programming language.
It is based on modular software libraries: modules in isolation are quite simple,
but can be put together in useful and sophisticated ways.
PyCBC allows scientists to create complicated entire end-to-end pipelines for
performing \ac{CBC} searches. PyCBC also enables scientists to use \acp{GPU} in a
transparent manner. PyCBC builds on software from the LIGO Algorithms Library
\cite{LAL} used in previous LIGO searches.
The rest of this paper is organized as follows. Sec.~\ref{sec:pipeline}
introduces our search pipeline. Sec.~\ref{sec:pycbc} introduces the PyCBC
toolkit and the computational details of the pipeline; this section can be
read independently. The template banks are described in
Sec.~\ref{sec:templatebank} and further details of the search are in
Sec.~\ref{sec:pipelinedetails}. Sec.~\ref{sec:results} presents the main
results and Sec.~\ref{sec:conclusions} provides a summary and directions for
future work.
\section{Search method}
\label{sec:pipeline}
A fair comparison of the effects of spin in a search needs to take into
account all the details of a search of real \ac{GW} data. Therefore we
implement a prototype pipeline which can search for both spinning and
non-spinning systems. This prototype pipeline is applied to a
synthetic data set obtained by recoloring initial LIGO data as
described in \cite{Aasi:2014tra}.
We first summarize the basic matched-filtering method employed,
which is described in more detail in \cite{Allen:2005fk}.
Let $s(t)$ be the data stream from a \ac{GW} detector. Let $n(t)$ be the
noise and $h(t)$ a \ac{GW} signal which may or may not be present in the
data stream. Thus, $s(t)=n(t)$ in the absence of a signal, and $s(t)
= n(t) + h(t)$ otherwise. We denote the Fourier transform of a time
series $x(t)$ as $\tilde{x}(f)$ defined as
\begin{equation}
\tilde{x}(f) = \int_{-\infty}^\infty x(t)e^{-2\pi ift}\,dt\,.
\end{equation}
With the assumption that $n(t)$ is a stationary noise process, we define
its single-sided \ac{PSD} $S_n(f)$ as
\begin{equation}
\langle \tilde{n}(f)\tilde{n}^\star(f^\prime)\rangle = \frac{1}{2}S_n(|f|)\delta(f-f^\prime)\,.
\end{equation}
where $\langle \cdot\rangle$ denotes the expectation value over an ensemble
of noise realizations. While the assumption of stationarity is not a good
one for realistic data, this definition of the \ac{PSD} is still applicable
over short time scales. The non-stationarity is handled by continuously
estimating $S_n(f)$ from the data using a modification of the Welch method
\cite{PercivalWalden} as described in \cite{Allen:2005fk}.
The signal $h(t)$ as seen in the detector is a linear combination of
the two polarizations $h_+(t)$ and $h_\times(t)$:
\begin{eqnarray}
h(t) &=& F_+(\mathbf{n},\psi;t_0)h_+(t-t_0,\phi_0) \nonumber \\
&& + F_\times(\mathbf{n},\psi;t_0)h_\times(t-t_0, \phi_0) \nonumber \\
&=& A(t)\cos\left(\phi_0 + \phi(t-t_0)\right)\,.
\end{eqnarray}
Here the beam pattern functions $F_{+,\times}$ depend on the sky
position given by a unit-vector $\mathbf{n}$ pointing towards the
source, and on the polarization angle $\psi$ (see
e.g.~\cite{Apostolatos:1994mx}). The beam pattern functions
$F_{+,\times}$ can be taken to be constant for the duration that the
signal is seen by the detector. $t_0$ is a suitably defined arrival
time: in this case we will use an inspiral waveform described by the
post-Newtonian approximation, then a convenient choice for $t_0$ is
the termination time, such that the frequency of a signal with \ac{GW}
phase evolution $\phi(t-t_0)$ formally becomes infinite at $t_0$.
$\phi_0$ is the corresponding termination phase. In the restricted
post-Newtonian approximation, the slowly varying amplitude $A(t)$ is given by
\begin{equation}
A(t) = -\left(\frac{G\mathcal{M}}{c^2D_{\rm eff}}\right)\left(\frac{t_0-t}{5G\mathcal{M}/c^3}\right)^{-1/4}
\end{equation}
with $\mathcal{M} = M\eta^{3/5}$ being the chirp mass of the binary,
$M= m_1+m_2$ the total mass, $\eta = m_1m_2/M^2$ the symmetric mass ratio,
$D_{\rm eff} = D/\sqrt{F_+^2(1+\cos^2\iota)^2/4 + F_\times^2\cos^2\iota}$ the
effective distance, $\iota$ the angle between the line of sight from the
binary system to Earth and the orbital angular momentum, and $D$ the
distance to the binary system (see e.g.~\cite{Allen:2005fk}).
The termination time $t_0$ can be searched over by an inverse \ac{FFT}
and the search over $\phi_0$ can be handled by an analytic
maximization. As shown in \cite{Allen:2005fk}, this results in having
to compute the complex statistic
\begin{equation}
\label{eq:complexsnr}
z(t_0) = 4\int_0^\infty \frac{\tilde{s}(f) \tilde{h}^{*}(f) }{S_{n}(f)}e^{-i2\pi ft_0}\,df\,
\end{equation}
where $\tilde{h}$ is a suitably normalized inspiral waveform template
expressed in the frequency domain. The \ac{SNR} is then defined as
$\rho = |z|/\sigma$ where
\begin{equation}
\sigma^2 := 4\int_0^\infty \frac{|\tilde{h}(f)|^2 }{S_{n}(f)}\,df\,.
\label{eq:sigmasq}
\end{equation}
With this normalization, in Gaussian noise in the absence of a signal we would
have $\langle\rho^2\rangle = 2$. For practical purposes the integrations in
(\ref{eq:complexsnr}) and (\ref{eq:sigmasq}) are limited to a lower frequency
cutoff below which the detector is dominated by seismic noise and an upper
frequency cutoff beyond which the post-Newtonian waveform becomes unreliable.
This work focuses on a sensitivity curve that could reasonably represent the
early (2016) runs of advanced LIGO \cite{Aasi:2013wya} and uses a lower
frequency cutoff $f_L = 30$\,Hz. However, in Sec.~\ref{sec:templatebank} we
also investigate three different sensitivities, namely: i) the typical
sensitivity of initial LIGO during its sixth scientific run \cite{LIGO:2012aa}
with $f_L = 40$\,Hz; ii) the projected sensitivity from \cite{PhysRevD.49.2658},
used for ease of comparison with the results of \cite{Apostolatos:1996rf}, with
$f_L = 10$\,Hz; and iii) the zero-detuned, high-power design sensitivity of
the mature advanced LIGO detectors \cite{Aasi:2013wya} with $f_L = 10$\,Hz.
These curves are shown in
Fig.~\ref{fig:sensitivities}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_senscurves}
\caption{Sensitivity curves used in this work. The black solid
curve corresponds to the recolored data used for testing the search.}
\label{fig:sensitivities}
\end{figure}
The \ac{SNR} $\rho$ works well as a detection statistic in Gaussian noise. To deal
with non-Gaussian noise of realistic detectors and veto non-Gaussian transients
of non-astrophysical origin, other statistics have been developed. A widely
used signal-based veto is the reduced $\chi^2$-statistic \cite{Allen:2004gu},
which computes the partial \acp{SNR} $\rho_{\ell}$ in $p$ non-overlapping
frequency bands and combines them as
\begin{equation}
\chi_{r}^{2} = \frac{p}{2p-2}\sum\limits_{\ell=1}^p\left(\rho_{\ell}- \frac{\rho}{p}\right)^2.
\label{eq:chisquare}
\end{equation}
The bands are chosen so that a true signal with total \ac{SNR} $\rho$ would
have a partial \ac{SNR} of $\rho/p$ in each band; the union of the bands must
cover the full frequency range used to compute $\rho$. We note that computing
$\chi_r^2$ for each time sample requires $p$ inverse \acp{FFT} and is thus
computationally expensive. The exact computational method of calculating
$\rho$ and $\chi_r^2$ given a discretely sampled time series $x(t)$, the
\texttt{FindChirp} algorithm, is described in \cite{Allen:2005fk}. We
continue to use the same algorithm in this work.
In order to mitigate the effect of non-Gaussian transients that
plagued previous spinning studies \cite{VanDenBroeck:2009gd} we use a
modified detection statistic that extends the usual \ac{SNR} using the
$\chi^2$-veto, known as the \emph{re-weighted \ac{SNR}} statistic
\cite{Colaboration:2011np,Babak:2012zx}:
\begin{equation}
\hat{\rho} = \left\{
\begin{array}{ll}
\rho \left[\left(1+\left(\chi_r^2\right)^3\right)/2\right]^{-1/6} & \textrm{if } \chi_r^2 > 1 \\
\rho & \textrm{otherwise}
\end{array}
\right.
\label{eq:newsnr}
\end{equation}
We threshold on both the \ac{SNR} and re-weighted \ac{SNR} when generating
candidate events, and rank them via re-weighted \ac{SNR}; this choice was
found to be sufficient for our purposes, although it is possible that
other choices of ranking statistic would perform even better.
Our prototype search pipeline is sketched in Fig.~\ref{fig:pipeline}.
We choose ``standard'' values for most parameters in
the pipeline (such as the number of $\chi^2$ bands $p$, the coincidence windows
etc.) that have been commonly used in other searches \cite{Babak:2012zx}.
Notable differences are the coincidence method, and the use of a template bank
which is common for all detectors and fixed for the whole data set. A more
detailed tuning of the pipeline could improve the sensitivity further. The
next sections introduce the PyCBC toolkit and describe the main components of
the pipeline in detail.
\begin{figure}
\includegraphics[width=\columnwidth]{pipeline}
\caption{Flowchart of the search pipeline. Data from the Hanford
(H) and Livingston (L) detectors is processed by the main search
engine \texttt{pycbc\_inspiral} which computes the \ac{SNR} and $\chi_r^2$
time series for a common template bank which is also fixed in time.
This results in a list of unclustered single-detector triggers which
then pass through a
coincidence step followed by clustering over a suitable time
window. The single-detector triggers can additionally be clustered
independently in each detector.}
\label{fig:pipeline}
\end{figure}
\section{The PyCBC Toolkit}
\label{sec:pycbc}
Since 2004, when the first result of a \ac{CBC} search on LIGO data was
published, the bulk of the data analysis for \ac{CBC} searches has been
carried out using software from the LIGO Algorithms Library (LAL)
\cite{LAL}. This is a set of tools and applications written in the C
programming language. The computational landscape has diversified
significantly over the last 10 years. In particular, the use of \acp{GPU}
for general purpose computing is now more
widespread and even \ac{CPU} design is moving towards parallel architectures.
It is important that the software infrastructure for \ac{GW} searches is
flexible enough to keep up with this diversity.
As we saw in the previous section, the computational cost for \ac{CBC}
analyses is typically dominated by the cost of performing Fourier
transforms, primarily in computing the \ac{SNR} and the $\chi_r^2$
statistics, described earlier, for each inspiral waveform in a large
template bank. One strategy would then be to move the \ac{FFT}
calculations to \acp{GPU} and keep the remaining computations on the
usual \ac{CPU} of a computer. This would require relatively minor
modifications to the existing software in LAL and we could continue
using LAL without major modifications. While this does speed up the
analysis time somewhat, a detailed profiling of the code reveals that
the \acp{GPU} are very under-utilized and significant time is spent in data
transfers between the \ac{CPU} and the \ac{GPU}. This suggests that further development
should allow the majority of the computation to run on the \ac{GPU}.
In this section we describe a new toolkit, PyCBC \cite{pycbc},
which builds on the tools available within LAL and makes it easier to assemble
complex end-to-end pipelines, and also enables the use of \acp{GPU} in a
transparent and user-friendly manner. PyCBC is written in the Python
programming language \cite{python}, a convenient high-level scripting language
with a large user community. There are extensive collections of external
libraries in Python for a wide variety of tasks, including interfaces
to \acp{GPU} and general purpose scientific computing. The Python modules
of PyCBC need to be able to access the existing LAL software written
in C. This is important firstly because C can often be computationally
more efficient and secondly because LAL has an extensive collection of
\ac{GW}-specific functionality which has been well tested and widely used within
the LIGO and Virgo collaborations.
PyCBC uses the SWIG framework \cite{swig} to access LAL software for
\ac{CPU} computations. This enables one to perform computations within PyCBC
without sacrificing computational speed. PyCBC supports \ac{GPU} computation via
either the CUDA \cite{cuda} or OpenCL \cite{opencl} architectures, using
respectively the PyCUDA \cite{pycuda} or PyOpenCL packages \cite{pyopencl}.
An example will help us illustrate how these design choices lead to
a toolkit that is flexible and maintainable, easy for users to code in,
and transparently provides the performance capability of \acp{GPU}, while also
allowing the same code to run optimally on a \ac{CPU} when that platform is
chosen instead. A simplified script for the basic
matched-filtering operation that performs the convolution of a template with
a data segment in PyCBC is as follows:
\begin{lstlisting}
with CUDAScheme:
for data in segments:
for params in bank:
make_waveform(params, template)
template *= data
ifft(template, snr_time_series)
\end{lstlisting}
While a real code is somewhat more complex, particularly due to the
thresholding, clustering, and $\chi^2$ vetoes mentioned earlier, the above
sample code shows how the design of PyCBC achieves several important goals:
\begin{enumerate}
\item Transferring data between the \ac{CPU} and \ac{GPU} is transparent to the author of
the scripts: he or she need only perform the relevant calculations inside
the \texttt{with CUDAScheme} block (a context block in Python) and memory will
automatically be transferred as it is used in computations within the block.
In actual scripts, the context (in the example above, \texttt{CUDAScheme}) is
a variable determined at run-time, so that the same script may execute on any
\ac{CPU}, CUDA, or OpenCL platforms.
\item We leverage Python's object oriented capabilities to ``make simple things
simple.'' In the example above, the multiplication of the template by the data
requires only the single $*\!=$ operator, though in reality it represents an
element-by-element multiplication of two frequency series, which is also
transparently sanity-checked first to ensure the two series have the same
length and frequency resolution.
\item Simplicity for the user is mirrored by comparative simplicity for PyCBC
developers, because the basic PyCBC objects (vectors, time-series, and
frequency-series) leverage the uniform interface for arithmetic and basic
mathematical operations presented by Numpy (used for \ac{CPU}), PyCUDA, and PyOpenCL.
Considering
the wide variety of basic operations, many of which can have multiple instances
depending on the precision and type (real or complex) of their inputs, this
is a huge saving in development overhead, and immediately provides a functionality
not present in LAL.
\item The inverse \ac{FFT} is transparently dispatched to the appropriate library
(CUFFT \cite{cufft} for CUDA, FFTW \cite{fftw} or MKL \cite{mkl} for \ac{CPU})
which again is not written by the PyCBC developers. In the end, very little code
must be separately written for the three supported platforms; in the example
listing above, only the generation of the frequency domain waveform would be
written and maintained by the PyCBC project directly.
\end{enumerate}
As a result of this design, it is also simple to change which parts of the
computation are performed on the \ac{CPU}, and which on the \ac{GPU}. The code listing above
shows part of a script where the entire matched filter computation, and not
only the inverse \ac{FFT}, is performed on the \ac{GPU}. It therefore makes more
efficient use of the \ac{GPU} while at the same time requiring very little
additional coding.
Given the large number of templates in the spinning template bank, our
search is computationally costly; the vast majority of the cost is
represented by the matched-filtering stage, while coincidence and
clustering are comparatively trivial. The Atlas cluster \cite{atlas}
at the Albert Einstein Institute in Hannover is equipped with Nvidia
Tesla C2050 \acp{GPU} and PyCBC's flexibility allows us to accelerate
the search by running the matched-filter stage on these GPUs. Our
implementation of the \ac{CBC} matched filtering engine uses roughly 35\%
of the \ac{GPU} time as reported by the \texttt{nvidia-smi} tool
\cite{nvidiasmi}. A detailed profiling of the code, a performance
comparison between CPUs and GPUs, and further optimizations will be
presented elsewhere.
\section{Template banks}
\label{sec:templatebank}
To describe the template bank used in our search, we establish
some standard notation. The inner product between two signals $h_1(t)$ and
$h_2(t)$, also known as the \emph{overlap}, is defined as
\begin{equation}
(h_1|h_2) = 4\mathrm{Re}\int_0^\infty \frac{\tilde{h}_1(f)\tilde{h}^\star_2(f)}{S_n(f)}\,df\,.
\end{equation}
We define the normalized signal as $\hat{h}(t) := h(t)/\sqrt{(h|h)}$.
The \emph{match} between the two waveforms is defined by maximizing the
inner product between the two normalized waveforms over the time of arrival
and the phase of, say, $h_2$:
\begin{equation}
m(h_1,h_2) = \max_{t_0,\phi_0}(\hat{h}_1|\hat{h}_2(t_0,\phi_0)) \,.
\end{equation}
Consider a template bank of $N$ waveforms $h_I$ with $I=1\ldots N$
that is meant to cover a particular parameter space of masses and spins.
The \emph{fitting factor} for any waveform $h$ in the parameter space
with the template bank is defined as:
\begin{equation}
\label{eq:ff}
FF = \max_I m(h_I,h)\,.
\end{equation}
In constructing a template bank, a common requirement is that any waveform
$h$ in the target parameter space must have a fitting factor
larger than 0.97~\cite{Babak:2012zx}; thus, any waveform $h$ in the
parameter space must match some waveform in the template bank by at
least 0.97. In the actual spinning template bank employed we find that
matches can fall as low as $0.94$. This small deterioration of the
minimal match condition occurs only in a small region of parameter
space for low values of $\eta \sim 0.05$ and will not greatly affect
signals with \ac{BH} masses below $15M_{\odot}$ and \ac{NS}
masses around $1.35M_{\odot}$. The cause of these lower match values is
discussed towards the end of this section.
In this section we compare a spinning template bank with a non-spinning
template bank. Both banks are constructed using a stochastic placement
procedure that was previously presented in \cite{Brown:2012qf}; a general
introduction to stochastic template banks can be found in
\cite{Harry:2009ea,Babak:2008rb}. We use the stochastic bank algorithm
implemented within the PyCBC framework. The template waveforms use the
restricted frequency-domain TaylorF2 approximant containing 3.5 pN non-spinning
phase corrections \cite{PhysRevLett.74.3515, PhysRevLett.93.091101} and 2.5 pN
spinning phase corrections \cite{Arun:2008kb, PhysRevD.49.2658,
PhysRevD.52.821, PhysRevD.74.104034}. When calculating the matched-filter
\ac{SNR}, our template waveforms terminate at a frequency corresponding to the
\ac{ISCO} of a Schwarzschild \ac{BH} of the same total mass as the template,
i.e.~$f_{\rm ISCO} := c^3 (6\sqrt{6}\pi G M)^{-1}$. This was the standard
choice in past \ac{CBC} searches. However, in the construction of our banks,
templates are assumed to terminate at a fixed frequency of 1000 Hz, which is
close to the maximum \ac{ISCO} frequency in our parameter space. Past searches
also made a fixed-frequency assumption. Although PyCBC has the ability to
construct banks with a varying termination frequency, we do not explore the
effect of this choice in this study.
The template bank for the non-spinning search has a \ac{BH} mass $m_{\rm BH}$
ranging from 3 to 15 $M_{\odot}$ and a \ac{NS} mass $m_{\rm NS}$ ranging from 1 $M_{\odot}$
to the equal-mass boundary $m_{\rm BH} = m_{\rm NS}$. We also impose the
constraint $M \le 18$ $M_{\odot}$. Both spins are
constrained to zero. This results in $\sim 28000$ templates. The bank
for the spinning search is constructed instead with $m_{\rm BH} \in [3,15]$
$M_{\odot}$, $m_{\rm NS} \in [1,3]$ $M_{\odot}$, $\chi_{\rm BH} \in [-1,1]$ and $\chi_{\rm NS} \in
[-0.4,0.4]$. Such settings produce $\sim 150\ 000$ templates, which
turn out to be mostly clumped around extremal values of $\chi_{\rm BH}$.
Fig.~\ref{fig:massboundaries} shows the mass boundary of the two
banks. As can be seen, the non-spinning bank has a larger mass range
for the \ac{NS} than the spinning bank, in particular it includes part
of the binary \ac{BH} region. We make this choice partly because
this is how a traditional low-mass non-spinning search would be
carried out and partly to allow spinning signals to be recovered by
non-spinning templates with similar chirp mass but closer to the
equal-mass boundary, thanks to a degeneracy between spin and symmetric
mass ratio \cite{PhysRevD.87.024035}.
In other words, we explicitly favor the non-spinning search by tolerating
a bias in the recovered symmetric mass ratio. The fraction
of templates in the non-spinning bank with $m_{\rm NS} > 3 M_{\odot}$ is $\sim 6\%$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_mass_region}
\caption{Mass boundaries used in constructing our non-spinning and
spinning template banks. The non-spinning bank includes templates
with \ac{NS} masses above the usual \ac{NS} mass range. As explained in the text,
these templates are able to detect spinning NS-BH signals with a \ac{NS}
mass in the usual range.}
\label{fig:massboundaries}
\end{figure}
Fig.~\ref{fig:templatebank} shows the template density of the
spinning bank in the $(\tau_0,\tau_3)$ plane, where
\begin{eqnarray}
\tau_0 &=& \frac{5}{256\pi\eta f_0}(\pi M f_0)^{-5/3} \\
\tau_3 &=& \frac{1}{128\pi\eta f_0}(\pi Mf_0)^{-2/3} \times \nonumber \\
&& \times \left[16\pi - \frac{\chi_{\rm BH}}{6}(19\delta^2 + 113\delta + 94)\right]
\end{eqnarray}
are the \emph{chirp times} \cite{Sengupta:2003wk} extended to include
spin-orbit effects, $\delta = (m_{\rm BH} - m_{\rm NS}) / M$ and $f_0 = 20$
Hz is a fiducial frequency. The non-spinning part of the NS-BH parameter
space is shown as a black contour in the figure and the region covered
by the non-spinning bank corresponds to the black contour plus the
small area delimited by the dashed contour. As can be seen, including
the effect of spin broadens the covered region significantly. Moreover,
although the density remains approximately
constant inside the black contour, it increases noticeably outside; in
particular, a large amount of templates is concentrated at small
$\tau_3$ values. Better coordinates for representing spinning
templates in which the template density is nearly constant are given in
\cite{Brown:2012qf,Ohme:2013nsa}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_tb_density_tau0tau3}
\caption{Template density of the spinning stochastic bank in $(\tau_0, \tau_3)$
coordinates. The black contour delimits the non-spinning region of NS-BH
parameter space and the dashed lines show the additional \ac{NS} mass range
allowed by the non-spinning bank. Templates above and below the black
contour correspond to $\chi_{\rm BH} < 0$ and $\chi_{\rm BH} > 0$ respectively.}
\label{fig:templatebank}
\end{figure}
\subsection{Fitting-factor calculations}
\label{subsec:banksim}
The behavior of a template bank with respect to various signals can be studied
without the effect of detector noise by numerically evaluating the fitting
factors defined in (\ref{eq:ff}), which can be done by PyCBC. In order to get
a first characterization of the effect of spin on a few non-spinning banks
associated with the different sensitivity curves and lower-frequency cutoffs
shown in Fig.\ref{fig:sensitivities}, we calculate the fitting factors for such
banks using simulated signals with fixed masses ($m_{\rm BH} = 7.8\,M_{\odot}$ and $m_{\rm NS}
= 1.35\,M_{\odot}$) and a full range of physical spins for the heavier object ($-1
< \chi_{\rm BH} < 1$) and zero spin for the lighter object ($\chi_{\rm NS} = 0$). The signals
are simulated using the standard time-domain SpinTaylorT2 approximant available
in LALSimulation \cite{LAL}. In principle one could choose other approximants
such as SpinTaylorT1 or SpinTaylorT4, which treat
the Taylor expansions of the energy and flux differently. However, we choose
SpinTaylorT2 because it is essentially the time-domain version of our
frequency-domain templates, reducing issues related to agreement between signal
and template approximants which are outside the scope of this paper. The
waveform generation starts at 20 Hz (well outside the integration range of the
matched filter) and terminates at the \ac{MECO} after which the evolution of
the orbit is no longer expected to be adiabatic (see e.g.~\cite{Pan:2003qt}).
This choice is different from the termination condition assumed in the
construction of the template banks (1000\,Hz) as well as the upper frequency
limit used in matched filtering (the template \ac{ISCO} frequency). In reality,
a physical NS-BH waveform terminates with the merger and ringdown, typically at
frequencies higher than \ac{ISCO}, so any choice of abrupt termination of the
signal is artificial. Given that we do not consider NS-BH merger and ringdown
in this study, \ac{MECO} is a good choice both for implementation reasons and
because it is also almost always greater than \ac{ISCO}. Nevertheless, as
discussed later in this section, this discrepancy can affect the fitting
factor of binaries at high mass or high positive \ac{BH} spin. The performance
of our template banks for more realistic signal models including merger and
ringdown will be assessed in a future study.
The results for the different non-spinning banks are given in
Fig.~\ref{fig:nospinmatches}, showing similar behavior over different
sensitivity curves and lower frequency cutoffs.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_banksim_nospin_matches}
\caption{Fitting factor of SpinTaylorT2 NS-BH signals with fixed masses
($m_{\rm BH} = 7.8$ $M_{\odot}$, $m_{\rm NS} = 1.35$ $M_{\odot}$) and different \ac{BH}
spins and non-spinning template banks constructed for different
choices of sensitivity and lower cutoff frequency $f_L$
(see Fig.\ref{fig:sensitivities}). The non-spinning bank considered
in the rest of this paper corresponds to the early advanced-LIGO
sensitivity (top right plot).}
\label{fig:nospinmatches}
\end{figure}
As can be seen, in all banks there is a range of low \ac{BH} spins for which
the non-spinning bank is able to match the spinning signals fairly well, but
then a sharp fall-off in the fitting factor occurs above $|\chi_{\rm BH}| \sim 0.4$.
The shading (color online) of the points shows the recovered value of $\eta$ in
the non-spinning bank. Although the signals are simulated with $\eta=0.126$, as
the spin is increased the recovered value of $\eta$ also increases,
compensating for the larger spin. It can be seen that the sharp fall-off in the
fitting factor for positively aligned systems is associated with the maximum
physical value of $\eta=1/4$, corresponding to equal mass templates. Thus, if
we had injected signals with a different value of $\eta$, the fall-off in match
could happen at different values of $\chi$. For the equal mass case $\eta=1/4$,
for instance, we are already at the boundary and $\eta$ cannot increase any
further to compensate for the spin. The match then starts to decrease sharply
even for small positive spins. In the rest of the paper we will only consider
the early advanced LIGO sensitivity curve (top-right panel of
Fig.~\ref{fig:nospinmatches}) and references to ``the (non-)spinning bank''
will denote banks built for this case.
The loss of match at high \ac{BH} spins can be further understood by
comparing the true and recovered values of the masses in the
non-spinning bank, as is done in Fig.~\ref{fig:nospinmasses}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_banksim_nospin_masses}
\caption{Mass parameters recovered in the non-spinning bank for
SpinTaylorT2 signals with fixed masses ($m_{\rm BH} = 7.8\,M_{\odot}$,
$m_{\rm NS} = 1.35\,M_{\odot}$, dashed lines) and variable dimensionless
\ac{BH} spin (x axes).}
\label{fig:nospinmasses}
\end{figure}
The shading now shows the match and again it is clear that the
rapid fall-off in match for the positively aligned waveforms is due to
the boundary at $\eta=1/4$. The rapid fall-off in the match for
anti-aligned waveforms is due to a different effect, namely the fact
that the minimum $m_{\rm NS}$ in all template banks is 1 $M_{\odot}$. Unlike the
$\eta=1/4$ case, this is not a physical boundary and one could obtain
better matches for highly spinning anti-aligned systems by lowering
the minimum \ac{NS} mass in the template bank.
The recovered mass and spin parameters in the spinning bank are given in
Fig.~\ref{fig:spinmasses}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_banksim_spin_masses_spins}
\caption{Mass and spin parameters recovered in the spinning bank for
SpinTaylorT2 signals with fixed masses ($m_{\rm BH} = 7.8\,M_{\odot}$,
$m_{\rm NS} = 1.35\,M_{\odot}$) and variable dimensionless \ac{BH} spin
(x axes). Dashed lines show the true parameters.}
\label{fig:spinmasses}
\end{figure}
Here, as expected the templates are all well-matched, although there
is a slight bias in the recovered masses and $\chi_{\rm BH}$ values. The
recovered $\chi_{\rm NS}$ value is seen to be widely scattered and it is
clear that this does not have any significant impact on the match, nor
is its value well recovered by the bank. In other aligned-spin search
investigations \cite{Privitera:2013xza} a single effective spin
parameter was used and the minimal impact of the $\chi_{\rm NS}$ value seen
here is consistent with that approach. The fact that the matches
descend below $0.95$ despite the design choice that the bank should
have a minimal match of $0.97$ was also noted in \cite{Harry:2013tca}
and was explained there by an inconsistency between the termination
condition of template and signal waveforms.
As an overall test of the performance of the two banks over the NS-BH
parameter space, we calculate fitting factors with SpinTaylorT2
signals uniformly distributed across the parameter space. The result
is shown in Fig.~\ref{fig:overallbanksim},
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{paper_banksim_overall}
\caption{Performance of the two template banks across the whole
non-precessing NS-BH parameter space. The color scale shows
the lowest fitting factor found in each hexagonal bin (note the
different color scales). The region of poor match in the
spinning bank is likely due to the different
termination conditions of templates and test waveforms.}
\label{fig:overallbanksim}
\end{figure*}
illustrating the deficiency of the non-spinning bank over the
parameter space. It can be clearly seen that the values of $\chi_{\rm BH}$ at
which the match suddenly drops are a function of $\eta$; for equal-mass
systems the match starts to drop for $\chi_{\rm BH} \gtrsim 0$. Interestingly,
the spinning bank can have a mismatch as large as 15\% in some parts of the
parameter space. As noted already, this is likely an effect of the different
(mass- and spin-dependent) termination conditions of the template and test
waveforms. In fact, a template terminating before the signal loses the signal
power contained between \ac{ISCO} and \ac{MECO}, while in the case of a
template terminating after \ac{MECO} the \ac{SNR} normalization defined in
Eq.~\ref{eq:sigmasq} is too large; both cases result in an effective \ac{SNR}
loss. This hypothesis is supported by the fact that most of the residual
mismatch on the right plot of Fig.~\ref{fig:overallbanksim} covers the region
of $(\eta, \chi_{\rm BH})$ plane where a large difference exists between the \ac{MECO}
frequency and either the fixed termination at 1000\,Hz assumed in constructing
the bank or the \ac{ISCO} termination used for the template matched filtering.
Although our test signals and templates both use spinning phase corrections up
to 2.5 pN order, 3.5 pN corrections have been implemented during the
development of this paper and are now ready to be used in searches
\cite{Marsat:2012fn, PhysRevD.88.083002}. Unfortunately, including 3.5 pN
spinning terms in the \ac{MECO} definition can lead to very different
termination frequencies for our SpinTaylorT2 signals, introducing technical
difficulties which complicate our fitting factor calculation. Nevertheless, as
a rough characterization of the effect of 3.5 pN terms, we test our
non-spinning, 2.5 pN TaylorF2 bank against \ac{ISCO}-terminated TaylorF2
signals with 2.5 pN and 3.5 pN spinning terms. We find that the largest
variation in fitting factor when going from 2.5 to 3.5 pN signals is $\sim
0.05$, which is comparable to the maximum mismatch used for constructing the
bank and well below the loss due to neglecting spinning terms altogether. A
more detailed characterization of the inclusion of 3.5 pN spin terms represents
a separate study, but we see no reason for not using the best available phasing
in future searches.
\section{Details of the pipeline}
\label{sec:pipelinedetails}
After having described the template banks, the next step is evaluating the
performance of a full search pipeline running on realistic data. This section
describes in detail the various components of the pipeline that we implement
(see Fig.~\ref{fig:pipeline}) and the corresponding parameter choices.
\subsection{Inspiral trigger generation}
\label{subsec:trig-gen}
Strain data are first processed by a PyCBC implementation of the standard
\texttt{FindChirp} algorithm \cite{Allen:2005fk} used in previous \ac{CBC} searches.
After conditioning the data \cite{Babak:2012zx} and estimating the noise
\ac{PSD}, the \ac{SNR} time series is computed for each template with a
low-frequency cutoff of 30\,Hz. This is lower than past searches because
we are targeting the early advanced-LIGO sensitivity. Local maxima of the
\ac{SNR} time series that satisfy the condition $\rho > 5.9$ are identified.
The $\chi_r^2$-statistic is then
computed for each surviving maximum via (\ref{eq:chisquare}) using 16
frequency bands, as is typical in \ac{CBC} searches \cite{Abbott:2009tt,
oai:arXiv.org:1209.6533, Colaboration:2011np, Abadie:2011kd, Abadie:2010yb,
Abbott:2009qj, PRD.73.102002, PRD.69.122001, PRD.72.082001, PRD.72.082002,
0004-637X-715-2-1453}, and combined with the \ac{SNR} via (\ref{eq:newsnr}) to
obtain $\hat{\rho}$. In order to reduce the very large number of maxima produced
by glitches, only those with $\hat{\rho} > 5.9$ are kept as candidate triggers.
Such triggers are then stored in a MongoDB database \cite{mongodb}, where they
can be conveniently accessed by the next processing stages and also queried to
investigate the features of the data and the search.
The thresholds on \ac{SNR} and re-weighted \ac{SNR} used here are higher than
past \ac{CBC} searches (e.g.~\cite{Abbott:2009qj,Abadie:2010yb,Abadie:2011kd,
Colaboration:2011np,Babak:2012zx}) as they are chosen to fit the triggers
into the available database storage space, which is limited in our
prototype setup. There is, in principle, no technical barrier to
extending the storage space of the database to handle a larger number
of events, which would allow one to lower the threshold back to the
usual value.
Even if the noise \ac{PSD} is continuously estimated from the data in order to
evaluate the \ac{SNR}, our template banks are constructed using the early
advanced-LIGO model \ac{PSD} and thus are constant for the whole data set. This
is also a notable difference with past searches, where template banks were
regenerated on a time scale of $\sim$ 30 minutes to account for the
variability of the noise \ac{PSD}. Our choice is based on simplicity and the
relatively high computational cost of template bank generation. It is also
partly justified by the fact that the synthetic data we analyze is recolored to
the same noise curve used for constructing the banks. Although the impact of a
fixed or varying bank on the sensitivity of a search is not yet fully
understood, we expect our choice to have a small effect on the result of our
comparison.
\subsection{Coincidence and clustering}
\label{subsec:coinc}
The next stage of the pipeline is the identification of triggers in
coincidence between the Hanford and Livingston detectors. Although the
recolored strain data we analyze covers a two-month period, the different
duty cycles of the detectors reduce the amount of data analyzed
in coincidence to about 25 days.
Because we store the triggers in a centralized database, different
coincidence methods can be applied to them. We choose an exact-match
method, where a trigger in detector A can only form a coincidence with
a trigger in detector B if the two triggers share the same template,
similar to \cite{Privitera:2013xza}.
This method has the advantage of simplicity and is straightforwardly
applicable to parameter spaces of any dimensionality.
It requires however a common template bank for all detectors, another
difference with respect to past \ac{CBC} searches, as discussed above.
A systematic comparison of different coincidence methods is outside
the scope of this paper.
Considering that the maximum arrival time delay between Hanford and
Livingston is $\sim 10$\,ms, and that the uncertainty in coalescence time
is of the order of a few milliseconds \cite{Nielsen:2012sb}, we choose a
conservative coincidence window of $\pm15$\,ms. Each pair of coincident
triggers is stored in the database and is tagged with the \emph{combined}
\ac{SNR} and re-weighted \ac{SNR}, defined respectively by summing in
quadrature the single-detector \acp{SNR} and re-weighted \acp{SNR}.
In order to keep only the most representative trigger among all the
triggers produced by a single inspiral signal or glitch, a final
clustering step is performed on coincident triggers. A trigger is
defined as representative if no other triggers with higher combined
re-weighted \ac{SNR} exist within $\pm 0.5$\,s.
Finally, in order to study the distribution of single-detector false
alarms, we also perform clustering of single-detector triggers.
This works in the same way as coincidence clustering, but it
uses a window of $\pm 15$\,ms only.
\subsection{Background and sensitivity estimation}
\label{subsec:bg}
In order to estimate the sensitivity of the search, we need to
determine i) the background rate of candidates in the absence of
astrophysical signals, and ii) how well the pipeline is able to
detect simulated NS-BH signals.
As usual in \ac{CBC} searches \cite{Colaboration:2011np}, we sample the
background distribution via time slides, i.e.\ by repeating the coincidence
step many times, each with a different time delay applied to triggers
from one of the detectors. We use 800 time delays, all multiples of 5
seconds. To avoid the possibility of true signals contaminating the
background, coincident triggers with zero time delay are excluded
from the sample.
We estimate the sensitivity to a population of NS-BH binaries by simulating
each binary's gravitational waveform, adding it to the strain data, analyzing
the data and recovering coincident triggers (if any) corresponding to each
coalescence. We perform three separate analysis runs with simulated signals
spaced over the full duration of recolored data at intervals of $\sim$ 10
minutes, resulting in $\sim 3\times10^4$ signals in total. The source
population is chosen to cover the parameter space reasonably broadly, while
being astrophysically plausible. The \ac{BH} mass is assigned a Gaussian
distribution centered on 7.8\,$M_{\odot}$ with a standard deviation of 3\,$M_{\odot}$,
truncated to the $[3,12]\,M_{\odot}$ range. The mean value is motivated by
\cite{Ozel:2010su} which suggests a mass distribution $(7.8 \pm 1.2)\,M_{\odot}$
for low-mass X-ray binaries; we choose a broader distribution with the same
mean. The \ac{NS} mass is also Gaussian distributed with mean 1.35\,$M_{\odot}$
and standard deviation 0.13\,$M_{\odot}$ (following \cite{Kiziltan:2013oja}),
truncated to $[1,2]\,M_{\odot}$. Since this study ignores precession, both spins
are aligned with the orbital angular momentum. Both $\chi_{\rm BH}$ and $\chi_{\rm NS}$ are
distributed uniformly, over ranges $[-0.99,0.99]$ and $[-0.05,0.05]$
respectively; as described later, however, we also consider three subsets of
the $\chi_{\rm BH}$ range. The orbital angular momentum is distributed isotropically.
For the distance distribution, it is useful to introduce the notion of a
\emph{chirp distance}\footnote{See for instance \cite{Abbott:2009tt}, but note
that their definition uses the \emph{effective} rather than physical distance.}.
In the frequency domain, the amplitude of the signal in the restricted
post-Newtonian approximation is proportional to $\mathcal{M}^{5/6}/D$ with
$\mathcal{M}$ and $D$ being respectively the chirp mass and distance defined
in Sec.~\ref{sec:pipeline}. The chirp distance is then defined as
\begin{equation}
\mathcal{D} = D \left( \frac{\mathcal{M}_{\rm BNS}}{\mathcal{M}}\right)^{5/6}
\end{equation}
with $\mathcal{M}_{\rm BNS} \simeq 1.22$ $M_{\odot}$ being the chirp mass of a
canonical binary \ac{NS} system. This quantity conveniently absorbs all the
mass dependent terms in the amplitude: to a first approximation, the detection
efficiency should have no additional mass dependence. We then simulate a
uniform distribution of sources over chirp distance, in the interval
$[1,160]$\,Mpc. Though unphysical, the choice of uniform chirp distance ensures
that i) the efficiency-vs-distance curve is sampled accurately across its
variation from 1 to 0 and ii) the most massive sources do not dominate the
recovered sample simply because of their high mass. As for our fitting-factor
calculations in Sec.~\ref{sec:templatebank}, the signal waveforms are simulated
via the standard SpinTaylorT2 approximant from LALSimulation \cite{LAL},
starting at 20 Hz and terminating at the \ac{MECO}.
The sensitivity of the searches is estimated by applying a window around
the parameters of each source and recovering the most significant coincident
trigger within that window. Based on the results of the fitting-factor
simulations in Sec.~\ref{sec:templatebank}, we choose a coalescence-time
window of $\pm 0.5$\,s and a chirp-mass window of $\pm0.6$\,$M_{\odot}$. The
figure of merit we compute to compare the sensitivity of the two searches is
\begin{equation}
V(\rho^*) = \frac{\sum_i \mathcal{D}^2_i P_i(\rho^*)}{\sum_i \mathcal{D}^2_i}
\end{equation}
where $P_i(\rho^*) = 1$ if source $i$ is recovered with a ranking statistic
larger than $\rho^*$ and equals 0 otherwise, and $\mathcal{D}_i$ is the chirp
distance of source $i$. Here we use the quadrature sum of re-weighted SNRs
$\hat{\rho}$ over coincident triggers as ranking statistic.
The $\mathcal{D}^2$ weighting corrects the figure of merit for
the unphysical distance distribution of the simulated binaries, such that
$V(\rho^*)$ is proportional to the sensitive volume of the search, which in
turn is proportional to the expected rate of detections~\cite{Finn:1992xs}.
\section{Results}
\label{sec:results}
\subsection{Background}
Due to the increased dimensionality of the parameter space when going
from non-spinning to spinning templates, we expect a higher
false-alarm rate for the spinning search both in single-detector
triggers as well as in triggers coincident between the two detectors.
Single-detector background triggers associated with \ac{SNR} and
re-weighted \ac{SNR} are shown in Fig.~\ref{fig:background-single}.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{paper_bg_single}
\caption{Rate of single-detector false alarms for the spinning and
non-spinning searches as a function of the threshold on
\ac{SNR} and re-weighted \ac{SNR}.}
\label{fig:background-single}
\end{figure*}
The spinning search clearly has a higher false-alarm rate for both
detection statistics. As is well known from past \ac{CBC} searches
\cite{Blackburn:2008ah}, the \ac{SNR} background exhibits a large
tail associated with non-Gaussian transient glitches\footnote{Note that
in our test we analyze all available science-mode data, including a few
poor-quality data segments which a real search would exclude via
data-quality flags \cite{Slutsky:2010ff,Babak:2012zx}. Thus, the tail
in our \ac{SNR} background is likely exaggerated.}. The spinning
search seems to be affected more by glitches, as can be seen from the
much larger tail at high \ac{SNR}. Thanks to the effectiveness of the
$\chi^2$ test, however, the re-weighted \ac{SNR} is almost tail-free,
although we find that strong glitches can still lead to false
alarms noticeably stronger than what is typical in stationary Gaussian
noise. The increase of false-alarm rate associated with the re-weighted
\ac{SNR} is proportional to the increase in number of templates ($\sim 5
\times$) for almost all values of the threshold. Applying a fixed
threshold in false-alarm rate implies an increase in single-detector
re-weighted \ac{SNR} of 0.5 or less when going from the non-spinning to
the spinning search. The background distribution of
re-weighted \ac{SNR} falls approximately like $\exp(-k\hat{\rho})$ with
$k\sim 4$, such that if the total rate of triggers increases by a
factor $\alpha$, the increase in statistic threshold required to
compensate this increase (and thus preserve the same false-alarm rate)
is only $\Delta \hat{\rho} \sim \log(\alpha)/k$.
The coincident background distribution over the combined (quadrature sum)
$\hat{\rho}$ statistic is shown in Fig.~\ref{fig:background-coinc}.
As for single-detector backgrounds, the larger false-alarm rate of the
spinning search is consistent with the increase in template bank size
except at very low rate, where our background sample is likely affected
by a small number of loud glitches. Nevertheless, the increase in
ranking statistic required to maintain a fixed false-alarm rate from
non-spinning to spinning search is only about 0.3.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{paper_bg_coinc}
\caption{Rate of coincident false alarms for the spinning and non-spinning
searches as a function of the threshold on combined re-weighted
\ac{SNR}. The shaded dotted curve shows the non-spinning curve
multiplied by the relative number of templates of the two
searches ($\sim 5 \times$).}
\label{fig:background-coinc}
\end{figure}
\subsection{Signal recovery and sensitivity}
As a check of the correct behavior of the search, we first calculate the
\emph{optimal} \ac{SNR} of each simulated source, i.e.\ the \ac{SNR} obtained
for a vanishing noise realization and a perfectly matched template. For each
source producing a coincidence in both searches, we compare its
combined optimal \ac{SNR} with the combined \ac{SNR} and re-weighted \ac{SNR}
actually recovered by the searches (the $\chi^2$ value for a zero noise
realization and ideal template is also zero, thus the optimal re-weighted
\ac{SNR} is equal to the optimal \ac{SNR}). We find that the non-spinning
search fails to recover a noticeable fraction of both \ac{SNR} and re-weighted
\ac{SNR} for $|\chi_{\rm BH}| \gtrsim 0.5$, which is roughly consistent with the
fitting factor calculations (Fig.~\ref{fig:recovery}, top and middle rows).
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{paper_snr_prediction}
\caption{Top row: fraction of optimal combined \ac{SNR} and re-weighted
\ac{SNR} recovered by the non-spinning search for each simulated (and found)
source, as a function of the \ac{BH} spin (compare with
Fig.~\ref{fig:nospinmatches}). Middle row: combined \ac{SNR} and re-weighted
\ac{SNR} recovered by the non-spinning search for each found source versus the
optimal combined \ac{SNR}. The color distinguishes between high ($>0.4$) and
low ($<0.4$) \ac{BH} spin magnitude. Bottom row: same as middle row, for the
spinning search.}
\label{fig:recovery}
\end{figure*}
The impact of the $\chi^2$ test on spinning signals is particularly dramatic,
as the loss in re-weighted \ac{SNR} is much larger than the loss in \ac{SNR}.
The spinning search, instead, recovers the expected \ac{SNR} almost completely,
for all values of the \ac{BH} spin (Fig.~\ref{fig:recovery}, bottom-left plot).
Note however that sources with optimal \ac{SNR} larger than $\sim 100$ have a
significant loss in re-weighted \ac{SNR} even in the spinning bank; in fact,
the re-weighted \ac{SNR} appears to asymptote to a finite value when the
optimal \ac{SNR} becomes very large (Fig.~\ref{fig:recovery}, bottom-right
plot). This can be explained by the small but non-zero residual mismatch which
is also present in the spinning bank. In fact, with any non-zero mismatch, at
some (large) value of $\rho$ the $\chi^2$ statistic eventually starts growing
like $\rho^2$ \cite{Allen:2005fk}. Combining this fact with the definition of
re-weighted \ac{SNR} (Eq.~\ref{eq:newsnr}) results in a finite re-weighted
\ac{SNR} for arbitrarily large \ac{SNR}.
Considering the relative sensitivity of the two searches at fixed false-alarm
rate, we find that it depends strongly on the distribution of \ac{BH} spins.
Fig.~\ref{fig:rocs} shows the \ac{ROC} curves for four populations of NS-BH binaries
associated with different \ac{BH} spin distributions. As done throughout this paper,
all cases assume alignment between the \ac{BH} spin and the orbital angular momentum.
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{paper_rocs}
\caption{\ac{ROC} curves for the spinning and non-spinning searches, comparing
the relative sensitive volume (or number of detections) at fixed
false-alarm rate. The four panels assume NS-BH systems with different
limits on a uniform distribution of \ac{BH} spins.}
\label{fig:rocs}
\end{figure*}
Assuming \acp{BH} can have any spin magnitude within the limits of the Kerr
bound, we obtain an increase in sensitivity of the spinning search between 40\%
and 60\% depending on the false-alarm rate. A slightly larger improvement is
obtained if $\chi_{\rm BH}$ is restricted to be positive. If all \acp{BH} are highly
spinning and positively aligned with the orbital angular momentum, however, the
spinning search can be $\mathcal{O}(10)$ times more sensitive than the
non-spinning one at interesting false-alarm rates. This large difference can be
understood by considering the dramatic loss in re-weighted \ac{SNR} of the
non-spinning search, which is due in part to the large \ac{SNR} loss and in
part to the poor $\chi^2$ value of highly-spinning signals. If the search could
be carried out using the standard \ac{SNR} as the ranking statistic, or if we
tuned the $\chi^2$ veto differently, the improvement could be significantly
less dramatic, but likely still interesting; Fig.~\ref{fig:overallbanksim}
(left plot) and Fig.~\ref{fig:recovery} (top-left plot) both suggest a factor
of $0.6^{-3}\simeq 4.6$ when using the \ac{SNR} as the ranking statistic. We
also note that the improvement could be less dramatic if precession is included
in the simulated binaries, but this will be studied in a forthcoming paper. For
weakly spinning \acp{BH} ($|\chi_{\rm BH}|<0.4$), the spinning search is a few percent
less sensitive, as can be expected from the larger background, although the
difference is comparable with the statistical fluctuations of our \ac{ROC}
curves.
Our signals provide insufficient statistics for studying the case of very
small \ac{BH} spins. Nevertheless, we can conclude with a back-of-the-envelope
comparison of the searches assuming a worst-case population of exactly
non-spinning \acp{BH}. Using the background curves from
Fig.~\ref{fig:background-coinc} and assuming a non-spinning search with
detection threshold $\rho^{*} \gtrsim 9.5$, the relative sensitive volume of
the spinning search would be
\begin{equation}
V(\rho^{*}) \simeq \left( \frac{\rho^{*}}{\rho^{*} + 0.3} \right)^3 > 90\%.
\end{equation}
In the worst case, therefore, the spinning search would lose 10\% or less of
the signals; the major burden would be the larger computational cost.
\section{Conclusions}
\label{sec:conclusions}
We show for the first time how an aligned-spin search for NS-BH binaries can
be successfully implemented in the advanced detector era. We demonstrate a
prototype search in the PyCBC framework which contains
all the essential elements of a realistic \ac{CBC} search: matched
filtering with a template bank, signal-based vetoes, a suitable
ranking statistic, coincidence, clustering, background estimates using
time slides and sensitivity estimates via a simulated population of signals.
The use of re-weighted \ac{SNR} as event ranking statistic
\cite{Colaboration:2011np} is sufficient to reduce the background to a level
that makes the search more sensitive than traditional non-spinning
searches over the full range of \ac{BH} spins.
An important element in making this work is running the analysis on \acp{GPU} using
PyCBC. The design of PyCBC and the available tools make it easy to put
together such a pipeline, and the use of \acp{GPU} speeds up the analysis
relative to using \acp{CPU}. Such an analysis would have been much harder
with earlier technology.
Our spinning search has improved sensitivity relative to a traditional
non-spinning search. The improvement in sensitivity depends strongly on the
exact distribution of \ac{BH} spins. If the magnitude of the dimensionless
\ac{BH} spins (taken to be parallel to the orbital angular momentum) is mostly
below $\sim 0.4$, the spinning and non-spinning searches have approximately the
same sensitivity, despite the fact that the spinning template bank contains
many more templates. If the \ac{BH} spin is distributed uniformly in the range
$(-1,1)$, then the spinning search has approximately $50\%$ greater
astrophysical reach as measured by sensitive volume. The increase is one order
of magnitude if the spin is uniformly distributed in the interval $(0.7,1)$.
We stress that these results assume that i) all systems are
non-precessing, and ii) the \ac{BH} and \ac{NS} mass distributions are
Gaussians with means $7.8\,M_{\odot}$ and $1.35\,M_{\odot}$ and standard deviations
$3\,M_{\odot}$ and $0.13\,M_{\odot}$ respectively, with the additional constraints
$m_{\rm BH} \in [3,12]\,M_{\odot}$ and $m_{\rm NS} \in [1,2]\,M_{\odot}$.
Although a careful study of the effect of different mass distributions is
outside the scope of this paper, different distributions are unlikely to change
the fact that a spinning search is more sensitive than a non-spinning search;
this would require an unrealistic distribution restricted to the
parameter-space region where the non-spinning bank performs well (see
Fig.~\ref{fig:overallbanksim}).
The available X-ray data and population synthesis studies suggest that
the spin parameters of \acp{BH} may be reasonably large, greater
than $\sim 0.7$ in the mass range we used. If this is the case, then
for aligned systems an aligned spin search offers a significant
improvement in sensitive volume and hence event rate, relative to traditional
non-spinning searches. The improvement in search sensitivity could then
mean the difference between detection and non-detection, depending on
the astrophysical rate of NS-BH coalescence events.
We base our conclusions on an idealized noise \ac{PSD} which could represent
the early runs of advanced LIGO and we employ template banks fixed in time
and identical between different detectors. We also show that the \ac{SNR} loss
of a non-spinning bank relative to a spinning one depends on the noise
\ac{PSD}. If advanced LIGO's sensitivity has a significantly smaller bandwidth
than our model, or if its noise \ac{PSD} turns out to have a large variability
over a time scale of a few months, the sensitivity of a search to spin effects
could be smaller and thus our spinning search could be less beneficial. We
argue however that these are unlikely scenarios, as the evolution of advanced
LIGO will drive towards the large bandwidth of the final design sensitivity.
In addition, if the bandwidth is so narrow that spin effects are less
important than we find here, the spinning search would be at least as sensitive
as the non-spinning one for a uniform distribution of \ac{BH} spins and thus
would only produce a larger computational cost.
The search methodology we present is straightforward and based on previous
\ac{CBC} analyses, but it has not been fully optimized. We expect that
further improvements to the analysis will be possible. These include: i)
constructing a better template bank using the full 3.5 pN phasing, a geometric
placement algorithm and a mass- and spin-dependent upper frequency cutoff;
ii) correcting the event ranking statistic to reflect the non-uniform
distribution of templates over component masses and spins, as described in
\cite{PhysRevD.89.062002}; iii) improving the ranking statistic by
accounting for event distributions over extrinsic parameters such as
coalescence time and amplitude; and iv) tuning the coincidence and
clustering steps.
Including merger and ringdown effects should further improve the search
sensitivity at the higher mass end
of the parameter space. The impact of poor data quality on the
computing time of inspiral jobs deserves further research. Improved data
conditioning techniques such as gating, i.e.\ appropriately windowing
out the data segment in the vicinity of high-amplitude glitches, are
under investigation. In developing and testing such improvements to the
search, PyCBC will be an essential tool. Finally, investigating the
effects of precession will be important as well. This search should
be seen as an intermediate step towards a full precessing search; a
study is underway to quantify how well the current pipeline performs in
detecting precessing signals.
Our conclusions remain robust towards further tuning: when using a template
bank that includes the effect of spin, with a signal-based veto such as
$\chi^2$ and the infrastructure required to run a search to completion, the
gain in signal sensitivity easily outweighs the increase in background.
Thus we advise an aligned-spin search rather than a non-spinning search for
NS-BH binaries, even for the early advanced detectors.
\begin{acknowledgments}
We thank Gergely Debreczeni, Mate Ferenc Nagy, Frank Ohme, Gianluca Guidi
and Richard O'Shaughnessy for useful discussions and comments, and
Carsten Aulbert and Oliver Bock for valuable support in using the Atlas
cluster and MongoDB at the Albert Einstein Institute. We also thank the
LIGO collaboration for providing the recolored synthetic strain data we
analyzed. TDC is supported by the International Max-Planck Research
School on Gravitational-Wave Astronomy. AJM and JLW are supported
in part by the Pursuit program and Office of Undergraduate Research
of ACU. This work is supported by National Science Foundation
awards PHY-0847611, PHY-0854812, and a Cottrell Scholar award from
the Research Corporation for Science Advancement. Part of the
computations used in this work were performed on the Syracuse
University Gravitation and Relativity cluster, which is supported by
NSF awards PHY-1040231 and PHY-1104371.
This paper has LIGO document number LIGO-P1400053.
\end{acknowledgments}
|
\section{Introduction}
A complex-valued harmonic function $f=u+iv$, defined on the unit disk $\D$, is called a {\it harmonic mapping} if the coordinate function $u$ and $v$ are real harmonic, and it maps $\D$ univalently onto a domain $\symD \subset \C$. Note that it is not required that the real part and the imaginary part of $f$ are harmonic conjugate functions, i.e., satisfy the Cauchy-Riemann equations. In 1984, Clunie and Sheil-Small \cite{clunie} showed that many classical results for conformal mappings have natural analogues for harmonic mappings, and hence they can be regarded as a natural generalization of conformal mappings. Since then, this class of mappings has attracted considerable interest in complex analysis, see e.g. \cite{Dur}.
An important method for studying geometric properties of harmonic mappings is called harmonic shearing. Since its introduction in \cite{clunie}, it has been researched by many authors. For example, Greiner \cite{greiner} studied harmonic shears of conformal mappings from the unit disk onto infinite strips and other domains. Shearing of conformal mappings from the unit disk onto regular polygonal domains have been studied in the paper by Driver and Duren \cite{driver}, and by the author with Ponnusamy and Rasila \cite{pqr2}.
The shear construction makes use of a conformal mapping $\varphi$ and an analytic dilatation
$\omega$. For required assumptions for the dilatation $\omega$, see Section \ref{section: harmonic-mapping}.
An applet \cite{rolf} for exploring harmonic shears with user defined conformal mappings is written by Rolf and examples are given in \cite{dorff}. However, no accuracy tests against analytic form have been given in \cite{dorff}. Note that, the applet requires an analytic expression for the conformal mapping.
In this paper, the Schwarz-Christoffel toolbox by Driscoll \cite{driscoll} is used to provide a conformal mapping for the presented numerical method for harmonic shear. The toolbox computes a conformal mapping from the unit disk $\D$ onto a polygonal domain. See \cite{dt, trefethen} for details for the construction of a conformal mapping.
In principle, other methods can be used to obtain conformal mappings as well. An example of a numerical method, that does not involve the Schwarz-Christoffel formula, is the Zipper algorithm of Marshall \cite{Mar, MR}. A method involving the harmonic conjugate function is presented in \cite[pp. 371-374]{Hen3}.
An algorithm using the harmonic conjugate function and properties of quadrilateral is given by the author in a joint work with Hakula and Rasila \cite{hqr}. The algorithm is based on properties of the conformal modulus originating on the theory of quasiconformal mappings \cite{ahlfors,lv,ps}. The method is suitable for simply and doubly connected domains, which may have curved boundaries and even cusps. The implementation of the algorithm is based on the $hp$-FEM from \cite{hrv1}.
An important application of harmonic mappings is related to minimal surfaces. A harmonic function $f = h + \overline{g}$ can be lifted to a minimal surface if and only if the dilatation $\omega$ is the square of an analytic function. Suppose that $\omega = q^2$ for some analytic function $q$ in the unit disk $\D$. Then the corresponding minimal surface has the form
\[
\{u,v,w\} = \{\textrm{Re}\, f, \textrm{Im}\, f, 2\,\textrm{Im}\, \psi\},
\]
where
\[
\psi(z) = \int_0^z q(\zeta) \frac{\varphi(\zeta)}{1-\omega(\zeta)} \, d\zeta.
\]
The approach taken in this paper is based on the Weierstrass-Enneper representation. For further information about the relation between harmonic mappings and minimal surfaces can be found from \cite{Dur}. Minimal surfaces are closely related to many interesting phenomena in natural science and engineering, from mathematical models of soap bubble surfaces \cite{isenberg}, to topics in molecular engineering \cite{barth}, and tensile structures \cite{berger}. For a further reading on minimal surfaces and their applications see, e.g., \cite{nitsche,osserman}.
In this paper, the {\sc Matlab} visualisations of harmonic mappings are kept as close to the original images (cf. \cite{driver,pqr2}) as possible for easier comparison. Besides given illustrations, one can enhance the visualisations, e.g., by using phase portrait method \cite{wegert12, wegert13} or by using domain coloring method \cite{lundmark,poelke}.
\section{Harmonic Mappings} \label{section: harmonic-mapping}
A harmonic mapping in $\D$
has a canonical presentation $f=h+\overline{g}$, where $h$ and $g$ are analytic in $\D$ and
$g(0)=0$. A harmonic mapping $f=h+\overline{g}$ is called {\it sense-preserving}
if the Jacobian $J_f = |h'|^2-|g'|^2$ is positive in $\D$. Then $f$ has an {\it analytic dilatation}
$\omega =g'/h'$ such that $|\omega(z)| < 1$ for $z\in\D$. For basic properties of harmonic
mappings, see \cite{Dur}.
A domain $\symD \subset \C$ is said to be {\it convex in the horizontal direction} (CHD)
if its intersection with each horizontal line is connected (or empty).
A univalent harmonic mapping is called a CHD mapping if its range is a CHD domain.
Construction of a harmonic mapping $f$ with prescribed dilatation
$\omega$ can be done by the {\it shear construction} devised by
Clunie and Sheil-Small \cite{clunie}. For reader's convenience, we recall the construction along with its basic properties.
\begin{theorem}
Let $f = h+\overline{g}$ be a harmonic and locally univalent in the unit disk $\D$.
Then $f$ is univalent in $\D$ and its range is a {\rm CHD} domain if and only if $h-g$ is a conformal
mapping of $\D$ onto a {\rm CHD} domain.
\end{theorem}
\begin{proof}
See, e.g., \cite[p. 37]{Dur}.
\end{proof}
Suppose that $\varphi$ is a CHD conformal mapping. For a given dilatation $\omega$, the harmonic shear
$f=h+\overline{g}$ of $\varphi$ is obtained by solving the differential equations
\[
\left\{ \begin{split}
h'-g' &=\varphi', \\
\omega h' - g' & = 0.
\end{split} \right.
\]
From the above equations, we obtain
\begin{equation} \label{eqn: fun-h}
h(z) = \int_0^z \frac{\varphi'(\zeta)}{1-\omega(\zeta)} \, d\zeta.
\end{equation}
For the anti-analytic part $g$, we have
\begin{equation} \label{eqn: fun-g}
g(z) = \int_0^z \omega(\zeta) \frac{\varphi'(\zeta)}{1-\omega(\zeta)} \, d\zeta.
\end{equation}
Observe that $f$ can also be written as
\begin{equation} \label{eqn: fun-f}
f(z) = 2 \, \textrm{Re} \left[ \int_0^z \frac{\varphi'(\zeta)}{1-\omega(\zeta)} \, d\zeta \right] - \overline{\varphi(z)}.
\end{equation}
The last equation is useful if the conformal mapping $\varphi$ is known.
\section{Numerical Aspects}
Solving \eqref{eqn: fun-f} numerically, we shall use the change of variable $\zeta = zt$. Thus the analytic part of $f$ takes the following form
\begin{equation} \label{eqn: fun-h-int}
h(z) = \int_0^1 \frac{\varphi'(zt)}{1-\omega(zt)} z \, dt.
\end{equation}
Then for the harmonic shear $f$, we have
\begin{equation} \label{eqn: fun-f-int}
f(z) = 2 \, \textrm{Re} \left[ \int_0^1 \frac{\varphi'(zt)}{1-\omega(zt)} z \, dt \right] - \overline{\varphi(z)}.
\end{equation}
Above integrals can be computed numerically, for example, by using the Gauss quadrature.
\subsection{Gauss Quadrature}
The key idea behind the Gauss quadrature is to choose the interpolation nodes in order to maximize the degree of exactness of the quadrature rule. In the Gauss quadrature one will consider integrals of the form
\begin{equation} \label{eqn: gauss-int}
I(g) = \int_a^b g(x)\, dx = \int_a^b f(x)\eta(x)\, dx = \sum_{j=1}^N w_j f(x_j),
\end{equation}
where $\{ w_j, x_j \}_{j=1}^N$ is a quadrature rule corresponding to the weight function $\eta$. In \cite{golub-welsch}, Golub and Welsch gave algorithms to the Gauss quadrature for different weight functions. More in-depth discussion of the Gauss quadrature can be found, e.g., in \cite[Section 5.3]{gst}.
In our case, we have a singularity of the form $1/(1-\omega(z))$ caused by the dilatation. For some choices of dilatation, we can use the appropriate Gauss quadrature to deal with the singularity. However, in this article, we shall use the Gauss-Kronrod quadrature, which is also build in {\sc Matlab}. The Gauss-Kronrod quadrature is a generalization of a pure Gaussian quadrature and the method adds additional nodes to the Gauss rule with a way to control the error. It should also be noted that, the Gauss-Kronrod quadrature does not take possible singularities into account. For futher discussion of the Gauss-Kronrod quadrature, see \cite[pp. 299--300]{gst}.
\subsection{Setup of the Validation Test}
Numerical experiments are divided into two meshing of the unit disk. For the first part, the following mesh of the form $re^{i\theta}$, where $r = \{ k/20: k=0,1,2,\ldots, 20 \}$ and $\theta = \{ 2\pi k/40: k=0,1,2,\ldots, 40\}$, is used. The second mesh is refined near the boundary of the unit disk, where $r = \{ (990+k)/1000: k=0,1,2,\ldots, 10\}$ and meshing for the angle is the same as in the first mesh.
Validation of the numerical scheme is run against the analytic representation of the harmonic mappings. All conformal mappings obtained by the SC toolbox are computed with the precision setting: \texttt{precision = 1e-14}.
Note that, by using \eqref{eqn: fun-f}, one have two sources of errors. The first one comes from the integral presentation of $h$ and the second source of error comes from the conformal mapping $\varphi$ itself.
In the unit disk, the comparison of the harmonic mapping is done by using the following test function
\begin{equation} \label{eqn: test}
\textrm{test} = |f - Q(f)|,
\end{equation}
where $f$ is the analytic expression of the harmonic mapping and $Q(f)$ is given by the quadrature. Note that, by \eqref{eqn: fun-f-int}, we may write the test function as follows
\begin{equation} \label{eqn: test-by-part}
\textrm{test} = |f - Q(f)| \leq 2\,| \textrm{Re}(h - Q(h))| + |\varphi - Q(\varphi)|,
\end{equation}
where $Q(h)$ and $Q(\varphi)$ are obtained by the Gauss-Kronrod quadrature and the SC toolbox, respectively.
\section{Examples of Polygonal Mappings}
In this section, we consider polygonal examples. Let the conformal mapping be (see \cite[p. 196]{nehari})
\[
\varphi(z) = \int_0^z (1-\zeta^n)^{-2/n} \, d\zeta,
\]
which maps the unit disk $\D$ onto a regular $n$-gon. In \cite{driver}, Driver and Duren discussed the harmonic shear of $\varphi$ with the dilatation $\omega (z) = z^n$. They studied other dilatation choices as well. The author considered the dilatation $\omega (z) = z^{2n}$ in a joint work with Ponnusamy and Rasila \cite{pqr2}.
\subsection{Analytic Form}
In \cite{driver}, it was shown that for the dilatation $\omega(z)=z^n$, the harmonic shear of $\varphi$ is given by
\[ \left\{ \begin{split}
h(z) & = z F\left(1+\frac{2}{n}, \frac{1}{n}; 1+\frac{1}{n}; z^n \right), \\
g(z) & = \frac{z^{n+1}}{n+1} F\left(1+\frac{2}{n}, 1+\frac{1}{n}; 2+\frac{1}{n}; z^n \right),
\end{split} \right.
\]
where $F(a,b;c;z)$ if the Gaussian hypergeometric function. The function is defined as follows
\[
F(a,b;c;z) = 1 + \sum_{n=1}^\infty \frac{(a)_n (b)_n}{n! (c)_n} z^n, \quad |z| < 1,
\]
where
\[
(\alpha)_n = \alpha (\alpha+1) \cdots (\alpha+n-1) = \frac{\Gamma(\alpha+n)}{\Gamma(\alpha)}, \quad \alpha \in \C,
\]
is the Pochhammer symbol. For $\textrm{Re}\, c > \textrm{Re}\, b > 0$, this can also be written as the Euler integral
\[
F(a,b,c;z) = \frac{\Gamma(c)}{\Gamma(b)\Gamma(c-b)} \int_0^1 t^{b-1} (1-t)^{c-b-1} (1-zt)^{-a} \, dt.
\]
In case of $\omega(z)=z^{2n}$, the harmonic shear is shown, in \cite{pqr2}, to be
\[ \left\{ \begin{split}
h(z) & = z F_1\left( \frac{1}{n}, 1+\frac{2}{n}, 1; 1+\frac{1}{n}; z^n, -z^n \right), \\
g(z) & = \frac{z^{2n+1}}{2n+1} F_1\left( 2 +\frac{1}{n}, 1 +\frac{2}{n}, 1; 3 +\frac{1}{n}; z^n, -z^n \right),
\end{split} \right.
\]
where $F_1(a,b_1,b_2;c;x,y)$ is the first Appell hypergeometric function \cite[p. 73]{bailey}, which is defined by
\[
F_1(a,b_1,b_2;c;x,y) = \sum_{k=0}^\infty \sum_{l=0}^\infty \frac{(a)_{k+l} (b_1)_k (b_2)_l}{(c)_{k+l} \,k!\, l!} x^k y^l,
\]
Appell hypergeometric functions can be given by Euler's integral as follows \cite[p. 77]{bailey}:
\[
F_1(a,b_1,b_2;c;x,y) = \frac{\Gamma(c)}{\Gamma(a)\Gamma(c-a)} \int_0^1 t^{a-1} (1-t)^{c-a-1} (1-xt)^{-b_1} (1-yt)^{-b_2} \, dt,
\]
where $\textrm{Re}\, c > \textrm{Re}\, a > 0$.
\subsection{Example}
Let $\omega(z)=z^{2n}$, $n=4$. The reason we chose this dilatation is that, in \eqref{eqn: fun-h} we have eight singularities and it reveals the accuracy of our algorithm and as well as the accuracy of the SC toolbox. In Figure \ref{fig: sc-poly4-dila-z8}, we have reproduced the figures from \cite{pqr2} and numerically computed images using the presented algorithm.
\begin{figure}[!ht]
\begin{center}
\subfloat[Visualisation of an analytic form given by Mathematica.]{\parbox{\textwidth}{\centering\includegraphics[width=0.43\textwidth]{kuva/conf-polygon-4} \hspace*{.3cm}
\includegraphics[width=0.43\textwidth]{kuva/harmonic-polygon-4}}} \\
\subfloat[A numerical version computed by {\sc Matlab}. Note that, the radius of the outermost circle is chosen to be 0.99, to avoid extensive amount of visual artefacts.]{\parbox{\textwidth}{\centering\includegraphics[width=\textwidth]{kuva/conf-polygon-4-sc} \hspace*{.3cm}}} \\
\caption{Conformal mapping $\varphi$ of the unit disk $\D$ onto a square and its harmonic shears with the dilatation $\omega(z) = z^{8}$.} \label{fig: sc-poly4-dila-z8}
\end{center}
\end{figure}
In Figure \ref{fig: sc-poly-dila-z8-error-f}, we have the error of the test function \eqref{eqn: test} of the harmonic shear $f$ given in the logarithmic scale (with base 10). Note that, the error is given in the pre-image form, so that the figure can be shown in a more compact manner. The error corresponding to the second mesh can be given in polar coordinates as well, but the readability of the figure would be lost due to the poor resolution. Therefore, the error is given in Cartesian coordinates.
White areas in illustrations correspond to the loss of accuracy at singularities due to roots of the $1-z^8$. In the second mesh, near the boundary, we see that the rapid loss of accuracy occurs at the neighbourhood of the singularities.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.45\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-f-2014-04-03} \hspace*{.2cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-f-2014-04-04-reuna} \\
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03} \hspace*{0.3cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03-reuna}
\caption{Error analysis of the harmonic shear of conformal mapping $\varphi$ of the unit disk $\D$ onto a square with the dilatation $\omega(z) = z^{8}$. On the left hand side, we have the error of the first mesh, and on the right hand side, we the corresponding error near the boundary of the unit disk. Results are obtained by taking the logarithm (with base 10) of the test function \eqref{eqn: test}.} \label{fig: sc-poly-dila-z8-error-f}
\end{center}
\end{figure}
In Figure \ref{fig: sc-poly-dila-z8-error-h-phi}, we have errors corresponding to right hand side of the \eqref{eqn: test-by-part}. The conformal mapping itself has 4 singularities arising from the roots of $1-\zeta^4$. The loss of accuracy is around 4 digits in this particular example, which is nuisance but a not severe obstacle.
Unfortunately, for the analytic part $h$ of $f$, even the leading digit is wrong at singularities most of the time. This can be improve if the type of singularities are known and the quadrature is adapted to take this information into account. Elsewhere the performance shows no significant shortcomings.
\begin{figure}[!ht]
\begin{center}
\subfloat[Error of the analytic part $h$ of $f$ in form of $2\,| \textrm{Re}(h - Q(h))|$ .]{\parbox{\textwidth}{\centering\includegraphics[width=0.45\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-h-2014-04-03} \hspace*{.2cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-h-2014-04-07-reuna} \\
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03} \hspace*{0.3cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03-reuna} }} \\
\subfloat[Error analysis of the conformal mapping $\varphi$ from the unit disk onto a square.]{\parbox{\textwidth}{\centering\includegraphics[width=0.45\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-phi-2014-04-07} \hspace*{.2cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-phi-2014-04-07-reuna} \\
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03} \hspace*{0.3cm}
\includegraphics[width=0.43\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-03-reuna} }}
\caption{Error analysis of the analytic part $h$ of $f$ and the conformal mapping $\varphi$ in the case of the harmonic shear of conformal mapping $\varphi$ of the unit disk $\D$ onto a square with the dilatation $\omega(z) = z^{8}$. Results are obtained by taking the logarithm (with base 10) of the corresponding part of the test function \eqref{eqn: test-by-part}.} \label{fig: sc-poly-dila-z8-error-h-phi}
\end{center}
\end{figure}
\subsection{Polygonal Shears with Other Dilatations}
We shall reproduce the images from \cite{driver} for the dilation $\omega(z)=z^n$ and from \cite{pqr2} using dilatition $\omega(z)=z^{2n}$ with our numerical algorithm along with the error analysis. In Figure \ref{fig: sc-triangle-dila-z3-z6}, we have $n=3$. For illustrastions for $n=5$, see Figure \ref{fig: sc-pentagon-dila-z5-z10}.
\begin{figure}[!ht]
\begin{center}
\subfloat[Dilatation $\omega(z)=z^3$.]{\parbox{\textwidth}{\centering\includegraphics[width=0.44\textwidth]{kuva/harmonic-triangle-z3-sc-2014-04-14} \hspace*{.1cm}
\includegraphics[width=0.4\textwidth]{kuva/vertailu-sc-triangle-dila-z3-f-2014-04-14} \hspace*{.1cm}
\includegraphics[width=0.1\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-14-vertical} }} \\
\subfloat[Dilatation $\omega(z)=z^6$.]{\parbox{\textwidth}{\centering\includegraphics[width=0.44\textwidth]{kuva/harmonic-triangle-z6-sc-2014-04-22} \hspace*{.1cm}
\includegraphics[width=0.4\textwidth]{kuva/vertailu-sc-triangle-dila-z6-f-2014-04-22} \hspace*{.1cm}
\includegraphics[width=0.1\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-14-vertical} }}
\caption{Reproduction from \cite{driver} and \cite{pqr2} of the conformal mapping $\varphi$ of the unit disk $\D$ onto a triangle and its harmonic shears with dilatation $\omega(z) = z^3, z^6$, respectively. The error is given in the logarithmic (base 10) scale. Note that, the radius of the outermost circle is chosen to be $0.99$, to avoid extensive amount of visual artefacts.} \label{fig: sc-triangle-dila-z3-z6}
\end{center}
\end{figure}
\begin{figure}[!ht]
\begin{center}
\subfloat[Dilatation $\omega(z)=z^5$.]{\parbox{\textwidth}{\centering\includegraphics[width=0.44\textwidth]{kuva/harmonic-pentagon-z5-sc-2014-04-24} \hspace*{.1cm}
\includegraphics[width=0.4\textwidth]{kuva/vertailu-sc-pentagon-dila-z5-f-2014-04-24} \hspace*{.1cm}
\includegraphics[width=0.1\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-14-vertical} }} \\
\subfloat[Dilatation $\omega(z)=z^{10}$.]{\parbox{\textwidth}{\centering\includegraphics[width=0.44\textwidth]{kuva/harmonic-pentagon-z10-sc-2014-04-24} \hspace*{.1cm}
\includegraphics[width=0.4\textwidth]{kuva/vertailu-sc-pentagon-dila-z10-f-2014-04-24} \hspace*{.1cm}
\includegraphics[width=0.1\textwidth]{kuva/vertailu-sc-poly4-dila-z8-parempi-SC-legend-2014-04-14-vertical} }}
\caption{Reproduction from \cite{driver} and \cite{pqr2} of the conformal mapping $\varphi$ of the unit disk $\D$ onto a pentagon and its harmonic shears with dilatation $\omega(z) = z^5, z^{10}$, respectively. The error is given in the logarithmic (base 10) scale. Note that, the radius of the outermost circle is chosen to be $0.99$, to avoid extensive amount of visual artefacts.} \label{fig: sc-pentagon-dila-z5-z10}
\end{center}
\end{figure}
\section{Minimal Surfaces}
It is known that a harmonic function $f = h + \overline{g}$ can be lifted to a minimal surface if and only if the
dilatation $\omega$ is the square of an analytic function. Suppose that $\omega = q^2$ for some analytic
function $q$ in the unit disk $\D$. Then the corresponding minimal surface has the form
\[
\{u,v,w\} = \{\textrm{Re}\, f, \textrm{Im}\, f, 2\,\textrm{Im}\, \psi\},
\]
where
\[
\psi(z) = \int_0^z q(\zeta) \frac{\varphi(\zeta)}{1-\omega(\zeta)} \, d\zeta.
\]
In \cite{driver}, Driver and Duren computed $\psi$ for a conformal mapping $\varphi$ from unit disk onto a $n$-gon with dilatation $\omega(z)=z^n$. Computation is done with assumption that $n$ is even. In this case the minimal surface lifting is given by
\[
\psi(z) = \frac{2 z^{1+1/n}}{n+2} F\left(1+\frac{2}{n}, \frac{1}{2}+\frac{1}{n}; \frac{3}{2}+\frac{1}{n}; z^n \right).
\]
In Figure \ref{fig: square-minimal}, we have a illustration for $n=4$ given by the exact solution and the numerical scheme.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=0.4\textwidth]{kuva/minimal-mathematica-driver-duren-2014-04-28} \hspace*{.1cm}
\includegraphics[width=0.4\textwidth]{kuva/minimal-sc-driver-duren-n4-2014-04-28}
\caption{Minimal surface of the conformal mapping $\varphi$ of unit disk onto $n$-gon with a dilatation $\omega(z)=z^n$, $n=4$. On the left hand side, we have a illustration given by the exact solution. In comparison, on the right hand side, we have a numerically computed version of the minimal surface. To give a sensible illustration, we have chosen the outermost circle's radius to be $0.8$. This example is a reproduction from \cite{driver}. } \label{fig: square-minimal}
\end{center}
\end{figure}
\section{Conclusions}
In this article, we have given an algorithm to numerically shear CHD conformal mappings. Required integrations are done using a standard Gauss-Kronrod quadrature. From given examples against analytic expression, we found that our numerical method's performance is satisfactory on all mesh points we have considered except singularities. The accuracy can be improved if the type of the singularity is known and the algorithm is tweaked to work around it.
|
\section{Introduction}
The control of the dynamics of spins in solid state materials has direct implications at fundamental and applied levels. Research topics, in particular quantum computing, rely heavily on complex control techniques and long spin coherence times to achieve robust information control. In this context, molecular magnets have gained significant attention in recent years, both for their technological potential and as a test bed for quantum mechanics at a macroscopic scale. The compound K$_6$[V$^{IV}_{15}$As$^{III}_6$O$_{42}$(H$_2$O)] $\cdot$ 8H$_2$O, in short V$_{15}$, has a magnetic cluster anion \cite{Mueller_Angew1988} [Fig.~\ref{fig1}(a)] and it was among the first to show coherence via direct measurements of Rabi oscillations \cite{Sylvain,Shim} as well as interesting out-of-equilibrium spin dynamics due to phonon bottlenecking \cite{IC1,IC2,IC3,Lei1,Lei2}. The observation of spin coherence in molecular magnets is currently not limited to V$_{15}$; it has also recently been studied in systems like Cr$_7$Ni, Fe$_8$, and Fe$_4$, showing the high interest that these systems have gained\cite{Takahashi,Schlegel,Ardavan}. In this work, we study the spin Hamiltonian of V$_{15}$ by means of electron spin resonance (ESR) on a precisely oriented single crystal. We study the Dzyaloshinsky-Moriya (DM) coupling between a molecule's spins as well as the anisotropy of their exchange coupling. Finally, we find that the anisotropy of the $g$-tensor shows an interesting temperature dependence which may indicate the onset of short-range correlations as seen in other systems \cite{Lawrence}. These Hamiltonian parameters are given here with a high degree of precision. The spin Hamiltonian is instrumental to obtain the transition probabilities between certain levels of this multi-level quantum system and thus to predict the spin dynamics at low temperatures.
\begin{figure}
\includegraphics[width=3.25 in]{fig1.eps}
\caption{(Color online) (a) Ball-and-stick representation of the V$_{15}$ cluster anions, with the $s=1/2$ V ions in green. The $x$ axis is parallel to one of the triangle's sides, and the $z$ axis represents the $c$ axis of the crystal unit cell. (b) Approximate sketch of the energy levels in a field corresponding to a resonant frequency of 120 GHz. Arrows represent the allowed transitions and the numbers are the $S_z$ values. The dashed blue lines indicate the two $S=1/2$ doublets, and the solid red lines show the $S=3/2$ quartet. The resonance fields are averaged in Eq.~(\ref{B1}). (c) V$_{15}$ crystal with markings used to find the $c$ axis in a two-step process (see text). }
\label{fig1}
\end{figure}
The V$_{15}$ cluster anions are arranged in a crystal with $R\bar 3c$ symmetry. Each V ion has a spin $1/2$, and the 15 spins are arranged in two hexagons sandwiching a triangle. The couplings between ions V$_1-\textrm V_2$ and V$_3-\textrm V_{1,2}$ are antiferromagnetic and are of the order of hundreds of degrees kelvin \cite{Gatteschi,Barra}. The coupling $J$ between the triangle's corners is also antiferromagnetic \cite{IC2} and can be indirect (via the two hexagons). Therefore, the effective Hamiltonian of the molecule corresponds to that of three spins $s=$1/2 coupled antiferromagnetically \cite{IC2,IC3}. A more precise form of the Hamiltonian is
\begin{multline}
H=\sum_{<i,j>}[\vec D_{ij}(\vec S^{(i)}\times\vec S^{(j)})+ J_zS_z^{(i)}S_z^{(j)}+ J_t(S_x^{(i)}S_x^{(j)}\\
+S_y^{(i)}S_y^{(j)})]+g_a\mu_B\vec B_\triangle \vec S_\triangle + g_c\mu_B\vec B_z \vec S_z
\label{H1}
\end{multline}
where one sums over the $(i,j)$ pairs $(1,2),(2,3),(3,1)$ of triangle spins $\vec S^{(1),(2),(3)}$. The first term represents the DM coupling, and the second one shows the antiferromagnetic coupling between the triangle's corners which, in principle, can be anisotropic, $J_z\ne J_t$. The Zeeman part shows the anisotropy of the $g$ tensor, with $g_{a,c}$ being the values in the triangle's plane and along the molecule's $c$ axis [$z$ axis in Fig.~\ref{fig1}(a)] respectively; $\vec B_{\triangle,z}$ are the planar and vertical vector components of the external magnetic field, respectively, $\vec S_{\triangle}=\vec S_x+\vec S_y= \sum_{i=1,2,3}(\vec S^{(i)}_x+\vec S^{(i)}_y) $, $\vec S_{z}=\sum_{i=1,2,3}\vec S^{(i)}_z$, and $\mu_B$ is the Bohr magneton. The resulting eigenvalues describe an excited quartet $S=3/2$ and two ground doublets $S=1/2$. The DM coupling, although usually considered to be very small, can be sufficient to lift the degeneracy of the two doublets and, in second order, the zero-field degeneracy of the $S=3/2$ quartet (see theoretical Refs. \onlinecite{Tsukerblat_JCP2006,Boris2,Tarantul,Machida,Seiji,Uchiyama} for more details).
\section{Experimental methods}
The continuous-wave ESR experiments were performed at 120 GHz on the quasioptical superheterodyne spectrometer at the National High Magnetic Field Laboratory \cite{Morley,van Tol}. The magnet used is a sweepable 12.5-T superconducting magnet with a homogeneity of $10^{-5}$ over 1 cm$^3$. The experimental setup allows for a continuous change of the angle $\theta=\arctan(B_\triangle/B_z)$ between the $c$ axis of the molecule and the static field $\vec B_0=\vec B_\triangle +\vec B_z$. The temperature can be varied from room temperature down to 2.5~K. The field is scanned and when the induced Zeeman splittings are equal to 120~GHz, resonant absorption of the microwaves is observed in the power reflected by the crystal. The measured resonant fields are compared against values computed using the method of first moments \cite{Tsukerblat_JCP2006}. There are three transitions within the quartet [shown by red arrows in Fig.~\ref{fig1}(b)] and four within the two doublets whose fields are found by exact diagonalization of the three-spin Hamiltonian (\ref{H1}). Using transition probabilities as weights, these fields are thermally averaged to produce a temperature-dependent resonance field:
\begin{equation}
B=\sum_{i=1}^{7}I_iB_i/\sum_{i=1}^{7}I_i
\label{B1}
\end{equation}
with
\begin{equation}
I_i=\frac{|\langle b|\vec S_\phi|a\rangle|^2}{Z(B_i)}\left[\exp\left(-\frac{E_a}{k_BT}\right)- \exp\left(-\frac{E_b}{k_BT}\right)\right].
\end{equation}
Here $Z(B_i)$ is the partition function over the eight levels at fixed field $B_i$ and $E_{a,b}$ are eigenvalues for the states $|a\rangle$ and $|b\rangle$ involved in a given transition, that is, $E_a-E_b=hF_{mw}$, with $F_{mw}=$120~GHz and $k_B$ and $h$ being the Boltzmann and Planck constants, respectively. The operator $\vec S_\phi=\vec S_x\sin\phi-\vec S_y\cos\phi$ represents the microwave drive $\vec h_{mw}\perp \vec B_0$ (here $\phi$ is the angle between $\vec B_\triangle$ and the $x$ axis).
The experiments are performed on a single crystal, shown in Fig.~\ref{fig1}(c), with typical, well-defined facets. Such facets are not unusual among $R\bar 3c$ compounds; for example, calcite has a similar exterior aspect. It is of essence to be able to correctly identify the $c$ axis of the unit cell (in hexagonal representation \cite{Barra}) since the V single-site anisotropy is translated at the level of the entire molecule \cite{Gatteschi}. Thus, the $g$ factor measured along the $c$ axis has the maximum value ($g_c=g_\perp$), while in the triangle plane, it is averaged between the single-site directional values $g_{\perp,||}$ [$g_a=\sqrt{(g_\perp^2+g_{||}^2)/2}$, see Ref.~\onlinecite{Gatteschi} for details]. This information allows us to find the $c$ axis through a process of rotating the crystal in the applied magnetic field and recording the ESR spectrum (reflected power vs field) at each rotational step $\theta$. After a full rotation, the crystal is extracted and repositioned on the rotator stage. For each new position, the resonance field vs $\theta$ shows an oscillatory behavior [like in Fig.~\ref{fig2}(a)] with a certain amplitude. The repositioning is iterated until a maximum amplitude is obtained. The $c$ axis is parallel to the applied field when the minimum resonance field occurs, and $\theta$ is thus calibrated as $\theta=0^\circ$ with an uncertainty of $\sim 1^\circ$. Figure~\ref{fig1}(c) shows the final location of the $c$ axis. Since the crystal shape is common among V$_{15}$ crystals, the result can be generalized to quickly find the $c$ axis using a two-step process followed by fine-tuning using the iterative procedure above. In step 1, one looks straight down at the crystal, and a line is drawn at a 2.6$^\circ$ angle to the short diagonal clockwise and along the top face plane. The pivot point of this rotation must be the corner that overhangs the bottom of the crystal [shown with white dots in Fig.~\ref{fig1}(c)]. In step 2, the line is rotated 42$^\circ$ directly into the crystal; this rotation is exactly orthogonal to the first one. The final direction, shown by the dashed line starting at the overhanging corner of the short diagonal, is parallel to the $c$ axis of the unit cell.
\begin{figure}
\includegraphics[width=3.37 in]{fieldrotation.eps}
\caption{(Color online) (a) Angular dependence of the resonance field as a function of the angle $\theta$ with the $c$ axis, at four different temperatures. The solid lines show fits using Eq.~(\ref{B1}). (b) Example of ESR spectrum showing the first derivative of the absorption signal (in arbitrary units, a.u.) of the calibration standard (BDPA) and the molecule at 10~K. The dashed line shows the fit used to obtain the resonance field. (c) Variation of the minimum $B_{min}$ ($\theta=0^{\circ},B_0||c$) and maximum $B_{max}$ ($\theta=90^{\circ},B_0$ in the triangle plane) of the resonance fields with temperature.}
\label{fig2}
\end{figure}
Using the procedure described above, one can perform $\theta$ rotations in a plane containing the $c$ axis and identify the resonance field for each angle. Also, the magnetic field is found by using the well-known ESR calibration standard 1,3-bisdiphenylene-2 phenylallyl (BDPA \cite{bdpa}, $g=2.00263$) placed on the same sample holder as the V$_{15}$ crystal. The measurements are repeated for temperatures ranging from 3 to 90 K. The data shown in Fig.~\ref{fig2}(a) correspond to three such temperatures (dots), while the continuous fits are numerical simulations allowing one to identify the Hamiltonian parameters (see below). The resonance field shows a minimum $B_{min}$ when the field is $||c$ and reaches a maximum $B_{max}$ when the field is in the triangle ($\triangle$) plane. A typical ESR spectrum is shown in Fig.~\ref{fig2}(b) as the first derivative of the absorption signal vs field at 10~K. At a lower field one observes the signal from the BDPA standard, while the V$_{15}$ signal is larger and placed at higher fields. The dashed red line shows the fit used to extract the resonance fields. The fit uses a typical ESR model which sums up the absorption and dispersion components of the signal\cite{Poole} since there is no automatic frequency control to cancel the dispersion part. Note that the width of a typical ESR line is of the order of few tens of milliteslas (most likely due to hyperfine and dipolar interactions), thus allowing a precise fitting for resonance field identification with an uncertainty of $\sim$0.5~mT. The temperature dependence of $B_{min}$ and $B_{max}$ is shown in Fig.~\ref{fig2}(c) and is an essential result used below to discuss the $J$ and DM couplings.
\begin{figure}
\includegraphics[width=3.37 in]{Janisotropy.eps}
\caption{(Color online) The resonance fields along the c-axis ($B_{min}$) and $\perp$ to it ($B_{max}$) as a function of temperature for an isotropic J coupling (dashed line) and two anisotropic situations, $J_z$ larger (dotted line) and smaller (line) than $J_t$.}
\label{fig3}
\end{figure}
\section{Exchange coupling anisotropy}
In Hamiltonian~(\ref{H1}), the antiferromagnetic coupling between spins must lead to a spin crossover \cite{IC2} at 2.8~T when the ground state switches between $S=1/2, S_z=-1/2$ and $S=3/2, S_z=-3/2$. To account for a potential anisotropy of this coupling, we consider $J_t=-2.45$~K and the ratio $J_z/J_t$ to be variable (the conclusions of this discussion are the same if we fix $J_z$ instead of $J_t$). We can observe that such anisotropy can, in principle, generate a zero-field splitting for the quartet state $S=3/2$. Indeed, by writing this part of the Hamiltonian as $J_t\vec S^{(i)}\vec S^{(j)}+ (J_z-J_t)S_z^{(i)}S_z^{(j)}$ one can observe that $J_z-J_t$ plays a role similar to that of a crystal field term $DS_z^2$, thus potentially raising the degeneracy between the $S_z=\pm 3/2$ and $\pm 1/2$ states of the quartet. Using Eq.~(\ref{B1}), we simulate the temperature dependence of $B_{min,max}$ for some typical values of the $g$ factors \cite{Sylvain}, namely, $g_a=$1.95 and $g_c=$1.98. The qualitative behavior shown in Fig.~\ref{fig3} is compared to that in Fig.~\ref{fig2}(c) and we find that for $J_z < J_t$ ($J_z>J_t$), $B_{min}$ ($B_{max}$) contradicts the experimental behavior by increasing instead of decreasing with increasing temperature. If the ratio $J_{z}/J_{t}$ is sufficiently large, a crossing between $B_{max}$ and $B_{min}$ is observed. A crossing is measured in V$_{15}$ at very high temperatures (200~K) \cite{Ajiro} when all 15 spins of the molecule contribute to its magnetism. In contrast, our study deals with the low-temperature regime where the three-spin model is adequate. Such a crossing would imply a large increase in amplitude $B_{max}-B_{min}$ at very low temperatures, whereas, experimentally, the amplitude is relatively constant. Consequently, to properly explain the data we consider $J_{z,t}=-2.45$~K (dashed line).
\begin{figure}
\includegraphics[width=3.37 in]{DManisotropy.eps}
\caption{(Color online) The resonance fields along the $c$ axis ($B_{min}$) and perpendicular to it ($B_{max}$) as a function of temperature for different values of the DM vector.}
\label{fig4}
\end{figure}
\section{Dzyaloshinsky-Moriya interaction}
The DM term of Hamiltonian~(\ref{H1}) can mix states with different total spins and can lift the degeneracy of the two doublets. The DM vector components are considered to be the same for all three spin pairs and are defined as in Ref. \onlinecite{Tsukerblat_JCP2006}, with $D_z$ being the vertical component and $D_t$ being the effective in-plane DM interaction incorporating both components, perpendicular and along the side of the triangle. For instance, the spin crossing at 2.8~T can show a level repulsion if a $D_t$ term is present, while a $D_z$ term is effective in raising the doublets degeneracy. Consequently, there is a potential to split and identify all the resonance fields and their transition probabilities that enter Eq.~(\ref{B1}).
We studied the effect of the DM coupling on the value of the resonance fields and the results are presented in Fig.~\ref{fig4} for $D_t=0, 0.1, 0.35$~K. The second value (0.1~K) is somewhat large when compared to previous studies \cite{IC3} based on low-temperature magnetization reversal curves (DM estimated at 0.05~K) in order to show its effect. The last one (0.35~K) is very large for this molecule to exaggerate the effect on the resonance fields. The simulations show a small effect but opposite to the measured behavior [see Fig.~\ref{fig2}(c)] in that the resonance fields are increasing with temperature instead of decreasing. The effect is orders of magnitude smaller if one replaces $D_t$ with $D_z$ (not shown). Therefore, we conclude that $D_t$ has to be sufficiently small (up to $\sim$10 mK) or even nonexistent such that its effect on resonance fields is negligible (since it is opposing the observed behavior). A similar conclusion is drawn for the $D_z$ component or for a situation when the exchange couplings between the triangle's corners are not identical \cite{Chaboussant_epl_2004}. Both of these cases would generate satellite transitions, and we do not observe such signals within the experiment resolution (a few milliteslas). This behavior and the discussed effect of an anisotropic coupling $J_z\neq J_t$ persist for any angle $\phi$. Consequently, we consider in the following no DM coupling and isotropic $J$, in which case the seven resonance fields are all equal and varying the angle $\phi$ has no effect on the resonance fields. Thus, we can further simplify the Hamiltonian by taking $\phi=0$.
\begin{figure}
\includegraphics[width=3.37 in]{gfactors.eps}
\caption{(Color online) Values of the $g_a$ and $g_c$ $g$ factors obtained from the angular dependences shown in Fig.~\ref{fig1} for seven temperatures. The error bars are smaller than the symbol size. The $g$ factors increase with the temperature by $\sim 0.2\%$. }
\label{fig5}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}\hline
Temperature (K) & $g_a$ & $\sigma_{g_a}(\times 10^4)$ & $g_c$ & $\sigma_{g_c}(\times 10^4)$ \\ \hline
4 & 1.9509 & 1.6 & 1.9792 & 1.6 \\ \hline
10 & 1.9513 & 1.5 & 1.9802 & 1.1 \\ \hline
30 & 1.9531 & 1.2 & 1.9821 & 1 \\ \hline
50 & 1.9544 & 1.2 & 1.9822 & 1.2 \\ \hline
60 & 1.9547 & 1 & 1.9817 & 1.2 \\ \hline
90 & 1.9542 & 1 & 1.9812 & 1 \\ \hline
\end{tabular}
\caption{Values of the $g_a$ and $g_c$ $g$ factors shown in Fig.~\ref{fig5}.}
\label{Table1}
\end{center}
\end{table}
\section{Anisotropy of the $\textrm{g}$-tensor}
Using a value of $J_z=J_t=2.45$~K, universal constant values as recommended by the 2010 CODATA database, and the field calibrated with BDPA as explained above, we performed the fits presented in Fig.~\ref{fig2} with solid lines. To determine the anisotropy of the $g$ tensor at different temperatures, the $\chi^2$ parameter is minimized. The fitted $g$ factor values have a precision of four decimal places.
We note that the $g$ factor increases only slightly, by 0.2$\%$ when the temperature is increased (see Table~\ref{Table1} and Fig.~\ref{fig5}). Thermal variations in the $g$ factor are sometimes quite sharp, indicating magnetic or structural transitions, as is the case for some organic conductors \cite{Tokumoto_PRL2008} or for Cu and Mn spins in NH$_4$Cl (see Refs. \onlinecite{Pilbrow_PhysStatSol_1967,Watanabe_PhysLett_1975,Stibbe_Physica_1978}). The decrease of temperature can alter the local symmetry of $3d$ orbitals in the V ions, thus altering their energy splittings. Using first- and second-order perturbation theory, this is demonstrated \cite{Pryce_ProcRSoc_1950} to induce a variation of the anisotropic $g$ tensor in the case of anisotropic orbitals.
The smooth decrease in the $g$ components below $\sim$30-40~K is suggestive of changes in the molecular structure of the V$_{15}$ anion at a local length scale or of an ensuing phase transition involving a long-range lattice order. No earlier study down to 30 mK has observed any long-range ordering (see, for example, Ref.~\onlinecite{IC3}). The observed $g$ value thus clearly indicates that there is a change in the local molecular geometry which could be due to the onset of short-range correlations in the lattice. Our report on the possibility of such local ordering might be of significance in understanding the spin dynamics below 30 K.
Also, it is important to note that the effect of a small DM coupling can easily be absorbed into a small decrease of the $g$ factor. For instance, even if one considers a very large DM vector of size 0.35~K, the $g$ factors would have to decrease by only $\le 0.4\%$ to accommodate it (not shown). As mentioned above, such large values of the DM vector would not be justified, and any DM coupling would induce effects opposed to the measured temperature dependence of the resonance fields. Consequently, it is more suitable to consider the DM vector as being equal to zero and to use the $g$ factor values given in Table \ref{Table1}.
\section{Conclusions}
In conclusion, we present a determination of the spin Hamiltonian parameters of V$_{15}$, a prototypical molecular magnet which has gained a large amount of interest recently. Angular-variation and temperature-dependence studies have been carried out on a precisely oriented single crystal. These measurements show that the anisotropic component of the DM vector is $<10^{-2}$~K, smaller than earlier estimates. Also observed are significant changes in the $g$ tensor at temperatures below 30 K, possibly indicating local order. This information should be of great interest for understanding the low-temperature spin dynamics and electronic structure and bonding in V$_{15}$.
\acknowledgements
We thank Dr. H. B\"ogge (Bielefeld) for some crystallographic remarks. This work was supported by NSF Grant No. DMR-1206267 and CNRS-PICS CoDyLow. The NHMFL is supported by Cooperative Agreement Grant No. DMR-0654118 and the state of Florida. Partial support was received from JSPS KAKENHI Grant-in-Aid for Scientific Research (C) No. 25400391 from MEXT of Japan, and some numerical calculations were supported by the supercomputer center of ISSP of the University of Tokyo. B.T. acknowledges financial support of the Israel Science Foundation (ISF Grant no. 168/09) and COST Action CM1203 \textquotedblleft Polyoxometalate Chemistry for molecular Nanoscience\textquotedblright (PoCheMon).
|
\section{Introduction} \label{SINTRO}
The \emph{Graph Coloring Problem} (GCP) is a very well-studied $\mathcal{NP}$-Hard problem since it models many applications such as
scheduling, timetabling, electronic bandwidth allocation and sequencing problems.
Given a simple graph $G = (V, E)$, where $V$ is the set of vertices and $E$ is the set of edges, a \emph{$k$-coloring of $G$}
is a partition of $V$ into $k$ sets $V_1, V_2, \ldots, V_k$, called \emph{color classes}, such that the endpoints of any edge
lie in different color classes. The GCP consists of finding the minimum number $k$ such that $G$ admits a $k$-coloring,
called the \emph{chromatic number} of $G$ and denoted by $\chi(G)$.
Some applications impose additional restrictions. For instance, in scheduling problems, it may be required to ensure the
uniformity of the distribution of workload employees. Suppose that a set of tasks must be assigned to a set of workers so that
pairs of tasks may conflict each other, meaning that they should not be assigned to the same worker. The problem is modeled by
building a graph containing a vertex for every task and an edge for every conflicting pair of tasks. Workers are represented by colors.
Then, in order for a coloring of this graph to represent a valid assignment of tasks to workers, the same number of tasks must be assigned to each worker.
Since this is impossible when the number of tasks is not divisible by the number of workers, one can ask for the number of tasks
assigned to two arbitrary workers can not differ by more than one.
It is called \emph{equity constraint} and the resulting problem is called \emph{Equitable Coloring Problem} (ECP).
ECP was introduced in \cite{MEYER}, motivated by an application concerning \emph{garbage collection} \cite{EXAMPLE2}.
Other applications of the ECP concern \emph{load balancing problems} in multiprocessor machines \cite{EXAMPLE3}
and results in \emph{probability theory} \cite{EXAMPLE1}. An introduction to ECP and some basic results are provided in
\cite{KUBALE}.
Formally, an \emph{equitable $k$-coloring} (or just $k$-eqcol) of a graph $G$ is a $k$-coloring
satisfying the \emph{equity constraint}, i.e.~ the size of two color classes can not differ by more than one unit.
The \emph{equitable chromatic number} of $G$, $\chi_{eq}(G)$, is the
minimum $k$ for which $G$ admits a $k$-eqcol. The ECP consists of finding $\chi_{eq}(G)$ which is an $\mathcal{NP}$-Hard problem \cite{KUBALE}.
There exist some differences between GCP and ECP that make the latter harder to solve.
It is known that the chromatic number of a graph is greater than or equal to the chromatic number of any of its induced subgraphs.
Unfortunately, in the case of ECP, this property does not hold. For instance, if $G$ is the graph shown in Figure \ref{fig:grasu},
by deleting $v_5$ from $G$, $\chi_{eq}(G)$ increases from 2 to 3.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.2]{grafo.eps}
\end{center}
\vspace{-20pt}
\caption{}
\label{fig:grasu}
\end{figure}
As far as we know, there are few approximate and exact algorithms available in the literature related to ECP.
It was proved that, for any graph $G$, $\Delta(G) + 1$ is an upper bound of $\chi_{eq}(G)$ \cite{ERDOS}, where $\Delta(G)$ is the maximum degree of vertices
in $G$. Based on this fact, a polynomial time algorithm for obtaining a $k$-eqcol of a graph $G$ with $k \geq \Delta(G)+1$ is described in \cite{MARCELO}.
Two constructive heuristics called \textsc{Naive} and \textsc{SubGraph} are given in \cite{KUBALE} to generate greedily an equitable coloring of a graph. There also exist heuristic algorithms for constructing colorings that are ``nearly'' equitable
\cite{BRELAZEQUIT,OTRO}, making emphasis on achieving a small difference between the sizes of the biggest
class and the smallest one, although the equity constraint still might be violated.
The authors of \cite{BYCBRA} propose a tabu search heuristic to initialize an exact algorithm that solves ECP
via Integer Linear Programming (ILP) techniques. Other exact algorithms for solving ECP are given in \cite{PAPERDAM} and \cite{EQDSATUR}.
The first one also uses IPL techinques and the second one is based on a DSATUR enumeration scheme.
In this work, we propose a new heuristic based on the dynamic \textsc{TabuCol} version of Galinier and Hao \cite{GALINIER}, one of the best tabu
search algorithms for GCP \cite{SURVEY}. Then, computational experiments are carried out in order to find the best combination of parameters
involved in the dynamic tenure of our heuristic and to show the good performance of it over known benchmark instances.
The paper is organized as follows. In Section \ref{SGCP}, we present \textsc{TabuCol} and the dynamic variant of Galinier and Hao.
In Section \ref{SECP}, we give our variant for ECP which we call \textsc{TabuEqCol}. Finally, in Section \ref{SCOMPU} we report
computational experiences and conclusions.
\section{\textsc{TabuCol} and its variants} \label{SGCP}
\emph{Tabu search} is a metaheuristic method proposed by Glover \cite{GLOVER} that guides a local search algorithm equipped with additional mechanisms that prevent from visiting a solution twice and getting stuck in a local optimum.
Let $S$ be the solution space of the problem and $f : S \rightarrow \mathbb{R}$ be the objective function. The goal is to obtain a
solution $s \in S$ such that $f(s)$ is minimum.
For each solution $s \in S$, consider a \emph{neighborhood} $N(s) \subset S$ with two desirable (but not exclusionary)
properties: 1) two solutions $s$ and $s'$ are neighbors when it is easy (from the computational point of view) to
obtain $s'$ from $s$, and to obtain $f(s')$ from $f(s)$ (for instance, in constant time), and 2) for any $s, s' \in S$,
there exists a path $s = s_1, s_2, \ldots, s_m = s'$ such that $s_{i+1} \in N(s_i)$ for $i = 1,\ldots,m-1$.
In general, neighbor solutions are very similar in some sense, and the difference between them can be seen as \emph{features} that
both solutions do not share. Consider a set of \emph{features} $P$ and a set $R \subset S \times P$ such that
$(s, p) \in R$ if solution $s$ presents a feature $p$.
Starting from an initial solution $s_0 \in S$, tabu search consists of generating a sequence of solutions
$s_1, s_2, \ldots$ such that $s_{i+1} = \textrm{arg min}_{s \in N'(s_i)} f(s)$, where $N'(s_i)$ is a subset of $N(s_i)$
described below. In each iteration of this algorithm, a \emph{movement} from $s_i$ to $s_{i+1}$ is performed and
some feature of $s_i$ is stored in a \emph{tabu list} $L \subset P$. This list indicates whether a movement is allowed or
forbidden: a solution $s$ can be reached in the future only if $s$ does not present any feature from $L$
(this rule avoids from visiting a solution previously visited), except when $s$ is better than the best solution found so
far. This exception is called \emph{aspiration} and the aspiration criterion is usually to check if the objective value
of $s$ is less than the value of currently-known best solution. Now, the set of allowed movements from $s_i$, $N'(s_i)$, is defined as
\[ N'(s) = \{ s' \in N(s) : f(s') < f(s^*) ~\lor~ (s',p) \notin R ~~ \forall~p \in L \}, \]
where $s^*$ is the best solution found so far.
However, after several iterations, old features are no longer needed and it is better to remove them from the tabu list.
This mechanism is usually implemented by assigning a ``time of live'' to each feature of the tabu list.
Consider $live : L \rightarrow \mathbb{Z}$ and let $live(p)$ be the number of remaining iterations that $p$ belongs to $L$.
When a new feature $p$ is inserted into $L$, $live(p)$ is assigned a value referred to as \emph{tabu tenure} $t$.
Then, in each iteration, the value of $live(p)$ is decreased by one unit until it reachs zero and $p$ is removed from $L$.
Above, we sketch a generic tabu search algorithm.
\begin{algorithm}
\DontPrintSemicolon
\KwData{initial solution $s_0$}
\SetKwData{undefined}{undefined}
\KwResult{best solution found $s^*$}
\Begin{
$L \leftarrow \varnothing$ \;
$s, s^* \leftarrow s_0$ \;
\While{stopping criterion is not met}{
\For{$p \in L$}{
$live(p) \leftarrow live(p) - 1$ \;
if $live(p) = 0$ then $L \leftarrow L \backslash \{p\}$ \;
}
$N'(s) \leftarrow \{ s' \in N(s) : f(s') < f(s^*) ~\lor~ (s',p) \notin R ~~ \forall~p \in L \}$ \;
choose a feature $p \in P$ such that $(s,p) \in R$ \;
$L \leftarrow L \cup \{p\}$ \;
$live(p) \leftarrow t$ \;
$s \leftarrow \textrm{arg min}_{s' \in N'(s)} f(s')$ \;
if $f(s) < f(s^*)$ then $s^* \leftarrow s$ \;
}
}
\caption{\textsc{TabuSearch}}
\end{algorithm}
In order to implement a tabu search algorithm, some decisions must be taken: neighborhood of a solution, features of a solution,
stopping criterion, how to choose the feature $p$ to be stored in the tabu list and how to compute the tabu tenure $t$.
In particular, the value of tabu tenure directly impacts \emph{diversification} of the algorithm.
A tabu search with low tenures behaves as a standard local search, where it frequently get trapped in local minima.
On the other hand, a tabu search with high tenures tends to wander across solution space without converging towards the optimal
solution.\\
\textsc{TabuCol}, the first tabu search algorithm designed for solving GCP, was proposed by Hertz and de Werra \cite{TABUCOL}.
For a given graph $G = (V, E)$ and number $k \in \{1, \ldots, n\}$, where $n = |V|$, the goal of this algorithm is to find a
$k$-coloring of $G$.
In order to obtain a coloring that uses as few colors as possible, it is usual to embed \textsc{TabuCol} in a routine that,
once a $k$-coloring is found, the algorithm can be restarted with $k \leftarrow k - 1$ and so on, until some criterion is met.
Details of \textsc{TabuCol} are given below:
\begin{itemize}
\item \emph{Search space and objective function}. A solution $s$ is a partition $(V_1, V_2, \ldots, V_k)$ of the set of vertices.
Let $E(V_i)$ be the set of edges of $G$ with both endpoints in $V_i$. The objective function is defined as
\[ f(s) = \sum_{i=1}^k |E(V_i)|. \]
Clearly, $s$ is a $k$-coloring if and only if $f(s) = 0$.
\item \emph{Stopping criterion}. The algorithm stops when $f(s) = 0$ or when a maximum number of iterations is reached.
Sometimes, a time limit is imposed.
\item \emph{Initial solution}. It is generated randomly. A suitable procedure given in \cite{REACTTABU} is the following.
Start with empty sets $V_1, V_2, \ldots, V_k$ and, at each step, choose a non-considered vertex $v$ randomly and put it into
$V_i$ with the smallest possible $i$ such that $E(V_i)$ is not incremented. If it is not possible, choose a random number
$j \in \{1, \ldots, k\}$ and put $v$ into $V_j$.
\item \emph{Set of features}. It is $P = V \times \{1, \ldots, k\}$. A solution $s$ presents a feature $(v, i)$ if and only if
$v \in V_i$, i.e.~ if $v$ is assigned color $i$.
\item \emph{Neighborhood of a solution}. Let $C(s)$ be the \emph{set of conflicting vertices} of a solution $s$, i.e.~
\[ C(s) = \{ v \in V : \textrm{$v$ is incident in some edge of $E(V_1) \cup E(V_2) \cup \ldots \cup E(V_k)$} \}. \]
From a solution $s = (V_1, V_2, \ldots, V_k)$, a neighbor $s' = (V'_1, V'_2, \ldots, V'_k)$ is generated as follows. Choose
a conflicting vertex $v \in C(s)$. Let $i$ be the color of $v$ in $s$. Next, choose a color $j \in \{1, \ldots, k\} \backslash \{i\}$
and set
\[ V'_j = V_j \cup \{ v\},~~ V'_i = V_i \backslash \{v\},~~ V'_l = V_l ~~ \forall~l \in \{1,\ldots,k\} \backslash \{i, j\}. \]
In other words, $s'$ is a copy of $s$ except that $v$ is moved from class color $V_i$ to $V_j$. We denote such operation with
$s' = s(i \xrightarrow{v} j)$. Note that objective value can be computed in linear time from $f(s)$:
\[ f(s') = f(s) + |\{ vw \in E : w \in V_j \}| - |\{ vw \in E : w \in V_i \}|. \]
Note also that searching all the neighbors of $s$ requires exploring $(k-1)|C(s)|$ solutions.
Original \textsc{TabuCol} only explores a random subset of $N(s)$ while newer versions explore $N(s)$ completely.
\item \emph{Selection of feature to add in the tabu list}. Once a movement from $s$ to $s(i \xrightarrow{v} j)$ is performed,
$p=(v,i)$ is stored on tabu list and $live(p)$ is set to a fixed tabu tenure $t = 7$.
\end{itemize}
Later, Galinier and Hao \cite{GALINIER} improved \textsc{TabuCol} by using a dynamic tabu tenure that depends on the quality of
the current solution, encouraging diversification of the search when solution is far from optimal.
They proposed to assign a tenure of $t = \alpha|C(s)| + Random(\beta)$ where $Random(\beta)$ returns an integer randomly chosen from $\{0, \ldots, \beta-1\}$ with uniform distribution.
Based on experimentation, they suggest to use $\alpha = 0.6$ and $\beta = 10$.
Other variants of \textsc{TabuCol} are discussed in \cite{SURVEY,REACTTABU}.
\section{\textsc{TabuEqCol}: A tabu search for ECP} \label{SECP}
In this section, we present a new tabu search algorthm for ECP based on \textsc{TabuCol} with
dynamic tabu tenure, which we call \textsc{TabuEqCol}.
Given a graph $G = (V, E)$ and a number $k \in \{1,\ldots,n\}$, where $n = |V|$, the goal of \textsc{TabuEqcol}
is to find a $k$-eqcol of $G$.
Solution space consists of partitions of $V$ into $k$ sets $V_1, V_2, \ldots, V_k$ such that they satisfy
the equity constraint, i.e.~ for any pair of classes $V_i$ and $V_j$, $\bigl| |V_i|-|V_j| \bigr| \leq 1$.
Objective function $f$ is the same as in \textsc{TabuCol}, so any solution $s$ such
that $f(s) = 0$ is indeed an equitable coloring.
Also, set of features $P$ is the same as in \textsc{TabuCol}.
Stopping criterion depends on the experiment carried out. Usually, a time limit is imposed.\\
Let $s \in S$. Denote $W^+(s) = \{ i : |V_i| = \lfloor n/k \rfloor+1\}$ and
$W^-(s) = \{ i : |V_i| = \lfloor n/k \rfloor\}$, where $V_i$ are the color classes of $s$. Since $s$ satisfies the equity
constraint, we have that $W^+(s)$ and $W^-(s)$ determine a partition of $\{1,\ldots,k\}$ and, in particular, $|W^+(s)| = r$ where
$r = n - k\lfloor n / k \rfloor$. From now on, we just write $W^+$ and $W^-$. These sets will be useful in the development
of the algorithm.
We propose two greedy procedures for generating initial solution $s_0$.\\
\noindent \emph{Procedure 1}. Start with empty sets $V_1, V_2, \ldots, V_k$ and an
integer $\tilde{r} \leftarrow 0$ (this value will have the cardinal of $W^+$). At each step,
define set $I = \{ i : |V_i| \leq M-1 \}$, where $M$ is the maximum allowable size of a class:
\[ M = \begin{cases}
\lfloor n/k \rfloor + 1, & \textrm{if}~~\tilde{r} < r \\
\lfloor n/k \rfloor, & \textrm{if}~~\tilde{r} = r
\end{cases} \]
(once we already have $r$ class of size $\lfloor n/k \rfloor + 1$, the size of the remaining classes must not exceed
$\lfloor n/k \rfloor$). Then, choose a non-considered vertex $v$ randomly and put it into a class $V_i$ such that $i \in I$ is the smallest possible and
$E(V_i)$ is not incremented. If it is not possible, $i$ is chosen ramdonly from $I$.
To keep $\tilde{r}$ up to date, each time a vertex is added to a set $V_i$ such that $|V_i| = \lfloor n/k \rfloor$,
$\tilde{r}$ is incremented by one unit.\\
The previous procedure works fine for generating initial solutions from scratch. However, at this point it is common to know a
$(k+1)$-eqcol (i.e.~ in the cases where we previously ran tabu search with $k+1$ and reached an equitable coloring) and
we can exploit this coloring in order to improve the quality of the initial solution as follows.\\
\noindent \emph{Procedure 2}. Let $\mathfrak{p}:\{1,\ldots,k+1\} \rightarrow \{1,\ldots,k+1\}$ be a bijective function
(i.e.~ a random permutation) and let $V^*_1, V^*_2, \ldots, V^*_k, V^*_{k+1}$ be the color classes of the known $(k+1)$-eqcol.
Set $V_i = V^*_{\mathfrak{p}(i)}$ for all $i \in \{1, \ldots, k\}$, and $\tilde{r} = |W^+|$. Then, run Procedure 1 to assign a
color to the remaining vertices which are those belonging to $V^*_{\mathfrak{p}(k+1)}$.\\
Regarding neighborhood of a solution $s \in S$ notice that, if $n$ does not divide $k$, $W^+ \neq \varnothing$ and it is possible to move a vertex from a class of $W^+$ to $W^-$, keeping equity. That is, for all $v \in \cup_{i \in W^+} V_i$ and all $j \in W^-$,
we have $s(i \xrightarrow{v} j) \in S$.
However, the number of allowed movements is rather limited when $r$ is very low (for instance, $r = 1$) or very high ($r = k-1$),
so we need to add supplementary movements.
Swapping the colors of two vertices simultaneously seems to work fine and as well can be used when $n$ divides $k$.
From a solution $s = (V_1, V_2, \ldots, V_k)$, a neighbor $s' = (V'_1, V'_2, \ldots, V'_k)$ is generated with two schemes:
\begin{itemize}
\item \emph{1-move} (only applicable when $n$ does not divide $k$). Choose a conflicting vertex
$v \in C(s) \cap (\cup_{i \in W^+} V_i)$. Let $i$ be the color of $v$ in $s$. Next, choose a color $j \in W^-$. We have
$s' = s(i \xrightarrow{v} j)$. Searching all the neighbors of $s$ with
this scheme requires exploring $(k-r)|C(s) \cap (\cup_{i \in W^+} V_i)|$ solutions.
\item \emph{2-exchange}. Choose a conflicting vertex $v \in C(s)$. Let $i$ be the color of $v$ in $s$. Next, choose another vertex
$u$ such that either $i < j$ or $u \notin C(s)$, where $j$ is the color of $u$ in $s$ (the condition imposed to $u$ prevents from
evaluating 2-exchange on $u$ and $v$ twice). Then, set
\[ V'_j = (V_j \backslash \{u\}) \cup \{v\},~~ V'_i = (V_i \backslash \{v\}) \cup \{u\},~~ V'_l = V_l ~~ \forall~l \in \{1,\ldots,k\} \backslash \{i, j\}. \]
Note that objective value can be computed in linear time from $f(s)$:
\begin{multline*}
f(s') = f(s) + |\{ uw \in E : w \in V_i\backslash\{v\} \}| - |\{ uw \in E : w \in V_j \}| \\
+ |\{ vw \in E : w \in V_j\backslash\{u\} \}| - |\{ vw \in E : w \in V_i \}|.
\end{multline*}
Searching all the neighbors of $s$ with this scheme requires exploring a quadratic number of solutions.
\end{itemize}
Now, let $s'$ be the next solution in the sucession; $s'$ is obtained by applying either 1-move or 2-exchange to $s$, where vertex
$v \in V_i$ in $s$ and $v \notin V'_i$ in $s'$. In both schemes, $p=(v,i)$ is stored on tabu list and $live(p)$ is set to a dynamic tabu tenure
$t = \alpha|C(s)| + Random(\beta)$ where $\alpha$ and $\beta$ are parameters to be determined empirically. This is one of the purposes of the next section.
\section{Computational experiments and conclusions} \label{SCOMPU}
This section is devoted to perform and analyze computational experiments. They were carried out on an Intel i5 CPU [email protected] with Ubuntu Linux O.S.
and Intel C++ Compiler. We considered graphs from \cite{DIMACS}, which are benchmark instances difficult to color.
First, we test different combinations of values for parameters $\alpha$ and $\beta$ from the dynamic tabu tenure in order to determine the
combination that makes \textsc{TabuEqCol} perform better.
Then, we report the behaviour of \textsc{TabuEqCol} over known instances by using the best combination previously found.
We also compare its performance against tabu search algorithm given in \cite{BYCBRA}.\\
\noindent \emph{Tuning parameters}\\
We run \textsc{TabuEqCol} over 16 instances with a predetermined value of $k$ and an initial solution $s_0$ generated
with Procedure 1. The same initial solution is used in all executions of \textsc{TabuEqCol} for the same instance.
Results are reported in Table \ref{TABLE1}. First column is the name of the graph $G$. Second and third columns
are the number of vertices and edges of $G$. Fourth and fifth columns are known lower and upper bound of $\chi_{eq}(G)$ (obtained by other means).
The remaining columns are the time elapsed in seconds by the execution of \textsc{TabuEqCol} when
a $k$-eqcol is found within the term of 1 hour, for each combination. In the case \textsc{TabuEqCol} is not able to find a $k$-eqcol,
$f(s^*)$ is displayed between braces where $s^*$ is the best solution found.
Three last rows indicate the sum of objective function $f(s^*)$ over non-solved instances, percentage of instances \textsc{TabuEqCol}
solved successfully and the average time elapsed for these instances to be solved.
For the sake of simplicity, we refer to each combination with a capital letter.\\
Note that combination D has the least average time, however it has solved less instances than other combinations
and the sum of objective values is also worse. We discard A, B, C, D, E and H with this criterion. By comparing the three remaining combinations,
we have that G is faster than the other two. Even if we restrict the comparison to those 11 instances the 3 combinations solve simultaneously,
we have 807 seconds for F, 562 seconds for G and 730 seconds for I, so G is still better.
We consider combination G ($\alpha = 0.9$ and $\beta = 5$) for \textsc{TabuEqCol}.\\
\noindent \emph{Testing tabu search heuristic}\\
For each instance, the following process is performed. First, execute \textsc{Naive} algorithm (described in \cite{KUBALE}) in order
to find an initial equitable coloring $c$ of the current instance. Suppose that $k+1$ is the number of colors of $c$. Then, obtain an
initial solution $s_0$ of $k$ color classes generated from $c$ with Procedure 2, and run \textsc{TabuEqcol} with parameters $\alpha=0.9$ and $\beta=5$.
If a $k$-eqcol is found, start over the process with $k-1$ color classes by running Procedure 2 and \textsc{TabuEqcol} again.
This process is repeated until 1 hour is elapsed or a $\underline{\chi_{eq}}$-eqcol is reached, and the best coloring found so far
is returned.
In Table \ref{TABLE2} we report results over 76 benchmark instances with at least 50
vertices (75 from \cite{DIMACS} and one Kneser graph used in \cite{BYCBRA}). First 5 columns have the name of the graph $G$, number of vertices
and edges, and best known lower and upper bound of $\chi_{eq}(G)$. Sixth column displays the number
of colors of the initial equitable coloring $c$. Seventh and eighth columns display the value $k$ of the best $k$-eqcol found after 30 seconds
of execution of our algorithm and the time elapsed in seconds until such $k$-eqcol is reached. If the coloring is optimal, $k$ is displayed in boldface. Next two columns show the same information
after 1 hour of execution, but if the best coloring is found within the first 30 seconds, these columns are left empty.
Time spent by \textsc{Naive} is not considered in the computation. However, \textsc{Naive} rarely spent more than 1 sec. (and never more than 4 sec.).
Last two columns show the same information for the tabu search described in \cite{BYCBRA}. If such information is not available, these columns
are left empty. We recall that the values provided in \cite{BYCBRA} were computed on a different platform (1.8 Ghz AMD-Athlon with Linux and GNU C++ compiler).\\
Note that our approach reachs optimality in 29 instances and a gap of one unit between $\underline{\chi_{eq}}$
and the best solution in 7 instances. In other words, it reachs a gap of at most one unit in roughly a half of the evaluated instances. Note also that
\textsc{TabuEqcol} improves the initial solution given by \textsc{Naive} in most cases (precisely, 63 instances).
On those instances the value of the best solution given by tabu search of \cite{BYCBRA} is known, our algorithm gives the same value or a
better one. Despite the difference between platforms, it seems that our approach also runs faster.
An interesting fact is that each execution of \textsc{TabuEqCol} needs no more than 500000 iterations to reach the
best value since the largest number of iterations performed was 493204 and took place when \textsc{TabuEqcol} found a 18-eqcol
of \texttt{DSJC125.5}.
In the same sense, \textsc{TabuEqCol} needs no more than 30000 iterations in each execution and the overall process needs
no more than 30 seconds to reach the best value on 56 instances; justly those ones such that columns 9 and 10 are empty.
On these instances, the largest number of
iterations performed was 28791 and took place when \textsc{TabuEqcol} found a 10-eqcol of \texttt{queen9\_9}.\\
\noindent \emph{Conclusion}\\
The Equitable Coloring Problem is a variation of the Graph Coloring Problem that naturally arises from several applications where the cardinalities of color
classes must be balanced. Just like Graph Coloring, the need to solve applications associated to this new NP-Hard problem justifies the development of exact
and approximate algorithms.
On large instances, known exact algorithms are unable to address them and heuristics such as \textsc{Naive} delivers poor solutions.
Our tabu search heuristic based on \textsc{TabuCol} has shown to improve these solutions and
presented a fairly good performance, even if a limit of 30 seconds is imposed. In addition, an iteration limit of 30000
(for a time limit of 30 seconds) and 500000 (for a time limit of 1 hour) can be imposed in order to save time.\\
\noindent \textbf{Acknowledgements}. This work is partially supported by grants
UBACYT 20020100100666, PICT 2010-304, PICT 2011-817, PID-UNR ING416 and PIP-CONICET 241.
|
\section{Introduction}
\label{sec:intro} Topological modal logic was initiated by the works of A. Tarski and J.C.C. McKinsey in the 1940s.
They were first to consider both topological interpretations of the diamond modality: one as closure, and another as derivative.
Their studies of closure modal
logics were rather detailed and profound.
In particular, in the fundamental paper \cite{MT} they have shown that the logic of any metric separable dense-in-itself space is ${\bf S4}$.
This remarkable result also demonstrates a relative weakness of the closure operator to distinguish between interesting topological properties.
The derivational interpretation gives more expressive power. For example, the real line can be distinguished from the real plane (the observation made by K. Kuratowski as early as in 1920s, cf. \cite{K22}); the real line can be distinguished from the rational line \cite{Sh90}; $T_0$ and $T_D$ separation axioms become expressible \cite{BEG}, \cite{E1}. However, in \cite{MT} McKinsey and Tarski only gave basic definitions for derivational modal logics and put several problems that were solved much later.
The derivational semantics also has its limitations (for example, it is still impossible to distinguish ${\bf R}^2$ from ${\bf R}^3$). Further increase of expressive power can be provided by the well-known methods of adding universal or difference modalities \cite{GorPas}, \cite{GarGor}. In the context of topological semantics this approach also has proved fruitful --- for example, connectedness is expressible in modal logic with the closure and the universal modality \cite{Sh99}, and the $T_1$ separation axiom in modal logic with the closure and the difference modality \cite{kudinov_2005}.
Until the early 1990s, when the connections between topological modal logic and Computer Science were established, the interest in that subject was moderate. Leo Esakia was one of the enthusiasts of modal logical approach to topology, and he was probably the
first to appreciate the role of the derivational modality, in particular, in modal logics of provability \cite{E}. Another strong motivation for further studies of derivational modal logics (`d-logics') were the axiomatization problems left open in \cite{MT}.\footnote{The early works of the second author in this field were greatly influenced by Leo Esakia.} In recent years d-logics have been studied rather intensively, a brief summary of results can be found in section 3 below.
In this chapter the first thorough investigation is provided for logics in the most expressive language in this context\footnote{Some other kinds of topomodal logics arise when we deal with topological spaces with additional structures, e.g. spaces with two topologies, spaces with a homeomorphism etc. (cf. \cite{H}).},
namely the derivational modal logics with the difference modality (`dd-logics'). It unifies earlier studies by the first author in closure modal logics with the difference modality (`cd-logics') and by the second author in d-logics.
The diagram in section 12 compares the expressive power of different kinds of topomodal logics. Our conjecture
is that dd-logics are strictly more expressive than the others, but it is still an open question if the dd-language is stronger than the cd-language. Speaking informally, it is more convenient --- for example, the Kuratowski's axiom for ${\bf R}^2$ (Definition 9.1) is expressible in cd-logic as well, but in a more complicated form \cite{kudinov_2006}.
We show that still in many cases
properties of dd-logics are similar to those of d-logics: finite axiomatizability, decidability and the
finite model property (fmp).
Besides specific results characterizing logics of some particular spaces, our goal was to propose some general methods. In fact, nowadays in topomodal logic there are many technical proofs, but few general methods. In this chapter we propose only two simplifying novelties --- dd-morphisms (section 6) and the Glueing lemma \ref{Glue}, but we hope that much more can be done in this direction, cf. the recent paper \cite{Hodkinson}.
In more detail, the plan of the chapter is as follows. Preliminary sections 2--4 include standard
definitions and basic facts about modal logics and their semantics. Some general completeness results
for dd-logics can be found in sections 5, 7. In section 5 we show that every extension of the minimal
logic $\lgc{K4^\circ D^+}$ by variable-free axioms is topologically complete. In section 8 we prove
the same for extensions of $\lgc{DT_1}$ (the logic of dense-in-themselves $T_1$-spaces); the proof is
based on a construction of d-morphisms from the recent paper \cite{BLB}.
In section 6 we consider validity-preserving maps from topological to Kripke frames (d-morphisms and
dd-morphisms) and prove a modified version of McKinsey--Tarski's lemma on dissectable spaces. In
section 7 we prove that $\lgc{DT_1}$ is complete w.r.t. an arbitrary zero-dimensional dense-in-itself
separable metric space by the method from \cite{Sh90}, \cite{Sh00}.
Sections 8--10 study the axiom of connectedness $AC$ and Kuratowski's axiom $Ku$ related to local
1-componency. In particular we prove that the logic $\lgc{DT_1CK}$ with both these axioms has the
fmp. This is a refinement of an earlier result \cite{Sh90}, \cite{Sh00} on the fmp of the d-logic
$\lgc{D4}+Ku$ (the new proof uses a simpler construction).
Section 11 contains our central result: $\lgc{DT_1CK}$ is the dd-logic of ${\bf R}^n$ for $n>1$. The
proof uses an inductive construction of dd-morphisms onto finite frames of the corresponding logic,
and it combines methods from \cite{Sh90}, \cite{Sh00}, \cite{kudinov_2006}, with an essential
improvement motivated by \cite{LB2} and based on the Glueing lemma.
The final section discusses some further directions and open questions. The Appendix contains technical details of some proofs.
\section{Basic notions}
\label{sec:bas}
\renewcommand{\labelenumi}{(\arabic{enumi})}
The material of this section is quite standard, and most of it can be found in \cite{CZ97}.
We
consider {\em $n$-modal (propositional) formulas} constructed from a countable set of propositional
variables $PV$ and the connectives $\bot$, $\rightarrow$, $\square_1,\ldots,\square_n$. The derived connectives are
$\wedge,\;\vee,\;\neg,\;\top,\;{\leftrightarrow},\;\Diamond_1,\ldots,\Diamond_n$. A formula without occurrences of
propositional variables is called {\em closed}.
A {\em (normal) $n$-modal logic } is a set of modal formulas containing the classical tautologies,
the axioms $\square_i(p\rightarrow q)\rightarrow (\square_i p \rightarrow\square_i q)$ and closed under the standard inference rules:
Modus Ponens ($A,~A\rightarrow B/B$), Necessitation ($A/\square_i A$), and Substitution $(A(p_j)/A(B))$.
To be more specific, we use the terms `($\square_1,\ldots,\square_n$)-modal formula' and
`($\square_1,\ldots,\square_n$)-modal logic'.
${\bf K}_n$ denotes the minimal $n$-modal logic (and ${\bf K}={\bf K}_1$). An $n$-modal logic containing a
certain $n$-modal logic $\mathbf{\Lambda}$ is called an \emph{extension} of $\mathbf{\Lambda}$, or a \emph{$\mathbf{\Lambda}$-logic}. The
minimal $\mathbf{\Lambda}$-logic containing a set of $n$-modal formulas $\Gamma$ is denoted by $\mathbf{\Lambda}+\Gamma$. In
particular,
$${\bf K4}:={\bf K}+\square p\rightarrow\square\sqr p,~
{\bf S4}:={\bf K4}+\square p\rightarrow p, ~{\bf D4}:={\bf K4}+\Diamond\top,
$$
$$
{\bf K4}^\circ := {\bf wK4}:= {\bf K} + p\wedge \square p\rightarrow \square \square p.
$$
The {\em fusion} $L_1*L_2$ of modal logics $L_1,~L_2$ with distinct modalities is the smallest modal
logic in the joined language containing $L_1\cup L_2$.
A {\em (normal) $n$-modal algebra} is a Boolean algebra with extra $n$ unary operations preserving
${\bf 1}$ (the unit) and distributing over $\cap$; they are often denoted by $\square_1,\ldots,\square_n$,
in the same way as the modal connectives. A {\em valuation} in a modal algebra $\mathfrak{A}$ is a
set-theoretic map $\theta: PV \longrightarrow \mathfrak{A}$. It extends to all $n$-modal formulas by induction:
$$\theta(\bot)={\varnothing},~\theta(A\rightarrow B)=-\theta(A)\cup\theta(B),~
\theta(\square_i A)=\square_i\theta(A).$$ A formula $A$ is {\em true in} $\mathfrak{A}$ (in symbols:
$\mathfrak{A}\vDash A$) if $\theta(A)=\textbf{1}$ for any valuation $\theta$. The set ${{\bf L}}(\mathfrak{A})$ of
all $n$-modal formulas true in an $n$-modal algebra $\mathfrak{A}$ is an $n$-modal logic called the
{\em logic of $\mathfrak{A}$}.
An $n$-modal {\em Kripke frame } is a tuple $ F=(W,R_1,\ldots,R_n)$, where $W$ is a nonempty set
(of worlds), $R_i$ are binary relations on $W$. We often write $x\in F$ instead of $x\in W$. In this
chapter (except for Section 2) all 1-modal frames are assumed to be transitive. The associated $n$-modal algebra $MA(F)$ is $2^W$ (the Boolean
algebra of all subsets of $W$) with the operations $\square_1,\ldots,\square_n$ such that $\square_i V= \{x\mid R_i(x)\subseteq
V\}$ for any $V\subseteq W$.
A {\em valuation} in $F$ is the same as in $MA(F)$, i.e., this is a map from $PV$ to ${\cal P}(W)$ (the
power set of $W$). A {\em (Kripke) model} over $F$ is a pair $M=(F,\theta)$, where $\theta$ is a valuation
in $F$. The notation $M,x\vDash A$ means $x\in\theta(A)$, which is also read as `$A$ is \emph{true in $M$
at} $x$'. A (modal) formula $A$ is {\em true in} $M$ (in symbols: $M\vDash A$) if $A$ is true in $M$ at
all worlds. A formula $A$ is called {\em valid in} a Kripke frame $F$ (in symbols: $F\vDash A$) if $A$
is true in all Kripke models over $F$; this is obviously equivalent to $MA(F)\vDash A$.
The {\em modal logic} ${{\bf L}}(F)$ of a Kripke frame $F$ is the set of all modal formulas valid in $F$,
i.e., ${{\bf L}}(MA(F))$. For a class of $n$-modal frames ${\cal C}$, the {\em modal logic} of ${\cal C}$ (or \emph{the modal logic determined} by ${\cal C}$) is ${{\bf L}}({\cal C}):=\bigcap\{{{\bf L}}(F)\mid F\in{\cal C}\}$. Logics determined by classes of
Kripke frames are called {\em Kripke complete}. An $n$-modal frame validating an $n$-modal logic
$\mathbf{\Lambda}$ is called a {\em $\mathbf{\Lambda}$-frame}. A modal logic has the {\em finite model property (fmp)} if it
is determined by some class of finite frames.
It is well known that $(W,R)\vDash{\bf K4}$ iff $R$ is transitive;
$(W,R)\vDash{\bf S4}$ iff $R$ is reflexive transitive (a {\em quasi-order}).
A {\em cluster} in a transitive frame $(W,R)$ is an equivalence class under the relation $\sim_R:=(R
\cap R^{-1}) \cup I_W$, where $I_W$ is the equality relation on $W$. A {\em degenerate cluster} is an
irreflexive singleton. A cluster that is a reflexive singleton, is called {\em trivial}, or {\em simple}. A {\em chain} is a frame $(W,R)$ with $R$
transitive, antisymmetric and linear, i.e., it satisfies $\forall x\forall y~
(xRy\vee yRx\vee x=y)$. A point $x\in W$ is {\em strictly (R-)minimal} if $R^{-1}(x)={\varnothing}$.
A {\em subframe} of a frame $F=(W, R_1,\ldots,R_n)$ obtained by restriction to $V\subseteq W$, is
$F|V:=(V,R_1|V,\ldots,R_n|V)$. Then for any Kripke model $M=(F,\theta)$ we have a {\em submodel}
$M|V:=(F|V,\theta|V)$, where $(\theta|V)(q):=\theta(q)\cap V$ for each $q\in PV$. If $R_i(V)\subseteq V$ for any
$i$, the subframe $F|V$ and the submodel $M|V$ are called {\em generated}.
The {\em union} of subframes $F_j=F|W_j,~j\in J$ is the subframe
$\bigcup\limits_{j\in J}F_j:=F|\bigcup\limits_{j\in J}W_j$.
A {\em generated subframe (cone) with the root $x$} is
$F^x:=F|R^*(x)$, where $R^*$ is the reflexive transitive closure
of $R_1\cup\ldots\cup R_n$; so for a transitive frame $(W,R)$,~$R^*=R\cup I_W$ is the reflexive
closure of $R$ (which is also denoted by $\overline{R}$). A frame $F$ is called {\em rooted} with the root
$u$ if $F=F^u$. Similarly we define a cone $M^x$ of a Kripke model $M$.
Every finite rooted transitive frame $F=(W,R)$ can be presented as the union $(F|C)\cup
F^{x_1}\cup\ldots\cup F^{x_m}$ ($m\geq 0$), where $C$ is the root cluster, $x_i$ are its successors
(i.e., $x_i\not\in C,~\overline{R}^{-1}(x_i)=\sim_R(x_i)\cup C$). If $C$ is non-degenerate, the frame
$F|C$ is $(C,C^2)$, which we usually denote just by $C$. If $C=\{a\}$ is degenerate, $ F|C$ is
$(\{a\},{\varnothing})$, which we denote by $\breve{a}$.
Let us fix the propositional language (and the number $n$) until the end of this section.
\begin{lem}\label{gl} {\em (Generation Lemma)}
\begin{enumerate}
\item
${{\bf L}}(F)=\bigcap\{{{\bf L}}(F^x) \mid x\in F\}$.
\item
If $F$ is a generated subframe of $G$, then ${{\bf L}}(G)\subseteq {{\bf L}}(F)$.
\item
If $M$ is a generated submodel of $N$, then for any formula $A$ for any $x$ in $M$
\[
N,x\vDash A\;\mbox{iff}\; M,x\vDash A.
\]
\end{enumerate}
\end{lem}
\begin {lem}\label{root}
For any Kripke complete modal logic $\mathbf{\Lambda}$,
\[
\mathbf{\Lambda}={{\bf L}}(\mbox{all }\mathbf{\Lambda}\mbox{-frames})={{\bf L}}(\mbox{all rooted }\mathbf{\Lambda}\mbox{-frames}).
\]
\end{lem}
A {\em p-morphism } from a frame $(W,R_1,\ldots,R_n)$ onto a frame $
(W^\prime,R_1^\prime,\ldots,R_n^\prime)$ is a surjective map $f:W\longrightarrow W^\prime $ satisfying the
following conditions (for any $i$):
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item
$\forall x\forall y~ (xR_i y\Rightarrow f(x) R_i'f(y))$ (monotonicity);
\item
$\forall x\forall z~ (f(x)R_i' z\Rightarrow\exists y
(f(y) = z ~\&~ xR_iy))$ (the lift property).
\end{enumerate}
If $xR_iy$ and $f(x)R_i'f(y)$, we say that $xR_iy$ {\em lifts} $f(x)R_i'f(y)$.
Note that (1)$~\&~$(2) is equivalent to
$$\forall x~f(R_i(x))=R_i'(f(x)).$$
$f:\; F\twoheadrightarrow F'$ denotes that $f$ is
a p-morphism from $F$ onto
$F^\prime $.
Every set-theoretic map $f:W\longrightarrow W'$ gives rise to the dual morphism of Boolean algebras
$2^f:2^{W'}\longrightarrow 2^W$ sending every subset $V\subseteq W'$ to its inverse image $f^{-1}(V)\subseteq W$.
\begin {lem}\label{p1} {\em (P-morphism Lemma)}~
\renewcommand{\labelenumi}{(\arabic{enumi})}
\begin{enumerate}
\item
$f:\; F\twoheadrightarrow F'$ iff $2^f$ is an embedding of $MA(F')$ in $MA(F)$.
\item
$f:\; F\twoheadrightarrow F'$ implies ${\text{\bf {\em L}}}(F)\subseteq{\text{\bf {\em L}}}(F').$
\item
If $f:\; F\twoheadrightarrow F'$, then $F\vDash A\Leftrightarrow F'\vDash A$ for any closed formula $A$.
\end{enumerate}
\end{lem}
In proofs of the fmp in this chapter we will use the well-known filtration method \cite{CZ97}. Let us recall the construction we need.
Let $\Psi$ be a set of modal formulas closed under subformulas.
For a
Kripke model $M =
(F,\varphi )$ over a frame $F = (W,R_1,\ldots,R_n)$, there is the equivalence relation on $W$
$$ x
\equiv_\Psi y \Longleftrightarrow \forall A\in\Psi (M,x \vDash A \Leftrightarrow M,y \vDash A).$$ Put
$W^{\prime}:=W/\equiv_\Psi;~~x^\sim\,:=\; \equiv_\Psi(x)$ (the equivalence class of $x$),\\
$\varphi^\prime (q)
:=\{x^\sim\mid x\in \varphi (q)\}$ for $q \in PV\cap\Psi$ (and let $\varphi'(q)$ be arbitrary for $q
\in PV-\Psi$).
\begin{lem}\label{L26}{\em(Filtration Lemma)}
Under the above assumptions,
consider the relations ${\underline{R}}_i, R'_i $ on $W^\prime$ such that
$$a{\underline{R}}_ib \hbox{ iff } \exists x\in a~ \exists y\in b~ xR_iy ,$$
\[
R'_i=
\begin{cases}
\mbox{the transitive closure of }{\underline{R}}_i & \mbox{ if }R_i\mbox{ is transitive,}\nonumber\\
{\underline{R}}_i & \mbox{otherwise.}\nonumber
\end{cases}
\]
Put $M^\prime := (W^\prime, R'_1,\ldots,R'_n, \varphi^\prime)$.
Then
for any $x\in W, ~A\in\Psi $ :
$$M,x \vDash A\mbox{ iff }M',x^\sim \vDash A.$$
\end{lem}
\begin{defi}
An {\em $m$-formula} is a modal formula in propositional variables $\{ p_1,\dots,p_m\}$. For a
modal logic $\mathbf{\Lambda}$ we define the {\em $m$-weak} (or {\em $m$-restricted) canonical frame
$F_{\mathbf{\Lambda}\lceil m} := (W, R_1,\ldots, R_n)$ and canonical model} $M_{\mathbf{\Lambda}\lceil m}:= (F_{\mathbf{\Lambda}\lceil
m}, \varphi)$, where $W$ is the set of all maximal $\mathbf{\Lambda}$-consistent sets of $m$-formulas,
$
x R_i y \mbox{ iff for any }m\mbox{-formula } A$\\
$(\square_i A\in x \Rightarrow A\in y), $
\[
\varphi(p_i): =
\begin{cases}
\{ x\mid p_i\in x\}& \mbox{ if }i\leq m,\nonumber\\
{\varnothing}& \mbox{ if }i> m.
\nonumber
\end{cases}
\]
$\mathbf{\Lambda}$ is called {\em weakly canonical} if $ F_{\mathbf{\Lambda}\lceil m}\vDash\mathbf{\Lambda}$ for any finite $m$.
\end{defi}
\begin{propo}\label{cm}
For any $m$-formula $A$ and a modal logic $\mathbf{\Lambda}$
\begin{enumerate}
\item
$M_{\mathbf{\Lambda}\lceil m},x \vDash A \mbox{ iff }A\in x$;
\item
$M_{\mathbf{\Lambda}\lceil m}\vDash A \mbox{ iff }A\in\mathbf{\Lambda}$;
\item
if $\mathbf{\Lambda}$ is weakly canonical, then it is Kripke complete.
\end{enumerate}
\end{propo}
\begin{coro}\label{cm1}
If for any $m$-formula $A$, $M_{\mathbf{\Lambda}\lceil m},x\vDash A{\Leftrightarrow} M_{\mathbf{\Lambda}\lceil m},y\vDash A$, then $x=y$.
\end{coro}
\begin{defi}
A cluster $C$ in a transitive frame $(W,R)$ is called
\emph{maximal} if \mbox{$\overline{R}(C)
= C$.}
\end{defi}
\begin{lem}\label{L82}\label{L84}
Let $ F_{\mathbf{\Lambda}\lceil m} =(W,R_1,\ldots,R_n)$ and suppose $\mathbf{\Lambda}\vdash\square_1p\rightarrow\square_1\square_1p $ (i.e.,
$R_1$ is transitive). Then every generated subframe of $(W,R_1)$ contains a maximal cluster.
\end{lem}
The proof is based on the fact that the general Kripke frame corresponding to a canonical model is
descriptive; cf. \cite{CZ97}, \cite{F85} for further details\footnote{For the 1-modal case this lemma
has been known as folklore since the 1970s; the second author learned it from Leo Esakia in 1975.}.
\section{Derivational modal logics}
\renewcommand{\labelenumi}{(\arabic{enumi})}
We denote topological spaces by $\mathfrak{X}, \mathfrak{Y},\ldots$ and the corresponding sets by
$X,Y,\dots$.\footnote{Sometimes we neglect this difference.} The interior operation in a space $\mathfrak X$
is denoted by ${\bf I}_X$ and the closure operation by ${\bf C}_X$, but we often omit the subscript $X$. A
set $S$ is a {\em neighbourhood} of a point $x$ if $x\in {\bf I} S$; then $S-\{x\}$ is called a {\em
punctured neighbourhood} of $x$.
\begin{defi}
Let $\mathfrak X$ be a topological space, $V\subseteq X$. A point $x\in X$ is said to be {\em limit} for $V$ if
$x\in{\bf C}(V-\{x\})$; a non-limit point of $V$ is called {\em isolated}.
The {\em derived set of $V$} (denoted by ${\bf d} V$,
or by ${\bf d}_X V$) is the set of all limit points of $V$. The unary operation $V \mapsto {\bf d} V$ on
${\cal P}(X)$ is called {\em the derivation} (in $\mathfrak X$).
A set without isolated points is called
{\em dense-in-itself}.
\end{defi}
\begin{lem}\label{dY}\cite{K66}
For a subspace $\mathcal Y\subseteq \mathfrak X$ and $V\subseteq X$
${\bf d}_Y(V\cap Y)={\bf d}_X(V\cap Y)\cap Y$; if $Y$ is open, then
${\bf d}_Y(V\cap Y)={\bf d}_XV\cap Y$.
\end{lem}
\begin{defi}
The \emph{derivational algebra of a topological space $\mathfrak X$} is $DA(X):=(2^X, {\bf \tilde{d}} )$, where $2^X$ is the Boolean algebra of all
subsets of $X$, ${\bf \tilde{d}} V := -{\bf d} (-V)$\footnote{There is no common notation for this operation; some authors use
$\mathbf{\tau}$.}.
The \emph{closure algebra of a space $\mathfrak X$} is $CA(\mathfrak X):=(2^X,{\bf I} )$.
\end{defi}
\begin{rem}\rm
In \cite{MT} the derivational algebra of $\mathfrak X$ is defined as $(2^X,{\bf d} )$, and the closure algebra as
$(2^X,{\bf C} )$, but here we adopt equivalent dual definitions.
\end{rem}
It is well known that $CA(\mathfrak X), ~DA(\mathfrak X)$ are modal algebras, $CA(\mathfrak X)\vDash {\bf S4}$ and
$DA(\mathfrak X) \vDash {\bf K4}^\circ$ (the latter is due to Esakia).
Every Kripke ${\bf S4}$-frame $F=(W,R)$ is associated with a topological space $N(F)$ on $W$, with the {\em Alexandrov} (or {\em right}) {\em topology} $\{V\subseteq W\mid R(V)\subseteq V\}$. In $N(F)$
we have ${\bf C} V=R^{-1}(V), ~{\bf I} V=\{x\mid R(x)\subseteq V\}$; thus $MA(F)=CA(N(F))$.
\begin{defi}
A modal formula $A$ is called {\em d-valid in a topological space} $\mathfrak X$ (in symbols, $\mathfrak X\vDash^d A$) if it is true in the algebra
$DA(\mathfrak X)$. The logic ${{\bf L}}(DA(\mathfrak X))$ is called the {\em derivational modal logic} (or the {\em d-logic}) of $\mathfrak X$ and denoted by ${{\bf L}}{\bf d} (\mathfrak X)$.
A formula $A$ is called {\em c-valid} in $\mathfrak X$ (in symbols, $\mathfrak X\vDash^c A$) if it is true in $CA(\mathfrak X)$.
${{\bf L}}{\bf c} (\mathfrak X):={{\bf L}}(CA(\mathfrak X))$ is called the {\em closure modal logic,} or the
\emph{c-logic} of $\mathfrak X$.
\end{defi}
\begin{defi}
For a class of topological spaces ${\cal C}$ we also define the
{\em d-logic}
${{\bf L}}{\bf d}({\cal C}):=\bigcap\{{{\bf L}}{\bf d}(\mathfrak X)\mid \mathfrak X\in{\cal C}\}$ and the
{\em c-logic}
${{\bf L}}{\bf c}({\cal C}):=\bigcap\{{{\bf L}}{\bf c}(\mathfrak X)\mid \mathfrak X\in{\cal C}\}$. Logics of this form
are called {\em d-complete} (respectively, {\em c-complete} ).
\end{defi}
\begin{defi}
A {\em valuation} in a topological space $\mathfrak X$ is a map
$\varphi : PV \longrightarrow {\cal P}(\mathfrak X)$. Then $(\mathfrak X,\varphi)$ is called a {\em topological model} over $\mathfrak X$.
\end{defi}
So valuations in $\mathfrak X$, $CA(\mathfrak X)$, and $DA(\mathfrak X)$ are the same. Every valuation $\varphi$ can be prolonged to
all formulas in two ways, according either to $CA(\mathfrak X)$ or $DA(\mathfrak X)$. The corresponding maps are
denoted respectively by $\varphi_c$ or $\varphi_d$. Thus
\begin{align*}
\varphi_d (\square A) &= {\bf \tilde{d}}\varphi_d (A),~&
\varphi_d (\Diamond A) &= {\bf d}\varphi_d (A),\\
\varphi_c (\square A) &= {\bf I}\varphi_c (A),~&
\varphi_c (\Diamond A) &= {\bf C}\varphi_c (A).
\end{align*}
A formula $A$ is called {\em d-true} (respectively, {\em c-true}) in $(\mathfrak X,\varphi)$ if $\varphi_d ( A) =X$ (respectively, $\varphi_c ( A) =X$ ).
So $A$ is d-valid in $\mathfrak X$ iff $A$ is d-true in every topological model over $\mathfrak X$, similarly for c-validity.
\begin{defi}
A modal formula $A$ is called \emph{d-true at a point} $x$ in a topological model $(\mathfrak X,\varphi)$ if $x\in\varphi_d (A)$.
\end{defi}
Instead of $x\in\varphi_d(A)$, we write $x \vDash^d A$ if the model is clear from the context.
Similarly we define the c-truth at a point and use the corresponding notation.
From the definitions we obtain
\begin{lem}
For a topological model over a space $\mathfrak X$
\begin{itemize}
\item
$x \vDash^d \square A$ iff $\exists U\ni x~ (U$ is open in $\mathfrak X~ \&~ \forall
y\in U-\{ x\}~ y \vDash^d A)$;
\item
$x \vDash^d \Diamond A$ iff $\forall U\ni x~ (U$ is open in $\mathfrak X~ \Rightarrow~ \exists
y\in U-\{ x\}~ y \vDash^d A)$.
\end{itemize}
\end{lem}
\begin{defi}\label{lt1}
A \emph{local $T_1$-space} (or a {\em $T_D$-space} \cite{AT62}) is a topological space,
in which every point is {\em locally closed}, i.e, closed in some
neighbourhood.
\end{defi}
Note that a point $x$ in an Alexandrov space $N(W,R)$ is closed iff it is minimal (i.e., $R^{-1}(x)=\{x\}$); $x$ is
locally closed iff $R(x)\cap R^{-1}(x)=\set{x}$. Thus $N(F)$ is local
$T_1$ iff $F$ is a poset.
\begin{lem}\label{L310}\cite{E1}
For a topological space $\mathfrak X$
\begin{enumerate}
\item
$\mathfrak X \vDash^d {\bf K4}\mbox{ iff } \mathfrak X\mbox{ is local }T_1;$
\item
$\mathfrak X \vDash^d \Diamond\top\mbox{ iff } \mathfrak X\mbox{ is dense-in-itself}.$
\end{enumerate}
\end{lem}
\begin{defi}
A Kripke frame (W,R) is called {\em weakly transitive} if
$R\circ R\subseteq \overline{R}$.
\end{defi}
It is obvious that the weak transitivity of $R$ is equivalent to the transitivity of $\overline{R}$.
\begin{propo}\label{P32}\cite{E1}~
(1) $(W,R) \vDash {\bf K4}^\circ\mbox{ iff } (W,R)\mbox{ is weakly transitive};$\\
(2) ${\bf K4}^\circ$ is Kripke-complete.
\end{propo}
\begin{lem}\label{L33}\cite{E1}~
(1) Let $F = (W,R)$ be a Kripke ${\bf S4}$-frame, and let $R^\circ := R-I_W$, $F^\circ := (W,R^\circ)$.
Then ${\bf Ld}(N(F)) = {{\bf L}}(F^\circ)$.\\
(2)
Let $F=(W,R)$ be a weakly transitive irreflexive Kripke frame, and let $\overline{F}=: (W, \overline{R})$ be its
reflexive closure. Then $ {{\bf L}}{\bf d} (N(\overline{F}))={{\bf L}}(F) $.\\
(3)
If $\mathbf{\Lambda}={{\bf L}}({\cal C})$, for some class ${\cal C}$ of weakly transitive irreflexive Kripke frames, then $\mathbf{\Lambda}$
is d-complete.
\end{lem}
\begin{proof}
(1) Note that $R^\circ(x)$ is the smallest punctured neighbourhood of $x$ in the space $N(F)$. So the inductive d-truth definition in a topological model $(N(F),\varphi)$ coincides with the inductive truthdefinition in the Kripke model $(W,R^\circ,\varphi)$.
(2) Readily follows from (1), since $\overline{R}$ is transitive and $(\overline{R})^\circ=R$.
(3) Follows from (2). {\hspace*{\fill} }
\end{proof}
\begin{defi}
For a 1-modal formula $A$ we define
$A^\sharp$ as the formula obtained by replacing every occurrence of every subformula $\square B$ with $\overline{\square}B:=\square B\wedge B$.
For a 1-modal logic $\mathbf{\Lambda}$ its {\em reflexive fragment} is $^\sharp\mathbf{\Lambda}:=\{A\mid\mathbf{\Lambda}\vdash A^\sharp\}$.
\end{defi}
\begin{propo}\label{cldl}\cite{BEG}~
(1) If $\mathbf{\Lambda}$ is a ${\bf K4}^\circ$-logic, then $^\sharp\mathbf{\Lambda}$ is an ${\bf S4}$-logic.
\\
(2) For any topological space $X$, ${\bf Lc}(\mathfrak X)= \,^\sharp{\bf Ld}(\mathfrak X)$,
\\
(3) For any weakly transitive Kripke frame $F$, ${\bf L}(\overline{F})=\, ^\sharp{\bf L}(F)$.
\end{propo}
\begin{proof}
(1) It is clear that for a weakly transitive $\mathbf{\Lambda}$, $\overline{\square}$ satisfies the axioms of ${\bf S4}$, so $^\sharp\mathbf{\Lambda}$ contains these axioms. Since $^\sharp$ distributes over implication, it follows that $^\sharp\mathbf{\Lambda}$ is closed under Modus Ponens. For the substitution closedness, note that for any variable $p$ and formulas $A,B$ $([B/p]A)^\sharp=[B^\sharp/p] A^\sharp$; thus $A\in\, ^\sharp\mathbf{\Lambda}$ implies $[ B/p]A\in\, ^\sharp\mathbf{\Lambda}$.Finally, since $(\square A)^\sharp= \overline{\square}\,A^\sharp$, it is clear that $A\in\, ^\sharp\mathbf{\Lambda}$ only if $\square A\in\,^\sharp\mathbf{\Lambda}$.
(2) By definitions,
$${\bf Lc}(X)\vdash A \;\mbox{iff}\; CA(X)\vDash A,$$
$$^\sharp{\bf Ld}(X)\vdash A \;\mbox{iff}\; {\bf Ld}(X)\vdash A^\sharp\;\mbox{iff}\; DA(X)\vDash A^\sharp.$$
Let us show that that $ CA(X){\,\not\mo\,} A$ iff $ DA(X){\,\not\mo\,} A^\sharp$. In fact, consider a topological model $(X,\varphi)$. We claim that
\[
\varphi_c(B)=\varphi_d(B^\sharp)\eqno(*)
\]
for any formula $B$. This is easily checked by induction, the crucial case is when $B=\square B_1$; then by definitions and the induction hypothesis we have:
\[
\varphi_c(B)={\bf I}\varphi_c(B_1)={\bf I}\varphi_d(B_1^\sharp)= \boxdot\varphi_d(B_1^\sharp)\cap \varphi_d(B_1^\sharp)=\varphi_d(\overline{\square}\,B_1^\sharp)= \varphi_d(B^\sharp).
\]
The claim (*) implies that $\varphi_c(A)\neq X$ iff $\varphi_d(A^\sharp)\neq X$ as required.
(3) On the one hand,
$${\bf L}(\overline{F})={{\bf L}}(MA(\overline{F}))= {\bf L} (CA(N(\overline{F}))= {\bf Lc}(N(\overline{F})).$$
On the other hand, by Lemma \ref{L33}(2),
\[
{\bf L}(F)= {\bf Ld}(N(\overline{F})),
\]
and we can apply (2) to $N(F)$.{\hspace*{\fill} }
\end{proof}
Let us give some examples of d-complete logics.
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item ${\bf Ld}(\mbox{all ~topological ~spaces})=
{\bf K4}^\circ$. This was proved by L. Esakia in the 1970s and published in \cite{E1}.
\item
${\bf Ld}(\mbox{all ~local~}T_1\mbox{-spaces})= {\bf K4}$. This is also a result from \cite{E1}.
\item
${\bf Ld}(\mbox{all }T_0\mbox{-spaces})= {\bf K4}^\circ + p \land \Diamond (q \land \Diamond p) \to \Diamond p \lor \Diamond(q \land \Diamond q)$. This result is from \cite{BEG2011}.
\item
L. Esakia \cite{E} also proved that G\"odel - L\"ob logic
${\bf GL} := {\bf K} + \square (\square p\supset p)\supset \square p$
is the derivational logic of the class of all topological scattered spaces (a space is {\em
scattered} if each its nonempty subset has an isolated point).
\item
The papers \cite{A87}, \cite{A88}, \cite{Bl} give a complete description of
d-logics of
ordinals with the interval topology: ${\bf Ld}(\alpha)$ is
either ${\bf
GL}$ (if $\alpha\geq\omega^\omega$), or
${\bf GL}+ \square^n \bot$ (if $\omega^{n-1}\leq\alpha<\omega^n$). In particular,
${\bf Ver} := {\bf K} + \square \bot$ is the d-logic of any finite ordinal (and of any discrete space).
\item
The well-known ``difference logic"
\cite{S80}, \cite{DR93}
${\bf DL} := {\bf K4}^\circ + \Diamond \square p \supset p,$
is determined by Kripke frames with the difference relation:
${\bf DL} = {{\bf L}}(\{(W,\neq_W)\mid W\neq{\varnothing}\}),$
where $\neq_W:=W^2-I_W$; hence by \ref{cldl}, ${\bf DL}$ is the d-logic of the class of all
trivial topological spaces. However, for any particular trivial space $\mathfrak X$, ${\bf Ld}(\mathfrak X)\neq {\bf
DL}$. Moreover, ${\bf Ld}(\mathfrak X)$ is not finitely axiomatizable for any infinite trivial $\mathfrak X$\cite{KudShap}; this surprising result is easily proved by a standard technique using Jankov formulas (cf. \cite{Kudinov08}).
\item
In \cite{Sh00} it was proved that
${\bf Ld}\mbox{(all 0-dimensional separable metric spaces)}={\bf K4}.$
All these spaces are embeddable in ${\bf R}$ \cite{K66}.
\item
In \cite{Sh00} it was also proved that for any dense-in itself separable metric space $\mathfrak X$, ${\bf
Ld}(\mathfrak X)={\bf D4}$; this was a generalization of an earlier proof \cite{Sh90} for $\mathfrak X={\bf Q}$. A more elegant proof for ${\bf Q}$ is in \cite{LB1}.
\item
Every extension of ${\bf K4}$ by a set of closed axioms is a d-logic of some subspace of ${\bf Q}$ \cite{BLB}. This gives us a
continuum of d-logics of countable metric spaces.
\item
In \cite{Sh90} ${\bf Ld}({\bf R}^2)$ was axiomatized and it was also proved that the d-logics of ${\bf R}^n$ for $n\geq 2$ coincide.
We will simplify and extend that proof in the present chapter.
\item
$\mathbf{Ld}({\bf R})$ was described in \cite{Sh00}; for a simpler completeness proof cf. \cite{LB2}.
\item $\mathbf{Ld} (\mbox{all Stone spaces})= {\bf K4}$ and $\mathbf{Ld} (\mbox{all weakly scattered Stone spaces})= {\bf K4} + \Diamond \top \to \Diamond \Box \perp$,
cf. \cite{BEG2010}.
\item
d-logics of special types of spaces were studied in \cite{BEG}, \cite{LB1}. They include submaximal, perfectly disconnected, maximal, weakly scattered and some others.
\end{enumerate}
However, not all extensions of ${\bf K4}^\circ$ are d-complete. In fact, the formula $p\supset \Diamond p$ never
can be d-valid, because ${\bf d} Y = {\varnothing}$ for a singleton $Y$. So every extension of ${\bf S4}$ is
d-incomplete, and thus Kripke completeness does not imply d-completeness.
\begin{propo}\label{P42}
Let $F=(\omega^*,\prec)$ be the ``standard irreflexive transitive tree", where $\omega^*$ is the set of all finite sequences in
$\omega$; $\alpha\prec\beta$ iff $\alpha$ is a proper initial segment of $\beta$. Then
$${\bf D4} = {{\bf L}}(F)={{\bf L}}{\bf d} (N(\overline{F})) = {{\bf L}}{\bf d} ({\cal D}),$$
where ${\cal D}$ denotes the class of all dense-in-themselves local $T_1$-spaces.
\end{propo}
\begin{proof}
The first equality is well known \cite{VB83}; the second one holds by \ref{L33}. By \ref{L310},
${\bf D4}$ is d-valid exactly in spaces from ${\cal D}$. So $N(\overline{F})\in{\cal D}$, ${\bf D4}\subseteq{{\bf L}}{\bf d} ({\cal D})
$, and the third equality follows.{\hspace*{\fill} }
\end{proof}
\section{Adding the universal modality and the difference modality}
\label{sec:basics}
\renewcommand{\labelenumi}{(\arabic{enumi})}
Recall that the {\em universal modality} $[\forall]$ and the {\em difference modality} $[\neq]$ correspond to Kripke frames with the universal and the difference relation.
So (under a valuation in a set $W$)
these modalities are interpreted in the standard way:
\begin{align*}
x\vDash[\forall]A &\;\mbox{iff}\;~ \forall y\in W~y\vDash A;~ &x\vDash[\neq]A &\;\mbox{iff}\; \forall y\in W~(y\neq x \Rightarrow y\vDash A).
\end{align*}
The corresponding dual modalities are denoted by $\langle\exists\rangle$ and $\DE$.
\begin{defi}
For a $[\forall]$-modal formula $A$
we define the $[\neq]$-modal formula $A^u$
by induction:
\begin{equation*}
A^u:=A\mbox{ for }A\mbox{ atomic},\ \
(A\supset B)^u:=A^u\supset B^u,\ \
([\forall]B)^u :=[\neq]B^u\wedge B^u.
\end{equation*}
\end{defi}
We can consider 2-modal topological logics obtained from ${\bf Lc}(\mathfrak X)$ or ${\bf Ld}(\mathfrak X)$ by adding
the universal or the difference modality\footnote{So we extend the definitions of the d-truth or the
c-truth by adding the item for $[\forall]$ or $[\neq]$.}. Thus for a topological space $\mathfrak X$ we obtain
four 2-modal logics : ${\bf Lc}_\forall(\mathfrak X)$ (the {\em closure universal (cu-) logic}), ${\bf
Ld}_\forall(\mathfrak X)$ (the {\em derivational universal (du-) logic}), ${\bf Lc}_{\neq}(\mathfrak X)$ (the {\em closure
difference (cd-) logic}), ${\bf Ld}_{\neq}(\mathfrak X)$ (the {\em derivational difference (dd-) logic}).
Similar notations (${\bf Lc}_\forall({\cal C})$ etc.) are used for logics of a class of spaces
${\cal C}$, and respectively we can define four kinds of topological completeness (cu-, du-, cd-, dd-) for
2-modal logics.
cd-logics were first studied in \cite{Gab01},
cu-logics in \cite{Sh99}, du-logics in
\cite{LB2}, but dd-logics have never been addressed so far.
For a $\square$-modal logic ${{\bf L}}$ we define the 2-modal logics
\begin{align*}
{{\bf L}}{\bf D} &:={{\bf L}}*{\bf D}{{\bf L}}+[\neq]p\wedge p\rightarrow\square p,
~
&{{\bf L}}{\bf D}^+ &:={{\bf L}}*{\bf D}{{\bf L}}+[\neq]p\rightarrow\square p,
\\
{{\bf L}}{\bf U}&:={{\bf L}}*{\bf S5}+[\forall]p\rightarrow\square p.
\end{align*}
Here we suppose
that ${\bf S5}$ is formulated in the language with $[\forall]$ and ${\bf D}{{\bf L}}$ in the language with $[\neq]$.
The following is checked easily:
\begin{lem}\label{L41}
For any topological space $\mathfrak X$,
\[
{\bf Lc}_\forall(\mathfrak X)\supseteq {\bf S4}{\bf U},\quad {\bf Ld}_\forall(\mathfrak X)\supseteq{\bf K4}^\circ{\bf U},\quad {\bf Lc}_{\neq}(\mathfrak X)\supseteq
{\bf S4}{\bf D},\quad {\bf Ld}_{\neq}(\mathfrak X) \supseteq{\bf K4}^\circ{\bf D}^+.
\]
\end{lem}
\begin{defi}
For a 1-modal Kripke frame $F=(W,R)$ we define 2-modal frames
$F_\forall:=(F,W^2),~ F_{\neq}:=(F,\neq_W)$
and modal logics ${{\bf L}}_\forall(F):={{\bf L}}(F_\forall),~ {{\bf L}}_{\neq}(F):={{\bf L}}(F_{\neq})$.
\end{defi}
Sahlqvist theorem \cite{CZ97} implies
\begin{propo}~
The logics ${\bf S4}{\bf U},~ {\bf K4}^\circ{\bf U}, ~ {\bf S4}{\bf D},~{\bf K4}^\circ{\bf D}^+ $ are Kripke complete.
\end{propo}
Using the first-order equivalents of the modal axioms for these logics (in particular, Proposition
\ref{P32}) we obtain
\begin{lem}\label{L46}~ For a rooted Kripke frame $G=(W,R,S)$
\begin{enumerate}
\item
$
G\vDash{\bf S4}{\bf U}\;\mbox{iff}\; R\mbox{ is a quasi-order }\&~S=W^2,
$
\item
$
G\vDash{\bf K4}^\circ{\bf U} \;\mbox{iff}\; R\mbox{ is weakly transitive }\&~S=W^2,
$
\item
$ G\vDash{\bf S4}{\bf D} \;\mbox{iff}\; R\mbox{ is a quasi-order }\&~\overline{S}=W^2, $
\item
$ G\vDash{\bf K4}^\circ{\bf D}^+\;\mbox{iff}\; R\mbox{ is weakly transitive }\&~\overline{S}=W^2\&~R\subseteq S. $
\end{enumerate}
Also note that $\overline{S}=W^2\;\mbox{iff}\;\neq_W\subseteq S$.
\end{lem}
\begin{defi}\label{F0}
A rooted Kripke $\lgc{K4^\circ D}^+$-frame described by
Lemma \ref{L46} (4) is called {\em basic}. The class of these frames is denoted by $\mathfrak{F}_0$.
\end{defi}
Next, we easily obtain the 2-modal analogue to Lemma \ref{L33}.
\begin{lem}\label{L47}~
(1)
Let $F$ be an ${\bf S4}$-frame. Then
\[
{\bf Ld}_{\neq}(N(F))={{\bf L}}_{\neq}(F^\circ),~ {\bf Ld}_\forall(N(F))={{\bf L}}_\forall(F^\circ).
\]
(2)
Let $F$ be a weakly transitive irreflexive Kripke frame. Then
\[
{\bf Ld}_{\neq}(N(\overline{F}))={{\bf L}}_{\neq}(F),~ {\bf Ld}_\forall(N(\overline{F}))={{\bf L}}_\forall(F).
\]
(3)
Let ${\cal C}$ be a class of weakly transitive irreflexive Kripke 1-frames. Then ${{\bf L}}_{\neq}({\cal C})$
is dd-complete, ${{\bf L}}_{\forall}({\cal C})$
is du-complete.
\end{lem}
Let us extend the translations $(-)^\sharp$, $(-)^u$ to 2-modal formulas.
\begin{defi}
$(-)^u$ translates $(\square, [\forall])$-modal formulas to $(\square, [\neq])$-modal formulas so that $([\forall]B)^u =[\neq]B^u\wedge B^u$
and $(-)^u$ distributes over the other connectives.
Similarly, $(-)^\sharp$ translates $(\square,[\neq])$-modal formulas and
$(\square,[\forall])$-modal formulas to formulas of the same kind, so that
$(\square\,B)^\sharp =\square\,B^\sharp\wedge B^\sharp$
and $(-)^\sharp$ distributes over the other connectives.
\begin{align*}
^u\mathbf{\Lambda}&:=\{A\mid A^u\in\mathbf{\Lambda}\}\mbox{ for a }(\square,[\forall])\mbox{-modal logic } \mathbf{\Lambda}\mbox{ (the {\em
universal fragment}),}\\
^\sharp\mathbf{\Lambda}&:=\{A\mid A^\sharp\in\mathbf{\Lambda}\}\mbox{ for a }(\square,[\neq])\mbox{- or a }
(\square,[\forall])\mbox{-modal }\mathbf{\Lambda}\mbox{ (the {\em reflexive fragment}),}\\
\,^\sharp\phantom{}^u\mathbf{\Lambda}&:= \,^\sharp(^u\mathbf{\Lambda})\mbox{ for a }(\square,[\neq])\mbox{-modal } \mathbf{\Lambda}\mbox{
(the {\em reflexive universal fragment}).}
\end{align*}
\end{defi}
\begin{propo}\label{P49}~
(1)
The map $\mathbf{\Lambda}\mapsto\,^\sharp\mathbf{\Lambda}$ sends $\lgc{K4^\circ D^+}$-logics to $\lgc{S4U}$-logics.\\
(2)
The map $\mathbf{\Lambda}\mapsto\,^u\mathbf{\Lambda}$ sends $\lgc{K4^\circ D^+}$-logics to
$\lgc{ K4^\circ U}$-logics and $\lgc{S4D}$-logics to $\lgc{S4U}$-logics.
\\(3)
The map $\mathbf{\Lambda}\mapsto\,^\sharp\phantom{}^u\mathbf{\Lambda}$ sends $\lgc{K4^\circ D^+}$-logics to
$\lgc{S4U}$-logics.
\\(4)
For a topological space $\mathfrak X$
\[
{{\bf L}}{\bf c}_{\neq}(\mathfrak X)=\,^\sharp{{\bf L}}{\bf d}_{\neq}(\mathfrak X),~ {{\bf L}}{\bf d}_{\forall}(\mathfrak X)=\,^u{{\bf L}}{\bf d}_{\neq}(\mathfrak X),~ {{\bf L}}{\bf c}_{\forall}(\mathfrak X)=\,^u{{\bf L}}{\bf c}_{\neq}(\mathfrak X)={} ^\sharp{{\bf L}}{\bf d}_{\forall}(\mathfrak X).
\]
\\(5)
For a weakly transitive Kripke frame $F$
\[
{{\bf L}}_{\neq}(\overline{F})=\,^\sharp{{\bf L}}_{\neq}(F),~
~{{\bf L}}_{\forall}(F)=\,^u{{\bf L}}_{\neq}(F),
~{{\bf L}}_{\forall}(\overline{F})=\,^\sharp{{\bf L}}_{\forall}(F)=
\,^\sharp\,^u{{\bf L}}_{\neq}(F).
\]
\end{propo}
Proposition \ref{P49}\,(4) implies that dd-logics are the most
expressive of all kinds of the logics we consider.
\begin{coro}
If ${{\bf L}}{\bf d}_{\neq}(\mathfrak X)= {{\bf L}}{\bf d}_{\neq}(\mathfrak Y)$ for spaces $\mathfrak X,\mathfrak Y$, then all the other
logics (du-, cu-, cd-, d-, c-) of these spaces coincide.
\end{coro}
Let
\[
AT_1:= \DA p \rightarrow \DA \square p, \quad
AC:= \left[\forall\right](\square p \lor \square\neg p) \rightarrow \left[\forall\right] p \lor \left[\forall\right] \lnot p.
\]
\begin{propo}\label{P412}
For a topological space $\mathfrak X$
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item $\mathfrak X \models^d \Diamond\top$ iff $\mathfrak X$ is dense-in-itself;
\item $\mathfrak X \models^d AT_1$ iff $\mathfrak X \models^c AT_1$ iff $\mathfrak X$ is a $T_1$-space;
\item $\mathfrak X\vDash^d AC^{\sharp}$ iff $\mathfrak X \models^c AC$ iff $\mathfrak X$ is connected.
\end{enumerate}
\end{propo}
\begin{proof}
(1) and the first equivalence in (2) are trivial. The first equivalence in (3) follows from
\ref{P49}(4). The remaining ones are checked easily, cf. \cite{kudinov_2006},
\cite{Sh99}.{\hspace*{\fill} }
\end{proof}
For a $\square$-modal logic $\lgc{L}$ put
\[
\begin{array}{rcl}
\lgc{LD^+T_1} := \lgc{LD^+} + AT_1, &\qquad&
\lgc{LD^+T_1C} := \lgc{LD^+} + AT_1 + AC^{\sharp u}.
\end{array}
\]
Also put
\[
\lgc{KT_1}:=\lgc{K4D^+T_1},~\lgc{DT_1}:=\lgc{D4D^+T_1},
~\lgc{DT_1C}:=\lgc{D4D^+T_1C}.
\]
\begin{propo}\cite{kudinov_2006}\label{pr:DS_AT1}
If $F = (W, R, R_D)$ is basic, then \mbox{$F \models AT_1$} iff all $R_D$-irreflexive points are strictly $R$-minimal iff $R_D \circ R \subseteq R_D$.
\end{propo}
\begin{rem}\rm
Density-in-itself is expressible in cd-logic and dd-logic by the formula $DS:= \DA p \supset \Diamond p$,
So for any space $\mathfrak X$,
$\mathfrak X\vDash^c DS\;\mbox{iff}\;\mathfrak X\vDash^d DS\;\mbox{iff}\; \mathfrak X\vDash^d\Diamond\top$.
It is known that $DS$ axiomatizes dense-in-themselves spaces in cd-logic \cite{kudinov_2006}.
However, in dd-logic this axiom is insufficient:
${\bf Ld}_{\neq}$(all dense-in-themselves spaces) $= \lgc{D4^\circ D^+}=\lgc{K4^\circ D^+}+\Diamond\top$, and it is
{\em stronger} than $\lgc{K4^\circ D^+}+DS$. (To see the latter, consider a
singleton Kripke frame, which is $R_D$-reflexive, but $R$-irreflexive.) Therefore
$\lgc{K4^\circ D^+}+DS$ is dd-incomplete.
\end{rem}
\begin{rem}\rm
Every $T_1$-space is a local $T_1$-space, so the dd-logic of all $T_1$-spaces contains $\square
p\rightarrow\square\sqr p$. However, $\lgc{K4^{\circ}D^+T_1}\,{\not\vdash}\,\square p\rightarrow\square\sqr p$.
In fact, consider a 2-point frame $F:=(W,\neq_W,W^2)$. It is clear that $F\vDash\lgc{K4^{\circ}D^+}$.
Also $F\vDash AT_1$, by Proposition \ref{pr:DS_AT1}, but $F{\,\not\mo\,}\square p\rightarrow\square\sqr p$, since $\neq_W$ is
not transitive.
It follows that $\lgc{K4^{\circ}D^+T_1}$ is dd-incomplete; $T_1$-spaces are actually axiomatized by ${\bf KT_1}$ (Corollary \ref{C713}).
\end{rem}
\newcommand{\mathbf{Lc_\forall}}{\mathbf{Lc_\forall}}
\newcommand{\mathbf{Lc_{\ne}}}{\mathbf{Lc_{\ne}}}
\newcommand{\mathbf{Ld_\forall}}{\mathbf{Ld_\forall}}
\newcommand{\mathbf{Ld_{\ne}}}{\mathbf{Ld_{\ne}}}
Let us give some examples of du-, cu- and cd-complete logics.
\begin{enumerate}
\item $\mathbf{Lc_\forall}(\hbox{all spaces}) = \lgc{S4U}$.
\item $\mathbf{Lc_\forall}(\hbox{all connected spaces}) = \mathbf{Lc_\forall}(\mathbf R^n) = \lgc{S4U} + AC$ for any $n\ge 1$ \cite{Sh99}\footnote{The paper \cite{Sh99} contains a stronger claim: $\mathbf{Lc_\forall}(\mathfrak X)=\lgc{S4U} + AC$ for any connected dense-in-itself separable metric $\mathfrak X$. However, recently we found a gap
in the proof of Lemma 17 from that paper. Now we state the main result only for the case ${\mathfrak X}=\mathbf R^n$; a proof can be obtained by applying the methods of the present Chapter, but we are planning to publish it separately.}
\item $\mathbf{Ld_\forall}(\hbox{all spaces}) = \lgc{S4D}$ \cite{DR93}.
\item $\mathbf{Lc_{\ne}}(\mathfrak X) = \lgc{S4DT_1+DS}$, where $\mathfrak X$ is a 0-dimensional separable metric space \cite{kudinov_2006}.
\item $\mathbf{Lc_{\ne}}({\bf R}^n)$ for any $n\ge 2$ is finitely
axiomatized in \cite{kudinov_2005}; all these logics coincide.
\item $\mathbf{Ld_\forall}({\bf R})$ is finitely axiomatized in \cite{LB2}.
\end{enumerate}
\section{dd-completeness of $\lgc{K4^\circ D^+}$ and some of its extensions}
This section contains some simple arguments showing that there are many dd-complete bimodal logics.
All formulas and logics in this section are $(\square,\DA)$-modal. An arbitrary Kripke frame for
$(\square,\DA)$-formulas is often denoted by $(W,R,R_D)$.
\begin{lem}\label{L35}~
(1)
Every weakly transitive Kripke 1-frame is a p-morphic image of some irreflexive weakly transitive Kripke 1-frame. \\
(2)
Every rooted $\lgc{K4^\circ D^+}$-frame is a p-morphic image of some $R$- and $R_D$-irreflexive rooted
$\lgc{K4^\circ D^+}$-frame.
\end{lem}
\begin{proof}
(1) Cf. \cite{E1}.
(2) Similar to the proof of (1).
For $F=(W,R,R_D)\in \mathfrak{F}_0$ put
\[
W_r:=\{a\mid aR_Da\},~ W_i=W-W_r, \quad \tilde{W}:= W_i\cup (W_r \times \{ 0,1\}).
\]
Then we define the relation $\tilde{R}$ on $\tilde{W}$ such that
\begin{align*}
(b,j) \tilde{R}a &\mbox{ ~~iff ~~ } bRa,&
a \tilde{R} (b,j) &\mbox{ ~~iff ~~ } aRb,\\
(b,j) \tilde{R} (b',k) &\mbox{ ~~iff ~~ } bRb'\ \&\ b\ne b' \vee b= b'\ \&\ j\ne k,&
a\tilde{R} a' &\mbox{ ~~iff ~~ } aRa'.
\end{align*}
Here $a,a'\in W_i;~ b,b'\in W_r; ~j,k\in\{ 0,1\}$. So we duplicate all $R_D$-reflexive points making them irreflexive (under
both relations).
It follows that $\tilde{F}:=(\tilde{W},\tilde{R},\neq_{\tilde{W}})\in \mathfrak{F}_0$ and $\tilde{R}$ is
irreflexive; the map $f:\tilde{W} \to W$ sending $(b,j)$ to $b$ and $a$ to itself (for $b\in W_r, ~ a\in
W_i$) is a p-morphism $\tilde{F}\twoheadrightarrow F$. {\hspace*{\fill} }
\end{proof}
\begin{propo}\label{P53}
Let $\Gamma$ be a set of closed 2-modal formulas, $\mathbf{\Lambda}:= \lgc{K4^\circ D^+}+\Gamma$. Then
\begin{enumerate}
\item
$\mathbf{\Lambda}$ is Kripke complete.
\item
$\mathbf{\Lambda}$ is dd-complete.
\end{enumerate}
\end{propo}
\begin{proof}
(1) $\lgc{K4^\circ D^+}$ is axiomatized by Sahlqvist formulas. One can easily check that (in the
minimal modal logic) every closed formula is equivalent to a positive formula, so
we can apply Sahlqvist theorem.
(2) Suppose $A\not\in\mathbf{\Lambda}$. By (1) and the Generation lemma there exists a rooted Kripke 2-frame $F$
such that $F\vDash L$ and $F{\,\not\mo\,} A$. Then by Lemma \ref{L35}, for some irreflexive weakly transitive
1-frame $G=(W,R)$ there is a p-morphism $(G,\neq_W)\twoheadrightarrow F$. By the p-morphism lemma
$(G,\neq_W){\,\not\mo\,} A$ and $(G,\neq_W)\vDash\mathbf{\Lambda}$ (since $\Gamma$ consists of closed formulas). Hence by Lemma
\ref{L47}, $\mathbf{\Lambda}\subseteq\lgc{Ld}_{\neq}(N(\overline{G}))$, $A\not\in\lgc{Ld}_{\neq}(N(\overline{G}))$.
{\hspace*{\fill} }
\end{proof}
\begin{rem}\rm
Using Proposition \ref{P53} and the construction from \cite{BLB} one can prove that there is a continuum of dd-complete logics. Such a claim is rather weak, because Proposition \ref{P53} deals only with Alexandrov spaces. In section 7 we will show how to construct many dd-complete logics of metric spaces.
\end{rem}
\section{d-morphisms and dd-morphisms; extended McKinsey - Tarski's Lemma}
In this section we recall the notion of a d-morphism (a
validity-preserving map for d-logics) and introduce dd-morphisms, the analogues of d-morphisms for dd-logics. This is the main technical tool in the present chapter. Two basic lemmas are proved here, an analogue of
McKinsey--Tarski's lemma on dissectability for d-morphisms and the Glueing lemma.
The original McKinsey--Tarski's lemma \cite{MT} states the existence of a c-morphism (cf. Remark \ref{R61} ) from an arbitrary separable dense-in-itself metric space onto a certain quasi-tree of depth 2. The
separability condition is actually redundant \cite[Ch.~3]{RS} (note that the latter proof is quite different from \cite{MT}\footnote{\label{foot:7}Recently P. Kremer \cite{Kremer} has showed that ${\bf S4}$ is {\em strongly complete} w.r.t. any dense-in-itself metric space. His proof uses much of the construction from \cite{RS}.}). But c-morphisms preserve validity only for c-logics, and unfortunately, the constructions by
McKinsey--Tarski and Rasiowa--Sikorski cannot be used for d-morphisms. So we need another construction to prove a stronger form of McKinsey--Tarski's lemma.
\begin{defi}\label{D51}
Let $\mathfrak X$ be a topological space, $F=(W,R)$ a transitive Kripke frame. A map $f: X\longrightarrow W$ is called a {\em
d-morphism} from $\mathfrak X$ to $F$ if $f$ is open and continuous as a map $\mathfrak X\longrightarrow N(\overline{F})$ and also satisfies
\begin{align*}
\mbox{r-density}:&~~\forall w\in W( wRw\Rightarrow f^{-1} (w)\subseteq {\bf d} f^{-1} (w)),\\
\mbox{i-discreteness}:&~~\forall w\in W( \neg wRw\Rightarrow f^{-1} (w)\cap {\bf d} f^{-1} (w)={\varnothing}).
\end{align*}
If $f$ is surjective, we write $f: \mathfrak X\twoheadrightarrow^d F$.
\end{defi}
\begin{propo}\label{P52}\cite{BEG}~
(1)
$f$ is a d-morphism from $\mathfrak X$ to $F$ iff $2^f$ is a homomorphism from $MA(F)$ to $DA(\mathfrak X)$.\\
(2)
If $f: \mathfrak X\twoheadrightarrow^d F$, then ${{\bf L}}{\bf d} (\mathfrak X) \subseteq {{\bf L}}(F)$.
\end{propo}
\begin{coro}\cite{Sh90}\label{L52}
A map $f$ from a topological space $\mathfrak X$ to a finite transitive Kripke frame $F$ is a d-morphism iff
$$\forall w\in W~ {\bf d} f^{-1} (w) = f^{-1}(R^{-1}(w)).$$
\end{coro}
\begin{proof}
$2^f$ preserves Boolean operations. It is a homomorphism of modal algebras iff it preserves diamonds, i.e., iff for any $V\subseteq W$,
$$f^{-1}(R^{-1}(V))={\bf d} f^{-1}(V).$$
Inverse images and ${\bf d}$ distribute over finite unions, so the above equality holds for any (finite) $V$ iff it holds for singletons, i.e.,
\[\hspace{3.8cm}\ f^{-1}(R^{-1}(w))= {\bf d} f^{-1}(w).\qedhere \]
\end{proof}
\begin{rem}\label{R61}\rm
For a space $\mathfrak X$ and a Kripke ${\bf S4}$-frame $F=(W,R)$ one can also define a {\em c-morphism} $\mathfrak X\longrightarrow F$ just as an open and continuous map $f: \mathfrak X\longrightarrow N(F)$. So every d-morphism to an ${\bf S4}$-frame is a c-morphism. It is well known \cite{RS} that $f: X\longrightarrow W$ is a c-morphism iff $2^f$
is a homomorphism $MA(F) \longrightarrow CA(\mathfrak X)$.
Again for a finite $F$ this is equivalent to
$$
\forall w\in W~{\bf C} f^{-1}(w) = f^{-1}(R^{-1}(w)).
$$
\end{rem}
\begin{lem}\label{Restlem}
If $f:\mathfrak X\twoheadrightarrow^d F$ for a finite frame $F$ and $\mathcal Y\subseteq\mathfrak X$ is an open subspace, then $f|Y$ is a d-morphism.
\end{lem}
\begin{proof}
We apply Proposition \ref{P52}. Note that $f|Y$ is the composition $f \cdot j$, where
$j:Y\hookrightarrow X$ is the inclusion map. Then $2^{ f|Y}=2^j\cdot 2^f$. Since $2^f$ is a
homomorphism $MA(F)\longrightarrow DA(\mathfrak X)$, it remains to show that $2^j$ is a homomorphism $DA(\mathfrak X)\longrightarrow
DA(\mathcal Y)$, i.e., it preserves the derivation:
$j^{-1}({\bf d} V)={\bf d}_Yj ^{-1}(V)$, or
${\bf d} V\cap Y={\bf d}_Y(V\cap Y)$,
which follows from \ref{dY}.
{\hspace*{\fill} }
\end{proof}
\begin{defi}\label{D53}
A set $\gamma$ of subsets of a topological space
$\mathfrak X$ is called \emph{dense} at $x\in X$ if every neighbourhood of $x$
contains a member of $\gamma$.
\end{defi}
\begin{propo}\label{P54}
For $m>0,~ l>0$ let
$\Phi_{ml}$ be a ``quasi-tree" of height 2, with singleton maximal clusters and
An
$m$-element
root cluster (Fig. \ref{fig:Phiml}). For $l=0, ~m>0, ~ \Phi_{ml}$ denotes an $m$-element cluster.
Let $\mathfrak X$ be a dense-in-itself separable metric space, $B\subset X$ a closed nowhere dense set.
Then there exists a d-morphism $g: \mathfrak X\twoheadrightarrow^d \Phi_{ml}$ with the following properties:
\begin{enumerate}
\item
$B \subseteq g^{-1}(b_1)$;
\item
every $g^{-1}(a_i)$ (for $i\leq l$ ) is a union of a set $\alpha_i$ of disjoint
open balls, which is
dense at any point of $g^{-1}(\{ b_1,...,b_m\}$).
\end{enumerate}
\end{propo}
\begin{proof}
Let $X_1,\dots, X_n,\dots$ be a countable base of $\mathfrak X$ consisting of open balls. We construct sets
$A_{ik}, ~B_{jk}$ for $1\leq i\leq l, ~1\leq j\leq m, ~k\in\omega$, with the following properties:
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item
$A_{ik}$ is the union of a finite set $\alpha_{ik}$ of nonempty open balls whose closures are disjoint;
\item
${\bf C} A_{ik} \cap {\bf C} A_{i'k} = {\varnothing}$ for $i\not= i'$ ;
\item
$\alpha_{ik} \subseteq \alpha_{i,k+1};~A_{ik} \subseteq A_{i,k+1}$;
\item
$B_{jk}$ is finite;
\item
$B_{jk} \subseteq B_{j,k+1};$
\item
$A_{ik} \cap B_{jk} = {\varnothing} ;$
\item
$X_{k+1} \subseteq\bigcup\limits_{i=1}^l A_{ik}\Rightarrow \alpha_{i,k+1} = \alpha_{ik}, ~B_{j,k+1}
=
B_{jk};$
\item
if $X_{k+1}\not\subseteq \bigcup\limits_{i=1}^l A_{ik}$, there are closed
nontrivial balls
$P_1,\dots,P_l$ such that for any $i$, $j$
$$P_i \subseteq X_{k+1}-A_{ik},~
\alpha_{i,k+1} = \alpha_{ik} \cup \{ {\bf I} P_i\},~
(B_{j,k+1} - B_{jk}) \cap X_{k+1}\not= {\varnothing}; $$
\item
$A_{ik} \subseteq X-B$;
\item
$B_{jk} \subseteq X-B$;
\item
$j\not= j'\Rightarrow B_{j'k} \cap B_{jk} = {\varnothing}$ .
\end{enumerate}
We carry out both the construction and the proof by induction on $k$.
Let $k=0$. $(X-B)$ is infinite, since
it is nonempty and open in a
dense-in-itself
$\mathfrak X$.
Take distinct points $v_1,\dots,v_l \not\in B$ and disjoint closed
nontrivial
balls
$ Z_1,\dots,Z_l \subset X-B$ with centres at $v_1,\dots,v_l$ respectively (see Fig.\ref{fig:McT-constraction}).
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{fig-2.pdf}
\caption{Case k = 0}
\label{fig:McT-constraction}
\end{figure}
Put
$$\alpha_{i0}: =\{ {\bf I} Z_i\}; ~A_{i0}: = {\bf I} Z_i;$$
then $Z_i = {\bf C} A_{i0}$. As above, since $(X-B) -\bigcup\limits^l_{i=1}Z_i$ is nonempty and open, it
is infinite. Pick distinct $w_1,\dots,w_m\in X-B$ and
put $B_{j0}: = \{ w_j\}$. Then the required properties hold for
$k=0$.
At the induction step we construct
$A_{i,k+1}, B_{j,k+1}$.
Put $Y_k :=\bigcup\limits ^l_{i=1}A_{ik}$ and consider two cases.\\
(a) $X_{k+1}\subseteq Y_k$. Then
put:
$$\alpha_{i,k+1}: = \alpha_{ik};~ A_{i,k+1}: = A_{ik}; ~B_{j,k+1}: = B_{jk}.$$
(b) $X_{k+1}\not\subseteq Y_k$. Then
$X_{k+1} \not\subseteq {\bf C} Y_k$.
In fact, $X_{k+1} \subseteq {\bf C} Y_k$ implies $X_{k+1} \subseteq {\bf I} {\bf C} Y_k=Y_k$, since $X_{k+1}$ is open and by (1) and (2).
So we put
$$W_0 := X_{k+1} - {\bf C} Y_k -\bigcup^m_{j=1}B_{jk},~
W := W_0 - B.$$
Since $(X_{k+1} - {\bf C} Y_k )$ is nonempty and open and every $B_{jk}$ is finite by
(4), $W_0$ is also open and nonempty (by the density of $\mathfrak X$). By the assumption of
\ref{P54}, $B$ is closed, and thus $W$ is open.
$W$ is also nonempty. In fact, otherwise $W_0 \subseteq B$, and then $W_0 \subseteq {\bf I} B={\varnothing}$ (since
$B$ is nowhere dense by the assumption of \ref{P54}).
Now we argue similarly to the case $k=0$.
Take disjoint closed nontrivial balls $P_1,\dots,P_l\subset W$. Then $W - \bigcup\limits^l_{i=1}P_i$
is infinite, so we pick distinct $b_{1,k+1},\dots,
b_{m,k+1}$ in this set and put
$$B_{j,k+1}: = B_{jk} \cup \{ b_{j,k+1}\},~
\alpha_{i,k+1}: = \alpha_{ik} \cup \{ {\bf I} P_i\},~ A_{i,k+1}: = A_{ik} \cup {\bf I} P_i.$$
In the case (a) all the required properties hold for $(k+1)$ by the construction.
In the case (b) we have to check only (1), (2), (6), (8)--(11).
(8) holds, since by construction we have
\begin{align*}
P_i\subset W\subset& X_{k+1}-{\bf C} Y_k\subset X_{k+1}-A_{ik};\\
b_{j,k+1}\in W \subseteq& X_{k+1},~ b_{j,k+1}\in (B_{j,k+1}-B_{jk}).
\end{align*}
(1): From IH it is clear that $\alpha_{i,k+1}$ is a finite set of open balls and
their closures are disjoint; note
that $P_i \cap {\bf C} A_{ik} = {\varnothing}$, since
$P_i \subseteq W \subseteq -{\bf C} A_{ik}.$
(2): We have
\begin{align*}
&{\bf C} A_{i,k+1} \cap {\bf C} A_{i^\prime,k+1} = ({\bf C} A_{ik} \cup P_i) \cap ({\bf C}
A_{i^\prime k}\cup P_{i^\prime}) =\\
=&({\bf C} A_{ik} \cap {\bf C} A_{i^\prime k}) \cup ({\bf C} A_{ik} \cap P_{i'}) \cup ({\bf C}
A_{i^\prime k} \cap P_i) \cup (P_i \cap P_{i'})
={\bf C} A_{ik} \cap {\bf C} A _{i^\prime k} = {\varnothing}
\end{align*}
by IH and by the construction; note that
$P_i, P_i' \subseteq W \subseteq -{\bf C} Y_k$.
(6): We have
$$
A_{i,k+1} \cap B_{j,k+1} = (A_{ik} \cap B_{jk} ) \cup ({\bf I} P_i \cap
\set{b_{j,k+1}}) \cup ( A_{ik} \cap \set{ b_{j,k+1}}) \cup ({\bf I} P_i \cap B_{jk}) = {\varnothing}
$$
by IH and since $b_{j,k+1}\not\in P_i,~ b_{j,k+1}\in W \subseteq X-Y_k$, $P_i \subset W \subseteq
X-B_{jk}$ .
(9): We have
$A_{i,k+1} = A_{ik} \cup {\bf I} P_i \subseteq -B,$
since $A_{ik} \subseteq -B$ by IH, and $P_i \subset W \subseteq -B$ by the construction.
Likewise, (10) follows from $B_{jk} \subseteq -B$ and $b_{j,k+1}\in W \subseteq -B$.
To check (11), assume $j \not= j'$. We have
$B_{j',k+1} \cap B_{j,k+1} = B_{j'k} \cap B_{jk}$,
since
$b_{j^\prime,k+1} \not= b_{j,k+1} , ~ b_{j,k+1}\in W \subseteq -B_{j^\prime k}$
and
$b_{j^\prime,k+1}\in W\subseteq -B_{jk}$. Then apply IH.
Therefore the required sets $A_{ik}, B_{jk}$ are constructed. Now put
$$\alpha_i: =\bigcup_k \alpha_{ik},~
A_i:= \bigcup \alpha_i=\bigcup_k A_{ik}, ~
B_j: =\bigcup_k B_{jk},$$
$$ B^\prime _1: = X - (\bigcup_i A_i \cup \bigcup_j B_j),$$
and define a map
$g: X\longrightarrow \Phi_{ml}$ as follows:
\[
g(x):=\left\{ \begin{array}{ll}
a_i & \mbox{if } x\in A_i, \\
b_j & \mbox{if } x\in B_j,~j\neq 1,\\
b_1 & \mbox{otherwise (i.e., for } x\in B^\prime _1). \\
\end{array} \right.
\]
By (2), (3), (5), (6), (11), $g$ is well defined; by (9), (10) $B \subseteq g^{-1}(b_1)$.
To prove that $g$ is a d-morphism, we check some other properties.
$$\leqno(12)\quad X - \bigcup\limits^l_{i=1}A_i \subseteq {\bf d} B_j.$$
In fact, take an arbitrary $x\not\in \bigcup\limits^l_{i=1}A_i $ and show that $x\in {\bf d} B_j$, i.\/e.,
$$\leqno(13)\quad(U-\{ x\}) \cap B_j \not= {\varnothing} . $$
for any neighbourhood $U$ of $x$. First assume that $x\not\in B_j$. Take a basic open $X_{k+1}$ such that $x\in X_{k+1} \subseteq U$. Then
\mbox{$X_{k+1}\not\subseteq \bigcup\limits^l_{i=1}A_i$,}
and (8) implies
$B_{j,k+1} \cap X_{k+1} \not= {\varnothing} .$
Thus
$B_j \cap U \not= {\varnothing}$.
So we obtain (13).
Suppose $x\in B_j$; then $x\in B_{jk}$ for some $k$. Since $\mathfrak X$ is dense-in-itself and
$\set{X_1,\,X_2,\dots}$ is its open base, $\setdef[X_{s+1}]{s\geq k}$ is also an open base
(note that every ball in $\mathfrak X$ contains a smaller ball).
So $x\in X_{s+1}\subseteq U$ for some $s\ge k$. Since $x\not\in \bigcup\limits^l_{i=1}A_i $, we have
$X_{s+1}\not\subseteq\bigcup\limits^l_{i=1}A_i$,
and so $(B_{j,s+1}-B_{js}) \cap X_{s+1} \not= {\varnothing}$ by (8); thus $(B_j-B_{js}) \cap U \not= {\varnothing}$.
Now $x\in B_{jk} \subseteq B_{js}$ implies (13).
$$\leqno(14)\qquad {\bf d} B_j \subseteq X -\bigcup\limits^l_{i=1}A_i.$$
In fact,
$B_j\subseteq -A_i$, by (3), (5), (6). So ${\bf d} B_j \subseteq {\bf d} (-A_i) \subseteq -A_i$, since $A_i $ is open.
Similarly we obtain
$$\leqno(15)\qquad {\bf d} B^\prime _1\subseteq X -\bigcup\limits^l_{i=1}A_i, \qquad {\bf d} A_i \subseteq X- \bigcup\limits_{r\not=i}A_r.$$
Also note that
$$\leqno(16)\qquad A_i \subseteq {\bf d} A_i,$$
since $A_i$ is open, $\mathfrak X$ is dense-in-itself. Similary to (12) we have
$$\leqno(17)\qquad \alpha_i \mbox{ is dense at every point of }B_j, B'_1\ \mbox{ (and thus $B_j,\ B'_1 \subseteq {\bf d} A_i$}).$$
To conclude that $g$ is a d-morphism, note that
$$g^{-1}(a_i) = A_i, ~g^{-1}(b_j) = B_j~ (\mbox{for }j\not= 1),~ g^{-1}(b_1) =
B^\prime _1,$$
and so by (15), (16), (17)
\begin{align*}
{\bf d} g^{-1}(a_i) &= {\bf d} A_i = X - \bigcup\limits_{r\not=i}A_r = g^{-1}(R^{-
1}(a_i)),
\end{align*}
and by (12), (14), (15)
\begin{align*}
{\bf d} g^{-1}(b_j) &= {\bf d} B_j = X - \bigcup\limits^l_{i=1}A_i = g^{-1}(R^{-
1}(b_j))\ \ \mbox{(for $j\ne 1$)},\\ \qquad\qquad\qquad
{\bf d} g^{-1}(b_1) &= {\bf d} B^\prime _1= X - \bigcup\limits^l_{i=1}A_i = g^{-1}(R^{-
1}(b_1)). \qedhere
\end{align*}
\end{proof}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{fig-1.pdf}
\caption{Frame $\Phi_{ml}$.}
\label{fig:Phiml}
\end{figure}
For the proof see Appendix.
\begin{lem}\label{refroot}
Assume that
\begin{enumerate}
\item
$\mathfrak X$ is a dense-in-itself separable metric space,
\item
$B\subset X$ is closed nowhere dense,
\item
$F = C\cup F_1\cup \dots\cup F_l$ is a $\lgc{D4}$-frame, where $C = \{ b_1,\dots,b_m\} $ is a non-degenerate
root cluster,
$F_1,\dots,F_l$ are the subframes generated by the successors of $C$,
\item
for any nonempty open ball $U$ in $\mathfrak X$, for any $i\in\{1,\ldots,l\}$ there exists a d-morphism $f_i^U:U\twoheadrightarrow^d F_i$.
\end{enumerate}
Then there exists $f:\mathfrak X\twoheadrightarrow^d F$ such that $f(B)=\{b_1\}$.
\end{lem}
\begin{proof}
First, we construct
$g: \mathfrak X\twoheadrightarrow^d \Phi_{ml}$ according to Proposition \ref{P54}. Then $B\subseteq g^{-1}(b_1)$ and
$A_i = g^{-1}(a_i)$ is the union of a set $\alpha_i$ of disjoint open balls.
Then put
\begin{equation}\label{Eq:fPhiml}
f(x):=\left\{ \begin{array}{ll}
g(x) & \mbox{if } g(x)\in C, \\
f_i^U(x)& \mbox{if } x\in U,~U\in\alpha_i.\\
\end{array} \right.
\end{equation}
Since $g$ and $f_i^U$ are surjective, the same holds for $f$. So let us show
$${\bf d} f^{-1}(a) = f^{-1}(R^{-1}(a))$$
($R$ is the accessibility relation on $F$).
First suppose $a\in C$. Then (since $g$ is a d-morphism)
$${\bf d} f^{-1}(a) = {\bf d} g^{-1}(a) = g^{-1}(C) = f^{-1}(C) = f^{-1}(R^{-1}(a)). $$
Now suppose $a \notin C$, $I = \setdef[i]{a \in F_i}$, and let $R_i$ be the accessibility relation on $F_i$. We have:
\begin{align*}
f^{-1}(a) =&\bigcup\limits_{i\in I}\bigcup\limits_{U\in\alpha_i}(f_i^U)^{-1}(a),&
R^{-1}(a) =& C \cup \bigcup\limits_{i\in I} R_i^{-1}(a),
\end{align*}
and so
$$ f^{-1}(R^{-1}(a)) =
g^{-1}(C) \cup \bigcup\limits_{i\in I}\bigcup\limits_{U\in\alpha_i} (f_i^U)^{-1}(R_i^{-1}(a)).
$$
Since $f_i^U$ is a d-morphism,
\begin{equation}\label{eq:L69:2}
f^{-1}(R^{-1}(a)) =
g^{-1}(C) \cup\bigcup\limits_{i\in I}\bigcup\limits_{U\in\alpha_i} {\bf d}_U((f_i^U)^{-1}(a))
\subseteq g^{-1}(C) \cup {\bf d} f^{-1}(a).
\end{equation}
Let us show that
\begin{equation}\label{eq:L69:3} g^{-1}(C) \subseteq {\bf d} f^{-1}(a).
\end{equation}
In fact, let $x\in g^{-1}(C)$. Since
$\alpha_i$ is dense at
$x$, every neighbourhood of $x$
contains some $U\in\alpha_i$. Since $f^U_i$ is
surjective,
$f(u) = f^U_i (u) = a$ for some $u\in U$. Therefore, $x\in{\bf d} f^{-1}(a)$.
(\ref{eq:L69:2}) and (\ref{eq:L69:3}) imply $f^{-1}(R^{-1}(a)) \subseteq {\bf d} f^{-1}(a)$.
Let us prove the converse:
\begin{equation}\label{eq:L69:5} {\bf d} f^{-1}(a) \subseteq f^{-1}(R^{-1}(a)).
\end{equation}
We have $A_j \cap f^{-1}(a)={\varnothing}$ for $j\notin I$ and $A_j$ is open,
hence
$A_j \cap {\bf d} f^{-1}(a)={\varnothing}$.
Thus ${\bf d} f^{-1}(a) \subseteq g^{-1}(C) \cup A_i$.
Now $g^{-1}(C)\subseteq f^{-1}(R^{-1}(a))$ by (\ref{eq:L69:2}), so it remains to show that for any $i\in I$
\begin{equation}\label{eq:L69:7}
{\bf d} f^{-1}(a)\cap A_i\subseteq f^{-1}(R^{-1}(a)).
\end{equation}
To check this, consider any $x\in{\bf d} f^{-1}(a)\cap A_i$. Then $x\in U$ for some $U\in \alpha_i$, and thus by \ref{dY} and (\ref{eq:L69:2})
$x\in{\bf d} f^{-1}(a)\cap U={\bf d}_U(f^{-1}(a)\cap U)={\bf d}_U(f^U_i)^{-1}(a)\subseteq f^{-1}(R^{-1}(a))$.
This implies (\ref{eq:L69:7}) and completes the proof of (\ref{eq:L69:5}).
{\hspace*{\fill} }
\end{proof}
Recall that $\partial$ denotes the boundary of a set in a topological space: $\partial A := {\bf C} A - {\bf I} A$.
\begin{lem}\label{Glue}{\em (Glueing lemma)}
Let $\mathfrak X$ be a local $T_1$-space satisfying
\noindent
(a) $X=X_1\cup Y\cup X_2$ for closed nonempty subsets $X_1,Y,X_2$ such that
\begin{itemize}
\item
$X_1\cap X_2=X_1\cap{\bf I} Y= X_2\cap{\bf I} Y ={\varnothing}$,
\item
$\partial X_1\cup\partial X_2=\partial Y$,
\item
${\bf d}{\bf I} Y=Y$ (i.e., $Y$ is regular and dense in-itself).
\end{itemize}
or
(b) $X=X_1\cup X_2$ is a nontrivial closed partition.
Let $F=(W,R)$ be a finite ${\bf K4}$-frame,
$F_1=(W_1,R_1),~ F_2=(W_2,R_2)$ its generated subframes such that $W=W_1\cup W_2$ and suppose there are
d-morphisms $f_i:\mathfrak X_i\twoheadrightarrow^d F_i$, $i=1,\,2$, where $\mathfrak X_i$ is the subspace of $\mathfrak X$ corresponding to $X_i$.
In the case (a) we also assume that $F_1, F_2$ have a common maximal cluster $C$, $f_i(\partial
X_i)\subseteq R^{-1}(C)$ for $i=1,2$ and there is $g:{\bf I} Y\twoheadrightarrow^d C$ (where $C$ is regarded as
a frame with the universal relation, ${\bf I} Y$ as a subspace of $\mathfrak X$). Then
$f_1\cup f_2\cup g:\mathfrak X\twoheadrightarrow^d F$ in the case (a), $f_1\cup f_2:\mathfrak X\twoheadrightarrow^d F$ in the case
(b).\footnote{$f_1\cup f_2$ is the map $f$ such that $f|X_i=f_i$; similarly for $f_1\cup f_2\cup g$.}
\end{lem}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{fig-Glue.pdf}
\caption{Case (a)}
\label{fig:glueing_lemma}
\end{figure}
\begin{proof}
Let $f:= f_1\cup f_2\cup g $ (or $f:= f_1\cup f_2$), $ F_i=(W_i,R_i)$,
${\bf d}:={\bf d}_X,~{\bf d}_i:={\bf d}_{X_i}$.
For $w\in W$ there are four options.
(1) $w\in W_1-W_2$. Then
${\bf d} f^{-1}(w)={\bf d} f_1^{-1}(w)={\bf d}_1 f_1^{-1}(w) = f_1^{-1}(R_1^{-1}(w))$ (since $X_1$ is closed and $f_1$ is a d-morphism).
It remains to note that $R_1^{-1}(w)=R^{-1}(w)\subseteq W_1-W_2$, and thus
$f_1^{-1}(R_1^{-1}(w))= f^{-1}(R^{-1}(w)).$
(2) $w\in W_2-W_1$. Similar to the case (1).
(3) $w\in (W_1\cap W_2)-C$ in the case (a) or $w\in W_1\cap W_2$ in the case (b). Then $f^{-1}(w)=f_1^{-1}(w)\cup f_2^{-1}(w)$, so similarly to (1),
\[
{\bf d} f^{-1}(w)= {\bf d}_1 f_1^{-1}(w)\cup {\bf d}_2 f_2^{-1}(w) = f_1^{-1}(R_1^{-1}(w))\cup f_2^{-1}(R_2^{-1}(w))= f^{-1}(R^{-1}(w)).
\]
(4) $w\in C$ in case (a). First note that ${\bf d} g^{-1}(w)=Y$. In fact,
$g$ is a d-morphism onto the cluster $C$, so
${\bf d}_{{\bf I} Y} g^{-1}(w)= g^{-1}(C)={\bf I} Y$. Hence
${\bf I} Y\subseteq {\bf d} g^{-1}(w )\subseteq {\bf d}{\bf I} Y=Y$, and thus
$$Y={\bf d}{\bf I} Y\subseteq{\bf d}\Bd g^{-1}(w ) \subseteq {\bf d} g^{-1}(w )$$
by \ref{L310}(2).
Next, since $X_1$, $X_2$ are closed and $f_1$, $f_2$ are d-morphisms we have
\begin{align*}
{\bf d} f^{-1}(w) &= {\bf d} f_1^{-1}(w)\cup {\bf d} f_2^{-1}(w)\cup {\bf d} g^{-1}(w)
=
{\bf d}_1 f_1^{-1}(w)\cup {\bf d}_2 f_2^{-1}(w)\cup Y =&\\
&= f_1^{-1}(R_1^{-1}(w))\cup f_2^{-1}(R_2^{-1}(w))\cup Y= f^{-1}(R^{-1}(w)). &\qedhere
\end{align*}
\end{proof}
The case (b) of the previous lemma can be generalized as follows.
\begin{lem}\label{L67}
Suppose a topological space $\mathfrak X$ is the disjoint union of open subspaces:
$\mathfrak X=\bigsqcup\limits_{i\in I}\mathfrak X_i$. Suppose a Kripke ${\bf K4}$-frame $F$ is the
union of its generated subframes: $F=\bigcup\limits_{i\in I}F_i$
and suppose $f_i:\mathfrak X_i\twoheadrightarrow^{d}F_i$. Then $\bigcup\limits_{i\in I}f_i:\mathfrak X\twoheadrightarrow^{d}F$.
\end{lem}
\begin{defi}\label{def_pmorphism}
Let $\mathfrak X$ be a topological space,
$F= (W, R, R_D)$ be a frame. Then a surjective map $f: X \longrightarrow W$ is called a
{\em dd-morphism}
(in symbols, $f:\mathfrak X \twoheadrightarrow^{dd} F$) if
\begin{enumerate}
\item $f:\mathfrak X \pmor^d{} (W, R)$ is a d-morphism ;
\item $f:(X, \ne_X) \twoheadrightarrow (W, R_D)$ is a p-morphism of Kripke frames.
\end{enumerate}
\end{defi}
\begin{lem}\label{lem_pmorphism}
If $f:\mathfrak X \twoheadrightarrow^{dd} F$, then ${{\bf L}}{\bf d}_{\neq}(\mathfrak X) \subseteq {{\bf L}}(F)$ and for any closed 2-modal $A$
\[
\mathfrak X\vDash A\Leftrightarrow F\vDash A.
\]
\end{lem}
\begin{proof}
Similar to \ref{P52} and \ref{p1}.
{\hspace*{\fill}
}
\end{proof}
\begin{defi}
A set-theoretic map $f:X\longrightarrow Y$ is called {\em n-fold at} $y\in Y$ if $|f^{-1}(y)|=n$;\footnote{$|\ldots|$ denotes the cardinality.}
$f$ is called {\em manifold at} $y$ if it $n$-fold for some $n>1$.
\end{defi}
\begin{propo}\label{lem:dmortodDmor}~
(1) Let $G = (X, \ne_X)$, $F = (W, S)$ be Kripke frames such that $\overline{S}=W^2$, and
let $f : X \longrightarrow W$ be a surjective function. Then
$$f : G \twoheadrightarrow F \quad\;\mbox{iff}\;\quad f\mbox{ is manifold exactly at }S\mbox{-reflexive points of }F.$$
(2)
Let $\mathfrak X$ be a $T_1$-space, $F= (W, R, R_D)$ a rooted
$\lgc{KT_1}$-frame, $f:\mathfrak X\twoheadrightarrow^d(W,R)$. Then $f:\mathfrak X\twoheadrightarrow^{dd}F$ iff for any strictly $R$-minimal $v$
\[
v R_D v\Leftrightarrow f\mbox{ is manifold at }v.
\]
(3)
If $\mathfrak X$ is a $T_1$-space, $f: \mathfrak X \pmor^d{} F=(W, R)$ and $R^{-1}(w)\neq{\varnothing}$ for any $w\in W$, then $f:\mathfrak X \twoheadrightarrow^{dd} F_{\forall}$,
where $F_{\forall}:= (W, R, W^2)$.
\end{propo}
\begin{proof}
(1) Note that
$f$ is a p-morphism iff for any $x\in X$
\[
f(X-\{x\})=S(f(x))=
\begin{cases}
W & {\rm if~}f(x)Sf(x),\nonumber\\
W-\{f(x)\} & {\rm otherwise.}\nonumber
\end{cases}
\]
(2) By (1), $f:\mathfrak X\twoheadrightarrow^{dd}F$ iff
\[
\forall v\in W(v R_D v\Leftrightarrow \abs{f^{-1}(v)} > 1).
\]
The latter equivalence holds whenever $ R^{-1}(v)\neq{\varnothing}$. In fact, then by Corollary \ref{L52},
${\bf d} f^{-1} (v) = f^{-1}(R^{-1}(v)) \neq{\varnothing}$, and thus $f^{-1}(v)$ is not a singleton (since $\mathfrak X$ is
a $T_1$-space). $ R^{-1}(v)\neq{\varnothing}$ also implies $vR_Dv$, by Proposition \ref{pr:DS_AT1}.
(3) follows from (2).
{\hspace*{\fill}
}
\end{proof}
After we have proved the main technical results, in the next sections we will study dd-logics of specific spaces.
\section{$\lgc{D4}$ and $\lgc{DT_1}$ as logics of zero-dimensional dense-in-themselves spaces}
In this section we will prove the d-completeness of $\lgc{D4}$ and
dd-completeness of $\lgc{DT_1}$ w.r.t. zero-dimensional spaces. The proof follows rather easily from the previous section and an additional technical fact (Proposition \ref{P61}) similar to the McKinsey--Tarski lemma.
Recall that a (nonempty) topological space $\mathfrak X$ is called
{\em zero-dimensional}
if clopen sets
constitute its open base \cite{A77}.
Zero-dimensional $T_1$-spaces with a countable base are subspaces of the Cantor
discontinuum, or of
the
set of irrationals \cite{K66}.
\begin{lem}\label{L71}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself Hausdorff space.
Then for any $n$ there exists a nontrivial open
partition $\mathfrak X=\mathfrak X_1\sqcup\ldots\sqcup\mathfrak X_n$, in which every $\mathfrak X_i$ is also a zero-dimensional dense-in-itself Hausdorff space.
\end{lem}
\begin{proof}
It is sufficient to prove the claim for $n=2$ and then apply induction.
A dense-in-itself space cannot be a singleton, so there are two different points $x,y\in X$. Since $\mathfrak X$ is $T_1$ and zero-dimensional, there exists a clopen $U$ such that $x\in U,~y\not\in U$. So
$X=U\cup (X-U)$ is a nontrivial open partition.
The Hausdorff property, density-in-itself, zero-dimensionality are inherited for open subspaces.
{\hspace*{\fill}}
\end{proof}
\begin{propo}\label{P61}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself metric space, $y\in X$. Let
$\Psi_l$ be the
frame consisting of an irreflexive root $b$ and its reflexive successors
$a_0,\dots,a_{l-1}$ (Fig. \ref{fig:Phiml_irreflexive_root}).
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{fig-3.pdf}
\caption{Frame $\Psi_l$.}
\label{fig:Phiml_irreflexive_root}
\end{figure}
Then there exists $f: \mathfrak X\twoheadrightarrow^d \Psi_l$ such that $f(y) = b$ and for every $i$
there is an open partition of $f^{-1}(a_i)$, which is dense at y.
\end{propo}
\begin{proof}
Let $O(a,r) := \setdef[x\in X]{\rho(a,x)<r}$, where $\rho$ is the distance in $\mathfrak X$.
There exist clopen sets $Y_0, Y_1, \dots$ such that
$$\set{y}\subset\dots\subset Y_{n+1} \subset Y_n \subset\dots Y_1 \subset Y_0 = X$$ \
and
$Y_n \subseteq O(y, 1/n)$ for $n >0$.
These $Y_n$ can be easily constructed by induction. Then
$$\bigcap\limits_n Y_n =\set{y},
\ \hbox{and}\
X-\{ y\} =\bigsqcup\limits_n X_n, $$
where $X_n = Y_n-Y_{n+1}$. Note that the $X_n$ are nonempty and
open, $X_n \subseteq O(y, 1/n)$ for $n>0$.
Now define a map $f: X\longrightarrow\Psi_l$ as follows:
\[
f(x)=\left\{ \begin{array}{ll}
a_{r(n)} & \mbox{if } x\in X_n; \\
b & \mbox{if } x=y,\\
\end{array} \right.
\]
where $r(n)$ is the remainder of dividing $n$ by $l$; it is clear that $f$ is surjective.
Let us show that for any $x$,
\[
x\in {\bf d} f^{-1}(u) \mbox{ iff }f(x)Ru.
\eqno(*)
\]
(i) Assume that $u=a_j$. Then
$f^{-1}(u)=\bigcup\limits_n X_{nl+j}$,
and
$$f(x)Ru \mbox{ iff }(f(x)=b \mbox{ or } f(x)=u).$$
To prove `if' in (*), consider two cases.
1. Suppose $f(x)=u, ~x\in X_{nl+j}$. Since $X_{nl+j}$ is nonempty and open, it is
dense-in-itself, and thus $x\in {\bf d} X_{nl+j}\subseteq {\bf d} f^{-1}(u)$.
2. Suppose $f(x)=b$, i.e. $x=y$. Then $x\in {\bf d} f^{-1}(u)$, since
$X_{nl+j}\subseteq O(y, 1/n)$.
The previous argument also shows that
$\{ X_{nl+j}\mid n\geq 0\}$ is an open partition of $f^{-1}(a_j)$, which
is dense at $y$.
To prove `only if', suppose $f(x)Ru$ is not true. Then
$f(x) = a_k$ for some $k \not= j$, and so for some $n$, $x\in
X_n$, $X_n\cap f^{-1}(u)={\varnothing}$. Since $X_n$ is open, $x\not\in {\bf d} f^{-1}(u)$.
(ii) Assume that $u = b$. Then $f^{-1}(u) = \{ y\}$, and so ${\bf d} f^{-1}(u) =
{\varnothing} = f^{-
1}(R^{-1}(u))$.
{\hspace*{\fill} }\end{proof}
\begin{propo}\label{P64}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself separable metric space, $F$ a
finite rooted ${\bf D4}$-frame.
Then there exists a d-morphism $\mathfrak X\twoheadrightarrow^dF$, which is 1-fold at the root of $F$
if this root is irreflexive.
\end{propo}
\begin{proof}
By induction on the size of $F$.
(i) If $F$ is a finite cluster, the claim follows from Proposition \ref{P54}.
(ii) If $F = C\cup F_1\cup \dots\cup F_l$, where $C = \{ b_1,\dots,b_m\} $ is a non-degenerate
root cluster,
$F_1,\dots,F_l$ are the subframes generated by the successors of $C$, we can apply Lemma \ref{refroot}. In fact, every open ball $U$ in $\mathfrak X$ is zero-dimensional and dense-in-itself.
(iii) Suppose $F=\breve{b}\cup F_0\cup\dots\cup F_{l-1}$, where $b$ is an irreflexive
root of $F$, $F_i$ are the subframes generated by the successors of $b$.
There exists
$g: X\twoheadrightarrow^d \Psi_l$ by \ref{P61}, with an arbitrary
$y\in X$. Then $g^{-1}(a_i)$ is a union of a set $\alpha_i$ of disjoint open
sets, and $\alpha_i$ is dense
at
$y$. If $U\in\alpha_i$, then by IH, there exists
$f^U_i: U\twoheadrightarrow^d F_i$. Put
\[
f(x)=\left\{ \begin{array}{ll}
b & \mbox{ if } x=y; \\
f_i^U(x)& \mbox{ if } x\in U,~U\in\alpha_i.\\
\end{array} \right.
\]
Then similarly to Lemma \ref{refroot} it follows that $f: X\twoheadrightarrow^d F$.
Finally note that if the root of $F$ is irreflexive, the first step of the construction is case (iii),
so the preimage of the root is a singleton.
{\hspace*{\fill} }
\end{proof}
\begin{theorem}\label{T65}
If $\mathfrak X$ is a zero-dimensional dense-in-itself separable metric space, then ${{\bf L}}{\bf d} (\mathfrak X)={\bf D4}$.
\end{theorem}
\begin{proof}
By Propositions \ref{P64}
and \ref{P52} ${{\bf L}}{\bf d} (\mathfrak X) \subseteq {{\bf L}} (F)$ for any finite rooted ${\bf D4}$-frame $F$, thus ${{\bf L}}{\bf d} (\mathfrak X) \subseteq {\bf D4}$, since ${\bf D4}$ has the fmp. By Lemma \ref{L310}
${\bf D4}\subseteq{{\bf L}}{\bf d} (\mathfrak X)$ . {\hspace*{\fill} }\end{proof}
\begin{lem}\label{P741}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself separable metric
space, $F$ a finite ${\bf D4}$-frame. Then there exists a d-morphism $\mathfrak X\twoheadrightarrow^d F$, which is 1-fold at all strictly minimal points.
\end{lem}
\begin{proof}
$F = F_1\cup\ldots\cup F_n$ for different finite rooted ${\bf D4}$-frames $F_i$.
By Lemma \ref{L71}, $\mathfrak X=\mathfrak X_1\sqcup\ldots\sqcup\mathfrak X_n$ for zero-dimensional dense-in-themselves subspaces
$\mathfrak X_i$, which are also metric and separable.
By Proposition \ref{P64}, we construct $f_i:\mathfrak X_i\twoheadrightarrow^d F_i$. Then by Lemma \ref{L67}, $\bigcup\limits_{i=1}^nf_i:\mathfrak X\twoheadrightarrow^d F$. Every strictly minimal point of $F$ is an irreflexive root of a unique $F_i$, so its preimage is a singleton.
{\hspace*{\fill} }
\end{proof}
\begin{propo}\label{P:dd-morph_zero_dem}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself separable metric space, $F \in
\mathfrak{F}_0$ a finite $\lgc{DT_1}$-frame. Then there exists a dd-morphism $\mathfrak X\twoheadrightarrow^{dd}F$.
\end{propo}
\begin{proof}
We slightly modify the proof of the previous lemma.
Let $F=(W,R,R_D)$, $G=(W,R)$. Then $G = G_1\cup\ldots\cup G_n$ for different cones $G_i$.
We call $G_i$ {\em special} if its root is strictly $R$-minimal
and $R_D$-reflexive.
We may assume that exactly $G_1,\ldots,G_m$ are special.
Then we count them twice and present $G$ as
$G_1\cup G'_1\cup\ldots\cup G_m\cup G'_m\cup G_{m+1}\cup\ldots\cup G_n$, where $G'_i=G_i$ for $i\leq m$ (or as $G_1\cup G'_1\cup\ldots\cup G_m\cup G'_m$ if $m=n$).
Now we can argue as in the proof of Lemma \ref{P741}. By Lemma \ref{L71},
$\mathfrak X=\mathfrak X_1\sqcup\mathfrak X'_1\sqcup\ldots\sqcup\mathfrak X_m\sqcup \mathfrak X'_m\sqcup \mathfrak X_{m+1}\sqcup\ldots\sqcup \mathfrak X_n$ for
zero-dimensional dense-in-itself separable metric $\mathfrak X_i, \mathfrak X'_i$. By Proposition \ref{P64}, we
construct the maps $f_i:\mathfrak X_i\twoheadrightarrow^d G_i$, $f'_i:\mathfrak X'_i\twoheadrightarrow^d G'_i$, which are 1-fold at irreflexive
roots; hence by Lemma \ref{L67}, $f:\mathfrak X\twoheadrightarrow^d G$ for $f:= \bigcup\limits_{i=1}^nf_i\cup
\bigcup\limits_{i=1}^mf'_i $.
Every strictly minimal point $a \in G$ is an irreflexive root of a unique $G_i$.
If $a$ is $R_D$-irreflexive, then $G_i$ is not special, so
$f^{-1}(a)=f_i^{-1}(a)$ is a singleton.
If $a$ is $R_D$-reflexive, then $G_i$ is special, so $f^{-1}(a)= f_i^{-1}(a)\cup (f'_i)^{-1}(a)$, and thus $f$ is 2-fold at $a$.
Therefore, $f:\mathfrak X\twoheadrightarrow^{dd} F$
by Proposition \ref{lem:dmortodDmor}.
{\hspace*{\fill} }
\end{proof}
\begin{lem}\label{LF1}
Let $M = (W, R, R_D, \varphi)$ be a rooted Kripke model over a basic frame\footnote{Basic frames were defined in Section 4.} validating
$ AT_1$, $\Psi$ a set of 2-modal formulas closed under subformulas. Let $M'=(W',R',R_D',\theta')$ be a
filtration of $M$ through $\Psi$ described in Lemma \ref{L26}\footnote{Recall that $R^\prime$ is
the transitive closure of ${\underline{R}}$, $R_D'=\underline{R_D}$.}. Then the frame $(W',R',R_D')$ is also
basic and validates $AT_1$.
\end{lem}
\begin{proof}
In fact, $R'$ is transitive by definition. For any two different $a, b\in W'$ we have $a R_D'b$,
since $xR_Dy$ for any $x\in a,~y\in b$ (as $F\in\mathfrak{F}_0$).
Next, note that if $a$ is $ R_D'$-irreflexive, then $a=\{x\}$ for some $R_D$-irreflexive $x$. In this case,
since $(W, R, R_D)\vDash AT_1$, there is no $y$ such that $yRx$ (Proposition \ref{pr:DS_AT1}),
hence $(R')^{-1}(a)={\varnothing}$, and thus $(W',R',R_D')\vDash AT_1$.
Finally, $R'\subseteq R_D'$. In fact, all different points in $F'$ are $ R_D'$-related, so it remains to
show that every $ R_D'$-irreflexive point is $R'$-irreflexive. As noted above, such a point is a
singleton class $x^\sim=\{x\}$, where $x$ is $R_D$-irreflexive. Then $x$ is $R$-minimal, so in $W'$
there is no loop of the form $x^\sim{\underline{R}} x_1{\underline{R}}\ldots{\underline{R}} x^\sim$, and thus $x^\sim$ is
$R'$-irreflexive. {\hspace*{\fill} }
\end{proof}
By a standard argument Lemma \ref{LF1} implies
\begin{theorem}\label{Tfmp}
Every logic of the form $\lgc{KT_1}+A$, where $A$ is a closed
2-modal formula, has the finite model property.
\end{theorem}
\begin{proof}
Let $L$ be such a logic and suppose $L\,{\not\vdash}\, B$. By Proposition \ref{P53} $L$ is Kripke complete,
so by the Generation lemma there is a rooted Kripke frame $F=(W,R,R_D)$ such that $F\vDash L,~F{\,\not\mo\,} B$.
Then $F$ is basic by definition. Let $M=(F,\theta)$ be a Kripke model over $F$ refuting $B$. Let $\Psi$
be the set of all subformulas of $A$ or $B$, and let us construct the filtration
$M'=(W',R',R_D',\theta')$ of $M$ through $\Psi$ as in Lemmas \ref{L26}(2) and \ref{LF1}. By the previous
lemma, $F':= (W',R',R_D')\vDash\lgc{KT_1}$.
By the Filtration lemma, $M'{\,\not\mo\,} B$. By the same lemma, the truth of $A$ is preserved in $M'$, so $F'\vDash A$, since $A$ is closed. Therefore, $F'\vDash L$.
\end{proof}
\begin{theorem}\label{thm:DLogic_zerospace}
Let $\mathfrak X$ be a zero-dimensional dense-in-itself separable metric space. Then ${{\bf L}}{\bf d}_{\neq}
(\mathfrak X)=\lgc{DT_1}$.
\end{theorem}
\begin{proof}
For any finite $\lgc{DT_1}$-frame $F$ we have ${{\bf L}}{\bf d}_{\neq} (\mathfrak X) \subseteq {{\bf L}} (F)$ by Proposition
\ref{P:dd-morph_zero_dem} and Lemma \ref{lem_pmorphism}. By the previous theorem, $\lgc{DT_1}$ has
the fmp, so ${{\bf L}}{\bf d}_{\neq} (\mathfrak X) \subseteq \lgc{DT_1}$. Since $\mathfrak X\vDash^d \lgc{DT_1}$ (Proposition
\ref{P412}), it follows that ${{\bf L}}{\bf d}_{\neq} (\mathfrak X) = \lgc{DT_1}$. {\hspace*{\fill} }
\end{proof}
\begin{propo}{\cite[Lemma 3.1]{BLB}}\label{prop:QtoF}
Every countable\footnote{In this chapter, as well as in \cite{BLB}, `countable' means `of cardinality at most $\aleph_0$'.} rooted $\lgc{K4}$-frame is a
d-morphic image of a subspace of ${\bf Q}$.
\end{propo}
To apply this proposition to the language with the difference
modality, we need to examine the preimage of the root
for the constructed morphism.
Fortunately, in the proof of Proposition \ref{prop:QtoF} in \cite{BLB}
the preimage of a root $r$ is a singleton iff $r$ is
irreflexive.
\begin{lem}\label{prop:QtoF_1fold}
Let $F$ be a countable $\lgc{K4}$-frame. Then there exists a
d-morphism from a subspace of ${\bf Q}$ onto $F$, which is 1-fold at
all strictly minimal points.
\end{lem}
\begin{proof}
Similar to Lemma \ref{P741}. We can present $F$ as a countable union of different cones
$\bigcup\limits_{i\in I}F_i$ and ${\bf Q}$ as a disjoint union $\bigsqcup\limits_{i\in I}\mathfrak X_i$ of
spaces homeomorphic to ${\bf Q}$. By Proposition \ref{prop:QtoF} (and the remark after it), for
each $i$ there exists $f_i: \mathcal Y_i \pmor^d{} F_i$ for some subspace $\mathcal Y_i \subseteq \mathfrak X_i$ such that $f_i$ is
1-fold at the root $r_i$ of $F_i$ if $r_i$ is irreflexive. Now by Lemma \ref{L67} $f:=
\bigcup\limits_{i\in I}f_i: \bigsqcup\limits_{i\in I}\mathcal Y_i \twoheadrightarrow^{d}F$, and $f$ is 1-fold at all
strictly minimal points of $F$ (i.e., the irreflexive $r_i$) --- since every $r_i$ belongs only to
$F_i$, so $f^{-1}(r_i)=f_i^{-1}(r_i)$. {\hspace*{\fill} }
\end{proof}
\begin{propo}\label{prop:subsetQtoF_ddmorph}
Let $F$ be a countable $\lgc{KT_1}$-frame. Then there exists a
dd-morphism from a subspace of ${\bf Q}$ onto $F$.
\end{propo}
\begin{proof}
Similar to Proposition \ref{P:dd-morph_zero_dem}. If $F = (W, R, R_D)$, the frame $G=(W,R)$ is a
countable union of different cones. There are two types of cones: non-special $G_i~ (i\in I)$ and
special (with strictly $R$-minimal and $R_D$-reflexive roots) $H_j~ (j\in J)$:
\[
G = \bigcup\limits_{i \in I} G_i \cup \bigcup\limits_{j \in J} H_j.
\]
Then we duplicate all special cones
\[
G = \bigcup\limits_{i \in I} G_i \cup \bigcup\limits_{j \in J} H_j \cup \bigcup\limits_{j \in J} H'_j
\]
and as in the proof of \ref{prop:QtoF_1fold}, construct $f: \bigsqcup\limits_{i\in
I}\mathcal Y_i\sqcup\bigsqcup\limits_{j\in J}\mathcal Z_j\sqcup\bigsqcup\limits_{j\in J}\mathcal Z'_j \twoheadrightarrow^{d}F$. This map is
1-fold exactly at all $R_D$-irreflexive points, so it is a dd-morphism onto $F$. {\hspace*{\fill}
}
\end{proof}
\begin{coro}\label{C713}
$\mathbf{Ld}_{\ne}(\mbox{all $T_1$-spaces}) = \lgc{KT_1}$.
\end{coro}
\begin{proof}
Note that $\lgc{KT_1}$ is complete w.r.t.~countable frames and every subspace of ${\bf Q}$ is $T_1$.
{\hspace*{\fill} }
\end{proof}
\begin{propo}\label{P714}
Let $\mathbf{\Lambda} = \lgc{KT_1} + \Gamma$ be a consistent logic, where
$\Gamma$ is a set of closed formulas. Then $\mathbf{\Lambda}$ is
dd-complete w.r.t.\/ subspaces of ${\bf Q}$.
\end{propo}
\begin{proof}
Since every closed formula is canonical, $\mathbf{\Lambda}$ is Kripke complete. So for every formula $A
\notin \mathbf{\Lambda}$ there is a frame $F_A$ such that $F_A \models \mathbf{\Lambda}$ and $F_A \nvDash A$. By
Proposition \ref{prop:subsetQtoF_ddmorph}, there is a subspace $\mathfrak X_A \subseteq {\bf Q}$ and $f_A: \mathfrak X_A \twoheadrightarrow^{dd} F_A$. Then $\mathfrak X_A{\,\not\mo\,} A,~\mathfrak X_A\vDash\mathbf{\Lambda}$ by Lemma \ref{lem_pmorphism}.
Therefore ${{\bf L}}{\bf d}_{\neq} (\mathcal K) = \mathbf{\Lambda}$ for
$\mathcal K := \setdef[\mathfrak X_A]{A \notin \mathbf{\Lambda}}$.
{\hspace*{\fill} }
\end{proof}
\begin{rem}\rm
A logic of the form described in Proposition \ref{P714} is dd-complete w.r.t. a set of subspaces of ${\bf Q}$. This set may be non-equivalent to a single subspace. For example,
there is no subspace $\mathfrak X\subseteq{\bf Q}$ such that $\lgc{KT_1}={{\bf L}}{\bf d}_{\neq}(\mathfrak X)
$. In fact, consider
\[
A:=[\neq]\square\bot\wedge\square\bot.
\]
Then $A$ is satisfiable in $\mathfrak X$ iff $\mathfrak X\vDash^d A$ iff $\mathfrak X$ is discrete. So $A$ is consistent in
$\lgc{KT_1}$. Now if $\lgc{KT_1}={{\bf L}}{\bf d}_{\neq}(\mathfrak X)$, then $A$ must be satisfiable in $\mathfrak X$, hence
$\mathfrak X\vDash^d A$; but $\lgc{KT_1}\not\vdash A$, and so we have a contradiction.
\end{rem}
\section{Connectedness}
Connectedness was the first example of a property expressible in
cu-logic, but not in c-logic. The corresponding connectedness axiom from \cite{Sh99} will be essential for our further studies.
In this section we show that it is weakly canonical, i.e., valid in weak canonical frames --- a fact not mentioned in \cite{Sh99}.
\begin{lem}\label{L80}\cite{Sh99}
A topological space $\mathfrak X$ is connected iff $\mathfrak X\vDash^c AC$, where
\[
AC :=[\forall](\square p\vee\square\neg p)\to [\forall] p\vee[\forall]\neg p.
\]
\end{lem}
For the case of Alexandrov topology there is an equivalent definition of connectedness in relational terms.
\begin{defi}\label{D75}
For a transitive Kripke frame $F = (W,R)$ we define the
{\em comparability} relation $R^\pm:=R\cup R^{-1}\cup I_W$.
$F$ is called {\em connected} if the transitive closure of $R^\pm$ is universal.
A subset $V\subseteq W$ is called {\em connected} in $F$ if the frame $F|V$ is connected.
A 2-modal frame $(W,R,S)$ is called \emph{(R)-connected} if $(W,R)$ is connected.
\end{defi}
Thus $F$ is connected iff every two points $x,y$ can be connected by a {\em non-oriented path}
(which we call just a {\em path}), a sequence of points $x_0x_1\ldots x_n$ such that
$x=x_0 R^\pm x_1\ldots R^\pm x_n=y$.
From \cite{Sh99} and Proposition \ref{P49} we obtain
\begin{lem}\label{L801}
(1) For an ${\bf S4}$-frame $F$, the associated space $N(F)$ is connected iff $F$ is connected.\\
(2) For a ${\bf K4}$-frame $F$, $F_{\forall}\models AC^{\sharp u}$ iff $F$ is connected.
\end{lem}
\begin{lem}\label{mcl}
Let $M=(W,R,R_D,\theta)$ be a rooted generated submodel of m-weak canonical model for a modal logic
$\mathbf{\Lambda}\supseteq{\bf K4D^+}$. Then
\begin{enumerate}
\item Every $R$-cluster in $M$ is finite of cardinality at most $2^m$.
\item
$(W,R)$ has finitely many $R$-maximal clusters.
\item
For each $R$-maximal cluster $C$ in $M$ there exists an $m$-formula $\beta(C)$ such that:
$$\forall x\in M~ (M,x \vDash \beta(C)\Leftrightarrow x\in\overline{R}^{-1} (C)).$$
\end{enumerate}
\end{lem}
The proof is similar to \cite[Section 8.6]{CZ97}.
\begin{lem}\label{lem_KripkeCompletAC}
Every rooted generated subframe of a weak canonical
frame for a logic
$\mathbf{\Lambda}\supseteq{\bf K4D^+} + AC^{\sharp u} $ is connected.
\end{lem}
\begin{proof}
Let $M$ be a weak canonical model for $\mathbf{\Lambda}$, $M_0$ its rooted generated submodel with the frame
$F=(W,R,R_D)$, and suppose $F$ is disconnected. Then there exists a nonempty proper clopen subset $V$
in the space $N(W,\overline{R})$. Let $\Delta$ be the set of all $R$-maximal clusters in $V$ and put
\[
B := \bigvee\limits_{C \in \Delta} \beta(C).
\]
Then $B$ defines $V$ in $M_0$, i.e., $V=\overline{R}^{-1}(\bigcup{\Delta})$.
In fact, $\bigcup{\Delta}\subseteq V$ implies $\overline{R}^{-1}(\bigcup{\Delta}) \subseteq V $, since $V$ is closed. The other way round, $V\subseteq\overline{R}^{-1}(\bigcup{\Delta})$, since for any $v\in V$, $\overline{R}(v)$ contains an $R$-maximal cluster $C\in{\Delta}$, and $\overline{R}(v)\subseteq V$ as $V$ is open.
So $w \models B$ for any $w \in V$, and since $V$ is open, $w \models \overline{\square} B$. By the same reason,
$w\models \overline{\square}\lnot B$ for any $w \not \in V $ . Hence
\[
M_0 \models \left[\forall\right] (\overline{\square} B \lor \overline{\square} \lnot B).
\]
By Proposition \ref{cm} all substitution instances of $AC$ are true in $M_0$. So we have
\[
M_0 \models \left[\forall\right] (\overline{\square} B \lor \overline{\square} \lnot B) \rightarrow \left[\forall\right] B \lor \left[\forall\right] \lnot B,
\]
and thus
\[
M_0 \models \left[\forall\right] B \lor \left[\forall\right] \lnot B.
\]
This contradicts the fact that $V$ is a nonempty proper subset of
$W$. {\hspace*{\fill} }
\end{proof}
In d-logic instead of connectedness we can express some its local versions; they will be considered in
the next section.
\section{Kuratowski formula and local 1-componency}
In this section we briefly study Kuratowski formula distinguishing ${\bf R}$ from ${\bf R}^2$ in d-logic. Here the main proofs are similar to the previous section, so most of the details are left to the reader.
\begin{defi}
We define {\em
Kuratowski formula} as
\[
Ku:= \quad \square (\overline{\square} p \vee \overline{\square} \neg p) \rightarrow \square p \vee \square \neg p.
\]
\end{defi}
The spaces validating $Ku$ are characterized as follows \cite{LB2}.
\begin{lem}\label{lem:th32_from_LB2}
For a topological space $\mathfrak X$, $\mathfrak X\vDash^d Ku$ iff
\smallskip
\noindent
for any $x\in X$ and any open neighbourhood $U$ of $x$, if $U-\{x\}$ is a disjoint union
$V_1\cup V_2$ of sets open in the subspace $U-\{x\}$, then there exists a neighbourhood\footnote{In
\cite{LB2} neighbourhoods are supposed open, but this does not matter here, since every neighbourhood
contains an open neighbourhood.} $V \subseteq U$ of $x$ such that $V-\{x\} \subseteq V_1$ or $V-\{x\}
\subseteq V_2$.
\end{lem}
\begin{defi}
A topological space $\mathfrak X$ is called {\em locally connected} if every neighbourhood of any point $x$
contains a connected neighbourhood of $x$. Similarly, $\mathfrak X$ is called {\em locally 1-component} if
every punctured neighbourhood of any point $x$ contains a connected punctured neighbourhood of $x$.
\end{defi}
It is well known \cite{A77} that in a locally connected space every neighbourhood $U$ of any point $x$
contains a connected {\em open} neighbourhood of $x$ (e.g. the connected component of $x$ in ${\bf I} U$).
\begin{lem}\label{LK1}
If $\mathfrak X$ is locally 1-component, then
$\mathfrak X\vDash^d Ku$.
\end{lem}
The proof is straightforward, and we leave it to the reader.
\begin{lem}\label{LK2}~
(1) Every space d-validating $Ku$ has the following \emph{non-splitting property}:
(NSP) If an open set $U$ is connected, $x\in U$ and $U-\{x\}$ is open, then $U-\{x\}$ is connected.\\
(2)
Suppose $\mathfrak X$ is locally connected and local $T_1$. Then (NSP) holds in $\mathfrak X$ iff $\mathfrak X$ is locally 1-component iff $\mathfrak X\vDash^d Ku$.
\end{lem}
\begin{proof}
(1) We assume $\mathfrak X\vDash^d Ku$ and check (NSP). Suppose $U$ is open and connected, $U^\circ:=U-\{x\}$ is
open, and consider a partition $U^\circ=U_1\cup U_2$ for open $U_1,U_2$. By
\ref{lem:th32_from_LB2} there exists an open $V\subseteq U$ containing $x$ such that $V\subseteq\{x\}\cup U_1$
or $V\subseteq\{x\}\cup U_2$. Consider the first option (the second one is similar). We have a partition
\[
U=(\{x\}\cup U_1)\cup U_2,
\]
and $\{x\}\cup U_1= V\cup U_1$, so $\{x\}\cup U_1$ is open. Hence by connectedness, $U= \{x\}\cup U_1$, i.e.,
$U^\circ = U_1$. Therefore,
$U^\circ $ is connected.
(2) It suffices to show that (NSP) implies the local 1-componency. Consider $x\in X$ and its neighbourhood $U_1$. Since $\mathfrak X$ is local
$T_1$, $U_1$ contains an open neighborhood $U_2$, in which $x$ is closed, i.e., ${\bf C}\{x\}\cap
U_2=\{x\}$. By the local connectedness, $U_2$ contains a connected open neighbourhood $U_3$, and
again ${\bf C}\{x\}\cap U_3=\{x\}$; thus $U_3-\{x\}$ is open. Eventually, $U_3-\{x\}$ is connected, by
(NSP). {\hspace*{\fill} }
\end{proof}
\begin{rem}\rm
The {\em
($n$-th)
generalized Kuratowski formula} is the following formula in variables $p_0,\dots,p_n$
$$Ku_n := \square \bigvee\limits^n_{k=0}\overline{\square} Q_k \rightarrow
\bigvee\limits^n_{k=0}\square\neg Q_k,$$
where $Q_k := p_k \wedge\bigwedge\limits_{j\not= k}\neg p_j$.
The formula $Ku_1$ is related to the equality found by Kuratowski \cite{K22}:
$$(*)\quad{\bf d} ((x \cap {\bf d} (-x)) \cup (-x \cap {\bf d} x)) = {\bf d} x \cap {\bf d} (-x),$$
which holds in every algebra $DA({\bf R}^n)$ for
$n>1$, but
not in $DA({\bf R})$. This equality corresponds to the modal formula
$$Ku^\prime:=\quad \Diamond ((p \wedge \Diamond \neg p) \vee (\neg p \wedge \Diamond p)) {\leftrightarrow}\Diamond p
\wedge \Diamond \neg p,$$
and one can show that ${\bf D4} + Ku^\prime = {\bf D4} + Ku_1={\bf D4} + Ku.$
\end{rem}
\begin{rem}\rm
The class of spaces validating $Ku_n$ is described in \cite{LB2}. In particular, it is valid in all locally $n$-component spaces defined as follows.
A neighbourhood $U$ of a point $x$ in a topological space is called
{\em $n$-component
at $x$} if the punctured neighbourhood $U-\set{x}$ has at most $n$
connected components.
A topological space is called {\em locally $n$-component} if
the $n$-component neighbourhoods at each of its point constitute a local base (i.e., every neighbourhood contains an $n$-component neighbourhood).
\end{rem}
\begin{lem}\label{L77}\cite{LB2}
For a transitive Kripke frame $(W,R)$
$(W,R) \vDash Ku$ iff for any $R$-irreflexive $x$, the subset $R(x)$ is connected (in the sense of Definition \ref{D75}).
\end{lem}
\begin{theorem}\label{th:D4KU1_KripCompl}
The logics ${\bf K4}+Ku$, ${\bf D4}+Ku$ are weakly canonical, and thus Kripke complete.
\end{theorem}
A proof of \ref{th:D4KU1_KripCompl} based on Lemma \ref{L77} and a 1-modal version of Lemma \ref{mcl} is straightforward, cf.
\cite{Sh90} or \cite{LB2} (the latter paper proves the same for $Ku_n$).
Hence we obtain
\begin{theorem}
The logic $\lgc{DT_1K}:=\lgc{DT_1}+Ku$ is weakly canonical, and thus Kripke complete.
\end{theorem}
\begin{proof}
(Sketch.) For the axiom $Ku$ the argument from the proof of \ref{th:D4KU1_KripCompl} is still valid due to definability of all maximal clusters (Lemma \ref{mcl}). The remaining axioms are
Sahlqvist formulas.
{\hspace*{\fill} }\end{proof}
\begin{theorem}\label{th:DT_1CK_KripCompl
The logic $\lgc{DT_1CK}:=\lgc{DT_1K}+AC^{\sharp u}$ is weakly canonical, and thus Kripke complete.
\end{theorem}
\begin{proof}
We can apply the previous theorem and
Lemma \ref{lem_KripkeCompletAC}.
{\hspace*{\fill} }\end{proof}
Completeness theorems from this section can be refined: in the next section we will prove the fmp for the logics considered above.
\section{The finite model property of ${\bf D4K}$, $\lgc{DT_1K}$, and $\lgc{DT_1CK}$}
For the logic ${\bf D4}+Ku$ the first proof of the fmp was given in \cite{Sh90}. Another proof (also for ${\bf D4}+Ku_n$) was proposed by M. Zakharyaschev \cite{Z}; it is based on a general and powerful method.
In this section we give a simplified version of the proof from \cite{Sh90}. It is based on a standard filtration method, and the same method is also applicable to 2-modal logics $\lgc{DT_1K}$, $\lgc{DT_1CK}$.
\begin{theorem}\label{thm:fmp_DT_1CK}
The logics $\lgc{DT_1K}$ and $\lgc{DT_1CK}$ have the finite model property.
\end{theorem}
\begin{proof}
Let $\mathbf{\Lambda}$ be one of these logics.
Consider an $m$-formula $A\not\in\Lambda$.
Take a generated submodel $M = (W, R, R_D, \varphi)$ of the $m$-restricted canonical model of $\Lambda$ such that $M,u{\,\not\mo\,} A$ for some $u$. As we know, its frame is basic and its $R$-maximal clusters are definable (Lemma \ref{mcl}).
Put
\begin{align*}
\Psi_0 &:= \setdef[\beta(C)]{C\ \hbox{is an $R$-maximal cluster in $M$}}, \\
\Psi_1 &:= \set{A} \cup \setdef[\overline{\square}\gamma]{\gamma\ \hbox{is a Boolean combination of formulas from $\Psi_0$}}, \\
\Psi &:= \hbox{the closure of $\Psi_1$ under subformulas.}
\end{align*}
The set $\Psi$ is obviously finite up to equivalence in $\mathbf{\Lambda}$.
Take the filtration $M' = (W', R', R'_D, \varphi')$ of $M$ through $\Psi$ as in Lemma \ref{LF1}. By that lemma, $F':=(W', R', R'_D)\vDash {\bf KT_1}$.
The seriality of $R'$ easily follows from the seriality of $R$.
Next, if $\mathbf{\Lambda}=\lgc{DT_1CK}$, the frame $(W, R, R_D)$ is connected by Lemma \ref{lem_KripkeCompletAC}. So for any $x,y \in W$ there is an $R$-path from $x$ to $y$. $aRb$ implies $a^\sim R'b^\sim$, so
there is an $R'$-path from $x^\sim$ to $y^\sim$ in $F'$. Therefore
$F'\vDash AC^{\sharp u}$.
It remains to show that $F'\vDash Ku$.
Consider an $R'$-irreflexive point $x^\sim \in W'$ and assume that
$R' (x^\sim )$ is disconnected.
Let $V$ be a nonempty proper connected component of $R'(x^\sim)$. Consider
\begin{align*}
\Delta &:= \setdef[C]{\exists y (y^\sim \in V \;\&\; C\subseteq R(y) \;\&\; C \hbox{ is an $R$-maximal cluster in $M$)}}; \\
B &:= \bigvee\limits_{C \in \Delta} \beta (C),
\end{align*}
where $\beta (C)$ is from Lemma \ref{mcl}.
Note that
\[
\leqno(1)\quad
z\in C~\&~ C\in{\Delta}\Rightarrow z^\sim \in V.
\]
In fact, if $C\in{\Delta}$, then for some $y^\sim \in V$ we have $yRz$; hence
$y^\sim R'z^\sim$, so $z^\sim \in V$, by the connectedness of $V$.
Let us show that for any $y^\sim \in R'(x^\sim)$
\[
\leqno(2)\quad
M', y^\sim \models B \ \ \hbox{iff}\ \ M, y \models B \ \ \hbox{iff}\ \ y^\sim \in V,
\]
i.e., $B$ defines $V$ in $ R'(x^\sim)$.
The first equivalence holds by the Filtration Lemma, since $B \in \Psi_1$.
Let us prove the second equivalence. To show `if', suppose $y^\sim\in V$.
By Lemma \ref{L82}, in the restricted canonical model there is a maximal cluster $C$ $R$-accessible from $y$; then $M, y \models \beta(C)$. We have $C \in \Delta$, and thus $M, y \models B$.
To show `only if', suppose $y^\sim \not\in V$, but $M, y \models B$. Then $M,y \models \beta (C)$, for some $C \in \Delta$,
hence $C \subseteq R(y)$, i.e., $yRz$ for some (and for all) $z\in C$; so it follows that $y^\sim R' z^\sim$.
Thus $y^\sim$ and $z^\sim$ are in the same connected component of $R'(x^\sim)$, which implies $z^\sim \not\in V$. However, $z^\sim \in V$ by (1), leading to a contradiction.
By Proposition \ref{cm} all substitution instances of $Ku$ are true in $M$. So
\[
M\vDash Ku(B):=\square (\overline{\square} B \lor \overline{\square}\lnot B) \to \square B \lor \Box \lnot B.
\]
Consider an arbitrary $y\in R(x)$. Then for any $z\in R(y)$, $y^\sim$ and $z^\sim$ are in the same connected component of $R'(x^\sim)$.
Thus $y^\sim$ and $z^\sim$ are both either in $V$ or not in $V$, and so by (2), both of them satisfy either $B$ or $\lnot B$.
Hence $M, y\models \overline{\square} B \lor \overline{\square}\lnot B$. Therefore, $x$ satisfies the premise of $Ku(B)$. Consequently, $x$ must satisfy the conclusion of $Ku(B)$. Thus $M,x \models \square B$ or $M,x \models \square \lnot B$.
Since $\square B, \square \lnot B \in\Psi_1$, the Filtration Lemma implies
$M', x^\sim\vDash \square B$ or $M', x^\sim\vDash \square\neg B$. Eventually by (2),
$V=R'(x^\sim)$ or $V={\varnothing}$, which
contradicts the assumption about $V$.
To conclude the proof, note that $A\in \Psi$, so by the Filtration Lemma $M',u^\sim\nvDash A$.
As we have proved, $F' \models \mathbf{\Lambda}$.
Therefore $\mathbf{\Lambda}$ has the fmp.
{\hspace*{\fill} }\end{proof}
\begin{theorem}\label{T101}
The logic ${\bf D4K}$ has the finite model property.
\end{theorem}
\begin{proof}
Use the argument from the proof of \ref{thm:fmp_DT_1CK} without the second relation.
{\hspace*{\fill} }\end{proof}
Thanks to the fmp, we have a convenient class of Kripke frames for the logic $\lgc{DT_1CK}$. This will allow us to prove the topological completeness result in the next section.
\section{The dd-logic of $\mathbf R^n$, $n\ge 2$.}
This section contains the main result of the Chapter. The proof is based on the fmp theorem from the previous section and a technical construction of a dd-morphism presented in the Appendix.
In this section $\norm{\cdot}$ denotes the standard norm in $\mathbf R^n$, i.e. for $x\in \mathbf R^n$
\[
\norm{x} = \sqrt{x_1^2+\ldots+x_n^2}.
\]
We begin with some simple observations on connectedness. For a path $\alpha = w_0 w_1 \ldots
w_{n}$ in a $\lgc{K4}$-frame $(W,R)$ we use the notation $ \overline{R}(\alpha) :=
\bigcup\limits_{i=0}^n \overline{R} (w_{i})$. A path $\alpha$ is called {\em global} (in $F$) if
$\overline{R} (\alpha)=W$.
\begin{lem}\label{globpath}
Let $F = (W, R)$ be a finite connected ${\bf K4}$-frame, $w,v\in W$. Then there exists a global path from $w$ to $v$.
\end{lem}
\begin{proof}
In fact, in the finite connected
graph $(W, R^\pm)$ the vertices $w,v$ can be connected by a path visiting all the vertices (perhaps,
several times). {\hspace*{\fill} }
\end{proof}
\begin{lem}\label{LPF0}
Let $F = (W, R, R_D)$ be a finite rooted $\lgc{DT_1CK}$-frame. Then the set of all $R_D$-reflexive points in $F$ is connected.
\end{lem}
\begin{proof}
Let $x,y$ be two $R_D$-reflexive points. Since $(W,R)$ is connected, there exists a path connecting
$x$ and $y$. Consider such a path $\alpha$ with the minimal number $n$ of $R_D$-irreflexive points, and
let us show that $n=0$.
Suppose not. Take an $R_D$-irreflexive point $z$ in $\alpha$; then
$\alpha=x\ldots uzv\ldots y$, for some $u,v$, and it is clear that $zRu$, $zRv$, since $z$ is strictly $R$-minimal. By Lemma \ref{L77}, $R(z)$ is connected, so $u$, $v$ can be connected by a path $\beta$ in $R(z)$. Thus in $\alpha$ we can replace the part $uzv$ with $\beta$, and
the combined path $x\ldots\beta\ldots y$ contains $(n-1)$ $R_D$-irreflexive points, which contradicts the minimality of $n$.
{\hspace*{\fill} }
\end{proof}
\begin{lem}\label{lem:pathinaFrame}
Let $F = (W, R, R_D)$ be a finite rooted $\lgc{DT_1CK}$-frame and let $w', w'' \in W$ be
$R_D$-reflexive. Then there is a global path $\alpha = w_0 \ldots w_n$ in $(W,R)$ such that $w' =
w_0$, $w_n = w''$ and all $R_D$-irreflexive points occur only once in $\alpha$.
\end{lem}
\begin{proof}
Let $\set{u_1, \,\ldots, \;u_k}$ be the $R_D$-irreflexive points.
By connectedness there exists paths $\alpha_0$, \ldots, $\alpha_{k}$
respectively from $w'$ to $u_1$, from $u_1$ to $u_2$, \ldots, from $u_k$ to $w''$.
By Lemma \ref{LPF0}, the set $W' := W - \set{u_1, \ldots, u_k}$ is connected.
Hence we may assume that each $\alpha_i$ does not contain $R_D$-irreflexive points except its ends.
Also there exists a loop $\beta$ in $F':=F|W'$ from $w''$ to $w''$ such that
\hbox{$W - \bigcup\limits_{i=1}^{k-1} \overline{R}(\alpha_i) \subseteq \overline{R}(\beta).$}
\begin{figure}[h]
\centering
\includegraphics[width=0.85\textwidth]{fig-4.pdf}
\caption{Path $\alpha$.}
\label{fig:PathAlpha}
\end{figure}
Then we can define $\alpha$ as the joined path $\alpha_0\ldots \alpha_k
\beta$, (Fig. \ref{fig:PathAlpha}). {\hspace*{\fill} }\end{proof}
\begin{propo}\label{pr:dd-morphismRntoF}
For a finite rooted $\lgc{DT_1CK}$-frame $F = (W, R, R_D)$ and $R$-reflexive points $w', w'' \in
W$, the following holds.
\begin{description}
\item[(a)] If $X=\set{ x\in {\bf R}^n \mid ||x|| \le r }$, $n \ge 2$,
then there exists $f: X \twoheadrightarrow^{dd} F$ such that $f(\partial X) = \set{w'}$;
\item[(b)] If $0\leq r_1<r_2$ and
\begin{align*}
X& = \setdef[x\in {\bf R}^n]{r_1\leq ||x||\leq r_2 }, \\
Y'& =\setdef[x\in {\bf R}^n]{||x||=r_1 },\ Y'' =\setdef[x\in {\bf R}^n]{||x||=r_2 },
\end{align*}
then there exists $f: X \twoheadrightarrow^{dd} F$ such that
$f(Y') = \set{w'}$, $f(Y'') = \set{w''}$.
\end{description}
\end{propo}
\begin{proof}
By induction on $\abs{W}$. Let us prove (a) first.
There are five cases:
\textbf{(a1)} $W = R(b)$ (and hence $b R b$) and $b = w'$.
Then there exists $f:X\twoheadrightarrow^d (W,R)$.
In fact, let $C$ be the cluster of $b$ (as a subframe of $(W,R)$). Then $(W,R)=C$ or $(W,R)=C\cup F_1\cup\ldots\cup F_l$, where the
$F_i$ are generated by the successors of $C$. If $(W,R)=C$, we apply Proposition \ref{P54};
otherwise we apply Lemma \ref{refroot} and IH.
By \ref{pr:DS_AT1} it follows that $R_D$ is universal. And so by \ref{lem:dmortodDmor}(3) $f$ is a dd-morphism.
\medskip
\textbf{(a2)} $W = R(b)$ and not $w'Rb$. We may assume that $r=3$. Put
\[
X_1 := \set{x \mid ||x|| \le 1},~
Y := \set{ x \mid 1\le ||x|| \le 2 },~
X_2 := \set{ x \mid 2 \le ||x|| \le 3 }.
\]
By the case (a1), there is $f_1: X_1 \twoheadrightarrow^{dd} F$ with $f_1(\partial X_1) = \set{b}$. Let $C$ be a
maximal cluster in $R(w')$. By \ref{P54} there is $g: {\bf I} Y \twoheadrightarrow^d C$. Since $R(w')\neq W$,
we can apply IH to the frame $F':=F^{w'}_{\forall}$ and construct a dd-morphism $f_2:
X_2 \twoheadrightarrow^{dd} F'$ with $f_2(\partial X_2) = \set{w'}$. Now since $f_i(\partial X_i)\subseteq R^{-1}(C)$, the
Glueing lemma \ref{Glue} is applicable. Thus $f:X\twoheadrightarrow^d F$ for $f:=f_1\cup f_2\cup g$ (See Fig. \ref{fig:dd-morphism_a23}, Case (a2)). Note that
$\partial X\subset
\partial X_2$, so $f(\partial X)=f_2(\partial X)=\{w'\}$.
As in the case (a1), $f$
is a dd-morphism by \ref{lem:dmortodDmor}.
\medskip
\textbf{(a3)} $(W,R)$ is not rooted. By Lemma \ref{lem:pathinaFrame} there is a global path $\alpha$ in $F$ with a single occurrence of every $R_D$-irreflexive point.
We may assume that $ \alpha =
b_0 c_0 b_1 c_1\ldots c_{m-1} b_m$, $b_m = w'$ and for any $i<m$ $c_i \in C_i \subseteq R(b_i) \cap R(b_{i+1})$, where $C_i$ is an $R$-maximal cluster.
Such a path is called \emph{reduced}.
For $0 \le j \le m$ we put $F_j := F|\overline{R}(b_j)$.
Since $(W,R)$ is not rooted, each $F_j$ is of smaller size than $F$, so we can apply the induction
hypothesis to $F_j$.
We may assume that
$$X = \set{ x \mid ||x||\leq 2m+1}, \ Y=\set{ x \mid ||x||=2m+1}.$$
Then put
\[
X_i := \set{ x \mid ||x|| \leq i+1 }\mbox{ for }0 \le i \le 2m,
\]
\[
Y_i := \partial X_i,~ {\Delta} _i := {\bf C} (X_{i} -X_{i-1})\mbox{ for }0 \le i \le 2m.
\]
By IH and Proposition \ref{P54} there exist
\begin{align*}
f_0:& X_0 \twoheadrightarrow^{dd} F_0 \hbox{ such that } f_0(Y_0) = \{c_0\},\\
f_{2j}:& {\Delta} _{2j} \twoheadrightarrow^{dd}F_{j} \hbox{ such that } f_{2j}(Y_{2j}) = \{c_j\}, \ f_{2j}(Y_{2j-1}) = \{c_{j-1}\}\ \mbox{ for }1 \le j \le m,\\
f_{2j-1}:& \Phi {\Delta}_{2j+1} \pmor^d{} C_j \mbox{ for }0 \le j \le m-1.
\end{align*}
\begin{figure
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.3\textwidth]{fig-2a.pdf}
&\qquad\qquad&
\includegraphics[width=0.5\textwidth]{fig-3a.pdf}\\
Case (a2)&& Case (a3)
\end{tabular}
\caption{dd-morphism $f$}
\label{fig:dd-morphism_a23}
\end{figure}
One can check that $f:X\twoheadrightarrow^{dd} F$ for $f:=\bigcup\limits_{j=0}^{2m}f_j$ (Fig. \ref{fig:dd-morphism_a23}).
\textbf{(a4)} $W = \overline{R}(b)$, $\lnot b R_D b$ (and so $\lnot b R b$).
We may assume that
$$X = \set{ x \mid ||x||\leq 2}, \ Y=\set{ x \mid ||x||=2}.$$
Then similar to case (a3) put
\[X_0 := X, ~Y_0 := Y, ~
X_i := \set{ x \mid ||x|| \leq \frac 1i },~ Y_i := \partial X_i,~ {\Delta} _i := {\bf C} (X_{i} -X_{i+1}), ~ (i>0).
\]
Consider the frame $F' := F|W'$, where $W' = W - \set{b}$. Note that $w'\in W'$, since $w'Rw'$, by
the assumption of \ref{pr:dd-morphismRntoF}. By Lemma \ref{L77} $F'$ is connected, and
thus $F'\vDash \lgc{DT_1CK}$. By Lemma \ref{lem:pathinaFrame} there is a reduced global path $\alpha = a_1
\ldots a_m$ in $F'$ such that $ a_1 = w'$. Let
\[
\gamma = a_1 a_2 \ldots a_{m-1} a_m a_{m-1} \ldots a_2 a_1 a_2 \ldots
\]
be an infinite path shuttling back and forth through $\alpha$. Rename the points in $\gamma$:
\begin{equation}
\gamma = b_0 c_0 b_1 c_1 \ldots b_m c_m b_{m+1} \ldots
\end{equation}
Again as in the case (a3) we put $F_j := F|\overline{R}(b_j)$, and assume that $c_j \in C_j$ and $C_j$ is an $R$-maximal cluster.
By IH there exist
\begin{align*}
f_0: & {\Delta} _0 \twoheadrightarrow^{dd} F_0 \hbox{ such that } f_0(Y_0) = \{b_0\} = \{w'\}, \ f_1(Y_1) =\{ c_0\},\\
f_{2j}: & {\Delta} _{2j} \twoheadrightarrow^{dd}F_{j} \hbox{ such that } f_{2j}(Y_{2j}) = \{c_{j-1}\}, \
f_{2j}(Y_{2j+1}) = \{c_{j}\}\hbox{ for }j>0,
\end{align*}
and by Proposition \ref{P54} there exist
$f_{2j+1}: \Phi {\Delta} _{2j+1} \pmor^d{} C_j$.
Put
\[
f(x) := \left\{
\begin{array}{ll}
b &\hbox{ if } x = {\bf 0},\\
f_{2j}(x)&\hbox{ if } x \in \Delta_{2j},\\
f_{2j+1}(x)&\hbox{ if } x \in \Phi\Delta_{2j+1},
\end{array}
\right.
\]
One can check that $f$ is d-morphic (Fig. \ref{fig:dd-morphism_a45}).
\medskip
\textbf{(a5)} $W = \overline{R}(b)$, $\lnot b R b$ and $b R_D b$. Then $R_D$ is universal,
$w'\ne b$. Put
\[
X' := \set{x \mid ||x|| < 1}, ~
X_4 := \set{ x \mid 1 \le ||x|| \le 2},
\]
and let $X_1,X_2$ be two disjoint closed balls in $X'$,
$X_3 := X'- X_1 - X_2$.
Let $C$ be a maximal cluster in $R(w')$,
$F' := F|R(w')$. Then there exist:
\begin{align*}
f_i:& X_i \twoheadrightarrow^d (W,R) \mbox{ for }i=1,2\mbox{ such that } \ f_i(\partial X_i) = \set{w'}, \hbox{ by the case (a4),}\\
f_3:& X_3 \twoheadrightarrow^d C, \hbox{ by Proposition \ref{P54},}\\
f_4:& X_4 \twoheadrightarrow^{dd} F' \mbox{ such that } \ f_4(\partial X_4) = \set{w'}, \hbox{ by the induction hypothesis.}
\end{align*}
\begin{figure
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.4\textwidth]{fig-4a.pdf}
&\qquad\qquad&
\includegraphics[width=0.4\textwidth]{fig-5a.pdf}\\
Case (a4)&& Case (a5)
\end{tabular}
\caption{dd-morphism $f$}
\label{fig:dd-morphism_a45}
\end{figure}
Put $f:=f_1\cup f_2\cup f_3\cup f_4$ (Fig. \ref{fig:dd-morphism_a45}). Then $f(\partial \mathfrak X) = \set{w'}$.
By Lemma \ref{Glue} (b) $f_1\cup f_2:X_1\cup X_2\twoheadrightarrow^d F$, and hence $f:X\twoheadrightarrow^d F$ by
Lemma \ref{Glue} (a).
$f$ is manifold at $b$, thus it is a dd-morphism by
\ref{lem_pmorphism}.
\medskip
Now we prove (b). There are three cases.
\textbf{(b1)} $w' = w'' = b$ and $W = R(b)$.
The argument is the same as in the case (a1), using
Proposition \ref{P54},
Lemma \ref{refroot}, the induction hypothesis, and
Proposition \ref{lem:dmortodDmor}.
\medskip
\textbf{(b2)} $w' = w'' = b$, but $W \neq R(b)$. Consider a maximal cluster $C \subseteq R (b)$.
Since all spherical shells for different $r_1$ and $r_2$ are homeomorphic, we assume that $r_1 = 1$, $r_2 = 4$. Consider the sets
\begin{equation*}
X_1 := \set{x \mid 1 \le ||x|| \le 2},\ \
X' := \set{x \mid 2 < ||x|| < 3},\ \
X_3 := \set{x \mid 3 \le ||x|| \le 4},
\end{equation*}
and let $X_0\subset X'$ be a closed ball, $X_2 := X' - X_0$.
Let $F' :=F| R (b)$. There exist
\begin{align*}
f_1:& X_1 \twoheadrightarrow^{dd} F'\mbox{ such that } f_1(\partial X_1) = \set{b}, \hbox{ by the case (b1)},\\
f_2:& X_2 \twoheadrightarrow^d C, \hbox{ by Proposition \ref{P54}},\\
f_3:& X_3 \twoheadrightarrow^{dd} F'\mbox{ such that } \ f_3(\partial X_3) = \set{b}, \hbox{ by the case (b1)},\\
f_0:& X_0 \twoheadrightarrow^{dd} F\mbox{ such that } \ f_4(\partial X_0) = \set{b},
\hbox{ by the statement (a) for }F.
\end{align*}
One can check that $f:X\twoheadrightarrow^{dd} F$ for $f:=f_0\cup f_1\cup f_2\cup f_3$.
\textbf{(b3)} $w' \ne w''$ and for some $b\in W$, $W = R(b)$, so $F$ has an $R$-reflexive root. Let
\[
F_1 := F|R(w'), ~ F_2 := F|R(w''),
\]
and let $C_i$ be an $R$-maximal cluster in $F_i$ for $i\in \set{1,2}$.
We assume that $r_1 = 1$, $r_2 = 6$ and consider the sets
\begin{align*}
X_i &:= \set{x \mid i \le ||x|| \le i+1}, \ i \in \set{1, \ldots, 5}.
\end{align*}
By the case (b1) and Proposition \ref{P54} we have
\begin{align*}
f_1:&~ X_1 \twoheadrightarrow^{dd} F_1 \mbox{ such that } f_1(\partial X_1) = \set{w'}, &f_2:&~ \Phi X_2 \twoheadrightarrow^d C_1,\\
f_3:&~ X_3 \twoheadrightarrow^{dd} F\mbox{ such that } \ f_3(\partial X_3) = \set{b}, &f_4:&~ \Phi X_4 \twoheadrightarrow^d C_2, \\
f_5:&~ X_5 \twoheadrightarrow^{dd} F_2 \mbox{ such that } f_1(\partial X_5) = \set{w''}.
\end{align*}
\begin{figure
\centering
\begin{tabular}{c|c}
\includegraphics[width=0.45\textwidth]{fig-2b.pdf}
\qquad&\qquad
\includegraphics[width=0.3\textwidth]{fig-3b.pdf}\\
Case (b2)& Case (b3)
\end{tabular}
\caption{dd-morphism $f$}
\label{fig:dd-morphism_b23}
\end{figure}
One can check that $f:X\twoheadrightarrow^{dd} F$ for $f:= \bigcup\limits_{i=1}^5 f_i$ (Fig. \ref{fig:dd-morphism_b23}, Case (b3)).
\medskip
\textbf{(b4)} $w' \ne w''$ and $W \ne R(b)$ for any $b\in W$. By Lemma \ref{LPF0} there is a reduced path
$\alpha = b_0 c_0 b_1 \ldots c_{m-1} b_m$ from $b_0 = w'$ to $b_m = w''$ that does not contain $R_D$-irreflexive
points, $c_i \in C_i$, where $C_i$ is an $R$-maximal cluster. We may also assume that
\begin{equation}\label{eq:Rbi}
\overline{R}(b_i) \ne W, \hbox{ for any $i\in \set{1, \ldots, m-1}$}.
\end{equation}
In fact, if the frame $(W,R)$ is not rooted, then (\ref{eq:Rbi}) obviously holds. If $(W,R)$
is rooted, then its root $r$ is irreflexive and by Lemma \ref{L77}, $R(r)$ is connected, so
there exists a path $\alpha$ in $R(r)$ satisfying (\ref{eq:Rbi}).
Put
\[
F_0 := F,~
F_j := F|R(b_j), 1 \le j \le m.
\]
Assuming that $r_1 = 1$, $r_2 = 2m+1$ we define
\begin{align*}
X_i &:= \setdef{||x|| \leq i +1 },~ Y_i := \partial X_i~ (\mbox{for }0 \le i \le 2m+1),\\
{\Delta} _i &:= {\bf C} (X_{i+1} -X_{i}) \ (\mbox{for }0\le i\le 2m).
\end{align*}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{fig-4b.pdf}
\caption{dd-morphism $f$, case (b4)}
\label{fig:dd-morphism_b4}
\end{figure}
By the cases (b2), (b1), Proposition \ref{P54}, and the induction hypothesis there exist
\begin{align*}
&f_0: {\Delta}_0 \twoheadrightarrow^{dd} F=F_0 \hbox{ such that } f_0(Y_0) = f_0(Y_1) = \set{w'}; \\
&f_{2j}: {\Delta} _{2j} \twoheadrightarrow^{dd} F_{j} \hbox{ such that } f_{2j}(Y_{2j+1}) = \set{c_{j}}, \ f_{2j}(Y_{2j}) = \{c_{j-1}\}\ (1 \le j \le m);\\
&f_{2j-1}: \Phi {\Delta} _{2j-1} \twoheadrightarrow^d C_{j-1} \ (1 \le j \le m),\\
&f_{2m}: {\Delta} _{2m} \twoheadrightarrow^{dd} F_m \hbox{ such that } f_{2m}(Y_{2m}) = \set{c_{m}}, \ f_{2m}(Y_{2m+1}) =
\set{w''}.
\end{align*}
We claim that $f:X\twoheadrightarrow^{dd} F$ for $f:= \bigcup\limits_{i=0}^{2m} f_i$ (Fig. \ref{fig:dd-morphism_b4}).
First, we prove by induction using Lemma \ref{Glue} (see previous cases) that $f$ is a d-morphism.
Note that $f(Y')= f(Y_0) = \set{w'}$ and $f(Y'')= f(Y_{2m+1}) = \set{w''}$.
Second, there are no $R_D$-irreflexive points in $\alpha$, so all preimages of $R_D$-irreflexive points
are in $\Delta_0$; since $f_0$ is a dd-morphism, $f$ is 1-fold at any $R_D$-irreflexive point and
manifold at all the others. Thus $f$ is a dd-morphism by Proposition \ref{lem:dmortodDmor}.
\end{proof}
\begin{theorem}\label{T125}
For $n \ge 2$, the dd-logic of ${\bf R}^n$ is $\lgc{DT_1CK}$.
\end{theorem}
\begin{proof}
Since ${\bf R}^n$ is a locally 1-component connected dense-in-itself metric space, ${\bf R}^n \models^d
\lgc{DT_1CK}$.
Now consider a formula $A \notin \lgc{DT_1CK}$. Due to the fmp (Theorem \ref{thm:fmp_DT_1CK}) there
exists a finite rooted Kripke frame $F = (W, R, R_D) \vDash \lgc{DT_1CK}$ such that $F \nvDash A$. By
Proposition $\ref{pr:dd-morphismRntoF}$ there exists $f:{\bf R}^n \twoheadrightarrow^{dd} F$. Hence ${\bf R}^n
\nvDash^d A$ by Lemma \ref{lem_pmorphism}. {\hspace*{\fill} }\end{proof}
\section{Concluding remarks}
\textbf{Hybrid logics.}
Logics with the difference modality are closely related to hybrid logics. The paper \cite{Litak06} describes a validity-preserving translation from the language with the topological and the difference modalities into the hybrid language with the topological modality, nominals and the universal modality.
Apparently a similar translation exists for dd-logics considered in our chapter.
There may be an additional option --- to use `local nominals', propositional constants that may be true not in a single point, but in a discrete set.
Perhaps one can also consider `one-dimensional nominals' naming `lines' or `curves' in the main topological space; there may be many other similar options.
\smallskip
\textbf{Definability.}
Among several types of topological modal logics considered in this chapter dd-logics are the most
expressive. The correlation between all the types are shown in Fig. \ref{fig:languages}.
A language ${\cal L}_1$ is \emph{reducible} to ${\cal L}_2$ (${\cal L}_1 \le {\cal L}_2$) if every
${\cal L}_1$-definable class of spaces is
${\cal L}_2$-definable; ${\cal L}_1<{\cal L}_2$ if ${\cal L}_1\le {\cal L}_2$ and ${\cal L}_2 \nleq {\cal L}_1$. The non-strict reductions 1--7 in Fig. \ref{fig:languages}
are rather obvious. Let us explain, why 1--6 are strict.
\begin{figure}[h]
\centering
\includegraphics[width=4.1cm]{fig-6.pdf}
\caption{Correlation between topomodal languages.}
\label{fig:languages}
\end{figure}
The relations 1 and 2 are strict, since the c-logics of ${\bf R}$ and ${\bf Q}$ coincide \cite{MT}, while the cu- and d-logics are different \cite{Sh99, E1}.
The relation 3 is strict, since in d-logic without the universal modality we cannot express connectedness (this follows from \cite{E1}). The relations 4 and 6 are strict, since the cu-logics of ${\bf R}$ and ${\bf R}^2$ are the same \cite{Sh99}, while the cd- and du-logics are different \cite{Gab01, LB2}.
In cd- and dd-logic we can express \emph{global 1-componency}: the formula
\[
\DA (\overline{\square} p \lor \overline{\square} \lnot p) \to \DA p \lor \DA \lnot p
\]
is c-valid in a space $\mathfrak X$ iff the complement of any point in $\mathfrak X$ is connected. So we can distinguish the line $\mathbf R$ and the circle ${\bf S^1}$. In du- (and cu-) logic
this is impossible, since there is a local homemorphism $f(t) = e^{it}$ from $\mathbf R$ onto
$S^1$. It follows that the relation 5 is strict.
Our conjecture is that the relation 7 is strict as well.
\smallskip
\textbf{Axiomatization.}
There are several open questions about axiomatization and completeness of certain dd-logics.
1. The first group of questions is about the logic of ${\bf R}$. On the one hand, in \cite{Kudinov08} it
was proved that $\mathbf{Lc}_{\ne}({\bf R})$ is not finitely axiomatizable. Probably, the same method
can be applied to $\mathbf{Ld}_{\ne}({\bf R})$. On the other hand, $\mathbf{Lc}_{\ne}({\bf R})$ has the fmp
\cite{Kud11}, and we hope that the same holds for the dd-logic. The decidability of
$\mathbf{Ld}_{\ne}({\bf R})$ follows from \cite{BG}, since this logic is a fragment of the universal
monadic theory of ${\bf R}$; and by a result from \cite{Rey} it is PSPACE-complete. However, constructing
an explicit infinite axiomatization of $\mathbf{Lc}_{\ne}({\bf R})$ or $\mathbf{Ld}_{\ne}({\bf R})$ might be a serious technical problem.
2. A `natural' semantical characterization of the logic $\lgc{DT_1C}+Ku_2$ (which is a proper sublogic of $\mathbf{Ld}_{\ne}({\bf R})$) is not quite clear. Our conjecture is that it is complete w.r.t. 2-dimensional cell complexes, or more exactly, adjunction spaces obtained from finite sets of 2-dimensional discs and 1-dimensional segments.
3. We do not know any syntactic description of dd-logics of 1-dimensional cell complexes (i.e., unions of finitely many segments in ${\bf R}^3$ that may have only endpoints as common).
Their properties are probably similar to those of $\mathbf{Ld}_{\ne}({\bf R})$.
4. It may be interesting to study topological modal logics with the graded difference modalities $\DA_n A$ with the following semantics: $x\models \DA_n A$ iff there are at least $n$ points $y\ne x$ such that $y \models A$.
5. The papers \cite{MT} and \cite{Kremer} prove completeness and strong completeness of ${\bf S4}$ w.r.t.~any dense-in-itself metric space. The corresponding result for d-logics is completeness of $\mathbf{D4}$ w.r.t. an arbitrary dense-in-itself separable metric space. Is separabilty essential here? Does strong completeness hold in this case? Similar questions make sense for dd-logics.
6. \cite{Gab01} presents a 2-modal formula cd-valid exactly in $T_0$-spaces. However, the cd-logic (and the dd-logic) of the class of $T_0$-spaces is still unknown. Note that the d-logic of this class has been axiomatized in \cite{BEG2011}; probably the same technique is applicable to cd- and dd-logics.
7. In footnote 7 we have mentioned that there is a gap in the paper \cite{Sh99}. Still we can prove that for any connected, locally connected metric space $\mathfrak X$ such that the boundary of any ball is nowhere dense, ${{\bf L}}{\bf c}_\forall(\mathfrak X)=\mathbf{S4U} + AC$. But for an arbitrary connected metric space $\mathfrak X$ we do not even know if ${{\bf L}}{\bf c}_\forall(\mathfrak X)$ is finitely axiomatizable.
8. Is it possible to characterize finitely axiomatizable dd-logics that are complete w.r.t.~Hausdorff spaces? metric spaces? Does there exist a dd-logic complete w.r.t. Hausdorff spaces, but incomplete w.r.t. metric spaces?
9. Suppose we have a c-complete modal logic $L$, and let ${\cal K}$ be the class of all topological spaces where $L$ is valid. Is it always true that ${\bf Lc}_\forall({\cal K})=LU$? and ${\bf Lc}_{\ne}({\cal K})=LD$?
Similar questions can be formulated for d-complete modal logics and their du- and dd-extensions.
10. An interesting topic not addressed in this chapter is the complexity of topomodal logics. In particular, the complexity is unknown for
the d-logic (and the dd-logic) of ${\bf R}^n$ ($n>1$).
We would like to thank anonymous referee who helped us improve the first version of
the manuscript.
The work on this chapter was supported by RFBR grants 11-01-00281-a,
11-01-00958-a, 11-01-93107-CNRS-a and the Russian President's grant
NSh-5593.2012.1.
|
\section{Introduction and main results}
\label{sec:intro}
We study spatially periodic solutions of the Degasperis-Procesi
equation
\begin{equation}
\label{DP0}
\begin{cases}
u_t-u_{txx}+4uu_x=3u_xu_{xx}+uu_{xxx}, &t>0,\quad x\in\R\\
u(t,x)=u(t,x+1), &t\ge0,\quad x\in\R.
\end{cases}
\end{equation}
Such equation attracted a considerable interest in the past few years, both for its remarkable mathematical
properties (see, {\it e.g.\/} \cites{DP99, CIL2010}), and for its physical interpretation as an asymptotic
model obtained from the water-wave system in shallow water regime. In this setting, the equation models
moderate amplitude waves and $u$ stands for a horizontal velocity of the water at a fixed depth,
see \cites{AConL09, John02} and the references therein for further physical motivations.
The Cauchy problem associated with~\eqref{DP0} can be more conveniently reformulated as
\begin{equation}
\label{DP}
\begin{cases}
u_t+uu_x+\partial_x p*\biggl(\displaystyle\frac32 u^2\biggr)=0, &t>0,\quad x\in\S\\
u(0,x)=u_0(x), &x\in\S\\
\end{cases}
\end{equation}
where $\S$ is the circle and $p$ the kernel of $(1-\partial_x^2)^{-1}$, given by
the continuous $1$-periodic function
\begin{equation}
\label{kernel}
p(x)=\frac{\cosh(x-[x]-1/2)}{2\sinh(1/2)}.
\end{equation}
Another possible reformulation of~\eqref{DP0} is the momentum-velocity equation,
\begin{equation}
\label{eq:pot}
y_t+uy_x+3u_xy=0, \qquad y(t,x)=y(t,x+1),\qquad t>0,\;x\in\R,
\end{equation}
where $y=u-u_{xx}$ is the associated potential, and $u$ can be recovered from~$y$ from the convolution relation
$u=p*y$.
It is well known (see, {\it e.g.\/}, \cite{Yin03}) that if $u_0\in H^s(\S)$, with $s>3/2$,
then the problem~\eqref{DP} possess a unique solution
\begin{equation}
\label{funspa}
u\in C([0,T),H^s(\S))\cap C^1([0,T),H^{s-1}(\S)),
\end{equation}
for some $T>0$, depending only $u_0$.
The maximal time $T^*$ of the above solution can be finite or infinite.
For instance, if the initial potential $y_0=y(0,\cdot)$ does not change sign, then it is known that $T^*=+\infty$,
see \cite{ELY07}.
On the other hand, several different blowup criteria were established, {\it e.g.\/}, in~\cites{Yin03, ELY07,GGL-2011}:
in the shallow water interpretation the finite time blowup corresponds to a wave breaking mechanism, as near the blowup time
solutions remain bounded, but have an unbounded slope in at least one point.
The purpose of this short paper it to establish the following Liouville-type theorem:
\begin{theorem}
\label{th:liou}
The only global solution $u\in C([0,+\infty),H^s(\S))\cap C^1([0,+\infty),H^{s-1}(\S))$, with $s>3/2$,
to the Degasperis-Procesi equation vanishing at some point $(t_0,x_0)$ is the identically zero solution.
\end{theorem}
The Degasperis-Procesi equation is often written with the additional dispersive term $3\kappa u_x$ in the left-hand side of equation~\eqref{DP0}, where $\kappa\in\R$ is the dispersion parameter.
In this more general setting the above theorem can be reformulated as follows:
\begin{theorem}
\label{th:k}
Let $s>3/2$. If $v\in C([0,+\infty),H^s(\S))\cap C^1([0,+\infty),H^{s-1}(\S))$ is a global solution
to the Degasperis-Procesi equation with dispersion
\begin{equation}
\label{eq:v}
v_t+vv_x+\partial_x p*\biggl(\displaystyle\frac32 v^2+3\kappa v\biggr)=0, \qquad t>0,\quad x\in\S,
\end{equation}
such that $v(t_0,x_0)=-\kappa$ at some point $(t_0,x_0)$, then $v(t,x)\equiv -\kappa$ for all $(t,x)$.
\end{theorem}
In the next section we will compare these theorems with earlier related results.
The main idea of the present paper will be remark that, in the dispersionless case,
for all time $t\in\R^+$, at least one of the two functions
$x\mapsto e^{\pm \sqrt{3/2}\,q(t,x)} u(t,q(t,x))$, where $q(t,x)$ is the flow of the global solution~$u$,
must be monotonically increasing.
\medskip
As a byproduct of our approach,
we will get a simple and natural blowup criterion for periodic solutions to the Degasperis-Procesi equation,
with or without dispersion, that is of independent interest:
\begin{proposition}
\label{prop:blow}
Let $v_0\in H^s(\S)$, with $s>3/2$, be such that $v_0'(a)<-\textstyle\sqrt{\frac32} \,\bigl|v_0(a)+\kappa\bigr|$
for some $a\in\S$. Then the solution $v\in C([0,T^*),H^s(\S))\cap C^1([0,T^*),H^s(\S))$ of~\eqref{eq:v} arising from~$v_0$
blows up in finite time.
\end{proposition}
\section{Comparison with some earlier results}
\label{sec:comparison}
In \cite{ELY07}*{Theorem 3.8}, Escher, Liu and Yin established the blowup for equation~\eqref{DP0}
assuming that $u_0\in H^s(\S)$, $u_0\not\equiv0$,
and that the corresponding solution $u(t,x)$ \emph{vanishes in at least one point $x_t\in\S$ for all $t\in[0,T^*)$}.
Theorem~\ref{th:liou} improves their result (and the corresponding corollaries)
by providing the same conclusion $T^*<\infty$ with a shorter proof, and
under an assumption that is easier to check.
Applying Proposition~\ref{prop:blow} with $\kappa=0$ and $a=0$
improves Yin's blowup criterion~\cite{Yin03}*{Theorem 3.2}, establishing the blowup for \emph{odd} initial data
with negative derivative at the origin.
In the particular case $\kappa=0$, Proposition~\ref{prop:blow} improves and simplifies
the wave-breaking criterion of \cite{ELY07}*{Theorem~4.3} (and its corollaries), that established the blowup under a condition of the form
$v_0'(a)<-\bigl(c_0\|v_0\|_{L^\infty}+c_1\|v_0\|_{L^2}\bigr)$, with suitable $c_0,c_1>0$.
In fact, Proposition~\ref{prop:blow} shows that one can take $c_1=0$, and more importantly, one only needs to check the behavior of $u_0$ in a neighborhood of a single point to get the blowup condition.
For general $\kappa\in\R$, Proposition~\ref{prop:blow} extends and considerably simplifies the blowup condition
$v_0'(a)<-M$ established in~\cite{GGL-2011}*{Theorem~4.1},
where $M=M(\kappa,\|v_0\|_{L^2},\|v_0\|_{L^\infty})$ was given by a quite involved expression.
In the same way, the pointwise estimates~\eqref{pointwise} below allow to improve results like
\cite{GGL-2011}*{Theorem~4.2} and its corollaries.
A Liouville-type theorem in the same spirit as Theorem~\ref{th:liou} has been established for periodic solutions
of the hyperelastic rod equation in~\cite{BraCor14}, when the physical parameter~$\gamma$ of the model belongs to a
suitable range (including $\gamma=1$, that corresponds to the dispersionless Camassa--Holm equation).
The specific structure of the nonlocal term
of the Degasperis-Procesi equation makes possible the much more concise proof presented here.
Neverthless, the Degasperis--Procesi equation remains worse understood than Camassa--Holm's.
First of all, the geometric picture between the two equations is different:
the Camassa--Holm equation can be realized as a metric Euler equation. On the other hand, there is no Riemannian metric on $\text{Diff}^\infty(\S)$ such that the corresponding geodesic flow is given by the Degasperis--Procesi equation.
In fact, the more subtle geometric picture for the latter equation has been disclosed only recently, see \cite{KolEsch11}.
Moreover, no \emph{necessary and sufficient\/} condition for the global existence of solutions to the Degasperis--Procesi equation is available. (Such a condition is instead known for the Camassa--Holm equation, see \cite{McKean04}).
For this reason, Proposition~\ref{prop:blow} provides valuable information.
\section{Proofs}
\begin{proof}[Proof of Theorem~\ref{th:liou}]
Equation~\eqref{DP0} is invariant under time translations and under the transformation
$\tilde u(t,x)=-u(-t,x)$. Therefore, it is enough to prove that if $u_0(x_0)=0$ at some point $x_0\in\S$, but
$u_0\not\equiv0$, then the solution~$u\in C([0,T),H^s(\S))\cap C^1([0,T),H^{s-1}(\S))$
arising from~$u_0$ must blow up in finite time.
Let $\alpha\in(x_0,x_0+1)$ be such that $u_0(\alpha)\not=0$.
We first consider the case $u_0(\alpha)>0$.
Let us introduce the map
\[\phi(x)=e^{\sqrt{\textstyle\frac32} x}u_0(x).\]
By the periodicity and the continuity of~$u_0$, we can find an open interval $(\alpha,\beta)\subset (x_0,x_0+1)$
such that $\phi(x)>0$ on the interval $(\alpha,\beta)$ and $\phi(\alpha)>0$, $\phi(\beta)=0$.
An integration by parts gives
\begin{equation}
\label{iip}
\int_\alpha^{\beta} e^{\sqrt{\textstyle\frac32}x}u_0'(x)\,{\rm d} x
=\phi(\beta)-\phi(\alpha)-\int_\alpha^{\beta}\sqrt{\textstyle\frac32}\phi(x)\,{\rm d} x.
\end{equation}
We deduce from this the existence of~$a\in(\alpha,\beta)$ such that
$u_0'(a)<-\sqrt{\textstyle\frac32}\,u_0(a)<0$.
Indeed, otherwise, we could bound the left-hand side in~\eqref{iip} from below by
$\displaystyle\int_\alpha^\beta -\sqrt{\textstyle\frac32}\phi(x)\,{\rm d} x$,
and get the contradiction $\phi(\alpha)\le \phi(\beta)$.
The second case to consider is $u_0(\alpha)<0$:
introducing now the map $\psi(x)=e^{-\sqrt{\textstyle\frac32} x}u_0(x)$ and arguing as before,
we get in this case the existence of a point $a$ such that
$u_0'(a)<\sqrt{\textstyle\frac32} \,u_0(a)<0$.
Notice that in both cases we get
\begin{equation}
\label{blow-cond}
\exists \,a\in\S \quad\mbox{such that}\quad
u_0'(a)<-\sqrt{\textstyle\frac32}\,\bigl|u_0(a)\bigr|.
\end{equation}
We thus reduced the proof of our claim to establishing the finite time blowup under condition~\eqref{blow-cond}, with
$u_0\in H^s(\R)$. In fact, by approximating $u_0$ with a sequence $(u_n)\subset H^3(\S)$, we can assume without loss of generality that $u_0\in H^3(\S)$. (Indeed, the argument below will provide un upper bound for $T^*$ independent on
the parameter~$n$).
\medskip
Let us introduce the flow map
\begin{equation}
\label{def:flow}
\begin{cases}
q_t(t,x)=u(t,q(t,x)), &t\in (0,T^*),\quad x\in\R\\
q(0,x)=x, &\qquad x\in\R.
\end{cases}
\end{equation}
We also introduce the $C^1$ functions, defined on $(0,T^*)$,
\[
f(t)=(-u_x+\sqrt{\textstyle\frac32} u)\bigl(t,q(t,a)\bigr),
\]
and
\[
g(t)=-(u_x+\sqrt{\textstyle\frac32} u)\bigl(t,q(t,a)\bigr).
\]
Taking the spatial derivative in equation~\eqref{DP},
recalling that $(1-\partial^2_x)p$ is the Dirac mass,
we get
\[
u_{tx}+uu_{xx}=-u_x^2+\frac32 u^2 -p*\biggl(\frac32 u^2\biggr).
\]
Using the definition of the flow map~\eqref{def:flow} we obtain
\begin{equation*}
\begin{split}
f'(t)
&=\biggl[ -(u_{tx}+uu_{xx})+\sqrt{\textstyle\frac32}(u_t+uu_x)\biggr](t,q(t,a))\\
&=\biggl[u_x^2-\textstyle\frac32 u^2 +(p-\sqrt{\textstyle\frac32}p_x)*(\textstyle\frac32 u^2)\biggr](t,q(t,a)).\\
\end{split}
\end{equation*}
From expression~\eqref{kernel} we easily get
\[
(p\pm\beta p_x)\ge0
\qquad
\mbox{if and only if}
\qquad |\beta|\le {\coth{(1/2)}},
\]
and so, in particular
\begin{equation}
\label{fail}
p\pm\sqrt{\textstyle\frac32}p_x\ge0.
\end{equation}
Hence we get
\begin{equation*}
\begin{split}
f'(t)\ge \bigl[ u_x^2 -\textstyle\frac32 u^2 \bigr](t,q(t,a)).
\end{split}
\end{equation*}
Factorizing the right-hand side leads to the differential inequality
\begin{equation}
\label{di1}
f'(t)\ge f(t)g(t), \qquad t\in(0,T^*).
\end{equation}
A similar computation yields
\begin{equation}
\label{di2}
g'(t)\ge f(t)g(t), \qquad t\in(0,T^*).
\end{equation}
Let
\[h(t)=\sqrt{f(t)g(t)}.\]
We first observe that
\[
h(0)=\sqrt{fg(0)}=\sqrt{u_0'(a)^2-\textstyle\frac32 u_0(a)^2}\,>0.
\]
Moreover, we deduce from the system~\eqref{di1}-\eqref{di2},
applying the geometric-arithmetic mean inequality, that
\[
h'(t)\ge h^2(t), \qquad t\in(0,T^*).
\]
This immediately implies
$T^*\le 1/h(0)<\infty$.
Theorem~\ref{th:liou} is completely established.
\end{proof}
\medskip
The above proof establishes also Proposition~\ref{prop:blow} in the particular case $\kappa=0$.
But $u(t,x)=v(t,x-\kappa t)+\kappa$,
is a global solution of~\eqref{DP} if and only if $v$ is a global solution of~\eqref{eq:v} with $u_0=v_0+\kappa$.
Hence, Theorem~\ref{th:k} follows immediately from Theorem~\ref{th:liou}. In the same way, we see that the claim of
Proposition~\ref{prop:blow} holds true for general~$\kappa$.
\medskip
Our arguments also reveal that global solutions must satisfy quite stringent pointwise estimates.
Indeed, assume that $u\in C([0,\infty),H^s(\S))\cap C^1([0,\infty),H^{s-1}(\S))$ is a given global solution of~\eqref{DP0}.
Then, by our theorem, $\mbox{sign(u)}=1,0$ or $-1$ is well defined and independent on $(t,x)$.
Moreover, $u'(t,x)\ge -\sqrt{\frac32} |u(t,x)|$ for all $t\ge0$ and $x\in\S$.
Then, arguing as in~\eqref{iip}, we deduce that, for all $t\ge0$, the map
$x\mapsto e^{\mbox{\scriptsize sign}(u)\sqrt{\frac32} x} \,u(t,x)$ is increasing.
Combining this with the periodicity, we get the pointwise estimates for $u(t,x)$,
for all $t\ge0$, all $\alpha\in\R$ and $\alpha\le x\le \alpha+1$:
\begin{equation}
\label{pointwise}
e^{\mbox{\scriptsize sign}(u)\sqrt{\frac32}(\alpha-x)}{u(t,\alpha)} \le u(t,x)
\le e^{\mbox{\scriptsize sign}(u)\sqrt{\frac32}(\alpha+1-x)}{u(t,\alpha)}.
\end{equation}
From~\eqref{pointwise} one immediately deduces the corresponding estimates for global
solutions to the Degasperis-Procesi
equation with dispersion.
\medskip
We conclude observing that our results
seem to be specific to periodic solutions and have no analogue in the case of solutions in~$ H^s(\R)$.
A reason for this is that in the non-periodic case the expression of $p$ should be
modified into $p(x)=\frac12 e^{-|x|}$ and the fundamental inequality~\eqref{fail} is no longer true.
|
\section{Introduction}
The advent of large scale cluster survey programs \citep[][]{PLANCK,SPT,XXL,ACT} will soon provide large catalogs of galaxy clusters to test the standard cosmological scenario based on the Cold Dark Matter paradigm with Cosmological Constant ($\Lambda$CDM).
Galaxy clusters are the largest observable structures in the universe which reside inside massive Dark Matter (DM) halos. These are gravitationally bound objects that result from the gravitational collapse of small matter density fluctuations which were present in the early universe. Because of this the number density of DM halos carries complementary information on both the statistics of the primordial matter density field and the growth rate of structures. Measurements of the abundance of cluster of galaxies aim to probe such features, but their effectiveness to constrain cosmological models depends upon the precise understanding of the survey selection function \citep{P2006,XXL,C2012} as well as the availability of accurate measurements of cluster masses \citep[see e.g.][]{MM2004,LH2004,C2009}. However, measuring the mass of several hundreds of clusters is a challenging task that requires costly follow-up observations of individual objects, thus inevitably spanning these projects over long periods of time. While completing these programs, cosmological information can still be inferred from reduced datasets consisting of the most massive objects. These probe the high-mass end of the halo mass function and thus have the potential to rule out entire classes of cosmological models. The advantage is that such systems, being also the most luminous, are easier to detect and being limited in number makes their follow-up observations more readily accessible.
In recent years this complementary approach to cluster cosmology has received lots of attention due to the discovery of very massive clusters at high-redshift \citep[][]{J2009,R2009,F2011,Menanteau2012a,S2012}. These detections have lead several authors to question the basic assumptions of the concordance $\Lambda$CDM model \citep[][]{JimenezVerde2009,Hoyle2011,HP2012}. However, estimating the probability that such extreme clusters occur is far from being trivial. As pointed out by \cite{Hotchkiss2011} assessing the rareness of clusters in terms of the probability of observing at least one cluster of mass larger than that observed and/or at a higher redshift can lead to biased conclusions.
A natural framework to address these questions is given by the Extreme Value Statistics (EVS). This can be used to predict the probability distribution of the mass of the most massive halo in the sample from prior knowledge of the halo mass function in a given cosmological model. This has been the subject of several studies \citep[see e.g.][]{Davis2011,Waizmann2011,HarrisonColes2011} which have focused on the mass as measure of cluster extremeness. However, a careful analysis of the most massive systems so far observed suggests that there are other characteristics that are indicative of the extremeness of these objects. For instance, 1E0657-56 \citep{Markevitch2002a,Markevitch2006a}, MACS J0025.4-1222 \citep{Bradac2008a}, ACT-CL \mbox{J0102-4915} \citep{Menanteau2012a} and AS1063 \citep{Gomez2012a} are merging clusters with high relative velocities. Among these 1E0657-56 is one of the most well documented. Known as the \enquote{Bullet Cluster}, it is composed of two clusters which have undergone a nearly head on collision. The main cluster has a mass $\simeq 10^{15}$ h$^{-1}$ M$_\odot$, while the smaller one has mass $\simeq 10^{14}$ h$^{-1}$ M$_\odot$ \citep{Clowe2004a}. The system is located at $z=0.296$ \citep{Clowe2006,Bradac2006} and the clusters are separated by a distance of \mbox{$\simeq 0.51$ h$^{-1}$ Mpc}. X-ray observations have shown that the collision has ripped off the clusters of their intra-cluster gas which is trapped in a shock with Mach number close to $\sim 3$. This implies a velocity of the shock front of \mbox{$\sim 4700$ km s$^{-1}$} \citep{Markevitch2006a}, which has been usually interpreted as being also the relative velocity of the colliding clusters. Under this hypothesis \citet{Hayashi2006a} analysed the velocities of the sub-halo distribution from the Millenium Simulation \citep{Springel2005a} and concluded that the existence of the Bullet Cluster is consistent with the standard $\Lambda$CDM cosmology. However, due to the limited volume of Millenium Simulation their result do not rely on direct measurement but rather on extrapolating the sub-halo velocity probability distribution to host halos with mass $>10^{14}$ h$^{-1}$ M$_\odot$. Furthermore, the relative velocity of the colliding clusters may well be different from that of the gas. As shown by \citet{Milosavljevic2007a} using 2-D hydrodynamical simulations this can be up to $\simeq 15 \%$ smaller \citep[see also][]{SpringelFarrar2007}.
To date the most detailed study of the Bullet Cluster has been performed by \citet{Mastropietro2008a} who have used 3-D non-cosmological hydrodynamical simulations to determine the initial physical configurations of the colliding clusters resulting in a system whose properties reproduce those observed in the Bullet Cluster. These include the displacement between the X-ray peaks and the mass distribution, the morphology of the shock velocity, the surface brightness and the projected temperature profiles across the shock. \citet{Mastropietro2008a} have shown that the colliding halos must have an initial velocity \mbox{v$_{12} \sim 3000$ km s$^{-1}$} with a mass ratio of $6:1$, an initial separation of $5$ Mpc (implying an initial redshift of $z=0.486$) and a collision angle $\theta_{12}$ close to 0 (i.e. head on collision). The identification of these parameters has provided criteria to select Bullet-like halo pairs in cosmological simulations, a crucial step to estimate the probability of finding the Bullet Cluster in a given cosmological setup.
\citet{Lee2010a} analysed the halo catalog from the MICE simulations \citep{Crocce2010} of $\Lambda$CDM cosmology at $z=0$ and $0.5$ to infer the pairwise velocities probability distribution for different mass ratios, distance separation and relative velocity. Quite remarkably they found that none of the analysed catalogues contains a system with parameters corresponding to that of the Bullet Cluster. Nevertheless, by fitting the probability density distribution to a Gaussian, they were able to extrapolate the probability to high relative velocities. They found the rate of occurrence of Bullet Cluster-like systems to be $P({\rm v}_{12}>3000\,\text{km s}^{-1})=3.3\times10^{-11}$ and $3.3\times10^{-9}$ at $z=0$ and $z=0.5$ respectively. \citet{Thompson2012a} performed a similar analysis of a set of $\Lambda$CDM simulations with DM mass resolution varying from 9$\times 10^9$ to 5.7$\times 10^{11}$ h$^{-1}$ M$_\odot$ and box sizes ranging from $200$ h$^{-1}$ Mpc to $2$ h$^{-1}$ Gpc. By extrapolating the cumulative distribution to high relative velocities these authors obtained $P({\rm v}_{12}>3000\,\text{km s}^{-1})=2.76\times10^{-8}$ at $z=0.489$ for masses $\ge 10^{14}$ M$_\odot$. \textcolor{black}{These probabilities imply that in the observable cosmic volume the existence of the Bullet Cluster pair is either incompatible with the $\Lambda$CDM scenario} or as argued by \citet{Thompson2012a} that the initial conditions determined from the analysis of non-cosmological hydrodynamical simulations by \citet{Mastropietro2008a} have to be revised to much lower values of v$_{12}$ \cite[see e.g.][for a recent study]{Lage2013a}.
A critical point is that all these analyses rely on extrapolating the tail of the pairwise velocity probability distribution to high-velocities. This is a direct consequence of the limited volumes of the numerical simulations from which these results have been derived. Furthermore, the probability of observing the Bullet Cluster has been directly estimated from the tail of the probability density distribution fitted to a Gaussian \textcolor{black}{which may suffer of potential biases especially if the tail of the distribution is non-Gaussian.}
In the work presented here we improve these studies in several ways. We use the catalog of halos from the Dark Energy Universe Simulation - Full Universe Runs (DEUS-FUR) which cover the entire observable cosmic volume with a box-length of $21$h$^{-1}$ Gpc and $8192^3$ dark matter particles \citep{Alimi2012,Rasera2014}. These simulations provide an unprecedented large statistical sample to test the rareness of halo properties for different cosmological models. The large simulation volume allows us to perform a detailed analysis of the pairwise velocity especially in the high-velocity tail. Building upon this study we use the Extreme Value Statistics (EVS) to quantify the probability of observing the Bullet Cluster. This enables us for the first time to perform a cosmological model comparison of the Bullet Cluster observation.
The paper is organized as follows. In Section \ref{Sec2}, we introduce the N-body simulation data, the halo finder algorithm and the halo pair selection method. In Section \ref{dndv_fof}, we described the dependence of the pairwise velocity function on the halo finder parameter and discuss the physical implications, while in Section \ref{sec_cosmoz} we study the redshift evolution and cosmology dependence. In Section \ref{evsbull} we introduce the Extreme Value Statistics and present the results of its application to the Bullet Cluster. Finally in Section \ref{conclu} we discuss our conclusions.
\section{N-Body Simulation Dataset}
\label{Sec2}
\subsection{DEUS Full Universe Runs}
We use numerical data issued from the DEUS-FUR project. This consists of three N-body simulations with $8192^3$ dark matter particles and box size of \mbox{(21000 h$^{-1}$ Mpc)$^3$} enclosing the observable volume of a flat $\Lambda$CDM cosmology and two dark energy models with different expansion histories. The simulations have been realized using the application AMADEUS - A Multi-purpose Application for Dark Energy Universe Simulation - expressly developed for the DEUS-FUR project \citep{Alimi2012}. This includes the generator of Gaussian initial conditions for which we use an optimized version of the code MPGRAFIC \citep{prunet08}, the N-body solver which is a version of the RAMSES code \citep{teyssier02} specifically improved to run on a large number of cores (\mbox{$>$ 40000} MPI tasks in production mode) and an optimized Friend-of-Friend halo finder \citep{Roy2014}. RAMSES solves the Vlasov-Poisson equations using an Adaptive Mesh Refinement (AMR) Particle Mesh method with the Poisson equation solved with a multigrid technique \citep{guillet11}. A detailed description of the algorithms, optimization schemes and the computing challenges involved with the realization of DEUS-FUR is given in \cite{Alimi2012}. All simulations share the same phase of the initial conditions. The coarse grid of the AMR hierarchy contains $8192^3$ resolution elements and is allowed to refined six times, reaching a formal spatial resolution of 40 h$^{-1}$ kpc, while the particle mass resolution for the different models is \mbox{$m_p \simeq 10^{12}$ h$^{-1}$ M$_\odot$}. Such resolution roughly corresponds to the size and mass of the Milky Way.
\begin{figure*}
\begin{minipage}{17cm}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=8cm,height=6cm]{DEUS_w_om_cont.eps}&
\includegraphics[width=8cm,height=6cm]{DEUS_w_sig_cont.eps}
\end{tabular}
\caption{\label{contours}Left panel: Marginalized $1$ and $2\sigma$ contour plots in the $\Omega_m-w$ planes from the likelihood analysis of the CMB data from WMAP-7yr observations (solid lines) and in combination with luminosity distance measurements to SN Ia from the UNION dataset (filled contours). The dashed line indicates the luminosity distance degeneracy curve. Right panel: Marginalized $1$ and $2\sigma$ contour plots in the $\sigma_8-w$ plane from CMB data only (solid lines). The dashed line indicates the degeneracy curve from the CMB anisotropy power spectra.}
\end{center}
\end{minipage}
\end{figure*}
The simulated cosmologies consist of a flat $\Lambda$CDM model best-fit to the WMAP-7yr data \citep[$\Lambda$CDM-W7,][]{spergel07}, a quintessence model with Ratra-Peebles potential \citep[RPCDM,][]{RatraPeebles} and a ``phantom'' dark energy model with constant equation of state $w<-1$ \citep[$w$CDM,][]{KamionCald}.
The model parameters have been calibrated to fit the Cosmic Microwave Background anisotropy power spectra from WMAP 7-year observations \citep{spergel07} and luminosity distances measurements to Supernova Type Ia from the UNION dataset \citep{union}. In particular, the mean cosmic matter density, $\Omega_m$, has been chosen within the marginalized $\sim 1\sigma$ contour in the $\Omega_m-w$ plane along the degeneracy line of the SN Ia data (see left panel in Fig. \ref{contours}); while the values of the root-mean-square fluctuation amplitude of the density contrast at $8$ h$^{-1}$ Mpc, $\sigma_8$, has been chosen within the marginalized $\sim1\sigma$ confidence contours in the $\sigma_8-w$ plane nearly parallel to the degeneracy line of the CMB data (see right panel in Fig. \ref{contours}). This particular choice is motivated by the fact that through the analysis of the DEUS-FUR simulations we aim to test whether observables of the non-linear clustering of matter can break the degeneracies affecting current cosmological parameter constraints. A summary of the cosmological model parameter values and the simulation characteristics are reported in Table \ref{DEUSSALL1}.
\setcounter{table}{0}
\begin{table}
\begin{center}
\begin{tabular}{cccc}
\hline
Parameters & RPCDM & $\Lambda$CDM-W7 & $w$CDM \\
\hline
\hline
$\Omega_m$ & 0.23& 0.2573 & 0.275 \\
$\Omega_b h^2$ & 0.02273& 0.02258 & 0.02258 \\
$\sigma_8$& 0.66 & 0.8 & 0.852 \\
$w_0$& -0.87 & -1 & -1.2 \\
$w_1$ & 0.08 & 0 & 0 \\
\hline
$z_{ini}$ & 94 & 106 & 107 \\
$m_p ($h$^{-1}$ M$_\odot$) & $1.08 \times 10^{12}$ & $1.20 \times 10^{12}$ & $1.29 \times 10^{12}$ \\
$\Delta x$ (h$^{-1}$ kpc) & 40 & 40 & 40 \\
\hline
\end{tabular}
\caption{Cosmological parameter values of the DEUS-FUR simulated cosmologies. For all models the scalar spectral index is set to $n_s=0.963$ and the Hubble parameter $h=0.72$. We also report the values of a linear equation of state parameterization $w(a) = w_0 + w_1(1- a)$ for the different models (see \citet{Alimi2010} for details). In the bottom table we list the values of the initial redshift of the simulations $z_{ini}$, the particle mass $m_p$ and the comoving spatial resolution $\Delta x$. For all three simulations the box-length is L$_{\textrm {box}}=21000$~h$^{-1}$ Mpc and the number of dark matter particles is $8192^3$.}
\label{DEUSSALL1}
\end{center}
\end{table}
\subsection{Halo Finder and Halo Pair Selection}
The detection of halos in the DEUS-FUR simulations is performed with a highly-scalable parallelized version of the Friend-of-Friend halo finder algorithm \citep[][]{Roy2014} implemented in the AMADEUS application. This algorithm detects halos as groups of particles characterized by an interparticle distance smaller than a given linking length (in units of the mean interparticle distance) or percolation parameter $b$.
In the study presented in Section \ref{dndv_fof} we consider halos detected with $b=0.1,0.15$ and $0.2$ respectively. In order to limit the effect of numerical artifacts due to low number of particles we only consider halos with $>100$ particles, which corresponds to halo masses \mbox{{$M\gtrsim 10^{14}$h$^{-1}$ M$_\odot$.}}
\begin{table}
\begin{center}
\begin{tabular}{ccccc}
\hline
&Redshift& RPCDM & $\Lambda$CDM-W7 & $w$CDM \\
\hline
\hline
\multirow{3}{*}{\rotatebox{90}{Single}} & $0$ & 76,180,615 & 144,630,773 & 169,186,215\\
& $0.3$& 40,613,387 & 90,788,115 & 109,227,390\\
&$0.5$& 24,554,151 & 61,804,451 & 74,966,075 \\
\hline
\hline
\multirow{3}{*}{\rotatebox{90}{Pair}} & $0$ & 47,727,489 & 125,555,136 & 156,600,237 \\
& $0.3$& 17,297,267 & 58,501,507 & 76,027,601 \\
&$0.5$& 7,700,836 & 31,192,185 & 40,877,856 \\
\hline
\end{tabular}
\caption{Total number of FoF(b=0.2) single halos with $>100$ particles and pair of halos with separation $<15$ h$^{-1}$ Mpc detected in the comoving volume of the three cosmological DEUS-FUR simulations at redshift $z=0,0.3$ and $0.5$ respectively.}
\label{NZ}
\end{center}
\end{table}
In Table \ref{NZ} we report the total number of FoF(b=0.2) halos detected in the DEUS-FUR simulations at different redshifts. These are vast halo catalogues for which the selection of halo pairs and the calculation of relevant quantities poses a challenging computational problem. In fact, the complexity of a standard brute force computation of the relative velocities for all pair separations grows with the square of the number of halos. This leads to a prohibitive computational time as soon as the number of halos exceeds $\sim 10^6$. However, we are interested only in halo pairs characterized by a small distance separation such as the Bullet Cluster initial configuration found in \citet{Mastropietro2008a}. Therefore, we can significantly reduce the number of computations by using an octree space decomposition which enables the computation of pairs up to a maximum distance $d_{12}^{\rm max}$. To be conservative we set $d_{12}^{\rm max}\approx3\,R_{\rm max}$, where $R_{\rm max}$ is the radius of the most massive halo in the catalogues. This gives a maximal distance separation \mbox{$d_{12}^{\rm max}\sim 15$ h$^{-1}$ Mpc}. Thanks to such a space decomposition we compute all relevant quantities for all pairs within the pruning radius, thus avoiding the most time consuming long-range computations. The implementation of this algorithm allows us to compute the relative velocities of about 100 million pairs in less than 2 minutes on a 64 core (Intel Xenon CPU X7550$@$2.00Ghz) local machine. The number of pairs within 15 h$^{-1}$ Mpc for the three cosmological catalogs at three redshifts of interest is given in Tabel \ref{NZ}. As we can see at $z=0$ the number of pairs is a significant fraction of the total number of halos detected in the simulations which is indicative of the high level of clustering of such objects compared to that of the average density field with a mean inter-halo separation of 40-70 h$^{-1}$ Mpc.
\section{Pairwise Velocities and Friend-of-Friend Halos}\label{dndv_fof}
In this section we study the dependence of the pairwise velocity function on the halo definition specified by the value of the percolation parameter $b$. It is usually understood that for a given value of $b$ the FoF algorithm selects halos whose boundary has approximatively a fixed isodensity surface. For instance, the local surface overdensity with respect to the mean matter density of two particles within a sphere of radius $b$ is $\delta_{\textrm{FOF}}\sim 1/ 2b^3$ \citep{Summers1995,Audit1998}. In the case $b=0.2$ this gives $\delta_{\textrm{FOF}}\sim 60$ which assuming an isothermal density profile, $\rho(r) \propto r^{-2}$, corresponds to an enclosed overdensity with respect to the mean matter density of $\Delta_m \sim 180$. This is close to the value of the virial overdensity predicted by the spherical collapse model in the Einstein-De Sitter cosmology. That is why the percolation parameter is commonly set to $b=0.2$. However, \cite{More2011} have shown that the boundary of FoF halos is not associated to a unique local surface overdensity but is distributed around a characteristic value (for $b=0.2$ this is $\sim 80$). In addition the profile is not isothermal. As a consequence the enclosed overdensity is much higher than 180 and is found to vary in the range $\sim 250$ to $\sim 600$ at $z=0$. It is important to keep this in mind when comparing the properties of N-body halos to observations. In fact, the mass of clusters is usually estimated in terms of the enclosed spherical overdensity with respect to the critical cosmic density that, depending on redshift and cosmology, may correspond to different percolation parameter values.
\begin{figure}
\includegraphics[width=0.48\textwidth]{dndv12_cosmo_bfof_HR.eps}
\caption{Probability density function of the pairwise velocity from the DEUS-FUR $\Lambda$CDM-W7 simulation at $z=0$ for pairs with separation $d_{12} < 15$ h$^{-1}$ Mpc detected assuming linking-length values $b=0.1$ (light grey), $b=0.15$ (grey) and $b=0.2$ (black) respectively.}
\label{bfof_dep}
\end{figure}
In Fig. \ref{bfof_dep} we plot the pairwise velocity distribution $dn/d{\rm v}_{12}$ for halo pairs with a separation $d_{12}<15$ h$^{-1}$Mpc from the DEUS-FUR $\Lambda$CDM-W7 simulation at $z=0$ detected with $b=0.1,0.15$ and $0.2$ respectively (corresponding to a typical variation of the enclosed overdensity $\Delta$ by a factor of $\sim$8). We can see that for increasing values of $b$ the peak of \textcolor{black}{the distribution increases while its width decreases}. This trend is a direct consequence of the halo definition corresponding to the different values of $b$. Such a difference implies a change of the halo mass function that manifests in the pairwise velocity distribution.
\begin{figure*}
\includegraphics[width=0.98 \textwidth]{Multivariate_bfof_lobes_HR.eps}
\caption{Isocontours of the multivariate pairwise velocity probability density function of halo pairs from the DEUS-FUR $\Lambda$CDM-W7 simulation at redshift $z=0$ shown as function of the relative velocity ${\rm v}_{12}$ and the average pair mass $\log_{10} \langle M_b \rangle$, for separation distances $d_{12}=2,5,10$ and $15$ h$^{-1}$Mpc (panels from top to bottom) and linking-length $b=0.2,0.15$ and $0.1$ (panels from left to right) respectively. The isocontours are composed of ten logarithmically spaced bins spanning the range $[10^{-8},10^{-3}]$. For \mbox{$d_{12}<5$ h$^{-1}$Mpc} the pairwise velocity function is characterised by a single population of halos, while for $d_{12}>5$ h$^{-1}$Mpc we notice the emergence of a second population of halo pairs in the low-average-mass range. The former is indicated with Lobe 1 (red dashed ellipse) and its tail consists high-mass low-velocity mergers, the latter which we mark as Lobe 2 (blue dash-dotted line) has a tail consisting of low-mass high-velocity mergers.}
\label{multi_bfof_dep}
\end{figure*}
We may gain a better insight from Fig. \ref{multi_bfof_dep} where we plot isocontours of the multivariate pairwise velocity distribution as function of the relative velocity ${\rm v}_{12}$ and the arithmetic mean mass $\langle $M$_b \rangle$ of halo pairs detected with $b=0.2,0.15$ and $0.1$ (panels from left to right) and distance separation $d_{12}=2,5,10$ and $15$ h$^{-1}$ Mpc (panels from top to bottom) respectively.
In the case of pairs with $d_{12}=2$ h$^{-1}$ Mpc (top panel) the size of the isocontours increases for decreasing values of $b$. This is because FoF(b=0.1) not only detects individual groups of particles but also sub-structures within halos detected by FoF(b=0.2), thus the former selects a greater number of halo pairs. This population of pairs is distributed in a lobe structure (Lobe 1) which extends from small average masses with low relative velocities to large average masses with moderate relative velocities. The spread of Lobe 1 increases for distance separations $d_{12}>2$ h$^{-1}$ Mpc, simply because FoF detects a greater number of halo pairs. Notice that the tails of the distribution does not exceeds $\approx 2000$ km s$^{-1}$. This indicates that halo pairs in the tail of Lobe 1 correspond to low velocity massive mergers. This could be the case of some observed systems such as MACS J0025.4-1222 \citep{Bradac2008a} or \mbox{ACT-CL J0102-4915} \citep{Menanteau2012a} which \textcolor{black}{are examples of pairs of massive interacting clusters with Bullet-like baryonic features.}
For $d_{12}>2$ h$^{-1}$ Mpc we can see the emergence of a second lobe (Lobe 2) that causes the multivariate probability density distribution to become bimodal. This second lobe consists of small average mass pairs with velocities extending up to $\approx 3000$ km s$^{-1}$. This is the population that seems to better correspond to the characteristics of 1E0657-56, the Bullet Custer. Notice that the velocity tail of Lobe 2 shifts to larger values for decreasing values of $b$. This is because FoF detects a greater number of \textcolor{black}{satellite halos that translates into an increase of the number of} small mass high-velocity pairs as $b$ decreases.
Although the differences of the pairwise velocity probability density function between the case $b=0.15$ and $b=0.2$ might look minor in the high-velocity tail, these may have an important impact on the evaluation of the probability of high-velocity colliding clusters. \textcolor{black}{Hence, when comparing to observations an important point concern the choice of the percolation parameter which has to be as consistent as possible with the observational mass definition. For instance in \citet{Mastropietro2008a} the colliding halos are spherical objects with a Navarro-Franck-White (NFW) profile \citep{Navarro96,Navarro97} characterized by a concentration parameter $c_{\textrm NFW}\sim 7$ for which the mass is defined in terms of the virial mass $M_{vir}=4\pi/3 r_{vir}^3\Delta_{c}\rho_c$, which is the mass contained in a spherical region of radius $r_{vir}$ enclosing an overdensity $\Delta_{c}=200$ with respect to the critical density of the universe $\rho_c$ at the redshift of the halo. These halos are sampled with $\sim 1000$~particles, thus following \citet{More2011} an enclosed overdensity of $200\rho_c$ at $z\sim 0.5$ roughly corresponds to applying FoF with a linking-length $b\sim 0.15$.}
\textcolor{black}{In the following, we therefore adopt a linking-length $b=0.15$ and limit our analysis to halo pairs with distance separation $d_{12}<10$ h$^{-1}$ Mpc which is the minimum distance for which two massive merging halos with virial radius $\sim 5$ h$^{-1}$ Mpc can be detected as distinct objects.}
\section{Statistics of pairwise velocities}\label{sec_cosmoz}
\subsection{Redshift evolution and cosmology dependence}
\textcolor{black}{We now focus on evaluating the dependence of the pairwise velocity function on redshift and cosmology.}
\textcolor{black}{In Fig. \ref{pairwise_redshift} we plot the probability density function associated with $dn/d{\rm v}_{12}$ for halo pairs from the DEUS-FUR $\Lambda$CDM-W7 simulation with distance separation $d_{12}<10$ h$^{-1}$ Mpc at $z=0.5,0.3$ and $0$ respectively. First, we may notice that the amplitude remains constant with redshift, this is the normalization of each curve is different as the number of halo pairs grows with cosmic time. Another effect concerns the tail of the velocity function which tends to slightly increase towards higher velocities from $z=0.5$ to $0$. For instance we find the maximal relative velocities to slowly evolve with redshift with ${\rm v}_{12}^{\rm max}=3609,3799$ and $4000$ km s$^{-1}$ at $z=0.5,0.3$ and $0$ respectively.} Because of this, we can expect that in a given cosmological model the probability (defined as the ratio of the velocity function to the total number of pairs) of finding a halo pair with a large relative velocity at redshift $z=0$ and $z=0.5$ is not significantly different.
\begin{figure}
\includegraphics[width=0.48 \textwidth]{dndv12_z_2_b01500_HR.eps}
\caption{Redshift evolution of the probability density function of the pairwise velocity for FoF(b=0.15) halo pairs from the DEUS-FUR $\Lambda$CDM-W7 simulation with distance separation $d_{12} < 10$ h$^{-1}$ Mpc at $z=0.5$ (black), $z=0.3$ (grey) and $z=0$ (light grey) respectively.}
\label{pairwise_redshift}
\end{figure}
The advantage of using the halo catalogues from the DEUS-FUR simulations can be appreciated from Fig. \ref{pairwise_redshift}. Despite statistical scatter, even at $z=0.5$, the high-velocity tail of the pairwise velocity function is resolved to $v_{12}\approx 3500$ km s$^{-1}$. In this range previous analyses had to strongly rely on extrapolation from fitting functions calibrated to lower relative velocities. For instance, \citet{Thompson2012a} approximated the cumulative pairwise velocity distribution of FoF(b=0.2) halo pairs in $\Lambda$CDM model simulations at $z=0.489$ with a quadratic fit which we plot in the top panel of Fig. \ref{cumulativev12} against the DEUS-FUR $\Lambda$CDM-W7 results. Quite remarkably we can see that it provides a very good approximation up to intermediate velocities.\footnote{The quadratic fit from \citet{Thompson2012a} has been derived from the analysis of FoF(b=0.2) halo pairs in simulations with mass and spatial resolution different from those of the DEUS-FUR simulations. Such differences may affect the halo mass function and introduce a systematic bias in the cumulative pairwise velocity distribution. However, for halo pairs with average masses $>10^{14}\,\textrm{M}_\odot$ this effect remains negligible.} \textcolor{black}{In contrast, large discrepancies occurs in the high-velocity tail as can be seen from the bottom panel of \mbox{Fig. \ref{cumulativev12}}}. In the same figure we also plot the cumulative distribution from our reference catalogue of halo pairs detected with FoF(b=0.15). As expected from the discussion in Section \ref{dndv_fof} this is characterised by both a greater number of halo pairs and a longer tail at high-velocity compared to $b=0.2$.
\begin{figure}
\includegraphics[width=0.48 \textwidth]{cumulativev12_b01500_2_HR.eps}
\caption{Top panel: cumulative pairwise velocity function of halo pairs with distance separation $d_{12} < 10$ h$^{-1}$ Mpc from the DEUS-FUR $\Lambda$CDM-W7 simulation at $z=0.5$ detected with FoF(b=0.2) (blue line) and FoF(b=0.15) (red line). The black solid line corresponds to the quadratic fit from \citet{Thompson2012a}. \textcolor{black}{Bottom panel: ratio of the cumulative pairwise velocity functions to the \citet{Thompson2012a} fit.}
}
\label{cumulativev12}
\end{figure}
\begin{figure}
\includegraphics[width=0.48 \textwidth]{dndv12_cosmo_2_b01500_HR.eps}
\caption{Pairwise velocity function at $z=0$ for FoF(b=0.15) halo pairs with distance separation $d_{12} < 10$ h$^{-1}$ Mpc from the DEUS-FUR RPCDM, $\Lambda$CDM-W7 and $w$CDM simulations respectively.
}
\label{cosmoend}
\end{figure}
Let us now turn to the cosmological dependence of the velocity function. In Fig. \ref{cosmoend} we plot $dn/d{\rm v}_{12}$ at $z=0$ for a distance separation $d_{12} < 10$ h$^{-1}$ Mpc for RPCDM, $\Lambda$CDM-W7 and $w$CDM. The first noticeable difference is the overall amplitude of the various curves which is essentially caused by the difference of the mass function of the DEUS-FUR simulated cosmologies. We can also notice that in the high-velocity interval the velocity functions have slightly different slopes, with the $w$CDM and $\Lambda$CDM-W7 models showing a heavier tail than RPCDM. This is indicative of the fact that halo pairs with extreme relative velocities are a sensitive probe of the underlying cosmological model. Understanding the mechanisms responsible for this dependence requires a physical analysis that is beyond the scope of this paper.
Even in the lack of such study we can have an idea of the dominant cosmological parameter dependence by evaluating the average pairwise velocity $\bar{\rm v}_{12}$. This can be directly inferred from the pairwise probability distribution function obtained from $dn/d{\rm v}_{12}$ and confronted with predictions from the following relation derived in the context of stable clustering \citep{Juszkiewicz1999,Caldwell2001}:
\begin{equation}
-\frac{\bar{\rm v}_{12}(x,a)}{Hr}=\frac{a}{3[1+\xi(x,a)]}\frac{\partial \bar \xi(x,a)}{\partial a} \ ,
\label{eqn1}
\end{equation}
where $\xi$ is the two-point correlation function of the density field, $a$ is the expansion factor, $r=ax$ is the proper separation, $H$ is the Hubble rate and
\begin{equation}
\bar \xi(x,a)=\frac{3}{x^3}\int_0^x \xi(y,a) y^2 dy
\label{eqn2}
\end{equation}
is the two-point correlation function averaged over a sphere of radius $x$. Evaluating Eqs. (\ref{eqn1}) and (\ref{eqn2}) using the correlation function $\xi$ of the density field of each of the DEUS-FUR simulations, we obtain the following average pairwise velocities: \mbox{$\bar{\rm v}_{12} = 439$ km s$^{-1}$} for the RPCDM model, $\bar{\rm v}_{12} = 490$ km s$^{-1}$ for the $\Lambda$CDM-W7 and $\bar{\rm v}_{12} = 507$ km s$^{-1}$ for the $w$CDM model. These values are within $5\%$ of those directly estimated from the numerical data. Their variation is essentially due to the different values of $\sigma_8$ of the DEUS-FUR cosmologies. This can be understood by taking the ratio of the average pairwise velocity given by Eq. (\ref{eqn1}) for a given $\sigma_8$ with respect to a reference one $\bar{\rm v}_{12}(\sigma_{8,{\rm ref}})$:
\begin{equation}
\frac{\bar{\rm v}_{12} \ (\sigma_8)}{\bar{\rm v}_{12} \ (\sigma_{8,{\rm ref}})}= \frac{\sigma_8}{\sigma_{8,{\rm ref}}} \frac{1+ \xi(r)}{1+\frac{\sigma_8}{\sigma_{8,{\rm ref}}}\xi(r)} \ .\label{eqn3}
\end{equation}
\textcolor{black}{Assuming that the shape of the power spectrum does not change over the range of scales where the stable clustering regime occurs \citep{Smith2003a} and considering as a reference case the $\Lambda$CDM-W7 values of $\bar{\rm v}_{12}$ and $\sigma_8$, we recover from Eq. \ref{eqn3} the estimated values of the average pairwise velocity of both RPCDM and $w$CDM to better than a few per cent.}
\textcolor{black}{In Fig. \ref{sigma8_dep} we plot the normalized probability density distribution of the pairwise velocity of the three simulated cosmologies. The $\sigma_8$ dependence described above can be seen here on the fact that the distributions have very similar average and overall amplitude. This is because their respective normalizations encode differences of the mass function of the underlying cosmological models that are mostly due to the different $\sigma_8$ values. On the other hand we can see that the high-velocity tail of the distributions is where the cosmological models differ the most. This indicates that the high-velocity tail carry information not only on $\sigma_8$, but also on the cosmic matter density and the properties of the Dark Energy which characterize the simulated models.}
\begin{figure}
\includegraphics[width=0.48 \textwidth]{dndv12_cosmo_2_b01500_HR2.eps}
\caption{Probability density pairwise velocity function at $z=0$ for FoF(b=0.15) halo pairs with distance separation $d_{12} < 10$ h$^{-1}$ Mpc from the DEUS-FUR RPCDM, $\Lambda$CDM-W7 and $w$CDM simulations respectively.}
\label{sigma8_dep}
\end{figure}
\subsection{Bullet-like halo pairs in DEUS-FUR cosmologies}
In order to identify extreme halo pairs in the DEUS-FUR simulation it is instructive to consider the redshift and cosmology dependence of the multivariate pairwise velocity distribution (see Appendix \ref{Appendix2}). This is shown in Fig. \ref{mvcos} for halo pairs with distance separation $d_{12}<10$ h$^{-1}$ Mpc at $z=0.5,0.3$ and $0$ (panels top to bottom) for DEUS-FUR RPCDM, $\Lambda$CDM-W7 and $w$CDM simulations (panels left to right) respectively. As expected from Section \ref{sec_cosmoz}, for a given cosmological model the range of relative velocities does not vary significantly as function of the redshift compared to interval variation of the average mass of the pairs. Indeed, this is due to the different redshift evolution of the halo mass function compared to that of the pairwise velocity function. Furthermore, we can see that in the RPCDM case the tail of velocity distribution at all redshift remains confined to low velocities. This is not the case of $\Lambda$CDM-W7 and $w$CDM for which we have a large number of pairs with average mass larger than $10^{15.2}$ M$_\odot$ and relative velocity v$_{12}>2100$ km s$^{-1}$. For these models the bimodality of the multivariate pairwise distribution is more pronounced than in RPCDM. In particular, the halo pairs in the tail of Lobe 1 and Lobe 2 of the multivariate distribution clearly indicate that observations of high-mass moderate-velocity merging clusters and low-mass high-velocity ones can provide powerful cosmological probes.
We are especially interested in extremal halos belonging to the latter category. In Appendix \ref{Appendix1} we summarize the characteristics of the highest velocity pairs and the most massive ones detected in the DEUS-FUR simulations. We find that the properties of such extremal pairs vary \textcolor{black}{with the cosmological model}. As an example, from the values quoted in Table \ref{cosvel} we notice that even at $z=0$ the RPCDM simulation has no pairs with relative velocities exceeding \mbox{$\approx 3000$ km s$^{-1}$} and an average mass $M>2 \times 10^{14}$~h$^{-1}$M$_{\odot}$. \textcolor{black}{More generally from Tables \ref{cosvel} and \ref{cosmass} we can infer that the deficiency of high-velocity halo pairs with mass above $2 \times 10^{14}$ h$^{-1}$ M$_\odot$ and that of massive pairs with relative velocities above \mbox{$1500$ km s$^{-1}$} tend to disfavor such a cosmological model.} The extremal halo pairs in RPCDM are low-velocity massive mergers with large distance separation. In contrast, in $w$CDM the highest velocity pairs all exceed \mbox{$\approx 4000$ km s$^{-1}$}, though their average masses remain small.
For comparison, in the top panels of Fig. \ref{mvcos} we plot the average mass and relative velocity of the Bullet Cluster (blue solid lines) as inferred from the analysis of \cite{Mastropietro2008a}. Using this as a reference of the extreme halo pairs we can see that RPCDM has no candidate pairs reproducing the Bullet Cluster characteristics. In the $\Lambda$CDM-W7 case the best candidate pair has a main halo with mass \mbox{$M_1=5.95 \times 10^{14}$ h$^{-1}$ M$_\odot$} and a smaller halo with mass $M_2=1.22 \times 10^{14}$ h$^{-1}$ M$_\odot$ (corresponding to a mass ratio $\sim 5:1$) separated by a distance of 8.4 h$^{-1}$ Mpc, which experience an head on interaction with a relative velocity of 3011 km s$^{-1}$. Notice that such an object is absent from the list of extreme halo pairs shown in Appendix since it has neither an extreme velocity nor a very high mass. In the $w$CDM model the best candidate is characterised by a main halo with a mass $M_1=8.40 \times 10^{14}$ h$^{-1}$ M$_\odot$ and a lighter halo of mass $1.97 \times 10^{14}$ h$^{-1}$ M$_\odot$ (corresponding to a mass ratio of $\sim 4:1$) separated by a distance of 8.8 h$^{-1}$ Mpc and a relative velocity 2839 km s$^{-1}$. \textcolor{black}{This candidate is no better than that of $\Lambda$CDM. This seems contradictory given the cosmological dependence of the high-velocity tail of the multivariate distribution shown in Fig. \ref{mvcos}. However, such discrepancy can be simply a consequence of the specific realization of the simulation run, such as the phase of the initial conditions. Hence, an object by object comparison is no meaningful in assessing the extremeness of the Bullet Cluster. Such estimation can only be performed through a statistical analysis of extreme halo pairs. In the next Section we will discuss the use of these pairwise velocity catalogs to infer the probability of observing the Bullet Cluster.}
\begin{figure*}
\includegraphics[width=0.98\textwidth]{multivariate_zcosmo_HR.eps}
\caption{Isocontours of the multivariate pairwise velocity probability density function for halo pairs with distance separation $d_{12}<10$ h$^{-1}$Mpc from the DEUS-FUR simulations as function of the relative velocity ${\rm v}_{12}$ and the average pair mass $\log_{10} \langle M \rangle$ for $z=0.5,0.3$ and $0$ (panels from top to bottom) and RPCDM, $\Lambda$CDM-W7 and $w$CDM (panels left to right) respectively. The isocontours are composed of ten logarithmically spaced bins spanning the range $[10^{-8},10^{-3}]$. In the top panels the solid blue lines specify the average mass and relative velocity of the Bullet Cluster from \citep{Mastropietro2008a}. \textcolor{black}{For illustrative purposes we also show the characteristics of two other Bullet-like systems: the red dashed lines in the top panels corresponds to the characteristics of MACS J0025.4-1222 at the observed redshift of the system \citep{Bradac2008a}, while in the bottom panel the green dot-dashed lines corresponds to the initial characteristics of the Abell 3376 system \citep{Machado2013a}}.
}
\label{mvcos}
\end{figure*}
\section{Extreme Value Statistics of Pairwise Velocities}\label{evsbull}
\subsection{Methodology}
Extreme Value Statistics, originally pioneered by \citet{Frechet1927}, \citet{FisherTippett1928}, \citet{Gumbel1935} and \citet{Gnedenko1943}, has been applied to a wide variety of problems to model the probability of extreme events. Here, we briefly review the basic formalism.
Consider a set of independent identically distributed $N$ random variates $\{X_1,...,X_N\}$ drawn from a cumulative distribution $F(x)$ and $X_\textrm{max}=\textrm{max}\{X_1,...,X_N\}$. It is easy to show that in such case the cumulative distribution function of the maximum of the first $N$ observations is given by
\begin{equation}
P(X_\textrm{max}\le x)=F^N(x).
\end{equation}
This is the so called \enquote{exact} EVS approach in which $F(x)$ is well known. In the large $N\rightarrow\infty$ limit, it is possible to show that the cumulative distribution function of extreme observations tends to the Generalized Extreme Value (GEV) distribution:
\begin{equation}
P_{[\mu,\sigma,\xi]}(x)=\exp\left\{-\left[1+\xi \left(\frac{x-\mu}{\sigma}\right)\right]^{-1/\xi}\right\},\label{gev}
\end{equation}
defined for $1+\xi(x-\mu)/\sigma>0$, where $\mu$ is the location parameter, $\sigma$ is the scale parameter and $\xi$ is the tail index (or shape parameter), which generalizes the central-limit theorem to extremal subset of data. Depending on the value of $\xi$, Eq. (\ref{gev}) reduces to three possible functional forms: the Gumbel (or type I) distribution ($\xi=0$), the Fr\'echet (or type II) distribution ($\xi>0$) and the Reversed Weibull (or type III) distribution ($\xi<0$).
Contrary to the exact EVS approach, the use of the Generalized Extreme Value distribution does not require prior knowledge of the underlying cumulative function of the random variates. Instead, it uses these observations to infer the GEV distribution parameters. This is done by classifying the data into blocks of arbitrary size, determining the maxima in each block and inferring the GEV parameters by best fitting the GEV function to the distribution of maxima. A potential disadvantage of this block maxima method is the fact that data need to be sampled. This may cause some loss of valuable rare information or inclusion of non-extremal events.
A complementary approach that is more suited to our purposes consists in using the Generalized Pareto distribution (GPD). This corresponds to Taylor expanding the tail of Eq. (\ref{gev}) to obtain the cumulative distribution function of observing extreme events above a fixed threshold $\mu$. This reads as
\begin{equation}
P_{[\mu,\sigma,\xi]}(x)=1- \left [ 1+ \xi \left ( \frac{x-\mu}{\sigma} \right)\right ]^{-1/\xi},\label{gpd}
\end{equation}
defined for $1+\xi(x-\mu)/\sigma>0$. Again depending on the value of $\xi$ we have different probabilities of the extreme events. In particular $\xi>0$ ($\xi<0$) corresponds to a long (short) tail distribution. Instead the case $\xi=0$ corresponds to the distribution of events in the tail of a Gaussian distribution. Hence, studies that have extrapolated the probability of the relative velocity of Bullet Cluster in terms of a Gaussian pairwise velocity probability distribution can be seen as a limiting case of the EVS approach described here, with $\xi$ fixed to zero. Notice also that since the tail of the pairwise velocity function depends on the underlying cosmology, we can expect the GPD parameters to carry a strong cosmological dependence.
In the GPD approach the issue of sampling the data is replaced by the problem of choosing a suitable value of the threshold. A high threshold would result in a drastic reduction of the data sample, whereas a low threshold may include non-extremal data and thus bias the results toward a gaussian behaviour. This can be seen in Fig. \ref{MEF} for a subset of the pairwise velocities in the DEUS-FUR $\Lambda$CDM-W7 catalog at $z=0.5$ where we have classified the halo pairs according to three different velocity thresholds. Using pairs above the lowest threshold would lead to a gaussian biased estimation of the GPD parameters, while using points above the highest threshold provides a too small sample of extremal events to determine the GPD.
\begin{figure}
\includegraphics[width=0.48\textwidth]{MEF_function_HR.eps}
\caption{Pairwise velocity exceedances defined as the mean of the data minus the chosen threshold for three different thresholds. The corresponding exceedances are shown with different color codes.}
\label{MEF}
\end{figure}
Several statistical diagnostics have been considered to estimate a suitable threshold that segregates common events from extreme ones \citep[for a review see e.g.][]{Scarrott2012}. \textcolor{black}{Here, we focus on the mean residual life method and the threshold stability plot that have been developed in relation with EVS data analysis problems.}
The mean residual life method consists in plotting the mean excess, defined as the mean of the exceedances of the data minus the threshold, as function of the threshold itself. An optimal choice is then given by the lowest threshold value for which all higher thresholds give a sequence of mean excesses that is consistent with a straight line \citep[][]{Scarrott2012}.
\textcolor{black}{A mean residual life plot is shown in Fig. \ref{MED} for the halo pairs of the full DEUS-FUR $\Lambda$CDM-W7 simulation volume (blue solid line) and three subvolumes of boxlength $10500$ h$^{-1}$ Mpc (red dashed line), $5184$ h$^{-1}$ Mpc (orange dot-dashed line) and \mbox{$2592$ h$^{-1}$ Mpc} (yellow squared solid line) respectively. We can see a characteristic trend with the mean of the exceedances rapidly decreasing at low threshold values, while increasing linearly at intermediate thresholds and then sharply decreasing at large values. The main difference among the various volume catalogs is the interval extent and statistical uncertainty of the linear trend. In the case of the full DEUS-FUR volume this interval has the maximum extent implying a precise selection of the GPD threshold which also guarantee a stability of GPD inferred results. For smaller volumes the interval ranges is much smaller and more uncertain such that it becomes impossible to reliably select a threshold value.}
The threshold stability plot is joint diagnostic that consists in plotting the GPD shape and scale parameter values best fitting the data as function of the threshold. Then, an optimal threshold value is chosen such that for higher values the GPD parameters remain stable \citep[][]{Scarrott2012}. We show such a plot in Fig. \ref{fig:stability}.
\begin{figure}
\includegraphics[width=0.48\textwidth]{stability_3_b01500_2_HR.eps}
\caption{\textcolor{black}{Mean residual life plot of the mean excess as function of the threshold for the full DEUS-FUR $\Lambda$CDM-W7 simulation volume (blue solid line) and three subvolumes of boxlength $10500$ h$^{-1}$ Mpc (red dashed line), $5184$ h$^{-1}$ Mpc (orange dot-dashed line) and \mbox{$2592$ h$^{-1}$ Mpc} (yellow squared solid line) respectively. For the full volume the region of stability, where the mean excess evolves as a straight line, spans the range \mbox{$\sim 1500$ km s$^{-1}$} to \mbox{$\sim 2500$ km s$^{-1}$}. The selected threshold \mbox{$\mu=2100$ km s$^{-1}$} corresponds to the largest threshold value with smallest statistical errors. We can see that for decreasing volumes the stability region rapidly shrinks while becoming more uncertain such that for small simulation volumes it is not possible to reliably select a threshold for which the results inferred from the GPD remain stable}.
}
\label{MED}
\end{figure}
From Fig. \ref{MED} and Fig. \ref{fig:stability} we can see that the curves are nearly constant straight lines in the threshold range $2000$ to \mbox{$2500$ km s$^{-1}$}, thus we set $\mu=2100$ km s$^{-1}$. This guarantees the stability of the results with respect to the choice of the threshold. Performing a similar analysis for the other DEUS-FUR cosmological simulations we set $\mu=1870$ km s$^{-1}$ for the RPCDM case and \mbox{$\mu=2151$ km s$^{-1}$} for the $w$CDM model respectively.
\begin{figure}
\includegraphics[width=0.48\textwidth]{stability_4_b01500_HR.eps}
\caption{Tail index (top panel) and scale parameter (bottom panel) diagnostic plots for the generalized Pareto distribution distribution fit to the $\Lambda$CDM-W7 data as function of the threshold parameter.
}
\label{fig:stability}
\end{figure}
In Fig. \ref{GPfitting} we plot the Generalized Pareto density distributions best fitting the probability density distribution functions of halo pairs with distance separation $d_{12} < 10$ h$^{-1}$ Mpc and average pair mass $>$10$^{14}$ h$^{-1}$ M$_\odot$ from the DEUS-FUR cosmological simulation catalogs at $z=0.5$. \textcolor{black}{The selected thresholds for the different cosmological models are indicated by vertical dashed lines, while the dot-dashed lines correspond to the Gaussian tails (with $\xi=0$) for the same threshold and scale parameter values.}
\textcolor{black}{The best-fit values and the $68\%$ confidence interval of $\xi$ and $\sigma$ have been determined through a Monte Carlo Markov Chain likelihood analyses of the binned numerical data assuming Poisson errors. The inferred values are quoted in Table \ref{ParetoPar}. Notice that in all cases a Gaussian tail ($\xi=0$) is excluded at more than $99.7\%$ confidence level. Since the best-fit value of the shape parameter is positive this implies the probability distribution of extreme pairwise velocities is slightly heavy tailed, which increases the probability of finding high relative velocity pairs compared to previous studies that have simply assumed a Gaussian distribution \citep{Lee2010a,Thompson2012a}.}
\subsection{Application to the Bullet Cluster}
Having determined the GPD parameters from each of the DEUS-FUR halo pair catalogs, we are now able to estimate the probability of observing the Bullet Cluster for different DEUS-FUR cosmological models. This is obtained by integrating the probability density functions shown in Fig. \ref{GPfitting} from $\textrm{v}_{12}=3000\,{\rm km \,s^{-1}}$ to infinity. This probability has to be interpreted as the rate of occurrence of Bullet Cluster-like systems in comoving space.
In the $\Lambda$CDM-W7 case we find $P({\rm v}_{12}>3000\,{\rm km\,s^{-1}})=6.4\times 10^{-6}$, which is two orders of magnitude larger than previous estimates \citep{Lee2010a,Thompson2012a}. For the RPCDM model we obtain $P({\rm v}_{12}>3000\,{\rm km \,s^{-1}})=9.7\times 10^{-8}$, while we find $P({\rm v}_{12}>3000\,{\rm km \,s^{-1}})=1.7\times 10^{-5}$ for the $w$CDM case.
As shown in Section \ref{sec_cosmoz} the pairwise velocity distribution carries information on cosmological model parameters such as $\sigma_8$, as well as $\Omega_m$ and $w$ which differentiate the DEUS-FUR cosmologies. The value of these parameters have been selected along the $\sigma_8-w$ (and $\Omega_m-w$) degeneracy line of the CMB (and SN Ia) data. Henceforth, the Bullet Cluster inferred probabilities can be used to provide us with some qualitative constraints on these class of models. In particular, these suggest that the observation of the Bullet Cluster strongly disfavors Dark Energy models, such as RPCDM, which have an equation of state $w>-1$ for which CMB data enforce smaller $\sigma_8$ values to compensate for the greater amplitude of the integrated Sachs-Wolfe \citep{ISW} effect on the CMB temperature anisotropy power spectrum \citep{Kunzetal2004} while SN Ia data enforce a lower value of $\Omega_m$ to compensate for the shorter luminosity distance. These models are characterized by a lower level of matter clustering with respect to the standard $\Lambda$CDM-W7 model. In contrast, the probability of finding the Bullet Cluster increases in the case of Dark Energy models with more negative values of the equation of state $w\le -1$ for which CMB data enforces larger $\sigma_8$ values while the SN Ia data requires larger values of $\Omega_m$. These models are characterized by a higher level of matter clustering compared to the $\Lambda$CDM-W7 case. Henceforth, it is plausible that the statistical measurements of the rate of occurrence of bullet cluster-like systems which sample the tail of the pairwise velocity distribution have the potential probe Dark Energy and break degeneracy lines of the underlying cosmological parameters.
\begin{figure}
\includegraphics[width=0.48\textwidth]{GPfitting_b01500_HR.eps}
\caption{Probability density distribution of the pairwise velocity of halo pairs with average mass $>$10$^{14}$ h$^{-1}$ M$_\odot$ and distance separation \mbox{$d_{12} < 10$ h$^{-1}$ Mpc} from the DEUS-FUR simulations at $z=0.5$ for $\Lambda$CDM-W7 (blue points), $w$CDM scenario (red points) and RPCDM (green points) models respectively. The solid lines shows the tail of the Generalized Pareto distributions best fitting the numerical data with threshold values indicated by the vertical dashed lines. \textcolor{black}{The dot-dashed lines corresponds to the Gaussian tails with threshold and scale parameters set to that of the best-fit GPD tails of the three cosmologies.} The thin solid black line at \mbox{v$_{12}=3000$ km s$^{-1}$} corresponds to the Bullet Cluster relative velocity estimated in \citet{Mastropietro2008a}.
}
\label{GPfitting}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{cccc}
\hline
Parameters & RPCDM & $\Lambda$CDM-W7 & $w$CDM \\
\hline
\hline
$\xi$ & $0.073 \pm 0.010$ & $0.035 \pm 0.007$ & $0.020 \pm 0.008$ \\
$\sigma$ (km s$^{-1}$) & $159.4 \pm 0.5$ & $205.1 \pm 0.4$ & $218.3 \pm 0.4$ \\
\hline
$\mu$ (km s$^{-1}$) & $1870$ & $2100$ & $2151$ \\
\hline
\end{tabular}
\caption{Best-fit values and $1\sigma$ errors of the GPD parameters for the different DEUS-FUR cosmological models: $\xi$ is the tail index parameter, $\sigma$ is the scale parameter and $\mu$ is the threshold. \textcolor{black}{The constraints on $\xi$ indicates deviations from a Gaussian distribution at more than $99.7\%$ confidence level}.
}
\label{ParetoPar}
\end{center}
\end{table}
\section{Conclusion}\label{conclu}
We have explored the possibility of testing cosmological models through observations of extreme pairwise velocities of interacting galaxy clusters. To this purpose we have studied the properties of pairwise velocities from the halo catalogs of the DEUS-FUR cosmological simulations. Thanks to the large simulation volume we have been able to resolve the high-velocity tail of the pairwise velocity distribution. We have studied its dependence on the percolation parameter of the FoF halo finder, the distance separation, the redshift evolution and cosmology. We have shown that a particular attention has to be paid to the halo mass definition, since the choice of the percolation parameter especially alter the tail of pairwise velocity function. In the redshift range $z=0$ to $0.5$ the latter show minor evolution, while it significantly varies with cosmology. To have an idea of the cosmological model parameter dependence we have estimated the average pairwise velocity of the DEUS-FUR cosmologies using a model based on stable clustering. From the comparison with the mean value inferred from the DEUS-FUR halo pairs catalogs we have shown that most of the average cosmological dependence is driven by the value of $\sigma_8$ while the tail of the distribution carries information on $\sigma_8$, $\Omega_m$ and $w$ which differentiate the DEUS-FUR cosmologies. As such observations of extreme relative velocities can be used \textcolor{cyan}{as a different probe to measure the equation of state of Dark Energy and test cosmological models}. In particular, the analysis of the multivariate pairwise velocity distribution indicates that observations of low-mass high-velocity interacting clusters (e.g. Bullet Cluster) as well as massive systems with moderate relative velocities are most sensitive to the underlying cosmology.
Focusing on the Bullet Cluster system, we have found a number of halo pairs candidates in $\Lambda$CDM-W7 and $w$CDM catalogs respectively, while we have found none in the RPCDM case within the simulated volume of the observable universe. Built upon these results we have quantified the probability of observing the Bullet Cluster in the context of Extreme Value Statistics. To this end we have used Generalized Pareto distributions to model the probability distribution of the pairwise velocities. We find the probability of observing a halo pair with average mass $>$10$^{14}$ h$^{-1}$ M$_\odot$, distance separation \mbox{$d_{12} < 10$ h$^{-1}$ Mpc} and relative velocity $>3000$ km s$^{-1}$ to strongly vary across the DEUS-FUR simulated cosmologies with probabilities of $9.7\times 10^{-8}$, $6.4\times 10^{-6}$ $1.7\times 10^{-5}$ for RPCDM, $\Lambda$CDM-W7 and $w$CDM respectively. Thus, we can deduce that the observation of the Bullet Cluster strongly disfavours cosmologies with low value of the $\sigma_8$ (low level of matter clustering). In the case of Dark Energy models calibrated against CMB observations this occurs for $w>-1$ since CMB data enforce low $\sigma_8$ values primarily to compensate for the enhanced amplitude of the ISW effect. In contrast, the probability of the Bullet Cluster increases for models with larger $\sigma_8$ and thus a higher level of matter clustering.
The study presented here suggests that observations of extreme interactive clusters sampling the tail of the pairwise velocity distribution can provide complementary information on Dark Energy models potentially capable of breaking standard cosmological parameter degeneracies. A first step in this direction will be the development of an accurate theoretical model of the pairwise velocity distribution density function along the line of the original work by \citep{Sheth1996,DiaferioSheth2001}. This will help elucidating the cosmological dependence of the high-velocity tail and provide an estimate of the number of bullet-like systems in the sky that need to be observed to improve current parameter constraints on Dark Energy.
\section*{Acknowledgements}
We warmly thank Vincent Reverdy and Ir\`ene Balmes for fruitful discussions about EVS and their deep implication in the achievement of the DEUS-FUR project. This work was granted access to HPC resources of TGCC through allocations made by GENCI (Grand \'Equipement National de Calcul Intensif) in the framework of the \enquote{Grand Challenge} DEUS. The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n. 279954. We acknowledge support from the DIM ACAV of the Region Ile de France. V.~R. Bouillot is supported financially (AW) by the National Research Foundation of South Africa. Any opinion, findings and conclusions or recommendations expressed in this material are those of the authors and therefore the NRF does not accept any liability in regard thereto.
|
\section{Introduction}
Let $v(n)$ denote the binary sum-of-digits residue function, i.e. the sum of the digits in the binary expansion of $n$ modulo 2. For example, $v(7)=v(111_2)=3$ mod 2 = 1. Then it is well known that $v(n)$ defines the classical Prouhet-Thue-Morse (PTM) integer sequence, which can easily be shown to satisfy the recurrence
\begin{align*}
v(0) &=0 \\
v(2n) &=v(n) \\
v(2n+1) &=1-v(n)
\end{align*}
The first few terms of $v(n)$ are $0,1,1,0,1,0,0,1$. Observe that the PTM sequence can also be generated by starting with the value 0 and recursively appending a negated copy of itself (bitwise):
\[0 \rightarrow 01 \rightarrow 0110 \rightarrow 01101001 \rightarrow ...
\]
Another method is to iterate the morphism $\mu$ defined on the alphabet $\{0,1\}$ by the substitution rules $\mu(0)= 01$ and $\mu(1)= 10$ and applied to $x_0=0$ as described in \cite{AS}:
\begin{align*}
x_1 & =\mu(x_0)=01 \\
x_2 & = \mu^2(x_0)=\mu(x_1)=0110 \\
x_3 & = \mu^3(x_0)=\mu(x_2)=01101001 \\
...
\end{align*}
This {\em ubiquitous} sequence, coined as such by Allouche and Shallit in \cite{AS}, first arose in the works of three mathematicians: E. Prouhet involving equal sums of like powers in 1851 (\cite{P}), A. Thue on combinatorics of words in 1906 (\cite{T}), and M. Morse in differential geometry in 1921 (\cite{Mo}). It has found interesting applications in many areas of mathematics, physics, and engineering: combinatorial game theory (\cite{AS},\cite{P}), fractals (\cite{ASk},\cite{MH}), quasicrystals (\cite{MM},\cite{RSL}) and more recently Dopper tolerant waveforms in radar (\cite{CPH},\cite{PCMH},\cite{NC}).
Suppose we now replace the 0's and 1's in the PTM sequence with 1's and $-1$'s, respectively. Then it is easy to prove that this yields an equivalent binary $\pm 1$-sequence $w(n)$ satisfying the recurrence
\begin{align*}
w(0) &=1 \\
w(2n) &=w(n) \\
w(2n+1) &=-w(n),
\end{align*}
where $w(n)$ and $v(n)$ are related by
\begin{equation} \label{eq:wv}
w(n)=1-2v(n),
\end{equation}
or equivalently,
\begin{equation} \label{eq:wv2}
w(n)=(-1)^{v(n)}.
\end{equation}
Of course, $v(n)$ can be generalized to any modulus $p\geq 2$. Towards this end, we define $v_p(n)$ to be the sum of the digits in the base-$p$ expansion of $n$ modulo $p$. We shall call $v_p(n)$ the mod-$p$ PTM integer sequence. Then $v_p(n)$ satisfies the recurrence
\begin{align*}
v_p(0) &=0 \\
v_p(pn+r) &=(v(n)+r)_p
\end{align*}
where $(m)_p \equiv m$ mod $p$. More interestingly, it is well known that $v_p(n)$ provides a solution to the famous Prouhet-Tarry-Escott (PTS) problem (\cite{P},\cite{L},\cite{W}): given a positive integer $M$, find $p$ mutually disjoint sets of non-negative integers $S_0$, $S_1$,...,$S_{p-1}$ so that
\[
\sum_{n\in S_0} n^m = \sum_{n\in S_1} n^m = ... = \sum_{n\in S_{p-1}} n^m
\]
for $m=1,...,M$. The solution, first given by Prouhet \cite{P} and later proven by Lehmer \cite{L} (see also Wright \cite{W}), is to partition the integers $\{0,1,...,p^{M+1}-1\}$ so that $n \in S_{v_p(n)}$. For example, if $M=3$ and $p=2$, then
the two sets $S_0=\{0,3,5,6,9,10,12,15\}$ and $S_1=\{1,2,4,7,8,11,13,14\}$ defined by Prouhet's algorithm solve the PTS problem:
\begin{align*}
60 & = 0+3+5+6+9+10+12+15 \\
& = 1+2+4+7+8+11+13+14 \\
620 & = 0^2+3^2+5^2+6^2+9^2+10^2+12^2+15^2 \\
& = 1^2+2^2+4^2+7^2+8^2+11^2+13^2+14^2 \\
7200 & = 0^3+3^3+5^3+6^3+9^3+10^3+12^3+15^3 \\
& = 1^3+2^3+4^3+7^3+8^3+11^3+13^3+14^3
\end{align*}
In this paper, we address the following question: what is the natural generalization of $w(n)$ to modulus $p\geq 2$? Which formula should we look to extend, (\ref{eq:wv}) or (\ref{eq:wv2})? Is there any intuition behind our generalization? One answer is to define $w_p(n)$ by merely replacing $v(n)$ with $v_p(n)$ in say (\ref{eq:wv2}). However, to discover a more satisfying answer, we consider a modified form of (\ref{eq:wv2}):
\begin{equation} \label{eq:wv-digit}
w(n)=(-1)^{d_{1-v(n)}}
\end{equation}
Here, $d_{1-v(n)}$ takes on one of two possible values, $d_0=1$ or $d_1=0$, which we view as the first two digits in the binary expansion (base 2) of the number 1, i.e. $1=d_12^1+d_02^0$. Thus, formula (\ref{eq:wv-digit}) involves the digit opposite in position to $v(n)$.
To explain how this formula naturally generalizes to any positive modulus $p\geq 2$, we begin our story with two arbitrary elements $a_0$ and $a_1$. Define $A=(a_n)=(a_0,a_1,...)$ to be what we call a mod-2 PTM sequence generated from $a_0$ and $a_1$, where the elements of $A$ satisfy the aperiodic condition
\[
a_n=a_{v(n)}
\]
Thus, $A=(a_0,a_1,a_1,a_0,a_1,a_0,a_0,a_1,...)$. Since formula (\ref{eq:wv-digit}) holds, it follows that $a_n$ can be decomposed as
\begin{equation} \label{eq:decomposition-mod2}
a_n=\frac{1}{2}(a_0+a_1)+\frac{1}{2}w(n)(a_0-a_1)
\end{equation}
In some sense, $w(n)$ plays the same role as $v(n)$ in defining the sequence $A$, but through the decomposition (\ref{eq:decomposition-mod2}).
We argue that formula (\ref{eq:decomposition-mod2}) leads to a natural generalization of $w(n)$. For example, suppose $p=3$ and
consider the mod-3 PTM sequence $A=(a_0,a_1,a_2,...)$ generated by three elements $a_0,a_1,a_2$ so that $a_n=a_{v_3(n)}$. The following decomposition generalizes (\ref{eq:decomposition-mod2}):
\begin{align*}
a_n & =\frac{1}{4}w_0(n)(a_0+a_1+a_2)+\frac{1}{4}w_1(n)(a_0+a_1-a_2) \\
& \ \ \ \ +\frac{1}{4}w_2(n)(a_0-a_1+a_2)+\frac{1}{4}w_3(n)(a_0-a_1-a_2)
\end{align*}
Here, $w_0(n),w_1(n),w_2(n),w_3(n)$ are $\pm 1$-sequences that we shall call the weights of $a_n$. Since $a_n=a_{v_3(n)}$, these weights are fully specified once their values are known for $n=0,1,2$. It is straightforward to verify in this case that $W(n)=(w_0(n),...,w_3(n))$ takes on the values
\begin{align*}
W(0) & =(1,1,1,1) \\
W(1) & =(1,1,-1,-1) \\
W(2) & =(1,-1,1,-1)
\end{align*}
Thus, the weights $w_i(n)$ are a natural generalization of $w(n)$.
More generally, if $p\geq 2$ is a positive integer and $A=(a_n)$ is a mod-$p$ PTM sequence generated from $a_0,a_1,...,a_{p-1}$, i.e. $a_n=a_{v_p(n)}$, then the following decomposition holds:
\begin{equation} \label{eq:aB}
a_n=\frac{1}{2^{p-1}} \sum_{i=0}^{2^{p-1}-1} w_i^{(p)}(n) B_i
\end{equation}
Here, the weights $w_i^{(p)}$ are given by
\begin{equation} \label{eq:wi}
w_i(n):=w_i^{(p)}(n)=(-1)^{d^{(i)}_{p-1-v_p(n)}}
\end{equation}
for $0\leq i \leq 2^{p-1}-1$ and $i=d_{p-2}^{(i)}2^{p-2}+...+d_1^{(i)} 2^1+d_0^{(i)}2^0$ is its binary expansion. Moreover, $B_i$ is calculated by the formula
\begin{equation} \label{eq:B-Rademacher}
B_i=\sum_{n=0}^{p-1}w_i(n) a_{n}
\end{equation}
Observe that we can extend the range for $i$ to $2^p-1$ (and will do so), effectively doubling the number of weights by expanding $i$.
In that case we find that
\[
w_i(n)=-w_{2^p-1-i}(n).
\]
With this extension, we demonstrate in Theorem \ref{th:recurrence} that each $w_i(n)$ satisfies the recurrence
\[
w_i(pn+r)=w_{x_r(i)}(n)w_i(n)
\]
where $x_r(i)$ denotes a quantity that we define in Section 4 as the `xor-shift' of $i$ by $r$, where $0\leq x_r(i) \leq 2^p-1$. For example, if $p=2$, we find that
\begin{align*}
w_1(2n) &=w_0(n)w_1(n) \\
w_1(2n+1) &=w_3(n)w_1(n)
\end{align*}
Since $w_0(n)=1$ and $w_3(n)=-1$ for all $n$, this yields the same recurrence satisfied by $w(n)=w_1(n)$ as described in the beginning of this section.
Next, we note that the set of values $R(n)=(w_0(n),...,w_{2^p-1}(n))$ represent those given by the Rademacher functions $\phi_n(x)$, $n=0,1,2,...$, defined by (see \cite{R}, \cite{F})
\begin{align*}
\phi_0(x) & =
1 \ \ (0 \leq x < 1/2) & \phi_0(x+1) & =\phi_0(x) \\
\phi_0(x) & =-1 \ \ (1/2 \leq x < 1) &
\phi_n(x) & =\phi_0(2^kx)
\end{align*}
In particular,
\[w_i(n)=\phi_n(i/2^p)
\]
so that the right-hand side of (\ref{eq:aB}) can be thought of as a discrete {\em Rademacher} transform of $(B_0,B_1,...,B_{2^{p-1}-1})$. Moreover, formula (\ref{eq:B-Rademacher}) can be viewed as the inverse transform, which follows from the fact that the Rademacher functions form an orthogonal set. Thus, weight sequences can be viewed as a mixing of Prouhet-Thue-Morse sequences and Rademacher functions.
It is known that the Rademacher functions generate the Walsh functions, which have important applications in communications and coding theory (\cite{B},\cite{Tz}). Walsh functions are those of the form (see \cite{F}, \cite{Wa})
\[
\psi_m(x)=\phi_{n_k}(x)\phi_{n_{k-1}}(x)\cdot \cdot \cdot \phi_{n_1}(x)
\]
where $m=2^{n_k} + 2^{n_{k-1}}+...+2^{n_1}$ with $n_i < n_{i+1}$ and $0\leq m \leq 2^p-1$. This allows us to generalize our weights $w_i(n)$ to sequences
\[
\tilde{w}_i(m)=w_i(n_k)\cdot \cdot \cdot w_i(n_1)
\]
which we view as a discrete version of the Walsh functions in the variable $i$. In that case, we prove in Section 3 that
\[
\sum_{i=0}^{2^p-1} \tilde{w}_i(m)B_i=\begin{cases}
a_n, & \text{if } m=2^n, 0\leq n \leq p-1 \\
0, & \text{otherwise}
\end{cases}
\]
We also prove in the same section a result that was used in \cite{NC} to characterize these weight sequences as sidelobes of Doppler tolerant radar waveforms (motivated by \cite{CPH} and \cite{PCMH}).
\section{The Prouhet-Thue-Morse Sequence}
Denote by $S(L)$ to be the set consisting of the first $L$ non-negative integers $0,1,...,L-1$.
\begin{definition}
Let $n=n_1n_2...n_k$ be the base-$p$ representation of a non-negative integer $n$. We define the {\em mod-$p$ sum-of-digits function} $v_p(n) \in \mathbb{Z}_p$ to be the sum of the digits $n_i$ modulo $p$, i.e.
\[
v_p(n)\equiv \sum_{i=1}^k n_i \mod p
\]
\end{definition}
\noindent Observe that $v_p(n)=n$ if $0\leq n < p$.
\begin{definition}
We define a sequence $A=(a_0,a_1,...)$ to be a {\em mod-$p$ Prouhet-Thue-Morse (PTM)} sequence if it satisfies the aperiodic condition
\[
a_n=a_{v_p(n)}
\]
\end{definition}
\begin{definition} Let $p$ and $M$ be positive integers and set $L=p^{M+1}$. We define $\{S_0,S_1,...,S_{p-1}\}$ to be a {\em Prouhet-Thue-Morse} (PTM) $p$-{\em block partition} of $S(L)=\{0,1,...,L-1\}$ as follows: if $v_p(n)=i$, then
\[
n\in S_i
\]
\end{definition}
The next theorem solves the famous Prouhet-Tarry-Escott problem.
\begin{theorem}[\cite{P}, \cite{L}, \cite{W}] \label{th:gptm-sequence}
Let $p$ and $M$ be positive integers and set $L=p^{M+1}$. Suppose $\{S_0,S_1,...,S_{p-1}\}$ is a PTM $p$-block partition of $S(L)=\{0,1,...,L-1\}$. Then
\[
P_m:=\sum_{n\in S_0} n^m = \sum_{n\in S_1} n^m = ... = \sum_{n\in S_{p-1}} n^m
\]
for $m=1,...,M$. We shall refer to $P_m$ as the $m$-th Prouhet sum corresponding to $p$ and $M$.
\end{theorem}
\begin{corollary}
Let $A=(a_0,a_1,...,a_{L-1})$ be a mod-$p$ PTM sequence of length $L=p^{M+1}$, where $M$ is a non-negative integer. Then
\begin{equation} \label{eq:sum-An}
\sum_{n=0}^{L-1}n^m a_n = P_m (a_0+a_1+...+a_{p-1})
\end{equation}
for $m=0,...,M$.
\end{corollary}
\begin{proof} We have
\begin{align*}
\sum_{n=0}^{L-1}n^m a_n & = \sum_{n\in S_0} n^m a_{v_p(n)} + \sum_{n\in S_1} n^m a_{v_p(n)} + ...+ \sum_{n\in S_{p-1}} n^m a_{v_p(n)} \\
& = a_0\sum_{n\in S_0} n^m + a_1\sum_{n\in S_1} n^m + ...+ a_{p-1} \sum_{n\in S_{p-1}} n^m \\
& = P_m (a_0+a_1+...+a_{p-1})
\end{align*}
\end{proof}
\section{Weight Sequences}
In this section we develop a generalization of the PTM $\pm1$-sequence $w(n)$ and derive orthogonality and recurrence relations for these generalized sequences that we refer to as {\em weight} sequences.
\begin{definition}
Let $i=d_{p-1}^{(i)}2^{p-1}+d_{p-2}^{(i)}2^{p-2}+...+d_1^{(i)} 2^1+d_0^{(i)}2^0$ be the binary expansion of $i$, where $i$ is a non-negative integer with $0\leq i \leq 2^p-1$. Define $w_0(n),w_1(n),...w_{2^p-1}(n)$ be binary $\pm 1$-sequences defined by
\[
w_i(n)=(-1)^{d_{p-1-v_p(n)}^{(i)}}
\]
\end{definition}
\begin{example} Let $p=3$. Then
\begin{align*}
w_0(n) & = (\mathbf{1, 1, 1}, 1, 1, 1, 1, 1, 1,...) \\
w_1(n) & = (\mathbf{1, 1, -1}, 1, -1, 1, -1, 1, 1,...) \\
w_2(n) & = (\mathbf{1, -1, 1}, -1, 1, 1, 1, 1, -1,...) \\
w_3(n) & = (\mathbf{1, -1, -1}, -1, -1, 1, -1, 1, -1,...) \\
w_4(n) & = (\mathbf{-1, 1, 1}, 1, 1, -1, 1, -1, 1,...) \\
w_5(n) & = (\mathbf{-1, 1, -1}, 1, -1, -1, -1, -1, 1,...) \\
w_6(n) & = (\mathbf{-1, -1, 1}, -1, 1, -1, 1, -1, -1,...) \\
w_7(n) & = (\mathbf{-1, -1, -1}, -1, -1, -1, -1, -1, -1,...)
\end{align*}
\end{example}
\noindent Observe that the first three values of each weight$w_i(n)$ (displayed in bold) represent the binary value of $i$ if we replace 1 and $-1$ with 0 and 1, respectively. Morever, we have the following symmetry:
\begin{lemma} \label{le:w-reverse}
For $i=0,1,...,2^p-1$, we have
\[
w_i(n)=-w_{2^p-1-i}(n)
\]
\end{lemma}
\begin{proof} If $i=d_{p-1}^{(i)}2^{p-1}+d_{p-2}^{(i)}2^{p-2}+...+d_0^{(i)}2^0$, then $j=2^p-1-i$ has expansion
\[
j=\bar{d}_{p-1}^{(j)}2^{p-1}+\bar{d}_{p-2}^{(j)}2^{p-2}+...+\bar{d}_0^{(j)}2^0
\]
where $\bar{d}_k^{(j)}=1-d_k^{(i)}$. It follows that
\[
w_i(n)=(-1)^{d^{(i)}_{p-1-v_p(n)}} = (-1)^{1-d^{(j)}_{p-1-v_p(n)}}=-w_{2^p-1-i}(n)
\]
\end{proof}
\begin{theorem} \label{th:W-orthogonal}
Let $p\geq 2$ be a positive integer. Then the vectors $W_p(0),W_p(1),..,W_p(p-1)$ defined by
\[
W_p(n)=(w_0^{(p)}(n),w_1^{(p)}(n),...,w_{2^{p-1}-1}^{(p)}(n))
\]
form an orthogonal set, i.e.
\[
W_p(n)\cdot W_p(m) = \sum_{i=0}^{2^{p-1}-1} w_i(n)w_i(m)= 2^{p-1} \delta_{n-m} =
\begin{cases}
2^{p-1}, & n = m \\
0, & n\neq m
\end{cases}
\]
for $0\leq n,m \leq p-1$. Here, $\delta_n$ is the Kronecker delta function.
\end{theorem}
\begin{proof} It is straightforward to check that the lemma is true for $p=2$. Thus, we assume $p\geq 3$ and define $k(n)=p-1-n$ so that
\[
W_p(n)\cdot W_p(m)= \sum_{i=0}^{2^{p-1}-1} (-1)^{d^{(i)}_{k(n)}+d^{(i)}_{k(m)}}
\]
Assume $n\neq m $ and without loss of generality, take $n<m$ so that $k(n)>k(m)$. Assume $0\leq i \leq 2^{p-1}-1$ and expand $i$ in binary so that
\[
i=d^{(i)}_{p-1}2^{p-1}+...+d^{(i)}_{k(n)}2^{k(n)}+...+d^{(i)}_{k(m)}2^{k(m)}+...+d^{(i)}_02^0
\]
where $d_{p-1}^{(i)}=0$. Suppose in specifying $i$ we fix the choice of values for all binary digits except for $d^{(i)}_{k(n)}$ and $d^{(i)}_{k(m)}$. Then the set $S=\{(0,0), (0,1),(1,0), (1,1)\}$ consists of the four possibilities for choosing these two remaining digits, which we express as the ordered pair $d=(d^{(i)}_{k(n)},d^{(i)}_{k(m)})$. But then the contribution from this set of four such values for $i$ sums to zero in the dot product $W_p(n)\cdot W_p(m)$, namely
\[
\sum_{d\in S}(-1)^{d^{(i)}_{k(n)}+d^{(i)}_{k(m)}}=0
\]
Since this holds for all cases in specifying $i$, it follows that $W_p(n)\cdot W_p(m)=0$ as desired. On the other hand, if $n=m$, then $k(n)=k(m)$ and so $d^{(i)}_{k(n)}=d^{(i)}_{k(m)}$ for all $i$. It follows that
\[
W_p(n)\cdot W_p(m)=\sum_{i=0}^{2^{p-1}-1} (-1)^{2d^{(i)}_{k(n)}}=\sum_{i=0}^{2^{p-1}-1}1=2^{p-1}
\]
\end{proof}
In fact, we have the more general result, which states a discrete version of the fact that the Walsh functions form an orthogonal set.
\begin{theorem} \label{th:w-tilde}
Let $m$ be an integer and expand $m=2^{n_k} + 2^{n_{k-1}}+...+2^{n_1}$ in binary with $n_i < n_{i+1}$ and $0\leq m \leq 2^p-1$. Define
\[
\tilde{w}_i(m)=w_i(n_k)\cdot \cdot \cdot w_i(n_1)
\]
for $i=0,1,...,2^p-1$. Then
\begin{equation} \label{eq:w-tilde}
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m)=0
\end{equation}
for all $m=0,1,....,2^p-1$.
\end{theorem}
\begin{proof} Let $m=2^{n_k} + 2^{n_{k-1}}+...+2^{n_1}$. We argue by induction on $k$, i.e. the number of distinct powers of $2$ in the binary expansion of $m$. Suppose $k=1$ and define $q=p-1-v_p(n_1)$. Then given any value of $i$ where the binary digit $d_{q}^{(i)}=0$, there exists a corresponding value $j$ whose binary digit $d_{q}^{(j)}=1$. It follows that
\begin{align*}
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m) & =\sum_{\substack {i=0 \\ d_{q}^{(i)} =0}}^{2^{p-1}-1} (-1)^{d_{q}^{(i)}} +\sum_{\substack {i=0 \\ d_{q}^{(i)} =1}} ^{2^{p-1}-1} (-1)^{d_{q}^{(i)}} \\
& = 2^{p-2}-2^{p-2} = 0
\end{align*}
Next, assume that (\ref{eq:w-tilde}) holds for all $m$ with $k-1$ distinct powers of 2. Define $q_k=p-1-v_p(n_k)$. Then for $m$ with $k$ distinct powers of 2, we have
\begin{align*}
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m) & = \sum_{i=0}^{2^{p-1}-1}(-1)^{d_{q_k}^{(i)}+d_{q_{k-1}}^{(i)}+...+d_{q_1}^{(i)}} \\
& = (-1)^0\sum_{\substack {i=0 \\ d_{q_k}^{(i)} =0}}^{2^{p-1}-1} (-1)^{d_{q_{k-1}}^{(i)}+...+d_{q_1}^{(i)}} +(-1)^1 \sum_{\substack {i=0 \\ d_{q_k}^{(i)} =1}} ^{2^{p-1}-1} (-1)^{d_{q_k}^{(i)}+d_{q_{k-1}}^{(i)}+...+d_{q_1}^{(i)}} \\
& = \frac{1}{2}\sum_{i=0}^{2^{p-1}-1} (-1)^{d_{q_{k-1}}^{(i)}+...+d_{q_1}^{(i)}} -\frac{1}{2} \sum_{i=0} ^{2^{p-1}-1} (-1)^{d_{q_k}^{(i)}+d_{q_{k-1}}^{(i)}+...+d_{q_1}^{(i)}} \\
& = \frac{1}{2}\cdot 0 - \frac{1}{2} \cdot 0 = 0
\end{align*}
\end{proof}
In \cite{Ri}, Richman observed that the classical PTM sequence $v(i)$ (although he did not recognize it by name in his paper) can be constructed from the product of all Radamacher functions up to order $p-1$, where $0 \leq i \leq 2^p-1$. This result easily follows from our formulation of weight sequences since
\begin{align*}
\tilde{w}_i^{(p)}(2^{p}-1) & =w_i^{(p)}(0)w_i^{(p)}(1)\cdot \cdot \cdot w_i^{(p)}(p-1) \\
& = (-1)^{d_{p-1}^{(i)}+d_{p-2}^{(i)}+...+d_{0}^{(i)}} \\
& = v(i)
\end{align*}
Next, we relate weight sequences with PTM sequences. Since $w_i(n)=-w_{p-1-i}(n)$ from Lemma \ref{le:w-reverse}, the following lemma is immediate.
\begin{lemma} \label{le:weights} Let $A=(a_0,a_1,...)$ be a mod-$p$ PTM sequence. Define
\[
B_i=\sum_{n=0}^{p-1}w_i(n)a_n
\]
for $i=0,1,...,2^{p}-1$. Then
\[
B_i(n)=-B_{2^p-1-i}(n)
\]
\end{lemma}
\begin{theorem} The following equation holds for all $n\in \mathbb{N}$:
\begin{equation} \label{eq:weights}
a_n=\frac{1}{2^{p-1}} \sum_{i=0}^{2^{p-1}-1} w_i(n) B_i
\end{equation}
\end{theorem}
\begin{proof} Since $a_n=a_{v(n)}$ for a PTM sequence, it suffices to prove (\ref{eq:weights}) for $n=0,1,...,p-1$. It follow from Theorem \ref{th:W-orthogonal} that
\begin{align*}
\frac{1}{2^{p-1}} \sum_{i=0}^{2^{p-1}-1} w_i(n) B_i
& = \frac{1}{2^{p-1}} \sum_{i=0}^{2^{p-1}-1} w_i(n) \left( \sum_{m=0}^{p-1}w_i(m)a_m \right) \\
& = \frac{1}{2^{p-1}}\sum_{m=0}^{p-1} \left( \sum_{i=0}^{2^{p-1}-1} w_i(n)w_i(m) \right) a_{m} \\
& = \frac{1}{2^{p-1}}\sum_{m=0}^{p-1} 2^{p-1}\delta_{n-m} a_{m} \\
& = a_n
\end{align*}
\end{proof}
\noindent NOTE: Because of the lemma above, we shall refer to $w_0(n),w_1(n),...,w_{2^{p-1}-1}(n)$ as the PTM weights of $a_n$ with respect to the basis of sums $(B_0,B_1,...,B_{2^{p-1}-1})$.
\begin{example} \
\noindent 1. $p=2$:
\begin{align*}
B_0 & = a_0+a_1 \\
B_1 & = a_0-a_1
\end{align*}
\noindent 2. $p=3$:
\begin{align*}
B_0 & = a_0+a_1+a_2 \\
B_1 & = a_0+a_1-a_2 \\
B_2 & = a_0-a_1+a_2 \\
B_3 & = a_0-a_1-a_2
\end{align*}
\end{example}
\begin{theorem} For $0\leq m \leq 2^p-1$, we have
\begin{equation} \label{eq:w-tilde-B}
\sum_{i=0}^{2^p-1} \tilde{w}_i(m)B_i=\begin{cases}
a_n, & \text{if } m=2^n, 0\leq n \leq p-1 \\
0, & \text{otherwise}
\end{cases}
\end{equation}
\end{theorem}
\begin{proof} If $m=2^n$, then $\tilde{w}_i(n)=w_i(n)$ and thus formula (\ref{eq:w-tilde-B}) reduces to (\ref{eq:weights}). Therefore, assume $m=2^{n_k}+...+2^{n_1}$ where $k>1$. Define $S_m=\{0,1,...,p-1\}-\{n_1,n_2,...,n_k\}$. Then
\begin{align*}
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m) B_i
& =\sum_{i=0}^{2^{p-1}-1} w_i(n_k)\cdot \cdot \cdot w_i(n_1) \left( \sum_{j=0}^{p-1}w_i(j)a_j \right) \\
& =\sum_{j=0}^{p-1} \left( \sum_{i=0}^{2^{p-1}-1} w_i(n_k)\cdot \cdot \cdot w_i(n_1)w_i(j) \right) a_{j}
\end{align*}
Next, isolate the terms in the outer summation above corresponding to $S_m$:
\begin{align*}
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m) B_i & = a_{n_1} \sum_{i=0}^{2^{p-1}-1} w_i(n_k)\cdot \cdot \cdot w_i(n_2) w_i(n_1)^2+... \\
& \ \ \ \ + a_{n_k} \sum_{i=0}^{2^{p-1}-1} w_i(n_k)^2w_i(n_{k-1})
\cdot \cdot \cdot w_i(n_1) \\
& \ \ \ \ \ \ + \sum_{\substack {j\in S_m}} \left( \sum_{i=0}^{2^{p-1}-1} w_i(n_k)\cdot \cdot \cdot w_i(n_1)w_i(j) \right) a_{k} \\
& = a_{n_1} \sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m^-_1)+...+ a_{n_k} \sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m^-_{k}) \\
& \ \ \ \ + \sum_{\substack {j\in S_m}} \left( \sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m^+_j) \right) a_{k}
\end{align*}
where $m^-_j=m-2^j$ and $m^+_j=m+2^j$. Now observe that all three summations above with index $i$ must vanish because of Theorem \ref{th:w-tilde}. Hence,
\[
\sum_{i=0}^{2^{p-1}-1} \tilde{w}_i(m) B_i = 0
\]
as desired.
\end{proof}
We end this section by presenting a result that is useful in characterizing sidelobes of Doppler tolerant waveforms in radar (\cite{PCMH},\cite{CPH},\cite{NC}).
\begin{theorem} \label{th:aBS}
Let $A=(a_0,a_1,...,a_{L-1})$ be a mod-$p$ PTM sequence of length $L=p^{M+1}$, where $M$ is a non-negative integer. Write
\begin{equation} \label{eq:aBS}
a_n = \frac{1}{2^{p-1}} w_0(n)B_0+\frac{1}{2^{p-1}} S_p(n)
\end{equation}
where
\[
S_p(n)=\sum_{i=1}^{2^{p-1}-1} w_i(n)B_i
\]
Then
\begin{equation} \label{eq:aBSN}
\sum_{n=0}^{L-1}n^m S_p(n)=N_m(L) B_0
\end{equation}
for $m=1,...,M$ where
\[
N_m(L)=2^{p-1}P_m-\sum_{n=0}^{L-1}n^m
\]
\end{theorem}
\begin{proof} We apply (\ref{eq:sum-An}):
\begin{align*}
\sum_{n=0}^{L-1}n^m S_p(n) & = 2^{p-1} \sum_{n=0}^{L-1}n^m a_n - B_0\sum_{n=0}^{L-1}n^m w_0(n)\\
& = 2^{p-1}P_m (a_0 + a_1+...+a_{p-1}) - B_0 \sum_{n=0}^{L-1}n^m \\
& = (2^{p-1}P_m-\sum_{n=0}^{L-1}n^m)B_0 \\
& = N_m(L) B_0
\end{align*}
\end{proof}
\section{XOR-Shift Recurrence}
In this section we develop a recurrence formula for our weight sequences. Towards this end, we introduce the notion of an {\em xor-shift} of a binary integer.
\begin{definition}
Let $a,b\in \mathbb{Z}_2$. We define $a\oplus b$ to be the exclusive OR (XOR) operation given by the following Boolean truth table:
\begin{align*}
0\oplus 0 & = 0 \\
0\oplus 1 & = 1 \\
1\oplus 0 & = 1 \\
1\oplus 1 & = 0
\end{align*}
More generally, let $x=a_k...a_0$ and $y=b_k...b_0$ be two non-negative integers expressed in binary. We define $z=x\oplus y=c_k..c_0$ to be the {\em xor bit-sum} of $x$ and $y$, where
\[
c_k=a_k\oplus b_k
\]
\end{definition}
\begin{definition}
We shall say that two integers $a$ and $b$ are congruent modulo 2 and write $a\cong b$ to mean $a = b$ mod 2.
\end{definition}
The following lemma, which is straightforward to prove, will be useful to us.
\begin{lemma} Let $a,b \in \mathbb{Z}$. Then
\[
a\pm b\cong a\oplus b
\]
\end{lemma}
\begin{definition} Let $p$ be a positive integer and $i$ a non-negative integer with $0\leq i \leq 2^{p}-1$. Expand $i$ in binary so that
\[
i=d_{p-1}2^{p-1}+...+d_02^0
\]
We define the {\em degree-$p$ xor-shift} of $i$ by $r\geq 0$ to be the decimal value given by the xor bit-sum
\[
x_r(i):=x_r^{(p)}(i)=d_{p-1}...d_rd_{r-1}...d_0 \oplus d_{p-1-r}...d_0d_{p-1}...d_{p-r}
\]
i.e.
\[
x_r(i) = e_{p-1}2^{p-1}+...+e_02^0
\]
where
\[
e_k=
\begin{cases}
d_k\oplus d_{k-r}, & k\geq r \\
d_k\oplus d_{d+(p-r)}, & k < r
\end{cases}
\]
for $k=0,1,...,p-1$.
\end{definition}
\begin{example}
\noindent Here are some values of $x_i^{(p)}(n)$ for $p=3$:
\begin{align*}
x_1^{(3)}(0) & = 000_2\oplus 000_2 = 000_2 = 0, & x_2^{(3)}(0) & =000_2\oplus 000_2 = 000_2 = 0 \\
x_1^{(3)}(1) & =001_2\oplus 010_2 = 011_2 = 3, & x_2^{(3)}(1) & =001_2\oplus 100_2 = 101_2 = 5 \\
x_1^{(3)}(2) & =010_2\oplus 100_2 = 110_2 = 6, & x_2^{(3)}(2) & =010_2\oplus 001_2 = 011_2 = 3 \\
x_1^{(3)}(3) & =011_2\oplus 110_2 = 101_2 = 5, & x_2^{(3)}(3) & =011_2\oplus 101_2 = 110_2 = 6
\end{align*}
In fact, when $n=p-1$, the sequence
\[
x_1^{(n+1)}(n) = (0, 3, 6, 5, 12, 15, 10, 9, 24, 27,...)
\]
generates the xor bit-sum of $n$ and $2n$ (sequence A048724 in the Online Encyclopedia of Integer Sequences (OEIS) database: http://oeis.org).
\end{example}
\begin{lemma}
Define
\begin{align*}
E_p(i,n):= d_{p-1-v_p(n)}^{(i)}
\end{align*}
so that $w_i(n)=(-1)^{E_p(i,n)}$. Then for $0\leq r < p$, we have
\[
E_p(i,pn+r) =
\begin{cases}
d_{p-1-v_p(n)-r}, & \text{if } v_p(n)+r < p\\
d_{p-1-s}, & \text{if } v_p(n)+r\geq p
\end{cases}
\]
where $s=v_p(n)+r-p$. Moreover,
\begin{equation} \label{eq:Ep}
E_p(i,pn+r)-E_p(i,n)\cong E_p(x_r(i),n)
\end{equation}
\end{lemma}
\begin{proof} Since $v_p(pn+r)=(v_p(n)+r)_p$, we have
\[
E_p(i,pn+r) = d_{p-1-(v_p(n)+r)_p}
\]
Now consider two cases: either $v(n)+r < p$ or $v(n)+p\geq p$. If $v(n)+r < p$, then
\[
E_p(i,pn+r) = d_{p-1-v_p(n)-r}
\]
On the other hand, if $v(n)+r\geq p$, then set $s=v_p(n)+r-p$ so that $(v_p(n)+r)_p=s$. It follows that
\[
E_p(i,pn+r) = d_{p-1-s}
\]
To prove (\ref{eq:Ep}), we again consider two cases. First, assume $v_p(n)+r < p$ so that $p-1-v_p(n)\geq r$. Then
\begin{align*}
E_p(i,pn+r)-E_p(i,n) & = d_{p-1-v_p(n)-r} - d_{p-1-v_p(n)} \\
& \cong d_{p-1-v_p(n)}\oplus d_{p-1-v_p(n)-r} \\
& \cong E_p(x_r(i),n)
\end{align*}
On the other hand, if $v_p(n)+r\geq p$, then set $s=v_p(n)+r-p$ so that $(v_p(n)+r)_p=s$. Since $p-1-v_p(n)<r$, we have
\begin{align*}
E_p(i,pn+r)-E_p(i,n) & = d_{p-1-s} - d_{p-1-v_p(n)} \\
& \cong d_{p-1-v_p(n)}\oplus d_{p-1-s} \\
& \cong d_{p-1-v_p(n)}\oplus d_{p-1-v_p(n)+(p-r)} \\
& \cong E_p(x_r(i),n)
\end{align*}
\end{proof}
\begin{theorem} \label{th:recurrence}
Let $p$ be a positive integer. The weight sequences $w_i(n)$, $0\leq i \leq 2^p-1$, satisfy the recurrence
\begin{equation}
w_i(pn+r)=w_{x_r(i)}(n)w_i(n)
\end{equation}
where $n\in \mathbb{N}$ and $r \in \mathbb{Z}_p$.
\end{theorem}
\begin{proof} The recurrence follows easily from formula (\ref{eq:Ep}):
\begin{align*}
\frac{w_i(pn+r)}{w_i(n)} & = (-1)^{E_p(i,pn+r)-E_p(i,n)} \\
& = (-1)^{E_p(x_r(i),n)} \\
&= w_{x_r(i)}(n)
\end{align*}
\end{proof}
\begin{example} Let $p=3$. Then $w_0(n)=1$ for all $n\in \mathbb{N}$ and the other weight sequences, $w_1(n)$, $w_2(n)$, $w_3(n)$, satisfy the following recurrences:
\begin{align*}
w_1(3n) = w_0(n)w_1(n), \ w_1(3n+1) = w_3(n)w_1(n), \ w_1(3n+2) = w_5(n)w_1(n) \\
w_2(3n) = w_0(n)w_2(n), \ w_2(3n+1) = w_6(n)w_2(n), \ w_2(3n+2) = w_3(n)w_2(n) \\
w_3(3n) = w_0(n)w_3(n), \ w_3(3n+1) = w_5(n)w_3(n), \ w_3(3n+2) = w_6(n)w_3(n)
\end{align*}
\end{example}
\section{Conclusion.}
In this paper we presented what appears to be a novel generalization of the Prouhet-Thue-Morse sequence to weight sequences by considering the Rademacher transform of a given set of elements. These weight sequences were shown to satisfy interesting recurrences and orthogonality relations. Moreover, they were used in \cite{NC} to describe sidelobes of Doppler tolerant waveforms to radar.
\vskip 5pt
\noindent {\em Acknowledgment.}
The authors wish to thank Greg Coxson (Naval Research Laboratory) for many useful discussions on radar and complementary code matrices.
|
\section{Introduction}
The most prominent and researched example of quantized spacetimes is the case of constant noncommutativity, the so called Moyal-Weyl plane.
Where did we get this plane from? Is it possible to derive it from anything we know, or is it just another experimentally non-realizable idea of theoretical physics? \\\\ From a mathematical point of view, one might argue that the Moyal-Weyl plane is just the coordinate generalization of the quantum-mechanical symplectic structure between the momentum and the coordinate. From a physical point of view and guided by simplicity, nontrivial commutation relations of the coordinates emerge from the fact that localization with high accuracy at Planck scale causes a gravitational collapse, \cite{DFR}. The commutation relations found are identified with those of the Moyal-Weyl. \\\\
Another {beautiful} physical point of view is the emergence of such a plane in the Landau-quantization. In this effect the physical constant describing noncommutativity of space is given by the inverse of a magnetic field, \cite{EZ, SZ}.\\\\ Nonetheless, a conceptual mathematical derivation of such a spacetime would lead to a deeper understanding of the physical nature of quantum-spacetimes. Such a concept was developed for the quantum-mechanical (QM) case in \cite{Muc1}. It was done in a rigorous mathematical fashion by using the deformation technique known under warped convolutions \cite{BLS}. In the QM-case, we were able to identify the quantum plane of the Landau quantization by taking the commutator of the deformed coordinate operator. The generators used for the deformation of the coordinate operator are the momentum operators. These techniques were also used rigorously in quantum measurement theory \cite{AA}.
\\\\ The idea developed in the QM-case is extended in this paper to a relativistic quantum field theoretical context. This is done by using the creation and annihilation operators of the free scalar field. Hence, we first define a quantum field theoretical version of an operator that is conjugate to the second-quantized momentum.
This is done by taking the QM-definition using the momentum operator and performing a second-quantization onto the bosonic Fock space. Moreover, the definition is extended to a temporal component and the deformed commutator of these operators is calculated. \\\\
Now we recognized with regards to the spatial part of the conjugate operator an equivalence to the Newton-Wigner-Pryce (NWP) operator. In the context of relativistic particles, the NWP operator is usually
mentioned as the rightful position or center-of-mass operator. To obtain this operator,
certain physical requirements where imposed on localized states, \cite{NW49}, while in
\cite{PR48} the same object was found by a relativistic generalization of the Newtonian definition of the
mass-center. However, localization in terms of the position operator is beset by problems regarding relativity. It is in clear conflict with relativistic covariance and causality in quantum physics. This fact can be verified by using the spectral theory of self-adjoint operators (see for example \cite[Theorem1]{Yng14}).\\\\
An exit strategy concerning localization is to shift the focus to localized fields, where the spacetime coordinates are common variables of such fields. In this context, it is worth mentioning that a quantum field deformed with the coordinate operator belongs to the class of wedge-local fields (\cite[Proposition 5.4]{Muc2}). Wedge-locality is a weaker form of the common point-wise locality and it seems the appropriate notion of causality in noncommutative QFT, \cite{GL1, Sol}. \\\\ Although the NW-localization violates relativistic covariance and causality, it holds at least in an approximate sense for distances of the order of the Compton wave length or smaller such as the Planck-length, \cite{Mund05} or \cite[Chapter 3c]{Sch}. And it is precisely around the Planck-length scale where noncommutative spacetimes are supposed to be non-negligible, \cite{LS, SZ}.
\\\\
Hence, following the QM-case the paper is organized as follows: We first review the most important definitions of warped convolutions. In Section \ref{s3}, we define the second quantized coordinate operator and prove its equivalence to the NWP-operator. Moreover, we show deformations of such unbounded operators are mathematically well-defined. By calculating the deformed commutator of the coordinate operators, we yield terms resembling relativistic corrections to the standard Moyal-Weyl.
\section{Preliminaries }\label{s2}
\subsection{Warped Convolutions}
In this section we write all basic definitions and lemmas of the deformation known under the name of warped convolutions. For proofs of the respective lemmas we refer the reader to the original works \cite{BLS, GL1}. \newline\newline
The authors start their investigation with a $C^{*}$-dynamical system $(\mathcal{A},\mathbb{R}^d)$, \cite{P97}.
It consists of a $C^{*}$-algebra $\mathcal{A}$ equipped with a strongly continuous automorphic action of the group $\mathbb{R}^d$ which will be denoted by $\alpha$.
Furthermore, let $\mathcal{B}(\mathscr{H})$ be the Hilbert space of bounded
operators on $\mathscr{H}$ and let the adjoint action $\alpha$ be implemented by the weakly continuous unitary representation $U$. Then, it is argued that since the unitary representation $U$ can be extended to the algebra $\mathcal{B}(\mathscr{H})$, there is no lost of generality when one proceeds to the $C^{*}$-dynamical system $(C^{*}\subset \mathcal{B}(\mathscr{H}),\mathbb{R}^d)$. Here $C^{*}\subset \mathcal{B}(\mathscr{H})$ is the $C^{*}$-algebra of all operators on which $\alpha$ acts strongly continuously.
\newline\newline
Hence, we start by assuming the
existence of a strongly continuous unitary group $U$ that is a representation of the additive
group $\mathbb{R}^{d}$, $d\geq2$, on some separable Hilbert space $\mathscr{H}$.
Moreover, to define the deformation of operators belonging to a $C^{*}$-algebra $C^{*}\subset \mathcal{B}(\mathscr{H})$, we consider
elements belonging to the subalgebra $C^{\infty}\subset C^{*} $. The subalgebra $C^{\infty}$ is
defined to be the $*-$algebra of smooth elements (in the norm topology) with respect to $\alpha$, which
is the adjoint action of a weakly continuous
unitary representation $U$ of $\mathbb{R}^{d}$ given by
\begin{equation*}
\alpha_{x}(A)=U(x)\,A\,U(x)^{-1}, \quad x \in \mathbb{R}^{d}.
\end{equation*}
Let $\mathcal{D}$ be the dense domain of vectors in $\mathscr{H}$ which transform
smoothly under the adjoint action of $U$.
Then, the warped convolutions for operators $A\in C^{\infty}$ are given by
the following definition.\\
\begin{definition}
Let $\theta$ be a real skew-symmetric matrix on $\mathbb{R}^{d}$ w.r.t. the Minkowski metric, let $A\in C^{\infty}$ and let $E$ be
the spectral resolution of the unitary operator $U$. Then, the corresponding \textbf{warped convolution} $A_{\theta}$
of $A$ is defined on the dense domain $\mathcal{D}$ according to
\begin{equation}\label{WC}
A_{\theta }:=\int \alpha_{\theta x}(A)dE(x),
\end{equation}
where $\alpha$ denotes the adjoint action of $U$ given by $\alpha_{k}(A)=U(k)\,A\,U(k)^{-1}$.
\end{definition}
The restriction in the choice of operators is owed to the fact that the deformation is performed
with operator valued integrals. Furthermore, one can represent the warped
convolution of $
{A} \in {C}^{\infty}$ by $\int \alpha_{\theta x}(A) dE(x)$, on the dense domain
$\mathcal{D}\subset\mathscr{H}$ of vectors smooth w.r.t. the action of $U$, in terms of
strong limits
\begin{equation*}
\int\alpha_{\theta x}(A) dE(x)\Phi=(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint dx\, dy \,\chi(\epsilon x,\epsilon y )\,e^{-ixy}\, U(y)\, \alpha_{\theta x}(A)\Phi,
\end{equation*}
where $\chi \in\mathscr{S}(\mathbb{R}^{d}\times\mathbb{R}^{d})$ with $\chi(0,0)=1$. In an intermediate step of showing this equivalence it was as well proven that $\mathcal{D}$ is \textbf{stable} under the deformation.
The latter representation makes calculations and proofs concerning the existence of integrals
easier.
During this work we make use of both representations. \\\\
In the following lemma we introduce the deformed product, known as the Rieffel product \cite{RI},
by using warped convolutions. The circumstance that the two are interrelated is due to
the fact that warped convolutions supply isometric
representations of Rieffel's strict deformations of $C^{*}$-dynamical systems with
actions of $\mathbb{R}^{d}$. The definition of the Rieffel product, given by warped convolutions, is used in the next chapter to calculate
the deformed commutators. \newline
\begin{lemma}\label{dpl}
Let $\theta$ be a real skew-symmetric matrix on $\mathbb{R}^{d}$ w.r.t. the Minkowski metric, $ {A}, {B} \in
C^{\infty}$ and let $\Phi\in\mathcal{D}$. Then, the product of two deformed operators $A, B$ is given by
\begin{equation*}
{A}_{\theta} {B}_{\theta} \Phi= (A\times_{\theta}B)_{\theta}\Phi,
\end{equation*}
where the \textbf{deformed product} $\times_{\theta}$ is the Rieffel product
on $ {C}^{\infty}$ defined as,
\begin{equation}\label{dp0}
(A\times_{\theta}B )\Phi=(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint dx\, dy\,\chi(\epsilon x,\epsilon y )e^{-ixy} \alpha_{\theta x}(A)\alpha_{y}(B)\Phi.
\end{equation}
\end{lemma}
Next, we shall give the definition of a deformed commutator using the deformed product.
\begin{definition}\label{dpl1}
Let $\theta$ be a real skew symmetric-matrix on $\mathbb{R}^{d}$, $ {A}, {B} \in
{C}^{\infty}$ and let $\Phi\in\mathcal{D}$. Then, the \textbf{deformed commutator} is defined as,
\begin{equation}\label{dp1}
[A\stackrel{\times_{\theta}}{,} B ]\Phi:= (A\times_{\theta}B-B\times_{\theta}A )\Phi.
\end{equation}
\end{definition}
The Rieffel product, i.e. an associative product, was used for the definition and hence the commutator has the standard properties. Furthermore, it is interesting to note that in \cite{BC}, a similar commutator was given and dubbed the Q-commutator. It differs from our current definition only by an additional term, which is the commutative product of the two operators $ {A}, {B} \in
{C}^{\infty}$.
\\\\ As a matter of fact, we intend to deform \textbf{unbounded operators} and hence we are obliged to prove that the deformation
formula, given as an oscillatory integral, is well-defined. This is done in Section \ref{s3}.
\subsection{ Bosonic Fock
space} The ($d=n+1 $)-dimensional relativistic bosonic Fock
space is defined in the following. A particle with momentum $\mathbf{p} \in \mathbb{R}^n$ has in the massless case the energy $\omega_{\mathbf{p}}=+\sqrt{\mathbf{p}^2}$. Moreover the Lorentz-invariant measure is given by $d^n\mu(\mathbf{p} )=d^n\mathbf{p}/( {2\omega_{\mathbf{p}}})$.
\begin{definition}\label{bf}
The \textbf{bosonic Fock space} $\mathscr{F^{+}({H})}$ is defined
as in \cite{BR}:
\begin{equation*}
\mathscr{F^{+}({H})}=\bigoplus_{k=0}^{\infty}\mathscr{H}_{k}^{+},
\end{equation*}
where $\mathscr{H}_{0}=\mathbb{C}$ and the symmetric $k$-particle subspaces are given as
\begin{align*}
\mathscr{H}_{k}^{+} =\{\Psi_{k}: \underbrace{\partial V^{+} \times \dots \times
\partial V^{+}}_{k-times} \rightarrow \mathbb{C}\quad \mathrm{symmetric}
| \left\Vert \Psi_k \right\Vert^2 <\infty\},
\end{align*}
with $
\partial V^{+}:=\{p\in \mathbb{R}^{d}|p^2=0,p_0>0\}.$
\end{definition}
The particle annihilation and creation operators $a,a^{*}$ of the bosonic
Fock space satisfy the following commutator relations
\begin{align}\label{pccr}
[a_c(\mathbf{p}), a_c^{*}(\mathbf{q})]=2\omega_{\mathbf{p}
}\delta^n(\mathbf{p}-\mathbf{q}), \qquad
[a_c(\mathbf{p}), a_c(\mathbf{q})]=0=[a_c^{*}(\mathbf{p}), a_c^{*}(\mathbf{q})].
\end{align}
By using $a_c,a_c^{*}$ the particle number operator and the momentum operator are defined in the following manner.
\begin{equation}\label{pcaopm}
N=\int d^n\mu(\mathbf{p}) a_c^{*}(\mathbf{p})a_c(\mathbf{p}), \qquad P_{\mu}=\int
d^n\mu(\mathbf{p})\, p_{\mu} a_c^{*}(\mathbf{p})a_c(\mathbf{p}),
\end{equation}
where $p_{\mu}=(\omega_{\mathbf{p}}, \mathbf{p})$. Moreover, we shall use the terms of the annihilation and creation operators in the noncovariant representation,
\begin{equation*}
{a}(\textbf{p}):= \frac{{a}_c (\textbf{p})}{\sqrt{2\omega_{\mathbf{p}}}},\qquad {a}^* (\textbf{p}):= \frac{{a}_c^* (\textbf{p})}{\sqrt{2\omega_{\mathbf{p}}}}.
\end{equation*}
Throughout this work we use $\mu, \,\nu=0,\dots,n$ and we shall use Latin letters for the spatial components which run from $1,\dots,n$. Moreover, we choose the following convention for the Minkowski scalar product of $ d $-dimensional vectors, $a\cdot b=a_0b^0+a_kb^k=a_0b^0- \vec{a}\cdot\vec{b}$.
\section{QFT-Moyal-Weyl from Deformation}\label{s3}
In the context of QM the deformation of the coordinate operator with the
momentum operator gave as the quantum plane of the Landau-quantization, (see \cite[Lemma 4.3]{Muc1}).
Since the commutator commutes with the generators of deformation, one can view the result as
the deformed commutator of the coordinate operators, (see Definition \ref{dpl1}). This is the point of view taken in this paper.
\\\\ Hence, in the present section
we follow the idea found in a QM context and calculate the deformed commutator of the \textbf{quantum field theoretical}
conjugate operator by using the momentum for deformation. The resulting quantum spacetime is called the \textbf{QFT-Moyal-Weyl spacetime}.
\subsection{Second-quantized Coordinate Operator}
In the so called \textbf{Schr\"odinger
representation}, \cite{RS1, BEH} the pair of operators
$(P_{i},Q_{j})$, satisfying the \textbf{canonical commutation relations} (CCR)
\begin{equation}\label{ccr}
[Q_i,P_{k}]=-i\eta_{ik},
\end{equation}
are represented as essentially self-adjoint operators on the dense domain
$\mathscr{S}(\mathbb{R}^n)$. Here $Q_{i}$ and $P_{k}$ are the closures of
$x_{i}$ and multiplication by $i {\partial}/{\partial x^k}$ on
$\mathscr{S}(\mathbb{R}^n)$ respectively.
Now in order to give the spatial coordinate operator in a \textbf{QFT-context} we take the quantum mechanical unitary equivalence of the momentum and coordinate operator into account and perform a second quantization.
\begin{lemma}\label{xsa}
Let $U_{\mathscr{F}}$ be the unitary operator of the $n$-dimensional Fourier transform, which is given by the following action on a function $\Psi\in\mathscr{S}(\mathbb{R}^n)$,
\begin{equation}\label{ft}
(U_{\mathscr{F}}\Psi)(p):= {(2\pi)^{-n/2}}\int\limits d^n\mathbf{x}\,\,
e^{ ip_kx^k}\Psi(x).
\end{equation}
Then, the Fourier transformation gives the unitary equivalence of operator $\mathbf{Q}$ to the operator $\mathbf{P}$ as follows,
\begin{equation}\label{uexp}
Q_{j}=U_{\mathscr{F}}^{-1}P_{j}U_{\mathscr{F}}.
\end{equation}
$\mathbf{Q}$ is an essentially self-adjoint operator on the dense domain $ D(\mathbf{Q})=
U_{\mathscr{F}}^{-1}D(\mathbf{P})=U_{\mathscr{F}}^{-1}\mathscr{S}(\mathbb{R}^n)=\mathscr{S}
(\mathbb{R}^n)$. Furthermore, its second-quantization $d\Gamma(.)$ is an essential self-adjoint operator on the dense domain $D(\mathbf{Q})_{\otimes}\subset \mathscr{F ({H})}$, which is the set of $\psi=\{\psi_{0},\psi_{1},\cdots\}$ such
that
$\psi_{k}=0$ for $k$ large enough and $\psi_{k}\in \bigotimes_{i=1}^k D(\mathbf{Q})$ for each $k$. The second-quantization of $\mathbf{Q}$ is given as
\begin{equation}\label{uexp2}
X_j:= d\Gamma(Q_{j})=\Gamma(U_{\mathscr{F}}^{-1})P_{j}\Gamma(U_{\mathscr{F}}),
\end{equation} with $\Gamma(U_{\mathscr{F}})=\bigoplus_{k=0}^{\infty} U_{\mathscr{F}}^{\otimes k}$.
\end{lemma}
\begin{proof}
The theorem for the one-particle subspace can be found in \cite[Equation 7.4]{BEH}.
Now the self-adjointness of the second quantized operator can be found in \cite[Theorem VIII.33]{RS1} and \cite[Chapter X.7]{RS}. By a slight abuse of notation we use the same symbol for the second-quantized momentum operator.
\end{proof}
\begin{remark}\label{ru}
Note that the quantum mechanical equivalence is demonstrated for a one-particle subspace w.r.t. Lebesgue measure. To change the Lebesgue to the Lorentz-invariant measure of QFT one combines the former with a multiplication operator. i.e. $(2P_0)^{-1}$. Hence, to have unitarity of the Fourier transformation w.r.t. the Lorentz-invariant measure we combine it with the unitary multiplication operator as follows. We consider the adjoint action of the operator $U_{\mathscr{F}}$ on the annihilation operator smeared with a function $f \in\mathscr{H}_1 $,
\begin{align*}
(\Gamma(U_{\mathscr{F}})a (\overline{f})\Gamma(U_{\mathscr{F}})^{-1})&=
\int \frac{d^n \mathbf{p}}{\sqrt{2\omega_{\mathbf{p}}}}
f(\mathbf{p})(\Gamma(U_{\mathscr{F}}) {a} (\mathbf{p})\Gamma(U_{\mathscr{F}})^{-1})
\\&=
\int \frac{d^n \mathbf{p}}{\sqrt{2\omega_{\mathbf{p}}}}
f(\mathbf{p}) \left( { {U}}_{\mathscr{F}} {a} \right)(\mathbf{p})
\\&=
\int \frac{d^n \mathbf{p}}{\sqrt{2\omega_{\mathbf{p}}}}\underbrace{\left(
\sqrt{2\omega} {U}_{\mathscr{F}}\frac{1}{\sqrt{2\omega}}
f\right)}_{=: {U}_{\mathscr{F},\omega}f}(\mathbf{p}) {a} (\mathbf{p}) = a(\overline{{U}_{\mathscr{F},\omega}f} ).
\end{align*}
The operator ${U}_{\mathscr{F},\omega}$ is unitary on the one-particle subspace w.r.t. the Lorentz-invariant measure.
\end{remark}
\begin{lemma}\label{xpr}
The coordinate operator has the following bosonic Fock space representation
\begin{equation}\label{uexp3}
X_j=-i\int d^n \mathbf{p}\, {a}^*(\textbf{p}) \frac{\partial}{\partial p^j} {a}(\textbf{p}).
\end{equation}
Moreover, let the dense domain $D(\mathbf{Q})_{\otimes}^{+}\subset \mathscr{F^{+} ({H})}$ be the projection onto the symmetric subspace of $D(\mathbf{Q})_{\otimes}$. Then, the second-quantized operator in the bosonic Fock space representation is an essential self-adjoint operator on $D(\mathbf{Q})_{\otimes}^{+}$.
\end{lemma}
\begin{proof} The proof for the Fock space representation (see \cite{SS}) can be found in \cite[Theorem 5.1]{Muc2}. Nevertheless, we outline it in more detail to ease to the reader the proof of the next somewhat more complicated lemma.
Since we intend to represent the second quantized coordinate operator on the symmetric Fock space, we use the momentum operator given in Equation (\ref{pcaopm}) and calculate the unitary equivalence explicitly on the creation and annihilation operators. Also note that the second quantized operator of the Fourier transformation commutes with the projections on to the symmetric Fock space (see for example \cite[Chapter VIII.10, Example 2]{RS1}). Therefore, we calculate the action of $\Gamma(U_{\mathscr{F}}^{-1})$ directly on the momentum operator in its symmetric Fock space representation, i.e.
\begin{align*}
X_j&=\Gamma(U_{\mathscr{F}}^{-1})P_{j}\Gamma(U_{\mathscr{F}})= \int
d^n \mathbf{p}\, p_{j} {(\hat{U}^{-1}_{\mathscr{F}}{a} )}^{*}(\mathbf{p})(\hat{U}^{-1}_{\mathscr{F}}{a})(\mathbf{p})\\&=
\int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(\hat{U}_{\mathscr{F}} p_{j} \underbrace{(\hat{U}^{-1}_{\mathscr{F}}{a} )}_{=:\tilde{a}} \right)(\mathbf{p}) \\&= \int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(
(2\pi)^{-n/2} \int d^n\mathbf{x}\,e^{ip_k x^k}x_j\,\tilde{a}(\mathbf{x})
\right) \\
&= -i(2\pi)^{-n/2}\int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(
\frac{\partial}{\partial p^j} \int d^n\mathbf{x}\,
e^{ip_k x^k} \,\tilde{a}(\mathbf{x})
\right) \\&= -i \int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})
\frac{\partial}{\partial p^j} {a} (\mathbf{p}),
\end{align*}
where for the Fourier transform we used \cite[Lemma IX.1]{RS}. The essential self-adjointness of the second-quantized operator on $D(\mathbf{Q})_{\otimes}^{+}$ follows from
\cite[Theorem VIII.33]{RS1}.
\end{proof}
\begin{remark}\label{fu}
The proof for the representation of the coordinate operator was done by using the unitarity of the Fourier transformation w.r.t. the Lebesgue measure, since the momentum operator was given in the noncovariant representation. However, the proof can be done analogously by using the momentum operator in covariant representation and the operator ${U}_{\mathscr{F},\omega}$ for the equivalence. This is merely a question of covariant versus noncovariant representation.
\end{remark}
Which quality of $\mathbf{X}$ makes it an appropriate coordinate operator? Well, first of all it is an essentially self-adjoint operator which is the second quantization of the QM-coordinate. Moreover in the following we stage the equivalence to the Newton-Wigner-Pryce operator (NWP). The NWP operator can be given by the product of generators
of the Poincar\'e group in the following way, \cite{ BAC2, BERG65, CA65, J80}
\begin{equation*}
Q_{j}^{NWP}= \frac{1}{2P_0}M_{0j}+M_{0j}\frac{1}{2P_0}.
\end{equation*}
On a one particle wave function the position operator is given by the following action,
\cite[Chapter 3c, Equation 35]{Sch}
\begin{equation}\label{NWP}
(Q_{j}^{NWP}\varphi)(\mathbf{p})=-i \left( \frac{p_j}{2\omega_{\mathbf{p}}^2}
+ \frac{\partial}{\partial p^j }
\right)\varphi(\mathbf{p}).
\end{equation}
\begin{lemma}\label{enwp}
The second quantization of the \textbf{Newton-Wigner-Pryce} operator onto the bosonic Fock space, defined on the dense domain $D(\mathbf{Q})_{\otimes}^{+}\subset \mathscr{F^{+} ({H})}$, is given in terms of the covariant annihilation and creation operators as follows,
\begin{equation*}
X_{j}^{NWP} =
-i \int
d^n \mu(\mathbf{p})\,
{{a}^{*}_c (\textbf{p})}
\left(\frac{p_j}{2\omega_{\mathbf{p}}^2}
+ \frac{\partial}{\partial p^j} \right){a}_c (\textbf{p}).
\end{equation*}
Moreover, it is equivalent to the position
operator $X_j$ given in Equation (\ref{uexp3}).
\end{lemma}
\begin{proof}
The first statement is fairly obvious by taking the action of the NWP operator (see Equation (\ref{NWP})) on the the dense domain $ D(\mathbf{Q})$ and performing a second quantization onto the bosonic Fock space. Now the equivalence can be as well easily seen by rewriting the spatial conjugate operator in terms of the covariant annihilation and creation operators, i.e.
\begin{align*}
X_j&= -i \int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})
\frac{\partial}{\partial p^j} {a} (\mathbf{p}) \\&=
-i \int
d^n \mathbf{p}\,
\frac{{a}^{*}_c (\textbf{p})}{\sqrt{2\omega_{\mathbf{p}}}}
\frac{\partial}{\partial p^j} \left(
\frac{{a}_c (\textbf{p})}{\sqrt{2\omega_{\mathbf{p}}}}\right)\\&=
-i \int
d^n \mu(\mathbf{p})\,
{{a}^{*}_c (\textbf{p})}
\left(\frac{p_j}{2\omega_{\mathbf{p}}^2}
+ \frac{\partial}{\partial p^j} \right){a}_c (\textbf{p}),
\end{align*}
Hence, the second quantization of the NWP operator, given by $d\Gamma(X_{j}^{NWP})$ onto the bosonic Fock space, is equivalent to $X_j$.
\end{proof}
A few comments are in order. First, the Newton-Wigner-Pryce operator is often given as the
product of generators of the Poincar\'e group. This representation is only true for one-particle
states and not for $n$-particle states. The reason is the simple observation that the product of second quantized
operators is not equal to the second quantized product of the operators, i.e. $d\Gamma(M_{0j}
P_{0}^{-1})\neq d\Gamma(M_{0j})d\Gamma(P_{0}^{-1})$. Therefore, the representation of the
NWP-operator as the product of the boost operator and the inverse of the Hamiltonian must be
discarded for an $n$-particle system.\newline\newline Second, from Lemma \ref{enwp} it follows
that the Newton-Wigner-Pryce
operator is equivalent to the position operator that we also obtained by second quantization
of the quantum mechanical position operator. One reason why this fact may have not been
obvious, is owed to the representation of the operator. The second quantized coordinate operator is
given in a non-covariant manner, while the NWP-operator is given in a covariant fashion. Of
course, the
difference of representation is merely a normalization feature and thus for the physical
interpretation, and specially for the second quantization not relevant.
\\\\ Since in QFT we work in a relativistic setting we \textbf{define} the zero component of the coordinate operator in the same fashion as the spatial part, i.e.
\begin{equation}\label{x0}
X_0:=\Gamma(U_{\mathscr{F}}^{-1})\,P_{0}\,\Gamma(U_{\mathscr{F}}).
\end{equation}
Note that the domain of essential self-adjointness of $P_0$ on the one-particle subspace is $\mathscr{S}
(\mathbb{R}^n)$.
\begin{lemma} The operator $X_0$ is given as an essentially self-adjoint operator on the dense domain $ D(\mathbf{Q})_{\otimes}\subset\mathscr{F ({H})}$ and has the following bosonic Fock space representation
\begin{equation}\label{uexpm3}
X_0=c_{d}
\int
d^n \mathbf{p} \,
{a}^*(\mathbf{p}) \left({\omega}^{ -{d}} \ast
{a}\right) (\mathbf{p} ),
\end{equation}
with constant $c_{d}:=-\pi^{-\frac{d}{2}}\Gamma(\frac{d}{2})$ and $ \ast$ denoting the convolution. Moreover, the second-quantized operator $X_{0}$ in the bosonic Fock space representation is an essential self-adjoint operator on $D(\mathbf{Q})_{\otimes}^{+}$.
\end{lemma}
\begin{proof}As for the spatial part the argument of essential self-adjointness is based on the unitary equivalence, given by the second quantized Fourier transformation, to the essential self-adjoint operator $P_0$. Let us now turn to the result in terms of the massless creation and annihilation operators,
\begin{align*}
X_0&=\Gamma(U_{\mathscr{F}}^{-1})P_{0}\Gamma(U_{\mathscr{F}})= \int
d^n \mathbf{p}\, \omega_{\mathbf{p}} {(\hat{U}^{-1}_{\mathscr{F}}{a} )}^{*}(\mathbf{p})(\hat{U}^{-1}_{\mathscr{F}}{a})(\mathbf{p})\\&= \int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(\hat{U}_{\mathscr{F}} {|\mathbf{p}|}{\tilde{a}} \right)(\mathbf{p}) \\& =\int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(
(2\pi)^{-n/2} \int d^n\mathbf{x}\,e^{ip_k x^k} {|\mathbf{x}|}\,\tilde{a}(\mathbf{x})
\right)
\\
&=(2\pi)^{-n}
\int
d^n \mathbf{p}\,{a}^{*}(\mathbf{p})\left(
\int d^n\mathbf{x}\,e^{ip_k x^k} {|\mathbf{x}|}\, \int d^n\mathbf{q} \,e^{-iq_k x^k} {a}(\mathbf{q})
\right)\\ &=(2\pi)^{-n}
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\,{a}^{*}(\mathbf{p})\left(
\int d^n\mathbf{x}\,e^{i(p-q)_k x^k}{|\mathbf{x}|}\, \right) {a}(\mathbf{q})\\&=-\pi^{-\frac{n+1}{2}}\Gamma(\frac{n+1}{2})
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\,{{|\mathbf{p}-\mathbf{q}|}^{ -(n+1)}}{a}^{*}(\mathbf{p}) {a}(\mathbf{q})
,
\end{align*}where the Fourier transformation performed in the last step can be found in \cite[Chapter III, Section 2.6]{GS1}. As for the spatial part, essential self-adjointness of $X_{0}$ on $D(\mathbf{Q})_{\otimes}^{+}$ follows from \cite[Theorem VIII.33]{RS1}.
\end{proof}
\subsection{Translations of the Conjugate Operators}
Before deforming the conjugate operator, we calculate the unitary transformation of these objects under
translations. This is done on an operator-valued distributional level by considering the
unitary transformation of the translations on particle creation and annihilation operators.
Let us first define the unitary operator of translations as follows,
\begin{equation}\label{bp}
U(b):=e^{ib_{\mu}P^{\mu}}.
\end{equation}
The unitary operator $ U(b)$ with $b\in\mathbb{R}^{d}$ transforms the particle creation
and annihilation operators $a,a^*$ in the following way, \cite{Sch}
\begin{equation}\label{caut}
U(b)\, {a}(\mathbf{p}) U(b)^{-1}=
e^{-ib_{\mu}p^{\mu}}{a}(\mathbf{p}), \qquad U(b) \,{a}^{*}(\mathbf{p})
U(b)^{-1}=
e^{ i b_{\mu}p^{\mu}}a^{*} (\mathbf{p}).
\end{equation}
The transformation property (\ref{caut}) is used in the next lemma to calculate the unitary
transformations of the coordinate operator. \newline$\,$
\begin{lemma}\label{utco}
Let $V_j$ be the velocity operator given as
\begin{equation}\label{vel}
V_j= \int d^n\mathbf{p}\,\frac{p_{j}}{\omega_{\textbf{p}}}\,{a}^*(\mathbf{p})
{a}(\mathbf{p}).
\end{equation}
Then, under the adjoint action of the unitary transformation $U(b):=e^{ib_{\mu}P^{\mu}}$, with
$b\in\mathbb{R}^{ {d}}$, the coordinate operators transform in the following way
\begin{align}\label{x0p}\nonumber
\alpha_{ b }\left(X_{0}\right)&= c_n
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\, {e^{ib_{\mu}(p-q )^{\mu}}}{\omega_{\mathbf{p}-\mathbf{q}}^{ - d}}{a}^{*}(\mathbf{p}) {a}(\mathbf{q})
\\&=: X_0- b_\mu \widetilde{V}^{\mu}+\mathcal{O}(b^2),
\end{align}
\begin{align}\label{xip}
\alpha_{ b }\left(X_{j}\right)&= e^{ib_{\mu}P^{\mu}}X_{j}
e^{-ib_{\mu}P^{\mu}}=X_{j}
+b_{0}V_j
- b_{j}N.
\end{align}
Moreover, the spatial part of $\widetilde{V}^{\mu}$ is the Fourier transform of the velocity operator and the second quantization of $\mathbf{Q}\cdot |\mathbf{Q}|^{-1}$, i.e.
\begin{equation*}
\widetilde{\mathbf{V}} =-d\Gamma(\mathbf{Q}\cdot |\mathbf{Q}|^{-1}).
\end{equation*}
\end{lemma}
\begin{proof}
By using the transformation property (see Equation (\ref{caut})) of the creation and
annihilation operator under the unitary
operator $U(b)$, we obtain for the zero component
\begin{align*}
\alpha_{ b }\left(X_{0}\right)&= c_n
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\,{\omega_{\mathbf{p}-\mathbf{q}}^{ - d}}e^{ib_{\mu}p^{\mu}} {a}^{*}(\mathbf{p}) e^{-ib_{\mu}q^{\mu}} {a}(\mathbf{q})\\&=
c_n
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\,\biggl(1+ib_\mu(p-q)^\mu\biggr) {\omega_{\mathbf{p}-\mathbf{q}}^{ - d}}{a}^{*}(\mathbf{p}) {a}(\mathbf{q})+\mathcal{O}(b^2) \\&=
X_0+ ib_\mu\,c_n
\iint
d^n \mathbf{p} \,d^n\mathbf{q}\, (p-q)^\mu {\omega_{\mathbf{p}-\mathbf{q}}^{ - d}} {a}^{*}(\mathbf{p}) {a}(\mathbf{q})+\mathcal{O}(b^2)
\\&=: X_0- b_\mu \widetilde{V}^{\mu}+\mathcal{O}(b^2).
\end{align*}
For the spatial component of the coordinate operators the adjoint action of the translations acts as follows,
\begin{align*}
e^{ib_{\mu}P^{\mu}}X_{j}
e^{-ib_{\mu}P^{\mu}}& =
-i \int d^n\mathbf{p}e^{ib_{\mu}p^{\mu}} {a}^*(\mathbf{p})
\partial_{j}\left(e^{-ib_{\mu}p^{\mu}} {a}(\mathbf{p})\right)
\\&= X_{j}
- b_{0}\int d^n\mathbf{p}(-\frac{p_{j}}{\omega_{\textbf{p}}}){a}^*(\mathbf{p})
{a}(\mathbf{p})
- b_{k}\int d^n\mathbf{p}\,\eta_{j}^{\,\,k}
{a}^*(\mathbf{p})
{a}(\mathbf{p}) \\&=X_{j}
+b_{0}V_j
- b_{j}N.
\end{align*}
By using the Baker-Campbell-Hausdorff formula it can be seen that the adjoint action of $X_0$ gives in first oder the Fourier transform of the velocity operator, i.e.
\begin{align*}
\alpha_{ \mathbf{b} }\left(X_{0}\right)&= X_0+ib_j\left[P^{j}, X_0\right]+\mathcal{O}(b^2)\\&=
X_0+ib_j\left(\left[\Gamma(U_{\mathscr{F}}) X^{j}\Gamma(U_{\mathscr{F}}^{-1}) , X_0\right]\right)+\mathcal{O}(b^2)\\&=
X_0+ib_j\left(\Gamma(U_{\mathscr{F}}) \left[X^{j} ,\Gamma(U_{\mathscr{F}}^{-1}) X_0\Gamma(U_{\mathscr{F}})\right]\Gamma(U_{\mathscr{F}}^{-1})\right)+\mathcal{O}(b^2)\\&=
X_0+ib_j \left(\Gamma(U_{\mathscr{F}}) \left[X^{j} , P_0 \right]\Gamma(U_{\mathscr{F}}^{-1})\right)+\mathcal{O}(b^2)\\&=
X_0-b_j \left(\Gamma(U_{\mathscr{F}})V^j\Gamma(U_{\mathscr{F}}^{-1})\right)+\mathcal{O}(b^2)\\&=
X_0-b_j\widetilde{V}^{j}+\mathcal{O}(b^2),
\end{align*}
where in the last lines we used the unitary equivalence and the velocity operator given by the Heisenberg equation $iV^{j}=[X^j,P_0]$. Moreover, the velocity operator in Equation (\ref{vel}) is the second quantization of $\mathbf{P}\cdot |\mathbf{P}|^{-1}$. Hence, its Fourier transform is given as,
\begin{align*}
\Gamma(U_{\mathscr{F}})d\Gamma(\mathbf{P} \cdot |\mathbf{P}|^{-1})\Gamma(U_{\mathscr{F}}^{-1})&= d\Gamma(U_{\mathscr{F}}\left(\mathbf{P}\cdot |\mathbf{P}|^{-1}\right)U_{\mathscr{F}}^{-1})\\&= -
d\Gamma(U^{-1}_{\mathscr{F}}\left(\mathbf{P}\cdot |\mathbf{P}|^{-1}\right)U_{\mathscr{F}})\\&=-
d\Gamma(\mathbf{Q} \cdot |\mathbf{Q}|^{-1}),
\end{align*}
where in the last lines we used the properties of the Fourier transform, the properties of the second quantization and the defining Equation (\ref{uexp2}). Of course, the equivalence can as well be directly shown on the level of the creation and annihilation operators, as was done for $X_{\mu}$.
\end{proof}
\subsection{QFT-Moyal-Weyl from deformation}Since we work with unbounded operators, we are obliged to prove that the deformed product of
the conjugate operators is well-defined. In particular, the task is to establish smoothness of the oscillatory integral
in a suitable locally convex topology and to demonstrate that all derivatives are polynomially bounded (see \cite[Theorem 1]{AA} for a similar proof).
After this is established, we build on results of \cite{LW} where smoothness and polynomial boundedness are the requirements needed in order to prove the well-definedness of the warped convolutions integral. \\\\ Now for the spatial part a simpler route will be chosen. It is easier to calculate the deformation of the spatial conjugate operator and take the commutator than to calculate the deformed product. The two results are equivalent since the outcome of the deformed product commutes with the momentum operator. This is a consequence of the fact that the product of the deformed operators is equal to the deformation of the deformed product of the operators (see Lemma \ref{dpl}). Therefore, we consider for the spatial part the oscillatory integral as follows
\begin{align*}
\langle \Psi , \mathbf{X}_{\theta}\Phi \rangle&=
(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y) {\langle \Psi,
U(y)\alpha_{\theta x}(\mathbf{X})\Phi \rangle} \\&=
(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi_{\theta}( \epsilon x,\epsilon
y) {\langle \Psi,
U(\theta y)\alpha_{ x}(\mathbf{X})\Phi \rangle}
,
\end{align*}
$\forall \Psi, \Phi \in D(\mathbf{Q})_{\otimes}^{+}$ and $ \chi_{\theta}( \epsilon x,\epsilon
y):= \chi (\epsilon\theta^{-1} x,\epsilon
\theta y)$ .
\begin{lemma}\label{lfpx}The deformation of the spatial conjugate operators, i.e. $\mathbf{X}_{\theta}$, is given as a well-defined
oscillatory integral on the dense domain $D(\mathbf{Q})_{\otimes}^{+}$.
\end{lemma}
\begin{proof}
Following the arguments discussed above we demonstrate the smoothness of the scalar product ${\langle \Psi,
U(\theta y)\alpha_{ x}(\mathbf{X})\Phi \rangle}$ in $x$ and $y$ and the polynomially boundedness of all derivatives. Hence, we start with the following expression
\begin{align*}
|\langle
\Psi, \partial_y^{r}\partial_x^{s}
U(\theta y)\alpha_{ x}(\mathbf{X})\Phi\rangle|=
|\langle
\Psi, (\theta P)^r
U(\theta y)\partial_x^{s}\alpha_{ x}(\mathbf{X})\Phi\rangle|. \end{align*}
where, $r,s \in\mathbb{N}^n_0$ are multi-indices. After applying the derivative w.r.t. the variable $x$ we obtain finite linear combinations of the following terms
\begin{align*}
|\langle
\Psi, (\theta P)^r
U(\theta y)P^{s_1} \alpha_{ x}(\mathbf{X})\underbrace{P^{s_2}\Phi}_{=:\Phi_s}\rangle|&
=
|\langle U(-\theta y)(\theta P)^rP^{s_1}
\Psi,
\alpha_{ x}(\mathbf{X})\Phi_s\rangle| \\&
\leq \underbrace{\| (\theta P)^r P^{s_1}
\Psi\|}_{=:c_{1,\Psi}} \| \alpha_{ x}(\mathbf{X})\Phi_s\|
\\&= c_{1,\Psi }\| \left( \mathbf{X}
+x_{0}\mathbf{V}
- \mathbf{x}N\right)\Phi_s\| \\&\leq c_{1,\Psi }\left( \underbrace{\| \mathbf{X}\Phi_s\| }_{=:c_2}
+|x_{0}|\underbrace{\| \mathbf{V}\Phi_s\| }_{=:c_3}
+\| \mathbf{x}\|\underbrace{\| N\Phi_s\| }_{=:c_4} \right)\\&\leq c_{ \Psi }\left(1+|x_{0}|+\| \mathbf{x}\|\right)
, \end{align*}
where in the last lines ${s_1} ,{s_2} \in\mathbb{N}^n_0$ are multi-indices and we used Cauchy-Schwarz. The constant $c_{1,\Psi}$ is finite
due to the properties of the vectors $\Psi$ and $\Phi$. Moreover due to the fact that $\Phi_s \in D(\mathbf{Q})_{\otimes}^{+}$ the constants $c_2, c_3, c_4$ are finite and a
constant $c_{ \Psi }$ exists such that the last inequality holds. Now
since $\Phi,\,\Psi$ are arbitrary elements
of a dense set of vectors we conclude that the deformation of the unbounded spatial conjugate operator is well-defined on $D(\mathbf{Q})_{\otimes}^{+}$.
\end{proof}$\,$\\
For the deformed product of the spatial and temporal conjugate operator we consider the oscillatory integral as follows
\begin{align*}
\langle \Psi , (X_0\times_{\theta}\mathbf{X}) \Phi \rangle&=
(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y) {\langle \Psi,\alpha_{\theta x}(X_0)\alpha_{y}(\mathbf{X}) \Phi \rangle}
,
\end{align*}
$\forall \Psi, \Phi \in D(\mathbf{Q})_{\otimes}^{+}$.
\begin{lemma}\label{lfpx1}The deformed product of the temporal and spatial conjugate operator, i.e. $X_0\times_{\theta}\mathbf{X}$, is given as a well-defined
oscillatory integral on the dense domain $D(\mathbf{Q})_{\otimes}^{+}$.
\end{lemma}
\begin{proof}
For this proof we demonstrate that the expression in the oscillatory integral, after getting rid of the boundary terms, is polynomially bounded.
Hence, we start by considering the following expression
\begin{align*} (2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y)
\langle \Psi ,\alpha_{\theta x}(X_0) \left(\mathbf{X}
+y_{0}\mathbf{V}
-\mathbf{y}N\right) { \Phi } \rangle ,
\end{align*}
where in the last line we used the adjoint action of the translations on $\mathbf{X}$ (see Equation (\ref{xip})).
Since the adjoint action of the momentum operator on $X_0$ is from a calculational point of view hard to handle to all orders in $(\theta x)$, we mention that we only need to look at the expansion in first order. This is due to the fact that all other orders vanish since the terms in $y$ are only of first order. For a proof of this statement review \cite[Proof of Lemma 5.3]{MUc}. However, this fact can be easily seen by rewriting the terms in $(\theta x)$ as partial derivatives of the exponential term. After partial integration, terms of unequal order in $x$ and $y$ vanish since derivatives of $\chi(\epsilon x,\epsilon
y)$ contain powers of $\epsilon$ which vanish in the limit $\epsilon\rightarrow 0$.
Now by using adjoint action of $X_0$ (see Equation (\ref{x0p})) we obtain for the deformed product the following,
\begin{align*}&(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y)\langle \Psi , (X_0- (\theta x)_\mu \widetilde{V}^{\mu}) \left(\mathbf{X}
+y_{0}\mathbf{V}
-\mathbf{y}N\right)\Phi \rangle \\&=(2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y)\langle \Psi , \biggl(
X_0 \mathbf{X}-y_0(\theta x)_\mu \widetilde{V}^{\mu} \mathbf{V}
+ \mathbf{y}(\theta x)_\mu \widetilde{V}^{\mu} N
\biggr)\Phi \rangle\\&= (2\pi)^{-d}
\lim_{\epsilon\rightarrow 0}
\iint \, d x \, d y\, e^{-ixy} \, \chi(\epsilon x,\epsilon
y)\underbrace{\langle \Psi , \biggl(
X_0 \mathbf{X}+ i \theta _{\mu 0} \widetilde{V}^{\mu} \mathbf{V}
-i \theta_{\mu j} \widetilde{V}^{\mu} N
\biggr)\Phi \rangle}_{\leq \|\Psi\|\left(\|X_0 \mathbf{X}\Phi\|+\| \theta _{\mu 0} \widetilde{V}^{\mu} \mathbf{V} \Phi\|+\| \theta_{\mu j} \widetilde{V}^{\mu} N\Phi\|\right)=: C_{\Psi,\Phi} } \\&\leq C_{\Psi,\Phi}(2\pi)^{-d}
\lim_{\varepsilon_1\rightarrow 0} \left(
\int dx \lim_{\varepsilon_2\rightarrow 0}
\left(\int dy e^{-ixy}
\chi_2(\varepsilon_2 y)\right)\,\chi_1(\varepsilon_1 x)\,
\right)
\\
&= C_{\Psi,\Phi} (2\pi)^{-d/2}
\lim_{\varepsilon_1\rightarrow 0} \left(
\int dx \,
\delta( {x} )\,\chi_1(\varepsilon_1 x) \right)
= C_{\Psi,\Phi} .
\end{align*}
Here we used the fact that the
oscillatory integral does not depend on the cut-off function chosen. As in
\cite{RI}, we chose $\chi (\varepsilon k,\varepsilon y)= \chi_2(\varepsilon_2 k
)\chi_1(\varepsilon_1 y)$
with $\chi \in\mathscr{S}(\mathbb{R}^{d}\times\mathbb{R}^{d})$ and $\chi_{l}(0 )=1$, $l=1$, $2$,
and obtained the delta
distribution $\delta(y)$ in the limit $\varepsilon_2 \rightarrow
0$, \cite[Section 7.8, Equation 7.8.5]{H}.
Hence, the remaining term inside the oscillatory integral is bounded by a constant and the integral is well-defined.
Now the scalar product under the integral is estimated by using Cauchy-Schwarz. The finiteness of $C_{\Psi,\Phi} $ follows from the property of the vectors
$ \Psi$ and $\Phi$. Equivalent considerations can be done for the deformed product $\mathbf{X}\times_{\theta}X_0$ and hence the deformed commutator of the conjugate operators is given as a well-defined oscillatory integral.
\end{proof}
\begin{remark}
In addition to the polynomial boundedness, smoothness w.r.t. the variables $x$ and $y$ is a necessary requirement. However, derivatives produce finite linear combinations of the following terms $
\langle (\theta P)^{r_1}\Psi ,\alpha_{\theta x}(X_0)(\theta P)^{r_2} P^{s_1}\left(\mathbf{X}
+y_{0}\mathbf{V}
-\mathbf{y}N\right) {P^{s_1} \Phi } \rangle$. Hence, we follow the same arguments as before and end up with a similar expression with the exception of additional polynomials of the momentum operator. Now by estimating the term using Cauchy-Schwarz and by using the fact that the vectors $ \Psi, \Phi \in D(\mathbf{Q})_{\otimes}^{+}$ finiteness of the resulting constant is guaranteed.
\end{remark}
$\,$\\
Next we turn to the actual result of the deformed product of the coordinate operators.
\newline
\begin{lemma}\label{qftdefx}
The deformed coordinate operator $X^{j}_{\theta}$, obtained by warped convolutions, represented on the dense domain $D(\mathbf{Q})_{\otimes}^{+}$ is given by
\begin{equation*}
X^{j}_{\theta}=X^{j} +\left(\theta P \right)^{0}V^j - \left(\theta P \right)^{j}N.
\end{equation*}
\end{lemma}
\begin{proof}
We use the defining equation (\ref{WC}) and Lemma \ref{utco}. The calculation can be performed for the spatial coordinate operator $X_{j}$ on
$\Psi \in
D(\mathbf{Q})_{\otimes}^{+}$ as follows,
\begin{align*}
X^{j}_{\theta}\Psi &= \int\alpha_{\theta k}( X^{j}) dE(k)\Psi
\\&= \int \left(X^{j}
+(\theta k)_{0}V^j
-(\theta k)^{j}N
\right)dE(k)\Psi
\\&= \left(X^{j} +\left(\theta P \right)^{0}V^j - \left(\theta P \right)^{j} N\right)
\Psi.
\end{align*}
\end{proof}
In the next step we give the commutator of the deformed coordinate operator. In order to make relativistic corrections more apparent we do not set the speed of light $c$ equal to one. \newline
\begin{theorem}\label{qmwxc} \textbf{QFT-Moyal-Weyl.} The deformed commutator of the coordinate operators represented on the dense domain $D(\mathbf{Q})_{\otimes}^{+}$ is given by
\begin{align*}
[X_\mu\stackrel{\times_{\theta}}{,} X_\nu] &= i\widehat{\theta}_{\mu\nu},
\end{align*}
where the operator-valued matrix $ i\widehat{\theta}$ is in the $0j$-component given as
\begin{align*}
i\widehat{\theta}_{0j}= i \theta_{0k} \, \widetilde{V} ^{k}_{\,\,j}
-2i\theta_{jk} \,
d\Gamma( {Q}^k \cdot |\mathbf{Q}|^{-1})N ,
\end{align*}
and in the spatial section, i.e. in the $ij$-components given as
\begin{align*}
i\widehat{\theta}_{ij}= -2 i\left(
\theta_{0i} V_j /c-\theta_{0j} V_i/c \right)N
-2i\theta_{ij}N^2.
\end{align*}
The operator $\widetilde{V}^{k}_{\,\,j}$ is the observable given by the commutator of the velocity with the Fourier transformed velocity, i.e. $ \widetilde{V}^{k}_{\,\,\,j} =[ \widetilde{V}^{k},V_j]$.
\end{theorem}
\begin{proof}
To calculate the commutator of the deformed product we first calculate the deformed product of $X_0$ and $X_j$ on $\Psi \in D(\mathbf{Q})_{\otimes}^{+}$,
\begin{align*}
(X_0\times_{\theta}X_j)\Psi&=(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \alpha_{\theta x}(X_0)\alpha_{y}(X_j)\Psi\\&=
(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \alpha_{\theta x}(X_0)\left(X_{j}
+y_{0}V_j
- y_{j}N\right)\Psi \\
&=
(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} (X_0- (\theta x)_\mu \widetilde{V}^{\mu} )\left(X_{j}
+y_{0}V_j
- y_{j}N\right)\Psi \\&=(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left(X_0X_{j}- (\theta x)_\mu y_{0} \widetilde{V} ^{\mu}V_j
+(\theta x)_\mu y_{j}\widetilde{V} ^{\mu}N\right)\Psi \\&=
X_0X_j\Psi + (2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left( i\frac{\partial}{\partial x^0}(\theta x)_\mu \widetilde{V} ^{\mu}V_j
-i\frac{\partial}{\partial x^j}(\theta x)_\mu \widetilde{V} ^{\mu} N\right)\Psi
\\&=X_0X_j\Psi +i (2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left( \theta_{\mu0} \, \widetilde{V} ^{\mu}V_j
- \theta_{\mu j}\, \widetilde{V} ^{\mu} N\right)\Psi \\&=\left(X_0X_j +i \theta_{\mu 0} \, \widetilde{V} ^{\mu}V_j
-i \theta_{\mu j}\, \widetilde{V} ^{\mu} N\right)\Psi ,
\end{align*}
where in the last lines we used the fact that the only nonzero terms in the expansion are the terms of
equal order in $(\theta x)$ and $y$. Moreover, we rewrote $y_{\mu}$ as a derivative on the exponential and performed a partial integration.
To calculate the commutator we additionally need the deformed product in exchanged order, i.e.
\begin{align*} i&(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \alpha_{\theta x}(X_j)\alpha_{y}(X_0)\Psi\\&=
(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy}
\left(X_{j}
+(\theta x)_{0}V_j
- (\theta x)_{j}N\right)\left(X_0-y_\mu \widetilde{V} ^{\mu}\right)\Psi \\&=(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left(X_jX_{0}- (\theta x)_0 y_{\mu} V_j \widetilde{V} ^{\mu}
+(\theta x)_j y_{\mu}N \widetilde{V} ^{\mu}\right)\Psi \\&=X_jX_{0} \Psi +(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left( i \frac{\partial}{\partial x^\mu}(\theta x)_0 V_j \widetilde{V} ^{\mu}
-i \frac{\partial}{\partial x^\mu}(\theta x)_j N \widetilde{V} ^{\mu}\right)\Psi \\&=X_jX_{0}\Psi +i(2\pi)^{- d}
\iint dx\,dy \, e^{-ixy} \left( \theta_{0 \mu} V_j \widetilde{V} ^{\mu}
-\theta_{j\mu} N \widetilde{V} ^{\mu}\right)\Psi \\&=\left( X_jX_{0} +i
\theta_{0 \mu} V_j \widetilde{V} ^{\mu}
-i\theta_{j\mu} N \widetilde{V} ^{\mu}\right)\Psi .
\end{align*}
By using the last two calculations we obtain the commutator of the deformed product which is given as,
\begin{align*}
[X_0\stackrel{\times_{\theta}}{,} X_j] \Psi&=
\left([X_0,X_j] +i \theta_{\mu 0} \, [\widetilde{V} ^{\mu},V_j]
-i\theta_{\mu j}\, \widetilde{V} ^{\mu} N
+i\theta_{j\mu} N \widetilde{V} ^{\mu}
\right)\Psi\\&=
\left([X_0,X_j] +i \theta_{\mu 0} \, [\widetilde{V} ^{\mu},V_j]
+i\theta_{0 j}\, [N,\widetilde{V} ^{0} ]
+i\theta_{jk} \{N, \widetilde{V} ^{k}\}
\right)\Psi,
\end{align*}
where in the last lines we used the skew-symmetry of $\theta$ w.r.t. the Minkowski metric, i.e.
$\theta_{0j}=\theta_{j0}$, $\theta_{kj}=-\theta_{jk}$ and $\{. \,,\, . \}$ denotes the anticommutator.
Now the first commutator vanishes since,
\begin{align*}
[X_0,X_j]=\Gamma(U^{-1}_{\mathscr{F}})\underbrace{[P_0,P_j]}_{=0}\Gamma(U_{\mathscr{F}})=0.
\end{align*}
Also note that $N$ commutes with all particle number conserving operators and since $\widetilde{V}$ is of the form $a^*a$ it commutes with $N$. Nevertheless, this statement can also be easily proven by using the canonical commutation relations (\ref{pccr}) and the form of the particle number operator given in Equation (\ref{pcaopm}). Hence the only term that remains is the second one, and we calculate it in the following,
\begin{align*} [\widetilde{V} ^{\mu},V_j]&
=-i\iiint
d^n \mathbf{p} \,d^n\mathbf{q}\,d^n\mathbf{k}\left(\frac{ (p-q)^\mu}{{\omega_{\mathbf{p}-\mathbf{q}}^{ d}}} \frac{k_{j}}{\omega_{\textbf{k}}}\right)\underbrace{ [{a}^{*}(\mathbf{p}) {a}(\mathbf{q}), {a}^*(\mathbf{k})
{a}(\mathbf{k})]}_{-
\delta^n( \mathbf{p}-\mathbf{k})
{a}^*(\mathbf{k})
{a}(\mathbf{q})+\delta^n( \mathbf{q}-\mathbf{k})
{a}^{*}(\mathbf{p}){a}(\mathbf{k})
}\\&=-i
\iint
d^n \mathbf{p} \,d^n\mathbf{q}
\left(\frac{ (p-q)^\mu}{{\omega_{\mathbf{p}-\mathbf{q}}^{ d}} }\left( \frac{q_{j}}{\omega_{\textbf{q}}}
-\frac{p_{j}}{\omega_{\textbf{p}}}
\right)
\right)
{a}^{*}(\mathbf{p}){a}(\mathbf{q})=:\widetilde{V} ^{\mu}_{\,\,j}.
\end{align*}
Hence by summing up all the commutators we obtain
\begin{align*}
[X_0\stackrel{\times_{\theta}}{,} X_j] \Psi&= \biggl(
i \theta_{\mu 0} \, \widetilde{V} ^{\mu}_{\,\,j}
+2i\theta_{jk} \widetilde{V} ^{k} N \biggr)\Psi\\&= \biggl(
i \theta_{0k} \, \widetilde{V} ^{k}_{\,\,j}
+2i\theta_{jk} \widetilde{V} ^{k} N \biggr)\Psi.
\end{align*}
Next we turn our attention to the commutator of the deformed product of the spatial coordinate operators.
We use the algebra of the massless coordinate operators, the algebra of the momentum operators
and the fact that the particle number operator $N$ commutes with $X^{\mu}$ and $P^{\nu}$ to
calculate the commutator of the deformed spatial operator.
\begin{align*}
[X_{i}^{\theta},X_{j}^{\theta}]\Psi&= [ X_{i}, \left(\theta P \right)_{0}V_j]\Psi-[ X_{i}, \left(\theta P \right)_{j}N]\Psi-
i\leftrightarrow j
\\&= \left(\theta P \right)_{0}[ X_{i}, V_j]\Psi+ [ X_{i}, \left(\theta P \right)_{0}]V_j\Psi
-[ X_{i}, \left(\theta P \right)_{j}]N\Psi-
i\leftrightarrow j
\\& = -2
i\left(
\theta_{0i} V_j -\theta_{0j} V_i \right)N\Psi
-2i\theta_{ij}N^2\Psi
\end{align*}
In the last lines we used the skew-symmetry of $\theta$ w.r.t. the
Minkowski metric. Note that the commutator of the deformed product of the spatial part is equal to the commutator of the deformed operator
$\mathbf{X}_{\theta}$ since,
\begin{align*}
\left( [X^{j}_{\theta},X^{k}_{\theta}]\right)_{-\theta }\Psi= [X^j\stackrel{\times_{\theta}}{,} X^k] \Psi,
\end{align*}
and the result of $[X^{j}_{\theta},X^{k}_{\theta}]$ commutes with $\mathbf{P}$. To see this we only have to calculate the commutator of the momentum operator with $\mathbf{V}$ since $N$ commutes with $\mathbf{P}$.
\begin{align*}
[V^{j} ,P^{\mu} ]& =
\iint d^n\mathbf{p}d^n\mathbf{q}\, \frac{p_{j}}{\omega_{\textbf{p}}}\,q^{\mu}\underbrace{ [{a}^{*}(\mathbf{p}) {a}(\mathbf{p}), {a}^*(\mathbf{q})
{a}(\mathbf{q})]}_{-
\delta^n( \mathbf{p}-\mathbf{q})
{a}^*(\mathbf{q})
{a}(\mathbf{p})+\delta^n( \mathbf{p}-\mathbf{q})
{a}^{*}(\mathbf{p}){a}(\mathbf{q})
}\\&=
\int d^n\mathbf{p}\, \left(-\frac{p_{j}}{\omega_{\textbf{p}}}\,p^{\mu}+\frac{p_{j}}{\omega_{\textbf{p}}}\,p^{\mu}\right){a}^{*}(\mathbf{p}) {a}(\mathbf{p})=0.
\end{align*}
In the last lines we used the canonical commutation relations (\ref{pccr}) and therefore we obtain,
\begin{align*}
\left( [X^{j}_{\theta},X^{k}_{\theta}]\right)_{-\theta }\Psi= [X^{j}_{\theta},X^{k}_{\theta}]\Psi=[X^j\stackrel{\times_{\theta}}{,} X^k]\Psi.
\end{align*}
\end{proof}
It turns out that the relativistic correction terms appearing, for the spatial part, are of the order $v/c$ and hence not equal to the standard Moyal-Weyl. Even more striking is the commutator relation between the zero component and the spatial component. It depends in both terms on the Fourier transformed velocity and even on $\theta_{ik}$. This suggests that setting $\theta_{0k}$ equal to zero, as often done in order to simplify calculations, is no longer an option. Of course, it is important to point out that in the one-particle and nonrelativistic limit, i.e. $v/c<<1$, we obtain the standard Moyal-Weyl spacetime for the spatial section.
\section{Conclusion and Outlook}
We obtained the QFT-Moyal-Weyl by taking the route layed out in a QM-context. In QM the deformation of the coordinate operator gives us, after the identification of $\theta$ with the inverse of a magnetic field, the so called guiding center
coordinates. These coordinates satisfy the commutator relations of the Moyal-Weyl Plane. In the Landau problem this plane is not merely
an abstract construct but has the physical meaning, that the space coordinates can not be
measured simultaneously.\\\\ Now at least the spatial part of the QFT-Moyal-Weyl plane can be viewed as a first order relativistic correction and second quantization of the Landau plane, after making the appropriate identifications, i.e. $\theta_{\mu \nu}=F^{-1}_{\mu \nu}$ for a constant electromagnetic field strength tensor $F$. This fact is first of all supported by working in a relativistic second-quantized context. Secondly, in the one particle non-relativistic limit we obtain the well known Landau plane (see \cite{ Muc1}). Hence, the QFT - noncommutative plane seems to be an intermediate stepping stone from the non-relativistic one-particle Landau quantization to the second quantized relativistic one. However, the case of a fully relativistic Landau effect and an emerging quantum spacetime thereof should be studied in more detail. In this context deformation theory seems to be a suitable approach.
\\\\
The commutator relations of the temporal and spatial part shed a new light on the Moyal-Weyl plane. First of all, the $\theta_{0j}$ term comes with an observable which is the commutator of the velocity with the Fourier-transformed velocity. This observable is not equal to the Kronecker-delta, as one can easily see from the form of the expression. In a sense the observable resembles the canonical commutation relations due to the definitions of the velocities in the different spaces. Nevertheless, we are not aware of any appearance of such a term in literature. Moreover, we have a second term in the commutator relations of $\hat{\theta}_{0j}$. This second term resembles the second quantization of a Lie-algebraic structure and is closely related to noncommutative spacetimes dubbed $\kappa$-Minkowski, \cite{AJ}. This can be seen clearer if we set $\theta_{0j}=0$ and apply the commutator relations on square-integrable functions defined on a sphere. Hence, turning the norm of $\mathbf{Q}$ into a constant identified with the radius of the sphere.
\\\\
The second-quantization also plays a role in the appearance of the particle number operator. It is not clear if this is an integral part of the QFT-Moyal Weyl or merely an effect of the deformation. The outcome is connected to the strategy we have taken. If we would have first deformed our conjugate operators and then performed a second quantization we would not have any particle number appearing terms, except for the $N^2$-term which would be linear in $N$. However, one could also define the conjugate operators with an $N^{-1}$ as was done in \cite{BDFP} for the mean event coordinate operator. The change will only be visible at the level of the commutator relations since $N$ commutes with all particle number preserving operators. The constant part of the spatial QFT-Moyal Weyl would have no particle number appearing and hence not differ from the conventional Moyal-Weyl. Moreover, the velocity dependent term would have the inverse of the particle number operator. Since all these possibilities have their theoretical justification only an experiment or arguments from renormalization theory can settle the uniqueness of the QFT-Moyal Weyl.
\\\\
Our intuition in using the massless scalar field is guided by simplicity. Nevertheless, by using the massive free scalar field the commutation relations do not change their form. The only obvious exception are the velocity and the Fourier transformed velocity, which would be interchanged with their massive counterparts. However, in the massive case one cannot draw analogies to the $\kappa$-Minkowski spacetime.
\\\\
Since, the QFT-Moyal-Weyl spacetime has some new features as the incorporation of the velocity and the particle number it would be interesting to study it in the context of scattering.
\bibliographystyle{alpha}
|
\section{Introduction}
\label{sIntro}
Many rational invariants of a finite group $G$ are encoded in the ring $\Char_\Q(G)$
of rationally-valued characters, the ring $R_\Q(G)$ of rational representations,
and the ring $\Perm(G)$ of virtual permutation representations. All three have the same
$\Z$-rank, and there are natural inclusions with finite cokernels
$$
\Perm(G) \quad\lar\quad R_\Q(G) \quad\lar\quad \Char_\Q(G).
$$
The quotient $\Char_\Q(G)/R_\Q(G)$ is studied by the theory of Schur indices,
and the purpose of this paper is to investigate the other two,
$$
\C(G) = \frac{R_\Q(G)}{\Perm(G)} \qquad\text{and}\qquad
\CH(G) = \frac{\Char_\Q(G)}{\Perm(G)}.
$$
They have exponent dividing $|G|$ by Artin's induction theorem,
and Serre remarked that $\C(G)$ need not be trivial (\cite{SerLi} Exc. 13.4).
It is trivial for $p$-groups \cite{Feit,Ritter,Segal},
and it is known for nilpotent groups \cite{Ras} (see also \S\ref{sBasic}),
for Weyl groups of Lie groups
\cite{Sol,Kle} and in other special cases \cite{Berz,HT}.
It follows from the general results of Dress, Kletzing, and Hambleton-Taylor-Williams \cite{Dress1, Dress2, Kle, HTW},
that the study of $\C(G)$ for a group $G$ reduces, in principle,
to that of its quasi-elementary subgroups, or of its `basic' quasi-elementary subquotients.
Specifically, for subgroups the statement is that of the two maps
$$
\prod_{Q\leq G\atop\text{quasi-elem.}}\C(Q)
\quad\stackrel{\Ind}{\hbox to 2.5em{\rightarrowfill}}\quad
\C(G)
\quad\stackrel{\Res}{\hbox to 2.5em{\rightarrowfill}}\quad
\prod_{Q\leq G\atop\text{quasi-elem.}}\C(Q),
$$
the first one is surjective and the second one injective, and similarly for $\CH$.
This is also an immediate consequence of Solomon's induction theorem, see \S\ref{sMackey}.
Our first observation is that the composite map allows us to describe $\C(G)$ and $\CH(G)$
explicitly, in a way that bypasses the representation theory of $G$ --- purely
in terms of quasi-elementary subgroups and the `$\Res\Ind$' maps between them;
in fact,
it is enough to consider maximal quasi-elementary subgroups, i.e. $p$-normalisers of
cyclic subgroups of $ G$. In \S\ref{sMackey} we give a simple formula for the $\Res\Ind$ maps,
and in \S\ref{sQE} we prove one of the main results of the paper,
which describes $\C(Q)$ and
$\CH(Q)$ for a $p$-quasi-elementary group $Q=C\rtimes P$.
Its simplest formulation is:
\begin{theorem}[=Theorem \ref{thmqemain1}]
Let $\rho$ be an irreducible rational representation of
a $p$-quasi-elementary group $Q=C\rtimes P$.
(So $C$ is cyclic, $P$ a $p$-group, and $p\nmid|C|$.)
The order of $\rho$ in $\C(Q)$ is $\frac{\dim\hat\psi\dim{\hat\pi}}{\dim\rho}$,
where $\hat\psi$ is the (unique) rational irreducible constituent of $\Res_C\rho$
and ${\hat\pi}$ a rational irreducible constituent of $\Res_P\rho$ of minimal dimension.
\end{theorem}
Together with the aforementioned `$\Res\Ind$' formula, it gives a way to compute
$\C(G)$ and $\CH(G)$ efficiently in a given finite group $G$.
Incidentally, it also gives an algorithm to find
$\Perm(G) \subset R_\Q(G)$ without computing the subgroup lattice,
which is now implemented in Magma \cite{Magma}.
In \S\ref{sGL2} and \S\ref{sPSL} we illustrate
applications of this approach to proving both triviality and non-triviality of
$\C(G)$, as we shall now describe.
In general, $\C(G)$ remains somewhat mysterious, especially in non-soluble
groups. Already Frobenius knew that $\C(A_n)$ is trivial for
all $n$. It was announced by Solomon in \cite{Sol} that $\C(\PSL_2(\F_q))$
is trivial for all prime powers $q$. In \S\ref{sGL2} we explain how this,
and the same statement for $\GL_2(\F_q)$ and $\PGL_2(\F_q)$, follow from
the results of \S\ref{sMackey} and \S\ref{sQE}.
There is, to our knowledge, no example in the literature of a simple
group with non-trivial $\C(G)$. In \S\ref{sPSL} we show:
\begin{theorem}[=Theorem \ref{pslmain} and Corollary \ref{cor:psl}]
The exponent of the 2-part of $\C(G)$ is unbounded in the families $G=\PSL_k(\F_p)$
and $G=\SL_k(\F_p)$. Moreover,
$\CH(\PSL_{k}(\F_p))\ne\{1\}$ for all even $k\geq 4$ and all odd primes $p$.
\end{theorem}
\subsection*{Notation}
Throughout the paper, $G$ denotes a finite group. We write
\smallskip
\begin{tabular}{llllll}
$\Char(G)$ & = & the character ring $G$, \cr
$\Char_\Q(G)$ & = & the ring of $\Q$-valued characters, \cr
$R_{\Q}(G)$ & = & the ring of characters of virtual $\Q G$-representations, \cr
$\Perm(G)$ & = & the ring of characters of virtual permutation representations, \cr
$\C(G)$ & = & ${R_\Q(G)}/{\Perm(G)}$, \cr
$\CH(G)$ & = & ${\Char_\Q(G)}/{\Perm(G)}$, \cr
$\Q(\chi)$ & = & the field of character values of a complex character $\chi$,\cr
$\schur(\chi)$ & = & the Schur index of an irreducible complex character $\chi$ over $\Q(\chi)$.\\[2pt]
\end{tabular}
\noindent
For a complex character $\chi$ of $G$, define its \emph{trace} and,
when $\chi$ is irreducible, its \emph{rational hull} as
$$
\begin{array}{clllllllllll}
\tr\chi &=& \displaystyle \sum_{\scriptscriptstyle \sigma\in \Gal(\Q(\chi)/\Q)}\chi^{\sigma}&&\in \Char_\Q(G), \\[5pt]
\hat{\chi} &=& \schur(\chi)\tr\chi &&\in R_\Q(G).\cr
\end{array}
$$
If $\chi$ is irreducible, then $\tr\chi$ is a \emph{$\Q$-irreducible character} and
$\hat{\chi}$ is the character of an \emph{irreducible rational representation}. We write
\smallskip
\begin{tabular}{llllll}
$\Irr(G)$ & = & the set of (complex) irreducible characters of $G$, \cr
$\Irr_\Q(G)$ & = & the set of $\Q$-irreducible characters of $G$, \cr
$\mu(\alpha,\beta)$ & = & $\frac{\langle\alpha,\beta\rangle}{\langle\alpha,\alpha\rangle}$ =
multiplicity of $\alpha$ in $\beta$, \cr
&& used for characters $\alpha\in\Irr_\Q(G), \beta\in \Char_\Q(G)$, and\cr
&& also for rational representations $\alpha,\beta$ with $\alpha$ irreducible.\\[2pt]
\end{tabular}
\noindent
We write $x\sim y$ for conjugate elements. A $p$-quasi-elementary group is one of the form
$G=C\rtimes P$ with $C$ cyclic, and $P$ a $p$-group; throughout the paper we adopt the
convention that $p\nmid |C|$.
\begin{acknowledgements}
The first author is supported by a Research Fellowship from the Royal
Commission for the Exhibition of 1851, and the second author is supported by a Royal Society
University Research Fellowship. We would like to thank Alexandre Turull
for his help with Corollary \ref{cor:psl}. We are grateful to an
anonymous referee for a careful reading of the manuscript and many helpful
comments.
\end{acknowledgements}
\section{Basic facts}
\label{sBasic}
\begin{lemma}\label{lem:lifts}
An inclusion $N\normal G$ induces injections $\C(G/N)\injects \C(G)$,
$\CH(G/N)\injects \CH(G)$.
\end{lemma}
\begin{proof}
Suppose $\bar\rho$ is a representation of $G/N$, which lifts to $\rho\in\Perm G$. Write
$$
\rho = \bigoplus \bC[G/H_i]^{\oplus n_i}, \qquad n_i\in\Z.
$$
For a subgroup $H\<G$ recall that
$
\bC[G/H]^N \iso \bC[G/NH],
$
as $G$-representations (see e.g. \cite{tamroot}, proof of Thm. 2.8(5)).
Therefore,
$$
\bar\rho=\rho^N = \bigoplus \bC[G/NH_i]^{\oplus n_i}\in \Perm(G/N),
$$
as required.
\end{proof}
\begin{lemma}\label{lem:schurnoschur}
Let $\rho$ be an irreducible rational representation and $\tau\in\Irr G$ its
constituent, so $\tr\tau\in \Irr_\Q(G)$ and $\rho=\schur(\tau)\tr\tau$.
The order of $\tr\tau$ in $\CH(G)$ is $\schur(\tau)$ times the order of
$\rho$ in $\C(G)$.
\end{lemma}
\begin{proof}
Clear from the definitions of $\C(G)$ and $\CH(G)$.
\end{proof}
This allows us to immediately deduce results about $\CH(G)$ from those about $\C(G)$,
and conversely.
\subsection*{Nilpotent groups}
Some statements seem to have a cleaner formulation for $\C(G)$
than for $\CH(G)$, and for some it is the other way around. Let us briefly
illustrate this with an example of nilpotent groups:
\begin{theorem}[Rasmussen \cite{Ras} Thm 5.2]
Let $G=G_2\times G_{2'}$ be a nilpotent group, where $G_2$ is its Sylow 2-subgroup.
Then $\C(G)$ is trivial, unless $G_{2'}\neq \{1\}$ and there exists an
irreducible character $\chi$ of $G_2$ with $\schur(\chi)=2$ and such that
one of the following holds:
\begin{enumerate}
\item $\Q(\chi)\neq \Q$, or
\item $\Q(\chi)= \Q$ and there exists a prime divisor $q$ of $|G|$ such that the order
of 2 $\pmod q$ is even.
\end{enumerate}
\end{theorem}
The conditions turn out to be much simpler if one transforms this into a result
about $\CH(G)$. The following follows easily from \cite[Thm. 4.2]{Ras}
and standard facts about Schur indices:
\begin{theorem}
Let $\chi=\chi_2\chi_{2'}$ be an irreducible character of a nilpotent group
$G=G_2 \times G_{2'}$ as above. Then the order of $\tr\chi$ in $\CH(G)$ is
$\schur(\chi_2)$ (which is 1 or~2).
\end{theorem}
\subsection*{Metabelian and supersolvable groups}
The following theorem will be of central importance in what follows. It implies that knowing
the order of every $\Q$-irreducible representation in $\CH(G)$ determines the structure
of $\CH(G)$ completely when $G$ is metabelian or supersoluble (e.g. nilpotent or
quasi-elementary). It does not hold in arbitrary groups, as first noted by Berz \cite{Berz};
the smallest counterexample is $G=C_3\times\SL_2(\F_3)$.
\begin{theorem}[Berz \cite{Berz}]\label{thm:Berz}
If $G$ is metabelian or supersoluble,
then $\Perm(G)\subseteq R_\Q(G)$ is freely generated by
$n_\rho\rho$, as $\rho$ ranges over irreducible rational representations of $G$, and
$$
n_\rho =\gcd_{H\le G} \mu(\rho,\Q[G/H]).
$$
\end{theorem}
\begin{lemma}
\label{A:Cp}
If $G=A\rtimes V$ with $A$ abelian and $V$ an elementary abelian
$p$-group, then $\CH(G)=\{1\}$.
\end{lemma}
\begin{proof}
By Theorem \ref{thm:Berz}, it is enough to show that every complex irreducible
character $\tau$ of $G$ occurs exactly once in $\bC[G/H]$ for a
suitable $H\<G$. This is clear when $\dim\tau=1$.
Otherwise $\tau=\Ind_{AU}^G\chi$, for some
subgroup $U$ of $V$ and a 1-dimensional character $\chi$ of $AU$
(see \cite[Part II, \S 8.2]{SerLi}). Let
$H$ be a subgroup of $V$ that is complementary to $U$, i.e. $HU=V$ and $H\cap U=\{1\}$.
By Mackey's formula, we have
\smallskip
$
\hfill\langle \tau,\bC[G/H]\rangle = \langle \chi,\Res_{AU}\Ind_{H}^G\triv\rangle =
\langle \chi,\Ind_{AU\cap H}^{AU}\triv\rangle = \langle \chi, \bC[AU]\rangle = 1.
\hfill
\qedhere
$
\end{proof}
Recall that a $p$-quasi-elementary group $G=C\rtimes P$ is {\em basic} if
the kernel $K$ of $P\rightarrow \Aut(C)$ is trivial or isomorphic to $D_8$ or
has normal $p$-rank one.
\begin{proposition}[\cite{HT}, Proposition 5.2]\label{prop:mainresfaithful}
Let $G=C\rtimes P$ be basic $p$-quasi-elementary.
Let $A_p$ be a maximal cyclic
subgroup of $K=\ker(P\rightarrow \Aut(C))$ that is normal in $P$ (it is all of $K$ if $K$ is cyclic,
and has index 2 in $K$ otherwise), let $A=C A_p$,
and let $\chi$ be a faithful one-dimensional character of $A$.
Then $\rho=\tr \Ind^G_A\chi$ is a $\Q$-irreducible character, and
$$
\text{order of $\rho$ in $\CH(G)$} = \frac{|P|}{|A_p|\cdot
\max\limits_{H\leq P\atop H\cap A_p=1}|H|}.
$$
\end{proposition}
\section{$\CH(G)$ as a Mackey functor}
\label{sMackey}
Let $\cR$ be a Mackey subfunctor of the character ring
Mackey functor $\Char(G)$. This simply means that for any finite group
$G$, $\cR(G)$ is a subgroup of $\Char(G)$ such that
if $H\leq G$ are finite groups, then
\begin{itemize}
\item for all $\rho\in \cR(H)$, $\Ind^G_H\rho\in \cR(G)$,
\item for all $\tau\in \cR(G)$, $\Res_H\tau\in \cR(H)$,
\item for all $\rho\in \cR(H)$ and $g\in G$, $\rho^g\in \cR(H^g)$.
\end{itemize}
Here are some examples:
\begin{itemize}
\item $R_K(G)$, the representation ring
of $G$ over a fixed subfield $K$ of $\bC$,
\item $\Char_K(G)$, the ring generated by $K$-valued characters, with fixed $K\subset \bC$,
\item $\Perm(G)$, the ring generated by permutation characters,
\item the subgroup of $\Char(G)$ generated by
characters of degree divisible by a fixed integer $n$.
\end{itemize}
If $p$ is a prime number, write $\cR(G)_p$ for $\cR(G)\otimes \Z_p$, where
$\Z_p$ is the ring of $p$-adic integers.
\begin{proposition}\label{mainreductionp}
Let $G$ be a finite group, fix a prime number $p$, and let $\cF_p$ be a family
of subgroups of $G$ such that every $p$-quasi-elementary subgroup of $G$ is conjugate to
a subgroup of some $Q\in\cF_p$. Then
$$
\prod_{Q\in \cF_p}\Res_Q: \cR(G)_p \lar \prod_{Q\in\cF_p} \cR(Q)_p
$$
is injective.
Dually,
$$
\sum_Q\Ind_Q^G: \prod_{Q\in\cF_p} \cR(Q)_p \lar \cR(G)_p
$$
is surjective.
\end{proposition}
\begin{proof}
By Solomon's induction theorem,
a prime-to-$p$ multiple $d$ of the trivial representation can be written as
$$
d\triv_G = \sum_i n_i \Ind_{H_i}^G\triv_{H_i}
$$
for some $p$-quasi-elementary subgroups $H_i$ and integers $n_i$.
Because $\Ind_{H_i^g}^G\triv_{H_i^g}\iso \Ind_{H_i}^G\triv_{H_i}$, we may assume
that each $H_i$ is contained in some $Q_i\in\cF_p$. Taking tensor products with
any $\rho\in \cR(G)$ yields
$$
d\rho = \sum_i n_i \Ind_{H_i}^G\Res_{H_i}\rho.
$$
If all $\Res_{H_i}\rho$ were 0, then so would be $d\rho$, and therefore also $\rho$.
This proves injectivity. Also, the equation shows that $d\rho \in
\Im\left(\sum_Q\Ind_Q^G\cR(Q)\right)$, which proves surjectivity, since $d$ is invertible
in $\Z_p$.
\end{proof}
\begin{corollary}
\label{mainreductionindres}
For $S,T\in\cF_p$ write $\alpha_{S,T}=\Res_T \Ind^G_S: \CH(S)\lar\CH(T)$.
Then
$$
\CH(G)_p \iso \text{Image}\Bigl(\prod_T\sum_S \alpha_{S,T}: \prod_{S\in\cF_p} \CH(S) \lar \prod_{T\in\cF_p} \CH(T)\Bigr).
$$
In particular,
$\CH(G)_p=1$ if and only if
for all pairs $S,T\in \cF_p$ and all $\rho\in R_\Q(S)$ (equivalently, for those $\rho$ whose class in
$\C(S)$ is nontrivial), we have $\Res^G_T\Ind_S^G\rho \in \Perm(T)$. The same also
holds for $\C(G)$.
\end{corollary}
\begin{proof}
Apply Proposition \ref{mainreductionp} to $\cR$ being $\Perm,R_{\Q}$,
and $\Char_{\Q}$.
\end{proof}
\begin{corollary}
\label{mainreduction}
Let $\cF$ be a family of subgroups of $G$ such that every quasi-elementary
subgroup is conjugate to a subgroup of some $Q\in\cF$.
Then
$$
\CH(G) \longinjects \prod_{Q\in\cF} \CH(Q)
$$
via the (product of) restriction maps. Consequently,
the kernel of the composition
$$
R_\Q(G) \>{\buildrel\prod\Res\over{\hbox to 40pt{\rightarrowfill}}}\>
\prod_{Q\in\cF} R_\Q(Q) \lar \prod_{Q\in\cF} \CH(Q)
$$
is $\Perm(G)$. Dually, the composition
$$
\prod_{Q\in\cF} R_\Q(Q) \stackrel{\Ind}{\longrightarrow} R_\Q(G) \rightarrow \CH(G)
$$
is onto. The same holds with $R_\Q$ replaced by $\Char_\Q$ and $\hat C$ by $C$.
\end{corollary}
\begin{remark}
The theorem and the two corollaries give a very efficient way of
computing $\CH(G)_p, \CH(G)$, $\C(G)_p$, $\C(G)$ and of finding
$\Perm(G)$ as a subring of $R_\Q(G)\leq \Char_\Q(G)$,
without computing the full lattice of subgroups of $G$.
\end{remark}
\begin{remark}
One possible family $\cF_p$ is the set of maximal $p$-quasi-elementary
subgroups of $G$. These are of the form
$$
Q = C\rtimes \Syl_p(N_G(C)),
$$
where $C$ is cyclic of order prime to $p$. Possible families $\cF$ in Corollary
\ref{mainreduction} are $\cF=\bigcup_p \cF_p$,
as $p$ ranges over prime divisors of $|G|$,
or alternatively $\cF=\{N_G(C)\}$ as $C$ ranges over
(representatives of conjugacy classes of) cyclic subgroups of $G$.
\end{remark}
\begin{notation}
For the remainder of this section we use the following notation:
\smallskip
\begin{tabular}{lll}
$CC(G)$ & = & the set of conjugacy classes of $G$,\\
$CC_{\cyc}(G)$ & = & the set of conjugacy classes of cyclic subgroups of $G$,\\
$[x]$ & = & the conjugacy class of $x$, when $x$ is either an element
of $G$\\ & & or a cyclic subgroup,\\
$\tr^*\chi$ & = & the normalised trace
$\tr^*\chi=\frac{1}{[\Q(\chi):\Q]}\tr\chi$ of a character $\chi$,\\
$\tau(D)$ & = & $\tau(y)$, where $D\leq G$ is a cyclic subgroup,
$y$ is any generator\\
& & of $D$, and $\tau\in \Char_{\Q}(G)\otimes \Q$. The rationality
of $\tau$ ensures\\
& & that $\tau(y)$ only depends on $D$ and not on the generator $y$.
\end{tabular}
Note in particular, that
for any character $\chi$ of $G$ and any cyclic subgroup $D$ of $G$,
$\tr^*\chi(D)$ is the average value of $\chi$ on the generators of $D$.
\end{notation}
\begin{lemma}
\label{lem:innerprod}
Let $H_1$, $H_2$ be two subgroups
of $G$. Let $\tau_i$ be a character of $H_i$, $i=1,2$, and
assume that
$\tau_1$ is $\Q$-valued. Then
\beq
\lefteqn{\langle \Ind_{H_1}^G \tau_1, \Ind_{H_2}^G \tau_2\rangle=}\\
& & \displaystyle\frac{1}{|H_1||H_2|} \sum_{[C]\in CC_{\cyc}(G)}|N_G(C)|\phi(|C|)\cdot
\sum_{D_1\leq H_1\atop D_1\sim C}\tau_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\tr^*\tau_2(D_2).
\eeq
\end{lemma}
\begin{proof}
First, note that by definition of inner products and of induced class functions,
$$
\langle \Ind_{H_1}^G \tau_1, \Ind_{H_2}^G \tau_2\rangle
=
\frac{1}{|H_1||H_2|}
\sum_{[x]\in CC(G)} |Z_G(x)|
\overline{\Bigl(\sum_{y\in[x]\cap H_1} \tau_1(y)\Bigr)}
\Bigl(\sum_{y\in[x]\cap H_2} \tau_2(y)\Bigr).
$$
The idea of the proof is to partition the set of conjugacy classes of
elements of $G$ according to conjugacy classes of cyclic subgroups they generate,
and to use the fact that for a rational character $\tau$, $\tau(x) = \tau(x')$
whenever $x$ and $x'$ generate conjugate cyclic subgroups. We get
\beq
\lefteqn{\langle \Ind_{H_1}^G \tau_1, \Ind_{H_2}^G \tau_2\rangle}\\
& = & \displaystyle \frac{1}{|H_1||H_2|}
\sum_{[x]\in CC(G)} |Z_G(x)|
\overline{\Bigl(\sum_{y\in[x]\cap H_1} \tau_1(y)\Bigr)}
\Bigl(\sum_{y\in[x]\cap H_2} \tau_2(y)\Bigr)\\[10pt]
& = &\displaystyle \frac{1}{|H_1||H_2|}
\sum_{[C] \in CC_{\cyc}(G)}f(C),
\eeq
where
\beq
f(C) &=& \displaystyle |Z_G(C)|\cdot \sum_{[x]\in CC(G)\atop \langle x\rangle = C}
\Bigl(\sum_{y\in[x]\cap H_1} \tau_1(y)\Bigr)
\Bigl(\sum_{y\in[x]\cap H_2} \tau_2(y)\Bigr)\\[10pt]
& = & |Z_G(C)|\cdot \#\{k: x\sim x^k\}\cdot\\
& &\displaystyle\qquad \cdot \sum_{D_1\leq H_1\atop D_1\sim C}\tau_1(D_1)\cdot
\sum_{[x]\in CC(G)\atop [x]\sim C}\sum_{y\in[x]\cap H_2} \tau_2(y)\\[10pt]
& = & \displaystyle|N_G(C)|\cdot
\sum_{D_1\leq H_1\atop D_1\sim C}\tau_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\sum_{\text{generators}\atop y \text{ of }D_2}\tau_2(y)\\[10pt]
& = &\displaystyle |N_G(C)|\cdot
\sum_{D_1\leq H_1\atop D_1\sim C}\tau_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\phi(|C|)\tr^*_{\bar{\Q}/\Q}\tau_2(y),
\eeq
as claimed.
\end{proof}
\begin{corollary}\label{cor:mu}
Suppose $H_1\<Q_1\<G$, $H_2\<Q_2\<G$, and let
$\chi_i$ be irreducible characters of $H_i$.
Set $\tau_i = \Ind_{H_i}^{Q_i}\chi_i$,
and $\rho_i=\tr\tau_i$.
Assume that $\tau_2$ is irreducible.~Then
\be
\mu(\rho_2,\Res_{Q_2}\Ind^{G}_{Q_1}\rho_1) & = &
\displaystyle\frac{[\Q(\tau_1):\Q]}{|H_1|\cdot|H_2|}
\sum_{[C]\in CC_{\cyc}(G)}|N_G(C)|\phi(|C|)\cdot\\
& & \displaystyle\qquad\cdot\sum_{D_1\leq H_1\atop D_1\sim C}\tr^*\chi_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\tr^*\chi_2(D_2).
\eeq
\end{corollary}
\begin{proof}
\beq
\mu(\rho_2,\Res_{Q_2}\Ind^{G}_{Q_1}\rho_1)
& = & \langle \tau_2, \Res_{Q_2}\Ind^G(\sum \tau_1^\sigma)\rangle\\
&=& \langle \Ind^G_{H_2}\chi_2,
\Ind_{H_1}^G (\sum_{\sigma\in \Gal(\Q(\tau_1)/\Q)}
(\chi_1)^\sigma)\rangle\\
&=& \frac{1}{[\Q(\chi_1):\Q(\tau_1)]}\langle \Ind^G_{H_2}\chi_2,
\Ind_{H_1}^G
(\tr\chi_1)\rangle\\[10pt]
&=&\frac{1}{[\Q(\chi_1):\Q(\tau_1)]|H_1|\cdot|H_2|}
\sum_{[C]\in CC_{\cyc}(G)}|N_G(C)|\phi(|C|)\cdot \\
& & \qquad \cdot \sum_{D_1\leq H_1\atop D_1\sim C}\tr\chi_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\tr^*\chi_2(y)\\[10pt]
&=&\frac{[\Q(\tau_1):\Q]}{|H_1|\cdot|H_2|}
\sum_{[C]\in CC_{\cyc}(G)}|N_G(C)|\phi(|C|)\cdot\\
& & \qquad\cdot
\sum_{D_1\leq H_1\atop D_1\sim C}\tr^*\chi_1(D_1)\cdot
\sum_{D_2\leq H_2\atop D_2\sim C}\tr^*\chi_2(D_2).
\eeq
\end{proof}
\begin{lemma}\label{lem:trace}
If $C$ is a cyclic group, and $\chi$ is a 1-dimensional character of $C$,
then $(\tr^*\chi)(C) = \mu(\ord(\chi))/\phi(\ord(\chi))$,
where $\mu$ is the Moebius mu function, and $\ord(\chi)$ is the smallest natural number
$n$ such that $\chi^{n}=\triv$.
\end{lemma}
\begin{proof}
It is enough to prove the lemma for faithful characters $\chi$, since we
may, without loss of generality, replace $C$ by $C/\ker\chi$.
Let $g$ be a generator of $C$. Then
$$
(\tr^*\chi)(C) = \frac{1}{[\Q(\chi):\Q])}\tr\chi(g)=\frac{1}{\phi(\ord(\chi))}\tr\chi(g).
$$
If $|C|=n$, then $\chi(g)$ is a primitive $n$-th root of unity, and the fact that
its trace is $\mu(n)$ is classical.
\end{proof}
\begin{corollary}
\label{cor:muindres}
Let $G$ be a group and $p^r$ a prime power.
Then $\CH(G)$ has an element of order $p^r$
if and only if there exist two $p$-quasi-elementary subgroups $Q_1$,
$Q_2$ of $G$, irreducible monomial characters $\tau_i=\Ind_{H_i}^{Q_i}\chi_i$
of $Q_i$, and an integer $k$,
such that
\begin{itemize}
\item
the rational character $\tr\tau_2$ has order divisible by
$p^{k+r}$ in $\CH(Q_2)$, and
\item
the rational number
\beq
\lefteqn{\frac{[\Q(\tau_1):\Q]}{|H_1||H_2|}\cdot\!\!\!\!
\sum_{[U]\in CC_{\cyc}(G)}|N_G(U)|\phi(|U|)\cdot}\\
& &\displaystyle\cdot \sum_{D_1\leq H_1\atop D_1\sim U}
\frac{\mu([D_1:D_1\cap\ker\chi_1])}{\phi([D_1:D_1\cap\ker\chi_1])}\cdot
\sum_{D_2\leq H_2\atop D_2\sim U}
\frac{\mu([D_2:D_2\cap\ker\chi_2])}{\phi([D_2:D_2\cap\ker\chi_2])}
\eeq
has $p$-adic valuation at most $k$.
\end{itemize}
In this case, $\Ind^G_{Q_1}\tr{\tau}_1$ has order divisible by $p^r$ in $\CH(G)$.
\end{corollary}
\begin{remark}\label{simplification}\par\noindent
\begin{itemize}
\item
Note that it is enough to take the last two sums in the above
formula only over those $D_i$ for which $D_i\cap \ker \chi_i$ has square-free
index in $D_i$, since for the others $\mu(\ord(\Res_{D_i}\chi_i))=0$.
For example if $\chi_i$ are faithful, then the outer sum may be taken over $U$ of
square free order.
\item
If, say, $H_1$ is cyclic, the sum $\sum_{D_1\leq H_1\atop D_1\sim U}$ has
at most one term for every $U$.
\item
If $Q_1$, $Q_2$ are basic and $H_1$, $H_2$ are cyclic, then
Proposition~\ref{prop:mainresfaithful} gives a simple expression for the
order of $\tr\tau_2$ in $\CH(Q_2)$.
\end{itemize}
\end{remark}
\begin{proof}[Proof of Corollary \ref{cor:muindres}]
By Corollary \ref{mainreductionindres}, $\CH(G)_p$ has an element of order
$p^r$ if and only if there exist $p$-quasi-elementary subgroups $Q_1$,
$Q_2$, and characters $\rho_i\in \Irr_{\Q}(Q_i)$,
such that $\rho_2$ has order $p^{k+r}$ in $\CH(Q_2)$ for some
$k$, and $\mu(\rho_2, \Res_{Q_2}\Ind^G\rho_1)$ has $p$-adic valuation at most
$k$. Quasi-elementary groups are M-groups, so if $\tau_i$ is a complex
irreducible constituent of $\rho_i$, then there exist subgroups
$H_i\leq Q_i$ such that $\tau_i=\Ind_{H_i}^{Q_i}\chi_i$
for 1-dimensional characters $\chi_i\in\Irr(H_i)$.
The result therefore follows from Corollary \ref{cor:mu} in combination
with Lemma \ref{lem:trace}.
\end{proof}
\section{Quasi-elementary groups}\label{sQE}
The aim of this section is to provide several formulae of theoretical and
algorithmic interest for the orders of characters in
$\CH(G)$ and $\C(G)$ when $G$ is quasi-elementary. Let
$G=C\rtimes P$ with $P$ a $p$-group
and $C$ cyclic of order coprime to $p$; we identify $P$ with a Sylow
subgroup of $G$.
\begin{lemma}\label{etatheta}
Let $N$ be a normal subgroup of a finite group $G$, let $\eta$ be an
irreducible character of $N$, and let $\theta$ be a complex irreducible
constituent of $\Ind_N^G\eta$. Write $\cG_\eta=\Gal(\Q(\eta)/\Q)$, and similarly
for $\cG_\theta$. Then
$$
\frac{[\Q(\eta):\Q]}{[\Q(\theta):\Q]} =
\frac{\#\{\gamma\in \cG_\eta\;|\;\langle\eta^\gamma,\Res_N\theta\rangle\neq 0\}}
{\#\{\gamma\in \cG_\theta\;|\;\langle\Ind_N^G\eta,\theta^\gamma\rangle\neq 0\}}.
$$
In particular, if $\Ind_N^G\eta$ is irreducible, then
$$
\frac{[\Q(\eta):\Q]}{[\Q(\theta):\Q]} = \#\{\gamma\in \cG_\eta\;|\;\langle\eta^\gamma,\Res_N\theta\rangle\neq 0\}.
$$
\end{lemma}
\begin{proof}
The $G$-action on the characters of $N$ commutes with the Galois action.
Every Galois conjugate of $\theta$ is a constituent of $\Ind^G\eta^\gamma$
for some $\gamma\in \cG_\eta$, and moreover the
number of distinct Galois conjugates of $\theta$ in $\eta^\gamma$ is independent
of $\gamma$. Also,
the number of Galois conjugates of $\eta$ in $\Res_N\theta^\gamma$ is
independent of $\gamma\in\cG_{\theta}$. So an inclusion--exclusion count gives
$$
\#\cG_\theta=\#\cG_{\eta}\cdot
\frac{\#\{\gamma\in \cG_\theta\;|\;\langle\Ind_N^G\eta,\theta^\gamma\rangle\neq 0\}}
{\#\{\gamma\in \cG_\eta\;|\;\langle\eta^\gamma,\Res_N\theta\rangle\neq 0\}}.
$$
\end{proof}
\begin{lemma}\label{dimhull}
Let $\eta$ be an irreducible complex representation of $G$, with rational hull $\hat\eta$.
Then
$$
\dim\hat\eta = \dim \eta\cdot\schur(\eta)\cdot[\Q(\eta):\Q].
$$
\end{lemma}
\begin{proof}
The rational hull of $\eta$ is given by
$$
\hat\eta = \schur(\eta)\sum_{\gamma\in\Gal(\Q(\eta)/\Q)}\eta^\gamma,
$$
whence the claim follows.
\end{proof}
\begin{theorem}\label{thm:innerprod}
Let $G=C\rtimes X$ with $C$ cyclic of order coprime to $|X|$. Let
$\tau$ be a complex irreducible character of $G$ with rational hull
$\rho=\hat\tau$,
let $\pi$ be a complex
irreducible constituent of $\Res_X\tau$ with rational hull $\hat{\pi}$,
$\psi$ an irreducible constituent of $\Res_C\tau$ with rational hull
$\hat\psi$, $K_{\psi}$ the stabiliser of $\psi$
under the $X$-action on $\Irr(C)$, and let $\xi$ be a complex irreducible
constituent of $\Res_{K_{\psi}}\pi$.
Then
\begin{eqnarray*}
\lefteqn{\mu({\rho},\Ind_X^G{\hat\pi}) =}\\
& & = \frac{\schur(\pi)}{\schur(\tau)}
\langle\xi,\Res_{K_{{\psi}}}\pi\rangle\cdot
\#\{\text{Galois conjugates $\pi'$ of $\pi \;|\; \langle\Res_{K_{\psi}}\pi',\xi\rangle\neq 0$}\}\\
& & = \frac{\dim{{\hat\psi}}\dim{\hat\pi}}{\dim{\rho}}.
\end{eqnarray*}
\end{theorem}
\begin{proof}
We may assume that $\rho|_C$ is faithful, otherwise we prove the
result in the quotient $G/(\ker\rho\cap C)$. So
$K=K_{\psi}$ is assumed to be the kernel of the $X$-action on $C$.
Recall that $\psi$ denotes a complex constituent of $\tau|_C$.
In particular, $\tau=\Ind_{CK}^G \psi\xi$, as explained in \cite[Part II, \S 8.2]{SerLi}.
We have
\beq
\rho &=& \displaystyle\schur(\tau)
\sum\limits_{\scriptscriptstyle\gamma\in\Gal(\Q(\tau)/\Q)} \tau^\gamma; &&
\dim\rho=\schur(\tau)[\Q(\tau):\Q]\dim\tau,\cr
\hat\pi &=& \schur(\pi)\displaystyle
\sum\limits_{\scriptscriptstyle\gamma\in\Gal(\Q(\pi)/\Q)} \pi^\gamma; &&
\dim\hat\pi=\schur(\pi)[\Q(\pi):\Q]\dim\pi.\cr
\eeq
Thus
\beq
\mu(\rho,\Ind_X^G\hat\pi) & = & \frac{1}{\schur(\tau)} \langle \tau, \Ind_X^G \hat\pi\rangle
=
\frac{1}{\schur(\tau)} \langle \Ind_{CK}^G \psi\xi, \Ind_X^G \hat\pi\rangle\\[4pt]
&=& \frac{1}{\schur(\tau)} \langle \Res_X \Ind_{CK}^G \psi\xi, \hat\pi\rangle
= \frac{1}{\schur(\tau)} \langle \Ind_K^X \xi, \hat\pi\rangle
= \frac{1}{\schur(\tau)} \langle \xi, \Res_K \hat\pi\rangle,\cr
\eeq
where the last line follows from Mackey's formula, noting that
$CK\backslash G/X$ consists of one double coset, and that
$CK\cap X=K$.
Next, $X$ acts on the representations of $K$ by conjugation,
and there is a Clifford theory decomposition
\beql{piclif}
\Res_K \pi = e\sum\limits_{\scriptscriptstyle g\in X/\Stab_X\xi} \xi^g.
\eeql
Recall that the constituents of $\hat\pi$ are Galois conjugates
of $\pi$, and we select those whose restriction to $K$ contains $\xi$:
$$
\Omega = \bigl\{\gamma\in\Gal(\Q(\pi):\Q) \bigm|
\langle \Res_K\pi^\gamma,\xi\rangle\ne 0\bigr\}.
$$
The inner product $\langle \Res_K\pi^\gamma,\xi\rangle=\langle \Res_K\pi,\xi^{\gamma^{-1}}\rangle$
is the same (and equals $e$) for every $\gamma\in\Omega$, since
$\xi^{\gamma^{-1}}$ is irreducible and so must be one of $\xi^g$ in \eqref{piclif}.
So we have
$$
\frac{1}{\schur(\tau)} \langle \xi, \Res_K \hat\pi\rangle
= \frac{\schur(\pi)}{\schur(\tau)} |\Omega| \langle \xi, \Res_K \pi\rangle,
$$
which proves the first equality.
It remains to show that
\beql{mudims}
\frac{\schur(\pi)}{\schur(\tau)} |\Omega| \langle \xi, \Res_K \pi\rangle = \frac{\dim{\hat\psi}\dim\hat\pi}{\dim\rho}.
\eeql
By comparing dimensions in \eqref{piclif}, and since $\tau=\Ind_{CK}^G\psi\xi$,
we see that
$$
\langle \xi, \Res_K \pi\rangle = e = \frac{\dim\pi}{[X:\Stab_X\xi]\dim \xi}
= \frac{[X:K]\dim\pi}{[X:\Stab_X\xi]\dim \tau} = \frac{[\Stab_X\xi:K]\dim\pi}{\dim\tau},
$$
so
$$
\mu(\rho,\Ind^G\hat\pi) = \frac{\schur(\pi)}{\schur(\tau)} |\Omega| \langle \xi, \Res_K \pi\rangle =
|\Omega|\cdot [\Stab_X\xi:K]\frac{\schur(\pi)\dim\pi}{\schur(\tau)\dim\tau}.
$$
Consider the two groups
\begin{eqnarray*}
H_1 & = & \{\gamma\in \Gal(\Q(\psi\xi)/\Q) \; |\; \langle (\psi\xi)^\gamma, \Res_{CK}\tau\rangle\neq 0\},\\
H_2 & = & \{\gamma\in \Gal(\Q(\xi)/\Q)\;|\;\langle\xi^\gamma,\Res_K\pi\rangle\neq 0\}.
\end{eqnarray*}
There is a natural projection $H_1\surjects H_2$ given by the restriction of
Galois action to $\Q(\xi)$, whose kernel consists of precisely those elements
of $\Gal(\Q(\psi\xi)/\Q)$ that act trivially on $\xi$, and through the action
of some $g\in X$ on $\psi$ (this last condition is equivalent to the Galois
element being in $H_1$). Thus, the kernel is isomorphic to the subgroup
of $G/CK$ that acts trivially on $\xi$, i.e. to $\Stab_X\xi/K$. We
deduce that
$$
\mu(\rho,\Ind^G\hat\pi) = |\Omega| \frac{|H_1|}{|H_2|}\frac{\schur(\pi)\dim\pi}{\schur(\tau)\dim\tau}.
$$
Now, by applying Lemma \ref{etatheta} first to $CK\normal G$ with $\theta=\tau$,
$\eta=\psi\xi$,
and then to $K\normal X$ with $\theta = \pi$, $\eta=\xi$, we find that
$$
|H_1| = \frac{[\Q(\xi):\Q][\Q(\psi):Q]}{[\Q(\tau):\Q]} \qquad\text{and}\qquad
|H_2| = |\Omega|\frac{[\Q(\xi):\Q]}{[\Q(\pi):\Q]},
$$
so that
\begin{eqnarray*}
\mu(\rho,\Ind^G\hat\pi) & = & |\Omega| \cdot\frac{|H_1|}{|H_2|}\frac{\schur(\pi)\dim\pi}{\schur(\tau)\dim\tau}\\
& = &|\Omega|\cdot \frac{[\Q(\xi):\Q]\cdot[\Q(\psi):\Q]\big/[\Q(\tau):\Q]}
{|\Omega|[\Q(\xi):\Q]\big/[\Q(\pi):\Q]}
\cdot\frac{\schur(\pi)\dim\pi}{\schur(\tau)\dim\tau}\\
& = & \frac{[\Q(\psi):\Q]\cdot[\Q(\pi):\Q]\schur(\pi)\dim\pi}{[\Q(\tau):\Q]\schur(\tau)\dim\tau}
= \frac{\dim{\hat\psi}\dim\hat\pi}{\dim\rho},
\end{eqnarray*}
where the last equality follows from Lemma \ref{dimhull}.
\end{proof}
\begin{theorem}
\label{thmqemain1}
Let $G=C\rtimes P$ be $p$-quasi-elementary,
let $\rho$ be an irreducible rational representation of $G$. Let $\psi$ be a
complex irreducible constituent of $\Res_C\rho$ with rational hull $\hat\psi$,
and let $\hat\pi$ be a rational irreducible constituent of
$\Res_P\rho$ of minimal dimension.
Denote by $\pi$ a complex irreducible constituent of $\hat\pi$,
by $\xi$ a complex irreducible consitutent of $\pi|_{K_\psi}$, where
$K_{\psi}\leq P$ is the stabiliser in $P$ of $\psi$,
and by $\tau$ a complex irreducible constituent of $\rho$ such that $\Res_P\tau$ contains $\pi$. Then
\begin{eqnarray*}
\lefteqn{\text{order of $\rho$ in $\C(G)$}
=\mu(\rho,\Ind_P^G{\hat\pi})= \frac{\dim\hat\psi\dim{\hat\pi}}{\dim\rho}}\\
& & =\frac{\schur(\pi)}{\schur(\tau)}
\langle\xi,\Res_{K_\psi}\pi\rangle\cdot
\#\{\text{Galois conjugates $\pi'$ of $\pi \; |\; \xi\subset\Res_{K_\psi}\pi'$}\}.
\end{eqnarray*}
\end{theorem}
\begin{proof}
We may assume that $\rho|_C$ is faithful, otherwise we prove the
result in the quotient $G/(\ker\rho\cap C)$ (see Lemma \ref{lem:lifts}). Thus,
$K=K_\psi$ is assumed to be the kernel of the $P$-action on $C$.
Under this assumption, if $H\le G$ intersects $C$ non-trivially, then
$$
\langle\rho,\bC[G/H]\rangle_G = \langle \Res_H\rho, \triv\rangle_H=0.
$$
Write $o$ for the order of $\rho$ in $\C(G)$.
By Theorem \ref{thm:Berz}, we have
\beq
o\cdot\langle \rho,\rho\rangle &=&
\gcd\limits_{H\le G}\langle\rho,\bC[G/H]\rangle_G =
\gcd\limits_{H\le P}\langle\rho,\bC[G/H]\rangle_G \cr
&=&
\gcd\limits_{H\le P}\langle\rho|_P,\bC[P/H]\rangle_P.
\eeq
Because $\C(P)=1$ by the Ritter-Segal theorem \cite{Ritter, Segal}, we can replace the
permutation representations $\bC[P/H]$ by all rational representations of $P$
in the last term. This is clearly the same as just taking the
rational irreducible constituents ${\hat\pi}_1,...,{\hat\pi}_k$ of $\rho|_P$, so
\begin{equation}\label{eqo}
o =
\frac{1}{\langle\rho,\rho\rangle}\gcd\limits_j\langle \rho|_P,{\hat\pi}_j\rangle
= \gcd\limits_j \frac{\langle\rho,\Ind_P^G{\hat\pi}_j\rangle}{\langle\rho,\rho\rangle}
= \gcd\limits_j \mu(\rho,\Ind_P^G{\hat\pi}_j).
\end{equation}
The theorem will therefore follow from Theorem \ref{thm:innerprod}, once we show
that the gcd may be replaced by the term corresponding to any ${\hat\pi}$ of minimal
dimension. Now, by Theorem \ref{thm:innerprod} and by Lemma \ref{dimhull},
$$
\mu(\rho,\Ind_P^G{\hat\pi}_j) = \frac{\dim\hat\psi\dim{\hat\pi}_j}{\dim\rho}
= \frac{\dim\hat\psi\,\schur(\pi_j)\dim\pi_j[\Q(\pi_j):\Q]}{\dim\rho},
$$
where $\pi_j$ is a complex irreducible constituent of ${\hat\pi}_j$.
We argue as in \cite[\S 2]{TurGcdmin}: if $p=2$, then
all the terms $\schur(\pi_j)$, $\dim\pi_j$, $[\Q(\pi_j):\Q]$ are powers
of 2, so gcd and minimum are the same. If $p$ is odd, then $\schur(\pi_j)=1$,
and moreover, either some $\pi_j=\triv$, in which case the claim is clear,
or else all $\dim\pi_j$ are powers of $p$, while all $[\Q(\pi_j):\Q]$
are $(p-1)$ times powers of $p$ (\cite[Lemma 2.1]{TurGcdmin}), so again
gcd and minimum are the same.
\end{proof}
\section{Examples: $\GL_2(\F_q)$, $\PGL_2(\F_q)$, $\SL_2(\F_q)$ and $\PSL_2(\F_q)$}
\label{sGL2}
\begin{theorem}
\label{gl2thm}
For every prime power $q=p^n$, the group $G=\GL_2(\F_q)$ has $\CH(G)~=~\{1\}$.
\end{theorem}
\begin{proof}
By Corollary \ref{mainreduction}, it suffices to show that every maximal quasi-elementary
subgroup $Q=C\rtimes P$ of $G=\GL_2(\F_q)$ is contained in some $\bar Q\<G$ with $\CH(\bar Q)=1$.
Pick $C=\langle g\rangle$ cyclic, and let $P=\Syl_l(N_G(C))$ for some prime
number $l$. Write $f(t)$ for the characteristic polynomial of $g$.
{\bf Case 1} (split Cartan).
Suppose $f(t)$ has distinct roots $a,b\in \F_q^\times$. Then $g$ is conjugate
to $\smallmatrix a00b$, and its centraliser is the split Cartan subgroup:
$$
Z_G(C)\iso \F_q^\times \times \F_q^\times, \qquad
N_G(C)\>\<\>(\F_q^\times \times \F_q^\times)\rtimes C_2,
$$
with $C_2=\langle\smallmatrix 0110\rangle$.
Here $\bar Q=(\F_q^\times \times \F_q^\times)\rtimes C_2$
has trivial $\CH(\bar Q)$ by Corollary \ref{A:Cp}.
{\bf Case 2} (non-split Cartan).
Suppose $f(t)$ is irreducible over $\F_q$. Then the centraliser of $C$ is
the non-split Cartan subgroup:
$$
Z_G(C)\iso \F_q[g]^\times\iso\F_{q^2}^\times, \qquad
N_G(C)\>\<\>\F_{q^2}^\times\rtimes\Gal(\F_{q^2}/\F_q)\iso \F_{q^2}^\times\rtimes C_2.
$$
Again $\bar Q=\F_{q^2}^\times\rtimes C_2$ has trivial $\CH$ by Corollary \ref{A:Cp}.
{\bf Case 3} (scalars). Suppose $g=\smallmatrix a00a$ is a scalar matrix. Then
$Q=C\rtimes P$ can be embedded into one of the following
\begin{itemize}
\item if $l=p$:
$\bar Q=C\times U=C\times \{\smallmatrix 1*01\}$; in this case
$U$ is an elementary abelian $p$-group; or
\item if either $l$ is odd and $l|(q-1)$, or $l=2$ and $q\equiv 1\mod 4$:
$\bar Q=H\rtimes C_2$ with $H=$ split Cartan; or
\item if either $l$ is odd and $l|(q+1)$, or $l=2$ and $q\equiv 3\mod 4$:
$\bar Q=H\rtimes C_2$ with $H=$ non-split Cartan.
\end{itemize}
In all these cases, $\CH(\bar Q)$ is trivial by Corollary \ref{A:Cp}.
{\bf Case 4} (non-semisimple). Finally suppose that $g$ is not semisimple,
say $g=g_s g_u$ with $g_s$ central and $g_u=\smallmatrix 1u01$ non-trivial unipotent.
Then
\beq
N_G C=N_G \langle g_u \rangle &=& \Bigl\{ \smallmatrix ab0c \Bigm|
ac^{-1}\in\F_p^\times \Bigr\} \\[4pt]
&=& \Bigl\{\smallmatrix a00a|a\in\F_q^\times\Bigr\}\cdot
\Bigl\{\smallmatrix 1c01|c\in\F_q\Bigr\} \cdot
\Bigl\{\smallmatrix b001|b\in\F_p^\times\Bigr\}\\[5pt]
& \cong & \F_q^\times\times (\F_q\rtimes \F_p^\times).
\eeq
If $l=p$, then $Q$ can be embedded into $\bar Q=\langle g_s\rangle \times U$,
where $U=\{\smallmatrix 1*01\}\cong \F_q$ is an elementary abelian $p$-group.
In this case $\CH(\bar Q)$ is trivial by Corollary \ref{A:Cp}.
Otherwise, $Q$ can be embedded into
$\bar Q\cong(\F_q^\times\times \langle g_u\rangle)\rtimes \F_p^\times$,
where the action in the semi-direct product is faithful. If $\tau$ is
an irreducible character of $\bar Q$ such that $\Res_{\langle g_u\rangle}\tau$
is faithful, then $\bar Q/\ker\tau$ satisfies the assumptions of
Proposition \ref{prop:mainresfaithful} with
$K=\{1\}$, so $\tr\tau\in \Perm(\bar Q)$.
Otherwise, $\Res_{\langle g_u\rangle}\tau=\dim\tau\cdot\triv$,
so $\tau$ factors through an abelian quotient, and $\tr\tau\in \Perm(\bar Q)$
e.g. by Corollary \ref{A:Cp}.
\end{proof}
\begin{remark}
It is also not hard to deduce the structure of $\CH$ for the related classical groups:
\par\noindent
\begin{itemize}
\item $G=\PGL_2(\F_q)$. Combined with Lemma \ref{lem:lifts}, the theorem implies $\CH(G)=1$.
\item $G=\SL_2(\F_q)$.
In general, $\CH(G)\ne 1$. For example,
$\SL_2(\F_3)$ has $\C=1$ and $\CH\iso \Z/2\Z$
(it has a 2-dimensional irreducible symplectic representation),
and $\SL_2(\F_{17})$ has $\C\iso \Z/4\Z$.
\item $G=\PSL_2(\F_q)$.
It is a result of Solomon, announced in \cite{Sol}, that $\CH(G)=1$.
This can also be seen following the argument for $\GL_2$ in Theorem \ref{gl2thm}:
the analogues of $\bar Q$ are the images of $\bar{Q}\cap \SL_2(\F_q)$
in $\PSL_2(\F_q)$,
and they are dihedral in Cases 1 and 2 of the theorem, elementary abelian
or dihedral ($p=2$) in Case 3 and isomorphic to $\F_p\rtimes\F_p^\times$ in Case 4.
Again, all these groups have $\CH=1$, so $\CH(G)=1$.
\end{itemize}
\end{remark}
\section{$\PSL_n(\F_p)$}
\label{sPSL}
Let $\ord_2$ denote the 2-adic valuation of a rational number,
$\ord_2\left(2^x\cdot{\displaystyle\frac ab}\right)=x$, where $2\nmid ab$.
\begin{theorem}
\label{pslmain}
Let $k\geq 4$ be an integer, and $p$ a prime.
The groups $\PSL_k(\F_p)$, and therefore also $\SL_k(\F_p)$,
have $\CH(G)$ of exponent divisible by $2^{\min(\ord_2(k),\ord_2(p-1))}$.
\end{theorem}
In the remainder of the section we prove the theorem using Corollary
\ref{cor:muindres}.
We will construct a 2-quasi-elementary subgroup $Q=C\rtimes P$ of
$G=\PSL_k(\F_p)$ and a rational character
$\rho$ of $Q$ such that $\Ind_Q^G\rho$ has order divisible by
$2^{\min(\ord_2(k),\ord_2(p-1))}$ in $\CH(G)$.
\begin{lemma}\label{lem:Zsigmondy}
Let $p$ be an odd prime and $k\geq 4$ an integer. If $k=4$, assume that
$p\equiv 1\pmod 4$. Then there exists a prime number
$l$ that divides $p^{k-2}-1$ but does not divide $p^s-1$ for any $s<k-2$.
\end{lemma}
\begin{proof}
This is a special case of Zsigmondy's Theorem \cite{Zsi}.
\end{proof}
Write $Q_{2^N}$ for the generalised quaternion group of order $2^N$.
\begin{lemma}
\label{sl2syl2}
The group $\SL_2(\F_q)$, $q=p^k$ has a 2-Sylow subgroup of the form
\begin{itemize}
\item
$S=\{\smallmatrix 1*01\}\iso C_p^k$ if $p=2$;
\item
$
S=\langle c,h \rangle\iso Q_{2^{N}},
c=\smallmatrix \alpha00{\alpha^{-1}},
h=\smallmatrix 01{-1}0
$
with $\alpha\in\F_q^\times$ of exact order $2^{N-1}||q-1$,
if $q\equiv 1\mod 4$;
\item
$
S=\langle c,h \rangle\iso Q_{2^{N}},
c=\smallmatrix \alpha{-\beta}\beta\alpha,
h=\smallmatrix \gamma\delta\delta{-\gamma}
$
with $\alpha+\beta \sqrt{-1}\in\F_{q^2}^\times$ of exact order $2^{N-1}||q+1$
and any choice of $\gamma,\delta\in \F_q$ with $\gamma^2+\delta^2=-1$,
if $q\equiv 3\mod 4$.
\end{itemize}
Conjugation by the matrix $\iota=\smallmatrix{-1}001$ is an automorphism of
$S$, acting
as $-1$ in the first case, as $c\mapsto c, h\mapsto h^{-1}$ in the second case,
and as $c\mapsto c^{-1}, h\mapsto h c^{2m+1}$ for some $m$ in the last case.
\end{lemma}
\begin{proof}
Direct computation.
\end{proof}
From now on, $G$ will denote $\PSL_k(\F_p)$.
The theorem only has content when $k$ is even and $p$ is odd, so we will assume
this. Write
$$
n=\ord_2(p-1)\ge 1, \qquad
N=\ord_2(p^{k-2}-1)\ge 3, \qquad
m=\ord_2(k-2)\ge 1.
$$
\textbf{Case A:}
Either $k>4$ or $p\equiv 1\mod 4$.
Let $A$ be a generator of a non-split Cartan subgroup
$\F_{p^{k-2}}^\times=\GL_1(\F_{p^{k-2}})\<\GL_{k-2}(\F_p)$, and
$l$ a prime divisor of $p^{k-2}-1$ as in Lemma \ref{lem:Zsigmondy}. The conditions
on $l$ imply that the normaliser
of $\langle A^{\frac{p^{k-2}-1}l}\rangle\iso C_l$ in $\GL_{k-2}(\F_p)$ is generated by $A$ and
by the Frobenius automorphism $F\in\Gal(\F_{p^{k-2}}/\F_p)$ of order $k-2$.
Note that $F$ has determinant $-1$,
since it is an odd permutation on a normal basis of $\F_{p^{k-2}}/\F_p$.
Define
\beq
c_p = \mat{1&1&{} \cr {}&1&{} \cr {}&{}&{I_{k-2}}}, & &
c_l = \mat{1&{}&{} \cr {}&1&{} \cr {}&{}&{A^{\frac{p^{k-2}-1}l}}},\\
x = \mat{d^{-1}&{}&{} \cr {}&1&{} \cr {}&{}&{U}}, & &
f = \mat{-1&{}&{} \cr {}&1&{} \cr {}&{}&{F^{(k-2)/2^m}}},
\eeq
where $U=A^{\frac{p^{k-2}-1}{2^N}}$ and $d=\det U$.
We view these matrices as representing elements of $G=\PSL_k(\F_p)$. Write
$$
C=\langle c_pc_l\rangle\iso C_{pl}, \;\;\;
P=\langle x,f\rangle\iso C_{2^N}\rtimes C_{2^m}, \;\;\;
Q=CP\iso (C_p\times C_l)\rtimes (C_{2^N}\rtimes C_{2^m}).
$$
Note that $C_{2^N}$ acts trivially on $C_l$, and through a $C_{2^n}$ quotient
on $C_p$, while $C_{2^m}$ acts through a $C_2$ quotient
on $C_p$ and faithfully on $C_l$.
\textbf{Case B:}
$p\equiv 3\mod 4$ and $k=4$. We take the same $c_p$ as in Case A, and $C=\langle c_p\rangle$.
A 2-Sylow of the centraliser of $C$ in $G$ is isomorphic to $\{1\}\times\Syl_2(\SL_2(\F_p))$,
which is isomorphic to $Q_{2^N}$ by the last case of Lemma \ref{sl2syl2}.
A 2-Sylow of the normaliser is
$$
P=\Syl_2 N_G(C) = \Syl_2 Z_G(C) \rtimes \mat{-1&{}&{}&{}\cr{}&1&{}&{}\cr{}&{}&{-1}&{}\cr{}&{}&{}&1}
\iso Q_{2^N}\rtimes \smallmatrix {-1}001,
$$
which is in fact isomorphic to the semi-dihedral group $SD_{2^{N+1}}$.
Again, we let $Q=CP$.
\smallskip
In both cases, write $K$ for the centraliser of $C$ in $P$.
Thus, $K\iso C_{2^{N-n}}$ in case A,
and $K\iso Q_{2^N}$ in case B, where the isomorphism is that of Lemma
\ref{sl2syl2}. Let $A_p$ be $K$ in case A, and a cyclic subgroup of index
2 in $K$ that is normal in $Q$ in case B.
Let $\chi$ be faithful irreducible characters of $CA_p$,
$\tau=\Ind_{CA_p}^Q\chi$ and $\rho=\tr\tau\in \Irr_{\Q}(Q)$.
\begin{lemma}\label{ordrho}
The character $\rho$ has order $2^n$ in $\CH(Q)$.
\end{lemma}
\begin{proof}
We will use Proposition \ref{prop:mainresfaithful}.
The biggest subgroup of $P$ that intersects $CA_p$ trivially is of order
2 in case B, and of order $2^m$ in case A.
So the order of $\rho$ in $\CH(G)$ is $2^{N+1-(N-1)-1}=2$ in case B, and
$2^{N+m-(N-n)-m}=2^n$ in case A.
\end{proof}
Finally, we show that
$\Ind_Q^G\rho$ has order divisible by $2^{\min(\ord_2(k),\ord_2(p-1))}$ in $\CH(G)$.
We will use Corollary \ref{cor:muindres} with $Q_1=Q_2=Q$ and $\chi_1=\chi_2=\chi$.
In view of Lemma \ref{ordrho}, it suffices
to show that
\begin{eqnarray}\label{muCaseA}
\sum_{[U]\in CC_{\cyc}(G)} S(U)
\end{eqnarray}
has 2-adic valuation at most $n-\min(\ord_2(k),\ord_2(p-1))$, where
for $U\leq CA_p$,
$$
S(U) = \frac{[\Q(\tau):\Q]}{|CA_p|^2}|N_G(U)|\phi(|U|)
\cdot \biggl(\sum_{D\leq CA_p\atop D\sim U}
\frac{\mu([D:D\cap\ker\chi])}{\phi([D:D\cap\ker\chi])}\biggr)^2.
$$
Note that since $CA_p$ is cyclic and $\chi$ is faithful, this simplifies to
$$
S(U) = \frac{[\Q(\tau):\Q]}{|CA_p|^2\phi(|U|)}|N_G(U)|\mu(|U|)^2,
$$
see Remark \ref{simplification}.
In particular, $S(U)=0$ if $U$ has non-square-free order.
\textbf{Case A}.
The subgroups of $CK$ of square-free
order are $C_{2lp}$, $C_{lp}$, $C_{2l}$, $C_l$, $C_{2p}$, $C_{p}$,
$C_2$, and $C_1$. We will show that $S(C_{lp}) + S(C_{2lp})$ has a strictly
lower 2-adic valuation than the rest of the sum, and that this valuation
is $n-\min(\ord_2(k),n)$. A summary of the calculations that follow is:
\begin{eqnarray*}
\ord_2 [\Q(\tau):\Q] &=& \ord_2(l-1)+N-n-1-m,\\
\ord_2 |CK|^2 &=& 2(N-n),\\
\phi(|C_{lp}|) = \phi(|C_{2lp}|) &=& (l-1)(p-1),\\
|N_G(C_{lp})|=|N_G(C_{2lp})| &=& \frac{(k-2)p(p^{k-2}-1)(p-1)}{\gcd (k,p-1)},\\
\ord_2(S(C_{lp})+S(C_{2lp})) &=& \ord_2(2S(C_{lp}))\\
& = &1+\ord_2(l-1)+N-n-1-m-2(N-2) + N+\\
& & n + m -\min(\ord_2(k),n) - \ord_2(l-1)+n\\
& = & n-\min(\ord_2(k),n).
\end{eqnarray*}
The assertions concerning $|CK|$ and $\phi(|C_{lp}|)$ are clear.
Since the conjugation action of $P$ on $\Irr(CK)$ is through Galois automorphisms,
and $\ker(P\rightarrow \Aut(CK))$ has index $2^{n+m}$ in $P$, we have
$$
[\Q(\tau):\Q]=2^{-n-m}[\Q(\chi):\Q]=\frac{p-1}{2^n}\frac{(l-1)2^{N-n-1}}{2^m},
$$
with 2-adic valuation $\ord_2(l-1)+N-n-1-m$.
The normaliser $N_{\GL_k(p)}$ of the preimage of $C_{lp}$ under $\SL\rightarrow \PSL$
consists of block diagonal matrices, with the normaliser of non-split Cartan
in the lower right corner (order $(k-2)(p^{k-2}-1)$), and a Borel subgroup in the top left
(order $p(p-1)^2$). The determinant is surjective on $N_{\GL_k(p)}$,
and $N_{\GL_k(p)}$ contains $Z(\GL_k(p))$, so the normaliser of $C_{lp}$ in $\PSL$ has order
$\frac{(k-2)(p^{k-2}-1)p(p-1)}{\gcd(k,p-1)}$, with 2-adic
valuation $N+n+m-\min(\ord_2(k),n)$. This is also the normaliser of $C_{2lp}$.
It remains to show that the rest of the sum in equation (\ref{muCaseA})
has strictly greater 2-adic valuation than $\ord_2(S(C_{lp})+S(C_{2lp}))$.
If $U\leq C$, then $|N_G(U)|$ and $|N_G(UC_2)|$ agree up to a power of $p$,
$\phi(|U|) = \phi(|UC_2|)$,
while $\mu(|U|)=-\mu(|UC_2|)$. It follows that the 2-adic valuation of $S(U) +
S(UC_2)$ is at least 1 greater than that of $S(U)$.
Moreover, for any $U\leq C_{lp}$, the normaliser of $U$ in $G$ contains
that of $C_{lp}$, while $1/\phi(|U|)$ has strictly greater 2-adic valuation than
$1/\phi(|C_{lp}|)$ whenever $U\neq C_{lp}$. This establishes the claim.
\newline
\textbf{Case B}.
The subgroups of $CA_p$ of square-free order are $C_1$, $C_2$, $C_p$, and $C_{2p}$.
We will show that $\ord_2(\sum S(U)) = \ord_2(S(C_p)+S(C_{2p}))=0$.
Again, we summarise the calculations as follows:
\begin{eqnarray*}
\ord_2 [\Q(\tau):\Q] &=& N-3,\\
\ord_2 |CA_p|^2 &=& 2N-2,\\
\phi(|C_{p}|) = \phi(|C_{2p}|) &=& p-1,\\
|N_G(C_{p})|=p^4|N_G(C_{2p})| &=& p^4\cdot\frac{(p-1)^3p^2(p+1)}{2},\\
\ord_2(S(C_{p})+S(C_{2p})) &=& \ord_2((1+p^4)S(C_{2p}))\\
& = &1+N-3-2N+2-1 + N+1=0.
\end{eqnarray*}
The assertions concerning $|CA_p|$ and $\phi$ are clear.
It follows from the description of the $P$-action on $\Irr(CK)$
that $[\Q(\tau):\Q]=\frac12[\Q(\chi):\Q]$, and has 2-adic valuation $2^{N-3}$.
The normaliser of $C_{2p}$ in $\GL_4$ is block diagonal, with all invertible matrices
in the bottom right corner, and Borel in the top left. So its order in $\PSL$ is
$\frac{(p-1)^3p^2(p+1)}{2}$ with 2-adic valuation $N+1$.
Finally, $|N(C_p)|=p^4|N(C_{2p})|$, e.g. see Murray \cite{Mur} \S4.
It remains to show that the 2-adic valuation of $S(C_1)+S(C_2)$ is positive.
The normaliser of $C_2$ in $\GL_4$ is $\GL_2\times \GL_2$, so the order of the normaliser
in $\PSL$ is $\frac{(p-1)^3p^2(p+1)^2}{2}$, with 2-adic valuation $2N$, and the
normaliser of $C_1$ is even bigger. So the 2-adic valuations of $S(C_1)$ and of $S(C_2)$
are positive.
\begin{corollary}\label{cor:psl}
As $G$ ranges over the simple groups $\PSL_{k}(\F_p)$, and therefore also
over $\SL_k(\F_p)$, the exponent of $\C(G)_2$ is unbounded.
\end{corollary}
\begin{proof}
If $\ord_2(k) > \ord_2(p-1)$, then by
\cite[Lemma 5.6(1)]{TurSchur} all Schur indices in $\PSL_k(\F_p)$ are trivial.
So the assertion follows from Theorem \ref{pslmain}.
\end{proof}
|
\section{Background on operads}
In this section, we review the material needed for this article. We refer to
Ginzburg and
Kapranov \cite{GLE} or J. L. Loday \cite{L}.
Let $K$ be a field of characteristic zero. An operad
(in the symmetric monoidal category of $k$-vector spaces) is given
by a collection of vector spaces $(\mathcal{O}(n)_{n\geq 0})$, a
right action of the symmetric group $S_n$ on $\mathcal{O}(n)$, and
a collection of compositions:
$$\begin{array}{clcll}
\circ_i&:&
\mathcal{O}(n)\otimes\mathcal{O}(m)&\longrightarrow&\mathcal{O}(n+m-1)~~~~~~~~
i=1,\dots,n;\\
&&(a,b)&\longmapsto &a\circ_i b ,\end{array}$$
with satisfy the following axioms:
\begin{itemize}
\item The two associativity conditions:
$$\begin{array}{cccc}
a\circ_i (b\circ_j c)&=&(a\circ_i b)\circ _{i+j-1}c,& \forall
i\in\{1,\dots,n\},~~\forall j\in\{1,\dots, m\}\\
(a\circ_i b)\circ_{m+j-1}c&=&(a\circ_j c)\circ_i b,&~~~\forall
i,j\in\{1,\dots,n\}, i<j,
\end{array}$$
called nested associativity and disjoint associativity
respectively.
\item The unit axiom: there exists an object $e\in\mathcal{O}(1)$
for which for any $a\in \mathcal{O}(n)$:
$$\begin{array}{ccll}
e\circ a&=&a,\\
a\circ_i e&=& a,~~~\forall i\in\{1,\dots,n\}
\end{array}$$
\item The equivariance axiom:
for any $\sigma\in S_n,~~\tau \in S_m$, we have:
$$a.\sigma \circ_{\sigma(i)}b.\tau=(a\circ_i b).\rho(\sigma,\tau),$$
where $\rho(\sigma,\tau)\in S_{n+m-1}$ is defined by letting $\tau$ permute the set
$E_i = \{i, i +
1,\cdots, i+m-1\}$ of cardinality $m$, and then by letting $\sigma$ permute the set
$\{1,\cdots, i-1,Ei,i +m,\cdots, m + n-1\}$
of cardinality $n$.
\end{itemize}
\begin{exam}
Any vector space $V$ yields an operad $End_V$ with $End_V (n)=
Hom(V^{\otimes n},V)$ where:
$$(f\circ_i
g)\big(a_1,\dots,a_{n+m-1}\big)=f\big(a_1,\dots,a_{i-1},g(a_i,\dots,a_{i+m-1}),a_{i+m},\dots,a_{n+m-1}\big)$$
\end{exam}
An algebra over an operad $\mathcal{O}$, or $\mathcal{O}$-algebra,
is a vector space $V$ together with an operad morphism from
$\mathcal{O}$ to $End_V$. This is equivalent to giving linear maps
$$\mathcal{O}(n)\otimes_{S_n} V^{\otimes n}\rightarrow V,$$
satisfying associativity conditions with respect to the
compositions.\\
An important point in the theory of operad
is the following theorem :
\begin{thm}
\cite[Chapter 5, sect 5.7.1]{LV}. The free $\mathcal{O}$-algebra generated by $V$ is the space
$\mathcal{O}(V)=\bigoplus_{n\geq 0}\mathcal{O}(n)\otimes_{S_n}
V^{\otimes n}.$
\end{thm}
In the remainder of this article, we describe operads
by species formalism, i.e: we replace the set $\{1,2,...,n\}$
by any finite set $A$ of cardinal $n$. For more details see sections $2$ and $3$ of
\cite{S}.
\section{A description of Pre Lie and Brace operads}
\subsection{Rooted trees and planar rooted trees}
\begin{itemize}
\item A rooted tree $T$ is a finite graph, without loops, with a
special vertex called the root of $T$. The set of rooted trees
will be denoted by $\mathcal{T}$. Let $D$ be a nonempty set. A
rooted tree decorated by $D$ is a rooted tree with an application
from the set of its vertices into $D$. The set of rooted trees
decorated by $D$ will be denoted by $\mathcal{T}^{D}$. Following
the notation of Connes and Kreimer \cite{CK1}, any tree $T$ writes
$T=B_{+}(r,T_1\cdots T_k)$ where $r$ is the root (or the decoration
of the root) and $T_1,\dots, T_k$ are trees. So we have
$$B_{+}(r,T_1\cdots T_k)=B_{+}(r,T_{\sigma(1)}\cdots
T_{\sigma(k)})~~~\forall \sigma\in S_k.$$
The vector space spanned
by $\mathcal{T}^D$ will be denoted by $\mathcal{RT}^D$. We denote by $\mathcal{RT}$ the species of rooted trees: for any finite set $A$ the vector space $\mathcal{RT}(A)$ is spanned by the rooted trees with $\vert A\vert$ vertices, together with a bijection from the set of vertices onto $A$.
\item A planar rooted tree $T$ is a rooted tree with an
embedding into the plane. The set of planar rooted trees will be
denoted by $\mathcal{T}_P$. Let $D$ be a nonempty set. A planar
rooted tree decorated by $D$ is a planar tree with an application
from the set of its vertices into $D$. The set of planar rooted
trees decorated by $D$ will be denoted by $\mathcal{T}_{P}^{D}.$
If $T_1 \neq T_2$, then:
$$B_+ (r,T_1T_2\cdots T_k)\neq B_+ (r,T_2 T_1\cdots T_k).$$
We draw the planar tree in the disk:
$$\mathcal{D}_+ =\{(x,y)\in\mathbb{R}^2; y>0~~\text{and}~~ x^2 +y^2<1\}~,$$
but the root is drawn in $x=y=0$.
The vector space spanned by $\mathcal{T}_{P}^{D}$ will be denoted
by $\mathcal{PRT}^D$.We denote by $\mathcal{PRT}$ the species of planar rooted trees: for any finite set $A$ the vector space $\mathcal{PRT}(A)$ is spanned by the rooted trees with $\vert A\vert$ vertices, together with a bijection from the set of vertices onto $A$.
\end{itemize}
\subsection{The pre-Lie operad}
We describe the pre-Lie operad in terms of non-planar labelled rooted trees, following
\cite{ChaLiv}. Let $A$ and $B$ be two finite sets. Let $v\in A$.
We define the partial composition
$\circ_v:\mathcal{RT}(A)\otimes\mathcal{RT}(B)\rightarrow\mathcal{RT}\big((A-\{v\})\amalg
B\big)$, as follows:
\begin{equation}
T\circ_v S=\sum_{f: E(T,v)\rightarrow B}{T\circ_{v}^{f}S},
\end{equation}
where $T\circ_{v}^{f}S$ is the tree of
$\mathcal{RT}((A-\{v\})\amalg B\big)$ obtained by replacing the
vertex $v$ of $T$ by the tree $S$ and connecting each edge $a$ in
$E(T,v)$ at the vertex $f(a)$ of $S$. If $v$ is not the root of
$T$, the edge going down from $v$ is now going down from the root
of $S$. The root of the new tree is the root of $T$ if it is
different from vertex $v$, and of $S$ else (see details in
\cite{ChaLiv}). The unit is the tree with a single vertex.
These partial compositions define an operad which is the pre-Lie operad.
\subsection{The Brace operad}
We describe the brace operad by the planar labelled rooted trees
(for more details see \cite{Chap1}). Let $T$ be a labelled planar rooted
tree. Let $s$ be a vertex of $T$. Let $B_{e}(s)$ be a little
disk of center $s$. The pair $(s,\alpha)$ is called an angle of
$T$ if $\alpha$ is a connected component of $B_{e}(s)\bigcap
(\mathcal{D}_+ \backslash T).$ We denote by $Ang(T)$ the set of
angles of $T$. Naturally, from left to right we set a total
order on $Ang(T)$ as follows: considering an angle as a direction from a vertex,
one can draw a path from every angle to a point of the upper part of the unit
circle. We order then these points clockwise.\\
\begin{exam}
$$\angleex$$
\begin{center}
$-~-~--$ Angles of of planar tree.
\end{center}
\end{exam}
Let $T, S$ be labelled planar rooted trees. Let $v$ be a
vertex of $T$. We denote by $E(T,v)$ the totally ordered set
(from left to right) of the incoming edges on $v$. We can
consider the set of increasing functions from $E(T,v)$ to $Ang
(S)$. We define
\begin{equation}
T\diamond_v S=\sum_{f:E(T,v)\to Ang(S)}{T\diamond_{v}^{f}S},
\end{equation}
where $f$ is an increasing function and $T\diamond_{v}^{f}S$ is the
planar tree obtained by substitution of $S$ on vertex $v$ of
$T$, plugging the incoming edges on $S$ according to map $f$.\\
These partial compositions defined above, define a structure of an
operad which is the Brace operad \cite{Chap1}.
\section {A description of the NAP-operad and its counterpart in the planar rooted
trees setting }
\subsection{The NAP operad}
we describe the NAP operad by the non-planar labelled rooted
trees \cite{Liv}. We define the partial compositions $\circ_v
:\mathcal{RT}(A)\otimes\mathcal{RT}(B)\rightarrow\mathcal{RT}\big((A-\{v\})\amalg
B\big)$, as follows:
\begin{equation}
T\circ_v S={T\circ_{v}^{f_0}S},
\end{equation}
where $T\circ_{v}^{f_0}S$ is the labelled rooted tree of
$\mathcal{RT}\big((A-\{v\})\amalg B\big)$ obtained by replacing
the vertex $v$ of $T$ by the tree $S$ and connecting each edge $a$
in $E(T,v)$ at the root of $S$. The unit is the tree with a single
vertex.
\subsection{An operad of planar rooted trees analogous to NAP }
In this section, we describe an operad of planar labelled
rooted trees. This operad is the planar analogue of NAP.\\
We define the partial compositions
$\diamond_v
:\mathcal{PRT}(A)\otimes\mathcal{PRT}(B)\rightarrow\mathcal{PRT}\big((A-\{v\})\amalg
B\big)$, as follows:
\begin{equation}
T\diamond_v S=\sum_{f_0:E(T,v)\to Ang^0(S)}{T\diamond_{v}^{f}S},
\end{equation}
where $Ang^0(S)$ is the set of angles starting from the root of $S$ and $f_0$ is
an increasing function from $E(T,v)$ to $Ang^0(S)$.
\begin{prop}
The partial compositions introduced above define a structure of an
operad. We will denote this operad by $\mathcal{B}^0$. The unit is the tree with a single vertex.
\end{prop}
\begin{proof}
We easily verify the unity, associativity and equivariance axioms. We omit the proof, as we will give a proof of a more general result later on.
\end{proof}
\begin{defn} \label{symm}
We will denote the symmetrization operator of trees by $\varphi$
from the space of non-planar labelled rooted trees to the space of planar labelled rooted trees,
by induction we define $\varphi$:\\
$\varphi(\racine)=\racine$ and if $T=B_+ (r,T_1\cdots T_k)$,
then
$$\varphi(T)=B_+\big(r,\varphi(T_1 )\sqcup \varphi(T_2)\sqcup\cdots\sqcup
\varphi(T_k)\big),$$ where $\sqcup$ is the shuffle product.
\end{defn}
i.e: $\varphi(T)$ is the sum of all planar representations of $T$.
\begin{thm}
$\varphi$ is a morphism of operads from Pre-Lie to Brace \cite{Chap1}.
Similarly $\varphi$ is a morphism of operads from NAP to $\mathcal{B}^0$.
\end{thm}
\begin{proof}
The first assertion is proved by F. Chapoton \cite[Prop 4]{Chap1}, the second assertion follows immediately by considering only the terms of minimal potential energy (see Definition \ref{potentiel} below).
\end{proof}
\begin{defn} We define
\begin{equation}
T\star S=(\echelvw\diamond_v T)\diamond_w S.
\end{equation}
Equivalently, if $T=B_+ (r,T_1\ldots T_n)$ then
$$T\star S=\sum_{i=0}^{n}{B_+ (r,T_1\ldots T_i ST_{i+1}\ldots T_n)}.$$
\end{defn}
\begin{prop}
The space $(\mathcal{T}_{P}^{D},\star)$ is a right non-associative permutative
algebra \cite{F}. i.e: for any planar rooted trees $T,S,U$, we have:
$$(T\star S)\star U=(T\star U)\star S$$
\end{prop}
\section{The notion of current-preserving operads}
\subsection{Structure of current-preserving operads}
Let $\mathcal{O}$ be an operad and $G$ be a commutative semigroup,
with additively denoted binary law. We say that $\mathcal{O}$ has a structure
of $G$-current-preserving operad, if moreover
$\mathcal{O}_A=\prod\mathcal{O}_{A,W},~~~W:A\to G$ where :
\begin{itemize}
\item The right action of the symmetric group $AutA$ verifies:
$$\mathcal{O}_{A,W}.\sigma=\mathcal{O}_{A,W\circ\sigma},~~\forall \sigma\in AutA.$$
\item For any finite sets $A, B$ and $v\in A,$ we have:
$$\circ_v: \mathcal{O}_{A,W}\otimes\mathcal{O}_{B,X}\to\mathcal{O}_{A-\{v\}\amalg
B,W\amalg X},$$
with image zero if $\sum_{b\in B}{X(b)}\neq W(v)$. Here, $W\amalg X$
is defined by: $W\amalg X(a)=W(a),\forall a\in A$, and $W\amalg X(b)=X(b),~~\forall
b\in B$.
\end{itemize}
\begin{exam}
The model of current-preserving operads is the operad $Endop(V)$ where $V$ is a $G$-graded
vector space \cite{S}.
Current-preserving operads are colored operads, with an extra structure given by
semigroup law on the set of colors.
\end{exam}
\subsection{Current-preserving operads associated to ordinary operads} \label{ordinaryoperad}
Given an operad $\mathcal{O}$ and any commutative semigroup $G$, we define
a $G$-current-preserving operad $\mathcal{O}^{G}$ as follows: for any finite
set $A$, we have:
$$\mathcal{O}_{A}^{G}:=\prod_{W:A\to G}{\mathcal{O}_{A,W},}$$
where $\mathcal{O}_{A,W}$ is nothing but a copy of $\mathcal{O}_A$. The partial compositions
of $(\alpha,W)\in \mathcal{O}_{A,W}$ and $(\beta,X)\in \mathcal{O}_{B,X}$ are defined for any $a\in A$ by:
$$(\alpha,W)\circ_{a}^{G}(\beta,X)=(\alpha\circ_{a}\beta,W\amalg X),$$
if $\sum_{b\in B}{X(b)}=W(a)$, and $(\alpha,W)\circ_{a}^{G}(\beta,X)=0$
if $\sum_{b\in B}{X(b)}\neq W(a)$. There is a natural morphism of operads
(in the ordinary sense) $\phi^G :\mathcal{O}\rightarrow \mathcal{O}^{G}$ given for
any finite set $A$ and for any $\alpha\in \mathcal{O}_A$ by:
$$\phi^G (\alpha):=\sum_{W:A\to G}{(\alpha,W)}.$$
The algebras on $\mathcal{O}^G$ are nothing but $G$-graded algebras on $\mathcal{O}$.
The morphism of operads $\phi$ simply reflects the forgetful functor from $G$-graded $\mathcal{O}$-algebras
to $\mathcal{O}$-algebras. Hence, any ordinary operad gives rise to a $G$-current-preserving operad naturally associated with it.
\section{A family of current-preserving operads}
\subsection{Interpolation between NAP and Pre-Lie}\label{NAPP}
We give here a family of $\mathbb N^*$-current-preserving operads
$(\mathcal{O}^{\lambda})_{\lambda\in K}$, where $\mathbb N^*$ is
the additive semi-group $\{1,2,3,\ldots\}$ of positive integers.
We have shown that this family interpolates between the $\mathbb
N^*$-current-preserving version of the NAP operad and the $\mathbb
N^*$-current-preserving version of the pre-Lie operad (see details in \cite{S}).
\begin{defn}
We introduce non-planar rooted trees with weights on their vertices:
$$\racineun ,\racinedeux ,\racinetrois,\ldots$$
For a non-planar rooted tree $T$ and a weight function $W:v\mapsto
W(v)\in\mathbb N^*$, we define the weight of $(T,W)$ by:
\begin{equation}
\left|T\right|=\sum_{v\in v(T)}W(v),
\end{equation}
where $v(T)$ denotes the set of vertices of $T$. Sometimes we will also use the
notation $|v|$ instead of $W(v)$.
\begin{exam}
$\racineun ,\racinedeux ,\racinetrois
,\echelunun,\echelundeux,\echeldeuxun,\couronne,\echelununun$ are
the non-planar rooted trees with weight less or equal to $3$.
\end{exam}
\end{defn}
We draw non-planar rooted trees with labels and numbers on their vertices, each number refers
to the weight of the vertex.
\begin{defn} \label{potentiel}
We define the potential energy of a weighted non-planar rooted tree $(T,W)$ by:
\begin{equation}
d(T)=\sum_{v\in v(T)}W(v)h(v),
\end{equation}
where $h(v)$ is the height of $v$ in $T$, i.e. the distance from $v$ to the root of
$T$ counting the number of edges.
\end{defn}
This notion of potential energy matches the physical intution: if a branch
of a tree is moved down, the potential energy decreases by a multiple of its
weight.\\
For any finite set $A$, let $\mathcal{O}_A$ be the completed
vector space spanned by the non-planar rooted trees with $\left|A\right|$
vertices of any weight, labellized by $A$. Namely:
\begin{equation}
\mathcal{O}_A:=\prod_{W:A\to\mathbb N^*}\mathcal{O}_{A,W},
\end{equation}
where $\mathcal{O}_{A,W}$ is the vector space spanned by the
non-planar rooted trees with $\left|A\right|$ vertices, labellized by $A$ and
with weight function $W$. For any weighted non-planar rooted tree $S\in
\mathcal{O}_{A}$ and any vertex $v$ of $S$, $E(S,v)$ denotes the
set of edges of $S$ arriving at the vertex $v$ of $S$. Let $B$ be
another finite set and $T\in \mathcal{O}_{B}$ another weighted
rooted tree with $|B|$ vertices. Let $\lambda$ be an element of
the field $K$. We define the partial compositions by:
$$S\circ_{v,\lambda} T=\left\{\begin{array}{l}
\sum_{f:E(S,v)\rightarrow v(T)}{\lambda^{d(S\circ_{v}^{f}
T)-d(S\circ_{v}^{f_0}T)}S\circ_{v}^{f}T}~~~~\text{if}~~\left|~T\right|=\left|v\right|\\
0~~~~~\text{otherwise},
\end{array}\right.$$
where $S\circ_{v}^{f}T$ is the element of
$\mathcal{O}_{(A-\{v\})\amalg B}$ obtained by replacing the vertex
$v$ by $S$ and connecting each edge of $E(S,v)$ to its
image by $f$ in $v(T)$. Here $f_0$ is the map from $E(S,v)$ to
$v(T)$ which sends each edge $a$ of $E(S,v)$ to the root of $T$.
The tree $S\circ_{v}^{f_0}T$ has therefore the smallest potential
energy in the above sum. Unit is given by:
$$e=\sum_{n\geq1}{\racinen},$$
where $\racinen$ is the tree with one single vertex of weight $n$
(this infinite sum makes sense as ${\mathcal O}_1$ is a direct
product). The right action of the symmetric groups is given by
permutation of the labels.
\begin{exam}
Let us consider $S=\arbaa$ and $T=\arbab$. Here letters $a,b,c...$ are labels of
vertices, which are of weight $1,2$ or $3$. We have:
\begin{align*}
S\circ_{b,\lambda} T=\\
&&\arbac&+\lambda\arbad&+\lambda^{2}\arbae &\hskip 6mm+\lambda^{3}\arbaf\\
& &\hskip -8mm f_0
:\left\{\begin{array}{ccc}c&\rightarrow&e\\d&\rightarrow&e\end{array}\right.\hbox
to
10mm{}&f:\left\{\begin{array}{ccc}c&\rightarrow&e\\d&\rightarrow&h\end{array}\right.&f:\left\{\begin{array}{ccc}c&\rightarrow&h\\d&\rightarrow&e\end{array}\right.&\hskip
8mm
f:\left\{\begin{array}{ccc}c&\rightarrow&h\\d&\rightarrow&h\end{array}\right.&&
\end{align*}
\end{exam}
\begin{thm}\label{principal}
The partial compositions defined above \cite{S} yield a structure of $\,\mathbb
N^*$-current-preserving operad on the species $A\mapsto \Cal O_A$, denoted by
$\mathcal{O}^{\lambda}$.
\end{thm}
\begin{rmk}
With the notations of Sec \ref{ordinaryoperad}, the current-preserving operad $\mathcal{O}^{\lambda}$ is naturally associated with $NAP$ for $\lambda=0$
and with Pre-Lie for $\lambda=1$ \cite{S}.
\end{rmk}
\subsection{Interpolation between Brace and $\mathcal{B}^0$}
We keep the notations of the sec. \ref{NAPP} but we replace non-planar rooted trees by planar
rooted trees and $\mathcal{O}$ by $\mathcal{B}$.\\
Let $A$ be a finite set. Let $S$ be a weighted planar rooted tree with $|A|$ vertices.
Let $B$ another finite set and $T$ another weighted planar rooted tree with $|B|$
vertices.
Let $\lambda$ be an element of the field $K$. We define the partial compositions:
$$S\diamond_{v,\lambda} T=\left\{\begin{array}{l}
\sum_{f:E(S,v)\rightarrow~ \text{Ang}(T)}{\lambda^{d(S\diamond_{v}^{f}
T)-d(S\diamond_{v}^{f_0}T)}S\diamond_{v}^{f}T}~~~~\text{if}~~\left|~T\right|=\left|v\right|\\
0~~~~~\text{otherwise},
\end{array}\right.$$
where $S\diamond_{v}^{f}T$ is the weighted planar rooted tree of
$\mathcal{B}_{A-\{v\}\amalg B}$
obtained by replacing the vertex $v$ by $T$ and connecting each edge of $E(S,v)$ to
its image
by $f$ in $Ang(T)$. Here, $f_0$ is any increasing map from $E(S,v)$ to $Ang^0(T)$.
The trees $S\diamond_{v}^{f_0}T$
have, therefore,
the smallest potential energy in the above sum.
\begin{rmk}
$f_0$ is not unique but the energy $d(S\diamond_{v}^{f_0}T)$ is the same for any $f_0
:E(S,v)\to Ang^0(T).$
\end{rmk}
\begin{exam}
Let us consider $S=\arbaa$ and $T=\arbab$. Here letters $a,b,c...$ are labels of
vertices,
which are of weight $1,2$ or $3$. We have:
\begin{align*}
S\diamond_{b,\lambda} T=\\
&&\arbacc&+ \arbac& +\arbaccc\\
&& +\lambda\arbaff&+ \lambda^{2}\arbae& +\lambda^{3}\arbaf
\end{align*}
\end{exam}
\begin{thm}
The partial compositions defined above yield a structure
of $\mathbb{N}^*$- current-preserving operad on the species $A\longmapsto
\mathcal{B}_A$,
denoted by $\mathcal{B}^{\lambda}$.
\end{thm}
\begin{proof}
We prove nested associativity first, and then disjoint associativity.
\begin{itemize}
\item \textbf{Nested associativity:}
Let $S,T,U$ be three weighted planar trees, let $v$ be a vertex of $S$
and $w$ be a vertex of
$T$ such that $\left|T\right|=\left|v\right|$ and
$\left|U\right|=\left|w\right|$.\\
Show $(S\diamond_{v,\lambda} T)\diamond_{w,\lambda} U=S\diamond_{v,\lambda}
(T\diamond_{w,\lambda} U)$
where $v$ is a vertex of $S$ and $w$ a vertex of $T$.\\
We have:
\begin{eqnarray*}
(S\diamond_{v,\lambda} T)\diamond_{w,\lambda} U&=&\sum_{f:E(S,v)\rightarrow
Ang(T)}{\lambda^{d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T)}(S\diamond_{v}^{f}T)\diamond_{w,\lambda}
U}\\
&\hskip -20mm=&\hskip -10mm \sum_{f:E(S,v)\rightarrow
Ang(T)}{\sum_{g:E(S\diamond_{v}^{f}T, w)\rightarrow
Ang(U)}{\lambda^{d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T)+d((S\diamond_{v}^{f}T)\diamond_{w}^{g}U)-d((S\diamond_{v}^{f}T)\diamond_{w}^{g_0
}U)}(S\diamond_{v}^{f}T)\diamond_{w}^{g}U}}\\
&=&\sum_{f:E(S,v)\rightarrow Ang(T)}{\sum_{g:E(S\diamond_{v}^{f}T,w)\rightarrow
Ang(U)}{\lambda^{A(f,g)}(S\diamond_{v}^{f}T)\diamond_{w}^{g}U}},
\end{eqnarray*}
where:
\begin{equation}
A(f,g)=d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T)+d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g}U\big)-d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g_0}U\big).
\end{equation}
Similarly we have:
\begin{eqnarray*}
S\diamond_{v,\lambda}(T\diamond_{w,\lambda}U)&=&\sum_{\wt g:E(T,w):\rightarrow
Ang(U)}{\lambda^{d(T\diamond_{w}^{\wt g}U)-d(T\diamond_{w}^{\wt
g_0}U)}S\diamond_{v,\lambda}(T\diamond_{w}^{\wt g}U)}\\
&\hskip -30mm=&\hskip -15mm \sum_{\wt f :E(S,v)\rightarrow Ang(T\diamond_{w}^{\wt
g}U)}\ {\sum_{\wt g :E(T,w)\rightarrow Ang(U)}{\lambda^{d(T\diamond_{w}^{\wt
g}U)-d(T\diamond_{w}^{\wt g_0}U)+d\big(S\diamond_{v}^{\wt f}(T\diamond_{w}^{\wt
g}U)\big)-d\big(S\diamond_{v}^{\wt f _{0}}(T\diamond_{w}^{\wt
g}U)\big)}S\diamond_{v}^{\wt f}(T\diamond_{w}^{\wt g}U)}}\\
&=&\sum_{\wt g : E(T,w)\rightarrow Ang(U)}\ {\sum_{\wt f :E(S,v)\rightarrow
Ang(T\diamond_{w}^{\wt g}U)}{\lambda^{B(\wt f,\wt g)}}S\diamond_{v}^{\wt
f}(T\diamond_{w}^{\wt g}U)},
\end{eqnarray*}
where we have set:
\begin{equation}
B(\wt f,\wt g)=d(T\diamond_{w}^{\wt g}U)-d(T\diamond_{w}^{\wt
g_0}U)+d\big(S\diamond_{v}^{\wt f}(T\diamond_{w}^{\wt
g}U)\big)-d\big(S\diamond_{v}^{\wt f_{0}}(T\diamond_{w}^{\wt g}U)\big).
\end{equation}
In order to show $(S\diamond_{v,\lambda} T)\diamond_{w,\lambda}
U=S\diamond_{v,\lambda} (T\diamond_{w,\lambda} U)$, we have to prove the following
lemma:
\begin{lem}\label{lemme emboite}
There is a natural bijection $(f,g)\longmapsto (\wt f,\wt g)$ such that
\begin{equation}\label{equation emboite}
(S\diamond_{v}^{f}T)\diamond_{w}^{g}U=S\diamond_{v}^{\wt f}(T\diamond_{w}^{\wt g}U).
\end{equation}
\end{lem}
\begin{proof}
Let $v$ be a vertex of $S$ and $w$ be a vertex of $T$ such that
$\left|T\right|=\left|v\right|$ and $\left|U\right|=\left|w\right|$.
We denoted by $Ang_w(T)$ the set of angles of $T$ issued from the vertex $w$.
Let $f:E(S,v)\rightarrow Ang(T)$ and $g:E(S\circ_{v}^{f}T,w)\rightarrow Ang(U)$ be a
two increasing functions.
We look for $\wt g:E(T,w)\rightarrow Ang(U)$ and $\wt f:E(S,v)\rightarrow
Ang(T\diamond_{w}^{\wt g}U)=Ang(U)\cup Ang(T)\backslash\{Ang_w(T)\}$
such that the equation \eqref{equation emboite} is checked.\\
Let $e$ be an edge of $T$ arriving at $w$, thus $e$ is an edge of $S\diamond_v T$
arriving at $w$. We set $\tilde{g}(e)=g(e)$. Similarly we define $\wt f$ in a unique
way:
$$\begin{array}{ccccl}
\wt f&:&E(S,v)&\longrightarrow &Ang(T\diamond_{w}^{\wt g}U)=Ang(U)\cup
Ang(T)\backslash\{Ang_w(T)\}\\
& & e &\longmapsto &\wt f(e)=\left\{\begin{array}{ccc}f(e)& if
&f(e)\notin Ang_w(T)\\g(e)&if& f(e)\in Ang_w(T)\end{array}\right.
\end{array}$$
Conversely, we assume that we have the pair $(\wt f,\wt g)$ and look for the pair
$(f,g)$ such that
equation \eqref{equation emboite} is verified.
We have $\wt f:E(S,v)\rightarrow Ang(U)\cup Ang(T)\backslash\{Ang_w(T)\}$ and $\wt
g:E(T,w)\rightarrow Ang(U).$
We then define:
$$\begin{array}{ccccl}
f&:&E(S,v)&\longrightarrow &Ang(T)\\
& & e & \longmapsto &f(e)=\left\{\begin{array}{ccc}\wt f(e)&if&\wt f(e)\notin
Ang(U)\\ w&if& \wt f(e)\in Ang(U)\end{array}\right.
\end{array}$$
and
$$\begin{array}{ccccl}
g&:&E(S\diamond_{v}^{f}T,w)&\longrightarrow&Ang(U)\\
& &e &\longmapsto
&g(e)=\left\{\begin{array}{ccl}\wt g(e)&if&e\in T\\\wt f(e)&if&e\in S\hbox{ and
}f(e)\in Ang_w(T).\end{array}\right.
\end{array}$$
\end{proof}
\textit{Proof of Theorem \ref{principal} (continued) :} To show
$(S\diamond_{v,\lambda} T)\diamond_{w,\lambda} U=S\diamond_{v,\lambda}
(T\diamond_{w,\lambda} U)$,
it remains to show the equality $A(f,g)=B(\wt f,\wt g)$. We set:
\begin{eqnarray*}
A'(f,g)&=&A(f,g)+d((S\diamond_{v}^{f_0}T)\diamond_{w}^{g_0}U),\\
B'(\wt f,\wt g)&=&B(\wt f,\wt g)+d(S\diamond_{v}^{\wt f_{0}}(T\diamond_{w}^{\wt
g_{0}}U)),\\
\epsilon(f)&=&d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T),\\
\epsilon(\wt g)&=&d(T\diamond_{w}^{\wt g}U)-d(T\diamond_{w}^{\wt g_0}U).
\end{eqnarray*}
We have:
\begin{equation}
\epsilon(f)=\sum_{e\in E(S,v)}{h\big(f(e)\big).\left|B_e\right|},
\end{equation}
where $h\big(f(e)\big)$ is the distance between $f(e)$ and the root of $T$ in the
new tree $S\diamond_{v}^{f}T$
and $\left|B_e\right|$ is the weight of the branch above $e$. Similarly:
\begin{eqnarray*}
d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g_0}U\big)-d\big((S\diamond_{v}^{f_0}T)\diamond_{w}^{g_0}U\big)&=&\sum_{e\in
E(S,v)}{h\big(f(e)\big)\left|B_e\right|}\\
&=&\epsilon(f).
\end{eqnarray*}
Here $g_0$ is not involved because it was connected to the root of $U$, so
$A'(f,g)=d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g}U\big)$.
By the same computation with $\wt g$ instead of $f$ we
show $B'(\wt f, \wt g)=d\big(S\diamond_{v}^{\wt f}(T\diamond_{w}^{\wt g}U)\big)$. So by
Lemma \ref{lemme emboite} we have $A'(f,g)=B'(\wt f, \wt g)$ that is to say
$A(f,g)+d\big((S\diamond_{v}^{f_0}T)\diamond_{w}^{g_0}U\big)=B(\wt f, \wt
g)+d\big(S\diamond_{v}^{\wt f_0}(T\diamond_{w}^{\wt g_0}U)\big)$,
which proves that $A(f,g)=B(\wt f, \wt g)$ because by Lemma \ref{lemme emboite} we have
$d\big((S\diamond_{v}^{f_0}T)\diamond_{w}^{g_0}U\big)=d\big(S\diamond_{v}^{\wt
f_0}(T\diamond_{w}^{\wt g_0}U)\big).$\\
\item \textbf{Disjoint associativity:} let $v,w$ be two disjoint vertices of $S$
such that $\left|v\right|=\left|T\right|$ and $\left|w\right|=\left|U\right|$, show
that:
\begin{equation}\label{equation disjointe}
(S\diamond_{v,\lambda} T)\diamond_{w,\lambda} U=(S\diamond_{w,\lambda}
U)\diamond_{v,\lambda} T.
\end{equation}
We have
\begin{eqnarray*}
(S\diamond_{v,\lambda} T)\diamond_{w,\lambda} U&=&\sum_{f:E(S,v)\rightarrow
Ang(T)}{\lambda^{d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T)}(S\diamond_{v}^{f}T)\diamond_{w,\lambda}
U}\\
&=&\sum_{f:E(S,v)\rightarrow Ang(T)}{\sum_{g:E(S\diamond_{v}^{f}T,w)\rightarrow
Ang(U)}{\lambda^{k(f)+d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g}U\big)-d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g_0}U\big)}}(S\diamond_{v}^{f}T)\diamond_{w}^{g}U}\\
&=& \sum_{f:E(S,v)\rightarrow Ang(T)}{\sum_{g:E(S\diamond_{v}^{f}T,w)\rightarrow
Ang(U)}{\lambda^{C(f,g)}}(S\diamond_{v}^{f}T)\diamond_{w}^{g}U},
\end{eqnarray*}
where:
\begin{eqnarray*}
k(f)&=&d(S\diamond_{v}^{f}T)-d(S\diamond_{v}^{f_0}T),\\
C(f,g)&=&k(f)+d\big((s\diamond_{v}^{f}T)\diamond_{w}^{g}U\big)-d\big((S\diamond_{v}^{f}T)\diamond_{w}^{g_0}U\big).
\end{eqnarray*}
Similarly we find:
$$(S\diamond_{w,\lambda} U)\diamond_{v,\lambda}T=\sum_{\wt g:E(S,w)\rightarrow
Ang(U)}{\sum_{\wt f:E(S\diamond_{w}^{\wt g}U,v)\rightarrow Ang(T)}{\lambda^{D(\wt f,
\wt g)}}(S\diamond_{w}^{\wt g}U)\diamond_{v}^{\wt f}T},$$
where:
\begin{eqnarray*}
D(\wt f,\wt g)&=&k(\wt g)+d\big((S\diamond_{w}^{\wt g}U)\diamond_{v}^{\wt
f}T\big)-d\big((S\diamond_{w}^{\wt g}U)\diamond_{v}^{\wt f_{0}}T\big), \hbox{
with}\\
k(\wt g)&=&d(S\diamond_{w}^{\wt g}U)-d(S\diamond_{w}^{\wt g_0}U).
\end{eqnarray*}
In order to prove \eqref{equation disjointe}, we need the following lemma:
\begin{lem}\label{lemme disjointe}
We have a natural bijection $(f,g)\longmapsto (\wt f,\wt g)$ such that
\begin{equation}\label{equa}
(S\diamond_{v}^{f}T)\diamond_{w}^{g}U=(S\diamond_{w}^{\wt g}U)\diamond_{v}^{\wt f}T.
\end{equation}
\end{lem}
\begin{proof}
Let $f:E(S,v)\rightarrow Ang(T)$ and $g:E(S\diamond_{v}^{f}T,w)\rightarrow
Ang(U)$. We look for $\wt g:E(S,w)\rightarrow Ang(U)$ and
$\wt f:E(S\diamond_{w}^{\wt g}U,v)\rightarrow Ang(T)$ such that the
Equation \eqref{equa} is verified. Let $\wt g$ be the restriction of $g$
on the edges $e$ from $S$ and $\wt f=f$. Here $E(S\diamond_{w}^{g}U,v)=E(S,v)$
because the vertices $v$ and $w$ are disjoint.
\end{proof}
\textit{Proof of Theorem \ref{principal} (end):} Thus to show disjoint
associativity, it remains to show that for any pair $(f,g)$ and $(\wt f,\wt g)$ we
have $C(f,g)=D(\wt f, \wt g)$. We set
$C'(f,g)=C(f,g)+d\big( (S\diamond_{v}^{f_0}T)\diamond_{w}^{g_0}U\big)$ and $D'(\wt
f,\wt g)=D(\wt f,\wt g)+d\big((S\diamond_{w}^{\wt g_0}U)\diamond_{v}^{\wt
f_0}T\big)$.
We have:
\begin{eqnarray*}
k(\wt g)&=&d(S\diamond_{w}^{\wt g}U)-d(S\diamond_{w}^{\wt g_0}U)\\
&=&\sum_{e\in E(S,w)}h\big(\wt g(e)\big).|B_e|,
\end{eqnarray*}
where $h\big(\wt g(e)\big)$ is the distance between the root of $U$ and $\wt g(e)$
in the new tree $S\diamond_w^{\wt g}U$.
We also have:
\begin{equation*}
d\big((S\diamond_w^{\wt g}U)\diamond_v^{\wt f_0}T\big)-d\big((S\diamond_w^{\wt
g_0}U)\diamond_v^{\wt f_0}T\big)=k(\wt g)
\end{equation*}
because we changed the vertex $v$ by a tree of the same weight, and because $\wt
f_0$ is defined by grafting onto the root. This proves:
\begin{equation*}
D'(\wt f, \wt g)=d\big((s\diamond_w^{\wt g}U)\diamond_v^{\wt f}T\big).
\end{equation*}
By Lemma \ref{lemme disjointe} again, $C'(f,g)=D'(\wt f,\wt g)$ and $C(f,g)=D(\wt
f,\wt g)$,
which proves disjoint associativity.
\end{itemize}
The partial compositions defined on $\mathcal{O}^{\lambda}$ hence verify the
axioms of a current-preserving operad.
\end{proof}
\begin{rmk}
With the notations of Sec \ref{ordinaryoperad}, the current-preserving
operad $\mathcal{B}^\lambda$ is naturally associated with $({\mathcal{B}^0})^{\mathbb{N}^*}$
for $\lambda=0$ and with $Brace^{\mathbb{N}^*}$ for $\lambda=1$.
\end{rmk}
\begin{thm} The map $\varphi$ introduced in Definition \ref{symm} is a
morphism of current-preserving operad from $\mathcal{O}^\lambda$ to $\mathcal{B}^\lambda$ i.e:
$$\varphi(S\circ_{v,\lambda}T)=\varphi(S)\diamond_{v,\lambda}\varphi(T).$$
\end{thm}
\begin{proof}
By definition of partial compositions on planar trees and non-planar trees,
and as the shuffle product permute the branches in all possible ways, then we verify that $\varphi$ is a current-preserving operad morphism. The key point is the following: for any $f:E(S,v)\to v(T)$, choosing a planar representative of $S\circ_{v}^{f}T$ amounts to choosing planar representatives $\tilde{S}$ and $\tilde{T}$ of $S,T$ respectively, together with an increasing map $\tilde{f}: E(\tilde{S},v)\to Ang(\tilde{T})$ above $f$, i.e such that the following diagram commutes :
\diagramme{
\xymatrix{E(\tilde{S},v)\ar[r]^{\tilde{f}}\ar[dr]_f & Ang(\tilde{T})\ar@{->>}[d]\\
&v(T)
}
}
\end{proof}
{\bf{Acknowledgments:}} I would like to thank Dominique Manchon for valuable discussions and comments.
|
\subsection{Modeling the system noise}
\label{21cm_Pk_noise}
We now compute the sensitivity of 21 cm power spectrum that will possibly be achieved for the telescopes such as SKA1-low and SKA1-mid. We will assume that the only noise which contributes to the measurement is the system noise and sample variance, and we will ignore effects arising from astrophysical foregrounds, variations in the ionosphere, radio-frequency interference etc. Accounting for these effects is essential for detecting the 21 cm signal, however, we make the assumption that they can be separately identified and removed from the data before comparing with theoretical models. The $1\sigma$ error in the power spectrum for a single mode $\vec{k}$ arising from the system temperature can be written as (see Appendix A of \cite{geil} or \cite{mcquinn} for the detailed derivation)
\begin{equation}
\delta P_{N}(\vec{k},\nu)=\frac{T_{\rm sys}^2}{B t_0} \left ( \frac{\lambda^2}{A_e} \right)^2 \frac{r_{\nu}^2 L}{n_b(\vec{U},\nu)},
\end{equation}
where $\nu$ and $\lambda$ are the observing frequency and wavelength respectively. The quantities $T_{\rm sys}$, $B$ and $t_0$ denote the system temperature of the instrument, total frequency bandwidth and observation time, respectively. The effective collecting area of an individual antenna is denoted as $A_e$ which in turn can be written as $A_e=\epsilon A$, where $\epsilon$ and $A$ are the antenna efficiency and physical collecting area respectively. The comoving distance to an observer at redshift $z$, corresponding to an observing frequency $\nu$, is denoted as $r_{\nu}$, while $L$ is the comoving length associated to the bandwidth $B$. The number density of baselines $\vec{U}$ is written as $n_b(\vec{U},\nu)$. The baseline vector $\vec{U}$ is related to $\vec{k}$ by the relation $\vec{U}=\vec{k}_{\perp}r_{\nu}/2 \pi$, where $\vec{k}_{\perp}$ is the component of $\vec{k}$ perpendicular to the line of sight. We write $n_b(\vec{U},\nu)$ as
\begin{equation}
n_b(\vec{U},\nu)=\frac{N(N-1)}{2}\rho_{2D}(\vec{U},\nu),
\end{equation}
where $N$ is the total number of antennae and $\rho_{2D}(\vec{U},\nu)$ is the two-dimensional normalized baseline distribution function which follows the condition $\int_0^{\infty} U d U \int_{0}^{\pi} d\phi ~\rho_{2D}(\vec{U},\nu)=1$.
We now assume that the baseline distribution is circularly symmetric, i.e, it is only a function of $U=|\vec{U}|$. This assumption simplifies the calculations considerably and holds true for many cases of interest. The normalized baseline distribution $\rho_{2D}(U,\nu)$ can be calculated for a given antenna distribution $\rho_{ant}(l)$ using the relation \cite{datta,petrovic,geil}
\begin{equation}
\rho_{2D}(U,\nu)=B(\nu) \int_0^{\infty} 2 \pi l d l \, \rho_{\rm ant}(l) \int_0^{2\pi} d \phi~ \rho_{\rm ant}(|\vec{l}-\lambda \vec{U}|),
\label{eq:rhoant2rho2d}
\end{equation}
where $|\vec{l} - \lambda \vec{U}| = \sqrt{l^2 + \lambda^2 U^2 - 2 l \lambda U \cos \phi}$ and the constant $B(\nu)$ is determined from the above normalization condition for a given frequency $\nu$.
Now we consider a three-dimensional cell in momentum-space having coordinates between $k$ to $k+dk$ and $\theta$ to $\theta+d \theta$, with $\theta$ being the angle between $\vec{k}$ and the line of sight. Clearly, the perpendicular component is given by $k_{\perp}=k \sin \theta$ while the parallel component is $k_{\parallel}=k \cos \theta$. If we average the power spectrum over all the modes which lie in the range $[k,k+dk]$ and $[\theta,\theta+d \theta]$, the error in the power spectrum is reduced to
\begin{equation}
\delta P_{N}(k,\theta)=\frac{T_{sys}^2}{B t_0} \left ( \frac{\lambda^2}{A_e}\right)^2 \frac{r_{\nu}^2 L}{[N(N-1)/2]\, \rho_{2D}(U)} \frac{1}{\sqrt{N_m(k,\theta)}}.
\label{eq:sigmap}
\end{equation}
Note that we have omitted the frequency $\nu$ in the above expression, i.e., the frequency dependence is implicit. The quantity $N_m(k,\theta)$ is the number of independent modes between $[k, k+dk]$ and $[\theta, \theta+d \theta]$ and can be written as
\begin{equation}
N_m(k,\theta)=\frac{2 \pi k^2 dk \sin \theta d \theta}{V_{\rm one-mode}},
\end{equation}
where
\begin{equation}
V_{\rm one-mode}=\frac{(2 \pi)^3 A}{r_{\nu}^2 L \lambda^2}
\end{equation}
is the volume for a single independent mode in $k$-space.
The noise error is a function of both $k$ and $\theta$, or equivalently, of both $k_{\perp}$ and $k_{\parallel}$. To calculate the noise error for the spherically averaged power spectrum we need to average over $\theta$ for a fixed $k$. Since for a fixed $k$ the noise error varies with $\theta$, simple averaging is not optimum. We rather use an inverse-variance weighting scheme for averaging so as to minimize the noise error in the power spectrum. Under this scheme the noise error can be written as
\begin{equation}
\delta P_{N}(k)=\left [ \sum_{\theta}\frac{1}{\delta P^2_{N}(k, \theta)} \right]^{-1/2}~.
\label{eq:inv-variance}
\end{equation}
In the continuum limit, the system temperature error in the 21 cm power spectrum can be written as \cite{geil,mcquinn}
\begin{equation}
\delta P_{N}(k)=\frac{T_{sys}^2}{B t_0} \left ( \frac{\lambda^2}{A_e}\right)^2 \frac{r_{\nu}^2 L}{[N(N-1)/2]\, \rho_{3D}(k)} \frac{1}{\sqrt{N_k}},
\label{eq:sigmap2}
\end{equation}
where $N_k$ is number of observable modes in the spherical shell between $k$ and $k+dk$ and can be calculated as
\begin{equation}
N_k=\frac{2 \pi k^2 dk}{V_{\rm one-mode}}.
\end{equation}
The quantity $\rho_{3D}(k,\nu)$ is a distribution of measurements in 3D $k$-space which is related to the 2D normalized baseline distribution as
\begin{equation}
\rho_{3D}(k)=\left [ \int_{0}^{\pi/2} d\theta \sin \theta \, \rho_{2D}^2 \left ( \frac{r k}{2 \pi} \sin \theta \right ) \right ]^{1/2}.
\label{eq:2dto3d}
\end{equation}
In the above expression, we have replaced the baseline $U$ by $ r_{\nu} k \sin \theta/2 \pi$. Although Eq. \ref{eq:sigmap2} is an expression for the noise error in the continuum limit, it gives an accurate estimate of the noise error even when the measurements are in discrete points \cite{jensen13}.
Finally, the 21 cm power spectrum error arising from sample variance can be written as:
\begin{equation}
\delta P_{SV}(k)=\left[ \sum_\theta \frac{N_m(k,\theta)}{P^2_{\rm 21cm}(k,\theta)} \right]^{-1/2}
\end{equation}
Notice that for convenience we have isolated both contributions to the total 21 cm power spectrum error (system temperature and sample variance) to investigate the scales at which each of them are most important.
\subsection{Detectability of the 21 cm signal}
\label{detectability}
We now address the detectability of the 21 cm power spectrum by the future SKA radio telescope. The SKA will have three different instruments working at different frequency bands and will be built in two phases. Here we consider two instruments called the SKA1-mid and SKA1-low which will be built in phase 1 in South Africa and Australia respectively. The SKA instrument specifications such as the antenna distribution, total number of antennae, total collecting area, frequency coverage, etc., have not been completely finalized yet and are possibly subject to change. Here we use the specifications described in one of the most recent documents, i.e., the `Baseline Design Document'\footnote{\tt http://www.skatelescope.org/wp-content/uploads/2012/07/SKA-TEL-SKO-DD-001-1\_BaselineDesign1.pdf} which is available on the SKA website\footnote{\tt https://www.skatelescope.org/home/technicaldatainfo/key-documents/}. Let us briefly summarize the main properties of the two instruments which we have considered in this work:
{\bf SKA1-mid}: Based on the above document we assume that the SKA1-mid will cover a frequency range from $350$ MHz to $14$ GHz and will have a total of $250$ antennae of $15$ meters diameter each. This also includes $60$ antennae from the MeerKAT instrument. We use the baseline density given in the above document in (violet line in their Fig. 10) to calculate the normalized baseline distribution function $\rho_{2D}(U)$. Note that this baseline distribution is consistent with the proposed antenna distribution with $40\%$, $54\%$, $70\%$, $81\%$ and $100\%$ of the total antennae being within $0.4$ km, $1$ km, $2.5$ km, $4$ km and $100$ km radius respectively (see table 6 in the above document). The baseline density given in the document is essentially the total number of baselines measurements lying in the annulus, i.e., it is proportional to $U\rho_{2D}(U)$ for a fixed annulus size. We use the normalization condition $\int_0^{\infty} U d U \int_{0}^{\pi} d\phi \rho_{2D}(\vec{U},\nu)=1$ to calculate the amplitude of the normalized baseline distribution function $\rho_{2D}(U)$. We then use Eq. \ref{eq:2dto3d} to compute $\rho_{3D}(U,\nu)$. We plug everything in Eq. \ref{eq:sigmap2} which gives the required noise error in the 21 cm power spectrum. We assume $T_{\rm sys}$ to be $30$K for the redshifts $z=2.4$ and $3$, corresponding to frequencies of $417.6$ MHz and $355$ MHz respectively. We also consider $100$ hours of observations over $32$ MHz bandwidth and a typical antenna efficiency equal to $0.7$.
We show the system noise for SKA1-mid in the top panels of Fig. \ref{Pk_21cm}. On comparing the noise estimates with the amplitude of the 21 cm power spectra at $z = 2.4$ and $3.6$, we find that the SKA1-mid will be able to resolve the 21 cm power spectrum up to $k\sim 1 h~ {\rm Mpc}^{-1}$ with 100 hours of observations provided the HI distribution is modeled using the halo-based model 1 or the particle-based method. In case the HI distribution is modeled with the halo-based model 2 we find that the prospect of detecting the 21 cm power spectrum will be much better. One can detect it for relatively smaller scales: $k\sim 3 h~{\rm Mpc}^{-1}$, and even for larger scales, we expect to achieve a much better signal-to-noise ratio.
{\bf SKA1-low}: The proposed SKA1-low will operate in the frequency range $50-300$ MHz and will consist of $911$ stations with a diameter of $35$ meters each. The SKA1-low has total collecting area of $0.88 \, {\rm km^2}$ and thus is much more powerful compared to the SKA1-mid which has a collecting area of only $0.044 \, {\rm km^2}$. Although the maximum baseline for SKA1-low is expected to be $\sim 100$ km, most of the stations will be in the centre (see figure 3 \& 4 in the baseline design document). We find that the antenna distribution can be nicely approximated by a simple functional form $\rho_{ant}(l)=(A/l) \exp[-0.5(l/1000 \, m)^2]$ with no antenna within a $50$ m radius and $866$ stations within $5$ km radius. We ignore the larger baselines as their contribution to the sensitivity is negligible. We note however that the large baselines will be useful for accurate measurements of foreground sources and for removing them from the observed data. Given the antenna distribution, we can use Eq. \ref{eq:rhoant2rho2d} and the normalization condition to calculate the normalized baseline distribution function $\rho_{2D}(U)$. We then use the procedure described above to calculate the noise error in the power spectrum at redshift $z=4$ for SKA1-low. We assume $T_{\rm sys}$ to be $110$K for the redshift $z=4$, corresponding to a frequency of $284$ MHz. As before, we assume the antenna efficiency to be $0.7$.
In the bottom-left panel of Fig. \ref{Pk_21cm} we show the SKA1-low system noise for $100$ hours of observations with a $32$ MHz bandwidth. We find that the 21 cm power spectrum will be detected by this telescope up to much smaller scales, i.e., $k\sim 5 h~{\rm Mpc}^{-1}$ for the halo-based 1 and particle-based models, while it can be detected up to $k\sim 20 h~{\rm Mpc}^{-1}$ for halo-based model 2. It is thus clear that SKA1-low, as per current specifications, would be quite sensitive for studying the HI power spectrum at $z \approx 4$.
The different panels of Fig. \ref{Pk_21cm} also show the sample variance errors in the 21 cm power spectrum. On small scales, we find that errors arising from sample variance are much below those coming from the instrument noise, whereas they dominate on large scales. Although sample variance errors reduce the signal to noise ratio on which the 21 cm power spectrum can be determined on large scales, our conclusions do not change: the largest scales that we can probe with our simulations will be detected both by SKA1-mid and SKA1-low at the redshift studied here with 100 hours of observations.
Interestingly, for all the models studied in this paper, the 21 cm power spectra would be detectable at $k < 1 h~{\rm Mpc}^{-1}$ with $\sim 100$ hours of observations for a wide range of redshifts $2.4 < z < 4$ using the SKA1 telescopes. Since the signal at these scales essentially traces the dark matter fluctuations (see Fig. \ref{Pk_21cm}), it might be possible constrain cosmological parameters with future telescopes. We plan to study these issues in a different paper.
\section{21 cm signal}
\label{21cm_section}
We now investigate the redshifted 21 cm signal generated by the distribution of neutral hydrogen and its detectability with the future telescopes like the SKA. We split this section into three subsections: In subsection \ref{21cm_Pk_subsection} we describe the method used to compute the 21 cm power spectrum. Our estimates of the noise expected in the 21 cm power are presented in subsection \ref{21cm_Pk_noise}. Finally, the detectability of the 21 cm power spectrum with the future SKA1-low and SKA1-mid telescopes is discussed in subsection \ref{detectability}.
\subsection{21 cm power spectrum from simulations}
\label{21cm_Pk_subsection}
The 21 cm power spectrum is computed from the spatial distribution of neutral hydrogen which in turn is obtained from the hydrodynamical simulations using the halo-based and particle-based methods. In real space, the brightness temperature excess due to the 21 cm emission from neutral hydrogen located in the real-space coordinate $\vec{r}$ and having a HI density $\rho_{\rm HI}(\vec{r})$ is given by \cite{Furlanetto_2006,Mao_2012}
\begin{equation}
\delta T_b(\nu)=\overline{\delta T_b}(z)\left(\frac{\rho_{\rm HI}(\vec{r})}{\bar{\rho}_{\rm HI}}\right)
\left[ 1-\frac{T_\gamma(z)}{T_s(\vec{r})}\right],
\end{equation}
where
\begin{equation}
\overline{\delta T_b}(z)=23.88~\bar{x}_{\rm HI}\left( \frac{\Omega_{\rm b}h^2}{0.02}\right)\sqrt{\frac{0.15}{\Omega_{\rm m}h^2}\frac{(1+z)}{10}}~{\rm mK},
\end{equation}
$\bar{\rho}_{\rm HI}$ is the mean density of neutral hydrogen and $\bar{x}_{\rm HI}=\bar{\rho}_{\rm HI}/\bar{\rho}_{\rm H}$ is the average neutral hydrogen fraction.
The quantity $T_\gamma(z)$ is the CMB temperature at redshift $z$ and $T_s$ is the spin temperature characterizing the relative population of HI atoms in different states.
Note that we have not accounted for the fact that the observed frequency not only depends on the cosmological redshift, but also on the peculiar velocity of the HI gas along the line of sight of observation. Since the radio telescope measurements will probe the signal in redshift-space, it is essential that the above equation is re-written in terms of the redshift space coordinates. If we assume the gas to be optically thin, and the spin temperature to be much larger than the CMB temperature ($T_s\gg T_\gamma$), the brightness temperature excess in redshift-space can be expressed as \cite{Mao_2012}
\begin{equation}
\delta T_b^s(\nu)=\overline{\delta T_b}(z)\left[\frac{\rho_{\rm HI}(\vec{s})}{\bar{\rho}_{\rm HI}}\right],
\label{delta_Tb}
\end{equation}
where the superscript $s$ in $\delta T_b$ indicates that the quantity is calculated in redshift-space and $\vec{s}$ is the redshift-space coordinate corresponding to $\vec{r}$. The relation between $\vec{s}$ and $\vec{r}$ is given by
\begin{equation}
\vec{s} = \vec{r} + \frac{1+z}{H(z)} \vec{v}_{\parallel}(\vec{r}),
\end{equation}
where $z$ should be interpreted as the redshift of observation and $\vec{v}_{\parallel}$ is the component of peculiar velocity along the line of sight. The 21 cm power spectrum in redshift space is then defined as
\begin{equation}
P^s_{\rm 21cm}(k)=\langle \delta T_b^s(\vec{k}) (\delta T_b^s)^*(\vec{k}) \rangle,
\label{Pk_21cm_s}
\end{equation}
where $\delta T_b^s(\vec{k})$ is simply the Fourier transform of the brightness temperature excess.
We should mention here that the assumption $T_s \gg T_\gamma$ holds in the redshifts of our interest as the spin temperature is usually coupled to the gas kinetic temperature which, in turn, is expected to be much higher than $T_\gamma$. The other approximation we have made, i.e., the assumption that the gas is optically thin, is also a good approximation because of the forbidden nature of the hyperfine transition. There could be some regions (e.g., highly overdense regions) where the peculiar velocity effects can make the optical depth quite large in redshift space \cite{Mao_2012}. However, the fraction of such regions is expected to be quite small and hence we ignore them in this work.
Given the above formalism we compute the 21 cm power spectrum by assigning neutral hydrogen to the gas particles of simulation $\mathcal{B}60$ using the three different methods investigated in this paper. We choose simulation $\mathcal{B}60$ because it allows us to study the 21 cm power spectrum for sufficiently low values of $k$ (i.e., large scales) which can be accessed by telescopes like the SKA. We should, however, keep in mind that the HI power spectrum is converged only for the halo-based models (see Figs. \ref{Pk_HI_Bagla} and \ref{Pk_HI_Paco}), while the convergence is not satisfactory for the particle-based method (see Fig. \ref{Pk_HI_Dave}).
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/Pk_21cm.pdf}\\
\end{center}
\caption{21 cm power spectrum in redshift-space. We compute the dimensionless 21 cm power spectrum, $\bigtriangleup^2_{\rm 21cm}(k)=k^3P_{\rm 21cm}(k)/2\pi^2$, by
assigning HI to the gas particles of the simulation $\mathcal{B}60$ using the halo-based models 1 and 2 (solid blue and dashed green respectively) and the particle-based method (red dot-dashed) at $z=2.4$ (top-left), $z=3$ (top-right) and $z=4$ (bottom-left). The black stars and black triangles represents the expected SKA-1 mid and SKA-1 low system temperature noises for an observation time of 100 hours in a 32 MHz bandwidth and for intervals in $k$ with a width of $dk=k/5$ (see text for further details). The solid, dashed and dot-dashed black lines represent the sample variance noise for the halo-based model 1, the halo-based model 2 and the particle-based model, respectively. The vertical lines display the position of the Nyquist frequency. Below each panel we show the 21 cm bias defined as $b^2_{\rm 21cm}(k)=\bigtriangleup^2_{\rm 21cm}(k)/\bigtriangleup^2_{\rm m}(k)$ where $\bigtriangleup^2_{\rm m}(k)$ is the dimensionless matter power spectrum in real-space.}
\label{Pk_21cm}
\end{figure}
The plots for the 21 cm power spectrum are shown in Fig. \ref{Pk_21cm}. We find that the power spectrum obtained using the halo-based model 1 and the particle-based method are almost identical. This follows from the fact that the HI power spectra obtained from these two methods are quite similar as can be seen from Fig. \ref{Pk_HI_filaments}. We however note that the values of $\Omega_{\rm HI}$ from the two cases are different, with the particle-based method having a lower value. This implies that the HI in the particle-based method is more strongly clustered than that of the halo-based model 1. Should both methods be normalized to the same value of $\Omega_{\rm HI}$, the HI power spectrum from the particle-based method would have a higher amplitude. The 21 cm power spectrum obtained by using the halo-based model 2 has a significantly higher amplitude than those obtained by employing the other two methods. This is because of the much stronger HI clustering in that model, which is required to match the bias measurements of DLAs \cite{Font_2012}.
In the bottom of each panel of Fig. \ref{Pk_21cm} we display the 21 cm bias: $b^2_{\rm 21cm}(k)=P_{\rm 21cm}(k)/P_{\rm m}(k)$, where $P_{\rm m}(k)$ is the matter power spectrum in real-space whereas the 21 cm power spectrum is computed in redshift-space. Interestingly, on large scales ($k \lesssim 0.5 h~{\rm Mpc}^{-1}$), the slope of both the 21 cm power spectra and the 21 cm bias is the same in all the three models. This is because the HI power spectrum at large scales essentially traces the underlying dark matter fluctuations, the only difference being given in terms of the (scale-independent) linear bias parameter. Different prescriptions for generating the HI distribution only result in different values of this bias. Hence, it should be possible to constrain the dark matter power spectrum at large-scales using the 21 cm power spectrum signal. Surprisingly, we find that in the fully non-linear regime, the 21 cm bias exhibits a very weak scale dependence.
\section*{Acknowledgements}
We thank Paramita Barai, Aritra Basu, Simeon Bird, Jayaram Changalur, Federica Govoni, Tae-sun Kim,
and Narendra Nath Patra for useful discussions. We thank Jamie Bolton for help with VPFIT and for having provided
us with the reference thermal IGM model.
Calculations were performed on SOM2 and SOM3 at IFIC and
on the COSMOS Consortium supercomputer within the DiRAC
Facility jointly funded by STFC, the Large Facilities Capital Fund of BIS and the
University of Cambridge, as well as the Darwin Supercomputer of the
University of Cambridge High Performance Computing Service (http://
www.hpc.cam.ac.uk/), provided by Dell Inc. using Strategic Research
Infrastructure Funding from the Higher Education Funding Council for
England. FVN and MV are supported by the ERC Starting Grant
``cosmoIGM'' and partially supported by INFN IS PD51 "INDARK".
KKD thanks the Department of Science \& Technology (DST), India for the research grant SR/FTP/PS-119/2012 under the Fast Track Scheme for Young Scientist.
We acknowledge partial support from "Consorzio per la Fisica - Trieste".
\section{Column density distribution: computation}
\label{column_density_appendix}
Here we describe in detail the method used to compute the column density distribution, $f_{\rm HI}$, from the N-body simulations. Once the neutral hydrogen has been assigned to the gas particles, using any of the methods investigated in this paper, the value of the column density along an arbitrary line of sight can be computed using the physical properties of the gas particles: mass, HI/H fraction and SPH smoothing length.
We start by projecting all the gas particle positions onto the XY plane (we have checked that results do not change if the projection is performed onto a different plane). We then draw lines of sights
(LOS) from a regular grid in the XY plane that go from $Z=0$ to $Z=L$, where $Z$ is the cartesian coordinate, perpendicular to the XY plane, and $L$ is the size of the simulation box. For any given LOS we compute the minimum distance between any gas particle and the LOS, $b$. If that distance is smaller than the gas particle smoothing length, $h^i$, then we integrate its density along the path that the LOS intersects the physical size of the gas particle (see Fig. \ref{integration}):
\begin{figure}
\begin{center}
\includegraphics[width=0.3\textwidth]{Figs/Figures.pdf}\\
\end{center}
\caption{Scheme showing a gas particle together with its boundary (circle black line). The SPH density has to be integrated along the orange line.}
\label{integration}
\end{figure}
\begin{equation}
N_{\rm HI}^i=\frac{0.76\left(\frac{\rm HI}{H}\right)^i}{m_H}\int_{-l_{max}}^{+l_{max}}\rho^i(r)dl=2\frac{0.76\left(\frac{\rm HI}{H}\right)^i}{m_H}m^i\int_{0}^{l_{max}}W(r,h^i)dl\, ,
\end{equation}
where $N_{\rm HI}^i$ is the column density due to the particle $i$, having mass $m^i$, neutral hydrogen fraction ${(\rm HI/H})^i$ and SPH smoothing length $h^i$. $m_H$ is the mass of the hydrogen atom. The relation between the integration variable $l$ and the radius $r$ is given by $r^2=b^2+l^2$, with $l_{max}^2=(h^i)^2-b^2$. For each LOS, we sum the $N_{\rm HI}$ of all the gas particles contributing to it. We repeat the procedure for all the LOS and finally we compute HI column density distribution function as:
\begin{equation}
f_{\rm HI}(N_{\rm HI})=\frac{d^2n(N_{\rm HI})}{dN_{\rm HI}dX}\,,
\end{equation}
where $n(N_{\rm HI})$ is the number of lines with column densities equal to $N_{\rm HI}$ and $dX=H_0(1+z)^2/H(z)dz$ is the absorption distance.
Since we compute the column density along a LOS that spans along the whole box, by using this method we are implicitly assuming that the column density of a given LOS is due to a single absorber. We have tested the validity of this assumption by dividing the simulation box into $N$ slices of width $L/N$ and computing the column densities along a grid of LOS that spans within any of those slices. In other words, for given LOS that goes from $Z=0$ to $Z=L$, we compute the column density for $N$ LOS that span among $[Z=0,Z=L/N]$, $[Z=L/N,Z=2L/N]$ and so on.
The distribution of the column densities for the different values of $N$ are shown in Fig. \ref{f_HI_convergence_test} when the HI is assigned used the halo-based model 1. We find that for column densities larger than $\sim~10^{19}$ cm$^{-2}$ the column density distribution is insensitive to the absorption distance of the LOS used to compute the values of the column densities. We have explicitly checked that by assigning the HI using the other two methods the column density distribution is also converged above $\sim~10^{19}$ cm$^{-2}$.
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{Figs/f_HI_Bagla_15Mpc_convergence_test.pdf}
\includegraphics[width=0.49\textwidth]{Figs/f_HI_Bagla_60Mpc_convergence_test.pdf}\\
\end{center}
\caption{Impact of the absorption distance on the column density distribution. We show the results of computing the column density distribution using lines of sights that span along the entire simulation box (solid blue lines) and line of sights that span along $1/30$ (dashed green) and $1/100$ (dash-dotted red) of the box size for the simulations $\mathcal{B}15$ (left) and $\mathcal{B}60$ (right) at $z=3$. The residuals are shown in the bottom panels. We have used the halo-based model 1 to assign the neutral hydrogen to the gas particles.}
\label{f_HI_convergence_test}
\end{figure}
We thus conclude that this method produces converged results for high column densities values ($N_{\rm HI}\gtrsim10^{19}$cm$^{-2}$). Notice that the advantage of this way of computing the column density distribution is that the calculations are very fast, allowing us to compute the column density of a very large number of LOS.
\section{Summary and conclusions}
\label{Conclusions}
The aim of this paper is to model the neutral hydrogen content of the Universe and investigate its detectability by future radio telescopes like the SKA. We have run high-resolution hydrodynamical N-body simulations using the code {\sc GADGET-III} and we have modeled the distribution of HI using two different techniques: the \textit{halo-based method} and the \textit{particle-based method}.
The halo-based method is built on the assumption that all the HI in the Universe resides within dark matter halos. According to this scheme, the neutral hydrogen mass assigned to a particular gas particle residing in a given dark matter halo depends on the halo HI mass, $M_{\rm HI}(M)$ and on the density profile of the HI in the halo $\rho_{\rm HI}(r|M)$. For simplicity, we have neglected any dependence of the above quantities on the environment.
We have investigated two different models that rely on this methodology: the halo-based model 1 and the halo-based model 2. Each model uses a different $M_{\rm HI}(M)$ and $\rho_{\rm HI}(r|M)$ functions. In particular, the halo-model 1 uses a simple prescription for the function $M_{\rm HI}(M)$ \cite{Bagla_2010} based on the observations whereas the halo-based model 2 is constructed to reproduce the recent estimate of DLAs bias obtained by SDSS-III/BOSS \cite{Font_2012}. We find that both models reproduce very well the DLAs column density distribution, although they over-predict the abundance of Lyman Limit Systems. In terms of the HI power spectrum our results indicate that the amplitude of the HI power spectrum is significantly higher in the halo-based model 2. The reason is that model has been constructed to reproduce the DLAs bias measurements \cite{Font_2012} and therefore the HI in that model is much more strongly clustered than the HI spatial distribution obtained by employing the halo-based model 1. Overall, these two models, whose primarily difference is in the bias between DLAs and matter, bracket a conservative and physical range for assigning neutral hydrogen to dark matter halos.
In the particle-based method, we instead do not make any assumption on the location of the neutral hydrogen: HI (and H$_2$ for star forming particles) is assigned to all gas particles in the simulation according to their physical properties. In this scheme, we first compute the HI/H fraction associated to each gas particle assuming photo-ionization equilibrium with the external UV background. Next, the HI/H fraction is corrected to account for
self-shielding effects and finally, for star forming particles, a further correction to the HI/H fraction is carried out to account for molecular hydrogen. This method not only predicts the spatial distribution of HI but also its amount, i.e. $\Omega_{\rm HI}$: we find a typical value of $\Omega_{\rm HI}$ equal to $0.6\times10^{-3}$, about a $40\%$ smaller than the observational measurements ($\Omega_{\rm HI}\sim10^{-3}$). This method reproduces fairly well the abundance of LLS and DLAs even though it fails to reproduce the abundance of the absorbers with the highest column densities. In terms of the HI power spectrum the predictions of this model are very similar to those obtained by using the halo-based model 1. However, since the value of $\Omega_{\rm HI}$ predicted by this method is below the observational measurements, we conclude that the HI is more clustered in this model than in the halo-based model 1. We notice that this model is not able to reproduce the DLAs bias measurements of \cite{Font_2012} (bias is lower than in observations) and therefore the amplitude of the HI power spectrum in this model is below the one computed from the halo-based model 2.
By using the particle-based method we have investigated the contribution of HI outside dark matter halos (an environment that we denominate \textit{filaments}) to both the total amount of HI in the Universe and to the HI power spectrum. We find that the amount of HI in filaments contributes to a very small fraction of the overall HI content, $\Omega_{\rm HI}^{\rm filaments}\sim10^{-6}$, with that fraction increasing significantly with redshift. Our results also point out that the contribution of HI in filaments to the total HI power spectrum is negligible for the redshifts studied in this paper.
We stress that our HI modeling reproduces extremely well the abundance of absorbers in the Lyman-$\alpha$ forest.
From our simulated HI distribution we have computed the 21 cm power spectrum. We find that the same features we observe in the HI power spectrum are also present in the 21 cm power spectrum, i.e. the amplitude and shape of the 21 cm power spectrum extracted from the HI distribution obtained by using the halo-based model 1 and the particle-based, are very similar. However, the amplitude of the 21 cm power spectrum calculated from the HI distribution found by employing the halo-based model 2 is much higher than those predicted by two other methods, reflecting the fact that the HI is much strongly clustered in this model in comparison to the other two. In order to asses the detectability of such quantity,
we have computed the system noise for the future SKA1-mid and SKA1-low radio-telescopes. Our results indicate that with 100 hours of observations the 21 cm power spectrum will be detected by these instrument up to very small scales: $k\sim1-3~h{\rm Mpc}^{-1}$ at redshifts $z=2.4$ and $z=3$ and $k\sim5-20~h{\rm Mpc}^{-1}$ at redshift $z=4$, depending on the particular model used to simulate the HI distribution. Since the 21 cm power spectra for all the three models have the same slope at large scale $k \lesssim 1 h~{\rm Mpc}^{-1}$, it should also be possible to constrain the underlying dark matter power spectrum at large scales using the 21 cm observations, though detailed investigations on this aspect will be done in a future work.
Furthermore, we have investigated the possibility of directly imaging rare peaks in the HI distribution. We find that SKA1-low may detect the brightest peaks with a SNR higher than $\sim5$ for a synthesized beam width of $2'$ at $z=4$ for 1000 hours of observation, if the HI distribution is described by the halo-based model 2. In case the HI distribution is modeled using the other two methods the SNR drops by a factor of $\sim2-3$ due to the fact that the HI content in halos is less those models. At redshifts $z=2.4$ and $z=3$ our results suggest that directly imaging large HI peaks with SKA1-mid would be challenging. By using the halo-based model 2 we find that the most massive HI peaks can be detected with a SNR close to 3 at $z=2.4$ for 1000 hours of observations and for a synthesized beam width of $2'$.
This work constitutes a first step in modeling the neutral hydrogen content in and outside halos by using semi-analytical recipes applied on top of hydrodynamical simulations and at the same time reproducing most of the relevant observational constraints.
\section{Visual comparison of the HI distribution}
\label{HI_distribution_appendix}
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/N_HI_z=3.png}\\
\end{center}
\caption{Distribution of the HI column densities, along lines of sights that expand along the whole simulation box, obtained by using the halo-based model 1 (left column), the halo-based model 2 (middle column) and the particle-based model (right column) on top of the simulation $\mathcal{B}60$ at $z=3$. The top row shows the distribution of HI on large scales whereas the bottom row represents a zoom into the region marked with a black square into each panel of the top row.}
\label{N_HI_map}
\end{figure}
Here we show the distribution of neutral hydrogen, obtained by assigning HI to the gas particles of the simulation $\mathcal{B}60$ at $z=3$, with the three methods investigated in this paper. In Fig. \ref{N_HI_map} we display the distribution of the column densities, computed along lines of sights that span
along the whole simulation box, when the halo-based model 1 (left column), the halo-based model 2 (middle column) and the particle based method is used. We find that the halo-based models over-predict the abundance of Lyman Limit Systems, whereas those are slightly under-predicted in the particle-based model. This is in agreement with the distribution of column densities from Figs. \ref{f_HI_Bagla}, \ref{column_density_Paco} and \ref{f_HI_Dave}.
The distribution of the DLAs is quite different among the models. Whereas the halo-based model 2 predicts a large cross-section for massive halos, this is saturated for massive halos in the halo-based model 1. The particle-based model is however in the intermediate situation, with a growing DLA cross-section with halo mass but with a lower amplitude than the one obtained from the halo-based model 2. We also stress the fact that the particle-based method also assign HI to particles outside halos (which give rise to the Lyman-$\alpha$ forest), that it is not present in the HI distribution generated by the other two models.
\section{HI modeling: halo-based}
\label{HI_halos}
In this paper we model the neutral hydrogen distribution using two
different techniques. In the first one, we assume that all the HI
resides in dark matter halos and we thus assign HI only to gas
particles belonging to dark matter halos. We refer to this method as
the \textit{halo-based} method and we describe it in detail in this
section. With the second technique we assign instead HI to all the gas
particles in the simulation, according to the physical properties of
the particles themselves. We label this second scheme the
\textit{particle-based} method and we shall depict it in section
\ref{HI_particles}.
The HI assignment in the halo-based model proceeds as follows: given
an N-body simulation snapshot we first identify all the dark matter
halos within it (FoF halos); then we compute, for each dark matter halo, the total
HI residing in it; finally we split the total HI mass among the gas
particles inside the halo.
In order to carry out the HI assignment using the halo-based method we
need two ingredients: 1) the HI mass within a halo as a function of
its total mass $M_{\rm HI}(M)$; 2) the scheme according to which the
halo HI mass is split among its gas particles. We neglect environment
effects and therefore, the HI mass that a halo host is only a function
of its mass. The function $M_{\rm HI}(M)$ completely determines the
value of the parameter $\Omega_{\rm HI}$:
\begin{equation}
\Omega_{\rm HI}(z)=\frac{1}{\rho_c}\int_0^\infty n(M,z)M_{\rm HI}(M)dM\,
\label{Omega_HI_constrain}
\end{equation}
where $n(M,z)$ is the halo mass function and $\rho_{\rm c}$ is the
Universe critical density at $z=0$. We have explicitly checked that
the abundance of halos in our simulations is well reproduced by the
Sheth \& Tormen halo mass function \cite{Sheth-Tormen}, that we use
through the paper. Notice that, even if our simulations do include gas cooling, we assume that the effects of baryons on
the halo mass function are negligible (see however \cite{Weiguang_2014}) and will not impact on our results.
We now investigate how the spatial properties of the neutral hydrogen
depend on the function $M_{\rm HI}(M)$ and on the HI distribution
within the dark matter halos. For that purpose we consider two
different models which use a different $M_{\rm HI}(M)$ function. We
model the spatial distribution of HI within halos to reproduce the
observed distribution of column densities.
\subsection{Model 1}
\label{Model_1}
In this model we follow the work of Bagla et al. 2010
\cite{Bagla_2010}, where three different schemes for the function
$M_{\rm HI}(M)$ were presented and applied to N-body dark matter only simulations.
In particular, we focus on their model number 3 which parametrizes the $M_{\rm HI}(M)$ function as:
\begin{equation}
M_{\rm HI}(M) = \left\{
\begin{array}{l l}
f_3\frac{M}{1+\left(\frac{M}{M_{\rm max}}\right)} & \quad \text{if $M_{\rm min}\leqslant M$}
\\
0 & \quad \text{otherwise}\\
\end{array} \right.
\label{M_HI_Bagla3}
\end{equation}
where $f_3$ is a free parameter describing the amplitude
of the function $M_{\rm HI}(M)$, $M_{\rm min}$ represents the minimum
mass of a dark matter halo able to contain neutral hydrogen and
$M_{\rm max}$ corresponds to the mass scale above which the fraction
of HI in a halo becomes suppressed.
We have chosen the Bagla et al. model number 3 to characterize the $M_{\rm HI}(M)$
function
since that function is physically better justified than the other two\footnote{We notice that
authors in \cite{Bagla_2010} also demonstrated that, on large scales,
the HI power spectrum obtained by using any of their three $M_{\rm HI}(M)$ functions is
almost the same and we have also verified their findings.} (see \cite{Bagla_2010} for
details). However, we have explicitly checked that our results do not significantly change
if we use the Bagla et al. model number 2.
Following \cite{Bagla_2010}, the value of $M_{\rm min}$ and
$M_{\rm max}$ can be obtained assuming that dark matter halos with
circular velocities below 30 km/s does not contain any HI and that
halos with circular velocities higher than 200 km/s will present a
suppressed HI mass fraction. The relationship between the halo
circular velocity ($v_c$) and its mass can be approximated by
\begin{equation}
M=10^{10}M_\odot\left(\frac{v_c}{60~{\rm km/s}}\right)^3\left(\frac{1+z}{4}\right)^{-3/2}~.
\end{equation}
The value of the free parameter $f_3$ is obtained by requiring that
$\Omega_{\rm HI}(z)=10^{-3}$, as suggested by observations.
Once the total HI mass within a given dark matter halo is
computed, we split it equally among all the gas particles belonging to
the halo. From now on we label this method as the \textit{halo-based model 1}.
We stress that unlike the original Bagla et al. implementation we rely on
hydrodynamical simulations and thus the HI is assigned to gas particles.
The values of both $\Omega_{\rm HI}$ and the DLAs line density,
$dN/dX$, that we obtain by assigning HI to the gas particles using the
halo-based model 1 are shown in table \ref{Omega_HI_tab_model1}.
For a particular simulation and redshift the value of
$\Omega_{\rm HI}$ is obtained by summing the HI content assigned
to all the gas particles, whereas the value of $dN/dX$ is computed
from the HI column density distribution (see below).
Notice that differently from \cite{Bagla_2010}, we obtain
the amplitude of the function $M_{\rm HI}(M)$, $f_3$, by using the
whole mass function range, not just the range that the simulations can
resolve. In this way we make sure that halos of the same mass contain
the same HI mass, independently of the resolution of the simulation. On the
other hand, low mass halos can not be resolved in the simulations with
the large box sizes. Thus, the value of $\Omega_{\rm HI}$ that we
obtain from the simulation $\mathcal{B}60$ is below the one from
observations, and has its origin on the fact that low mass dark matter
halos are not properly resolved in that simulation.
\begin{table}
\begin{center}
\resizebox{12cm}{!}{
\begin{tabular}{|c|c|c|c|c|}
\cline{1-5}
z & quantity & $\mathcal{B}60$ & $\mathcal{B}30$ & $\mathcal{B}15$ \\ \cline{1-5}
\multicolumn{1}{ |c| }{\multirow{2}{*}{2.4} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.738\times10^{-3}$ & $0.937\times10^{-3}$ & $0.988\times10^{-3}$ \\ \cline{2-5}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.026$ & $0.033$ & $0.037$ \\ \cline{1-5}
\multicolumn{1}{ |c| }{\multirow{2}{*}{3.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.692\times10^{-3}$ & $0.960\times10^{-3}$ & $1.012\times10^{-3}$ \\ \cline{2-5}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.028$ & $0.037$ & $0.043$ \\ \cline{1-5}
\multicolumn{1}{ |c| }{\multirow{2}{*}{4.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.585\times10^{-3}$ & $0.953\times10^{-3}$ & $1.006\times10^{-3}$ \\ \cline{2-5}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.027$ & $0.040$ & $0.047$ \\ \cline{1-5}
\end{tabular}
}
\end{center}
\caption{Values of the DLAs line density, $dN/dX$, and $\Omega_{\rm
HI}$ obtained by assigning the HI to the gas particles according
to the halo-based model 1.}
\label{Omega_HI_tab_model1}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/f_HI_Bagla.pdf}\\
\end{center}
\caption{Column density distribution obtained by using the halo-based
model 1 at $z=2.4$ (left), $z=3.0$ (middle) and $z=4$ (right) for
the simulations $\mathcal{B}15$ (solid blue), $\mathcal{B}30$
(dashed green) and $\mathcal{B}60$ (dot-dashed red). The
observational measurements of Noterdaeme et al. 2012
\cite{Noterdaeme_2012} are shown in black whereas those by Zafar
et al. 2013 \cite{Zafar_2013} are displayed in gray.}
\label{f_HI_Bagla}
\end{figure}
We compute (see appendix \ref{column_density_appendix} for details)
the column density distribution function, $f_{\rm HI}$, for the simulations
$\mathcal{B}15$, $\mathcal{B}30$ and $\mathcal{B}60$ at redshifts
$z=2.4$, $z=3$ and $z=4$ and show the results in
Fig.~\ref{f_HI_Bagla}, together with the measurements of
Noterdaeme\footnote{These results are obtained for DLAs in the
redshift range $z\in[2,3.5]$, with a mean redshift value of $\langle
z \rangle=2.5$.} et al. 2012 \cite{Noterdaeme_2012} and Zafar et al.
2013 \cite{Zafar_2013}. For simplicity
we assume that the HI column density distribution from observations
does not vary with redshift, i.e. we consider that the observed abundance
of DLAs and Lyman Limit Systems (LLS) at the redshift studied in this paper
($z=2.4,3,4$) are given by the measurements by Noterdaeme et al. (2012) and
Zafar et al. (2013).
Overall we find a good agreement between our HI distribution and
the observational measurements for absorbers with column densities
larger than $10^{21}~{\rm cm}^{-2}$. However,
we find that this model under-predicts the abundance of absorbers
with column densities in the range $[10^{20}-10^{21}]~{\rm cm^{-2}}$,
which gives rise to a lower value of the
DLA line density that the one observed (see table
\ref{Omega_HI_tab_model1}),
and over-predicts the abundance of Lyman Limit
Systems (LLS). The spatial distribution of the column densities,
obtained by assigning HI to the gas particles of the simulation $\mathcal{B}15$
at $z=3$ using the three methods investigated in this paper, is shown in the appendix
\ref{HI_distribution_appendix}. The column density distributions of the simulations
$\mathcal{B}15$ and $\mathcal{B}30$ are in good agreement among
themselves even though the resolution of the former is eight times larger
than the one of the latter. The column density distribution of the
simulation $\mathcal{B}$60 is below those obtained from the other two
simulations, pointing out the lack of the DLAs residing in low mass
halos that the simulation is not able to resolve. We find that the
column density distribution exhibits a very weak redshift dependence.
We notice that we obtain almost the same HI column density distribution
if we assign HI to the halos gas particles according to the Bagla et al. model 2.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/cross_section_Bagla.png}\\
\end{center}
\caption{DLA comoving cross-section for a subsample of halos at $z=2.4$ (left), $z=3$ (middle) and $z=4$ (right) from the simulations $\mathcal{B}60$ (red), $\mathcal{B}30$ (green) and $\mathcal{B}15$ (blue) when the neutral hydrogen is assigned to the gas particles using the halo-based model 1. A fit to the results is shown with a solid orange line.}
\label{cross_section_Bagla}
\end{figure}
For each dark matter halo we compute the DLA cross-section,
$\sigma_{\rm DLA}$, as follows: we select the gas particles belonging
to the halo and we throw random lines of sight within the halo virial
radius. For each line of sight we then compute its column density and we
estimate the DLA cross section as \cite{Pontzen_2008}:
\begin{equation}
\sigma_{\rm DLA}=\pi R_{\rm halo}^2\left(\frac{n_{\rm DLA}}{n_{\rm total}}\right)~,
\end{equation}
where $R_{\rm halo}$ is the halo virial radius, $n_{\rm DLA}$ is the
number of lines of sights with column densities higher than
$10^{20.3}~{\rm cm}^{-2}$ and $n_{\rm total}$ is the total number of
lines of sight used. We note that since we are using FoF halos the
quantity $R_{\rm halo}$ is not well defined. We use the value of the
radius, centered on the position of the particle with the minimum
energy, within which the mean density is 200 times the mean density of
the Universe as definition for $R_{\rm halo}$. We have explicitly
checked that our results do not change if we use a different
definition for the halo radius. We use 5000 random LOSs for each dark
matter halo; we have checked that our results do not significantly
change if we increase the former number. In
Fig. \ref{cross_section_Bagla} we show the comoving cross-section of
the DLAs for the simulations $\mathcal{B}60$, $\mathcal{B}30$ and
$\mathcal{B}15$ at redshifts 2.4, 3 and 4. We find that the DLA
cross-section increases with the halo mass, although for halos more
massive than $\sim 5\times10^{11}~h^{-1}{\rm M}_\odot$ the cross
section remains constant, almost independent of redshift. This is a
consequence of the $M_{\rm HI}(M)$ function used, which assigns, for
halos with masses above $M_{\rm max}$, a decreasing $M_{\rm HI}/M$
fraction. The mean DLA cross-section is well reproduced
by a function of the form:
\begin{equation}
\sigma_{\rm DLA}(M) = \left\{
\begin{array}{l l}
\sigma_0\left[\frac{M}{M_0}\right]^{\alpha}\left[1+\left(\frac{M}{M_0}\right)^{\beta}\right]^{-\alpha/\beta} & \quad \text{if $M_{\rm min}\leqslant M$}\\
\\
0 & \quad \text{otherwise}~.\\
\end{array} \right.
\label{cross_section_Bagla_fit}
\end{equation}
\begin{table}
\begin{center}
\resizebox{10cm}{!}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$z$ & $\sigma_0$ & $M_0$ & $\alpha$ & $\beta$ & $M_{\rm min}$ \\
& $(h^{-1}{\rm kpc})^2$ & $(h^{-1}M_\odot)$ & & & $(h^{-1}M_\odot)$ \\
\hline
$2.4$ & $1641$ & $3.41\times10^{11}$ & $0.79$ & 1.81 & $1.12\times10^9$\\
\hline
$3.0$ & $2429$ & $2.67\times10^{11}$ & $0.85$ & 1.89 & $8.75\times10^8$\\
\hline
$4.0$ & $4886$ & $2.81\times10^{11}$ & $0.79$ & 2.04 & $6.26\times10^8$ \\
\hline
\end{tabular}
}
\end{center}
\caption{Best fit values of the fitting formula \ref{cross_section_Bagla_fit} used to reproduce the average comoving cross-section obtained by assigning HI to the gas particles belonging to dark matter halos using the halo-based model 1.}
\label{cross-section_Bagla_table}
\end{table}
In table \ref{cross-section_Bagla_table} we present the best fit
values of the parameters of the above fitting function as a function of
redshift. Those fitting are also shown with a solid orange line in
Fig. \ref{cross_section_Bagla}. Next, we use mean DLAs cross-section to
estimate the DLAs bias as:
\begin{equation}
b_{\rm DLA}(z)=\frac{\int_0^\infty b(M,z)n(M,z)\sigma_{\rm DLA}(M,z)dM}{\int_0^\infty n(M,z)\sigma_{\rm DLA}(M,z)}~\, ,
\label{biaseq}
\end{equation}
with $n(M,z)$ being the halo mass function and $b(M,z)$ the halo
bias. For the former quantity we use the Sheth \& Tormen halo mass
function whereas for the latter we employ the Sheth, Mo \& Tormen halo
bias formula \cite{SMT}. We find the following values for the DLAs
bias: $b_{\rm DLA}=1.47$ at $z=2.4$, $b_{\rm DLA}=1.74$ at $z=3$ and
$b_{\rm DLA}=2.10$ at $z=4$. At $z=2.4$ our results are in strong
tension with recent observational measurements obtained with SDSS-III/BOSS
survey, which find a value of $b_{\rm DLA}=(2.17\pm0.20)\beta_F^{0.22}$, with
$\beta_F\sim 1-1.5$, at $\langle z \rangle=2.3$ \cite{Font_2012}. We notice that the values of the DLAs bias barely change by using the Bagla et al. model 2 to characterize the function $M_{\rm HI}$.
We try to simulate different HI density profiles within dark matter halos by distributing the neutral hydrogen according to the physical properties of the gas particles (see Eq. \ref{M_HI_Paco_sim}) to investigate whether we can fit at the same time the DLAs column density distribution and their bias. Our results indicate that this model is unable to fit simultaneously both quantities. The reason arises from the $M_{\rm HI}(M)$ used in this model, which assigns a constant HI mass to the most massive halos. In order to reproduce the DLAs bias we would need a HI density profile which only allocate DLAs to the most massive halos. On the other hand we find that such density profile can not reproduce the observations of the DLAs column density distribution function.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/Pk_HI_Bagla.pdf}\\
\end{center}
\caption{HI power spectrum obtained by assigning the HI to the gas
particles belonging to dark matter halos using the halo-based model
1. The results are displayed at $z=2.4$ (left), $z=3$ (middle) and
$z=4$ (right) for the simulations $\mathcal{B}60$ (dot-dashed red),
$\mathcal{B}30$ (dashed green) and $\mathcal{B}15$ (solid blue). The
value of the Nyquist frequency is displayed for each simulation with
a vertical line. We show the bias between the distributions of HI and
matter, $b_{\rm HI}^2(k)=P_{\rm HI}(k)/P_{\rm m}(k)$, in the bottom panels.}
\label{Pk_HI_Bagla}
\end{figure}
In Fig. \ref{Pk_HI_Bagla} we show the results of computing the HI
power spectrum at redshifts $z=2.4$, $z=3$ and $z=4$ for the different
simulations. The values of the HI over-density are computed on
a regular cubic grid with $1024^3$ points
using the Cloud-in-Cell (CIC) interpolation technique. The mean HI
mass per grid cell is computed using $\Omega_{\rm HI}=10^{-3}$
instead of the value of $\Omega_{\rm HI}$ from the simulation itself.
The reason for doing this arises because some simulations (for instance
$\mathcal{B}60$) do not have resolution enough to resolve the smallest
halos that host HI and therefore, if we use the value of $\Omega_{\rm HI}$
from those simulations, the HI power spectrum from will be artificially above
the real one. The values of the neutral hydrogen mass over-densities are then
Fourier transformed using the Fast Fourier Transform algorithm and the
HI power spectrum is estimated through:
\begin{equation}
P_{\rm HI}(k)=\frac{1}{N_{\rm modes}}\sum_{\vec{k}\in k}\delta_{\rm HI}(\vec{k})\delta^*_{\rm HI}(\vec{k})
\end{equation}
with $N_{\rm modes}$ being the number of modes lying within the
spherical shell in $k$ considered. For concreteness, we compute the HI
power spectrum within spherical shells of width $\delta k=2\pi/L$,
with $L$ being the simulation box size.
We have explicitly checked that our results are the same on
intermediate and large scales if we compute the HI power spectrum
taking into account the SPH kernel of the gas particles (see appendix
\ref{power_spectrum_method}). Whereas there are some differences on
small scales, a proper computation requires a mode amplitude
correction to take into account the mass assignment scheme (MAS),
which is beyond the scope of this paper. We therefore use the CIC
technique, correcting the modes amplitude to account for the
MAS, since the calculation is much faster.
We find that the HI power spectrum from the simulations
$\mathcal{B}15$, $\mathcal{B}30$ and $\mathcal{B}60$ are in very good
agreement among themselves at all redshifts. On large scales, the
amplitude of the HI power spectrum is slightly lower in the simulation
$\mathcal{B}60$ than in the other two. This is because this simulation can not resolve
the smallest dark matter halos that host HI, and therefore, the
contribution of those halos to the HI power spectrum is not accounted
for. Even though the value of $\Omega_{\rm HI}$ is substantially
below $10^{-3}$ for the simulation $\mathcal{B}60$, the HI power
spectrum is almost converged, pointing out that the amplitude of the
HI power spectrum is set by the most massive, and biased, dark matter
halos.
In the bottom panels of Fig. \ref{Pk_HI_Bagla} we display the bias
between the distribution of neutral hydrogen and the one of the underlying matter:
$b^2_{\rm HI}(k)=P_{\rm HI}(k)/P_{\rm m}(k)$. We note that even though the HI
power spectra from the different simulations are in good agreement with each other,
the HI bias do not show such good convergence. This is due to the matter power
spectrum, whose amplitude decreases slightly with resolution due to cosmic variance
and also because small box size simulations do not account for the large scale modes
that influence the structure growth. We expect this problem not to be present in large box-size
simulations, having resolution high enough to resolve the smallest halos capable of hosting HI.
As expected, we find that the HI bias increases with redshift. Our estimates for $b_{\rm HI}(k)$ agree quite well with previous works which use a prescription similar to ours
\cite{Bagla_2010,Guha_2012}.
\subsection{Model 2}
\label{Model_2}
In this model we follow the work of \cite{Barnes_2010, Barnes_2014}
and use a different $M_{\rm HI}(M)$ function, which assigns a constant
$M_{\rm HI}/M$ fraction for very massive halos:
\begin{equation}
M_{\rm HI}(M)=\alpha f_{\rm H,c}\exp{\left[ -\left(\frac{v_c^0}{v_c}\right)^\beta\right]}M~,
\label{M_HI_Paco}
\end{equation}
where $f_{\rm H,c}=0.76\Omega_{\rm b}/\Omega_{\rm m}$ is the cosmic
hydrogen mass fraction, $v_c$ is the halo circular velocity and
$\alpha$, $\beta$ and $v_c^0$ are the model free parameters. Our aim
is to reproduce the most important properties of the
DLAs (their bias, line density and column density distribution). We
first present a simple analytic model that is able to reproduce the
DLAs properties and then we use this as a framework to assign HI to the
gas particles belonging to the dark matter halos.
Our simple analytic model, based on the work of \cite{Barnes_2014},
consists of two ingredients: 1) the HI mass within a dark matter halo
of mass $M$; 2) the density profile of the neutral hydrogen within
it. For the former we use Eq. \ref{M_HI_Paco} whereas we
phenomenologically model the HI density profile within any dark matter
halo as:
\begin{equation}
\rho_{\rm HI}(r)=\frac{\rho_0}{\left( \frac{r}{r_s}+0.06\right)^2\left(\frac{r}{r_s}+1\right)^3}~\, ,
\label{HI_profile_Paco}
\end{equation}
in order to match observations.
We assume that the HI profile extends out to the halo virial radius,
$R$ and we take the value of the parameter $r_s$ to be $r_s=Rc$, with $c$
being the halo concentration following the law:
\cite{Bullock_2001,Maccio_2007}
\begin{equation}
c(M,z)=c_0\left(\frac{M}{10^{11}M_\odot}\right)^{-0.109}\left(\frac{4}{1+z}\right)~.
\end{equation}
The value of the parameter $\rho_0$ is found by requiring that the HI
mass obtained by integrating the density profile \ref{HI_profile_Paco}
is equal to this obtained from Eq. \ref{M_HI_Paco}. Given the HI
density profile and the function $M_{\rm HI}(M)$, the DLAs bias can be
computed using Eq.~\ref{biaseq}.
The DLA cross-section, $\sigma_{\rm DLA}=\pi s^2$, of
a given halo is calculated by finding the value $s$, for which
\begin{equation}
N_{\rm HI}=\frac{2}{m_H}\int_0^{\sqrt{R^2-s^2}}\rho_{\rm HI}(r=\sqrt{s^2+l^2})dl=10^{20.3}~{\rm cm}^{-2}~,
\label{N_HI_analytic}
\end{equation}
with $m_H$ being the mass of the Hydrogen atom. Similarly, the DLAs line density and the column density distribution can be obtained by computing:
\begin{eqnarray}
\frac{dN}{dX}&=&\frac{c}{H_0}\int_0^\infty n(M,z)\sigma_{\rm DLA}(M,z)dM~,\\
f_{\rm HI}(N_{\rm HI},X)&=&\frac{c}{H_0}\int_0^\infty n(M,z)\left|\frac{d\sigma(N_{\rm HI}|M,X)}{dN_{\rm HI}}\right|dM~,
\end{eqnarray}
with
\begin{equation}
\sigma_{\rm DLA}(M,X)=\int_{10^{20.3}}^\infty \left|\frac{d\sigma(N_{\rm HI}|M,X)}{dN_{\rm HI}}\right| dN_{\rm HI}~.
\label{eqdlabias}
\end{equation}
$\sigma(N_{\rm HI})=\pi\tilde{s}^2$, with $\tilde{s}$ obtained by
using Eq. \ref{N_HI_analytic} but taking the actual value for $N_{\rm HI}$,
rather than $10^{20.3}$\, cm$^{-2}$. At $z=2.4$ we tune the values of the model free
parameters to reproduce both the DLAs bias \cite{Font_2012} and
the HI column density distribution \cite{Noterdaeme_2012}. At $z=3$
and $z=4$ we calibrate the value of the parameters to reproduce the
column density distribution, keeping the same value for the parameters
$v_c^0$ and $\beta$. At $z=3$ and $z=4$ the DLAs bias arises thus as a
prediction of the model.
We have calibrated the free parameters of this simple analytic model using the
cosmological model of our N-body simulation. We have explicitly
checked that the values of the model parameters exhibit a weak
dependence with the cosmological parameters\footnote{Note that in
\cite{Font_2012} the value of the cosmological parameters are
slightly different to ours: $\Omega_{\rm m}=0.281$, $\Omega_{\rm
b}=0.049$, $\Omega_\Lambda=0.719$, $\sigma_8=0.8$, $n_s=0.963$ and
$h=0.71$.}. In table \ref{Paco_parameters} we show the value of the
model parameters together with the values of the DLAs line density,
$\Omega_{\rm HI}$ and the DLAs bias as a function of redshift.
The value of $\Omega_{\rm HI}$ has been computed using Eq.
\ref{Omega_HI_constrain}. We note that this model predicts a slightly
larger value for $\Omega_{\rm HI}$ than this obtained by observations.
We will see later that this happens because this model over-predicts the
abundance of LLS. In
Fig. \ref{column_density_analytic_Paco} we show the DLAs column
density distribution and its comparison with the observational
measurements. We remind the reader that in this paper
we assume no redshift dependence in the observations of the
HI column density distribution.
\begin{table}
\begin{center}
\resizebox{10cm}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
z & $\alpha$ & $\beta$ & $v_c^0 (km/s)$ & $c_0$ & $dN/dX$ & $\Omega_{\rm HI}$ & $b_{\rm DLA}$\\
\hline
2.4 & 0.18 & 3 & 37 & 3.4 & 0.07 & $1.18\times10^{-3}$ & 2.15\\
\hline
3.0 & 0.222 & 3 & 37 & 3.2 & 0.07 & $1.12\times10^{-3}$ & 2.37\\
\hline
4.0 & 0.34 & 3 & 37 & 2.7 & 0.07 & $1.12\times10^{-3}$ & 2.76\\
\hline
\end{tabular}
}
\end{center}
\caption{Value of the parameters $\alpha$, $\beta$, $v_c^0$ and $c_0$ that best reproduce the DLAs observational properties for the analytic model 2. The values of the DLAs line density and bias, together with the value of $\Omega_{\rm HI}$ are also shown in the table.}
\label{Paco_parameters}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{Figs/f_HI_Paco_analytic.pdf}\\
\end{center}
\caption{Column density distribution from the analytic model 2 and its comparison with the observational measurements of Noterdaeme et al. 2012 \cite{Noterdaeme_2012}.}
\label{column_density_analytic_Paco}
\end{figure}
\begin{table}
\begin{center}
\resizebox{14cm}{!}{
\begin{tabular}{|c|c|c|c|c|c|}
\cline{1-6}
z & quantity & $\mathcal{B}120$ & $\mathcal{B}60$ & $\mathcal{B}30$ & $\mathcal{B}15$ \\ \cline{1-6}
\multicolumn{1}{ |c| }{\multirow{2}{*}{2.4} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.886\times10^{-3}$ & $1.183\times10^{-3}$ & $1.208\times10^{-3}$ & $1.157\times10^{-3}$ \\ \cline{2-6}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & 0.057 & $0.067$ & $0.064$ & $0.064$ \\ \cline{1-6}
\multicolumn{1}{ |c| }{\multirow{2}{*}{3.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.711\times10^{-3}$ & $1.067\times10^{-3}$ & $1.136\times10^{-3}$ & $1.012\times10^{-3}$ \\ \cline{2-6}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.047$ & $0.066$ & $0.068$ & $0.043$ \\ \cline{1-6}
\multicolumn{1}{ |c| }{\multirow{2}{*}{4.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$} & $0.524\times10^{-3}$ & $0.991\times10^{-3}$ & $1.151\times10^{-3}$ & $1.115\times10^{-3}$ \\ \cline{2-6}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.036$ & $0.066$ & $0.074$ & $0.073$ \\ \cline{1-6}
\end{tabular}
}
\end{center}
\caption{Values of the DLA line density, $dN/dX$ and $\Omega_{\rm HI}$ obtained by assigning the HI to the gas particles according to the halo-based model 2.}
\label{Omega_HI_tab_model2}
\end{table}
Motivated by the fact that the above simple model is able to reproduce
the most important properties of the DLAs, we now proceed to populate
the dark matter halos with HI. We compute the neutral hydrogen mass
that a dark matter halo of mass $M$ hosts by using Eq. \ref{M_HI_Paco},
with the calibrated parameter values at the corresponding redshift.
We then split the total HI mass
within a given dark matter halo among its gas particles. If
we distribute equally the overall HI mass among all the gas particles
we are not able to reproduce the observed column density
distribution: we have thus developed a simple recipe to distribute the HI
among the gas particles that is able to reproduce the HI density
profile of Eq. \ref{HI_profile_Paco}. The HI distribution proceeds as
follows: given the total HI that a given dark matter halo host,
$M_{\rm HI}$, we identify the gas particles belonging to it and assign a
HI mass to them equal to:
\begin{equation}
M_{\rm HI}^{p}=\frac{\left(\frac{{\rm HI}}{{\rm H}}\right)_p^{0.17}\left(r_{{\rm SPH,}p}\right)^{0.35}}{\sum_i \left(\frac{{\rm HI}}{{\rm H}}\right)_i^{0.17}\left(r_{{\rm SPH,}i}\right)^{0.35}}M_{\rm HI}~,
\label{M_HI_Paco_sim}
\end{equation}
where HI/H is the neutral hydrogen fraction that {\sc GADGET} assigns to each
gas particle, $r_{\rm SPH}$ is the gas particle SPH radius and $M_{\rm HI}$ is the
total HI mass within the halo. We denominate this method of assigning HI to the gas
particles belonging to dark matter halos as the \textit{halo-based model 2}.
In table \ref{Omega_HI_tab_model2} we show the values of the DLAs line density
and $\Omega_{\rm HI}$, computed by summing the HI of all the gas particles,
obtained by assigning HI to the gas particles using the halo-based model 2
described above. The column density distribution and its comparison
with the observational measurements are displayed in
Fig. \ref{column_density_Paco}. We find an excellent agreement between
observations and our mock HI distribution in the DLA regime.
However, this model, as the halo-based model 1, fails to reproduce the abundance of
Lyman Limit systems, predicting a number significantly above the one
observed, which in turn produces an enhancement in the value of $\Omega_{\rm HI}$.
Moreover, we note that the results from the simulation $\mathcal{B}120$ are slightly
below the observational measurements: this is because this simulation
does not have resolution enough to resolve the smallest halos that
host DLAs.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/f_HI_Paco.pdf}\\
\end{center}
\caption{Column density distribution obtained by assigning HI to the
gas particles using the halo-based model 2 at $z=2.4$ (top), $z=3.0$
(middle) and $z=4$ (bottom) for the simulations $\mathcal{B}15$
(solid blue), $\mathcal{B}30$ (dashed green), $\mathcal{B}60$
(dot-dashed red) and $\mathcal{B}120$ (dotted cyan). The
observational measurements of Noterdaeme et al. 2012
\cite{Noterdaeme_2012} are shown in black.}
\label{column_density_Paco}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/cross_section_Paco.png}\\
\end{center}
\caption{DLAs comoving cross section for a subsample of halos at
$z=2.4$ (left), $z=3$ (middle) and $z=4$ (right) from the
simulations $\mathcal{B}120$ (cyan), $\mathcal{B}60$ (red),
$\mathcal{B}30$ (green) and $\mathcal{B}15$ (blue) when the neutral
hydrogen is assigned to the gas particles using the halo-based model
2. The orange line represents the cross-section of the analytic
model (see text for details).}
\label{cross_section_Paco}
\end{figure}
In order to verify that the HI density profile that we simulate
matches the one of the above analytic model we plot in
Fig. \ref{cross_section_Paco} the comoving cross-section of a subset
of dark matter halos from the different simulations. With a solid
orange line we show the cross-section obtained by using the above
analytic model. We find a good agreement between the analytic model
and cross-section of the halos populated with HI using the halo-based
model 2 at all redshifts. In other words, this model distributes the
HI in such a way that the spatial distribution of neutral hydrogen
reproduces the measurements of both the DLAs
column density distribution and bias (at $z=2.4$). We emphasize that the
orange line in Fig. \ref{cross_section_Paco} is not a fit to the DLAs
cross-section from the N-body but the prediction from the analytic model.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/Pk_HI_Paco.pdf}\\
\end{center}
\caption{HI power spectrum obtained by assigning the HI to the gas
particles belonging to dark matter halos using the halo-based model
2. The results are displayed at $z=2.4$ (top), $z=3$ (middle) and
$z=4$ (bottom) for the simulations $\mathcal{B}120$ (dotted cyan),
$\mathcal{B}60$ (dot-dashed red), $\mathcal{B}30$ (dashed green) and
$\mathcal{B}15$ (solid blue). The value of the Nyquist frequency is
displayed for each simulation with a vertical line.
We show the bias between the distributions of HI and matter,
$b_{\rm HI}^2(k)=P_{\rm HI}(k)/P_{\rm m}(k)$, in the bottom panels.}
\label{Pk_HI_Paco}
\end{figure}
In Fig. \ref{Pk_HI_Paco} we plot the HI power spectrum when the HI is
assigned to the gas particles using the halo-based model 2. As
discussed at the beginning of this section, this model assigns a
constant $M_{\rm HI}/M$ fraction for very massive halos: this implies
that a strong contribution of the overall HI power spectrum is
expected to arise from very massive, and thus biased, dark matter
halos. We find that the HI power spectrum is converged on large and
intermediate scales only when using box sizes larger than $\sim
60~h^{-1}$ com. Mpc. This clearly shows the impact that massive halos have
in terms of power spectrum for this model. We note that the results from
the simulation $\mathcal{B}120$ should not be fully converged on all
scales, since that simulation is not able to resolve the smallest
halos that host HI (see from table \ref{Omega_HI_tab_model2} that the
value of $\Omega_{\rm HI}$ is below the expected one). For this scheme,
the $\Omega_{\rm HI}$ value used to compute the average HI mass
per grid cell (for the HI power spectrum computation) is the one
expected from the analytic model (see table \ref{Paco_parameters}),
not the value obtained from the simulations themselves.
To corroborate that the discrepancy in the HI power spectrum arises
as a consequence of the different halo masses sampled by the different
simulations we have selected halos of the same mass range, converged
in terms of the halo mass function, and computed the power spectrum
of the HI residing within those halos. In this case we find that the HI power spectrum is
almost the same among the different simulations, pointing out that the
discrepancies are related to the different mass ranges explored in the
simulations.
The bottom panels of Fig. \ref{Pk_HI_Paco} display the HI bias. As expected, the
values of the HI bias are significantly above those obtained by using the halo-based
model 1 for the reasons depicted above. Notice that the HI bias at $z=2.4$ and
$z=3$ for the simulations $\mathcal{B}60$ and $\mathcal{B}120$ are in very good
agreement among themselves, since both the HI and the matter
power spectrum are very similar for those simulations.
In summary, whereas the distributions of HI obtained by using both halo-based
models can reproduce well the DLAs column density distribution, only the
halo-based model 2 is capable of distribute the HI in such a way that it matches
the recent measurements of the DLAs bias from SDSS-III/BOSS. This means that the HI in the halo-based
model 2 is more strongly clustered than in halo-based model 1. This is clearly
reflected in the amplitude of the HI power spectrum: on large scales the HI power
spectrum from the halo-based model 2 is a factor $\sim 3$ above this obtained
from the halo-based model 1.
\section{HI outside dark matter halos}
\label{HI_outside_halos_section}
It is well known that in the post-reionization era the majority of the neutral hydrogen in the Universe is expected to reside in dense environments such as galaxies. We now investigate the importance, in terms of abundance and spatial clustering, of the HI residing outside dark matter halos, and we define this environment, i.e. all the gas outside dark matter halos (as identified by the FoF algorithm), as \textit{filaments}.
Our filamentary cosmic-web is realistic in the sense that it gives rise to a Lyman-$\alpha$ forest whose statistical properties are in agreement
with observations both in terms of continuous statistics like flux power and discrete statistics as properties of lines.
Since, by construction, the halo-based method does not assign any HI to gas particles outside halos, we use the particle-based method to investigate the properties of HI in filaments. We compute the value of $\Omega_{\rm HI}^{\rm filaments}$, i.e. the value of $\Omega_{\rm HI}$ using only the gas particles outside halos, and show the results in table \ref{Omega_HI_tab_Dave}. For the simulations with the highest resolution we find a typical value of $\Omega_{\rm HI}^{\rm filaments}\sim10^{-6}$. However, this value increases by more than an order of magnitude as resolution decreases.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{Figs/Pk_HI_filaments.pdf}\\
\end{center}
\caption{Contribution of the HI in filaments, i.e. HI residing outside dark matter halos, to the total neutral hydrogen power spectrum. With solid lines we show the total HI power spectrum at $z=2.4$ (top-left), $z=3$ (top-right) and $z=4$ (bottom-left) when the HI is assigned to the gas particles of the simulation $\mathcal{B}60$ using the halo-based models 1 and 2 (red and green lines respectively) and the particle-based method (blue lines). The yellow lines represent the results obtained by using the simulation $\mathcal{B}60W$ (only applying the particle-based method). The dashed lines represent the power spectrum of the HI in filaments. The vertical lines display the value of the Nyquist frequency.}
\label{Pk_HI_filaments}
\end{figure}
We now quantify the contribution of the HI in filaments to the total neutral hydrogen power spectrum. In Fig. \ref{Pk_HI_filaments} we show with solid lines the total HI power spectrum from the simulation $\mathcal{B}60$ when the neutral hydrogen is assigned to the gas particles using the three different methods investigated in this paper. The dashed lines correspond to the power spectrum of the HI in filaments. We find that for all methods and redshifts studied in this paper the contribution of the HI outside halos to the total HI power spectrum is negligible in all cases. However, we have checked that this is no longer the case as the redshift increases: for the simulation $\mathcal{B}60$, at $z=6$, about 30\% of the total HI content is contributed by the filaments. We expect that the contribution of HI in filaments to the overall HI content to increase with redshift since at higher redshift the abundance of dark matter halos decreases, leaving much more gas available for hosting HI outside dark matter halos.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{Figs/f_HI_Lya.pdf}\\
\end{center}
\caption{HI column density distribution. The red solid line shows the HI column density distribution obtained by using VPFIT over the neutral hydrogen distribution obtained by assigning HI to the gas particles of the simulation $\mathcal{B}60$ at $z=3$ using the particle-based method. The red dashed line represents the HI column density distribution using our standard procedure of projecting the particles in a plane. The error points represent the measurements by Kim et al. 2013 \cite{Kim_2013} (black), Prochaska et al. 2010 \cite{Prochaska_2010} (gray area), Zafar et al. 2013 \cite{Zafar_2013} (dark gray) and Noterdaeme et al. 2012 \cite{Noterdaeme_2012} (light gray).}
\label{N_HI_vpfit}
\end{figure}
Finally, we investigate the properties of the HI in filaments in terms of the HI column density distribution. We assign HI to the gas particles of the simulation $\mathcal{B}60$ at $z=3$ using the particle-based method. We then generate 1000 QSO mock spectra and fit them using the VPFIT code \cite{Carswell_1987}. In Fig. ~\ref{N_HI_vpfit} we show with a solid red line the HI column density distribution obtained by using VPFIT together with the observational measurements by Kim et al. \cite{Kim_2013}. The agreement between simulations and observations is remarkable. We also plot with a dashed red line the column density distribution for high column density values obtained using our standard procedure. As discussed in Sec. \ref{HI_particles}, the particle-based method produces reasonable results for the LLS systems but fails to reproduce the abundance of the highest column density absorbers.
We emphasize that neither the value of $\Omega_{\rm HI}^{\rm filaments}$ nor the filaments HI power spectrum is converged and the results strongly depend on resolution, while the column density distribution function obtained with VPFIT is reasonably well converged. This happens because our definition of filaments not only includes the diffuse inter-galactic medium, but also the circum-galactic medium surrounding the dark matter halos and sub-resolution halos. Therefore, it is natural that some of the quantities shown exhibit such relatively large dependence with resolution. However, we can safely conclude that the amplitude of the HI power spectrum at $z<4$ is set by the HI residing in dark matter halos, with an almost negligible contribution coming from HI residing outside halos. These conclusions are no longer valid when considering higher redshift regimes.
\section{HI modeling: particle-based}
\label{HI_particles}
In this section we use a different approach to simulate the distribution of neutral hydrogen. The method, that we refer to as the \textit{particle-based} method, is based on assigning HI to all the gas particles of the N-body simulation on a particle-by-particle basis,
according to the individual physical properties.
Notice that one of the most important differences of this approach with respect to the halo-based method is that we do not make any assumption about the location of the HI (in the halo-based method we assume that all the HI reside within dark matter halos).
\begin{table}
\begin{center}
\resizebox{15.5cm}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|}
\cline{1-7}
z & quantity & $\mathcal{B}120$ & $\mathcal{B}60W$ & $\mathcal{B}60$ & $\mathcal{B}30$ & $\mathcal{B}15$ \\ \cline{1-7}
\multicolumn{1}{ |c| }{\multirow{2}{*}{2.4} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (total)} & $0.552\times10^{-3}$ & - & $0.647\times10^{-3}$ & $0.573\times10^{-3}$ & $0.510\times10^{-3}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (filaments)} & $7.265\times10^{-5}$ & - &$1.147\times10^{-5}$ & $3.231\times10^{-6}$ & $0.989\times10^{-6}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.033$ & - & $0.041$ & $0.042$ & $0.041$ \\ \cline{1-7}
\multicolumn{1}{ |c| }{\multirow{2}{*}{3.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (total)} & $0.524\times10^{-3}$ & $0.659\times10^{-3}$ & $0.667\times10^{-3}$ & $0.640\times10^{-3}$ & $0.619\times10^{-3}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (filaments)} & $9.163\times10^{-5}$ & $2.450\times10^{-5}$ &$2.154\times10^{-5}$ & $2.699\times10^{-6}$ & $1.884\times10^{-6}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.032$ & $0.045$ & $0.045$ & $0.047$ & $0.048$ \\ \cline{1-7}
\multicolumn{1}{ |c| }{\multirow{2}{*}{4.0} } &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (total)} & $0.351\times10^{-3}$ & $0.551\times10^{-3}$ & $0.536\times10^{-3}$ & $0.587\times10^{-3}$ & $0.619\times10^{-3}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$\Omega_{\rm HI}$ (filaments)} & $8.104\times10^{-5}$ & $3.706\times10^{-5}$ & $3.440\times10^{-5}$ & $3.589\times10^{-6}$ & $2.879\times10^{-6}$ \\ \cline{2-7}
\multicolumn{1}{ |c| }{} &
\multicolumn{1}{ c| }{$dN/dX$} & $0.022$ & $0.038$ & $0.038$ & $0.044$ & $0.048$ \\ \cline{1-7}
\end{tabular}
}
\end{center}
\caption{Values of the DLAs line density, $dN/dX$, $\Omega_{\rm HI}$ over the whole simulation and the value of $\Omega_{\rm HI}$ for the gas particles residing outside dark matter halos (we refer to that environment as filaments, see section \ref{HI_outside_halos_section} for details) in each N-body simulation obtained by assigning the HI to the gas particles according to the particle-based method.}
\label{Omega_HI_tab_Dave}
\end{table}
We follow pretty closely the method recently presented by Dav\'e et al. 2013 \cite{Dave_2013}. We now briefly review the method here and refer the reader to \cite{Dave_2013} for further details. Firstly, HI is assigned to each gas particle assuming photo-ionization equilibrium with the external UV background. For star forming particles we assume that hydrogen residing in the cold phase is fully neutral, while hydrogen in the hot phase is fully ionized. Thus, we set the HI/H fraction of star forming particles equal to its multi-phase cold gas fraction (see \cite{Springel-Hernquist_2003, Nagamine_2004} for further details).
The value of the HI photo-ionization rate, $\Gamma_{\rm HI}$, is tuned such as the Ly$\alpha$ mean transmission flux is reproduced. Next, the method assumes that each gas particle has a density profile given by the SPH kernel $W(r,r_{\rm SPH})$:
\begin{equation}
W(r,r_{\rm SPH}) = \frac{8}{\pi r_{\rm SPH}^3} \left\{
\begin{array}{l l}
1-6\left(\frac{r}{r_{\rm SPH}}\right)^2+6\left(\frac{r}{r_{\rm SPH}}\right)^3 & \quad 0\leqslant\frac{r}{r_{\rm SPH}}\leqslant\frac{1}{2}\\
\\
2\left(1-\frac{r}{r_{\rm SPH}}\right)^3 & \quad \frac{1}{2}<\frac{r}{r_{\rm SPH}}\leqslant 1\\
\\
0 & \quad \frac{r}{r_{\rm SPH}}> 1\\
\end{array} \right
\label{SPH_kernel}
\end{equation}
where $r_{\rm SPH}$ is the SPH smoothing length of the particle and a correction to account for self-shielding is applied if a radius $r_{\lim}$, such as
\begin{equation}
N_{\rm HI}(r_{\rm lim})=\frac{0.76m\left(\frac{\rm HI}{\rm H}\right)}{m_H}\int_{r_{\rm lim}}^{r_{\rm SPH}} W(r,r_{\rm SPH})dr=10^{17.2}~{\rm cm}^{-2}~,
\end{equation}
exists. In the above expression HI/H is the neutral hydrogen fraction obtained in the first step, $m$ is the gas particle mass and $m_H$ is the mass of the hydrogen atom. If such radius exists, then the method considers that the spherical shell from $r=0$ to $r=r_{\rm lim}$ is self-shielded againts the external radiation and assigns to it a HI/H fraction equal to 0.9. The HI/H fraction in the shell from $r=r_{\rm lim}$ to $r=r_{\rm SPH}$ is kept to the value obtained by assuming photo-ionization. Finally, the method takes into account the presence of molecular hydrogen, $H_2$, by assigning it to star forming particles using the observed ISM pressure relation found by the The HI Nearby Galaxy Survey (THINGS) \cite{Leroy_2008}
\begin{equation}
R_{\rm mol}=\frac{\Sigma_{H_2}}{\Sigma_{\rm HI}}=\left(\frac{P/k_B}{1.7\times10^{4}~{\rm cm^{-3}K}}\right)^{0.8}~.
\end{equation}
with $P$ being the pressure, $k_B$ the Boltzmann constant and $\Sigma_{\rm HI}$ and $\Sigma_{\rm H_2}$ are the HI and H$_2$ surface densities respectively.
We notice that our implementation of the above method only differs in the way in which the HI/H fractions are computed in photo-ionization equilibrium. Whereas Dav\'e et al. \cite{Dave_2013} uses a simplified hydrogen ionization balance computation, {\sc GADGET} uses a more refined method following \cite{Katz_1996}. We have however explicitly checked that differences are very small among the two methods.
In practice, we begin by extracting 5000 random QSO spectra from a given N-body simulation snapshot and by computing the mean transmitted flux. We tune the value of the photo-ionization rate such as we reproduce the observed mean transmitted flux from the Ly$\alpha$ forest \cite{Becker_2013} at the redshift of the N-body snapshot.
We then calculate, in case it exits, the radius $r_{\rm lim}$ for every gas particle in the snapshot and correct the neutral hydrogen content of it to account for self-shielding. Finally, for star forming particles, we correct their self-shielded HI mass to take into account the presence of molecular hydrogen. We compute the pressure of the gas particles using their density and internal energy.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\textwidth]{Figs/HI_halos.png}\\
\end{center}
\caption{For each dark matter halo of the simulation $\mathcal{B}120$ (cyan), $\mathcal{B}60$ (red), $\mathcal{B}30$ (green) and $\mathcal{B}15$ (blue), at $z=3$, we compute the neutral hydrogen mass, $M_{\rm HI}$, and the stellar mass $M_\star$. In the left panel we show the ratio $M_{\rm HI}/M_{\rm halo}$ as a function of the halo mass whereas the ratio $M_\star/M_{\rm halo}$ is displayed in the right panel.}
\label{M_HI_Dave}
\end{figure}
We apply the above method to the simulations $\mathcal{B}120$, $\mathcal{B}60$, $\mathcal{B}60$W, $\mathcal{B}30$ and $\mathcal{B}15$ and show in table \ref{Omega_HI_tab_Dave} the values we obtain for the DLAs line density and the parameter $\Omega_{\rm HI}$. In Fig. \ref{M_HI_Dave} we show the HI mass fraction and the stellar mass fraction for each dark matter halo in each simulation (for clearness we do not show the results for the simulation with galactic winds). We find that as the resolution of the simulation increases, halos of the same mass contain larger fractions of stellar mass, i.e. star formation is more efficient in the high resolution simulations than in the low ones; note however that the results seem to begin converging for the simulations $\mathcal{B}30$ and $\mathcal{B}15$. The consequence of this is that in the highest resolution simulations there are fewer gas particles, and thus, the neutral hydrogen mass that a dark matter halo host critically depends on resolution as it can be seen in the left panel of Fig. \ref{M_HI_Dave}. We notice that this problem was already discussed in \cite{Dave_2013} and can be also understood in terms of semi-analytic models of galaxy formation \cite{DeLucia_2008,Guo_2011}.
As can be seen from table \ref{Omega_HI_tab_Dave} the value of $\Omega_{\rm HI}$ critically depends on the size of the simulation box. This unphysical behavior arises due to two competing effects. On one hand we have that large box size simulations can only resolve the most massive halos, and therefore, in those simulations the contribution to $\Omega_{\rm HI}$ from low mass halos is not accounted for. On the other hand, small box size simulations (higher resolution simulations) convert gas to stars more effectively, thus, the amount of gas left in those simulations capable of host HI decreases with resolution. These two effect compete and at $z=2.4$ and $z=3$ we find that the value of $\Omega_{\rm HI}$ peaks for the simulations $\mathcal{B}60$.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/f_HI_Dave.pdf}\\
\end{center}
\caption{Column density distribution obtained by using the particle-based method at $z=2.4$ (left), $z=3.0$ (middle) and $z=4$ (right) on top of the simulations $\mathcal{B}15$ (solid blue), $\mathcal{B}30$ (dashed green, $\mathcal{B}60$ (dot-dashed red), $\mathcal{B}120$ (dotted cyan) and $\mathcal{B}60$W (purple triangles). The observational measurements of Noterdaeme et al. 2012 \cite{Noterdaeme_2012} are shown in black whereas those by Zafar
et al. 2013 \cite{Zafar_2013} are reported in gray.}
\label{f_HI_Dave}
\end{figure}
In Fig. \ref{f_HI_Dave} we show the HI column density distribution at redshifts $z=2.4$, $z=3$ and $z=4$. We find that this method fails to reproduce the abundance of the systems with the highest column densities. As expected, the results are sensitive to the resolution of the simulation. The formation of H$_2$ is reflected by a sudden suppression of the abundance of absorbers with column densities higher than $N_{\rm HI}\sim10^{21}-10^{21.5}~{\rm cm^{-2}}$. This transition is quite abrupt for the simulation $\mathcal{B}15$, in which the number of star forming particles with high pressure is significantly larger than in the other simulations.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/cross_section_Dave.png}\\
\end{center}
\caption{DLA comoving cross section for a subsample of halos at $z=2.4$ (left), $z=3$ (middle) and $z=4$ (right) from the simulations $\mathcal{B}120$ (cyan), $\mathcal{B}60$ (red), $\mathcal{B}30$ (green) and $\mathcal{B}15$ (blue) when the neutral hydrogen is assigned to the gas particles using the particle-based method. A fit to the results is presented with a solid orange line.}
\label{cross_section_Dave}
\end{figure}
We show the DLAs comoving cross-section obtained by using the particle-based model in Fig. \ref{cross_section_Dave}. In contrast to the results obtained by using the halo-based model 1,
where the DLA cross-section remains constant for very massive halos,
we find that the DLA cross-section increases with halo mass. However, the amplitude of the cross-section is significantly lower in this model in comparison to the one obtained by employing the halo-based model 2. Our results suggest that the minimum mass that a dark matter halo should have to host DLAs is about $\sim5\times10^{8}~h^{-1}{\rm M}_\odot$, almost independently of redshift. We find that the mean DLA cross-section can be well described by the following fitting formula
\begin{equation}
\sigma_{\rm DLA}(M)=\sigma_0\left(\frac{M}{M_0}\right)^\beta\exp{\left[-(M_0/M)^3\right]}~.
\label{fitting_formula_Dave}
\end{equation}
In table \ref{cross-section_Dave_table} we show the values that best fit our results and show these in Fig. \ref{cross_section_Dave} with orange lines. By using the average DLA cross-section and Eq. \ref{biaseq} we find a value for the bias of the DLAs equal to 1.48, 1.69 and 2.07 at $z=2.4, 3$ and 4 respectively. These values are very similar to those obtained by using the halo-based model 1, and thus, they are also in strong tension with the observational measurements by \cite{Font_2012}.
\begin{table}
\begin{center}
\resizebox{6cm}{!}{
\begin{tabular}{|c|c|c|c|}
\hline
$z$ & $\sigma_0$ & $M_0$ & $\beta$ \\
& $(h^{-1}{\rm kpc})^2$ & $(h^{-1}M_\odot)$ & \\
\hline
$2.4$ & $18.1$ & $7\times10^{8}$ & $0.721$ \\
\hline
$3.0$ & $29.1$ & $6\times10^{8}$ & $0.739$ \\
\hline
$4.0$ & $44.4$ & $5\times10^{8}$ & $0.750$ \\
\hline
\end{tabular}
}
\end{center}
\caption{Best fit values of the fitting formula \ref{fitting_formula_Dave} used to reproduce the average cross-section obtained by assigning HI to the gas particles using the particle-based model.}
\label{cross-section_Dave_table}
\end{table}
In Fig. \ref{Pk_HI_Dave} we display the power spectrum of the neutral hydrogen distribution for the different simulations at the redshifts $z=2.4$, $z=3$ and $z=4$. We stress that the mean HI mass per grid cell has been obtained by using $\Omega_{\rm HI}=10^{-3}$.
We find that the HI power spectrum depends on the simulation resolution, although results seem almost converged for the simulations $\mathcal{B}15$ and $\mathcal{B}30$. Our results indicate however that on large scales the HI power spectrum is converged even for the simulation $\mathcal{B}60$. The results of the simulation $\mathcal{B}120$ are slightly below those of the other simulations on large scales: this is due to the fact that this simulation can not resolve the smallest halos that can host HI, which is then reflected on the amplitude of the HI power spectrum on large scales. We show the HI bias in the bottom panels of Fig. \ref{Pk_HI_Dave}. The dependence of the results with resolution is clear from the plots, even though the HI bias seems almost converged for the simulations $\mathcal{B}15$ and $\mathcal{B}30$. Therefore, the differences we find in terms of the HI power spectrum for those simulations are mainly due to cosmic variance and box size effects rather than resolution.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/Pk_HI_Dave.pdf}\\
\end{center}
\caption{HI power spectrum at redshift $z=2.4$ (left), $z=3$ (middle) and $z=4$ (right) obtained by assigning neutral hydrogen to the gas particles using the particle-based method.
Results are shown for the simulations $\mathcal{B}15$ (solid blue), $\mathcal{B}30$ (dashed green), $\mathcal{B}60$ (dot-dashed red), $\mathcal{B}120$ (dotted cyan) and $\mathcal{B}60$W (purple triangles). The vertical lines represent the value of the Nyquist frequency for the different simulations. We display the bias between the distributions of HI and matter, $b_{\rm HI}^2(k)=P_{\rm HI}(k)/P_{\rm m}(k)$, in the bottom panels.}
\label{Pk_HI_Dave}
\end{figure}
We now discuss the role played by the galactic winds in our results. As can be seen from table \ref{Omega_HI_tab_Dave}, the simulation $\mathcal{B}60$W contains a similar amount of HI as the simulation $\mathcal{B}60$ (the resolution is exactly the same in both simulations) at all redshifts. The reason for this is due to two effects going in opposite directions. On one hand, galactic winds suppress the star formation rate which means less star forming particles and thus, less HI. On the other hand the gas particles that experience galactic winds are hydrodynamically decoupled for a certain time (or until their density is below some threshold). During that time, radiative cooling is very efficient, which means that the HI/H fraction computed assuming photo-ionization equilibrium is close to 1. Therefore, these two effects conspire to produce a similar total amount of HI. In terms of the HI column density distribution the simulations with and without winds produce very similar results. Finally,
galactic winds also leave their signature on the HI power spectrum as can be seen from Fig. \ref{Pk_HI_Dave}. We find that the amplitude of the HI power spectrum is slightly higher in the simulation with galactic winds than in the simulation without winds.
This is due to the fact that galactic winds can escape from low-mass halos but are not able to leave the most massive ones. In the simulation with galactic winds, the most massive (and biased) halos will contribute more to the HI power spectrum than the low mass halos, producing an enhancement on the HI power spectrum as the one we obtain.
\section{Introduction}
The cosmology standard model is able to reproduce reasonably well a variety of observables from large to small scales like the fluctuations in the cosmic microwave background (CMB), the spatial distribution of galaxies, the weak lensing statistical properties, the clustering of the cosmic-web filaments as seen in Lyman-$\alpha$ , the cluster number counts, etc.
Among the different tracers of the large scale structure, Hydrogen is the most abundant element with a contribution of $\sim75\%$ to the total baryonic mass. The Lyman-$\alpha$ forest, in the post-reionization era, is a powerful probe of the diffuse hydrogen atoms
that are ionized due to the presence of a rather strong ultraviolet (UV) background radiation flux. However, it is expected that the UV background radiation would not be able to penetrate dense hydrogen environments associated to galaxies. Thus, the neutral hydrogen within massive galaxies is believed to be self-shielded agains the external UV radiation. It is also thought that the neutral hydrogen within galaxies represents an intermediate phase between the diffuse inter-galactic medium and the high density molecular hydrogen, the fuel for the formation of stars. Therefore, any successful theory of galaxy formation/evolution should predict not only the galactic HI mass but also the interplay between the galactic properties and the HI content and distribution.
The amount of neutral hydrogen can be parametrized with the parameter $\Omega_{\rm HI}(z)=\bar{\rho}_{\rm HI}(z)/\rho_{\rm c}^0$, with $\bar{\rho}_{\rm HI}(z)$ being the average comoving density of neutral hydrogen (HI) at redshift $z$ and $\rho_{\rm c}^0$ the Universe critical density at $z=0$. Observations of the abundance and properties of the DLAs have been used to infer a value of $\Omega_{\rm HI}(z)\simeq10^{-3}$, almost independently of redshift \cite{Peroux_2003, Zwaan_2005, Rao_2006, Lah_2007, Martin_2010, Braun_2012, Noterdaeme_2012, Zafar_2013}.
Neutral hydrogen can be detected through its redshifted 21-cm emission (see \cite{Furlanetto_2006, Morales_2009} for a recent reviews). Even though galaxies can not
be detected individually at high redshift, variations in the 21 cm signal
on large scales can be used to measure the power spectrum: a technique which is known as intensity mapping \cite{Bharadwaj_2001A, Bharadwaj_2001B,Chang_2008,Loeb_Wyithe_2008}. Moreover, the fluctuations in the 21 cm intensity emission are believed to be one of the most powerful probes of the epoch of reoinization (EoR) (see for instance \cite{Furlanetto_2006, Zaldarriaga_2004, Datta_2007, Morales_2009, Pritchard_2011} and references therein). In the post-reionization era, the much larger volume available for 21 cm observations, with respect to those probed by galaxy surveys, is expected to allow us to measure the matter power spectrum with an unprecedented precision \cite{Loeb_Wyithe_2008, Wyithe_2008, Camera_2013,Bull_2014}.
The current radio-telescope technology only allows us to detect the 21 cm emission from galaxies only at relatively low redshift ($z\lesssim0.3$). On the other hand,
existing radio interferometers such as the Giant Meterwave Radio Telescope (GMRT)\footnote{http://gmrt.ncra.tifr.res.in/}, the Ooty Radio Telescope (ORT) \cite{Ooty}, the Low-Frequency Array (LOFAR)\footnote{http://www.lofar.org/}, the Murchison Wide-field Array (MWA)\footnote{http://www.mwatelescope.org/} or upcoming like ASKAP (The Australian Square Kilometer Array Pathfinder)\footnote{http://www.atnf.csiro.au/projects/askap/index.html}, MeerKAT (The South African Square Kilometer Array Pathfinder)\footnote{http://www.ska.ac.za/meerkat/} and SKA (The Square Kilometer Array)\footnote{https://www.skatelescope.org/} are designed to detect the fluctuations in the emission of 21 cm neutral hydrogen. We notice that the 21 cm signal has recently been detected at $z\sim1$ by cross-correlating intensity maps and the large scale structure (via catalogues of optically selected galaxies) \cite{Chang_2010,Masui_2013} and also forecasting (see for instance \cite{Ghosh_2010,Ghosh_2011}). Therefore, a deep understanding of the HI distribution is required from the theory/numerical side in order to extract the maximum possible information from the data.
In this paper we aim at modeling the distribution of HI using high resolution hydrodynamical N-body simulations. The HI distribution has been previously studied in several works \cite{Bharadwaj_2004,Nagamine_2004, Kohler_2006, Pontzen_2008, Razoumov_2008, Tescari_2009, Altay_2010, Popping_2009, Bagla_2010, Nagamine_2010, McQuinn_2011, Fumagalli_2011,Duffy_2012, Bird_2013, Rahmati_2013, Dave_2013, Marinacci_2014, Bird_2014}. Here, we use two different approaches to simulate the distribution of HI. With the first one, the so-called \textit{halo-based method}, we only assign HI to the gas particles belonging to dark matter halos. This approach is based on Bagla et al. 2010 \cite{Bagla_2010} method. We will then use two different schemes: one it is just the Bagla et al. method whereas with the other we try to reproduce the observational measurements of both the DLAs column density distribution and their bias as obtained recently by the SDSS-III/BOSS collaboration \cite{Font_2012}. The two schemes are quite different and by exploring both of them in terms of 21 cm power spectrum we are conservatively exploring the expected signal.
Our second approach, consists of assigning HI to the N-body gas particles using a particle-by-particle basis and make use of the Dav\'e et al. 2013 \cite{Dave_2013} method, which accounts for the most important effects that set the hydrogen phase: photo-ionization equilibrium in low density regions and self-shielding and HI-H$_2$ conversion in high density zones. Ref.~\cite{Dave_2013} showed that this method, applied to their simulations,
is able to reproduce many of the low-redshift HI observations \cite{Meyer_2004, Giovanelli_2005, Catinella_2010}. Moreover, this method allows us to quantify the contribution of HI residing outside halos to the overall Universe HI content and to the HI power spectrum.
We also investigate the sensitivity of future radio telescopes such as the SKA to the 21 cm signal arising from the spatial distribution of neutral hydrogen. Furthermore, we study the prospect of direct imaging the largest peaks in the HI distribution.
The paper is organized as follows. In Sec. \ref{Simulations} we present the hydrodynamical N-body suite of simulations used in this paper. The methods employed to model the distribution of neutral hydrogen are described in Sections \ref{HI_halos} and \ref{HI_particles}. In the first, we focus on the halo-based method, whereas in the latter we use the particle-based method. In Sec. \ref{HI_outside_halos_section} we quantify the contributions of HI outside dark matter halos to both the overall amount of HI in the Universe and the HI power spectrum. The 21 cm power spectrum from our simulated HI distribution is presented in Sec. \ref{21cm_section}. In that section we also discuss its detectability for future radio surveys as SKA. In Sec. \ref{imaging} we investigate the possibility, for the SKA, to detect bright, isolated, HI peaks. Finally, a summary and the main conclusions of this paper are given in Sec. \ref{Conclusions}.
\section{Prospects of imaging}
\label{imaging}
In this section we investigate the prospects of imaging the HI distribution using radio telescopes. In general, the 21 cm signal in an average region of the Universe may not be strong enough for imaging, hence we focus on regions of very high densities where the concentration of HI is expected to be considerably larger than average. In particular we investigate whether SKA1-low and SKA1-mid might be able to detect a few individual bright HI peaks which we find in our simulations.
We begin by creating brightness temperature maps from our simulated distribution of neutral hydrogen. The procedure used to produce these maps is as follows: Given a frequency channel $[\nu_0-\triangle \nu/2,~\nu_0+\triangle \nu/2]$ of width $\triangle \nu$, with $\nu_0=1420/(1+z)$ MHz, we take a slice of the $N$-body snapshot at redshift $z$ with a width equal to $L=r_{\nu_0-\triangle \nu/2}-r_{\nu_0+\triangle \nu/2}$, where $r_{\nu}$ is the comoving distance to redshift $z=(1420~{\rm MHz})/\nu-1$. We then divide the slice into $N_{\rm pixels}\times N_{\rm pixels}$ cells of equal volume, where $N_{\rm pixels}$ determines the resolution of the map. For each cell we compute the HI density and use Eq. \ref{delta_Tb} to calculate the brightness temperature excess within it. Next, we compute the specific intensity excess from the brightness temperature excess using the relation
\begin{equation}
I_\nu = \frac{2\nu^2}{c^2}k_B\delta{T_b},
\end{equation}
where $k_B$ is the Boltzmann constant and $c$ is the light speed. In order to account for instrumental resolution, we smooth the $I_\nu$ field with a Gaussian window of angular radius $\theta_{\rm beam}$, with $\theta_{\rm beam}$ being the synthesized beam width. Since radio-interferometers are not sensitive to the mean value of the specific intensity and can only measure deviations from the mean, we correct the value of the specific intensity in each pixel by making the transformation $I_\nu \rightarrow I_\nu - \langle I_\nu \rangle$. We then multiply the $I_\nu$ map with the beam solid angle $\Delta \Omega_{\rm beam}=\theta^2_{\rm beam}$ which essentially gives the total flux within a single beam, i.e., the resultant image maps will be in units of flux per beam. We repeat the above procedure with different slices taken from the simulation box until we find the slice containing the highest value of $I_\nu$.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/RM_500kHz_z=3.png}\\
\end{center}
\caption{Mock radio-maps, for a frequency channel of 500 kHz, created from the HI distribution obtained by using the halo-based model 1 (top row), the halo-based model 2 (middle row) and the particle-based method (bottom row) on top of the simulation $\mathcal{B}60$ at $z=3$. Three different synthesized beam widths has been used: $\theta_{\rm beam}=2'$ (left column), $\theta_{\rm beam}=1'$ (middle column) and $\theta_{\rm beam}=0.5'$ (right column). The maps do not include the system noise associated to radio-telescopes.}
\label{radio_maps}
\end{figure}
In Fig. \ref{radio_maps} we show mock radio-maps created using the above procedure where the HI distributions have been obtained using the halo-based and the particle-based methods on top of the simulation $\mathcal{B}60$ at $z=3$. The simulated maps are for a frequency channel width of 500 kHz and for three different synthesized beam widths of $2'$, $1'$ and $0.5'$. We should mention that these are not proper radio-maps as would be observed in telescopes, since we have used a very large number of pixels\footnote{Notice that the purpose of using a large number of pixels is just to show a smooth image rather than the pixelated one.} ($N_{\rm pixels}=1024$) and we have not included the system noise. In this figure, we only aim to show the specific intensity field that will be sampled by the radio-telescopes.
We find that the maps, for a given $\theta_{\rm beam}$, look quite similar. However, the peak amplitude obtained in the halo-based model 2 is significantly higher than the one found by using the other two methods (typically a factor $\sim2-3$). This is because the HI fraction in the large halos for halo-based model 2 is much higher than in the other two models.
We now investigate whether isolated peaks can be directly imaged by the future SKA1-low and SKA1-mid radio-telescopes, i.e, whether there will be enough sensitivity in the maps to identify the cosmological HI. The noise r.m.s. per synthesized beam in radio images for two polarizations can be written as
\begin{equation}
\Delta S_{\nu}= \frac{\sqrt{2}k_BT_{\rm sys}}{A_e\sqrt{N_b ~\triangle \nu~ t_0}},
\label{eq:noise-beam}
\end{equation}
where $N_b=N(N-1)/2$ is the total number of instantaneous baselines. We have already defined $\triangle \nu$ as the frequency channel width, $T_{\rm sys}$ as the instrument temperature and $t_0$ as the observation time. When the number of antennae is large ($N>>1$) the above equation can be written as
\begin{equation}
\Delta S_{\nu}= 65.7 \, {\mu \rm Jy}\left (\frac{T_{\rm sys}}{100 \, K}\right)\left (\frac{0.7}{\epsilon}\right)\left (\frac{100 \, {\rm m}^2}{A}\right)\left (\frac{100}{N}\right)\left (\frac{1 \, {\rm MHz}}{\triangle\nu}\right)^{0.5}\left (\frac{100 \, {\rm hrs}}{t_0}\right)^{0.5}.
\label{eq:noise-beam1}
\end{equation}
We calculate the noise r.m.s. values for three different synthesized beams of $0.5'$, $1'$ and $2'$. As we will see later, the sensitivity for beam sizes smaller than the $0.5'$ is very poor and therefore we do not consider smaller beams. Higher beam sizes smooth out the HI signal in peaks, hence we avoid using higher beam sizes. To achieve a desired synthesized beam width, $\theta_{\rm beam}$, we consider antennae only from the central region which will provide baselines up to $U_{\rm max}=1/\theta_{\rm beam}$. We simply discard antennae outside the central region which do not contribute to baselines $U \leq U_{\rm max}$. For example, we need a maximum baseline $U_{\rm max} \sim 6879, \, 3440 $ and $1720$ to achieve the synthesized beam sizes of $0.5'$, $1'$ and $2'$ respectively. Thus, at redshift $z=4$ (corresponding to the observing wavelength $\lambda=1.056$m) we take into account the antennae which are within a circular region of diameter $7264$ m, $3632$ m and $1816$ m from the core centre. Fig. \ref{fig:cum_frac} shows the cumulative fraction of antennae (stations) as a function of distance from the core centre for the SKA1-mid and SKA1-low. The cumulative fraction of antennae we use here is consistent with the antenna distribution given in the SKA baseline design document mentioned in the previous section (see table 6 and figure 3 in the document). We can calculate the total number of antennae lying within a given radius using Fig. \ref{fig:cum_frac}. For SKA1-low, at frequencies corresponding to $z=4$, there are $866$, $830$ and $600$ antennae within distances corresponding to synthesized beam widths of $0.5'$, $1'$ and $2'$ respectively. We note that we do not necessarily lose a large amount of sensitivity in the images while discarding antennae at large distances from the core. This is because the antenna distribution is highly condensed in the central region. The advantage in discarding antennae at large distances is that the noise calculation in images becomes considerably simpler. The same argument holds for the SKA1-mid at redshifts $z=2.4$ and $z=3$.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{Figs/cum_frac_stations.pdf}\\
\end{center}
\caption{The cumulative fraction of antennae (stations) as a function of distance from the core centre for the SKA1-mid and SKA1-low.}
\label{fig:cum_frac}
\end{figure}
\begin{table}[ht]
\caption{The brighest HI peak at redshift $z=4$ detection prospects for the SKA1-low with $1000$ hrs of observations}
\centering
\vspace{.2in}
\begin{tabular}{c rrrrrrrrr}
\hline
\hline
Frequency & \multicolumn{3}{c}{$\theta_{\rm beam}=0.5'$} & \multicolumn{3}{c}{$\theta_{\rm beam}=1'$} & \multicolumn{3}{c}{$\theta_{\rm beam}=2'$} \\
Channel& Peak flux & noise & SNR & Peak flux & noise & SNR & Peak flux & noise & SNR \\
(kHz)&($\mu$Jy) & ($\mu$Jy) & &($\mu$Jy) &($\mu$Jy)& &($\mu$Jy)&($\mu$Jy)& \\
\\ [0.5ex]
\hline
125 & 1.79 & 0.77 & 2.32 & 3.20 & 0.81 & 3.95 & 5.20 & 1.12 & 4.64 \\
\hline
500 & 1.34 & 0.39 & 3.44 & 2.57 & 0.40 & 6.42 & 4.19 & 0.56 & 7.48 \\
\hline
1000 & 0.71 & 0.27 & 2.63 & 1.40 & 0.29 & 4.83 & 2.47 & 0.40 & 6.18 \\
\hline
\end{tabular}
\label{tab:ska1-low}
\end{table}
\begin{table}[ht]
\caption{The brighest HI peak at redshift $z=2.4$ detection prospects for the SKA1-mid with $1000$ hrs of observations}
\centering
\vspace{.2in}
\begin{tabular}{c rrrrrrrrr}
\hline
\hline
Frequency & \multicolumn{3}{c}{$\theta_{\rm beam}=0.5'$} & \multicolumn{3}{c}{$\theta_{\rm beam}=1'$} & \multicolumn{3}{c}{$\theta_{\rm beam}=2'$} \\
Channel& Peak flux & noise & SNR & Peak flux & noise & SNR & Peak flux & noise & SNR \\
(kHz)&($\mu$Jy) & ($\mu$Jy) & &($\mu$Jy) &($\mu$Jy)& &($\mu$Jy)&($\mu$Jy)& \\
\\ [0.5ex]
\hline
125 & 5.82 & 5.25 & 1.11 & 9.48 & 6.44 & 1.47 & 14.3 & 8.22 & 1.74 \\
\hline
500 & 4.79 & 2.62 & 1.83 & 7.93 & 3.22 & 2.46 & 12.1 & 4.11 & 2.94 \\
\hline
1000 & 3.32 & 1.85 & 1.79 & 5.43 & 2.28 & 2.38 & 8.57 & 2.90 & 2.54 \\
\hline
\end{tabular}
\label{tab:ska1-mid}
\end{table}
Tables \ref{tab:ska1-low} and \ref{tab:ska1-mid} show the noise r.m.s. values for the SKA1-low at redshift $z=4$ and the SKA1-mid at $z=2.4$ for three different frequency channels for $1000$ hours of observations. Note that the time of observation is 10 times higher than what was used for calculating the noise in power spectra measurements. For the values of the parameters $T_{\rm sys}$, $A$, $\epsilon$, $N$ we simply use those quoted in the previous section. We also quote the peak HI flux obtained from our simulations using the halo-based model 2 in tables \ref{tab:ska1-low} and \ref{tab:ska1-mid}. As we have seen above, the peak flux would be a factor $\sim 2-3$ lower for the other two methods.
We find that at $z=4$, for synthesized beam widths of $\theta_{\rm beam}=1'$ and $2'$, the SKA1-low will be able to detect the brightest HI peaks with a high signal to noise ratio (SNR$> 4.5$), while the SNR is relatively smaller $\lesssim 3$ for $\theta_{\rm beam} = 0.5'$. For the parameter values we have explored, it seems that SKA1-low will achieve the highest SNR for a $500$ kHz channel with a $2'$ synthesized beam, resulting in a SNR higher than 7. For higher frequency channel width the SNR starts to decline as the peak flux drops rapidly. This is because of the fact that the HI density around high density peaks follow the density profile of the halo and thus does not extend very far.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\textwidth]{Figs/RM_noise.png}\\
\end{center}
\caption{Detectability of HI peaks in radio-maps. The top-left panel shows a radio-maps for a frequency channel of 500 kHz and for a synthesized beam width of $2'$ at z=4 obtained from the HI distribution generated by using the halo-based model 2 on top of the simulation $\mathcal{B}60$. In the top-right panel we show the same map but with a pixel size in correspondence with $\theta_{\rm beam}$ and with noise added for each pixel according to the noise r.m.s. The bottom panels display the same but at $z=2.4$.}
\label{radio_maps_noise}
\end{figure}
At lower redshift the SKA1-mid may be able to detect only the brightest peaks, albeit with a very low SNR. From our results we find that the most promising case would be to detect a peak using a synthesized beam width of $2'$ with a frequency channel width of 500 kHz. In this case, the signal to noise ratio would be close to 3. Hence, with the current specifications, the prospect of imaging cosmological HI with SKA1-mid at the redshifts studied in this paper is not very promising.
In order to visualize the difficulties in detecting peaks, we show in Fig. \ref{radio_maps_noise} radio-maps with instrumental noise included. The left hand panels in the figure show maps smoothed over the beam size, but using $1024 \times 1024$ pixels and without adding any noise (i.e, these maps are similar to those presented in Fig. \ref{radio_maps}). In the right column panels show the same maps made using a pixel size in correspondence with the synthesized beam width and with noise added according to the noise r.m.s. value. The top panels are for $z = 4$ (i.e., SKA1-low) while the bottom panels are for $z = 2.4$ (SKA1-mid). The regions of peak flux, as is clear from the left panels, lie in the top-right corner of the maps. From the figure it seems that we can possibly detect the HI in the density peaks for SKA1-low as it looks quite distinct from the noise fluctuations. On the other hand, it will be extremely difficult to image the bright HI peak with SKA1-mid as the signal is virtually indistinguishable from the noise fluctuations.
We should remind the reader that the results presented in tables \ref{tab:ska1-low} and \ref{tab:ska1-mid} have been obtained by using the halo-based model 2. The other two models, i.e., the halo-based model 1 and the particle-based model, predict lower peak flux values and hence are much more difficult to image. On the other hand it is expected that the mass of the most massive halos at those redshifts would be significantly higher than the one used in this analysis, due to the small simulation volumes of our simulations. An interesting consequence of this calculation is that the detection of isolated bright HI peaks by SKA can provide some indication on the amount of HI present in very high density peaks.
\section{Power spectrum: method}
\label{power_spectrum_method}
The power spectrum measurements presented in the paper have been obtained through the standard procedure of assigning the particle positions (or any other quantity associated to them as the HI mass) to a regular cubic grid using the Cloud-in-Cell (CIC) interpolation technique. By doing this, it is implicitly assumed that each particle represents a cube of size $L/N_{\rm grid}$, with $L$ being the size of the simulation box and $N_{\rm part}$ the number of points along one axis in the grid, with a uniform interior density. However, the actual density profile of the gas particles it is not given by a uniform cube, but it is instead described by the SPH kernel (Eq. \ref{SPH_kernel}). In order to investigate how much our results are affected by the fact of having computed the power spectrum using the CIC, i.e. by ignoring the internal structure of the gas particles, we have calculated the power spectrum taking into account SPH kernel of
the gas particles.
The procedure used is as follows: for each gas particle we select $N^3$ points (we have verified that with $N>7$ our results are converged) within the SPH smoothing length in such a way that each of them represent the same interior volume. We then assign each gas particle interior point to the grid cell that it belongs to. Finally, we follow the standard procedure of computing the FFT of the field obtained in that way and calculating the power spectrum $P(k)$. We have assigned HI to the gas particles of the simulation $\mathcal{B}30$ at $z=3$ using the halo-based model 1 and computed the HI power spectrum using the standard procedure (CIC) and the above one which takes into account the gas particles SPH kernel. We show the results in Fig. \ref{Pk_HI_test}. We find that the at large scales both procedures yield to identical results while at smaller scales the results, as expected, differ. We notice that whereas we have corrected the amplitude of the modes to take into account the CIC assignment, we have not implemented that correction when using the SPH kernel of the gas particles (for details about this correction see \cite{Jing_2005}). In Fig. \ref{Pk_HI_test} the results obtained by using the CIC interpolation procedure but not correcting the modes amplitude are represented by the red curves. It is clear that the mode correction plays a very important role when computing the power spectrum and thus, in order to quantify the scale at which both procedures begin to diverge it is needed to implement the mode correction to account for the SPH kernel. Unfortunately, the implementation of the amplitude mode correction to account for the SPH kernel is beyond the scope of this paper. We have explicitly checked that similar results are obtained by using different simulations and different HI assignments.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{Figs/Pk_HI_comparison.pdf}\\
\end{center}
\caption{HI power spectrum, obtained by using the halo-based model 1, for the simulation $\mathcal{B}30$ at $z=3$ using different methods. The black lines represent the results obtained using the standard procedure: interpolation through CIC, FFT, mode correction and measurement of the P(k). The red lines display the results of using the above procedure but without correct the modes amplitude to take into account the CIC assignment to the grid. The blue lines show the results of computing the power spectrum taking into account the SPH kernel of each gas particle (without mode correction). Two grids have been used: $1024^3$ (solid lines) and $512^3$ (dashed lines) points. The value of the Nyquist frequency for the two grids are represented by vertical lines.}
\label{Pk_HI_test}
\end{figure}
In summary, both procedures yield equivalent results on large scales. On small scales some differences show up, but in order to quantify the discrepancy a correction to the modes amplitude needs to be implemented. For the scales of interest in the present work, we can safely interpolate with CIC.
\section{N-body simulations}
\label{Simulations}
We run a set of five high-resolution hydrodynamical N-body simulations
including both cold dark matter (CDM) and baryon (gas + star) particles. We
use the TreePM-SPH N-body code {\sc GADGET-III}, which is an improved and
faster version of the code {\sc GADGET-II} \cite{Springel_2005}.
The starting redshift of the simulations is set to $z=99$. At this
redshift, the N-body initial conditions are generated by displacing
the particle positions from a cubic grid using the Zel'dovich
approximation. We make use of CAMB \cite{CAMB} to compute the CDM and
baryon transfer functions together with the matter power spectrum at
the starting redshift.
The simulations contain $512^3$ CDM and $512^3$ baryon particles
within a periodic box of linear comoving size 15, 30, 60 and 120
$h^{-1}$ Mpc. Thus, the resolution of the 15 $h^{-1}$ Mpc simulation
is 512 times higher than the one of the simulation with box size equal to
120 $h^{-1}$ Mpc. The reason of having different simulations with
different resolutions is because we will perform convergence tests
along the paper, investigating the stability of our results.
The gravitational softening of the particles (for both CDM and
baryons) is set to $1/30$ of the mean inter-particle linear spacing,
corresponding to $\sim1,2,4$ and 8 $h^{-1}$ com. kpc for the simulations
with boxes of 15, 30, 60 and 120 $h^{-1}$ com. Mpc, respectively. The values
of the cosmological parameters, in roughly good agreement with the latest
Planck results \cite{Planck_2013}, are the same for all the
simulations: $\Omega_{\rm m}=\Omega_{\rm cdm}+\Omega_{\rm b}=0.2742$,
$\Omega_{\rm b}=0.046$, $\Omega_\Lambda=0.7258$, $h=0.7$, $n_s=0.968$
and $\sigma_8=0.816$.
The code includes radiative cooling by hydrogen and helium and heating
by a uniform Ultra Violet (UV) background, while star formation is
implemented using the effective multiphase sub-grid model of Springel
\& Hernquist \cite{Springel-Hernquist_2003}. The UV background and
the cooling routine have been modified in order to achieve a wanted
thermal history which corresponds to the reference model of
\cite{viel13} which has been shown to provide a good fit to the
statistical properties of the transmitted Lyman-$\alpha$ flux. In this
model, hydrogen reionization takes place at $z\sim 12$ and the
temperature-density relation for the low-density IGM
$T=T_0(z)(1+\delta)^{\gamma(z)-1}$ has $\gamma(z)=1.3$ and
$T_0(z=2.4,3,4)=(16500,15000,10000)$ K.
One of the simulations is run with feedback in the form of energy
driven galactic wind as implemented in Ref. \cite{Springel-Hernquist_2003}:
star forming particles receive 'a kick' of $\sim480$ km/s according to
a probabilistic criterion. Hydrodynamical forces are then turned off
for a fixed time interval or until the density of the wind particle
reaches a value of $0.1\rho_{\rm th}$, whichever happen earlier, with
$\rho_{\rm th}$ is the density above which star formation takes place.
For further details we refer the reader to Refs.
\cite{Springel-Hernquist_2003, Barai_2013}.
A summary of the simulation suite is presented in table
\ref{tab_sims}.
\begin{table}
\begin{center}
\resizebox{14cm}{!}{
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Name & Box & $m_{\rm CDM}$ & $m_{\rm b}$ & wind model & $z_{\rm end}$ \\
& ($h^{-1}\rm{Mpc}$) & ($h^{-1}M_\odot$) & ($h^{-1}M_\odot$) & & \\
\hline
\hline
$\mathcal{B}120$ & 120 & $8.16\times10^{8}$ & $1.64\times10^{8}$ & no winds & 2.4\\
\hline
$\mathcal{B}60$W & 60 & $1.02\times10^{8}$ & $2.05\times10^{7}$ & constant velocity winds & 3.0\\
\hline
$\mathcal{B}60$ & 60 & $1.02\times10^{8}$ & $2.05\times10^{7}$ & no winds & 2.4 \\
\hline
$\mathcal{B}30$ & 30 & $1.28\times10^{7}$ & $2.56\times10^{6}$ & no winds & 2.4 \\
\hline
$\mathcal{B}15$ & 15 & $1.59\times10^{6}$ & $3.20\times10^{5}$ & no winds & 2.4\\
\hline
\end{tabular}
}
\end{center}
\caption{Summary of the simulations. The value of the cosmological
parameters is the same for all simulations: $\Omega_{\rm
m}=\Omega_{\rm cdm}+\Omega_{\rm b}=0.2742$, $\Omega_{\rm
b}=0.046$, $\Omega_\Lambda=0.7258$, $h=0.7$, $n_s=0.968$ and
$\sigma_8=0.816$. Each simulation contains $512^3$ CDM and $512^3$
baryon particles.}
\label{tab_sims}
\end{table}
The simulations without galactic winds have been run until $z=2.4$,
whereas the simulation with galactic winds has been run only until
$z=3$. We analyzed snapshots at redshifts $z=2.4,3,4,6$ (except for
the simulation with galactic winds where we only keep snapshots at
$z=3,4,6$), on top of which we have run the Friends-of-Friends (FoF)
\cite{FoF} algorithm with a value of the linking length parameter
equal to $b=0.2$.
For each gas particle, we consider its mass, SPH smoothing length and
HI/H fraction. The neutral hydrogen fraction, together with the
abundance of other ionization states of hydrogen and helium, is
computed by the code assuming optically thin gas and photo-ionization
equilibrium with the external UV background using the formalism
described in \cite{Katz_1996}. In the version used in the present
work, we stress that the code does not take into account neither
HI self-shielding effects nor the formation of molecular
hydrogen. However, the HI content outside halos and the IGM physical
state is realistic in the sense that it fits all the observational
constraints.
|
\section{Introduction}\label{sec:intro}
The first generation of stars are believed to have had a significant impact on the cosmic
evolution through their radiative, mechanical and chemical feedback~\citep[e.g.,][]{Ciardi2005}.
Stellar extreme-ultraviolet (EUV) photons~($> 13.6$ eV) ionize the circumstellar gas
to form an HII region, thereby initiating the process of the reionization of the intergalactic medium~(IGM).
Even the single supernova (SN) of a massive ($\ga 150{\rm M}_{\odot}$) first-star is able to totally
evacuate the gas and suppresses subsequent star formation in the same minihalo~\citep{Bromm2003,Wada2003}
although some of the materials blown out by the SN of a less-massive
($20\mbox{-}40{\rm M}_{\odot}$) progenitor may fall back again and
commence another episode of star formation~\citep{Ritter2012}.
The ejected gas by a SN contributes to the metal enrichment of the IGM.
The extent of such feedback depends crucially on the initial mass function (IMF), as well as
the star formation rate (SFR), of first stars.
Stars formed from the metal-free primordial gas, or Population III (Pop III) stars,
have been classified into two sub-populations depending on
whether the gas is still in an undisturbed pristine state
with a low electron fraction $\sim 10^{-4}$~(Pop III.1),
or the gas has already been affected by the radiative or kinematic feedback
from pre-existing stars~(Pop III.2), following the nomenclature proposed by O'Shea et al. (2008)\footnote{Previously, Pop III.2 stars were called in different names from authors to authors,
for example, Pop II.5 stars (Mackey, Bromm \& Hernquist 2003).}.
Despite with the same primordial composition, the thermal evolution of the gas
is fairly different in those cases, resulting in different characteristic stellar masses.
In the standard $\Lambda$CDM cosmology,
Pop III.1 stars are predicted to be formed in minihalos of $\sim 10^6\ {\rm M}_{\odot}$
~\citep{Couchman1986,Haiman1996b,Tegmark1997}.
During its protostellar collapse, cooling is entirely by ${\rm H}_2$, which is
produced by the following electron-catalyzed reaction~\citep[so-called the ${\rm H}^-$ channel;][]{Peebles1968, Hirasawa1969}:
\begin{equation}
{\rm H} + e \rightarrow {\rm H}^- + \gamma,
\label{eq:H-1}
\end{equation}
\begin{equation}
{\rm H}^- + {\rm H} \rightarrow {\rm H}_2 + e.
\label{eq:H-2}
\end{equation}
With the high value of the minimum temperature $200\ \rm K$ attained at the density $\sim 10^{4}\ {\rm cm^{-3}}$,
the fragmentation mass is as large as $\sim 10^{3}\ {\rm M_{\odot}}$
~\citep{Bromm1999, Bromm2002, Abel2002, Yoshida2006},
which is set by the Jeans mass at this epoch.
In this case, the final mass of formed stars is set by the radiative feedback from the growing protostar,
and its value is a few 10 - a few 100 ${\rm M_{\odot}}$, much higher than the present-day counterpart~\citep{Omukai_Palla2001,Omukai_Palla2003,McKee2008,Yoshida2008, Hosokawa2011, Stacy2012, Susa2013, Hirano2014}.
On the other hand, Pop III.2 stars are formed from the metal-free gas affected either by
the radiative or kinematic feedback from earlier generation of stars, typically in atomic-cooling halos
with virial temperatures $\gtrsim 10^4\ {\rm K}$~\citep{O'Shea2008b,Greif2008}.
An example of Pop III.2 stars is those formed in the relic HII regions of defunct Pop III
stars~\citep{Kitayama2004, Whalen2004, Yoshida2007a, Abel2007}.
In this case, a larger amount of ${\rm H}_2$ is produced~($\sim 10^{-3}$) via the ${\rm H}^-$ channel~(Eqs. \ref{eq:H-1} and \ref{eq:H-2}) than in the Pop III.1 case~($\sim 10^{-4}$) \citep[e.g.,][]{O'Shea2005}.
Once the temperature plummets below $150\ \rm K$ by this ${\rm H}_2$ cooling, deuterium is rapidly converted into HD via the exothermic reaction,
\begin{equation}
{\rm D}^+ + {{\rm H}_2} \rightarrow {\rm HD} + {\rm H}^+.
\label{eq:HD}
\end{equation}
This cooling by HD lowers the temperature further to $\sim 30\ \rm K$~\citep[e.g.,][]{Nagakura2005}.
The characteristic mass of stars formed in such an environment is predicted to be $\sim 10\ {\rm M_{\odot}}$~\citep{Uehara2000, Nakamura2002,Mackey2003,Machida2005,Nagakura2005,Johnson2006,Yoshida2007a, Yoshida2007b,Wolcott2011,Hosokawa2012}.
That is, Pop III.2 stars may be typically less massive
than preceding Pop III.1 stars although they are more massive than ordinary present-day stars
\footnote{It should be noted that recent numerical simulations suggest that there is another possible pathway of lower-mass primordial star formation in Pop III.1 case, like disk fragmentation~\citep[e.g.,][]{Greif2012}. So the typical mass-scale of Pop III.1 stars remains an open question.}.
With high effective temperatures ($\sim 10^5\ {\rm K}$)~\citep{Schaerer2002}, massive Pop III
stars emit a copious amount of far-UV (FUV) photons in the Lyman-Werner (LW) bands ($11.2 \mbox{-}13.6$ eV).
The FUV radiation field photodissociates ${\rm H}_2$, thereby prohibiting efficient cooling in the primordial gas.
Even a single massive Pop III star emits enough FUV photons to suppress the subsequent star formation
in the same halo~\citep{Omukai1999,Glover2001}.
In addition, the gas in a minihalo is fairly vulnerable to the external irradiation of FUV photons and
thus the number of Pop III.1 stars formed can be severely regulated by the FUV background~\citep[e.g.,][]{Haiman1997}.
On the other hand, Pop III.2 star formation in more massive halos still goes on by atomic cooling
even under the FUV background and is not strongly affected.~\citep[Haiman et al. 1997;][]{O'Shea2008b}.
This indicates the possibility that Pop III.2 stars can be the major population of primordial stars and
significant sources of radiative and chemical feedback to subsequent star formation~\citep{Trenti2009,Souza2011}.
Moreover, the Pop III.2 star-formation epoch may have lasted until redshifts as low as
$z \sim 6$~\citep{Tornatore2007,Souza2011,Johnson2013}.
If so, gamma-ray bursts (GRBs) or SNe of Pop III.2 stars are detectable with future facilities,
which allow us to peer into the nature of primordial stars~\citep{Nakauchi2012, Whalen2013, Tanaka2013}.
Later, however, it was found that HD formation is strongly suppressed in the presence of
even a weak FUV radiation field~\citep[Yoshida et al. 2007b;][]{Wolcott2011}
since the slight photodissociation of ${\rm H}_2$ causes the significant reduction of the HD fraction.
Without HD cooling, the thermal evolution of a Pop III.2 star-forming gas becomes identical
to the Pop III.1 case, resulting in Pop III.2 stars having similar masses to Pop III.1 stars.
However, this is not yet the conclusive result since the previous studies have not properly taken account of
the following effects, which favors HD formation/cooling.
In the galaxy formation epoch, shocks occur ubiquitously, for example, associating
with the virialization of halos or SN explosions.
In a shocked gas, the enhanced ionization degree enables efficient H$_2$/HD formation and cooling
~\citep{Uehara2000,Mackey2003,Johnson2006,Greif2008}.
For a strong shock, the post-shock gas is compressed isobarically with its cooling,
i.e., the temperature decreases with increasing the density ~\citep[e.g.,][]{Shapiro1987}.
On the other hand, for the free-falling cloud in a relic HII region,
rapid cooling makes the temperature to plummet almost isochorically,
i.e., the temperature decreases at a constant density~\citep[Yoshida et al. 2007b;][]{Johnson2006, Wolcott2011}.
With the higher column density, the FUV radiation field is more effectively shielded for the post-shock flow than in the relic HII region, resulting in the lower FUV flux in the former case.
Nonetheless, in the previous studies~\citep[e.g.,][]{Wolcott2011}, the post-shock flows are treated
without the proper account of the isobaric nature of evolution.
Associated with star-formation activities, not only the FUV background, but also the
cosmic ray~(CR) background is expected to be generated by way of
the CR acceleration in SN remnants.
CR irradiation promotes ${\rm H}_2$ and HD formation/cooling by
enhancing the ionization degree~\citep{Jasche2007,Stacy2007, Inayoshi2011}.
Thus, CR irradiation is able to counteract with the negative feedback of FUV radiation.
However, only either FUV or CR irradiation has been considered in the previous studies
~\citep{Jasche2007,Stacy2007,Wolcott2011} and their combined effects have not been fully understood.
In this paper, we study the conditions for HD cooling to become important in Pop III.2 star formation
under the presence of both FUV and CR backgrounds.
We calculate the temperature and chemical evolution of (i) a cloud compressed by plane-parallel steady
shocks~\citep{Shapiro1987}, and of (ii) a free-falling cloud in a relic HII region.
We find that the critical value for FUV intensity above which HD cooling is quenched
is about an order of magnitude higher in the former than in the latter case.
Also the critical FUV intensity increases with increasing CR intensity.
We also estimate the background FUV and CR intensities from theoretically predicted values for the SFR
in the high-redshift universe.
By comparing the background FUV intensity with the critical value,
we conclude that the HD mode of Pop III.2 star formation is fairly common at the galaxy formation epoch.
The rest of this paper is organized as follows.
In \S \ref{sec:model}, we describe the numerical model and initial conditions for our calculation.
In \S \ref{sec:result}, we present the results for the free-falling clouds in relic HII regions
and for the clouds compressed either by the virialization shock or by the SN blast wave.
We also present the critical FUV intensity for HD cooling as a function of the CR intensity.
In \S \ref{sec:estimate}, we estimate the background FUV and CR intensities in high-redshift universe,
and discuss whether the HD cooling condition is satisfied at these epochs.
After discussing the implications and uncertainties of our study in \S \ref{sec:discussion},
we summarize our findings in \S \ref{sec:conclusion}.
\section{Model}\label{sec:model}
In this section, we present our model for calculating the thermal and chemical evolution of primordial clouds.
\subsection{Formulation}
Here, we first describe the method of calculation for the free-falling cloud in a relic HII region
and then for the shock-compressed gas in structure formation or an SN explosion.
We use the one-zone model, where we focus on the evolution of the central core for the free-falling cloud and of the coolest gas layer for the shock-compressed gas before the self-gravity becomes effective by using the plane-parallel and steady shock approximation.
\subsubsection{Free-falling Cloud in a relic HII region}\label{subsec:free-fall_form}
When a Pop III star dies and collapses directly to a black hole (BH), the gas in the relic HII region surrounding the star begins to recombine and cool to form a subsequent star~\citep{Kitayama2004, Whalen2004, Yoshida2007a, Yoshida2007b, Abel2007}. Here, we assume that the formed BH has no influence on the evolution of the gas in the relic HII region for simplicity.
At first, the baryon mass in the relic HII region is too small for the gas to contract by self-gravity, but the gas loses pressure-support by efficient radiative cooling and contracts being attracted by the gravitational potential of the DM halo.
After the temperature decreases so that the Jeans mass becomes lower than the HII region mass, the gas begins to contract in the almost free-fall way by self-gravity.
Owing to its high pressure, the gas in the HII region is accelerated monotonically with radius up to a supersonic velocity and expands into the interstellar medium~(Whalen et al. 2004; Kitayama et al. 2004).
However, our model does not take into account the initial velocity profile within the HII region for simplicity.
The following prescription is the same as the previous studies~(Johnson \& Bromm 2006; Yoshida, Omukai \& Hernquist 2007; Wolcott-Green \& Haiman 2011).
An isolated cloud collapsing at the rate close to free-fall by its self-gravity
develops a core-envelope structure~\citep{Larson1969, Penston1969}, consisting of a nearly Jeans-scale core with the
constant density and an envelope where the density decreases radially as $\rho \propto r^{-2}$.
In our one-zone model, we follow the evolution of the physical quantities at the core center~\citep{Omukai2001}.
Since collapse proceeds roughly at the free-fall rate, the density evolves as
\begin{equation}
\frac{d \rho}{dt} = \frac{\rho}{t_{\rm ff}},
\label{eq:drho_dt}
\end{equation}
where $t_{\rm ff} = \sqrt{{3 \pi}/{32 G \rho}}$ is the free-fall time and $G$ the gravitational constant.
Temperature evolution follows the energy equation:
\begin{equation}
\frac{d e}{dt} = - P \frac{d}{dt}\left(\frac{1}{\rho}\right) - \frac{\Lambda_{\rm net}}{\rho},
\label{eq:energy}
\end{equation}
where $e$ is the internal energy per unit mass, $P$ the pressure,
$\Lambda_{\rm net}$ the net cooling rate per unit volume.
To supplement the equations above,
the ideal-gas equation of state $P = \rho k_{\rm B} T / \mu m_{\rm H}$ is used,
where $k_{\rm B}$ is the Boltzmann constant, $m_{\rm H}$ the proton mass and $\mu$ the mean molecular weight.
Using the ratio of specific heat $\gamma$, specific energy is related to pressure as
$e = P/\rho (\gamma -1)$.
Throughout this paper, we take $\mu = 1.2$ and $\gamma = 5/3$ neglecting molecular contribution.
Considering that the size of the contracting core is half the Jeans length $\lambda_{\rm J}=\sqrt{{\pi k_{\rm B} T}/{G \rho \mu m_{\rm H}}}$, we calculate the column density of the core as $N_{\rm J} = n_{\rm H} \lambda_{\rm J} / 2$~\citep{Inayoshi2011}.
In this case, the initial size of this region is estimated as $\sim 1$ kpc, which is an overestimation compared to the typical size of an HII region $R_{\rm HII} \sim 100$ pc~(Whalen et al. 2004; Kitayama et al. 2004).
However, at the beginning of collapse, owing to the low gas density and H$_2$ fraction, the H$_2$ column density will be too low for effectively shielding the gas from the FUV filed, and the above overestimation will not matter for estimating the shielding factor.
In our model of free-fall collapse, the collapse equation and the energy equation (Eqs. \ref{eq:drho_dt}, \ref{eq:energy}) are decoupled and feedback from thermal pressure is not included for simplicity.
In the runaway-collapse phase, an isothermal cloud collapses approximately at a free-fall rate~\citep{Larson1969, Penston1969}.
Although the pressure gradient is not negligible, the deviation of the collapse time from free fall is only by a factor of 1.58~\citep{Larson1969}.
For a non-isothermal case, the deviation is represented as a monotonically increasing function of the effective adiabatic index $\gamma$ of the cloud in equations 7-9 of \cite{Omukai2005}.
For example, the collapse time becomes $\sim 3$ times longer than the free-fall time for $\gamma = 1.2$.
If we use this generalized collapse time for the collapse equation, the collapse equation 4 is coupled with the energy equation 5.
However, we find that the effective adiabatic index becomes smaller than 1.2 for most of the evolutionary trajectory, and the deviation from free-fall may be at most by a factor of a few.
Thus, feedback from thermal pressure may change our results little.
Here, we consider that an overdense region collapsing within an HII region is large enough. Then, since the overdense region is strongly bounded by the gravitational field of DM and gas, it contracts by the gravitational force which exceeds the pressure force from its less-dense surroundings.
Therefore, the compression of the overdense region by external pressure may be neglected, and it will collapse roughly at a free-fall rate.
Although our formulation does not include DM gravity, it becomes significant at the beginning of gas evolution. Then, we recalculate the evolution of the gas in relic HII regions taking account of DM gravity by modifying $\rho$ in $t_{\rm ff}$ of Eq. \eqref{eq:drho_dt} to $\rho + \rho_{\rm DM}$, but we find little change in our results.
Here, $\rho_{\rm DM} = 1.3 \times 10^{-24}\ {\rm g}\ {\rm cm}^{-3}$ is the mean DM mass density of a virialized halo at $z=15$.
\subsubsection{Shock-Compressed Gas}\label{subsec:shock_form}
After the passage of a blast wave, which is generated in structure formation or an SN explosion, the post-shock gas has a very high temperature due to shock heating and it cools and contracts with the radiative cooling time. At first, self-gravity is not effective in the shocked gas layer. We focus on the evolution of the most cooled gas layer, assuming that the post-shock gas layer is plane-parallel and the flow is steady~\citep{Shapiro1987, Yamada1998, Inayoshi2012}.
The initial condition of a shock-compressed gas is characterized by two parameters: the pre-shock density $\rho_0$ and
the velocity of the shock front $v_0$.
In the strong shock limit, the physical quantities just behind the shock front are given by the Rankine-Hugoniot relation:
\begin{equation}
\rho_1 = 4 \rho_0,
\label{eq:rho1}
\end{equation}
\begin{equation}
v_1 = \frac{1}{4} v_0,
\label{eq:v1}
\end{equation}
\begin{equation}
T_1 = \frac{3}{16}\frac{\mu m_{\rm H}}{k_{\rm B}}v_0^2
\end{equation}
\begin{equation}
\sim 1.1 \times 10^4 \left(\frac{\mu}{1.2}\right) \left(\frac{v_0}{20\ {\rm km}/{\rm s}}\right)^2 {\rm K}, \notag
\label{eq:T1}
\end{equation}
where the subscripts 0 and 1 refer to the pre- and post-shock quantities, respectively.
Behind the shock front, the flow evolution is described by
\begin{equation}
\rho_1 v_1 = \rho v,
\label{eq:rho_v}
\end{equation}
\begin{equation}
\rho_1 v_1^2 + P_1 = \rho v^2 + P,
\label{eq:rho_v^2}
\end{equation}
along with the energy equation (\ref{eq:energy}).
The column density of the post-shock gas can be estimated from
\begin{equation}
N_{\rm H} (t) = n_{{\rm H}, 0} v_0 t,
\label{eq:clden_b}
\end{equation}
where $n_{{\rm H}, 0}$ is the number density in the pre-shock gas and $t$ is the time since the shock occurs.
Rapid cooling in the post-shock layer leads to the formation of a dense gas shell, which eventually becomes
gravitationally unstable and fragments.
For the SN shock, we assume that fragmentation occurs when the column density of the post-shock layer $N_{\rm H}$
exceeds that in the Jeans length $N_{\rm J}$ \citep{Safranek-Shrader2010, Inayoshi2012}:
\begin{equation}
N_{\rm H} (t) = N_{\rm J}.
\label{eq:cond_sn}
\end{equation}
For the structure-formation shock, the mass of the post-shock layer is doubled compared to the SN case, since in one dimension, two flows collide to form the shocked region and two shock fronts propagate in opposite directions to each other. In this case, fragmentation occurs under the condition:
\begin{equation}
2N_{\rm H} (t) = N_{\rm J}.
\label{eq:cond_sf}
\end{equation}
In addition to them, for fragmentation, the growth time of a perturbation, which is approximately given by
the local free-fall time $t_{\rm ff}$, must be shorter than the contraction time of the layer,
i.e., the cooling time $t_{\rm cool}$ (Yamada \& Nishi 1998).
We have confirmed this condition is always satisfied at the fragmentation epoch determined by
the column-density conditions given above.
After fragmentation, each fragment or clump begins to contract by its self-gravity.
We also follow its evolution by the same one-zone model as in Sec. \ref{subsec:free-fall_form}.
\subsection{Thermal and Chemical Processes}\label{subsec:thermal_chemical}
The net cooling rate $\Lambda_{\rm net}$ in Eq. \eqref{eq:energy} consists of
the following terms:
\begin{equation}
\Lambda_{\rm net} = \Lambda_{\rm H} + \Lambda_{{\rm H}_2} + \Lambda_{\rm HD} - \Gamma_{\rm CR},
\label{eq:lambda_net}
\end{equation}
where $\Lambda_{\rm H}$, $\Lambda_{{\rm H}_2}$ and $\Lambda_{\rm HD}$ are
the radiative cooling rates by H Lyman-$\alpha$ line emission, ${\rm H}_2$-line emission and HD-line emission,
respectively, and $\Gamma_{\rm CR}$ is the CR heating rate.
We adopt the cooling rates $\Lambda_{\rm H}$ from \cite{Glover2007}, $\Lambda_{{\rm H}_2}$ from \cite{Galli1998}, and $\Lambda_{\rm HD}$ from \cite{Galli2002}.
Since each CR ionization associates with heating of $\sim 6\ {\rm eV}$ \citep{Spitzer1969},
\begin{equation}
\Gamma_{\rm CR} \sim 9.6 \times 10^{-12}\ n({\rm H})\ \zeta_{\rm CR}\ {\rm erg}\ {\rm cm}^{-3}\ {\rm s}^{-1},
\label{eq:cr_heat}
\end{equation}
where $\zeta_{\rm CR}$ is the hydrogen ionization rate by CRs,
and $n({\rm H})$ is the number density of hydrogen atoms.
At densities we consider~($\lesssim 10^{7}\ {\rm cm}^{-3}$), the contribution from
chemical cooling and heating is small and is neglected in this paper~\citep{Omukai2001}.
Our chemical network includes following 11 species: H, ${\rm H}_2$, $e$, ${\rm H}^+$, ${\rm H}_2^+$, ${\rm H}^-$, D, HD, ${\rm D}^+$, ${\rm HD}^+$ and ${\rm D}^-$.
The deuterium fraction is set to $4.0 \times 10^{-5}$, and the He chemistry is neglected since it is thermally inert
in the temperature range we consider.
Chemical reactions include 18 hydrogen collisional reactions in Table 1 of \cite{Yoshida2006},
18 deuterium collisional reactions in Table 1 of \cite{Nakamura2002},
photo-dissociation of H$_2$, HD, and H$^{-}$, and CR ionization.
For collisional reactions, we use the updated rate coefficients by \cite{Glover2008}.
The photo-dissociation rate coefficients are taken from \cite{Omukai2001} and
\cite{Wolcott2011}, and the ${\rm H}_2$ and HD shielding factors are from~\citet{Wolcott2011}.
For the shielding factors, we take account of the Doppler shift due to the bulk motion of the cloud
as in \cite{Omukai2001} for the free-falling case and in \cite{Omukai2007} for the shock-compressed case.
The stellar radiation field is assumed to have the diluted black-body spectrum with the temperature of $10^4$\ K, resembling that from a Pop II star cluster, and its intensity is parameterized by $J_{21} \equiv J_{\rm \nu_{Ly}}/10^{-21}\ {\rm erg}\ {\rm cm}^{-2}\ {\rm s}^{-1}\ {\rm Hz}^{-1}$, where $J_{\rm \nu_{Ly}}$ is the intensity at the Lyman limit.
The CR ionization rate of neutral hydrogen $\zeta_{\rm CR}$ is treated as a free parameter,
since the CR intensity at the high-redshift universe is highly uncertain.
However, the CR intensity may be related with the star-forming activity, and the SFR at $z \sim 10$ is theoretically
estimated to be no more than the present-day value~(e.g., Tornatore et al. 2007; Trenti \& Stiavelli 2009; Johnson et al. 2013).
We thus consider the CR intensity in the range of $\zeta_{\rm CR}\lesssim 10^{-17}\ {\rm s}^{-1}$, which are smaller than the Galactic value $10^{-17}\ {\rm s}^{-1} \lesssim \zeta_{\rm CR} \lesssim 10^{-15}\ {\rm s}^{-1}$~\citep{Hayakawa1961,Spitzer1968,Webber1998,McCall2003,Indriolo2007}.
\subsection{Initial Conditions}
\subsubsection{Relic HII Region}\label{subsec:free-fall_IC}
In the relic HII region of a defunct Pop III star, a primordial cloud begins to collapse from a highly ionized state.
The temperature and density in the relic HII region are almost uniform with $1\mbox{-}3 \times 10^4\ {\rm K}$ and
$0.1\mbox{-}1\ {\rm cm}^{-3}$~(Kitayama et al. 2004; Whalen et al. 2004; Abel, Wise \& Bryan 2007; Yoshida et al. 2007a).
We thus adopt $T_0 = 3 \times 10^4\ {\rm K}$ and $n_{\rm H, 0} = 0.3\ {\rm cm}^{-3}$ as the fiducial initial values,
and also study the cases with higher and lower densities $n_{\rm H, 0} = 0.03, 3 \ {\rm cm}^{-3}$ to see
the dependence on the initial density.
We set the initial chemical abundances as $y_0({\rm H}^{+}) = 1.0$, $y_0({\rm D}^{+}) = 4.0 \times 10^{-5}$, $y_0(e) = y_0({\rm H}^{+}) + y_0({\rm D}^{+})$, and $y(i) = 0$ for other species, where
$y(i)$ is the fractional abundance of the $i$-th species to the hydrogen nuclei.
\subsubsection{Shock-Compressed Gas}\label{subsec:shock_IC}
We consider two kinds of shocks which occur in the virialization of halos and SN explosions.
Here, we need to specify the shock velocity $v_0$, pre-shock density $n_{{\rm H}, 0}$,
and initial chemical abundance $y_0(i)$ for each case.
For the virialization shock,
the mean gas number density within the virialized halo is estimated as \citep{Clarke2003}
\begin{equation}
n_{{\rm H}, 0} \sim 0.13 \left(\frac{1+z}{15}\right)^{3} {\rm cm}^{-3}.
\label{eq:n_H_halo}
\end{equation}
However, according to the recent studies on first galaxy formation,
the shock front does not sustain at the virial radius and its location shrinks inward owing to
efficient Ly-$\alpha$ emission~\citep{Wise2007,Wise2008}.
Protogalaxies accrete gas via cold flows in dense filaments, with the infall velocity of $\sim 20\ {\rm km}\ {\rm s}^{-1}$.
Such cold flows penetrate deep inside the halo and collide each other to form virialization shocks
with pre-shock densities $1\mbox{-}10^3\ {\rm cm}^{-3}$, much higher than in Eq. \eqref{eq:n_H_halo}.
Considering this effect, we adopt $(v_0, n_{\rm H, 0}) = (20, 1.0)$ as the fiducial values for the pre-shock gas.
The shocked gas behind a SN blast wave cools and is compressed significantly after the SN remnant enters
the pressure-driven expansion phase \citep{Machida2005, Nagakura2009, Chiaki2013}.
The shock velocity early in this phase is $\sim 100\ {\rm km}\ {\rm s}^{-1}$~(Machida et al. 2005; Nagakura et al. 2009; Chiaki et al. 2013).
The pre-shock density is comparable to the density in relic HII regions $0.1\mbox{-}1\ {\rm cm}^{-3}$, since core-collapse SNe explode within HII regions which their progenitors produce.
Hence, we adopt $(v_0, n_{\rm H, 0}) = (100, 0.1)$ as the fiducial values for the SN shock.
\cite{Shapiro1987} and \cite{Kang1992} studied the chemical composition in the primordial gas
just behind a shock front. We adopt their results as the initial composition in the shock-compressed gas,
which are summarized in Table \ref{tab:shock_comp} for the virialization and SN shocks.
The fractions of species not presented in Table \ref{tab:shock_comp} are set to zero,
except for the electron fraction $y_0(e) = y_0({\rm H}^{+}) + y_0({\rm D}^{+})$.
\begin{table
\caption{Initial chemical composition in the shock-compressed gas}
\begin{center}
\begin{tabular}{c|cccc}
\hline
$v_0$\ (km\ ${\rm s}^{-1}$) & $y_0({\rm H})$ & $y_0({\rm H}^{+})$ & $y_0({\rm D})$ & $y_0({\rm D}^{+})$ \\ \hline
20 & 0.99 & 0.01 & $3.96 \times 10^{-5}$ & $4.0 \times 10^{-7}$ \\
100 & 0.3 & 0.7 & $1.2 \times 10^{-5}$ & $2.8 \times 10^{-5}$ \\
\hline
\end{tabular}
\end{center}
\label{tab:shock_comp}
\end{table}
\section{Results}\label{sec:result}
\subsection{Relic HII region}\label{subsec:free-fall_result}
Here, we present the results for the free-falling cloud in a relic HII region.
We also find the critical FUV intensity above which HD cooling is suppressed
for the different values of the CR intensity.
\subsubsection{Thermal Evolution}\label{subsec:free-fall_evolve}
First, we see the results without CR irradiation to extract the FUV effect alone.
Fig. \ref{fig:h2_nT_CR0} shows the temperature evolution, and Fig. \ref{fig:h2_nT_CR0_chem} shows the fractions of (a) ${\rm H}_2$, (b) electron, and (c) HD, for five different FUV intensities: $J_{21}=0, 0.01, 0.1, 1$, and $10$, both as a function of
the number density $n_{\rm H}$.
The temperature first drops isochorically to
$\sim 8000\ {\rm K}$ via efficient Ly-$\alpha$ cooling in all the cases in Fig. \ref{fig:h2_nT_CR0}, and the subsequent behavior depends on the FUV intensities.
The same results are obtained in Johnson \& Bromm (2006), Yoshida, Omukai \& Hernquist (2007), and Wolcott-Green \& Haiman (2011).
In the absence of FUV irradiation ($J_{21} = 0$), abundant ${\rm H}_2$
with the fraction $\sim 10^{-3}$~forms via the H$^-$ channel (Eqs. \ref{eq:H-1} and \ref{eq:H-2})
owing to the high initial ionization degree~(solid line in Fig. \ref{fig:h2_nT_CR0_chem}a).
The cooling by this enhanced H$_2$ makes the temperature to fall below $\simeq 200\ {\rm K}$,
the minimum value attainable in the primordial pristine gas, where the ionization degree is
only $\sim 10^{-4}$~\citep{Omukai2001}.
Once the temperature reaches $\la 150\ {\rm K}$,
the exothermic reaction (Eq. \ref{eq:HD}) converts
most of the deuterium into HD (solid line in Fig. \ref{fig:h2_nT_CR0_chem}c).
HD cooling further drops the temperature to
the minimum value $\sim 30\ {\rm K}$ at density $\sim 10^{5}\ {\rm cm}^{-3}$,
which is the critical density for HD to reach local thermodynamic equilibrium.
Toward higher density, the temperature increases gradually by compressional heating.
At the temperature minimum, the parent cloud is considered to fragment into
clumps with the Jeans mass of $\la 100\ {\rm M}_{\odot}$~\citep[Yoshida et al. 2007b;][]{McGreer2008}.
We call hereafter this mode of star formation as {\it HD-Pop III star formation}, in contrast to
{\it H$_2$-Pop III star formation}, where H$_2$ is the sole coolant during proto-stellar collapse.
With FUV irradiation, the temperature becomes higher owing to the photodissociation of the coolants,
H$_2$ and HD~(Fig. \ref{fig:h2_nT_CR0_chem}a, c).
With $J_{21}$ as low as 0.01, the evolutionary track is similar to that of $J_{21} = 0$,
although with slightly higher temperature.
With intensity $J_{21} = 0.1$, the ${\rm H}_2$ fraction is reduced to $\sim 10^{-4}$,
and the temperature does not reach the regime of efficient HD formation/cooling
(short-dashed line in Fig. \ref{fig:h2_nT_CR0_chem}a, c).
With the even higher FUV intensity $J_{21} \geq 1.0$,
photodissociation strongly reduces the H$_2$ fraction at low densities
(dotted and dot-dashed lines in Fig. \ref{fig:h2_nT_CR0_chem}a).
The clouds contract isothermally at $\sim 8000\ {\rm K}$
by Ly-$\alpha$ cooling until $n_{\rm H} \sim 10\ {\rm cm^{-3}}$ for $J_{21}= 1.0$
($\sim 100\ {\rm cm^{-3}}$ for $J_{21} =10$, respectively), where the H$_2$ self-shielding
against FUV radiation becomes effective.
Thereafter, the temperature drops first almost vertically to $\sim 2000$ K and then
gradually to $\sim 300\ {\rm K}$ by ${\rm H}_2$ cooling.
In the cases with $J_{21} \geq 0.1$, without HD cooling,
the gas cools only to the minimum temperature $\sim 200\ {\rm K}$ by H$_2$ cooling
at $\sim 10^{3}\mbox{-}10^{4}\ {\rm cm^{-3}}$ as in the case of Pop III.1 star formation.
Namely, star formation proceeds in the H$_2$-Pop III mode.
In this mode, the clouds fragment into rather massive clumps of
$\sim 10^{3}\ {\rm M}_{\odot}$.
Note that, in the high density regime ($\ga 10^{5}\ {\rm cm^{-3}}$),
all the evolutionary tracks converge to either of the two tracks,
one by HD cooling for $J_{21} \leq 0.01$ or the other
by ${\rm H}_2$ cooling for $J_{21} \geq 0.1$.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{h2CR0_J_n03_e09.eps}
\caption{FUV effect on thermal evolution of clouds in relic HII regions.
No CR irradiation is considered.
The initial density is $n_{\rm H, 0} = 0.3\ {\rm cm}^{-3}$.
Different lines correspond to the different FUV intensities:
$J_{21}=0.0, 0.01, 0.1, 1.0$, and $10$ from bottom to top.
The thin dashed lines indicate the constant Jeans masses.}
\label{fig:h2_nT_CR0}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{h2CR0_J_n03_e09_chem.eps}
\caption{Chemical fractions of (a) ${\rm H}_2$, (b) $e$, and (c) HD in the same clouds
as in Fig. \ref{fig:h2_nT_CR0}.
In each panel, different lines correspond to the different FUV intensities: $J_{21}=0.0$ (solid), $0.01$ (long-dashed), $0.1$ (short-dashed), $1.0$ (dotted), and $10$ (dot-dashed).}
\label{fig:h2_nT_CR0_chem}
\end{center}
\end{figure}
Next, we see the results with CR irradiation on cloud evolution,
as shown in Figs. \ref{fig:h2_nT_J01} and \ref{fig:h2_nT_J01_chem},
for the cases with FUV intensity $J_{21} = 0.1$.
The different lines in the Figures show different values of the CR intensity:
$\log \zeta_{\rm CR} ({\rm s}^{-1})= - \infty$ (i.e., $\zeta_{\rm CR}=0$),
$-20, -19, -18$ and $-17$.
Recall that, with $J_{21}=0.1$, HD cooling is suppressed without CR irradiation.
As seen in Fig. \ref{fig:h2_nT_J01}, with higher $\zeta_{\rm CR}$, the temperature becomes lower.
This is because the elevated ionization degree by CR ionization
causes more efficient H$_2$ formation and cooling (see Fig. \ref{fig:h2_nT_J01_chem}),
which outweighs the effect of CR ionization heating.
For $\log \zeta_{\rm CR} \geq -18$, the temperature falls below
150 K, where the HD fraction increases significantly, owing to
the enhanced ${\rm H}_2$ cooling~(solid and long-dashed lines in Fig. \ref{fig:h2_nT_J01_chem}c).
By HD cooling, the temperature further drops to $\sim 30\ {\rm K}$, and HD-Pop III star formation
is realized in these cases.
In this way, CR ionization can compensate the negative feedback of FUV photodissociation.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{h2CR_J01_n03_e09.eps}
\caption{CR effect on thermal evolution of clouds in relic HII regions.
The FUV intensity is fixed to $J_{21} = 0.1$, and the initial density to $n_{\rm H, 0} = 0.3\ {\rm cm}^{-3}$.
Different lines correspond to the different CR intensities: $\log \zeta_{\rm CR} ({\rm s}^{-1})= - \infty, -20, -19, -18$, and $-17$. The thin dashed lines indicate the constant Jeans masses.}
\label{fig:h2_nT_J01}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{h2CR_J01_n03_e09_chem.eps}
\caption{(a) H$_2$, (b) $e$, and (c) HD fractions for the same clouds as in Fig. \ref{fig:h2_nT_J01}.
In each panel, the different lines correspond to the different values of
CR intensity:
$\log \zeta_{\rm CR}({\rm s}^{-1}) = - \infty$ (dot-dashed), $-20$ (dotted), $-19$ (short-dashed), $-18$ (long-dashed), and $-17$ (solid).}
\label{fig:h2_nT_J01_chem}
\end{center}
\end{figure}
Since in our model for the free-falling cloud, the collapse equation~(eq. \ref{eq:drho_dt}) is decoupled from the energy equation~(eq. \ref{eq:energy}), the collapse time becomes the same for the same initial density and is obtained as $t_{\rm collapse} \sim 200$ Myr for $n_{{\rm H},0} = 0.3\ {\rm cm}^{-3}$.
The Hubble time is estimated as $t_{\rm H}(z) = 500 \left((1+z)/10\right)^{-3/2}\ {\rm Myr}$ and this is roughly equal to the galaxy merger time. We find that the collapse timescale estimated above is smaller than the Hubble time for $z < 15$.
Moreover, in the early universe we consider, the halo does not have an angular momentum large enough to form a rotation-supported disk which is comparable to the protogalaxy scale~(Abel et al. 2002).
Thus, the gas cloud in a relic HII region may remain isolated without being affected by the protogalaxy rotation and merger, and star formation there proceeds within the local Hubble time.
\subsubsection{Critical FUV intensity for HD Cooling}\label{subsec:h2_zeta_Jcrit}
We have learned that with the FUV intensity exceeding a critical value
$J_{21, {\rm crit}}$,
HD cooling is totally suppressed even in initially ionized clouds.
In addition, the value of the critical FUV intensity is elevated under CR irradiation.
In Fig. \ref{fig:h2_zeta_Jcrit}, the critical FUV intensity $J_{21, {\rm crit}}$ is plotted
as a function of $\zeta_{\rm CR}$ for initial densities $n_{\rm H, 0} = 0.03$ (dashed), 0.3 (solid) and 3 (dot-dashed)$\ {\rm cm}^{-3}$.
In finding the value of $J_{21, {\rm crit}}$, we judge
HD cooling is effective if the following two conditions are satisfied:
(i) the minimum temperature at $> 10^{5}\ {\rm cm^{-3}}$ is less than 100\ K,
and (ii) the HD-cooling rate exceeds a half of the total (i.e. the sum of compressional and CR) heating rate at some moment.
These two conditions are imposed to discriminate whether the high-density temperature tracks are
either those by H$_2$ cooling or by HD cooling.
In the fiducial case of $n_{\rm H, 0}=0.3\ {\rm cm^{-3}}$ (solid line),
the critical intensity is $J_{21, {\rm crit}} = 0.03$ for the negligible CR intensity
($\zeta_{\rm CR} \sim 10^{-20}\ {\rm s}^{-1}$), in accord with previous
studies~\citep[Yoshida et al. 2007b;][]{Wolcott2011}.
The critical intensity $J_{21, {\rm crit}}$ increases with the CR strength $\zeta_{\rm CR}$,
since strong enough CR irradiation enables efficient HD cooling even under an intense FUV field.
In the weak CR regime $\zeta_{\rm CR} \la 3 \times 10^{-19}\ {\rm s}^{-1}$, the electrons present initially, rather than those produced by CR ionization, work as catalysts for the H$_2$ formation reaction.
Abundant H$_2$ is formed if the self-shielding against the FUV field becomes effective before the ionization degree
is reduced significantly by recombination.
Since the self-shielding is more effective for larger column densities and thus
for larger $n_{\rm H, 0}$, more H$_2$ is formed under the same FUV intensity.
This results in higher $J_{21, {\rm crit}}$ for higher $n_{\rm H, 0}$.
On the other hand, in the strong CR regime $\zeta_{\rm CR}$ $(\gtrsim 3 \times 10^{-19}\ {\rm s}^{-1})$, CR ionization provides the necessary electrons for abundant H$_2$ formation.
Moreover, with a FUV field as high as $J_{21, {\rm crit}}$, the shielding against the FUV field becomes effective
and ${\rm H}_2$ formation occurs only after the density is significantly increased from the initial value.
Therefore, the evolutionary track and critical FUV intensity $J_{21, {\rm crit}}$ do not depend on the initial density.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{h2_zeta_Jcrit.eps}
\caption{The critical FUV intensity $J_{21, {\rm crit}}$ for HD cooling as a function of the CR intensity $\zeta_{\rm CR}$.
Each curve corresponds to the different initial densities: $n_{\rm H, 0} = 0.03$ (dashed), 0.3 (solid) and 3 (dot-dashed)$\ {\rm cm}^{-3}$.
Below $J_{21, {\rm crit}}$, HD dominates the cooling in the course of cloud collapse.
The shaded region shows the background FUV and CR intensities estimated from Eqs. \eqref{eq:J_theory} and \eqref{eq:zeta_theory} with $6 \la z \la 15$ and $10^{-4} \lesssim \Psi_{\ast, \rm II}(z) / {\rm M}_{\odot} {\rm yr}^{-1} {\rm Mpc}^{-3} \lesssim 3 \times 10^{-2}$.
The blue thin-solid line shows the evolution of the background intensities evaluated using the Pop II SFR $\Psi_{\ast, \rm II}(z)$ of Johnson et al. (2013).
The blue points refer to values at $z = 15, 10$ and $6$ from left to right.}
\label{fig:h2_zeta_Jcrit}
\end{center}
\end{figure}
\subsection{Shock-compressed Gas}\label{subsec:shock_result}
Here, we present the results for a shock-compressed gas in the two cases of the shock velocity:
(i) $v_0 = 20\ {\rm km}\ {\rm s}^{-1}$ (virialization shock) and
(ii) $100\ {\rm km}\ {\rm s}^{-1}$ (SN shock), respectively.
\subsubsection{Virialization Shock}\label{subsec:virial_result}
First, to see the FUV effect on the gas compressed by the virialization shock
($v_{0}=20\ {\rm km}\ {\rm s}^{-1}$), the results are shown
for the cases without CR irradiation in Figs. \ref{fig:shock20_nT_CR0} and \ref{fig:shock20_nT_CR0_chem}.
Just behind the shock front, the gas is heated to $\gtrsim 10^4\ {\rm K}$ and then
cools immediately to $\sim 8000\ {\rm K}$ by Ly-$\alpha$ cooling.
With the negligible or weak FUV field ($J_{21} \la 0.01$), H$_2$ forms abundantly owing to the high ionization degree
($\sim 0.01$) in the post-shock gas (solid and long-dashed lines in Fig. \ref{fig:shock20_nT_CR0_chem}a, b).
As in the relic HII region, the gas enters the HD formation regime ($\la 150$ K)
by H$_2$ cooling (solid and long-dashed lines in Fig. \ref{fig:shock20_nT_CR0_chem}c),
and the temperature further decreases to $\sim 30\ \rm K$ by HD cooling.
As long as the cooling time is shorter than the free-fall time, the post-shock gas is compressed
almost isobarically~\citep[e.g.,][]{Shapiro1987, Yamada1998}.
However, below $\sim 30\ {\rm K}$, HD cooling is not efficient anymore,
and the cooling time exceeds the free-fall time.
When the sufficient mass for gravitational instability accumulates
in the dense layer, it fragments into clumps of approximately the Jeans mass
$\sim 100\ {\rm M}_{\odot}$, which then continue collapsing by HD cooling
in the free-fall manner.
This late-time temperature evolution is along the HD-cooling track, like the free-falling clouds in relic HII regions~($J_{21}=0.0$ in Fig. \ref{fig:h2_nT_CR0}).
For $J_{21} = 0.1$, the isobaric contraction proceeds in the same way until $\sim 30\ \rm K$
as with the weaker FUV case.
However, during the dense gas layer stays at $\sim 30\ \rm K$ until the Jeans instability sets in, the HD fraction is significantly reduced by electron recombination and H$_2$/HD photodissociation~(short-dashed lines in Fig. \ref{fig:shock20_nT_CR0_chem}c).
As a result, HD cooling does not become important in the subsequent free-fall phase and
temperature evolution eventually converges to the H$_2$-cooling track.
For $J_{21} \geq 1.0$, strong FUV radiation completely quenches HD cooling
and massive clumps of $\sim 10^4\ {\rm M}_{\odot}$ will be formed.
Especially, in the case of $J_{21} = 10$, fragmentation takes place early at
$\sim 8000\ \rm K$. The subsequent free-fall evolution is similar to that of the free-falling clouds
in relic HII regions with $J_{21} \geq 1.0$ (Fig. \ref{fig:h2_nT_CR0}).
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR0_J_n1_v20_e02.eps}
\caption{FUV effect on thermal evolution of shock-compressed clouds by virialization.
No CR irradiation is considered. The initial density and shock velocity are taken as fiducial values: $n_{\rm H, 0} = 1.0\ {\rm cm}^{-3}$ and $v_0 = 20\ {\rm km}\ {\rm s}^{-1}$. Different lines correspond to the different FUV intensities: $J_{21}=0.0, 0.01, 0.1, 1.0$, and $10$.}
\label{fig:shock20_nT_CR0}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR0_J_n1_v20_e02_chem.eps}
\caption{Fractions of (a) H$_2$, (b) $e$, and (c) HD in the same clouds as in Fig. \ref{fig:shock20_nT_CR0}.
Different lines correspond to the different FUV intensities: $J_{21}=0.0$ (solid), $0.01$ (long-dashed), $0.1$ (short-dashed), $1.0$ (dotted), and $10$ (dot-dashed).}
\label{fig:shock20_nT_CR0_chem}
\end{center}
\end{figure}
Next, we see the CR effect for the case with FUV intensity $J_{21}=1.0$.
The results are shown in Figs. \ref{fig:shock20_nT_J1} and \ref{fig:shock20_nT_J1_chem}.
Recall that, with $J_{21} = 1.0$, HD cooling is suppressed in the absence of CR irradiation.
As in the case of the relic HII region, more ${\rm H}_2$ is formed by CR ionization~(Fig. \ref{fig:shock20_nT_J1_chem}a, b) and the temperature becomes lower by the enhanced H$_2$ cooling.
With the CR strength exceeding $\zeta_{\rm CR} = 10^{-18}\ {\rm s}^{-1}$, most deuterium is converted
into HD~(long-dashed and solid lines in Fig. \ref{fig:shock20_nT_J1_chem}c), and temperature falls
below $\la 100$ K. Thermal evolution finally converges to the HD-cooling track in the free-falling phase.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR_J1_n1_v20_e02.eps}
\caption{CR effect on the thermal evolution of shock-compressed clouds by virialization.
The FUV intensity is fixed to $J_{21} = 1.0$.
The initial density and velocity are the fiducial values: $n_{\rm H, 0} = 1.0\ {\rm cm}^{-3}$ and $v_0 = 20\ {\rm km}\ {\rm s}^{-1}$.
The different lines correspond to the different CR intensities: $\log \zeta_{\rm CR}({\rm s}^{-1}) = - \infty, -19, -18$, and $-17$.}
\label{fig:shock20_nT_J1}
\end{center}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR_J1_n1_v20_e02_chem.eps}
\caption{Fractions of (a) H$_2$, (b) $e$, and (c) HD in the same clouds as in Fig. \ref{fig:shock20_nT_J1}.
The different lines correspond to the different CR intensities: $\log \zeta_{\rm CR}({\rm s}^{-1}) = - \infty$ (dot-dashed), $-19$ (short-dashed), $-18$ (long-dashed), and $-17$ (solid).}
\label{fig:shock20_nT_J1_chem}
\end{center}
\end{figure}
In our model for the shock-compressed gas, the contraction equations~(eqs. \ref{eq:rho_v}, \ref{eq:rho_v^2}) are coupled with the energy equation~(eq. \ref{eq:energy}) before the self gravity becomes effective, so that the timescale of the isobaric contraction phase differs among the lines.
However, the evolution time after self-gravity becomes effective is almost the same and the time to reach $n=10^5\ {\rm cm}^{-3}$ is $\sim 50$ Myr for the structure-formation shock, which is by an order of magnitude smaller than the local Hubble time $t_{\rm H}(z) = 500 \left((1+z)/10\right)^{-3/2}\ {\rm Myr}$ for $z < 15$.
Thus, the shock-compressed gas in structure formation may be unaffected by the protogalaxy rotation and merger, and star formation there proceeds within the local Hubble time.
\subsubsection{Supernova Shock}\label{subsec:SN_result}
In Fig. \ref{fig:shock100_nT_CR0}, we show the FUV effect on the shock-compressed gas in an SN explosion
($v_0=100\ {\rm km}\ {\rm s}^{-1}$) for cases without CR irradiation.
We find that, despite the different initial conditions, the evolutionary tracks are similar
to those in the virialization shock with the same FUV intensity~(see Fig. \ref{fig:shock20_nT_CR0}).
This is because, although the initial density $n_{{\rm H}, 0} = 0.1\ {\rm cm}^{-3}$ is ten times lower
in the SN case, the post-shock temperature is 25 times higher and thus the pressure in the isobaric
evolution phase differs only by a factor of 2.5~(see Fig. \ref{fig:shock20_nT_CR0}).
We also find that the chemical compositions vary in a similar way as in the case of virialization
shock~(see Fig. \ref{fig:shock20_nT_CR0_chem}),
except that the ionization degree reaches $\sim 1$ due to the very high temperature
$\sim 10^5\ {\rm K}$ behind the SN shock front.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR0_J_n01_v100_e07.eps}
\caption{FUV effect on the thermal evolution of shock-compressed clouds by SNe.
No CR irradiation is considered.
The fiducial initial conditions are adopted: $n_{\rm H, 0} = 0.1\ {\rm cm}^{-3}$ and $v_0 = 100\ {\rm km}\ {\rm s}^{-1}$.
The different lines correspond to the different FUV intensities: $J_{21}=0.0, 0.01, 0.1, 1.0$, and $10$.}
\label{fig:shock100_nT_CR0}
\end{center}
\end{figure}
The same trend holds also in the cases with CR irradiation, and the evolutionary tracks are similar
to those in the virialization shock with the same CR intensity~(see Figs. \ref{fig:shock20_nT_J1} and \ref{fig:shock100_nT_J1}).
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{shCR_J1_n01_v100_e07.eps}
\caption{CR effect on the same clouds as in Fig. \ref{fig:shock100_nT_CR0}.
The FUV intensity is fixed to $J_{21} = 1.0$.
The different lines correspond to the different CR intensities: $\log \zeta_{\rm CR}({\rm s}^{-1}) = - \infty, -19, -18$, and $-17$.}
\label{fig:shock100_nT_J1}
\end{center}
\end{figure}
For the SN-shock case, the evolution time after self-gravity becomes effective is almost the same and the time to reach $n=10^5\ {\rm cm}^{-3}$ is calculated as $\sim 250$ Myr, which is much longer than the lifetime of an SN remnant $\sim 1$ Myr.
Thus, the SN shock will dissolve into the ISM before it collects enough materials to trigger fragmentation and star formation in the post-shock region.
This result is consistent with Chiaki et al. (2013), where they find that the SN-shock-compressed gas can collapse within the lifetime of the SN remnant for the ISM density $\gtrsim 1\ {\rm cm}^{-3}$ and $E_{\rm SN} = 10^{51}$ erg.
However, we discuss the evolution of the SN-shock-compressed gas for comparison with that of structure formation.
\subsubsection{Critical FUV intensity for HD Cooling}\label{subsec:sh_zeta_Jcrit}
As in \S \ref{subsec:h2_zeta_Jcrit},
the critical FUV intensity for HD cooling $J_{21, {\rm crit}}$ is calculated
for a shock-compressed gas
and is presented in Fig. \ref{fig:sh_zeta_Jcrit} as a function of the CR intensity $\zeta_{\rm CR}$.
The solid and dashed lines are for the virialization shock in the fiducial case
$(v_0, n_{\rm H, 0}) = (20, 1.0)$ and in the high density case
$(v_0, n_{\rm H, 0}) = (20, 10)$, respectively, and
the dot-dashed line indicates the SN shock case $(v_0, n_{\rm H, 0}) = (100, 0.1)$.
From Fig. \ref{fig:sh_zeta_Jcrit}, we see that for $\log \zeta_{\rm CR} \lesssim -18$, the critical value $J_{21, {\rm crit}}$ becomes higher for a higher initial density $n_{\rm H, 0}$ or higher shock velocity $v_0$.
The reason is the same as in the case of a relic HII region:
in the weak CR regime, abundant H$_2$/HD formation is caused by the
high post-shock ionization degree, and thus the prompt FUV shielding
before significant recombination is needed.
This can be realized more easily with a higher initial density $n_{\rm H, 0}$ or higher shock velocity $v_0$,
since the column density at $\sim 10^4$ K, below which
recombination proceeds significantly, is higher in these cases.
On the other hand, for $\log \zeta_{\rm CR} \gtrsim -18$,
the critical value $J_{21, {\rm crit}}$ is solely determined by the CR strength $\zeta_{\rm CR}$ and
becomes independent of the initial condition $(n_{\rm H, 0}, v_0)$.
Again the reason is the same as in the relic-HII-region case:
the electrons from CR ionization, rather than those present initially, now work
as catalysts for H$_2$ formation and thus the H$_2$ formation rate does not depend on the
initial parameters.
The comparison between Figs. \ref{fig:h2_zeta_Jcrit} and \ref{fig:sh_zeta_Jcrit} tells us
that in the weak CR regime, $\log \zeta_{\rm CR} \la -19$, the critical intensity is $\sim 10$ times higher
for the shock-compressed gas ($\log J_{21,{\rm crit}}=-0.5...0.5$) than in the relic HII region
($\log J_{21,{\rm crit}}=-2...-1$).
In this case, for abundant H$_2$/HD formation, H$_2$ should be promptly shielded against the FUV field
before the electrons present initially recombine significantly.
This condition can be more easily satisfied in the shock-compressed gas than in the relic HII region.
The reason is that the shock-compressed gas evolves isobarically~(see Fig. \ref{fig:shock20_nT_CR0}) rather
than isochorically as in the relic HII region~(see Fig. \ref{fig:h2_nT_CR0}), which leads to the larger column density in the shock-compressed gas.
In addition, shocks tend to occur in denser environments~(e.g., cold-accretion flows) than
relic HII regions.
Consequently, in the shock-compressed gas, H$_2$ is shielded from the FUV field more effectively, and the critical FUV intensity becomes higher than that in the relic HII region.
\begin{figure
\begin{center}
\includegraphics[scale=1.0]{sh_zeta_Jcrit.eps}
\caption{The critical FUV intensity $J_{21, {\rm crit}}$ for HD cooling as a function of the CR intensity $\zeta_{\rm CR}$ in shock-compressed clouds~(see also Fig. \ref{fig:h2_zeta_Jcrit}).
The solid and dashed lines correspond to the virialization shock with the fiducial case, $(v_0, n_{\rm H, 0}) = (20, 1.0)$, and the high density case, $(v_0, n_{\rm H, 0}) = (20, 10)$, respectively.
The dot-dashed line shows the SN-shock case, $(v_0, n_{\rm H, 0}) = (100, 0.1)$.}
\label{fig:sh_zeta_Jcrit}
\end{center}
\end{figure}
\section{Estimate for background intensities}\label{sec:estimate}
We have seen that Pop III.2 star formation proceeds through the HD-cooling track,
i.e., in the HD-Pop III star formation mode, under a weak FUV field or strong enough CR irradiation.
In this section, we estimate the intensities of the FUV and CR backgrounds as a function of redshift.
Then, by comparing them with the critical value $J_{21, {\rm crit}}$ obtained in \S \ref{sec:result},
we discuss whether HD-Pop III star formation is realized at the epoch of galaxy formation~($z \lesssim 15$).
At this epoch, metal enrichment in an average galaxy may have proceeded to
the level for Pop II star formation to occur~(e.g., Wise et al. 2012; Johnson et al. 2013).
We thus consider massive Pop II stars and their SN remnants as the plausible sources of the FUV and CR backgrounds,
respectively.
Then, the intensity of the FUV background is estimated as~\citep{Greif2006}
\begin{equation}
J_{\rm LW, bg}(z) = \frac{hc}{4\pi m_{\rm H}} \frac{\bar{\nu}}{\Delta \nu} \eta_{\rm LW} \rho_*(z) (1+z)^3,
\label{eq:BFUV_1}
\end{equation}
where $h$ is the Planck constant, $c$ the speed of light, $\bar{\nu}$ the average frequency of the LW band ($h \bar{\nu} = 12.4$ eV), $\Delta \nu$ the LW band width ($h \Delta \nu = 2.4$ eV), $\eta_{\rm LW} = 4000$ the number of FUV photons emitted per stellar baryon, and $\rho_*(z)$ the stellar mass density at $z$.
Johnson et al. (2013) evaluated $\rho_*(z)$ by the relation
$\rho_*(z) = \Psi_{\ast, \rm II}(z)\ t_{\ast}$,
where $\Psi_{\ast, \rm II}(z)$ is the Pop II SFR density, and $t_{\ast}=5\ {\rm Myr}$ the lifetime of a massive star.
Then, Eq. \eqref{eq:BFUV_1} can be rewritten as
\begin{equation}
J_{21, \rm bg}(z) \sim 0.1 \left(\frac{\Psi_{\ast, \rm II}(z)}{10^{-3}\ {\rm M}_{\odot}\ {\rm yr}^{-1}\ {\rm Mpc}^{-3}}\right) \left(\frac{1+z}{11}\right)^{3}.
\label{eq:J_theory}
\end{equation}
Following \cite{Stacy2007}, the energy density of the background CRs $U_{\rm CR}(z)$ is estimated from
\begin{equation}
U_{\rm CR}(z) \sim p_{\rm CR} E_{\rm SN} f_{\rm SN} \Psi_{\ast, \rm II}(z) t_{\rm H}(z) (1+z)^3,
\label{eq:U_theory}
\end{equation}
\begin{equation}
\sim 4 \times 10^{-17} \left(\frac{\Psi_{\ast, \rm II}(z)}{10^{-3}\ {\rm M}_{\odot} {\rm yr}^{-1}\ {\rm Mpc}^{-3}}\right)\left(\frac{1+z}{11}\right)^{\frac{3}{2}}{\rm erg}\ {\rm cm}^{-3}, \notag
\end{equation}
where $E_{\rm SN}=10^{51}\ {\rm erg}$ is the SN energy, $p_{\rm CR}=0.1$ the
efficiency of $E_{\rm SN}$ converted to CR acceleration, $f_{\rm SN}$ the number of SNe per
unit mass of stars formed, and $t_{\rm H}(z)$ the Hubble time at $z$.
$f_{\rm SN}$ is obtained as $\sim 2 \times 10^{-2}\ {\rm M}_{\odot}^{-1}$
assuming that massive stars with $\gtrsim 8\ {\rm M}_{\odot}$ explode as SNe
for the Salpeter IMF.
The CR ionization rate of a hydrogen atom $\zeta_{\rm CR}$ is related with $U_{\rm CR}$ as
\citep{Inayoshi2011}:
\begin{equation}
\zeta_{\rm CR} = 1.7 \times 10^{-18} \left(\frac{U_{\rm CR}}{10^{-15}\ {\rm erg}\ {\rm cm}^{-3}}\right)\ {\rm s}^{-1}.
\label{eq:zeta_U}
\end{equation}
Combining Eqs. \eqref{eq:U_theory} and \eqref{eq:zeta_U},
\begin{equation}
\zeta_{\rm CR}(z) \sim 6.8 \times 10^{-20}\left(\frac{\Psi_{\ast, \rm II}(z)}{10^{-3}{\rm M}_{\odot}{\rm yr}^{-1}{\rm Mpc}^{-3}}\right)\left(\frac{1+z}{11}\right)^{\frac{3}{2}}{\rm s}^{-1}.
\label{eq:zeta_theory}
\end{equation}
The high-$z$ SFR density has been evaluated theoretically by some authors~(e.g., Tornatore et al. 2007; Trenti \& Stiavelli 2009; Johnson et al. 2013).
According to Johnson et al. (2013), $\Psi_{\ast, \rm II}(z)$ increases almost monotonically from $10^{-4}$ to $3 \times 10^{-2}\ {\rm M}_{\odot} {\rm yr}^{-1} {\rm Mpc}^{-3}$ as $z$ decreases from $15$ to $6$.
The blue thin-solid lines in Figs. \ref{fig:h2_zeta_Jcrit} and \ref{fig:sh_zeta_Jcrit} show the evolution of $J_{21, \rm bg}(z)$ and $\zeta_{\rm CR}(z)$ evaluated from Eqs. \eqref{eq:J_theory}, \eqref{eq:zeta_theory} and $\Psi_{\ast, \rm II}(z)$ of Johnson et al. (2013).
The background intensities move from left to right along the line, and the blue points indicate the values
at $z = 15, 10$ and $6$, respectively.
Although the high-$z$ SFR density is still uncertain, most studies concluded $10^{-4} \lesssim \Psi_{\ast, \rm II}(z)/{\rm M}_{\odot} {\rm yr}^{-1} {\rm Mpc}^{-3} \lesssim 3 \times10^{-2}$ for $6 \lesssim z \lesssim 15$~(e.g., Tornatore et al. 2007; Trenti \& Stiavelli 2009; Johnson et al. 2013).
The shaded regions in Figs. \ref{fig:h2_zeta_Jcrit} and \ref{fig:sh_zeta_Jcrit} indicate the range of $J_{21, \rm bg}(z)$ and $\zeta_{\rm CR}(z)$ corresponding to this parameter range.
First, we examine the case of free-falling clouds in relic HII regions~(Fig. \ref{fig:h2_zeta_Jcrit}).
For initial densities lower than the fiducial case $n_{\rm H, 0}=0.3\ {\rm cm^{-3}}$~(black solid line),
the critical FUV intensity for HD cooling $J_{21, {\rm crit}}$ is below the background level $J_{21, \rm bg}$ at $z \gtrsim 10$, and HD cooling is suppressed in Pop III.2 star formation.
As the background CR intensity $\zeta_{\rm CR}$ increases with cosmic star formation, the critical intensity
$J_{21, {\rm crit}}$ rises and eventually exceeds the background level $J_{21, \rm bg}$ at $z \lesssim 10$.
Below this redshift, HD cooling becomes efficient and HD-Pop III star formation occurs if some primordial environments still survive.
Although in a very dense HII region~(dot-dashed line), the critical intensity $J_{21, {\rm crit}}$ is above the background level $J_{21, \rm bg}$ even at $z \gtrsim 10$, such dense conditions appear to be hardly realized~(Kitayama et al. 2004; Whalen et al. 2004, Yoshida et al. 2007a,Abel2007).
We conclude that, in relic HII regions, HD-Pop III star formation is possible only at $z \lesssim 10$, where the intensity of the CR background is modest.
On the other hand, for the shock-compressed gas (Fig. \ref{fig:sh_zeta_Jcrit}), the critical value $J_{21, {\rm crit}}$ exceeds the background level $J_{21, \rm bg}$ for most of
the redshift we consider ($6 \la z \la 15$), and thus HD-Pop III star formation occurs.
Note that the pre-shock densities we adopt in Fig. \ref{fig:sh_zeta_Jcrit} are rather low,
and our results may give the lower bound of the critical FUV intensity $J_{21, {\rm crit}}$.
In fact, in the mass assembly process of first galaxies, cold-streaming shocks occur at the regions as dense as $n_{{\rm H}, 0} = 1\mbox{-}10^3\ {\rm cm}^{-3}$~(Wise \& Abel 2007; Wise et al. 2008).
In such a dense shock environment, HD-Pop III star formation would be more common than the cases we have studied.
In conclusion, the HD-Pop III mode is the main mode of Pop III.2 star formation in the cloud compressed either by a virialization or SN shock.
\section{Discussion}\label{sec:discussion}
So far, we have considered that the background FUV photons are uniformly distributed in the universe.
In reality, they have some spatial fluctuations caused by local stellar sources, like star-forming galaxies~(Dijkstra et al. 2008; Ahn et al. 2009; Agarwal et al. 2012; Johnson et al. 2013).
In the primordial clouds irradiated with the strong FUV field as $J_{21} > 3$, HD formation is suppressed even in the shock-compressed gas unless the initial density is high ($10\ {\rm cm}^{-3}$) or the CR intensity is strong
($> 10^{-18}\ {\rm s}^{-1}$: see Fig. 12).
However, according to Johnson et al. (2013), only $\la 10$ percent of the dense primordial cloud with $n_{\rm H} > 1\ {\rm cm}^{-3}$ is exposed to such a strong FUV field as $J_{21} > 3$.
Thus, our analysis with the uniform background FUV field is enough to discuss the major character of primordial star formation and the spatial fluctuations of it do not affect our main conclusions.
We also assume that CRs accelerated in SN remnants escape freely from their host galaxies and fill the universe homogeneously.
However, in reality, they may be trapped in the host galaxy by the interaction with the magnetic field, and propagate diffusively through the ISM of the galaxy~\citep{Longair2011}.
On the way of propagation, CRs also lose their kinetic energy by ionizing the neutral ISM.
For the efficient CR escape into the IGM, the CR escape time by diffusion $\tau_{\rm diff}$ should be shorter than both the ionization loss time $\tau_{\rm ion}$ and the age of the system $\tau_{\rm age}$, which is roughly equal to the age of the universe. Here, $\tau_{\rm diff}$ and $\tau_{\rm ion}$ depends on the kinetic energy of a CR particle $E$ and the redshift $z$.
\cite{Rollinde2008} estimated the above timescales for CRs with $E=30$ MeV and discussed the CR escape efficiency during the cosmic structure formation.
Here, we extend their arguments by examining the $E$ dependence of the timescales.
Following \cite{Stacy2007}, we assume that the energy spectrum of the CR number density has a power-law shape $dn_{\rm CR}/d E \propto E^{-x}$ with $E = 10^6\mbox{-}10^{15}$ eV.
We find that the inequality $\tau_{\rm diff}(E, z) < \tau_{\rm age}(z)$ holds for all the CR energy and redshift we consider~($6 \la z \la 15$).
On the other hand, the inequality $\tau_{\rm diff}(E, z) < \tau_{\rm ion}(E, z)$ holds only for high energy CRs with $E \ga 10^7$ eV at $6 \la z \la 15$.
Thus, the CR background is composed of CRs with $E \ga 10^7$ eV and its intensity varies with the spectrum index $x$.
In the standard Fermi acceleration theory, the spectrum index is given as $x = 2$~\citep[e.g.,][]{Bell1978}, and the background intensity is reduced only by 10\ \% due to the confinement of CRs with $E = 10^6\mbox{-}10^7$ eV.
On the other hand, if we adopt a steeper slope as $x = 3$ in~\cite{Rollinde2008}, it is reduced by an order of magnitude.
Even in this case, if a star-forming pristine gas resides in the same galaxy hosting SNe, it is exposed to the CR intensity in the original or even the enhanced level, because of the efficient confinement within the galaxy.
For more quantitive and definitive discussions, a detailed study of CR propagation in the ISM of high-$z$ galaxies is needed and is beyond the scope of this paper.
At the epoch of galaxy formation, high-mass X-ray binaries and miniquasars may be the plausible sources of high energy photons extending to the X-ray band~\citep{Glover2003}.
Since the photoionization cross section of neutral atoms generally decreases with frequency in the power-law way, the IGM becomes optically thin for X-ray photons with $\ga 1$ keV and the X-ray background develops at this epoch~(Haiman et al. 2000).
The injection of X-ray photons enhances the cooling efficiency of primordial clouds by indirectly promoting H$_2$ and HD formation through the photoionization of neutral atoms~\citep[e.g.,][]{Haiman1996a,Haiman2000,Glover2003,Machacek2003}.
Nonetheless, we did not include this X-ray feedback, and here we briefly discuss how it changes our results.
Following \cite{Glover2003}, we assume that the X-ray background has a power-law spectrum
\begin{equation}
J_{\rm X}(\nu) = J_{{\rm X}, 21} \times 10^{-21} \left(\frac{\nu}{1\ {\rm keV}}\right)^{-1.5}\ {\rm erg}\ {\rm cm}^{-2}\ {\rm s}^{-1}\ {\rm Hz}^{-1},
\label{eq:J_X}
\end{equation}
where $J_{\rm X, 21} \equiv J_{1 \rm keV}/10^{-21}\ {\rm erg}\ {\rm cm}^{-2}\ {\rm s}^{-1}\ {\rm Hz}^{-1}$ and $J_{1 \rm keV}$ is the intensity at 1 keV.
Then, using Eqs. (17-19) in \cite{Inayoshi2011}, the ionization rate of H is calculated as
\begin{equation}
\zeta_{\rm X}^{\rm H} \sim 2 \times 10^{-15}J_{\rm X, 21}\ {\rm s}^{-1},
\label{eq:zeta_X_J_X}
\end{equation}
for the hydrogen column density $N_{\rm H} \lesssim 10^{22}\ {\rm cm}^{-2}$.
Since the background field derives from the stellar activity, the X-ray intensity is related with the CR energy density as
\begin{equation}
J_{{\rm X}, 21} \sim 10^{-3} \left(\frac{U_{\rm CR}}{10^{-15}\ {\rm erg}\ {\rm cm}^{-3}}\right),
\label{eq:J_X_u_cr}
\end{equation}
where we suppose that CR sources are Pop II clusters with the Salpeter IMF and the mass range $1 \mbox{-}100\ {\rm M}_{\odot}$~\citep{Inayoshi2011}\footnote{We adopt $E_{\rm SN} = 10^{51}$ erg as the energy of a Pop II SN instead of $E_{\rm SN} = 10^{52}$ erg in \cite{Inayoshi2011}.}.
Substituting Eq. \eqref{eq:zeta_U} into Eq. \eqref{eq:J_X_u_cr} and Eq. \eqref{eq:J_X_u_cr} into Eq. \eqref{eq:zeta_X_J_X}, the X-ray ionization rate is related with the CR ionization rate as
\begin{equation}
\zeta_{\rm X}^{\rm H} \sim \zeta_{\rm CR}.
\label{eq:zetaX_zetaCR}
\end{equation}
Thus, we find that, at each redshift, X-ray feedback changes the ionization rate in pristine clouds not by an order of magnitude, but by only up to a factor of two.
Note that this is the upper limit of the X-ray effect since X-ray is more effectively shielded than CRs.
Therefore, our main conclusion still remains valid even if we consider X-ray feedback.
As we have seen above, Pop III star formation at the epoch of galaxy formation is affected significantly by the injection of CRs.
Nonetheless, some previous studies calculated the Pop III SFR by considering the regulation process
of star formation from FUV feedback alone, and the positive feedback from CRs is neglected~(e.g., Tornatore et al. 2007; Trenti \& Stiavelli 2009; de Souza et al. 2011; Johnson et al. 2013).
Future studies should take all these feedback into consideration.
We must admit that our treatment of the plane-parallel and steady shock
is too simplistic to capture possible complex phenomena induced by e.g.,
the turbulence, rotation, thermal instability, and magnetic field.
For example, \cite{Inoue2008,Inoue2009} studied the effect of magnetic fields on the thermal evolution of a shock-compressed gas, using the two-dimensional MHD simulation. They found that after the short isobaric contraction, the magnetic pressure is amplified to be comparable to the ram pressure of the pre-shock materials, and thereafter the shock-compressed gas cools almost isochorically.
This implies that the shielding of H$_2$ against a FUV field becomes less effective and the resultant H$_2$ fraction is lowered in the presence of magnetic fields.
In this case, we expect that more massive clumps may form after the gas becomes gravitationally unstable and fragments, compared to the isobaric contraction case without magnetic fields.
However, their calculation did not follow the evolution until the self-gravity of the shocked layer becomes important.
They did not consider the gas in the primordial composition as well.
Thus, the above effects on our results should be investigated through the 3D MHD simulations in the future.
Another example is that our assumption of the time-independent shock velocity becomes invalid at least in the SN-shock, according to the one-dimensional numerical calculations~\citep{Machida2005, Nagakura2009, Chiaki2013}.
They showed that the SN-shock velocity becomes lower than $100\ {\rm km}\ {\rm s}^{-1}$ at $\sim 10^5$ yr after the explosion.
Moreover, at $\gtrsim 6 \times 10^5$ yr, the ram pressure of the ambient medium decreases below the pressure in the shell so that the shell re-expands and decreases its temperature and density in an almost adiabatic manner.
Although in this case, H$_2$ is produced efficiently to $\sim 10^{-3}$ and the temperature decreases to $\sim 200$ K by H$_2$ cooling, HD is produced only to $\sim 10^{-6}$, which is $\sim 10$ times less than our results, in the adiabatic expansion phase owing to the low gas density.
In this case, the temperature decrease below $100$ K is not caused by HD cooling but by the adiabatic expansion of the shell~\citep{Nagakura2009}.
However, if the SN shock can collect enough materials to trigger gas fragmentation and contraction by self-gravity before dissolving into the ISM, HD formation becomes effective and HD cooling becomes dominated in the free-falling phase, as we can see from Fig. 6 in \cite{Machida2005} and Fig. 6 in \cite{Chiaki2013}.
Thus, whether this realistic evolution of the SN shock front significantly changes our results is not clear, and more detailed numerical calculations are needed for more concrete discussion.
\section{Conclusions}\label{sec:conclusion}
HD molecules form abundantly in the primordial gas with enhanced initial ionization and play a dominant role in gas cooling.
At the epoch of galaxy formation~($z \lesssim 15$), shocks occur ubiquitously associated with the mass assembly
as well as the SN explosions.
In such a shock-compressed gas, the ionization degree jumps up owing to collisional ionization.
Other examples include the cloud in the relic HII region of a defunct Pop III star and
that with modest CR irradiation.
However, even a low level of the FUV background suppresses HD formation/cooling via ${\rm H}_2$ photodissociation
in cases without CR irradiation.
In this paper, we have examined the conditions for efficient HD cooling in primordial star formation by calculating the thermal and chemical evolution of a gas cloud under both FUV and CR irradiation.
We have obtained the critical FUV intensity $J_{21, {\rm crit}}$ for HD cooling as a function of the CR intensity $\zeta_{\rm CR}$, and compared it with the estimated background level $J_{21, \rm bg}$.
We have considered both cases of the free-falling cloud in a relic HII region and a shock-compressed gas.
At $z \ga 10$, the background CR intensity is estimated as $\sim 0.1\ \%$ of the Galactic value. In this case, the critical FUV intensity $J_{21, {\rm crit}}$ in relic HII regions is below the background level $J_{21, \rm bg}$ and HD cooling is suppressed there.
However, as the background CR intensity increases with the cosmic star formation, the critical intensity $J_{21, {\rm crit}}$ is also elevated and eventually exceeds the background level $J_{21, \rm bg}$ at $z \la 10$.
Below this redshift, HD cooling becomes efficient and HD-Pop III star formation proceeds
if some primordial environments still survive.
On the other hand, the critical FUV intensity $J_{21, {\rm crit}}$ for the shock-compressed gas is $\sim 10$ times higher than that for the relic HII region even in the weak CR regime as $\sim 0.1\ \%$ of the Galactic value.
This is because the shock-compressed gas evolves isobarically~(i.e., gas cools with its density increasing), while the gas in a relic HII region evolves isochorically~(i.e., the gas cools at a constant density) due to the initial rapid cooling.
Moreover, shocks tend to occur in denser environments~(e.g., cold-accretion flows) than relic HII regions.
As a result, the shock-compressed gas is shielded from the FUV field more effectively, and has higher critical intensity.
Thus, for the shock-compressed gas, the critical value $J_{21, {\rm crit}}$ exceeds the background level $J_{21, \rm bg}$
and HD-Pop III star formation proceeds for most of the redshift we consider ($6 \la z \la 15$).
Our result suggests that HD-Pop III stars can be more common than previously considered and could be
even the dominant population of Pop III stars.
H$_2$-Pop III stars are born in clumps as massive as $\sim 10^3\ {\rm M}_{\odot}$, but the final stellar mass is set to $\sim 100\ {\rm M}_{\odot}$ by the stellar radiative feedback onto the accretion
flow~\citep{McKee2008, Hosokawa2011,Hirano2014}.
On the other hand, HD-Pop III stars form in clumps of $\sim 100\ {\rm M}_{\odot}$, and the final mass is typically set to a few $10\ {\rm M}_{\odot}$ by the stellar feedback \citep{Hosokawa2012}.
These previous results may suggest that the star formation efficiency, i.e., the mass ratio of formed
stars to the parent clouds, of HD-Pop III stars is several times higher
than that of H2-Pop III stars. They also imply the possibility that HD-Pop III stars could be the majority of Pop III.2 stars.
Then, HD-Pop III stars of a few 10 ${\rm M}_{\odot}$ can be important sources of radiative feedback
at the epoch of galaxy formation.
Moreover, according to the observations of the abundance pattens in the metal-poor halo stars in the Galaxy, the imprints of a pair-instability SN have not been discovered yet~\citep{Tumlinson2004, Frebel2009}.
Since stars with $10\mbox{-}40\ {\rm M}_{\odot}$ end their lives as core-collapse SNe~\citep{Heger2003}, HD-Pop III stars might be the major contributor to the metal enrichment in the ISM/IGM at that epoch, if there were no PISNe in the universe.
Recent discoveries of the almost pristine gas with metallicity $\lesssim 10^{-4}\ {\rm Z}_{\odot}$
in the damped Ly $\alpha$ systems at $z \sim 3$ and $7$~\citep{Fumagalli2011,Simcoe2012}
imply that metal enrichment in the IGM has proceeded quite inhomogeneously.
Theoretical models also show the inhomogeneous nature of metal enrichment in the IGM.
They predict that Pop III star formation can continue
up to $z \sim 6$ or even lower, and evaluate the ratio of the Pop III SFR to the total SFR as $\sim 10^{-3}\mbox{-}10^{-2}$ ($\sim 10^{-4}$) at $z=10$ ($z=6$)~(Tornatore et al. 2007; Johnson et al. 2013).
This indicates that Pop III stars can form to some extent even at $z \lesssim 10$, and our study show that
they are typically formed as HD-Pop III stars with a few $10\ {\rm M}_{\odot}$.
GRBs and SNe from these Pop III progenitors at $z \la 10$ will be detected with the current or
future facilities and may provide us with information on the mass, IMF and SFR of these stars~(Nakauchi et al. 2012; Whalen et al. 2013; Tanaka et al. 2013).
Moreover, if the HD absorption lines are detected in the spectrum of a GRB afterglow which
lacks metal absorption, HD is confirmed as the dominant coolant in that primordial gas~\citep{Inoue2007}.
\section*{Acknowledgements}
We thank the anonymous referee for helpful comments and improving the quality of this paper.
We also thank K. Kashiyama, T. Nakamura Y. Suwa, and K. Tanaka for fruitful discussions and comments.
This work is supported in part by the Grant-in-Aid from the Ministry
of Education, Culture, Sports, Science and Technology (MEXT) of Japan~(23-838 KI; 25287040 KO).
|
\section{Introduction}
A wormhole can be defined as a tunnel which can joint two
universes \cite{MorTho}. Since General Relativity does not
preclude the existence of (traversable) wormholes, a large number
of papers have been written which clarify, support, or contradict
much of the research about wormholes. Morris and Thorne
\cite{MorTho} have shown that in order to construct a traversable
wormhole, one needs to have extraordinary material, denoted as
exotic matter. Exotic matter can guarantee the flare-out condition
of the wormhole at its throat.
Unlike the classical form of matter \cite{HawkingLSS}, it is
believed that the exotic matter violates the well-known energy
conditions such as the null energy conditions (NEC), weak energy
conditions (WEC), strong energy conditions (SEC) and dominant
energy conditions (DEC). One of the open questions about the
exotic matter is that if it can be formed in macroscopic
quantities or not. We should note that these energy conditions are
violated by certain states of quantum fields, amongst which we may
refer to the Casimir energy, Hawking evaporation, and vacuum
polarization \cite{NegativeEnergy}. Furthermore, it has been shown
that one of the effective causes of the (late time) cosmic
acceleration is an exotic fluid \cite{Exotic1}. Hence, it will be
motivated to study wormhole solutions, at least geometrically.
Traversable wormholes in the Dvali-Gabadadze-Porrati theory with
cylindrical symmetry has been studied in Ref. \cite{Richarte}.
Higher dimensional Lorentzian wormholes have been analyzed by
several authors \cite{Kar}. Moreover, wormhole solutions of higher
derivative gravity with linear and nonlinear Maxwell fields have
been considered in \cite{WormHDG}. For other kinds of wormhole
solutions, we refer the reader to Refs. \cite{CUTpaste,WormHDG}
and references therein.
Many authors have extensively considered the nonlinear
electrodynamics and used their results to explain some physical
phenomena \cite{Cuesta,Marklund,BIpapers,PMIpapers,HendiJHEP}. A
charged system whose performance cannot be described by the linear
equations may be characterized with nonlinear electrodynamics.
From mathematical point of view, since Maxwell equations
originated from the empirical nature, we can consider a general
nonlinear theory of electrodynamics and state that the Maxwell
fields, are only approximations of nonlinear electrodynamics,
which the approximation breaks down for the small distances. From
physical viewpoint, generalizations of Maxwell theory to nonlinear
electrodynamics were introduced to eliminate infinite quantities
in theoretical analysis of the electrodynamics \cite{BIpapers}. In
addition, one may find some various limitations of the linear
electrodynamics in Ref. \cite{Delphenich}.
Recently, we take into account new classes of nonlinear electrodynamics,
such as Born-Infeld (BI) like \cite{HendiJHEP} and power-Maxwell invariant
(PMI) \cite{PMIpapers} nonlinear electrodynamics, in order to obtain new
analytical solutions in Einstein and higher derivative gravity.
Motivated above, in this paper we look for the analytical magnetic
horizonless solutions of Einstein gravity with nonlinear Maxwell source.
Properties of the solutions will be investigated.
\section{Field Equations and wormhole solutions:}
The field equations of Einstein gravity with an arbitrary $U(1)$ gauge field
as a source, may be written a
\begin{equation}
G_{\mu \nu }+\Lambda g_{\mu \nu }=\frac{1}{2}g_{\mu \nu
}L(\mathcal{F})-2L_{\mathcal{F}}F_{\mu \lambda }F_{\nu
}^{\;\lambda }, \label{Geq}
\end{equation}
\begin{equation}
\partial _{\mu }\left( \sqrt{-g}L_{\mathcal{F}}F^{\mu \nu }\right) =0,
\label{Maxeq}
\end{equation}
where $G_{\mu \nu }$ is the Einstein tensor, $\Lambda =-3/2l^{2}$ denotes
the four dimensional negative cosmological constant, $L(\mathcal{F})$ is an
arbitrary function of the closed $2$-form Maxwell invariant $\mathcal{F
=F_{\mu \nu }F^{\mu \nu } $ and $L_{\mathcal{F}}=\frac{dL(\mathcal{F})}{
\mathcal{F}}$.
In addition to PMI and BI theories, in this paper, we take into account the
recently proposed BI-like models \cite{HendiJHEP}, which we called them as
Exponential form of nonlinear electrodynamics theory (ENE) and Logarithmic
form of nonlinear electrodynamics theory (LNE), whose Lagrangians are
\begin{equation}
L(\mathcal{F})=\left\{
\begin{array}{ll}
\left( -\mathcal{F}\right) ^{s}, & \;~{PMI} \\
4\beta ^{2}\left( 1-\sqrt{1+\frac{\mathcal{F}}{2\beta ^{2}}}\right), & \;~{B
} \\
\beta ^{2}\left( \exp (-\frac{\mathcal{F}}{\beta ^{2}} )-1\right) , & \;~{EN
} \\
-8\beta ^{2}\ln \left( 1+\frac{\mathcal{F}}{8\beta ^{2}} \right) , & \;~{LNE
\end{array}
\right. , \label{Lnon}
\end{equation}
where $s$ and $\beta $ are two nonlinearity parameters. Expanding the
mentioned Lagrangians near the linear Maxwell case ($s\longrightarrow 1$ and
$\beta \longrightarrow \infty $), one can obtai
\begin{equation}
L(\mathcal{F})\longrightarrow L_{Max}+\left\{
\begin{array}{ll}
-\mathcal{F}\ln \left( -\mathcal{F}\right) (s-1)+O(s-1)^{2}, & \;~{PMI} \\
+\frac{\chi \mathcal{F}^{2}}{16\beta ^{2}}+O\left( \frac{\mathcal{F}^{3}}
\beta ^{4}}\right), & \;~{others
\end{array}
\right. , \label{Lbast}
\end{equation}
where Maxwell Lagrangian $L_{Max}=-\mathcal{F}$ and $\chi =1$, $2$ and $8$
for LNE, BI and ENE branches, respectively.
Investigation of the effects of the higher derivative corrections to the
gauge field seems to be an interesting phenomenon. These nonlinear
electrodynamics sources have different effects on the physical properties of
the solutions. For example in black hole framework, these nonlinearities may
change the electric potential, temperature, horizon geometry, energy density
distribution and also asymptotic behavior of the solutions. In what follows,
we study the effects of nonlinearity on the magnetic solutions.
Motivated by the fact that we are looking for the horizonless magnetic
solution (instead of electric one), one can start with the following $4
-dimensional spacetime
\begin{equation}
ds^{2}=-\frac{r^{2}}{l^{2}}dt^{2}+\frac{dr^{2}}{f(r)}+\Upsilon
^{2}l^{2}f(r)d\theta ^{2}+r^{2}d\phi ^{2}, \label{StaticMetric}
\end{equation
where $\Upsilon $ is a constant and will be fixed later. We should note
that, because of the periodic nature of $\theta $, one can obtain the
presented metric (\ref{StaticMetric}) with \emph{local} transformations
t\rightarrow il\Upsilon \theta $ and $\theta \rightarrow it/l$ in the
horizon-flat Schwarzschild metric, $ds^{2}=-f(r)dt^{2}+\frac{dr^{2}}{f(r)
+r^{2}\left( d\theta ^{2}+d\phi ^{2}\right) $. In other words, metric (\re
{StaticMetric}) may be \emph{locally} mapped to Schwarzschild spacetime, but
not \emph{globally}. Considering the mentioned local transformation, one can
find that the nonzero component of the gauge potential is $A_{\theta }
\begin{equation}
A_{\mu }=h(r)\delta _{\mu }^{\theta }, \label{PotStatic}
\end{equation
where $h(r)$ is an arbitrary function of $r$. Using Eq. (\ref{Maxeq}) with
the metric (\ref{StaticMetric}), we find $h(r)=\int E(r)dr$ in whic
\begin{equation}
E(r)=\left\{
\begin{array}{ll}
\frac{2ql^{2}\Upsilon ^{2}}{r^{2/(2s-1)}}, & \;~PMI \\
\frac{2ql^{2}\Upsilon ^{2}}{r^{2}\sqrt{1+\frac{4q^{2}l^{4}\Upsilon ^{2}}
\beta ^{2}r^{4}}}}, & \;~BI \\
\frac{2ql^{2}\Upsilon ^{2}}{r^{2}}\exp \left( -\frac{L_{W}}{2}\right) , & \;
{ENE} \\
\frac{\beta ^{2}r^{2}\left( \Gamma -1\right) }{ql^{2}}, & \;~{LNE
\end{array
\right. , \label{hp}
\end{equation
where $L_{W}=LambertW\left( X\right) ,$ $X=\frac{16l^{4}q^{2}\Upsilon ^{2}}
\beta ^{2}r^{4}}$ and $\Gamma =\sqrt{1+\frac{X}{4}}$ and therefore the
nonzero component of electromagnetic field tensor i
\begin{equation}
F_{r\theta }=E(r). \label{Frtheta}
\end{equation
We should note that the physical gauge potential should vanish for large
values of $r$. This condition is satisfied for $1/2<s<3/2$ and arbitrary
\beta $ (the mentioned constrain for $s$ is used throughout the rest of the
paper). Now, one can expand Eq. (\ref{hp}) to obtain the leading
nonlinearity correction of Maxwell field
\begin{equation}
\left. E(r)\right\vert _{\;near\;the\;linear\;case}=\frac{2ql^{2}\Upsilon
^{2}}{r^{2}}+\left\{
\begin{array}{ll}
\frac{8ql^{2}\Upsilon ^{2}\ln (r)}{r^{2}}(s-1)+O(s-1)^{2}, & \;~PMI \\
-\frac{2\chi q^{3}l^{6}\Upsilon ^{4}}{\beta ^{2}r^{6}}+O(\frac{1}{\beta ^{4}
), & \;other
\end{array
\right. . \label{Ebast}
\end{equation
In order to examine the effect of nonlinearity on the
electromagnetic field, we plot Figs. \ref{Epmi} and \ref{Eothers}.
Figure \ref{Epmi} shows that when we reduce the nonlinearity $s$,
the electromagnetic field of the PMI branch diverges for
$r\longrightarrow 0$ more rapidly and for large distances it goes
to zero more quickly. Figure \ref{Eothers} shows that for all
BI-like branches, the electromagnetic field (the same behavior as
in Maxwell case) vanishes for large $r$. Near the origin the
electromagnetic field of BI and LNE branches have finite values,
but for ENE branch, it diverges. Comparing this divergency with
Maxwell one, we find that ENE divergency is more slowly.
\begin{figure}[tbp]
\epsfxsize=7cm \centerline{\epsffile{Epm.eps}} \caption{$E(r)$
versus $r$ for $q=1$, $\Upsilon=1$, $l=1$ and $s=1.4$ (solid
line), $s=1.2$ (dashed line), $s=1$ "Maxwell field" (bold line),
$s=0.8$ (dotted line) and $s=0.6$ (dash-dotted line)} \label{Epmi}
\end{figure}
\begin{figure}[tbp]
\epsfxsize=7cm \centerline{\epsffile{Eoth.eps}}
\caption{$E(r)$ versus $r$ for $q=1$, $\Upsilon=1$, $l=1$ and $\protect\bet
=1$. BI (solid line), ENE (dashed line), LNE (dotted line) and Maxwell field
(bold line)}
\label{Eothers}
\end{figure}
Taking into account the electromagnetic field tensor, we are in a position
to find the function $f(r)$. In order to obtain it, one may use any
components of Eq. (\ref{Geq}). We simplify the components of Eq. (\ref{Geq})
and find that the nonzero independent components of Eq. (\ref{Geq}) are
\begin{eqnarray}
&&f^{\prime \prime }(r)+\frac{2f^{\prime }(r)}{r}+2\Lambda +\Delta _{1}(r)
=0, \label{FE1} \\
&&f^{\prime }(r)+\frac{f(r)}{r}+\Lambda r+\Delta _{2}(r) =0, \label{FE2}
\end{eqnarray}
wit
\begin{eqnarray}
\Delta _{1}(r) &=&\left\{
\begin{array}{ll}
-\left( \frac{2h^{\prime 2}(r)}{\Upsilon ^{2}}\right) ^{s}, & \;~PMI \\
4\beta ^{2}\left[ \sqrt{1-\frac{h^{\prime 2}(r)}{\beta ^{2}\Upsilon ^{2}}}-
\right], & \;~BI \\
\beta ^{2}\left[ 1-\exp \left( \frac{2h^{\prime 2}(r)}{\beta ^{2}\Upsilon
^{2}}\right) \right] , & \;~{ENE} \\
8\beta ^{2}\ln \left[ 1-\left( \frac{h^{\prime }(r)}{2\beta \Upsilon
\right) ^{2}\right] , & \;~{LNE
\end{array}
\right. , \label{delta1} \\
\Delta _{2}(r) &=&\left\{
\begin{array}{ll}
\frac{r}{2}(2s-1)\left( \frac{2h^{\prime 2}(r)}{\Upsilon ^{2}}\right) ^{s},
& \;~PMI \\
2r\beta ^{2}\left[ \left( 1-\frac{h^{\prime 2}(r)}{\beta ^{2}\Upsilon ^{2}
\right) ^{-1/2}-1\right], & \;~BI \\
\frac{r\beta ^{2}}{2}\left\{ 1-\left[ 1-\left( \frac{2h^{\prime }(r)}{\beta
\Upsilon }\right) ^{2}\right] \exp \left( \frac{ 2h^{\prime 2}(r)}{\beta
^{2}\Upsilon ^{2}}\right) \right\}, & \;~{ENE} \\
4r\beta ^{2}\left\{ \ln \left[ 1-\left( \frac{h^{\prime }(r)}{2\beta
\Upsilon }\right) ^{2}\right] -\frac{2}{1-\left( \frac{2\beta \Upsilon }
h^{\prime }(r)}\right) ^{2}}\right\}, & \;~{LNE
\end{array}
\right. , \label{delta2}
\end{eqnarray}
where prime and double prime are first and second derivatives with respect
to $r$ , respectively. After some cumbersome manipulation, the solutions of
Eqs. (\ref{FE1}) and (\ref{FE2}) can be written as
\begin{equation}
f(r)=\frac{-2m}{r}-\frac{\Lambda r^{2}}{3}+\left\{
\begin{array}{ll}
-\frac{r^{2}(2s-1)^{2}}{2(2s-3)}\left( \frac{8q^{2}l^{4}\Upsilon ^{2}}
r^{4/(2s-1)}}\right) ^{s}, & \;~PMI \\
\frac{2\beta ^{2}r^{2}}{3}-\frac{2\beta ^{2}}{r}\int \sqrt{1+\frac
4q^{2}l^{4}\Upsilon ^{2}}{\beta ^{2}r^{4}}}r^{2}dr, & \;~BI \\
-\frac{\beta ^{2}r^{2}}{6}+\frac{2\beta ql^{2}\Upsilon }{r}\int \left( \frac
1}{\sqrt{L_{W}}}-\sqrt{L_{W}}\right) dr, & \;~{ENE} \\
+\frac{8\beta ^{2}r^{2}}{3}-\frac{4\beta ^{2}}{r}\int r^{2}\ln \left( \frac
\beta ^{2}r^{4}\left( \Gamma -1\right) }{2q^{2}\Upsilon ^{2}l^{4}}\right) dr
\frac{16l^{4}q^{2}\Upsilon ^{2}}{r}\int \frac{dr}{r^{2}\left( \Gamma
-1\right) }, & \;~{LNE
\end{array}
\right. , \label{f(r)}
\end{equation}
where $m$ is the integration constant which is related to the mass
parameter. In order to investigate the effect of nonlinearity on the metric
function, simplistically, we expand $f(r)$\ for $s\longrightarrow 1$ for PMI
and $\beta \longrightarrow \infty $ for other branches. After some
manipulation, we obtain
\begin{equation}
f(r)=f_{Max}(r)+\left\{
\begin{array}{ll}
\frac{4q^{2}l^{4}\Upsilon ^{2}\left[ 6+\ln \left( 8q^{2}\Upsilon
^{2}l^{4}r^{4}\right) \right] }{r^{2}}(s-1)+O(s-1)^{2}, & \;~PMI \\
-\frac{2\chi q^{4}l^{8}\Upsilon ^{4}}{5\beta ^{2}r^{6}}+O(\frac{1}{\beta ^{4
}), & \;other
\end{array}
\right. , \label{Bast}
\end{equation}
where $f_{Max}(r)$ is the magnetic solution of Einstein-Maxwell gravity
\begin{equation}
f_{Max}(r)=\frac{-2m}{r}-\frac{\Lambda r^{2}}{3}+\frac{4q^{2}l^{4}\Upsilon
^{2}}{r^{2}}, \label{f(r)Maxwell}
\end{equation}
and the second term on the right hand side of Eq. (\ref{Bast}) is the
leading nonlinearity correction to the Einstein-Maxwell gravity solution.
\subsection{Properties of the solutions}
At first step, we should note that the presented solutions are
asymptotically anti-de Sitter (adS) and they reduce to asymptotically adS
Einstein-Maxwell solutions for $s\longrightarrow 1$ (PMI branch) or $\beta
\longrightarrow \infty $ (other branches).
The second step should be devoted to singularities and hence we should
calculate the curvature invariants. One can show that for the metric (\re
{StaticMetric}), the Kretschmann and Ricci scalars are
\begin{eqnarray}
R_{\mu \nu \rho \sigma }R^{\mu \nu \rho \sigma } &=&f^{\prime \prime 2}(r)
\frac{4f^{\prime 2}(r)}{r^{2}}+\frac{4f^{2}(r)}{r^{4}}, \label{Kretschmann}
\\
R &=&-f^{\prime \prime }(r)-\frac{4f^{\prime }(r)}{r}-\frac{2f(r)}{r^{2}}.
\label{Ricci}
\end{eqnarray}
Inserting Eq. (\ref{f(r)}) into the Eqs. (\ref{Kretschmann}) and (\ref{Ricci
), and using numerical calculations, one can show that the Ricci and
Kretschmann scalars diverge at $r=0$, are finite for $r>0$\ and for
r\rightarrow \infty $ one obtains
\begin{eqnarray}
\left. R_{\mu \nu \rho \sigma }R^{\mu \nu \rho \sigma }\right\vert _{
Large\; r} &=&\frac{8\Lambda ^{2}}{3}+\left\{
\begin{array}{ll}
O\left( \frac{1}{r^{\xi }}\right) & \;~PMI \\
O\left( \frac{1}{r^{6}}\right), & \;~other
\end{array}
\right. , \label{RRinfinity} \\
\left. R\right\vert _{Large\;r} &=&4\Lambda +\left\{
\begin{array}{ll}
O\left( \frac{1}{r^{\xi }}\right) & \;~PMI \\
O\left( \frac{1}{r^{8}}\right), & \;~other
\end{array}
\right. , \label{Rinfinity} \\
\xi &\in &\left( 3,\infty \right), \nonumber
\end{eqnarray}
which confirms that the asymptotic behavior of the solutions is adS.
Considering the divergency of the Ricci and Kretschmann scalars at the
origin, one may think that there is a curvature singularity located at $r=0
. This singularity will be naked if the function $f(r)$\ has no real root
(singularity is not covered with a horizon) and we are not interested in it.
Therefore, we consider the case in which the function $f(r)$ has at least a
non-extreme positive real root. It is notable that the function $f(r)$\ is
negative for $r=r_{+}-\epsilon $ ($\epsilon $ is an infinitesimal number),
and positive for $r>r_{+}$\ where $r_{+}$\ is the largest positive real root
of $f(r)=0$. Negativity of the function $f(r)$ leads to an apparent change
of metric signature and it forces us to consider $r_{+}\leq r<$\ $\infty $.
We should state that although the metric function $f(r)$\ vanishes at
r=r_{+}$, but we have $f^{\prime }(r=r_{+})\neq 0$ ($f^{\prime }(r=r_{+})>0
). In addition, there is no curvature singularity in the range $r_{+}\leq
r<\infty $. Following the procedure of Ref. \cite{HendiCQG}, one may find
that there is a conic singularity at $r=r_{+}$.
Removing this conical singularity with $\Upsilon =1/[lf^{\prime }(r_{+})]$
\cite{HendiCQG}, we desire to interpret the obtained solutions as wormholes.
In order to construct wormholes from the gluing, one requires to use the
cut-and-paste prescription \cite{CUTpaste}. In this method, we take into
account two geodesically incomplete copies of the solutions (removing from
each copy the forbidden region $\Omega $) with two copies of the boundaries
\partial \Omega $, where
\begin{eqnarray}
\Omega &\equiv &\left\{ r\mid r<r_{+}\right\} , \label{forbidden} \\
\partial \Omega &\equiv &\left\{ r\mid r=r_{+}\right\} . \label{Bforbidden}
\end{eqnarray
Now, we identify two copies of the mentioned boundaries to build a
geodesically complete manifold. This cut-and-paste method constructs a
wormhole with a throat at $r=r_{+}$. In order to confirm this claim, we
should check the so-called flare-out condition at the throat. To do this,
one can consider a $2$-dimensional submanifold of the metric (\re
{StaticMetric}), $ds_{2-\dim }^{2}$, (with constant $t$ and $\theta $) and
embed it in a $3$-dimensional Euclidean flat space in cylindrical
coordinates, $ds_{3-\dim }^{2}$, where
\begin{eqnarray}
ds_{2-\dim }^{2} &=&\frac{dr^{2}}{f(r)}+r^{2}d\phi ^{2}, \label{2dim} \\
ds_{3-\dim }^{2} &=&dr^{2}+r^{2}d\phi ^{2}+dz^{2}. \label{3dim}
\end{eqnarray}
Regarding the surface $z=z(r)$, we obtain
\begin{eqnarray}
\left. \frac{dr}{dz}\right\vert _{r=r_{+}} &=&\left. \sqrt{\frac{f(r)}{1-f(r
}}\right\vert _{r=r_{+}}=0, \label{drdz} \\
\left. \frac{d^{2}r}{dz^{2}}\right\vert _{r=r_{+}} &=&\left. \frac{f^{\prime
}}{2\left[ 1-f\right] ^{2}}\right\vert _{r=r_{+}}=\frac{1}{2}f^{\prime
}(r=r_{+})>0, \label{d2rdz2}
\end{eqnarray}
which shows that the flare-out condition may be satisfied for the surface
z=z(r)$ and therefore $r=r_{+}$ is the radius of the wormhole throat.
Now, we should discuss the energy conditions for the wormhole solutions. We
should note that, traversable wormhole may exist with exotic matter which
violates the null energy condition \cite{MorTho}. We use the following
orthonormal contravariant (hatted) basis to simplify the mathematics and
physical interpretations
\begin{equation}
\mathbf{e}_{\widehat{t}}=\frac{l}{r}\frac{\partial }{\partial t},\;~ \;~
\mathbf{e}_{\widehat{r}}=f^{1/2}\frac{\partial }{\partial r},\;~\;~ \mathbf{
}_{\widehat{\theta }}=\frac{1}{\Upsilon lf^{1/2}}\frac{\partial }{\partial
\theta },\;~\;~\mathbf{e}_{\widehat{\phi }}=r^{-1}\frac{\partial }{\partial
\phi }. \label{base}
\end{equation}
Using the mentioned basis, we can obtain
\begin{eqnarray}
T_{_{\widehat{t}\widehat{t}}} &=&-T_{_{\widehat{\phi }\widehat{\phi }
}}=\left\{
\begin{array}{ll}
-\frac{1}{2}\left( \frac{8\Upsilon ^{2}q^{2}l^{4}}{r^{4/(2s-1)}}\right) ^{s},
& \;~PMI \\
2\beta ^{2}\left( \Gamma ^{-1}-1\right), & \;~BI \\
\frac{\beta ^{2}}{2}\left( 1-\sqrt{\frac{X}{L_{W}}} \right), & \;~{ENE} \\
4\beta ^{2}\ln \left( \frac{8\left( \Gamma -1\right) }{X} \right), & \;~{LNE
\end{array}
\right. , \label{Tab1} \\
\;T_{_{\widehat{r}\widehat{r}}} &=&-T_{_{\widehat{\theta }\widehat{\theta
}}=\left\{
\begin{array}{ll}
-\frac{(2s-1)}{2}\left( \frac{8\Upsilon ^{2}q^{2}l^{4}}{r^{4/(2s-1)}}\right)
^{s}, & \;~PMI \\
2\beta ^{2}\left( 1-\Gamma \right), & \;~BI \\
\frac{\beta ^{2}}{2}\left( \sqrt{\frac{X}{L_{W}}}-\sqrt{ XL_{W}}-1\right), &
\;~{ENE} \\
\beta ^{2}\left[ 8-4\ln \left( \frac{8\left( \Gamma -1\right) }{X}\right)
\frac{X}{\left( \Gamma -1\right) }\right], & \;~{LNE
\end{array}
\right. , \label{Tab2}
\end{eqnarray}
and therefore
\begin{eqnarray}
T_{_{\widehat{t}\widehat{t}}}&<&0 \nonumber \\
T_{_{\widehat{t}\widehat{t}}}+T_{_{\widehat{r}\widehat{r}}}&<&0, \nonumber
\end{eqnarray}
which shows that all the energy conditions are violated as well
(see Fig. \ref{LWLN} for more clarification).
\begin{figure}[tbp]
\begin{array}{cc}
\epsfxsize=7cm \epsffile{Lamb.eps} & \epsfxsize=7cm \epsffile{LOG.eps
\end{array}
\caption{\textbf{left figure:} $\protect\sqrt{\frac{L_{W}}{X}}$ (solid line) and $\protect\sqrt
XL_{W}}$ (bold line) versus $X$. \textbf{right figure:} $\ln
\left( \frac{8(\Gamma -1)}{X}\right) $ versus $X$.} \label{LWLN}
\end{figure}
At the end of this section, we desire to study the effects of the
nonlinearity on energy density of the spacetime. At the start, we can expand
$T_{_{\widehat{t}\widehat{t}}}$ near the linear case to obtain
\begin{equation}
T_{_{\widehat{t}\widehat{t}}}=\left. T_{_{\widehat{t}\widehat{t}
}}\right\vert _{Maxwell}+\left\{
\begin{array}{ll}
-\frac{4q^{2}l^{4}\Upsilon ^{2}\ln \left( 8q^{2}\Upsilon
^{2}l^{4}r^{4}\right) }{r^{4}}(s-1)+O(s-1)^{2}, & \;~PMI \\
+\frac{6\chi q^{4}l^{8}\Upsilon ^{4}}{\beta ^{2}r^{8}}+O(\frac{1}{\beta ^{4}
), & \;other
\end{array}
\right. , \label{Tbast}
\end{equation}
where $\left. T_{_{\widehat{t}\widehat{t}}}\right\vert _{Maxwell}=\frac
-4\Upsilon ^{2}q^{2}l^{4}}{r^{4}}$ and the second term on the right hand
side of Eq. (\ref{Tbast}) is the leading nonlinearity correction to the
energy density of the Einstein-Maxwell theory. In addition, we plot the
energy density $T_{_{\widehat{t}\widehat{t}}}$\ versus $r$ for different
values of nonlinearity parameter $s$ and also various branches of BI-like
fields. Figures \ref{TttPMI} and \ref{TttOTHERS}, show that for the
arbitrary choices of $r$ the energy density is negative. Furthermore, Fig.
\ref{TttPMI} shows that the nonlinearity parameter, $s$, has effects on the
behavior of the energy density and when we reduce $s$, both divergency of
energy density near the origin and its vanishing for large values of
distance occur more rapidly. Moreover, considering Fig. \ref{TttOTHERS}, one
can find that $T_{_{\widehat{t}\widehat{t}}}$ has a finite value for an
arbitrary distance in BI branch and it diverges near the origin for other
branches. It is notable that, near the origin, the divergency of LNE branch
is stronger than ENE branch. Also, for BI-like branches, the nonlinearity
reduces the strength of energy density divergency.
\begin{figure}[tbp]
\begin{array}{cc}
\epsfxsize=7cm \epsffile{Tpm1.eps} & \epsfxsize=7cm \epsffile{Tpm2.eps
\end{array}
\caption{$T_{_{\widehat{t}\widehat{t}}}$ versus $r$ for $q=1$, $\Upsilon=1$,
$l=1$ and $s=1.4$ (solid line), $s=1.2$ (dashed line), $s=1$ "Maxwell field"
(bold line), $s=0.8$ (dotted line) and $s=0.6$ (dash-dotted line) "\textbf
different scales}"}
\label{TttPMI}
\end{figure}
\begin{figure}[tbp]
\epsfxsize=7.5cm \centerline{\epsffile{TOTH.eps}}
\caption{$T_{_{\widehat{t}\widehat{t}}}$ versus $r$ for $q=1$, $\Upsilon=1$,
$l=1$ and $\protect\beta=1$. BI (solid line), ENE (dashed line), LNE (dotted
line) and Maxwell field (bold line)}
\label{TttOTHERS}
\end{figure}
\subsection{Rotating solutions \label{Rotworm}}
In this section, we want to add angular momentum to the static spacetime
\ref{StaticMetric}). To do this, one can use the following rotation boost in
the $t-\theta $ plane
\begin{equation}
t\mapsto \Xi t-a\theta ,\hspace{0.5cm}\theta \mapsto \Xi \theta -\frac{a}
l^{2}}t, \label{Tr}
\end{equation}
where $\Xi =\sqrt{1+a^{2}/l^{2}}$ and $a$ is a rotation parameter. Taking
into account Eq. (\ref{Tr}) and applying it to static metric (\re
{StaticMetric}), one obtains
\begin{equation}
ds^{2}=-\frac{r^{2}}{l^{2}}\left( \Xi dt-ad\theta \right) ^{2}+\frac{dr^{2}}
f(r)}+\Upsilon ^{2}l^{2}f(r)\left( \frac{a}{l^{2}}dt-\Xi d\theta \right)
^{2}+r^{2}d\phi ^{2}, \label{RotatingMetric}
\end{equation
where $f(r)$ is the same as $f(r)$ given in Eq. (\ref{f(r)}). It is notable
that one can obtain the presented metric (\ref{RotatingMetric}) with \emph
local} transformations $t\rightarrow il\Upsilon \left( at/l^{2}-\Xi \theta
\right) $ and $\theta \rightarrow i\left( \Xi t-a\theta \right) /l$ in the
horizon-flat Schwarzschild metric, $ds^{2}=-f(r)dt^{2}+\frac{dr^{2}}{f(r)
+r^{2}d\theta ^{2}+r^{2}d\phi ^{2}$. Thus, the nonzero components of the
gauge potential are $A_{\theta }$ and $A_{t}
\begin{equation}
A_{\mu }=h(r)\left( \Xi \delta _{\mu }^{\theta }-\frac{a}{l^{2}}\delta _{\mu
}^{t}\right) , \label{PotRotating}
\end{equation
where $h(r)$ is the same as in the static case. Furthermore, the nonzero
components of electromagnetic field tensor are given by
\begin{equation}
F_{tr}=\frac{a}{\Xi l^{2}}F_{r\theta }=\frac{a}{l^{2}}E(r). \label{Ftr}
\end{equation
As we mentioned before, the periodic nature of $\theta $ helps us to
conclude that the transformation (\ref{Tr}) is not a proper coordinate
transformation on the entire manifold and therefore the metrics (\re
{StaticMetric}) and (\ref{RotatingMetric}) are distinct \cite{Sta}. In
addition, it is desired to note that rotating solutions have no horizon and
curvature singularity. Moreover, it is worth noting that besides the
magnetic field along the $\theta $ coordinate, there is also a radial
electric field ($F_{tr}\neq 0$) and therefore, unlike the static case, the
rotating wormhole has a nonzero electric charge which is proportional to the
rotation parameter.
\subsection{Conserved Quantities \label{Conserve}}
Here we desire to calculate finite conserved quantities. In order to obtain
a finite value for these quantities, we can use the counterterm method
inspired by the concept of (AdS/CFT) correspondence \cite{Mal}. It has been
shown that for asymptotically AdS solutions the finite energy momentum
tensor i
\begin{equation}
T^{ab}=\frac{1}{8\pi }\left( K^{ab}-K\gamma ^{ab}-\frac{2\gamma ^{ab}}{l
\right) , \label{Stress}
\end{equation
where $K$ is the trace of the extrinsic curvature $K^{ab}$ and $\gamma ^{ab}$
is the induced metric of the boundary. Taking into account the Killing
vector field $\mathcal{\xi }$, one may obtain the quasilocal conserved
quantities in the following form
\begin{equation}
\mathcal{Q}(\mathcal{\xi )}=\int_{\mathcal{B}}d^{2}\varphi \sqrt{\sigma
T_{ab}n^{a}\mathcal{\xi }^{b}, \label{charge}
\end{equation
where $\sigma $ is the determinant of the boundary metric in ADM
(Arnowitt-Deser-Misner) form $\sigma _{ij}$, and $n^{a}$ is the timelike
unit vector normal to the boundary $\mathcal{B}$. Considering two Killing
vectors $\xi =\partial /\partial t$ and $\zeta =\partial /\partial \theta $,
we can find their associated conserved charges which are mass and angular
momentum
\begin{equation}
{M}=4\pi ^{2}\left[ 3\left( \Xi ^{2}-1\right) +1\right] \Upsilon m,
\label{M}
\end{equation}
\begin{equation}
{J}=12\pi ^{2}\Upsilon m\Xi a, \label{J}
\end{equation}
where the former equation confirms that $a$ is the rotational parameter.
Finally, we are in a position to discuss the electric charge. In order to
compute it, we need a nonzero radial electric field $F_{tr}$ and therefore
one expects vanishing $F_{tr}$ (static case) leads to zero electric charge.
Taking into account the Gauss's law for the rotating solutions and computing
the flux of the electric field at infinity, one can find
\begin{equation}
Q=\left\{
\begin{array}{ll}
\frac{2^{3s+1}\pi ^{2}\Upsilon s}{4l}\left( \frac{(2s-1)q}{(3-2s)l}\right)
^{2s-1}a, & \;~PMI \\
\frac{4\pi ^{2}\Upsilon q}{l^{2}}a, & \;~other
\end{array}
\right. , \label{chden}
\end{equation}
which confirms that the static wormholes do not have electric charge.
\section{ Closing Remarks}
In this paper, we took into account a class of magnetic Einsteinian
solutions in the presence of nonlinear source. The magnetic spacetime which
we used in this paper, may be obtained from the horizon-flat Schwarzschild
metric with \emph{local} transformations $t\rightarrow il\Upsilon \theta $
and $\theta \rightarrow it/l$. It is notable that because of the periodic
nature of $\theta$, the mentioned transformations cannot be global.
We considered four forms of nonlinear electrodynamics, namely PMI, BI, ENE
and LNE theories, whose asymptotic behavior leads to Maxwell theory. We
investigated the effect of nonlinearity parameter on the electromagnetic
field and found that for PMI branch, if one reduces the nonlinearity
parameter $s$, then the electromagnetic field diverges near the origin more
rapidly and for large distances it goes to zero more quickly. In addition,
we found that for all BI-like branches, the behavior of the electromagnetic
field is the same as Maxwell case for large values of distance, but near the
origin, the electromagnetic field of the BI and LNE branches is finite and
it diverges for the ENE branch. It is interesting to note that the
divergency of the ENE branch has less strength in comparison to the the
Maxwell field.
Then, we obtained the metric function for all branches and found that they
reduce to asymptotically adS Einstein-Maxwell solutions for
s\longrightarrow 1$ (PMI branch) or $\beta \longrightarrow \infty $ (other
branches). We also expanded the metric function near the linear Maxwell
field and calculated the curvature scalars for large $r$ to find that
obtained solutions are asymptotically anti-de Sitter (adS). Taking into
account the presented metric, one can find that the function $f(r)$ cannot
be negative since its negativity leads to an apparent change of metric
signature. This limitation forced us to consider $r_{+}\leq r<$\ $\infty $.
Using numerical calculations, one can find that there is no curvature
singularity in the range $r_{+}\leq r<\infty $, but one may find a conic
singularity at $r=r_{+}$.
After that, we removed the mentioned conic singularity and used the
cut-and-paste prescription to construct a wormhole from the gluing and then
we checked the so-called flare-out condition at the throat $r=r_{+}$. Since
it has been stated before, that traversable wormhole may exist with exotic
matter \cite{MorTho}, we investigated the energy conditions for the obtained
wormhole solutions and found that the energy conditions are violated.
We also studied the effects of nonlinearity parameter on the energy density.
We found that when we reduce $s$, both divergency of energy density near the
origin and its vanishing for large values of distance occur more rapidly.
Moreover, one can find that energy density has a finite value for an
arbitrary distance only in BI branch and it diverges near the origin for
other BI-like branches. It is notable that, near the origin, the divergency
of LNE branch is stronger than ENE branch.
We generalized the static solutions to rotating ones and obtained the
conserved quantities. We found that, unlike the static case, for the
spinning spacetime, the wormhole has a net electric charge density. We also
found that in spite of the fact that the mentioned nonlinear theories change
the properties of the solutions significantly, but they do not have any
effect on mass and angular momentum.
\begin{acknowledgements}
We are indebted to H. Mohammadpour for reading the manuscript. We
also wish to thank Shiraz University Research Council. This work
has been supported financially by Research Institute for Astronomy
\& Astrophysics of Maragha (RIAAM), Iran.
\end{acknowledgements}
|
\section{Introduction}
Suppose that we observe equispaced samples (at the Nyquist sampling rate) of a number of $L$ sinusoidal signals:
\equ{y_{jt}^o=\sum_{k=1}^K s_{kt}e^{i2\pi j f_k}, \quad \sbra{j,t}\in \m{J}\times\mbra{L}, \label{formu:observmodel1}}
denoted by matrix $\m{Y}^o=\mbra{y_{jt}^o}\in\mathbb{C}^{N\times L}$, on the index set $\m{\Omega}\times\mbra{L}$, where $N$ is the number of equispaced samples per sinusoidal signal, $\m{\Omega} \subset \m{J} =\lbra{0,1,\dots,N-1}$, and $\mbra{L}=\lbra{1,2,\dots,L}$. That is, we have $L$ measurement vectors corresponding to the $L$ columns of $\m{Y}^o$.
Here $(j,t)$ indexes the entries of $\m{Y}^o$, $i=\sqrt{-1}$, $f_k\in\mathbb{T}\triangleq\left[0,1\right]$ denotes the $k$th normalized frequency (the starting point 0 and the ending point 1 are identical), $s_{kt}\in\mathbb{C}$ is the (complex) amplitude of the $k$th frequency component composing the $t$th sinusoidal signal, and $K$ is the number of the components which is unknown but typically small. Moreover, we let $M=\abs{\m{\Omega}}\leq N$ be the sample size of each measurement vector. Following from the literature of spectral analysis \cite{stoica2005spectral,wang2006spectral}, the observed data $\m{Y}_{\m{\Omega}}^o\triangleq\lbra{y_{jt}^o}_{\sbra{j,t}\in\m{\Omega}\times \mbra{L}}$ are called complete if $M=N$ (i.e., $\m{\Omega} = \m{J}$ or $\m{Y}_{\m{\Omega}}^o=\m{Y}^o$) and incomplete otherwise. In the latter case, the unobserved data on the complementary index set $\overline{\m{\Omega}}\times \mbra{L}$, $\overline{\m{\Omega}}=\m{J}\backslash \m{\Omega}$, are called missing data. Let $\mathcal{T}=\lbra{f_1,\dots,f_K}$ denote the set of the frequencies. The problem concerned in this paper is to recover $\mathcal{T}$ given the observed data, which is referred to as the problem of joint sparse frequency recovery in the sense that the multiple measurement vectors (MMVs) (i.e., the $L$ columns of $\m{Y}_{\m{\Omega}}^o$) share the same $K$ frequencies. After $\mathcal{T}$ is obtained, the amplitudes $\lbra{s_{kt}}$ and the missing data can be easily obtained by a simple least-squares method.
In the single measurement vector (SMV) case where $L=1$, this frequency recovery problem is known as line spectral estimation and has wide applications in communications, radar, sonar, seismology, astronomy and so on \cite{stoica2005spectral,wang2006spectral}. With complete data, Prony's method can recover the frequencies once $K\leq\frac{N}{2}$ by root-finding regardless of the values of the frequencies (see, e.g., \cite{blu2008sparse}). However, it does not work in the presence of missing data and is sensitive to measurement noise. With the development of sparse signal representation and later the compressed sensing (CS) concept \cite{candes2006robust, donoho2006compressed}, which studies the recovery of a sparse signal from a number of linear measurements much less than its ambient dimension, sparse methods for frequency recovery have been popular in the past decade. But unfortunately, the frequencies of interest are usually assumed to lie on a fixed grid of the frequency domain $\mathbb{T}$ because the development of CS so far has been focused on signals that can be sparsely represented under a finite discrete dictionary. Under the aforementioned assumption the observation model (\ref{formu:observmodel1}) can be written into an underdetermined system of linear equations and then sparse methods are applied to solve the sparse signal involved whose support corresponds to the frequency set $\mathcal{T}$. Typical sparse methods include combinatorial optimization or $\ell_0$ (pseudo-)norm minimization, its convex relaxation or $\ell_1$ norm minimization, and greed methods such as orthogonal matching pursuit (OMP) \cite{donoho2003optimally,tropp2007signal}. The $\ell_0$ minimization has the best theoretical guarantee, typically ensuring exact recovery once $K\leq\frac{M}{2}$, however, it is NP-hard and cannot be practically solved. In conventional wisdom, the maximal $K$ allowed in the $\ell_1$ minimization and OMP for guaranteed exact recovery is inversely proportional to a metric called coherence which, however, increases dramatically as the grid gets fine. On the other hand, due to the grid selection, basis mismatches become a major problem of CS-based methods and many modifications have been proposed to alleviate this drawback (see, e.g., \cite{hu2012compressed,yang2012robustly, yang2013off,austin2013dynamic}). A breakthrough came up recently. Cand\`{e}s and Fernandez-Granda \cite{candes2013towards} deal directly with the continuous frequency recovery problem and therefore completely eliminate the basis mismatches. In particular, they consider the complete data case and show that the frequencies can be exactly recovered via convex optimization once any two frequencies are separate by at least $\frac{4}{N}$. That means, up to $K=\frac{N}{4}$ frequencies can be recovered within a polynomial time under the frequency separation condition. The convex optimization is based on the so-called total variation norm or atomic norm which extends the $\ell_1$ norm to the continuous frequency setting and is formulated as semidefinite programming (SDP) \cite{aleksanyan1944real,chandrasekaran2012convex}. Inspired by \cite{candes2013towards}, Tang {\em et al.} \cite{tang2012compressed} study the problem of continuous frequency recovery from partial observations (i.e., incomplete data) based on the atomic norm minimization. Under the same frequency separation condition, they show that a number of $M\geq O\sbra{K\log K\log N}$ randomly located measurements is sufficient to guarantee exact recovery with high probability. Several subsequent works on this topic have been done and an incomplete list includes \cite{bhaskar2013atomic,candes2013super,chen2013robust,azais2013spike,tang2013near,yang2014gridless}.
In the MMV case, an example at hand is direction of arrival (DOA) estimation in array processing \cite{krim1996two,stoica2005spectral}. In particular, suppose that $K$ farfield, narrowband sources impinge on an array of sensors and one wants to know their directions. The output of the sensor array can be modeled by (\ref{formu:observmodel1}), where each frequency corresponds to one source direction. The sampling index set $\m{\Omega}$ therein represents the geometry of the sensor array. To be specific, $\m{\Omega}=\m{J}$ in the complete data case corresponds to an $N$-element uniform linear array (ULA) while $\m{\Omega}\subsetneq\m{J}$ denotes a sparse linear array (SLA). Each measurement vector is composed of the output of the sensor array at one snapshot, and the MMVs are obtained by taking measurements at multiple snapshots, where the source directions are assumed constant during the time window. MUSIC \cite{schmidt1981signal} is prominent for this joint sparse frequency recovery problem, however, it is sensitive to correlations of the sources and requires a sufficient number of snapshots such that the sample variance can capture the whole signal subspace. In lieu of sparse methods in the SMV case, joint sparse recovery techniques have been widely studied in the MMV case which, besides the sparse property, exploit the prior knowledge that the MMVs share the same sparsity profile, known as joint sparsity \cite{malioutov2005sparse,cotter2005sparse,chen2006theoretical, fornasier2008recovery,mishali2008reduce, gribonval2008atoms,kowalski2009sparse, ji2009multi,eldar2009robust, eldar2010average, hyder2010direction, van2010theoretical, kim2012compressive,lee2012subspace,davies2012rank}. It has been vastly reported that the performance of joint sparse recovery can be generally improved by increasing the number of measurement vectors. Theoretical results include \cite{chen2006theoretical} on the $\ell_{0}$ norm under a mild condition, and \cite{gribonval2008atoms} on greed methods and \cite{eldar2010average} on the $\ell_1$ norm under the assumption that the joint sparse signals are randomly drawn, in particular, the rows of the source signals $\mbra{s_{kt}}\in\mathbb{C}^{K\times L}$ in (\ref{formu:observmodel1}) are at general positions. However, it is worth noting that the theoretical guarantee of any joint sparse recovery technique cannot be improved without additional assumptions of the sparse signals of interest. To see this, suppose that all the columns of $\mbra{s_{kt}}$ are identical up to some scale factors (any two sources/rows of $\mbra{s_{kt}}$ are identical up to a scale factor as well and usually said to be coherent in array processing) and therefore, the MMVs are simply scaled replica of a SMV and do not provide additional information for the recovery. In this respect, the results of \cite{gribonval2008atoms} and \cite{eldar2010average} are referred to as the \emph{average case} analysis while those accounting for the aforementioned extreme case are called \emph{worst case} analysis. Similarly to the SMV case, the joint sparse recovery methods rely on discretization/gridding of the frequency domain and suffer from basis mismatches. Unlike the SMV case in which the continuous frequency recovery methods have been recently studied, results are rare on the joint sparse frequency recovery concerned in this paper. To the best of our knowledge, the only known discretization-free/gridless method is introduced in our previous work \cite{yang2014discretization} based on a statistical perspective and utilizing a weighted covariance fitting (WCF) criterion in \cite{stoica2011spice}. In the main context of this paper we will show that this WCF technique is related to the MMV atomic norm method to be proposed in this paper.
In this paper, we present optimization methods for joint sparse frequency recovery and theoretically analyze their performances in the noiseless setting. We firstly consider a continuous $\ell_0$ norm formulation, referred to as the atomic $\ell_0$ norm, and present its theoretical guarantees for exact frequency recovery, which extends the conventional discrete problem formulation, concept and result to the continuous setting. We next consider its convex relaxation, referred to as the (MMV) atomic norm, and investigate its theoretical guarantees with complete and incomplete data separately. In particular, given the complete data we prove that the frequencies can be exactly recovered by solving a convex optimization problem once they are separate by at least $\frac{4}{N}$. Under the same frequency separation condition, a number of $O\sbra{K\log K\log N}$ randomly located samples per measurement vector is sufficient to guarantee that the frequencies and the missing data can be exactly recovered with high probability by solving an atomic norm minimization problem. As a result, our analysis with the atomic norm extends the results of \cite{candes2013towards,tang2012compressed} to the MMV case. Since no or very mild assumptions are made for the source signals in our analysis, the \emph{worst case} theoretical guarantees above do not improve as the number of measurement vectors increases. Moreover, we formulate the atomic $\ell_0$ norm and the atomic norm problems as rank minimization and SDP, respectively. Extensive numerical simulations are carried out which validate our theoretical results and further demonstrate that the frequency separation condition required for exact recovery can be relaxed in general as the number of measurement vectors increases but cannot in the worst case. Our results provide theoretical guidance for the practical array processing applications and will inspire further studies such as the average case analysis.
Notations used in this paper are as follows. $\mathbb{R}$ and $\mathbb{C}$ denote the sets of real and complex numbers respectively. $\mathbb{T}$ denotes the unit circle $\left[0,1\right]$ by identifying the beginning and ending points. Boldface letters are reserved for vectors and matrices. For an integer $N$, $[N]\triangleq\lbra{1,\cdots,N}$. $\abs{\cdot}$ denotes the amplitude of a scalar or cardinality of a set. $\onen{\cdot}$, $\twon{\cdot}$ and $\frobn{\cdot}$ denote the
$\ell_1$, $\ell_2$ and Frobenius norms respectively. $\m{A}^T$ and $\m{A}^H$ are the matrix transpose and conjugate transpose of $\m{A}$ respectively. $x_j$ is the $j$th entry of a vector $\m{x}$, and $\m{A}_j$ denotes the $j$th row of a matrix $\m{A}$. Unless otherwise stated, $\m{x}_{\m{\Omega}}$ and $\m{A}_{\m{\Omega}}$ respectively reserve the entries of $\m{x}$ and the rows of $\m{A}$ in the index set $\m{\Omega}$. For a vector $\m{x}$, $\diag\sbra{\m{x}}$ is a diagonal matrix with $\m{x}$ being its diagonal. $\m{x}\succeq\m{0}$ means $x_j\geq0$ for all $j$. $\rank\sbra{\m{A}}$ denotes the rank of a matrix $\m{A}$ and $\tr\sbra{\m{A}}$ the trace. For positive semidefinite matrices $\m{A}$ and $\m{B}$, $\m{A}\geq\m{B}$ means that $\m{A}-\m{B}$ is positive semidefinite. $\mathbb{E}\mbra{\cdot}$ denotes the expectation and $\mathbb{P}\sbra{\cdot}$ the probability of an event. $\widehat{f}$ is an estimator of $f$. For notational simplicity, a random variable and its numerical value will not be distinguished.
The rest of the paper is organized as follows. Section \ref{sec:atomell0min} presents the atomic $\ell_0$ norm method and studies its theoretical guarantees. Section \ref{sec:convexrelaxation} turns to the convex relaxation method and its theoretical results. Section \ref{sec:connection} discusses connections of our methods to prior arts. Sections \ref{sec:proofcompletedata} and \ref{sec:proofincompletedata} present proofs of two main theorems in Section \ref{sec:convexrelaxation}. Section \ref{sec:simulation} provides numerical simulations and finally Section \ref{sec:conclusion} concludes this paper.
\section{Frequency Recovery via Nonconvex Optimization} \label{sec:atomell0min}
\subsection{Atomic $\ell_0$ Norm and Its Use for Frequency Recovery}
We exploit the sparsity to solve the problem of joint sparse frequency recovery. In particular, we seek a set of frequencies of the minimum length among the infinitely many candidates which can express the observed data $\m{Y}_{\m{\Omega}}^o$. To state it formally, we denote $\m{a}\sbra{f}=\mbra{1,e^{i2\pi f},\cdots,e^{i2\pi\sbra{N-1}f}}^T\in\mathbb{C}^N$ and $\m{s}_{k}=\mbra{s_{k1},\cdots,s_{kL}}\in\mathbb{C}^{1\times L}$. Then (\ref{formu:observmodel1}) can be written as
\equ{\m{Y}^o=\sum_{k=1}^K \m{a}\sbra{f_k}\m{s}_{k}=\sum_{k=1}^K c_k\m{a}\sbra{f_k}\m{\phi}_{k}=\sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_{k}}, \label{formu:observmodel}}
where $c_k=\twon{\m{s}_{k}}>0$ and $\m{\phi}_{k}=c_k^{-1}\m{s}_{k}$ with $\twon{\m{\phi}_{k}}=1$. Let $\mathbb{S}^{2L-1}=\lbra{\m{\phi}: \m{\phi}\in\mathbb{C}^{1\times L}, \twon{\m{\phi}}=1}$ denote the unit complex $L-1$-sphere or real $2L-1$-sphere. We define the continuous dictionary or the set of atoms
\equ{\mathcal{A}\triangleq \lbra{\m{a}\sbra{f,\m{\phi}}=\m{a}\sbra{f}\m{\phi}: f\in\mathbb{T}, \m{\phi}\in\mathbb{S}^{2L-1}}.\label{formu:atomset}}
We see by (\ref{formu:observmodel}) that $\m{Y}^o$ is a linear combination of a number of $K$ atoms in $\mathcal{A}$. In particular, we say that a decomposition of $\m{Y}^o$ as in (\ref{formu:observmodel}) is an atomic decomposition of order $K$ if $c_k>0$ and the frequencies $f_k$ are distinct. For $\m{Y}\in\mathbb{C}^{M\times L}$, we define the smallest number of atoms that can express it as its atomic $\ell_0$ (pseudo-)norm:
\equ{\norm{\m{Y}}_{\mathcal{A},0}
=\inf\lbra{\widehat K: \m{Y}=\sum_{k=1}^{\widehat K} c_k\m{a}_k, \m{a}_k\in\mathcal{A}, c_k>0}. \label{formu:AL0}}
So, we propose to recover the frequencies by minimizing the atomic $\ell_0$ norm of some $\m{Y}$ which is consistent with the observed data on $\m{\Omega}\times\mbra{L}$, i.e., to solve the following optimization problem:
\equ{\min_{\m{Y}} \norm{\m{Y}}_{\mathcal{A},0}, \text{ subject to } \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}. \label{formu:AL0min}}
For $\m{u}\in\mathbb{C}^N$, denote by $T\sbra{\m{u}}\in\mathbb{C}^{N\times N}$ a (Hermitian) Toeplitz matrix with
\equ{T\sbra{\m{u}}=\begin{bmatrix}u_1 & u_2 & \cdots & u_M\\ {u}_2^H & u_1 & \cdots & u_{M-1}\\ \vdots & \vdots & \ddots & \vdots \\ {u}_M^H & {u}_{M-1}^H & \cdots & u_1\end{bmatrix}, \notag}
where $u_j$ is the $j$th entry of $\m{u}$. We provide a finite dimensional formulation of $\norm{\m{Y}}_{\mathcal{A},0}$ in the following proposition.
\begin{prop} $\norm{\m{Y}}_{\mathcal{A},0}$ defined in (\ref{formu:AL0}) equals the optimal value of the following rank minimization problem:
\equ{\min_{\m{W},\m{u},\m{U}} \rank\sbra{\m{U}}, \text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix} \text{ and } \m{U}\geq\m{0}. \label{formu:AL0_rankmin}}
\label{prop:AL0_rankmin}
\end{prop}
The proof of Proposition \ref{prop:AL0_rankmin} is based on the classical Vandermonde decomposition lemma stated as follows.
\begin{lem}[\cite{grenander1958toeplitz,stoica2005spectral}] Any positive semidefinite Toeplitz matrix $T\sbra{\m{u}}\in\mathbb{C}^{N\times N}$ of rank $r$ has an order-$r$ Vandermonde decomposition:
\equ{T\sbra{\m{u}}=\m{A}\m{P}\m{A}^H, \notag}
where $\m{A}=\mbra{\m{a}\sbra{f_1},\dots,\m{a}\sbra{f_r}}$ and $\m{P}=\diag\sbra{p_1,\dots,p_r}$ with $p_j>0$. Moreover, the decomposition is unique if $r\leq N-1$. $\hfill \vrule height6pt width 6pt depth 0pt$ \label{lem:toeplitz}
\end{lem}
\begin{rem} The Vandermonde decomposition is not unique if $T\sbra{\m{u}}$ has full rank. In fact, we can arbitrarily choose $f_1\in\mathbb{T}$ and let $p_1=\mbra{\m{a}\sbra{f_1}^H T\sbra{\m{u}}^{-1}\m{a}\sbra{f_1}}^{-1}$. It follows that the residue $T\sbra{\m{u}}-p_1\m{a}\sbra{f_1}\m{a}\sbra{f_1}^H$ is of rank $N-1$ and remains a positive semidefinite Toeplitz matrix, which admits a unique Vandermonde decomposition of order $N-1$ and further results in a decomposition of $T\sbra{\m{u}}$ of order $N$. \label{rem:vanddecompfullrank}
\end{rem}
\emph{Proof of Proposition \ref{prop:AL0_rankmin}:} Let $K=\norm{\m{Y}}_{\mathcal{A},0}$ and $K^*=\rank\sbra{\m{U}^*}$, where $\sbra{\m{W}^*,\m{u}^*,\m{U}^*}$ denotes an optimal solution to the rank minimization problem (\ref{formu:AL0_rankmin}). We need to show that $K=K^*$. On one hand, since $\m{U}^*\geq\m{0}$ we have $T\sbra{\m{u}^*}\geq\m{0}$ and $r=\rank\sbra{T\sbra{\m{u}^*}}\leq K^*$. It follows from Lemma \ref{lem:toeplitz} that $T\sbra{\m{u}^*}=\m{A}\m{P}\m{A}^H$, where $\m{A}=\mbra{\m{a}\sbra{f_1},\cdots,\m{a}\sbra{f_{r}}}$, $\m{P}=\diag\sbra{p_1,\dots,p_r}$ and $p_j>0$. Moreover, $\m{Y}$ lies in the range space of $T\sbra{\m{u}^*}$ and therefore, there exists $\m{S}\in\mathbb{C}^{r\times L}$ such that $\m{Y}=\m{A}\m{S}=\sum_{j=1}^r\m{a}\sbra{f_j}\m{S}_{j}$, where $\m{S}_{j}$ denotes the $j$th row of $\m{S}$. By the definition of the atomic $\ell_0$ norm in (\ref{formu:AL0}) we have $K\leq r\leq K^*$.
On the other hand, let $\m{Y}=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}=\m{A}\m{S}$ be an atomic decomposition of $\m{Y}$, where $\m{A}$ is similarly defined as before and $\m{S}=\mbra{c_1\m{\phi}_1^T, \dots, c_K\m{\phi}_K^T}^T$. Let $T\sbra{\m{u}}=\m{A}\m{P}\m{A}^H$ and $\m{W}=\m{S}^H\m{P}^{-1}\m{S}$ for arbitrary $p_j>0$, $j\in\mbra{K}$. Then
\equ{\m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}=\begin{bmatrix}\m{S}^H\m{P}^{-1}\\ \m{A}\end{bmatrix} \m{P} \begin{bmatrix}\m{P}^{-1}\m{S} & \m{A}^H\end{bmatrix}\geq\m{0}. \notag}
As a result, $\sbra{\m{W},\m{u},\m{U}}$ defines a feasible solution of the rank minimization problem (\ref{formu:AL0_rankmin}), and we have that $K^*\leq \rank\sbra{\m{U}}\leq \rank\sbra{\m{P}}=K$.
\hfill\rule{2mm}{2mm}
Proposition \ref{prop:AL0_rankmin} presents a rank minimization problem to characterize the atomic $\ell_0$ norm. It follows that (\ref{formu:AL0min}) can be formulated as follows:
\equ{\begin{split}
&\min_{\m{Y},\m{W},\m{u},\m{U}} \rank\sbra{\m{U}},\\
&\text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}, \m{U}\geq\m{0}, \text{ and } \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}. \end{split} \label{formu:AL0_rankmin1}}
In (\ref{formu:AL0_rankmin1}) we need to complete a (structured positive semidefinite) low rank matrix $\m{U}$ with partial access to its entries. Therefore, we establish a link between the frequency recovery problem and low rank matrix completion \cite{candes2009exact,recht2011simpler}. Note that a similar rank minimization problem is presented in \cite{tang2012compressed} in the SMV case, where the rank is put on the matrix $T\sbra{\m{u}}$ rather than the full matrix $\m{U}$. Later we will see that (\ref{formu:AL0_rankmin1}) has a clearer connection to the convex relaxation method presented in Section \ref{sec:convexrelaxation}.
Suppose we could solve the rank minimization problem in (\ref{formu:AL0_rankmin1}). According to the proof of Proposition \ref{prop:AL0_rankmin}, we can obtain the solution of the frequencies by the Vandermonde decomposition of $T\sbra{\m{u}^*}$, where $\m{u}^*$ is an optimal solution. While Lemma \ref{lem:toeplitz} states the existence of the decomposition, readers are referred to \cite[Appendix A]{yang2014gridless} for a detailed approach to its realization. We next investigate theoretical guarantees of the proposed atomic $\ell_0$ norm method.
\subsection{Spark of Continuous Dictionary} \label{sec:spark}
To analyze the atomic $\ell_0$ norm minimization problem (\ref{formu:AL0min}) or equivalently (\ref{formu:AL0_rankmin1}), we generalize the concept of spark in \cite{kruskal1977three} to the case of continuous dictionary. We define the following continuous dictionary with respect to the index set $\m{\Omega}$ similarly to (\ref{formu:atomset}):
\equ{\begin{split}\mathcal{A}_{\m{\Omega}}^1
&\triangleq \lbra{\m{a}_{\m{\Omega}}\sbra{f}: f\in\mathbb{T}}.\end{split}}
\begin{defi}[Spark of continuous dictionary] Given the continuous dictionary $\mathcal{A}_{\m{\Omega}}^1$, the quantity spark of $\mathcal{A}_{\m{\Omega}}^1$, denoted by $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$, is the smallest number of atoms of $\mathcal{A}_{\m{\Omega}}^1$ which are linearly dependent.
\end{defi}
In general, it is NP-hard to compute $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ given some sampling index set $\m{\Omega}$. Some preliminary results are presented in the following proposition.
\begin{prop} The following results hold about $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ with $\abs{\m{\Omega}}=M$:
\begin{enumerate}
\item $2\leq\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}\leq M+1$,
\item $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= 2$ if and only if the elements of
\equ{\m{\mathcal{D}}\triangleq\lbra{m_1-m_2: m_1,m_2\in\m{\Omega},m_1\geq m_2}}
is not coprime, and
\item $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= M+1$ if $\m{\Omega}$ consists of $M$ consecutive integers.
\end{enumerate}
\label{prop:fullspark}
\end{prop}
\begin{proof}
1) $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}\geq 2$ since all atoms of $\mathcal{A}_{\m{\Omega}}^1$ are nonzero. $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}\leq M+1$ because $\abs{\m{\Omega}}=M$ and any $M+1$ atoms of $\mathcal{A}_{\m{\Omega}}^1$ are linearly dependent.
2) $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}=2$ if and only if there exist $f_1\neq f_2$ such that $\m{a}_{\m{\Omega}}\sbra{f_1}$ and $\m{a}_{\m{\Omega}}\sbra{f_2}$ are linearly dependent. We next prove the equivalence between the linear dependence and the non-coprimality.
\emph{Sufficiency:} If $\m{a}_{\m{\Omega}}\sbra{f_1}$ and $\m{a}_{\m{\Omega}}\sbra{f_2}$ are linearly dependent, i.e., $\m{a}_{\m{\Omega}}\sbra{f_1}=ce^{i\theta}\m{a}_{\m{\Omega}}\sbra{f_2}$ with $c>0$ and $\theta\in\mathbb{R}$, then it holds for any $m\in\m{\Omega}$ that
\equ{2\pi m\sbra{f_1-f_2}\equiv\theta \mod 2\pi \notag}
and thus $\sbra{m_1-m_2}\sbra{f_1-f_2}$ is an integer for any $m_1,m_2\in\m{\Omega}$, i.e., $d\sbra{f_1-f_2}$ is an integer for any $d\in\m{\mathcal{D}}$. It follows that $f_1-f_2$ is a rational number. Let $\abs{f_1-f_2}=\frac{b_1}{b_2}<1$, where $b_1,b_2$ are coprime positive integers with $b_2\geq2$. Consequently, $b_2$ is a common divisor of the elements of $\m{\mathcal{D}}$.
\emph{Necessity:} Suppose integer $b\geq2$ is a common divisor of the elements of $\m{\mathcal{D}}$. Consider any two $f_1,f_2\in\mathbb{T}$ satisfying that $f_1=f_2+\frac{1}{b}$. Then, for any $m\in\m{\Omega}$,
\equ{e^{i2\pi mf_1}= e^{i2\pi \frac{m}{b}} e^{i2\pi m f_2}=e^{i2\pi\frac{\Omega_1}{b}} e^{i2\pi m f_2}. \notag}
The last equality holds since $b$ evenly divides $m-\Omega_1\in\m{\mathcal{D}}$. It follows that $\m{a}_{\m{\Omega}}\sbra{f_1}$ and $\m{a}_{\m{\Omega}}\sbra{f_2}$ are linearly dependent.
3) If $\m{\Omega}$ consists of $M$ consecutive integers, say, $m, m+1, \cdots, m+M-1$ with $m\geq0$, then $\m{a}_{\m{\Omega}}\sbra{f_j}$, $j\in\mbra{M}$, are linearly independent for any $M$ distinct $f_j\in\mathbb{T}$ since the determinant
\equ{\begin{split}&\abs{\mbra{\m{a}_{\m{\Omega}}\sbra{f_1}, \dots , \m{a}_{\m{\Omega}}\sbra{f_M} }} \\
&= e^{i2\pi m\sum_{j=1}^Mf_j} \prod_{1\leq j\leq l\leq M}\sbra{e^{i2\pi f_j}-e^{i2\pi f_l}}\\
&\neq0.\end{split}}
As a result, any $M$ atoms of $\mathcal{A}_{\m{\Omega}}^1$ are linearly independent, which together with $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}\leq M+1$ concludes that $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= M+1$.
\end{proof}
Proposition \ref{prop:fullspark} presents the range of $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ with respect to the sampling index set $\m{\Omega}$. A necessary and sufficient condition is provided under which $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ achieves the lower bound 2, where there exist two atoms that are linearly dependent. In this case, in fact, for any atom in $\mathcal{A}_{\m{\Omega}}^1$ we can always find one such that they are linearly dependent following from the proof above. However, such $\m{\Omega}$ is rare. To see this, suppose that $\m{\Omega}$ is chosen uniformly at random. Then it satisfies the condition with probability no greater than
\equ{\frac{\sum_{p\leq\lfloor\frac{N}{M}\rfloor \text{ is prime}} p {\lceil\frac{N}{p}\rceil \choose M}}{{N\choose M}},}
where ${N\choose M}$ denotes the number of $M$-combinations given an $N$-element set. It is clear that the probability equals 0 when $M>\frac{N}{2}$. Numerically, the probability is less than $1.2\times10^{-3},\, 1.8\times10^{-7},\, 3.2\times10^{-12}$ when $N=100$ and $M=10,\,20,\,30$ respectively. A sufficient (but unnecessary) condition is also provided in the third part under which $\mathcal{A}_{\m{\Omega}}^1$ achieves the upper bound $M+1$. We say that $\mathcal{A}_{\m{\Omega}}^1$ is a full spark dictionary if $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}= M+1$ following from the discrete setting in \cite{alexeev2012full}. We will see the benefits of a full spark dictionary for frequency recovery in the ensuing subsection.
\subsection{Theoretical Guarantees of the Atomic $\ell_0$ Norm} \label{sec:l0exactrecovery}
We provide theoretical guarantees of the atomic $\ell_0$ norm minimization in (\ref{formu:AL0min}) or the rank minimization in (\ref{formu:AL0_rankmin1}) for frequency recovery in this subsection. In particular, we have the following result, which can be considered as a continuous version of \cite[Theorem 2.4]{chen2006theoretical}.
\begin{thm} $\m{Y}^o=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}$ is the unique optimizer to (\ref{formu:AL0min}) or (\ref{formu:AL0_rankmin1}) if
\equ{K< \frac{\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1+\rank \sbra{\m{Y}_{\m{\Omega}}^o}}{2}. \label{Kbound}}
Moreover, the atomic decomposition above is the unique one satisfying that $K=\norm{\m{Y}^o}_{\mathcal{A},0}$.
\label{thm:AL0_guanrantee}
\end{thm}
\begin{proof} Suppose that there exists $\widetilde{\m{Y}}\neq \m{Y}^o$ satisfying that $\widetilde{\m{Y}}_{\m{\Omega}}= \m{Y}^o_{\m{\Omega}}$ and $\norm{\widetilde{\m{Y}}}_{\mathcal{A},0}\leq \norm{\m{Y}^o}_{\mathcal{A},0}= K$, and $\widetilde{\m{Y}}= \sum_{k=1}^{\widetilde K} \widetilde c_j\m{a}\sbra{\widetilde f_j,\widetilde{\m{\phi}}_j}$ is an atomic decomposition of order $\widetilde K\leq K$. Let $\m{A}_1=\mbra{\m{a}\sbra{f}}_{f\in\mathcal{T}\backslash{\lbra{\widetilde f_j}}}$ (the matrix consisting of those $\m{a}\sbra{f}$, $f\in\mathcal{T}\backslash{\lbra{\widetilde f_j}}$), $\m{A}_{12}=\mbra{\m{a}\sbra{f}}_{f\in\mathcal{T}\cap{\lbra{\widetilde f_j}}}$ and $\m{A}_2=\mbra{\m{a}\sbra{f}}_{f\in{\lbra{\widetilde f_j}}\backslash\mathcal{T}}$. In addition, let $K_{12}=\abs{\mathcal{T}\cap{\lbra{\widetilde f_j}}}$ and $\m{A}=\begin{bmatrix}\m{A}_{1} & \m{A}_{12} & \m{A}_2\end{bmatrix}$. Then we have
{\setlength\arraycolsep{2pt}\equa{\m{Y}^o
&=& \begin{bmatrix}\m{A}_1 & \m{A}_{12}\end{bmatrix} \begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}\end{bmatrix}, \notag \\ \widetilde{\m{Y}}
&=& \begin{bmatrix}\m{A}_{12} & \m{A}_{2}\end{bmatrix} \begin{bmatrix}\m{S}_{21} \\ \m{S}_{2}\end{bmatrix}, \notag
}}where $\m{S}_1$, $\m{S}_{12}$, $\m{S}_{21}$ and $\m{S}_{2}$ are matrices of proper dimensions. It follows that
\equ{\m{Y}^o-\widetilde{\m{Y}}= \m{A} \begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}-\m{S}_{21}\\ -\m{S}_{2}\end{bmatrix}\neq\m{0}, \notag}
and thus
\equ{\begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}-\m{S}_{21}\\ -\m{S}_{2}\end{bmatrix}\neq\m{0}. \label{formu:diffnot0}}
On the other hand, it follows from $\widetilde{\m{Y}}_{\m{\Omega}}= \m{Y}^o_{\m{\Omega}}$ that
\equ{\m{A}_{\m{\Omega}} \begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}-\m{S}_{21}\\ -\m{S}_{2}\end{bmatrix}=\m{0}. \notag}
Note that $\m{A}_{\m{\Omega}}$ is composed of atoms in $\mathcal{A}_{\m{\Omega}}^1$, and has a nontrivial null space by (\ref{formu:diffnot0}). Then by the definition of spark we have
\equ{\rank\sbra{\m{A}_{\m{\Omega}}}\geq \spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1. \label{formu:rankbound}}
Moreover, for the nullity (dimension of the null space) of $\m{A}_{\m{\Omega}}$ it holds that
\equ{\begin{split}
\text{nullity}\sbra{\m{A}_{\m{\Omega}}}
&\geq \rank\sbra{\begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}-\m{S}_{21}\\ -\m{S}_{2}\end{bmatrix}}\\
&\geq \rank\sbra{\begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}-\m{S}_{21}\end{bmatrix}}\\
&\geq \rank\sbra{\begin{bmatrix}\m{S}_{1} \\ \m{S}_{12}\end{bmatrix}} - \rank\sbra{\begin{bmatrix}\m{0} \\ \m{S}_{21}\end{bmatrix}}\\
&\geq \rank\sbra{\m{Y}_{\m{\Omega}}^o} - K_{12}.\end{split} \label{formu:nullitybound}}
Consequently, the equality
\equ{\#\text{columns of }\m{A}_{\m{\Omega}}=\rank \sbra{\m{A}_{\m{\Omega}}}+ \text{nullity}\sbra{\m{A}_{\m{\Omega}}} \notag}
together with (\ref{formu:rankbound}) and (\ref{formu:nullitybound}) yields that
$K+\widetilde K-K_{12}\geq \spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1+ \rank\sbra{\m{Y}_{\m{\Omega}}^o} - K_{12}$,
and then
\equ{2K\geq K+\widetilde K\geq \spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1+\rank \sbra{\m{Y}_{\m{\Omega}}^o}, \notag}
which contradicts the condition in (\ref{Kbound}).
To show the uniqueness of the atomic decomposition, we observe that the condition in (\ref{Kbound}) implies that $K< \spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-1$ since $\rank \sbra{\m{Y}_{\m{\Omega}}^o}\leq K$. According to the definition of spark, any $K$ atoms in $\mathcal{A}_{\m{\Omega}}^1$ are linearly independent. Therefore, the atomic decomposition is unique given the set of frequencies $\mathcal{T}=\lbra{f_j}_{j=1}^K$. Now suppose there exists another decomposition $\m{Y}^o=\sum_{j=1}^{\widetilde K} \widetilde c_j\m{a}\sbra{\widetilde f_j,\widetilde{\m{\phi}}_j}$ with $\widetilde K\leq K$ and a different set of frequencies $\lbra{\widetilde f_j}$ (i.e., there exists some $\widetilde f_j\notin\mathcal{T}$). Note that we have repetitively used the same notations for notational simplicity and we similarly define the other notations. Once again we have (\ref{formu:diffnot0}) since $\m{A}_2$ is nonempty followed by $\m{S}_2\neq\m{0}$. The rest of the proof follows from the same arguments.
\end{proof}
It is generally difficult to compute $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ given $\m{\Omega}$, which makes the condition (\ref{Kbound}) hard to check in a particular scenario. However, it is checkable in the special complete data case, which is shown in the following corollary. Note that the rank minimization problem (\ref{formu:AL0_rankmin1}) can still help to recover the frequencies though (\ref{formu:AL0min}) admits a trivial solution.
\begin{cor}[Complete data] $\m{Y}^o=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}$ is the unique atomic decomposition satisfying that $K=\norm{\m{Y}^o}_{\mathcal{A},0}$ if
\equ{K< \frac{N+\rank \sbra{\m{Y}^o}}{2}. \label{Kbound_comp}}
\label{cor:AL0_guanrantee}
\end{cor}
\begin{proof} The result follows from $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}=N+1$ given $\m{\Omega}=\m{J}$, which holds according to the third part of Proposition \ref{prop:fullspark}.
\end{proof}
Theorem \ref{thm:AL0_guanrantee} shows that the frequencies can be exactly recovered by minimizing the atomic $\ell_0$ norm or equivalently solving the rank minimization problem (\ref{formu:AL0_rankmin1}) provided that the sparsity level $K$ is sufficiently small. In particular, the upper bound of $K$ depends on a particular sampling index set $\m{\Omega}$ and the observed data $\m{Y}_{\m{\Omega}}^o$. In the SMV case where $\rank \sbra{\m{Y}_{\m{\Omega}}^o}=1$ and therefore, the condition becomes $K< \frac{1}{2}{\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}}$ (or $K\leq \frac{N}{2}$ in the complete data case). As we take more measurement vectors, we have a chance to recover more complex signals by potentially increasing $\rank \sbra{\m{Y}_{\m{\Omega}}^o}$ (or $\rank \sbra{\m{Y}^o}$), which is practically relevant in array processing applications. When $\rank \sbra{\m{Y}_{\m{\Omega}}^o}$ achieves the maximum value $K$, the sparsity level $K$ can be as large as $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}-2$ (or $N-1$ in the complete data case), which is consistent with the result in array processing (see e.g., \cite{yang2014discretization}).
\begin{rem} \label{rem:L0_coherent}
An unpleasant scenario is in the presence of coherent sources. In an extreme case where all the sources are coherent, i.e., all the rows of $\mbra{s_{kt}}$ are identical up to some scale factors, it holds that $\rank\sbra{\m{Y}_{\m{\Omega}}^o}=1$ as in the SMV case and taking more measurement vectors does not improve the number of frequencies recoverable. In fact, we can easily verify that all the measurement vectors are identical up to scale factors as well and thus provide the same amount of information for frequency recovery as a SMV. In this case, it is easy to verify that the optimal solution of $\m{u}$ to (\ref{formu:AL0_rankmin1}) remains constant when we add or delete columns of $\m{Y}$.
\end{rem}
\begin{rem}
Since the upper bound increases with $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$, an interesting topic in the future is to study selection of the sampling index set $\m{\Omega}$, which in array processing corresponds to design of the geometry of the SLA, such that the spark is maximized. While preliminary results have been obtained in the discrete frequency setting (the frequencies are fixed on a uniform grid) in \cite{alexeev2012full}, it is worth noting that the continuous setting concerned here is more practical and challenging.
\end{rem}
\section{Frequency Recovery via Convex Relaxation} \label{sec:convexrelaxation}
\subsection{Convex Relaxation and Semidefinite Formulation} \label{sec:convexrelax}
The atomic $\ell_0$ norm exploits sparsity directly, however, it is nonconvex and the formulated rank minimization problem cannot be globally solved with a practical algorithm. To avoid the nonconvexity, we utilize convex relaxation to relax the atomic $\ell_0$ norm. In particular, it can be relaxed in two ways from two different perspectives. One is to relax the atomic $\ell_0$ norm to the atomic $\ell_1$ norm, or simply the atomic norm, which is defined as the gauge function of $\text{conv}\sbra{\mathcal{A}}$, the convex hull of $\mathcal{A}$ \cite{chandrasekaran2012convex}:
\equ{\begin{split}\atomn{\m{Y}}
&\triangleq\inf\lbra{t>0: \m{Y}\in t\text{conv}\sbra{\mathcal{A}}} \\
&= \inf\lbra{\sum_k c_k: \m{Y}=\sum_k c_k\m{a}_k, c_k>0, \m{a}_k\in\mathcal{A}}.\end{split} \label{formu:atomicnorm}}
The atomic norm $\atomn{\cdot}$ induced by $\mathcal{A}$ defined in (\ref{formu:atomset}) is a norm and thus convex by the property of the gauge function. The concept of atomic norm is firstly introduced in \cite{chandrasekaran2012convex} and the authors argue that the atomic norm minimization problem is the best convex heuristic for recovering simple models with respect to a given atomic set. The other way of convex relaxation is based on a perspective of rank minimization illustrated by (\ref{formu:AL0_rankmin}) and to relax the pseudo rank norm to the nuclear norm or equivalently the trace norm for a positive semidefinite matrix, i.e., to replace $\rank\sbra{\m{U}}$ by $\tr\sbra{\m{U}}$ in (\ref{formu:AL0_rankmin}). It is worthy noting that the nuclear norm has been extensively studied for low rank matrix completion and recovery \cite{recht2007guaranteed,candes2009exact,recht2011simpler}. Interestingly enough, the two ways of convex relaxation are equivalent, which is shown in the following result.
\begin{thm} $\norm{\m{Y}}_{\mathcal{A}}$ defined in (\ref{formu:atomicnorm}) equals the optimal value of the following SDP:
\equ{\min_{\m{W},\m{u},\m{U}} \frac{1}{2\sqrt{N}}\tr\sbra{\m{U}}, \text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix} \text{ and }\m{U}\geq\m{0}. \label{formu:AN_SDP}} \label{thm:AN_SDP}
\end{thm}
Before presenting the proof of Theorem \ref{thm:AN_SDP}, we clarify some notations first. We use the following identity whenever $\m{R}\geq\m{0}$:
\equ{\m{y}^H\m{R}^{-1}\m{y}=\min t, \text{ subject to } \begin{bmatrix}t & \m{y}^H \\ \m{y} & \m{R}\end{bmatrix}\geq0. \label{formu:yRy}}
It follows that $\m{y}^H\m{R}^{-1}\m{y}$ is finite if and only if $\m{y}$ is in the range space of $\m{R}$. In fact, (\ref{formu:yRy}) is equivalent to defining $\m{y}^H\m{R}^{-1}\m{y}\triangleq\lim_{\sigma\rightarrow0_+} \m{y}^H\sbra{\m{R}+\sigma\m{I}}^{-1}\m{y}$ when $\m{R}$ loses rank. The following result will be used to prove Theorem \ref{thm:AN_SDP}.
\begin{lem} Given $\m{R}=\m{A}\m{A}^H\geq\m{0}$, it holds that $\m{y}^H\m{R}^{-1}\m{y}=\min\twon{\m{s}}^2, \text{ subject to } \m{A}\m{s}=\m{y}$. \label{lem:lem1}
\end{lem}
\begin{proof} We need only to show that for any $\m{s}$ satisfying $\m{A}\m{s}=\m{y}$ it holds that $\m{y}^H\m{R}^{-1}\m{y}\leq\twon{\m{s}}^2$, or equivalently,
$\begin{bmatrix}\twon{\m{s}}^2 & \m{y}^H \\ \m{y} & \m{R}\end{bmatrix}\geq0$.
The conclusion follows from that
\equ{\begin{bmatrix}\twon{\m{s}}^2 & \m{y}^H \\ \m{y} & \m{R}\end{bmatrix} = \begin{bmatrix}\twon{\m{s}}^2 & \m{s}^H\m{A}^H \\ \m{A}\m{s} & \m{A}\m{A}^H\end{bmatrix}=\begin{bmatrix}\m{s}^H\\\m{A}\end{bmatrix}\begin{bmatrix}\m{s}^H\\\m{A}\end{bmatrix}^H\geq0. \notag}
\end{proof}
\emph{Proof of Theorem \ref{thm:AN_SDP}:} Provided $\m{U}\geq\m{0}$, we have equivalently $T\sbra{\m{u}}\geq\m{0}$ and $\m{W}\geq \m{Y}^H\mbra{T\sbra{\m{u}}}^{-1} \m{Y}$. So, we need to show that
\equ{\begin{split}\norm{\m{Y}}_{\mathcal{A}}=
&\min_{\m{u}} \frac{\sqrt{N}}{2}u_1 + \frac{1}{2\sqrt{N}}\tr\sbra{\m{Y}^H\mbra{T\sbra{\m{u}}}^{-1} \m{Y}},\text{ subject to } T\sbra{\m{u}}\geq\m{0}, \end{split} \label{formu:AN_sdp222}}
where $u_1$ is the first entry of $\m{u}$ as before.
By the Vandermonde decomposition we have $T\sbra{\m{u}}=\m{A}\m{P}\m{A}^H=\mbra{\m{A}\m{P}^{\frac{1}{2}}} \mbra{\m{A}\m{P}^{\frac{1}{2}}}^H$, where $\m{A}=\m{A}\sbra{\m{f}}$ is composed of columns $\m{a}\sbra{f_j}$, and $\m{P}=\diag\sbra{\dots,p_j,\dots}$ with $p_j>0$. Moreover, we can verify that $u_1=\sum p_j$. For the $t$th column of $\m{Y}$, say $\m{y}_{: t}$, it holds by Lemma \ref{lem:lem1} that
\equ{\begin{split}\m{y}_{: t}^H\mbra{T\sbra{\m{u}}}^{-1}\m{y}_{: t}
&= \min_{\m{v}} \twon{\m{v}}^2, \text{ subject to } \m{A}\m{P}^{\frac{1}{2}}\m{v}=\m{y}_{: t}\\
&= \min_{\m{s}} \twon{\m{P}^{-\frac{1}{2}}\m{s}}^2, \text{ subject to } \m{A}\m{s}=\m{y}_{: t}\\
&= \min_{\m{s}} \m{s}^H\m{P}^{-1}\m{s}, \text{ subject to } \m{A}\m{s}=\m{y}_{: t}.
\end{split} \notag}
Therefore,
\equ{\begin{split}
\tr\sbra{\m{Y}^H\mbra{T\sbra{\m{u}}}^{-1}\m{Y}}
&= \sum_{t=1}^N \m{y}_{: t}^H\mbra{T\sbra{\m{u}}}^{-1}\m{y}_{: t} \\
&= \min_{\m{S},\m{A}\sbra{\m{f}}\m{S}=\m{Y}} \tr\sbra{\m{S}^H\m{P}^{-1}\m{S}}.\\
\end{split} \notag}
We complete the proof via the following equalities:
\equ{\begin{split}
& \min_{\m{u}} \frac{\sqrt{N}}{2}u_1 + \frac{1}{2\sqrt{N}}\tr\sbra{\m{Y}^H\mbra{T\sbra{\m{u}}}^{-1} \m{Y}}\\
=& \min_{\substack {\m{f},\m{p}\succeq\m{0}, \m{S}\\ \m{A}\sbra{\m{f}}\m{S}=\m{Y}}}\frac{\sqrt{N}}{2}\sum_j p_j + \frac{1}{2\sqrt{N}}\tr\sbra{\m{S}^H\m{P}^{-1}\m{S}}\\
=& \min_{\substack {\m{f},\m{p}\succeq\m{0}, \m{S}\\ \m{A}\sbra{\m{f}}\m{S}=\m{Y}}} \frac{\sqrt{N}}{2}\sum_j p_j + \frac{1}{2\sqrt{N}}\sum_j \twon{\m{S}_j}^2p_j^{-1}\\
=& \min_{\m{f},\m{S}} \sum_j\twon{\m{S}_j}, \text{ subject to } \m{Y}=\m{A}\sbra{\m{f}}\m{S} \\
=& \min_{\m{f},\m{c}\succeq\m{0}} \sum_j c_j, \text{ subject to } \m{Y}=\sum_j c_j\m{a}\sbra{f_j,\m{\phi}_j}\\
=& \norm{\m{Y}}_{\mathcal{A}}.
\end{split} \label{formu:AN_equa}}
The second last equality holds by the substitutions $c_j=\twon{\m{S}_j}$ and $\m{\phi}_j=c_j^{-1}\m{S}_j$, followed by $\m{Y}=\m{A}\sbra{\m{f}}\m{S}=\sum_j\m{a}\sbra{f_j}\m{S}_j=\sum_jc_j\m{a}\sbra{f_j, \m{\phi}_j}$, where $\m{S}_j$ denotes the $j$th row of $\m{S}$. The last equality follows from the definition of the atomic norm.
\hfill\rule{2mm}{2mm}
\begin{rem} The semidefinite formulation (\ref{formu:AN_SDP}) of the atomic norm was firstly reported in our conference paper \cite{yang2014continuous}. When preparing this paper, we found that the same result was also independently obtained in \cite{chi2013joint}.
\end{rem}
\begin{rem} Other optimization methods can be studied in the future for joint sparse frequency recovery by borrowing ideas in low rank matrix recovery (see, e.g., \cite{mohan2012iterative,malek2014recovery}). As an example, the (nonconvex) log-determinant criterion $\log\abs{\m{U}+\epsilon \m{I}}$, where $\epsilon>0$ is a small number, has been studied in \cite{yang2014continuous} and implemented by a series of SDPs via a majorization-maximization (MM) process.
\end{rem}
Theorem \ref{thm:AN_SDP} shows that the atomic norm can be formulated as an SDP and thus can be globally solved using standard SDP solvers such as SDPT3 \cite{toh1999sdpt3}. Note that the semidefinite formulation (\ref{formu:AN_SDP}) generalizes the results in the SMV case in \cite{candes2013towards,bhaskar2013atomic,tang2012compressed}. As a result, by the convex relaxation we propose the following optimization problem for frequency recovery:
\equ{\min_{\m{Y}} \norm{\m{Y}}_{\mathcal{A}}, \text{ subject to } \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}, \label{formu:ANmin}}
or equivalently,
\equ{\begin{split}
&\min_{\m{Y},\m{W},\m{u},\m{U}} \tr\sbra{\m{U}},\\
&\text{ subject to } \m{U}=\begin{bmatrix}\m{W} & \m{Y}^H \\ \m{Y} & T\sbra{\m{u}}\end{bmatrix}, \m{U}\geq\m{0}, \text{ and } \m{Y}_{\m{\Omega}}=\m{Y}^o_{\m{\Omega}}. \end{split} \label{formu:AN_sdp}}
For a matrix $\m{V}\in\mathbb{C}^{N\times L}$, the dual norm of the atomic norm is defined as
\equ{\begin{split}\datomn{\m{V}}
&= \sup_{\atomn{\m{a}}\leq1}\inp{\m{V},\m{a}}_{\mathbb{R}}\\
&= \sup_{\m{a}\sbra{f,\m{\phi}}\in\mathcal{A}}\inp{\m{V},\m{a}\sbra{f,\m{\phi}}}_{\mathbb{R}}\\
&= \sup_{f\in\mathbb{T},\m{\phi}\in\mathbb{S}^{2L-1}}\inp{\m{a}(f)^H\m{V},\m{\phi}}_{\mathbb{R}}\\
&= \sup_{f\in\mathbb{T}}\twon{\m{a}(f)^H\m{V}},
\end{split} \label{formu:dualAN}}
where $\inp{\m{A},\m{B}}_{\mathbb{R}}=\Re\tr\sbra{\m{B}^H\m{A}}$ for two matrices (or column/row vectors) $\m{A}$ and $\m{B}$ of proper dimensions and $\Re$ takes the real part of a complex argument.
The dual problem of (\ref{formu:ANmin}) is thus
\equ{\begin{split}
&\max_{\m{V}\in\mathbb{C}^{N\times L}} \inp{\m{V}_{\m{\Omega}},\m{Y}_{\m{\Omega}}^o}_{\mathbb{R}},\\
&\text{ subject to } \datomn{\m{V}}\leq1, \m{V}_{\overline{\m{\Omega}}}=\m{0} \label{formu:dual1} \end{split}}
following from a standard Lagrangian analysis \cite{boyd2004convex}.
On the other hand, the dual problem of (\ref{formu:ANmin}) has the following semidefinite formulation
\equ{\begin{split}
&\max_{\m{V}\in\mathbb{C}^{N\times L},\m{H}\in\mathbb{C}^{N\times N}} \inp{\m{V}_{\m{\Omega}},\m{Y}_{\m{\Omega}}^o}_{\mathbb{R}},\\
&\text{ subject to } \begin{bmatrix}\m{I} & -\m{V}^H \\ -\m{V} & \m{H}\end{bmatrix}\geq\m{0}, \\
&\phantom{\text{ subject to }} \sum_{k=1}^{N-j} H_{k,k+j}=\left\{\begin{array}{ll} 1, & j=0, \\ 0, &j\in\mbra{N-1},\end{array}\right.\\
&\phantom{\text{ subject to }} \m{V}_{\overline{\m{\Omega}}}=\m{0} \end{split} \label{formu:dual}}
based on the semidefinite formulation (\ref{formu:AN_sdp}) and a standard Lagrangian analysis. That means, the first two constraints in (\ref{formu:dual}) can equivalently characterize $\datomn{\m{V}}\leq1$. We empirically find that the dual problem (\ref{formu:dual}) can be solved more efficiently than the primal problem (\ref{formu:AN_sdp}) using, for example, SDPT3, while most solvers can return an optimal solution of (\ref{formu:AN_sdp}) (as the dual of (\ref{formu:dual})) for free when solving (\ref{formu:dual}).
\subsection{Frequency Retrieval} \label{sec:freqretriev}
Given an optimal solution $\sbra{\m{Y}^*,\m{u}^*}$ of (\ref{formu:AN_sdp}), we can retrieve the frequencies of interest and obtain corresponding atomic decomposition(s) of $\m{Y}^*$ as follows. Consider first the case where $T\sbra{\m{u}^*}$ is rank deficient, i.e., $K^*\triangleq\rank\sbra{T\sbra{\m{u}^*}}\leq N-1$. Then we can obtain a unique Vandermonde decomposition $T\sbra{\m{u}^*}=\m{A}\sbra{\m{f}^*}\diag\sbra{\m{p}^*}\m{A}\sbra{\m{f}^*}^H$ of order $K^*$ by Lemma \ref{lem:toeplitz}. Since $\m{Y}^*$ is in the range space of $T\sbra{\m{u}^*}$, there exist unique $\m{S}^*$, $c_j^*$ and $\m{\phi}_j^*$ corresponding to the variables in the proof of Theorem \ref{thm:AN_SDP} such that
\equ{\m{Y}^*=\m{A}\sbra{\m{f}^*}\m{S}^*=\sum_{j=1}^{K^*} c_j^*\m{a}\sbra{f_j^*,\m{\phi}_j^*} \notag}
is an atomic decomposition of $\m{Y}^*$ with
\equ{\atomn{\m{Y}^*}=\sum c_j^* = \sum \twon{\m{S}_j^*}=\sqrt{N}\sum p_j^*=\sqrt{N} u_1^*. \notag}
That means, we have obtained a unique atomic decomposition that achieves the atomic norm. On the other hand, when $T\sbra{\m{u}^*}$ has full rank, there exist infinitely many Vandermonde decompositions of $T\sbra{\m{u}^*}$ of order $N$ according to Remark \ref{rem:vanddecompfullrank}. Consequently, we can obtain infinitely many atomic decompositions of $\m{Y}^*$ of order $N$ with each achieving the atomic norm. Therefore, it is impossible to exactly recover the frequencies.
\subsection{Theoretical Guarantees of Convex Relaxation: Complete Data}
To guarantee exact frequency recovery with the convex relaxation, we require that the frequencies of interest be appropriately separate. To quantify the separation, we define \emph{minimum separation} of a set of frequencies as follows.
\begin{defi}[Minimum separation] For a finite subset $\mathcal{T}\subset\mathbb{T}$, the minimum separation of $\mathcal{T}$ is defined as the closest wrap-around distance between any two elements,
\equ{\Delta_{\mathcal{T}}=\inf_{a,b\in\mathcal{T}:a\neq b}\min\lbra{\abs{a-b}, 1-\abs{a-b}}. \notag}
For example, the distance between $0$ and $\frac{3}{4}$ equals $\frac{1}{4}$.
\end{defi}
Inspired by \cite{candes2013towards} and \cite{tang2012compressed}, we study the cases of complete data and incomplete data separately. Our analysis in the former is deterministic and applicable once $N\geq257$ while that in the latter is based on non-asymptotic analysis of random matrices \cite{vershynin2010introduction}. Since the rows of the source signals $\mbra{s_{kt}}$ are allowed to be correlated or even coherent in our analysis, it is \emph{worst case} analysis as opposed to the \emph{average case} analysis in \cite{gribonval2008atoms,eldar2010average}.
Given the complete data, the proposed atomic norm minimization problem (\ref{formu:ANmin}) has a trivial solution, however, the semidefinite formulation (\ref{formu:AN_sdp}) can be used to recover the frequencies. Its theoretical guarantee is presented in the following result, which generalizes that in the SMV case in \cite{candes2013towards}.
\begin{thm}[Complete data] $\m{Y}^o=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}$ is the unique atomic decomposition satisfying that $\norm{\m{Y}^o}_{\mathcal{A}}=\sum_{j=1}^K c_j$ if $\Delta_{\mathcal{T}}\geq \frac{1}{\lfloor(N-1)/4\rfloor}$ and $N\geq257$.\footnote{The condition $N\geq257$ is more like a technical requirement but not an obstacle in practice (see numerical simulations in Section \ref{sec:simulation}).} \label{thm:completedata}
\end{thm}
\begin{proof} See Section \ref{sec:proofcompletedata}.
\end{proof}
\begin{rem}[Necessary separation for uniform recovery] Theorem \ref{thm:completedata} presents a result of uniform recovery in the sense that it holds for all source signals $\mbra{s_{kt}}$ once the frequency separation condition $\Delta_{\mathcal{T}}\geq \frac{1}{\lfloor(N-1)/4\rfloor}\approx 4N^{-1}$ is satisfied. Consider the observation model (\ref{formu:observmodel1}) which contains $2KL+K$ real variables ($2KL$ for $\mbra{s_{kt}}$ and $K$ for $\lbra{f_k}$) and $2NL$ real measurements. As a result, for guaranteed uniform recovery, a necessary condition is that $2KL+K\leq2NL$ or equivalently $K\leq\frac{2NL}{2L+1}$. On the other hand, given the sparsity level $K$, we can always select a set of equispaced frequencies which has a minimum separation of $\frac{1}{K}\geq \sbra{1+\frac{1}{2L}}N^{-1}$. Therefore, a necessary condition of the minimum separation for uniform recovery is $\Delta_{\mathcal{T}}\geq \sbra{1+\frac{1}{2L}}N^{-1}$. \label{rem:necessaryseparation}
\end{rem}
We do not impose any assumption on the source signals $\mbra{s_{kt}}$ in Theorem \ref{thm:completedata}. Therefore, it can be applied uniformly to all kinds of source signals including correlated or even coherent sources. When all the sources are coherent as discussed in Remark \ref{rem:L0_coherent}, taking more measurement vectors does not increase the information for frequency recovery. In this case, in fact, it is easy to show that the proposed atomic norm minimization problem (\ref{formu:AN_sdp}) produces the same solution of $\m{u}$ up to a positive scale factor and thus the same frequency solution when increasing/decreasing the number of measurement vectors. As a result, we cannot expect a better theoretical guarantee than that in the SMV case. Our contribution by Theorem \ref{thm:completedata} is showing that in the presence of MMVs we can confidently recover the frequencies via a single convex optimization problem by exploiting the joint sparsity therein. Though our \emph{worst case} analysis does not improve over that in \cite{candes2013towards} on the SMV case, it will be shown in Section \ref{sec:simulation} via numerical simulations that the proposed joint sparse frequency recovery method improves the recovery performance significantly when the source signals are at general positions. We pose this \emph{average case} analysis as a future work.
\subsection{Theoretical Guarantees of Convex Relaxation: Incomplete Data}
It is difficult to analyze the atomic norm minimization problem (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}) with a specific sampling index set $\m{\Omega}$. Instead, we assume that it is selected uniformly at random. Our result is presented in the following theorem, which generalizes that in the SMV case in \cite{tang2012compressed} with modifications.
\begin{thm}[Incomplete data] Suppose we observe $\m{Y}^o=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}$
on the index set $\m{\Omega}\times\mbra{L}$, where $\m{\Omega}\subset\m{J}$ is of size $M$ and selected uniformly at random. Assume that $\lbra{\m{\phi}_j}_{j=1}^K\subset \mathbb{S}^{2L-1}$ are independent random variables with $\mathbb{E}\m{\phi}_j=\m{0}$. If $\Delta_{\mathcal{T}}\geq \frac{1}{\lfloor(N-1)/4\rfloor}$, then there exists a numerical constant $C$ such that
\equ{M\geq C\max\lbra{\log^2\frac{\sqrt{L}N}{\delta}, K\log\frac{K}{\delta}\log\frac{\sqrt{L}N}{\delta}} \label{formu:AN_bound}}
is sufficient to guarantee that, with probability at least $1-\delta$, $\m{Y}^o$ is the unique optimizer to (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}) and $\m{Y}^o=\sum_{j=1}^K c_j\m{a}\sbra{f_j,\m{\phi}_j}$ is the unique atomic decomposition satisfying that $\norm{\m{Y}^o}_{\mathcal{A}}=\sum_{j=1}^K c_j$. \label{thm:incompletedata}
\end{thm}
\begin{proof} See Section \ref{sec:proofincompletedata}.
\end{proof}
\begin{rem} We emphasize that the dependence of the sample size $M$ per snapshot on $L$ is for controlling the probability of successful recovery. To make it clear, we consider the case where we seek to recover the columns of $\m{Y}^o$ independently via the SMV atomic norm minimization. Then with $M$ satisfying (\ref{formu:AN_bound}) at $L=1$ as in \cite{tang2012compressed}, each column of $\m{Y}^o$ can be recovered with probability $1-\delta$. It follows that $\m{Y}^o$ can be exactly recovered with probability at least $1-L\delta$. However, if we attempt to recover $\m{Y}^o$ via a single convex optimization problem that we propose, then with the same number of measurements the success probability is improved to $1-\sqrt{L}\delta$ (replacing $\delta$ in (\ref{formu:AN_bound}) by $\sqrt{L}\delta$).
\end{rem}
\begin{rem}~
\begin{enumerate}
\item Compared to \cite{tang2012compressed} on the SMV case, Theorem \ref{thm:incompletedata} relaxes the assumption of $\lbra{\m{\phi}_j}_{j=1}^K$. In particular, the phases are assumed in the former drawn i.i.d. from the uniform distribution on the unit sphere. In contrast, they are only required to be independent and have zero means.
\item The relaxation of the assumption has significant impact on the practical DOA estimation problem in two aspects. On one hand, each $\m{\phi}_j$ corresponds to one source and therefore, they do not necessarily obey an identical distribution. On the other hand, under the assumption of Theorem \ref{thm:incompletedata} the sources are allowed to be (spatially) coherent or temporally correlated, which can be encountered in practical scenarios. For example, the source signals $\mbra{s_{kt}}$ are temporarily (column-wise) correlated if the rows of $\mbra{s_{kt}}$ are i.i.d. Gaussian with mean zero and non-diagonal covariance. Two sources (or two rows of $\mbra{s_{kt}}$) are certain to be coherent (identical up to a scale factor) if they are i.i.d. Gaussian with mean zero and covariance of rank one, where the resulting $\m{\phi}_j$'s satisfy the assumptions in Theorem \ref{thm:incompletedata}. In contrast, any two sources with $\m{\phi}_j$ drawn i.i.d. from the uniform distribution on $\mathbb{S}^{2L-1}$ are uncorrelated.
\end{enumerate}
\end{rem}
Theorem \ref{thm:incompletedata} shows that, if we fix the number of measurement vectors $L$, then a number of $O\sbra{K\log K\log N}$ samples per measurement vector are sufficient to recover the frequencies and the missing data by solving the convex optimization problem (\ref{formu:AN_sdp}) provided that the frequencies are sufficiently separate. When applied to array processing, it means that a number of $O\sbra{K\log K\log N}$ sensors are sufficient to exactly determine the directions of $K$ sources.
Our proofs of Theorems \ref{thm:completedata} and \ref{thm:incompletedata}, which are deferred to Sections \ref{sec:proofcompletedata} and \ref{sec:proofincompletedata} respectively, are inspired by those in \cite{candes2013towards} and \cite{tang2012compressed} and follow similar procedures. The main challenge of our proofs is dealing with vector-valued dual polynomials induced by the MMV problem.
\section{Connection to Prior Arts} \label{sec:connection}
\subsection{Grid-based Joint Sparse Recovery}
The problem of joint sparse frequency recovery concerned in this paper has been widely studied within the framework of CS, typically under the topic of DOA estimation. Since CS has been focused on signals that can be sparsely represented under a finite discrete dictionary, discretization/gridding of the frequency domain has become standard, or equivalently, the frequencies are assumed to lie on a fixed grid. Now recall the atomic $\ell_p$ norm, $p\in\lbra{0,1}$, in (\ref{formu:AL0}) and (\ref{formu:atomicnorm}). It is easy to show that, for $p\in\lbra{0,1}$,
\equ{\norm{\m{Y}}_{\mathcal{A},p}=\inf\lbra{\sum_k \twon{\m{s}_k}^p: \m{Y}=\sum_k\m{a}\sbra{f_k}\m{s}_k, f_k\in\mathbb{T}},}
where $\m{s}_k\in\mathbb{C}^{1\times L}$. As a result, the atomic $\ell_0$ norm (or the atomic norm) generalizes the $\ell_{2,0}$ norm (or the $\ell_{2,1}$ norm) in grid-based joint sparse recovery methods (see, e.g., \cite{malioutov2005sparse,hyder2010direction}) to the continuous frequency setting. It is worth noting that, since the grid-based methods rely on the ideal assumption that the true frequencies lie exactly on a fixed grid, we cannot expect exact frequency recovery when it fails unlike the continuous recovery technique studied in this paper. Moreover, even if the assumption holds, existing coherence or RIP-based analysis for the $\ell_{2,1}$-based joint sparse recovery is very conservative, compared to the results in this paper, due to the high coherence in the presence of a dense grid. Readers are referred to \cite{candes2013towards} for detailed discussions in the SMV case.
\newpage
\subsection{Gridless Joint Sparse Recovery}
To the best of our knowledge, the only discretization-free/gridless technique for joint sparse frequency recovery is introduced in \cite{yang2014discretization} prior to this work, termed as the sparse and parametric approach (SPA).\footnote{Another related paper published online is \cite{tan2013direction}, however, in this paper the authors reformulate the MMV problem into a SMV one, instead of exploiting the joint sparsity, and then solve the problem within the framework in \cite{candes2013towards}. Due to the loss of data structure, the capability of frequency recovery might be degraded.} Differently from the atomic norm technique proposed in this paper, SPA is from a statistical perspective and based on a weighted covariance fitting (WCF) criterion. But we show next that the two methods are connected. Given the complete data case as an example. In the limiting noiseless case, SPA is equivalent to the following optimization problem:
\equ{ \min_{\m{u}\in\mathbb{C}^N, T\sbra{\m{u}}\geq\m{0}} \tr\sbra{\widehat{\m{R}}T\sbra{\m{u}}^{-1}\widehat{\m{R}}}+\tr\sbra{T\sbra{\m{u}}}, \label{formu:SPA}}
where $\widehat{\m{R}}=\frac{1}{L}\m{Y}^o{\m{Y}^o}^H$ denotes the sample covariance. Let $\m{V}=\sbra{\frac{1}{N}{\m{Y}^o}^H\m{Y}^o}^{\frac{1}{2}}\in\mathbb{C}^{L\times L}$. Then we have the following equalities/equivalences:
\equ{\begin{split} \text{(\ref{formu:SPA})}
&= \min_{\m{u}, T\sbra{\m{u}}\geq\m{0}} \frac{N}{L^2}\tr\sbra{\sbra{\m{Y}^o\m{V}}^H T\sbra{\m{u}}^{-1} \m{Y}^o\m{V}} + \tr\sbra{T\sbra{\m{u}}}\\
&= \min_{\m{W},\m{u}} \tr\sbra{\m{W}} + \tr\sbra{T\sbra{\m{u}}}, \text{ subject to } \begin{bmatrix} \m{W} & \frac{\sqrt{N}}{L}\sbra{\m{Y}^o\m{V}}^H \\ \frac{\sqrt{N}}{L}\m{Y}^o\m{V} & T\sbra{\m{u}} \end{bmatrix}\geq\m{0}\\
&= 2\sqrt{N}\atomn{\frac{\sqrt{N}}{L}\m{Y}^o\m{V}}\\
&= \frac{2N}{L} \atomn{\m{Y}^o\m{V}},
\end{split} \notag}
where the third equality follows from the semidefinite formulation of the atomic norm in Theorem \ref{thm:AN_SDP}.
As a result, SPA is equivalent to computing the atomic norm of $\m{Y}^o\m{V}$. Note by (\ref{formu:observmodel}) that
\equ{\m{Y}^o\m{V}=\sum_{k=1}^K \m{a}\sbra{f_k}\sbra{\m{s}_{k}\m{V}}. \label{formu:observmodel_SPA}}
So, SPA is the atomic norm method proposed in this paper with modifications of the source signals.
In the special SMV case where $L=1$ and $\m{V}$ is a positive scalar, the two techniques are exactly equivalent. We note that a similar result holds with incomplete data and will be omitted.
\section{Proof of Theorem \ref{thm:completedata}} \label{sec:proofcompletedata}
\subsection{Dual Certificate}
We have considered the general case where $\m{J}=\lbra{0,\cdots,N-1}$ in this paper. For technical reasons, our proof is mainly focused on the symmetric case where $\m{J}=\lbra{-2n,\dots,2n}$ with $n=\lfloor \frac{N-1}{4}\rfloor$, in which all our previous results hold as well with the substitution $N=\abs{\m{J}}$. We complete the proof by showing that the same result holds in the general case.
The following proposition presents the so-called \emph{dual certificate} which guarantees the optimality of a certain atomic decomposition.
\begin{prop} Suppose that the atomic set $\mathcal{A}$ is composed of atoms defined by $\m{a}\sbra{f,\m{\phi}}$ whose rows are indexed by the set $\m{J}$ being either $\lbra{-2n,\dots,2n}$ or $\lbra{0,\cdots,N-1}$. Then $\m{Y}^o=\sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_k}$ is the unique atomic decomposition satisfying that $\atomn{\m{Y}^o}=\sum_{k=1}^K c_k$ if $K\leq\abs{\m{J}}$ and there exists a vector-valued dual polynomial $Q: \mathbb{T}\rightarrow \mathbb{C}^{1\times L}$
\equ{Q(f)=\inp{\m{V},\m{a}\sbra{f}}\triangleq\m{a}(f)^H\m{V} \label{formu:dualpoly1}}
satisfying that
\setlength\arraycolsep{2pt}{\equa{ Q\sbra{f_k}
&=& \m{\phi}_k, \quad f_k\in\mathcal{T}, \label{formu:cons1}\\ \twon{Q\sbra{f}}
&<& 1, \quad f\in\mathbb{T}\backslash\mathcal{T}, \label{formu:cons2}
}}where the coefficient matrix $\m{V}\in\mathbb{C}^{\abs{\m{J}}\times L}$.\footnote{Here we have abused the notation of inner-product since $\m{V}\in\mathbb{C}^{\abs{\m{J}}\times L}$ and $\inp{\m{V},\m{a}\sbra{f}}=\m{a}(f)^H\m{V}\in\mathbb{C}^{1\times L}$.} \label{prop:dualpoly1}
\end{prop}
\begin{proof} For a vector-valued polynomial $Q(f)=\m{a}(f)^H\m{V}$ satisfying (\ref{formu:cons1}) and (\ref{formu:cons2}), we have by (\ref{formu:dualAN}) that
\equ{\datomn{\m{V}}=\sup_{f\in\mathbb{T}}\twon{\m{a}(f)^H\m{V}}=1. \notag}
Based on (\ref{formu:cons1}) and (\ref{formu:cons2}) it is easy to show that
\equ{\begin{split} \inp{\m{V}, \m{Y}^o}_{\mathbb{R}}
&= \inp{\m{V}, \sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_k}}_{\mathbb{R}} \\
&= \sum_{k=1}^K c_k\inp{\m{a}\sbra{f_k}^H\m{V}, \m{\phi}_k}_{\mathbb{R}} \\
&= \sum_{k=1}^K c_k\geq \atomn{\m{Y}^o}.
\end{split} \notag}
The last inequality follows from the definition of the atomic norm. On the other hand, it holds that $\inp{\m{V}, \m{Y}^o}_{\mathbb{R}}\leq \datomn{\m{V}}\atomn{\m{Y}^o}\leq \atomn{\m{Y}^o}$, implying that $\inp{\m{V}, \m{Y}^o}_{\mathbb{R}}= \atomn{\m{Y}^o} =\sum_{k=1}^K c_k$. We next show the uniqueness. Suppose that there exists another decomposition $\m{Y}^o=\sum_{k} \widetilde c_k\m{a}\sbra{\widetilde f_k,\widetilde{\m{\phi}}_k}$ satisfying also that $\atomn{\m{Y}^o}=\sum_{k} \widetilde c_k$. There must exist some $\widetilde f_k\notin\mathcal{T}$ due to linear independence between $\lbra{\m{a}\sbra{f_k}}_{k=1}^K$, $K\leq\abs{\m{J}}$, for both the settings of $\m{J}$, otherwise the two decompositions will be identical. We complete the proof by the following contradiction:
\equ{\begin{split}
&\sum_{k} \widetilde c_k=\atomn{\m{Y}^o}=\inp{\m{V}, \m{Y}^o}_{\mathbb{R}}\\
&= \inp{\m{V}, \sum_{k} \widetilde c_k\m{a}\sbra{\widetilde f_k,\widetilde {\m{\phi}}_k}}_{\mathbb{R}} \\
&= \sum_{\widetilde f_k\in\mathcal{T}} \widetilde c_k\inp{Q\sbra{\widetilde f_k}, \widetilde{\m{\phi}}_k}_{\mathbb{R}} + \sum_{\widetilde f_k\notin\mathcal{T}} \widetilde c_k\inp{Q\sbra{\widetilde f_k}, \widetilde{\m{\phi}}_k}_{\mathbb{R}} \\
&< \sum_{\widetilde f_k\in\mathcal{T}} \widetilde c_k + \sum_{\widetilde f_k\notin\mathcal{T}} \widetilde c_k =\sum_{k} \widetilde c_k.
\end{split} \notag}
\end{proof}
The rest of the proof is focused on construction of a dual polynomial as in Proposition \ref{prop:dualpoly1}, for which the condition $K\leq\abs{\m{J}}$ naturally holds. We revisit the proof in \cite{candes2013towards} for the SMV case in the ensuing subsection and present our proof for the general MMV case after that.
\subsection{Revisiting the SMV Case} \label{sec:detourcomplete}
We consider only the case where $\m{J}=\lbra{-2n,\dots,2n}$. The reason will be clear in the next subsection.
Cand\`{e}s and Fernandez-Granda \cite{candes2013towards} consider a polynomial $q: \mathbb{T}\rightarrow \mathbb{C}$ of the following form:
\equ{q\sbra{f}=\sum_{f_k\in\mathcal{T}}\alpha_k \mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}}\beta_k \mathcal{K}'\sbra{f-f_k}, \notag}
where $\mathcal{K}\sbra{f}$ is the squared Fej\'{e}r kernel
\equ{\mathcal{K}\sbra{f}=\mbra{\frac{\sin(\pi (n+1)f)}{(n+1)\sin\sbra{\pi f}}}^4=\sum_{j=-2n}^{2n}g_n\sbra{j}e^{-i2\pi jf} \notag}
with coefficients
\equ{g_n\sbra{j}=\frac{1}{n+1}\sum_{k=\max\sbra{j-n-1,-n-1}}^{\min\sbra{j+n+1,n+1}} \sbra{1-\frac{\abs{k}}{n+1}}\sbra{1-\frac{\abs{j-k}}{n+1}} \label{formu:gnj}}
obeying that $0<g_n\sbra{j}\leq1$, $j=-2n,\dots,2n$. Denote by $\mathcal{K}'$, $\mathcal{K}''$ and $\mathcal{K}'''$ the first three derivatives of $\mathcal{K}$. To obtain the coefficients $\lbra{\alpha_k}$ and $\lbra{\beta_k}$ such that $q$ satisfies the constraints of the dual polynomial, they require for $f_j\in\mathcal{T}$ that
\setlength\arraycolsep{2pt}{\equa{q\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\alpha_k \mathcal{K}\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}} \beta_k\mathcal{K}'\sbra{f_j-f_k} =\phi_j, \label{formu:dualpolycons11}\\ q'\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\alpha_k \mathcal{K}'\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}} \beta_k\mathcal{K}''\sbra{f_j-f_k} =0.\label{formu:dualpolycons21}
}}The equality (\ref{formu:dualpolycons11}) ensures that $q\sbra{f}$ satisfies the interpolation condition (\ref{formu:cons1}), while (\ref{formu:dualpolycons21}) is used to guarantee the inequality (\ref{formu:cons2}). (\ref{formu:dualpolycons11}) and (\ref{formu:dualpolycons21}) can be written more compactly as
\equ{\begin{bmatrix}\m{D}_0 & c_0^{-1}\m{D}_1 \\ -c_0^{-1}\m{D}_1 & -c_0^{-2}\m{D}_2\end{bmatrix} \begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix} = \begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}, \label{formu:linsys1}}
where $\mbra{D_0}_{jk}=\mathcal{K}\sbra{f_j-f_k}$, $\mbra{D_1}_{jk}=\mathcal{K}'\sbra{f_j-f_k}$, $\mbra{D_2}_{jk}=\mathcal{K}''\sbra{f_j-f_k}$, $c_0=\sqrt{\abs{\mathcal{K}''\sbra{0}}}=\sqrt{\frac{4\pi^2n(n+2)}{3}}$, $\m{\Phi}\in\mathbb{C}^{K}$ is the vector by stacking $\lbra{\phi_j}$ together and similarly for $\m{\alpha}, \m{\beta}\in\mathbb{C}^{K}$. The system of linear equations above is rescaled following from \cite{tang2012compressed} such that the coefficient matrix of (\ref{formu:linsys1}) is symmetric, positive definite, and very close to identity. Then we have
\equ{\begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix} = \begin{bmatrix}\m{I}\\ -c_0\m{D}_2^{-1}\m{D}_1\end{bmatrix}\m{D}_3^{-1}\m{\Phi}, \quad \m{D}_3\triangleq\m{D}_0-\m{D}_1\m{D}_2^{-1}\m{D}_1, \label{formu:alphabeta1}}
where $\m{D}_3$ is the Schur complement, and the solution $\begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix}$ is shown to be close to $\begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}$ with $\inftyn{\m{\alpha}}$ and $\inftyn{\m{\beta}}$ upper bounded.
To prove that the dual polynomial $q\sbra{f}$ with the specified coefficients satisfies the condition (\ref{formu:cons2}), the authors divide the the frequency domain $\mathbb{T}$ into two regions near to and far from the frequencies $\mathcal{T}$:
\setlength\arraycolsep{2pt}{\equa{ \mathcal{T}_{\text{near}}
&=& \cup_{k=1}^K\mbra{f_k-\nu,f_k+\nu}, \\ \mathcal{T}_{\text{far}}
&=& \mathbb{T}\backslash \mathcal{T}_{\text{near}}
}}with $\nu=8.245\times 10^{-2}\frac{1}{n}$. $\abs{q\sbra{f}}<1$ is shown directly on $\mathcal{T}_{\text{far}}$, and on each continuous interval of $\mathcal{T}_{\text{near}}$, $\abs{q\sbra{f}}$ is shown to be strictly concave and obtains the maximum 1 at each element in $\mathcal{T}$. We summarize some results in \cite{candes2013towards} which will be used later.
\begin{lem} Assume $\Delta_{\mathcal{T}}\geq\Delta_{\text{min}}\triangleq\frac{1}{n}$ and $n\geq64$.
\begin{enumerate}
\item
\setlength\arraycolsep{2pt}{\equa{ \inftyn{\m{I}-\m{D}_0}
&\leq& 6.253\times10^{-3}, \notag \\ \inftyn{\m{D}_1}
&\leq& 0.1528n, \notag \\ \inftyn{\abs{\mathcal{K}''\sbra{0}}\m{I}-\m{D}_2}
&\leq& 4.212n^2, \notag \\ \inftyn{\m{I}-\m{D}_3^{-1}}
&\leq& 8.824\times10^{-3}. \notag
}}
\item For $f\in\mbra{-\nu,\nu}$ (or $f\in\mbra{0,\nu}\cup \mbra{1-\nu,1}$),
\setlength\arraycolsep{2pt}{\equa{ 1\geq K\sbra{f}
&\geq& 0.9539, \notag\\ -11.69n^2\geq K''\sbra{f}
&\geq& -13.57n^2. \notag
}}
\item For $f\in\mathcal{T}_{\text{near}}$,
\setlength\arraycolsep{2pt}{\equa{ \sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}'\sbra{f-f_k}}
&\leq& 1.272n, \notag\\ \sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}'''\sbra{f-f_k}}
&\leq& 193.9n^3; \notag
}}for $f\in\mathcal{T}_{\text{near}}\backslash \mbra{f_j-\nu,f_j+\nu}$, $f_j\in\mathcal{T}$,
\setlength\arraycolsep{2pt}{\equa{ \sum_{f_k\notin\mathcal{T}\backslash\lbra{f_j}}\abs{\mathcal{K}\sbra{f-f_k}}
&\leq& 6.279\times10^{-3}, \notag \\ \sum_{f_k\notin\mathcal{T}\backslash\lbra{f_j}}\abs{\mathcal{K}''\sbra{f-f_k}}
&\leq& 4.212n^2; \notag
}}and for $f\in\mathcal{T}_{\text{far}}$,
\equ{\begin{split}
&1.008824\sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}\sbra{f-f_k}} \\
&+ \frac{0.01647}{n} \sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}'\sbra{f-f_k}} \leq0.99992. \end{split} \notag}
\end{enumerate}
\end{lem}
\subsection{Proof of Theorem \ref{thm:completedata}} \label{sec:proofcompdata}
Theorem \ref{thm:completedata} is a direct result of the proposition below.
\begin{prop} Under the assumptions of Theorem \ref{thm:completedata}, there exists a low-frequency vector-valued polynomial as in (\ref{formu:dualpoly1}) satisfying (\ref{formu:cons1}) and (\ref{formu:cons2}) for $\m{J}$ being either $\lbra{-2n,\dots,2n}$ or $\lbra{0,\cdots,N-1}$, where $K\leq\abs{\m{J}}$ naturally holds. \label{prop:dualpoly}
\end{prop}
To prove Proposition \ref{prop:dualpoly}, we first show that once we can construct such a dual polynomial in the symmetric case, then a similar polynomial can be obtained in the general case.
Assume $N=4n+n_0$ with $n=\lfloor \frac{N-1}{4}\rfloor$ and $n_0=1,2,3,4$. Suppose that $\m{Y}^o$ has a decomposition in the general case
\equ{\begin{split}\m{Y}^o
&= \sum_{k=1}^K c_k\begin{bmatrix}1\\ e^{i2\pi f_k} \\ \vdots \\ e^{i2\pi(N-1)f_k}\end{bmatrix}\m{\phi}_k\\
&= \sum_{k=1}^K c_k \begin{bmatrix}e^{i2\pi(-2n) f_k}\\ \vdots\\ e^{i2\pi(2n) f_k} \\ \vdots \\ e^{i2\pi(2n+n_0-1)f_k}\end{bmatrix} \underbrace{e^{i2\pi(2n)f_k}\m{\phi}_k}_{\widetilde{\m{\phi}}_k}. \end{split} \label{formu:Yotwodecompositions}}
Moreover, suppose that we can construct a dual polynomial $\widetilde Q\sbra{f}=\sum_{j=-2n}^{2n}\widetilde{\m{V}}_je^{-i2\pi jf}$ in the symmetric case satisfying that
\setlength\arraycolsep{2pt}{\equa{\widetilde Q\sbra{f_k}
&=&\widetilde{\m{\phi}}_k, \quad f_k\in\mathcal{T}, \notag\\ \twon{\widetilde Q\sbra{f}}
&<& 1, \quad f\in\mathbb{T}\backslash\mathcal{T}, \notag
}}and $K=\abs{\mathcal{T}}\leq4n+1\leq N$. Now define $\m{V}\in\mathbb{C}^{N\times L}$ with
\equ{\m{V}_j=\left\{\begin{array}{ll}\widetilde{\m{V}}_{j-2n}, & j=0,\dots,4n,\\ \m{0}, &\text{otherwise}, \end{array} \right. \notag}
and
\equ{\begin{split}Q\sbra{f}
&=\sum_{j=0}^{N-1}\m{V}_je^{-i2\pi j f}\\
&=\sum_{j=0}^{4n} \widetilde{\m{V}}_{j-2n} e^{-i2\pi j f}\\
&= e^{-i2\pi(2n)f}\widetilde Q\sbra{f}. \end{split} \notag}
It can be readily verified that
\setlength\arraycolsep{2pt}{\equa{Q\sbra{f_k}
&=&e^{-i2\pi(2n)f_k}\widetilde{\m{\phi}}_k = \m{\phi}_k, \quad f_k\in\mathcal{T}, \notag\\ \twon{Q\sbra{f}}
&=& \twon{\widetilde Q\sbra{f}}< 1, \quad f\in\mathbb{T}\backslash\mathcal{T}. \notag
}}That means, we have constructed a dual polynomial $Q\sbra{f}$ for the general case as in (\ref{formu:dualpoly1}) satisfying (\ref{formu:cons1}) and (\ref{formu:cons2}) with $K\leq N$.
We prove Proposition \ref{prop:dualpoly} in the symmetric case in the rest of this subsection and thus complete the proof of Theorem \ref{thm:completedata}. Inspired by \cite{candes2013towards}, we construct the following vector-valued dual polynomial:
\equ{Q\sbra{f}=\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}}\m{\beta}_k \mathcal{K}'\sbra{f-f_k}, \label{formu:dualpoly}}
where for each $k$, $\m{\alpha}_k,\m{\beta}_k\in\mathbb{C}^{1\times L}$ are vector-valued coefficients. Again, for $f_j\in\mathcal{T}$, we impose the following equalities:
\setlength\arraycolsep{2pt}{\equa{Q\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}} \m{\beta}_k\mathcal{K}'\sbra{f_j-f_k} =\m{\phi}_j, \label{formu:dualpolycons1}\\ Q'\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}'\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}}\m{\beta}_k\mathcal{K}''\sbra{f_j-f_k} =\m{0},\label{formu:dualpolycons2}
}}or equivalently,
\equ{\begin{bmatrix}\m{D}_0 & c_0^{-1}\m{D}_1 \\ -c_0^{-1}\m{D}_1 & -c_0^{-2}\m{D}_2\end{bmatrix} \begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix} = \begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}, \label{formu:linsysalpbet}}
where we denote $\m{\Phi}=\begin{bmatrix}\m{\phi}_1^T,\dots,\m{\phi}_K^T\end{bmatrix}^T\in\mathbb{C}^{K\times L}$, $\m{\alpha}=\begin{bmatrix}\m{\alpha}_1^T,\dots,\m{\alpha}_K^T\end{bmatrix}^T\in\mathbb{C}^{K\times L}$ and similarly for $\m{\beta}\in\mathbb{C}^{K\times L}$. Now we have $2KL$ equations with $2KL$ variables in total. Therefore, we can uniquely determine $\m{\alpha}$ and $\m{\beta}$ as in (\ref{formu:alphabeta1}) following from (\ref{formu:linsysalpbet}).
Note that the coefficient matrix in (\ref{formu:linsysalpbet}) is close to identity and hence, $\begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix}$ should be close to $\begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}$. Differently from that in the SMV case, however, $\lbra{\m{\alpha}_k}$ and $\lbra{\m{\beta}_k}$ are vector-valued. The difficulty of the remaining proof is to provide tight bounds for them. To do this, we introduce the concept of $\ell_{2,\infty}$ matrix norm and its induced operator norm as follows.
\begin{defi} We define the $\ell_{2,\infty}$ norm of a matrix $\m{X}\in\mathbb{C}^{d_1\times d_2}$ as
\equ{\twoinfn{\m{X}}=\max_{j} \twon{\m{X}_j} \notag}
and its induced norm of a linear operator $\m{\mathcal{P}}: \mathbb{C}^{d_1\times d_2}\rightarrow\mathbb{C}^{d_3\times d_2}$ as
\equ{\twoinfn{\m{\mathcal{P}}}=\sup_{\m{X}\neq\m{0}} \frac{\twoinfn{\m{\mathcal{P}}\m{X}}}{\twoinfn{\m{X}}}= \sup_{\twoinfn{\m{X}}\leq1} \twoinfn{\m{\mathcal{P}}\m{X}}, \notag}
where $d_1$, $d_2$ and $d_3$ are positive integers.
\end{defi}
By the definition, we have $\twoinfn{\m{\Phi}}=1$ and expect to bound $\twoinfn{\m{\alpha}}$ and $\twoinfn{\m{\beta}}$ using the induced norm of the operators $\m{D}_j$, $j=0,\dots,3$. So we need to calculate the induced norm first. Interestingly, the induced $\ell_{2,\infty}$ norm is identical to the $\ell_{\infty}$ norm, which is given in the following result.
\begin{lem} $\twoinfn{\m{\mathcal{P}}}=\inftyn{\m{\mathcal{P}}}$ for any linear operator $\m{\mathcal{P}}$ defined by a matrix $\m{P}$ such that $\m{\mathcal{P}}\m{X}=\m{P}\m{X}$ for any $\m{X}$ of proper dimension. \label{lem:twoinfinitynorm}
\end{lem}
\begin{proof} According to the definition of the induced $\ell_{\infty}$ norm, we have $\inftyn{\m{\mathcal{P}}}=\inftyn{\m{P}}=\max_j \onen{\m{P}_j}$, where $\m{P}_j$ is the $j$th row of $\m{P}$.
We first show that $\twoinfn{\m{\mathcal{P}}}\geq\inftyn{\m{\mathcal{P}}}$. Denote by $\m{X}_{:1}$ and $\m{X}_{:-1}$ respectively the first column and the rest of $\m{X}$. Note that if $\m{X}_{:-1}=\m{0}$, then $\twoinfn{\m{X}}=\inftyn{\m{X}_{:1}}$ and $\twoinfn{\m{\mathcal{P}}\m{X}}=\inftyn{\m{P}\m{X}_{:1}}$. Hence,
\equ{\begin{split}\twoinfn{\m{\mathcal{P}}}
&\geq \sup_{\twoinfn{\m{X}}\leq 1, \m{X}_{:-1}=\m{0}} \twoinfn{\m{\mathcal{P}}\m{X}} \\
&= \sup_{\inftyn{\m{X}_{:1}}\leq1} \inftyn{\m{P}\m{X}_{:1}}\\
&= \inftyn{\m{P}}=\inftyn{\m{\mathcal{P}}}. \end{split} \notag}
We next show that $\twoinfn{\m{\mathcal{P}}}\leq\inftyn{\m{\mathcal{P}}}$ to complete the proof. Note that
\equ{\begin{split} \twoinfn{\m{\mathcal{P}}\m{X}}^2
&= \max_j \twon{\m{P}_j\m{X}}^2 \\
&= \max_j \m{P}_j\m{X} \m{X}^H\m{P}_j^H \\
&= \max_j \tr\sbra{\m{R}\m{P}_j^H\m{P}_j},
\end{split} \notag}
where $\m{R}\triangleq\m{X}\m{X}^H$. It is easy to show that if $\twoinfn{\m{X}}\leq1$, then $\m{R}\geq\m{0}$ and $\m{R}\preccurlyeq \m{1}$, and vise versa, where $\m{R}\preccurlyeq \m{1}$ means $R_{jl}\leq 1$ for all the entries $R_{jl}$ of $\m{R}$. That means, the two constraints $\m{R}\geq\m{0}$ and $\m{R}\preccurlyeq\m{1}$ are equivalent to $\twoinfn{\m{X}}\leq1$. As a result,
\equ{\begin{split}\twoinfn{\m{\mathcal{P}}}^2
&= \sup_{\twoinfn{\m{X}}\leq1} \twoinfn{\m{\mathcal{P}}\m{X}}^2 \\
&= \sup_{\twoinfn{\m{X}}\leq1} \max_j \tr\sbra{\m{R}\m{P}_j^H\m{P}_j}\\
&= \max_j \sup_{\twoinfn{\m{X}}\leq1} \inp{\text{vec}\sbra{\m{R}}, \text{vec}\sbra{\m{P}_j^H\m{P}_j}} \\
&= \max_j \sup_{\m{R}\geq\m{0}, \m{R}\preccurlyeq\m{1}} \inp{\text{vec}\sbra{\m{R}}, \text{vec}\sbra{\m{P}_j^H\m{P}_j}} \\
&\leq \max_j \sup_{\m{R}\geq\m{0}, \m{R}\preccurlyeq\m{1}} \inftyn{\text{vec}\sbra{\m{R}}} \onen{\text{vec}\sbra{\m{P}_j^H\m{P}_j}} \\
&\leq \max_j \onen{\m{P}_j}^2\\
&= \inftyn{\m{\mathcal{P}}}^2,
\end{split} \notag}
where $\text{vec}\sbra{\cdot}$ denotes the vectorized form of a matrix argument, i.e., by stacking all its columns as a column vector.
The last inequality follows from that $\inftyn{\text{vec}\sbra{\m{R}}}\leq1$ provided $\m{R}\preccurlyeq\m{1}$, and $\onen{\text{vec}\sbra{\m{P}_j^H\m{P}_j}}=\onen{\m{P}_j}^2$.
\end{proof}
Following from Lemma \ref{lem:twoinfinitynorm}, we use only $\inftyn{\m{\mathcal{P}}}$ rather than $\twoinfn{\m{\mathcal{P}}}$ to denote the induced $\ell_{2,\infty}$ norm hereafter for notational simplicity, which also avoids ambiguities between the matrix norm and the induced operator norm of a matrix (a matrix can also denote a linear operator). The following lemma provides upper bounds for $\m{\alpha}$ and $\m{\beta}$.
\begin{lem} Under the assumptions of Proposition \ref{prop:dualpoly}, the matrices $\m{\alpha},\m{\beta}\in\mathbb{C}^{K\times L}$ determined by (\ref{formu:linsysalpbet}) satisfy that
\setlength\arraycolsep{2pt}{\equa{ \twoinfn{\m{\alpha}-\m{\Phi}}
&\leq& 8.824\times10^{-3}, \notag\\ \twoinfn{\m{\beta}}
&\leq& \beta^{\text{max}}\triangleq\frac{1.647}{n}\times10^{-2}. \notag
}}It follows that $1-8.824\times10^{-3}\triangleq \alpha^{\text{min}}\leq\twon{\m{\alpha}_j}\leq\alpha^{\text{max}}\triangleq1+8.824\times10^{-3}$, $j=1,\dots,K$.
\end{lem}
\begin{proof} As in (\ref{formu:alphabeta1}) we have
\equ{\begin{bmatrix}\m{\alpha}\\ \m{\beta}\end{bmatrix} = \begin{bmatrix}\m{I}\\ -\m{D}_2^{-1}\m{D}_1\end{bmatrix}\m{D}_3^{-1}\m{\Phi}. \label{formu:alphabeta}}
It follows that $\m{\alpha}-\m{\Phi}=\sbra{\m{D}_3^{-1}-\m{I}}\m{\Phi}$, and
\equ{\twoinfn{\m{\alpha}-\m{\Phi}}\leq \inftyn{\m{D}_3^{-1}-\m{I}}\twoinfn{\m{\Phi}}\leq 8.824\times10^{-3}. \notag}
The interval of $\twon{\m{\alpha}_j}$ is a direct result of the inequality
\equ{\twon{\m{\phi}_j}-\twon{\m{\alpha}_j-\m{\phi}_j}\leq \twon{\m{\alpha}_j} \leq \twon{\m{\phi}_j}+\twon{\m{\alpha}_j-\m{\phi}_j} \notag} provided $\twon{\m{\phi}_j}=1$, $j\in\mbra{K}$.
Utilizing $\inftyn{\m{A}\m{B}}\leq \inftyn{\m{A}}\inftyn{\m{B}}$ and $\inftyn{\m{A}^{-1}}\leq\frac{1}{1-\inftyn{\m{I}-\m{A}}}$, we have
\equ{\begin{split}\twoinfn{\m{\beta}}
&\leq \inftyn{\m{D}_2^{-1}\m{D}_1\m{D}_3^{-1}}\\
&\leq \frac{\inftyn{\m{D}_1}\inftyn{\m{D}_3^{-1}}} {\abs{\mathcal{K}''(0)}-\inftyn{\abs{\mathcal{K}''(0)}\m{I}-\m{D}_2}}\\
&\leq\frac{1.647}{n}\times10^{-2}. \end{split} \notag}
\end{proof}
We now complete the proof of Proposition \ref{prop:dualpoly} by showing that the polynomial $Q\sbra{f}$ constructed above satisfies (\ref{formu:cons2}). On $\mathcal{T}_{\text{far}}$,
\equ{\begin{split}
\twon{Q\sbra{f}}&= \twon{\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}}\m{\beta}_k \mathcal{K}'\sbra{f-f_k}}\\
&\leq \sum_{f_k\in\mathcal{T}}\twon{\m{\alpha}_k} \abs{\mathcal{K}\sbra{f-f_k}} + \sum_{f_k\in\mathcal{T}}\twon{\m{\beta}_k} \abs{\mathcal{K}'\sbra{f-f_k}}\\
&\leq \alpha^{\text{max}}\sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}\sbra{f-f_k}} + \beta^{\text{max}} \sum_{f_k\in\mathcal{T}}\abs{\mathcal{K}'\sbra{f-f_k}} \\
&\leq 0.99992.
\end{split} \label{formu:Qonfar}}
On each interval $\mbra{f_j-\nu, f_j+\nu}$, $f_j\in\mathcal{T}$, since $\twon{Q\sbra{f}}=1$ and $Q'\sbra{f}=\m{0}$, we need only to show that $\twon{Q\sbra{f}}^2$ is a concave function, i.e.,
\equ{\frac{1}{2}\frac{{\text{d}^2}\twon{Q\sbra{f}}^2}{\text{d}f^2}=\twon{Q'\sbra{f}}^2 +\Re\lbra{Q''\sbra{f}Q\sbra{f}^H}<0. \notag}
We observe that
\equ{\begin{split}
\twon{Q'\sbra{f}}
&= \twon{\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}'\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}} \m{\beta}_k\mathcal{K}''\sbra{f-f_k}}\\
&\leq \alpha^{\text{max}}\sum_{f_k\in\mathcal{T}} \abs{\mathcal{K}'\sbra{f-f_k}} + \beta^{\text{max}}\abs{\mathcal{K}''\sbra{f-f_j}}\\ &\quad+\beta^{\text{max}}\sum_{f_k\notin\mathcal{T}\backslash\lbra{f_j}} \abs{\mathcal{K}''\sbra{f-f_k}}\\
&\leq 1.576n.
\end{split} \notag}
Moreover,
\equ{\begin{split}
Q''\sbra{f}Q\sbra{f}^H
&= \mbra{\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}''\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}} \m{\beta}_k\mathcal{K}'''\sbra{f-f_k}} \\
&\quad \times \mbra{\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}} \m{\beta}_k\mathcal{K}'\sbra{f-f_k}}^H \\
&= \twon{\m{\alpha}_j}^2 \mathcal{K}''\sbra{f-f_j}\mathcal{K}\sbra{f-f_j}\\
&\quad +\mathcal{K}''\sbra{f-f_j}\sum_{f_k\notin \mathcal{T}\backslash\lbra{f_j}} \m{\alpha}_j\m{\alpha}_k^H\mathcal{K}\sbra{f-f_k} \\
&\quad +\mathcal{K}''\sbra{f-f_j}\sum_{f_k\in \mathcal{T}} \m{\alpha}_j\m{\beta}_k^H\mathcal{K}'\sbra{f-f_k} \\
&\quad +\mbra{\sum_{f_k\notin \mathcal{T}\backslash\lbra{f_j}} \m{\alpha}_k\mathcal{K}''\sbra{f-f_k}}Q\sbra{f}^H \\
&\quad +\mbra{\sum_{f_k\in \mathcal{T}} \m{\beta}_k\mathcal{K}'''\sbra{f-f_k}}Q\sbra{f}^H,
\end{split} \notag}
\equ{\begin{split}
\twon{\m{\alpha}_j}^2 \mathcal{K}''\sbra{f-f_j}\mathcal{K}\sbra{f-f_j}
&\leq \sbra{\alpha^{\text{min}}}^2 \times\sbra{-11.69n^2}\times0.9539\\
&= -10.96n^2,
\end{split} \notag}
and
\equ{\begin{split}
\twon{Q\sbra{f}}
&\leq \alpha^{\text{max}}\abs{\mathcal{K}\sbra{f-f_j}}+ \alpha^{\text{max}}\sum_{f_k\notin\mathcal{T}\backslash\lbra{f_j}}\abs{\mathcal{K}\sbra{f-f_k}} \\
&\quad +\beta^{\text{max}}\sum_{f_k\in\mathcal{T}} \abs{\mathcal{K}'\sbra{f-f_k}} \\
&\leq 1.036.
\end{split} \notag}
It follows that
\equ{\begin{split}
\Re\lbra{Q''\sbra{f}Q\sbra{f}^H}
&\leq -10.96n^2 \\
&\quad +\sbra{\alpha^{\text{max}}}^2 \abs{\mathcal{K}''\sbra{f-f_j}}\sum_{f_k\notin \mathcal{T}\backslash\lbra{f_j}} \abs{\mathcal{K}\sbra{f-f_k}} \\
&\quad +\alpha^{\text{max}}\beta^{\text{max}}\abs{\mathcal{K}''\sbra{f-f_j}}\sum_{f_k\in \mathcal{T}} \abs{\mathcal{K}'\sbra{f-f_k}} \\
&\quad +\alpha^{\text{max}}\sum_{f_k\notin \mathcal{T}\backslash\lbra{f_j}} \abs{\mathcal{K}''\sbra{f-f_k}}\twon{Q\sbra{f}} \\
&\quad +\beta^{\text{max}}\sum_{f_k\in \mathcal{T}} \abs{\mathcal{K}'''\sbra{f-f_k}}\twon{Q\sbra{f}} \\
&\leq -2.875n^2.
\end{split} \notag}
Finally, we conclude that
\equ{\begin{split} \frac{1}{2}\frac{{\text{d}^2}\twon{Q\sbra{f}}^2}{\text{d}f^2}
&=\twon{Q'\sbra{f}}^2 +\Re\lbra{Q''\sbra{f}Q\sbra{f}^H} \\
&\leq -0.3910n^2. \end{split} \notag}
\section{Proof of Theorem \ref{thm:incompletedata}} \label{sec:proofincompletedata}
\subsection{Dual Certificate}
For incomplete data, we also begin with the dual certificate stated in the proposition below, which generalizes Proposition \ref{prop:dualpoly1} for complete data.
\begin{prop} Suppose that the atomic set $\mathcal{A}$ is composed of atoms defined by $\m{a}\sbra{f,\m{\phi}}$ whose rows are indexed by the set $\m{J}$ being either $\lbra{-2n,\dots,2n}$ or $\lbra{0,\cdots,N-1}$. Then $\m{Y}^o=\sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_k}$ is the unique optimizer to (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}) if $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}\subset \mathcal{A}_{\m{\Omega}}^1$ are linearly independent and there exists a vector-valued dual polynomial $Q: \mathbb{T}\rightarrow \mathbb{C}^{1\times L}$
\equ{Q(f)=\m{a}(f)^H\m{V} \label{formu:dualpoly1_inc}}
satisfying that
\setlength\arraycolsep{2pt}{\equa{ Q\sbra{f_k}
&=& \m{\phi}_k, \quad f_k\in\mathcal{T}, \label{formu:cons1_inc}\\ \twon{Q\sbra{f}}
&<& 1, \quad f\in\mathbb{T}\backslash\mathcal{T}, \label{formu:cons2_inc}\\ \m{V}_j
&=& \m{0}_{L\times1}, \quad j\notin \m{\Omega}, \label{formu:cons3_inc}
}}where the coefficient matrix $\m{V}\in\mathbb{C}^{\abs{\m{J}}\times L}$. Moreover, $\m{Y}^o=\sum_{k=1}^K c_k\m{a}\sbra{f_k,\m{\phi}_k}$ is the unique atomic decomposition satisfying that $\atomn{\m{Y}^o}=\sum_{k=1}^K c_k$. \label{prop:dualpoly_inc}
\end{prop}
\begin{proof} The proof is similar to that of Proposition \ref{prop:dualpoly1}. Note that the dual problem of (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}) is (\ref{formu:dual1}), and strong duality holds between them. We can similarly show that $\datomn{\m{V}}\leq1$ and $\inp{\m{V}_{\m{\Omega}},\m{Y}_{\m{\Omega}}^o}_{\mathbb{R}}=\atomn{\m{Y}^o}$. Since $\sbra{\m{Y}^o,\m{V}}$ is primal-dual feasible, we conclude that $\m{Y}^o$ is a primal optimal solution due to the strong duality. Suppose that there exists another optimal solution $\widetilde{\m{Y}}\neq\m{Y}^o$ with an atomic decomposition $\widetilde{\m{Y}}=\sum_{k} \widetilde c_k\m{a}\sbra{\widetilde f_k,\widetilde{\m{\phi}}_k}$ satisfying that $\atomn{\widetilde{\m{Y}}}=\sum_{k} \widetilde c_k=\atomn{\m{Y}^o}$. Since $\widetilde{\m{Y}}_{\m{\Omega}} = \m{Y}_{\m{\Omega}}^o$ and that $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}\subset \mathcal{A}_{\m{\Omega}}^1$ are linearly independent, there must exist some $\widetilde f_k\notin\mathcal{T}$, otherwise the two decompositions will be identical and $\widetilde{\m{Y}}=\m{Y}^o$. Consequently, we have the following contradiction:
\equ{\begin{split}
\sum_{k} \widetilde c_k&=\atomn{\widetilde{\m{Y}}}=\inp{\m{V}, \widetilde{\m{Y}}}_{\mathbb{R}}\\
&= \inp{\m{V}, \sum_{k} \widetilde c_k\m{a}\sbra{\widetilde f_k,\widetilde {\m{\phi}}_k}}_{\mathbb{R}} \\
&= \sum_{\widetilde f_k\in\mathcal{T}} \widetilde c_k\inp{Q\sbra{\widetilde f_k}, \widetilde{\m{\phi}}_k}_{\mathbb{R}} + \sum_{\widetilde f_k\notin\mathcal{T}} \widetilde c_k\inp{Q\sbra{\widetilde f_k}, \widetilde{\m{\phi}}_k}_{\mathbb{R}} \\
&< \sum_{\widetilde f_k\in\mathcal{T}} \widetilde c_k + \sum_{\widetilde f_k\notin\mathcal{T}} \widetilde c_k =\sum_{k} \widetilde c_k,
\end{split} \notag}
showing that $\m{Y}^o$ is the unique optimizer. With similar arguments, we can conclude the uniqueness of the atomic decomposition of $\m{Y}^o$.
\end{proof}
\begin{rem} The condition that $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}\subset \mathcal{A}_{\m{\Omega}}^1$ are linearly independent is used to prove the uniqueness of the optimizer to (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}). From this point of view, Proposition II.4 of \cite{tang2012compressed} or at least its proof for the special SMV case is flawed, where this condition is omitted. However, we will show later that it can be satisfied for free when we construct the dual polynomial $Q\sbra{f}$.
\end{rem}
\subsection{Revisiting the SMV Case}
Similarly to the proof of Theorem \ref{thm:completedata}, we first revisit the SMV case in \cite{tang2012compressed}, where a scalar-valued dual polynomial is constructed inspired by \cite{candes2013towards} and proven to satisfy the three constraints listed in Proposition \ref{prop:dualpoly_inc}. While the uniform sampling model is difficult to analyze directly, an equivalent Bernoulli observation model is studied instead following from \cite{candes2006robust}, where we say ``equivalent'' in the sense that the probability that (\ref{formu:ANmin}) or (\ref{formu:AN_sdp}) fails to recover the original signal $\m{Y}^o$ under the uniform model is at most twice of that under the Bernoulli model. Assume in the Bernoulli model that the samples indexed by $\m{J}=\lbra{-2n,\dots,2n}$ are observed independently with probability $p=\frac{M}{4n}$, i.e., we observe about $M$ entries on average. In mathematics, we let $\lbra{\delta_j}_{j\in\m{J}}$ be i.i.d. Bernoulli random variables such that
\equ{\mathbb{P}\sbra{\delta_j=1}=p,}
where $\delta_j=1$ or $0$ indicates whether we observe the $j$th entry in $\m{J}$. It follows that the sampling index set $\m{\Omega}=\lbra{j:\delta_j=1}$.
To deal with the randomness of the observation model, \cite{tang2012compressed} considers a random dual polynomial
\equ{\overline q\sbra{f}=\sum_{f_k\in\mathcal{T}}\alpha_k \overline\mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}}\beta_k \overline\mathcal{K}'\sbra{f-f_k}, \label{formu:randpoly1}}
where $\overline\mathcal{K}\sbra{f}$, a random analog of $\mathcal{K}\sbra{f}$, denotes a random kernel as follows:
\equ{\begin{split}\overline\mathcal{K}\sbra{f}
&= \frac{1}{n+1}\sum_{j\in \m{\Omega}}g_n\sbra{j}e^{-i2\pi jf}\\
&= \frac{1}{n+1}\sum_{j=-2n}^{2n} \delta_jg_n\sbra{j}e^{-i2\pi jf}
\end{split} \label{formu:cKbar}}
with $g_n\sbra{j}$ defined in (\ref{formu:gnj}). It is clear that $\mathbb{E} \overline\mathcal{K}\sbra{f}=p\mathcal{K}\sbra{f}$ and similarly for its derivatives. The proof of \cite{tang2012compressed} is mainly based on that the random kernel $\overline\mathcal{K}\sbra{f}$ is concentrated tightly around its expectation $p\mathcal{K}\sbra{f}$ as the sample size is large enough. As in the complete data case, let us denote $\mbra{\overline D_0}_{jk}=\overline\mathcal{K}\sbra{f_j-f_k}$, $\mbra{\overline D_1}_{jk}=\overline \mathcal{K}'\sbra{f_j-f_k}$ and $\mbra{\overline D_2}_{jk}=\overline\mathcal{K}''\sbra{f_j-f_k}$, and then obtain the system of linear equations
\equ{\begin{bmatrix}\overline{\m{D}}_0 & c_0^{-1}\overline{\m{D}}_1 \\ -c_0^{-1}\overline{\m{D}}_1 & -c_0^{-2}\overline{\m{D}}_2\end{bmatrix} \begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix} = \begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix} \label{formu:linsys2}}
as in (\ref{formu:linsys1}) by imposing conditions as in (\ref{formu:dualpolycons11}) and (\ref{formu:dualpolycons21}), where $c_0$, $\m{\alpha}$, $\m{\beta}$ and $\m{\Phi}$ are defined in the same manner. The authors of \cite{tang2012compressed} show that the coefficient matrix $\overline{\m{D}}=\begin{bmatrix}\overline{\m{D}}_0 & c_0^{-1}\overline{\m{D}}_1 \\ -c_0^{-1}\overline{\m{D}}_1 & -c_0^{-2}\overline{\m{D}}_2\end{bmatrix}$ in (\ref{formu:linsys2}) is concentrated tightly around its expectation $p\m{D}$, where $\m{D}$ is the deterministic analog of $\overline {\m{D}}$. That means, $\overline{\m{D}}$ is close to $p\m{I}$ with high probability, and hence, $\begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix}$ is well defined by (\ref{formu:linsys2}) and close to $p^{-1}\begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}$ with high probability.
According to (\ref{formu:randpoly1}) and (\ref{formu:linsys2}), $\overline q\sbra{f}$ satisfies the interpolation condition (\ref{formu:cons1_inc}) and the support condition (\ref{formu:cons3_inc}). To further prove that it obeys (\ref{formu:cons2_inc}), the authors show that the random perturbation between $\overline q\sbra{f}$ (and its derivatives) and its deterministic analog $q\sbra{f}$ (and its derivatives) in Section \ref{sec:detourcomplete} can be arbitrarily small provided the sample size $\abs{\m{\Omega}}$ is sufficiently large, firstly on a set of grid points and then extended to the continuous domain $\mathbb{T}$. The proof is completed by showing that $\abs{\overline q\sbra{f}}$ (or its second derivative) can be arbitrarily close to $\abs{q\sbra{f}}$ (or its second derivative) on $\mathcal{T}_{\text{far}}$ (or $\mathcal{T}_{\text{near}}$).
We list some useful results in \cite{tang2012compressed} which are also necessary in our later proof. We start with some notations. For $\tau\in\left(0,\frac{1}{4}\right]$, define the event
\equ{\mathcal{E}_{1,\tau}=\lbra{\twon{p^{-1}\overline{\m{D}}-\m{D}}\leq\tau}. \notag}
On $\mathcal{E}_{1,\tau}$, $\overline{\m{D}}$ is guaranteed to be invertible. Then we introduce the partitions $\overline{\m{D}}^{-1}=\begin{bmatrix}\overline{\m{L}} & \overline{\m{R}}\end{bmatrix}$ and $\m{D}^{-1}=\begin{bmatrix}\m{L} & \m{R}\end{bmatrix}$, where $\overline{\m{L}}$, $\overline{\m{R}}$, $\m{L}$ and $\m{R}$ are of the same dimension.
Let
\equ{\overline{\m{v}}_l\sbra{f}=c_0^{-l}\begin{bmatrix} \overline\mathcal{K}^{\sbra{l}}\sbra{f-f_1}^H \\ \vdots \\ \overline\mathcal{K}^{\sbra{l}}\sbra{f-f_K}^H \\ c_0^{-1}\overline\mathcal{K}^{\sbra{l+1}}\sbra{f-f_1}^H \\ \vdots \\c_0^{-1} \overline\mathcal{K}^{\sbra{l+1}}\sbra{f-f_K}^H \end{bmatrix} \label{formu:vl}}
and similarly define its deterministic analog $\m{v}_l\sbra{f}$, where $\overline{\mathcal{K}}^{\sbra{l}}$ denotes the $l$th derivative of $\overline\mathcal{K}$ and $f_j\in\mathcal{T}$.
Let $\delta$ be a small positive number and $C$ a constant which may vary from instance to instance.
\begin{lem} Assume $\Delta_{\mathcal{T}}\geq \frac{1}{n}$. We have $\mathbb{P}\sbra{\mathcal{E}_{1,\tau}}\geq 1-\delta$ if
\equ{M\geq\frac{50}{\tau^2}K\log\frac{2K}{\delta}. \notag} \label{lem:invertibility}
\end{lem}
\begin{lem} Assume $\Delta_{\mathcal{T}}\geq \frac{1}{n}$. Let $\tau\in\left(0,\frac{1}{4}\right]$ and consider a finite set $\mathbb{T}_{\text{grid}}=\lbra{f_d}\subset\mathbb{T}$. Then, we have
\equ{\begin{split}
& \mathbb{P}\left[\sup_{f_d\in\mathbb{T}_{\text{grid}}} \twon{\overline{\m{L}}^H\sbra{\overline{\m{v}}_l\sbra{f_d} -p\m{v}_l\sbra{f_d}}}\right.
\left.\geq 4\sbra{2^{2l+3}\sqrt{\frac{K}{M}}+\frac{n}{M}a\overline{\sigma}_l}, l=0,1,2,3 \right] \\
& \leq 64\abs{\mathbb{T}_{\text{grid}}}e^{-\gamma a^2}+\mathbb{P}\sbra{\mathcal{E}_{1,\tau}^c}
\end{split} \notag}
for some constant $\gamma>0$, where $\overline\sigma_l^2= 2^{4l+1}\frac{M}{n^2}\max\lbra{1,\frac{2^4K}{\sqrt{M}}}$ and $0<a\leq \left\{\begin{array}{ll} \sqrt{2}M^{\frac{1}{4}}, & \text{if } \frac{2^4K}{\sqrt{M}}\geq1, \\ \frac{\sqrt{2}}{4}\sqrt{\frac{M}{K}}, & \text{otherwise.}\end{array}\right.$ \label{lem:I1bound}
\end{lem}
\begin{lem} Assume $\Delta_{\mathcal{T}}\geq \frac{1}{n}$. On the event $\mathcal{E}_{1,\tau}$, we have
\equ{\twon{\sbra{\overline{\m{L}}-p^{-1}\m{L} }^Hp\m{v}_l\sbra{f} } \leq C\tau \notag}
for some constant $C>0$. \label{lem:I2bound}
\end{lem}
\subsection{Proof of Theorem \ref{thm:incompletedata}} \label{sec:proofcompletedata1}
Theorem \ref{thm:incompletedata} is a direct result of the following proposition.
\begin{prop} Under the assumptions of Theorem \ref{thm:incompletedata}, then there exists a numerical constant $C$ such that
\equ{M\geq C\max\lbra{\log^2\frac{\sqrt{L}N}{\delta}, K\log\frac{K}{\delta}\log\frac{\sqrt{L}N}{\delta}} \notag}
is sufficient to guarantee that with probability at least $1-\delta$ there exists a vector-valued polynomial as in (\ref{formu:dualpoly1_inc}) satisfying (\ref{formu:cons1_inc})--(\ref{formu:cons3_inc}) for $\m{J}$ being either $\lbra{-2n,\dots,2n}$ or $\lbra{0,\cdots,N-1}$, where $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}\subset \mathcal{A}_{\m{\Omega}}^1$ are linearly independent. \label{prop:dualpoly1_inc}
\end{prop}
Similarly to the case of complete data, we first argue that we can consider only the symmetric case where $\m{J}=\lbra{-2n,\dots,2n}$. In particular, we attempt to show that, under the assumptions of Proposition \ref{prop:dualpoly1_inc}, if we can construct a vector-valued dual polynomial in the symmetric case, then a similar polynomial can be obtained in the general case.
Our arguments are similar to those in Subsection \ref{sec:proofcompdata} for complete data. We omit the details but emphasize some important features. For a fixed set of frequencies $\mathcal{T}$, $\lbra{\widetilde{\m{\phi}}_k=e^{i2\pi(2n)f_k}\m{\phi}_k}_{k=1}^K$ defined in (\ref{formu:Yotwodecompositions}) are mutually independent, with $\twon{\widetilde{\m{\phi}}_k}=1$ and $\mathbb{E} \widetilde{\m{\phi}}_k=\m{0}$ under the same assumptions of $\lbra{\m{\phi}_k}_{k=1}^K$. Moreover, the linear independence of $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}$ remains unchanged.
The rest of the proof is to show the existence of a dual polynomial as in Proposition \ref{prop:dualpoly1_inc} in the symmetric case. Based on Lemma \ref{lem:invertibility} we can easily obtain the following result, which shows the linear independence of $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}$.
\begin{lem} Assume $\Delta_{\mathcal{T}}\geq \frac{1}{n}$. Then there exists a numerical constant $C$ such that $M\geq CK\log \frac{K}{\delta}$ is sufficient to guarantee that, with probability at least $1-\delta$, $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}$ are linearly independent. \label{lem:linearindependence}
\end{lem}
\begin{proof} According to (\ref{formu:cKbar}), we have
\equ{\begin{split}\mbra{\overline{D}_0}_{jk}
&= \overline \mathcal{K}\sbra{f_j-f_k}\\
&= \frac{1}{n+1}\sum_{l\in \m{\Omega}}g_n\sbra{l}e^{-i2\pi l\sbra{f_j-f_k}}\\
&= \frac{1}{n+1}\m{a}_{\m{\Omega}}\sbra{f_j}^H\m{G}\m{a}_{\m{\Omega}}\sbra{f_k}, \end{split} \notag}
where $\m{G}=\diag\sbra{g_n\sbra{l}}_{l\in\m{\Omega}}$ is a diagonal matrix composed of the entries $\lbra{g_n(l)}$, $l\in\m{\Omega}$, and thus has full rank. Denote $\m{A}_{\m{\Omega}}=\mbra{\m{a}_{\m{\Omega}}\sbra{f_1}, \dots, \m{a}_{\m{\Omega}}\sbra{f_K}}\in\mathbb{C}^{M\times K}$, i.e., the matrix formed by $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}$. It follows that \equ{\m{D}_0=\frac{1}{n+1}\m{A}_{\m{\Omega}}^H\m{G}\m{A}_{\m{\Omega}}. \notag}
On the event $\mathcal{E}_{1,\tau}$, since $\m{D}$, as well as $\m{D}_0$, is invertible, $\m{A}_{\m{\Omega}}$ must have full column rank provided $M\geq K$, i,e., $\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{f_k\in\mathcal{T}}$ are linearly independent. Let us fix $\tau=\frac{1}{4}$. Then the bound of $M$ follows directly from Lemma \ref{lem:invertibility}.
\end{proof}
\begin{rem} For a specific sampling index set $\m{\Omega}$, it is generally infeasible to compute the quantity $\spark\sbra{\mathcal{A}_{\m{\Omega}}^1}$ as mentioned in Section \ref{sec:spark}. Here we show that if $\m{\Omega}$ is selected uniformly at random, then $M\geq O\sbra{K\log K}$ is sufficient to guarantee that with high probability any $K$ atoms in $\mathcal{A}_{\m{\Omega}}^1$ are linearly independent under a frequency separation condition. Define the \emph{conditional spark} of $\mathcal{A}_{\m{\Omega}}^1$ as
\equ{\begin{split}
&\spark_{\text{c}}\sbra{\mathcal{A}_{\m{\Omega}}^1,\Delta}= \inf\left\{\widehat K:\lbra{\m{a}_{\m{\Omega}}\sbra{f_k}}_{k=1}^{\widehat K}\subset \mathcal{A}_{\m{\Omega}}^1 \right.\left. \text{ are linearly dependent with }\Delta_{\lbra{f_k}}\geq \Delta \right\}. \end{split} \notag}
Then we have $\spark_{\text{c}}\sbra{\mathcal{A}_{\m{\Omega}}^1,\frac{1}{n}}\geq K+1$. This will be of independent interest.
\end{rem}
Inspired by \cite{tang2012compressed}, we construct the following vector-valued polynomial
\equ{\overline Q\sbra{f}=\sum_{f_k\in\mathcal{T}}\m{\alpha}_k \overline\mathcal{K}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}}\m{\beta}_k \overline\mathcal{K}'\sbra{f-f_k}, \label{formu:randpoly}}
where $\overline\mathcal{K}$ is the random kernel, and $\m{\alpha}_k$ and $\m{\beta}_k$ are vector-valued coefficients as in the complete data case. It is clear by (\ref{formu:cKbar}) that $\overline Q\sbra{f}$ above satisfies the support condition (\ref{formu:cons3_inc}). To make it satisfy (\ref{formu:cons1_inc}) and (\ref{formu:cons2_inc}), we impose again that
\setlength\arraycolsep{2pt}{\equa{\overline Q\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\m{\alpha}_k \overline\mathcal{K}\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}} \m{\beta}_k\overline\mathcal{K}'\sbra{f_j-f_k} =\m{\phi}_j, \notag\\ \overline Q'\sbra{f_j}
&=& \sum_{f_k\in\mathcal{T}}\m{\alpha}_k \overline\mathcal{K}'\sbra{f_j-f_k} + \sum_{f_k\in\mathcal{T}}\m{\beta}_k\overline\mathcal{K}''\sbra{f_j-f_k} =\m{0},\notag
}}i.e.,
\equ{\overline{\m{D}} \begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix} = \begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix}, \notag}
where $\m{\alpha}=\begin{bmatrix}\m{\alpha}_1^T,\dots,\m{\alpha}_K^T\end{bmatrix}^T\in\mathbb{C}^{K\times L}$ and similarly for $\m{\beta}\in\mathbb{C}^{K\times L}$. It follows that, on the event $\mathcal{E}_{1,\tau}$, $\begin{bmatrix}\m{\alpha}\\ c_0\m{\beta}\end{bmatrix}=\overline{\m{D}}^{-1}\begin{bmatrix}\m{\Phi}\\ \m{0}\end{bmatrix} =\overline{\m{L}}\m{\Phi}$, and
\equ{\begin{split} c_0^{-l}\overline Q^{\sbra{l}}\sbra{f}
&= \sum_{f_k\in\mathcal{T}} \m{\alpha}_k c_0^{-l} \overline\mathcal{K}^{\sbra{l}}\sbra{f-f_k} + \sum_{f_k\in\mathcal{T}} c_0\m{\beta}_k\cdot c_0^{-\sbra{l+1}}\overline\mathcal{K}^{\sbra{l+1}}\sbra{f-f_k} \\
&= \overline{\m{v}}_l\sbra{f}^H\overline{\m{L}}\m{\Phi} \triangleq \inp{\m{\Phi}, \overline{\m{L}}^H \overline{\m{v}}_l\sbra{f}},
\end{split}}
where $\overline{\m{v}}_l\sbra{f}$ is defined in (\ref{formu:vl}).
As in \cite{tang2012compressed}, we decompose $\overline{\m{L}}^H \overline{\m{v}}_l\sbra{f}$ into three parts:
\equ{\begin{split}\overline{\m{L}}^H \overline{\m{v}}_l\sbra{f}
&= \m{L}^H\m{v}_l\sbra{f} +\overline{\m{L}}^H\sbra{\overline{\m{v}}_l\sbra{f}-p\m{v}_l\sbra{f} } + \sbra{\overline{\m{L}}-p^{-1}\m{L}}^Hp\m{v}_l\sbra{f},
\end{split}}
which results in a decomposition on $c_0^{-l}\overline Q^{\sbra{l}}\sbra{f}$:
\equ{\begin{split}c_0^{-l}\overline Q^{\sbra{l}}\sbra{f}
&= \inp{\m{\Phi}, \overline{\m{L}}^H \overline{\m{v}}_l\sbra{f}} \\
&= \inp{\m{\Phi}, \m{L}^H\m{v}_l\sbra{f}} + \inp{\m{\Phi}, \overline{\m{L}}^H\sbra{\overline{\m{v}}_l\sbra{f}-p\m{v}_l\sbra{f} }}\\
&\quad + \inp{\m{\Phi}, \sbra{\overline{\m{L}}-p^{-1}\m{L}}^Hp\m{v}_l\sbra{f}}\\
&= c_0^{-l}Q^{\sbra{l}}\sbra{f} +I_1^l\sbra{f} + I_2^l\sbra{f},
\end{split} \label{formu:decomposition}}
where $Q^{\sbra{l}}\sbra{f}$ is the dual polynomial (\ref{formu:dualpoly}) in the complete data case with $c_0^{-l}Q^{\sbra{l}}\sbra{f}=\inp{\m{\Phi}, \m{L}^H\m{v}_l\sbra{f}}$. Here we have defined
\setlength\arraycolsep{2pt}{\equa{ I_1^l\sbra{f}
&=& \inp{\m{\Phi}, \overline{\m{L}}^H\sbra{\overline{\m{v}}_l\sbra{f}-p\m{v}_l\sbra{f} }}, \notag\\ I_2^l\sbra{f}
&=& \inp{\m{\Phi}, \sbra{\overline{\m{L}}-p^{-1}\m{L}}^Hp\m{v}_l\sbra{f}}.\notag
}}Denote the event
\equ{\mathcal{E}_2=\lbra{\sup_{f_d\in\mathbb{T}_{\text{grid}}} c_0^{-l} \twon{\overline Q^{\sbra{l}}-Q^{\sbra{l}}} \leq \frac{\epsilon}{3}, l=0,1,2,3}. \notag}
We attempt to prove that $\mathcal{E}_2$ happens, i.e., $c_0^{-l}\overline Q^{\sbra{l}}\sbra{f}$ and $c_0^{-l}Q^{\sbra{l}}\sbra{f}$ are close to each other on a fixed grid, with high probability provided $M$ is sufficiently large, which is summarized in the following proposition.
\begin{prop} Suppose $\mathbb{T}_{\text{grid}}\subset \mathbb{T}$ is a finite set of grid points. Under the assumptions of Proposition \ref{prop:dualpoly1_inc}, there exists a numerical constant $C$ such that if
\equ{M\geq C\frac{1}{\epsilon^2}\max\lbra{\log\frac{\abs{\mathbb{T}_{\text{grid}}}}{\delta} \log\frac{L\abs{\mathbb{T}_{\text{grid}}}}{\delta}, K\log\frac{K}{\delta} \log\frac{L\abs{\mathbb{T}_{\text{grid}}}}{\delta}}, \label{formu:Mboundgrid}}
then
\equ{\mathbb{P}\sbra{\mathcal{E}_2}\geq 1-\delta. \notag} \label{prop:boundgrid}
\end{prop}
To prove Proposition \ref{prop:boundgrid}, we need to show that $I_1^l\sbra{f}$ and $I_2^l\sbra{f}$ are small on $\mathbb{T}_{\text{grid}}$. Differently from the SMV case in \cite{tang2012compressed}, however, both $I_1^l\sbra{f}$ and $I_2^l\sbra{f}$ are vector-valued and the difficulty is to provide tight bounds on their norms. To do this, we present the following lemma which corresponds to the vector-form Heoffding's inequality.
\begin{lem}[Vector-form Heoffding's inequality] Let the rows of $\m{\Phi}\in\mathbb{C}^{K\times L}$ be sampled independently on the complex hyper-sphere $\mathbb{S}^{2L-1}$ with zeros means. Then, for all $\m{w}\in\mathbb{C}^K$, $\m{w}\neq\m{0}$, and $t\geq0$,
\equ{\mathbb{P}\sbra{\twon{\inp{\m{\Phi},\m{w}}}\geq t } \leq \sbra{L+1}e^{-\frac{t^2}{8\twon{\m{w}}^2}}. \notag} \label{lem:vectorHeoffding}
\end{lem}
\begin{proof} See Appendix \ref{sec:appendHeoffding}.
\end{proof}
The vector-form Heoffding's inequality generalizes the common scenario with $L=1$ which is used in \cite{tang2012compressed} to bound the scalars $I_1^l$ and $I_2^l$ by utilizing Lemma \ref{lem:I1bound} and Lemma \ref{lem:I2bound}, respectively. Following the same procedures as in \cite{tang2012compressed} with minor modifications of the coefficients, we can bound $I_1^l$ and $I_2^l$ on $\mathbb{T}_{\text{grid}}$ with high probability, which is summarized in the ensuing Lemma \ref{lem:I1} and Lemma \ref{lem:I2} respectively. We omit their proofs and interested readers are referred to the proofs of Lemmas IV.8 and IV.9 in \cite{tang2012compressed}.
\begin{lem} Under the assumptions of Proposition \ref{prop:dualpoly1_inc}, there exists a numerical constant $C$ such that if
\equ{\begin{split}M
&\geq C\max\left\{\frac{1}{\epsilon^2} \max\left( K\log\frac{L\abs{\mathbb{T}_{\text{grid}}}}{\delta}, \right.\right. \left.\left.\log\frac{\abs{\mathbb{T}_{\text{grid}}}}{\delta} \log\frac{L\abs{\mathbb{T}_{\text{grid}}}}{\delta} \right), K\log\frac{K}{\delta}\right\}, \end{split} \notag}
then we have
\equ{\mathbb{P}\lbra{\sup_{f_d\in\mathbb{T}_{\text{grid}}} \twon{I_1^l\sbra{f_d}}\leq\epsilon, l=0,1,2,3 } \geq 1-12\delta. \notag} \label{lem:I1}
\end{lem}
\begin{lem} Under the assumptions of Proposition \ref{prop:dualpoly1_inc}, there exists a numerical constant $C$ such that if
\equ{M\geq C\frac{1}{\epsilon^2} K\log\frac{K}{\delta} \log\frac{L\abs{\mathbb{T}_{\text{grid}}}}{\delta}, \notag}
then we have
\equ{\mathbb{P}\lbra{\sup_{f_d\in\mathbb{T}_{\text{grid}}} \twon{I_2^l\sbra{f_d}}<\epsilon, l=0,1,2,3 } \geq 1-8\delta.\notag} \label{lem:I2}
\end{lem}
Now we can conclude Proposition \ref{prop:boundgrid} by combining (\ref{formu:decomposition}), Lemma \ref{lem:I1} and Lemma \ref{lem:I2} with suitable redefinition of $C$, $\epsilon$ and $\delta$.
We have shown in Proposition \ref{prop:boundgrid} that $c_0^{-l}\overline Q^{\sbra{l}}\sbra{f}$ and $c_0^{-l}Q^{\sbra{l}}\sbra{f}$ are close to each other on a fixed grid $\mathbb{T}_{\text{grid}}\subset\mathbb{T}$ with high probability. We extend the result to the whole circle $\mathbb{T}$ in the following proposition.
\begin{prop} Under the assumptions of Proposition \ref{prop:dualpoly1_inc}, there exists a numerical constant $C$ such that if
\equ{M\geq C\frac{1}{\epsilon^2}\max\lbra{\log\frac{\sqrt{L}n^3}{\epsilon\delta}\log\frac{L^{\frac{3}{2}}n^3}{\epsilon\delta}, K\log\frac{K}{\delta} \log\frac{L^{\frac{3}{2}}n^3}{\epsilon\delta} }, \label{formu:Mboundcontinu}}
then with probability $1-\delta$, we have
\equ{\sup_{f\in\mathbb{T}}c_0^{-l}\twon{\overline Q^{\sbra{l}}\sbra{f} - Q^{\sbra{l}}\sbra{f}}\leq \epsilon, \quad l=0,1,2,3. \notag} \label{prop:appendcontinuous}
\end{prop}
\begin{proof} See Appendix \ref{sec:appendcontinuous}.
\end{proof}
Now we are ready to show that the vector-valued polynomial $\overline Q\sbra{f}$ constructed above satisfies (\ref{formu:cons2_inc}), and so complete the proof together with Lemma \ref{lem:linearindependence}.
\begin{lem} Under the assumptions of Proposition \ref{prop:dualpoly1_inc}, there exists constant $C$ such that if
\equ{M\geq C\max\lbra{\log^2\frac{\sqrt{L}n}{\delta}, K\log\frac{K}{\delta}\log\frac{\sqrt{L}n}{\delta}}, \notag}
then with probability $1-\delta$ we have
\equ{\twon{\overline Q\sbra{f}}<1, \quad f\in\mathbb{T}\backslash\mathcal{T}. \notag}
\end{lem}
\begin{proof}
The proof is based on Proposition \ref{prop:appendcontinuous} and the results in the complete data case in Section \ref{sec:proofcompletedata1}. On $\mathcal{T}_{\text{far}}$, we have
\equ{\twon{\overline Q\sbra{f}} \leq \twon{Q\sbra{f}} + \twon{\overline Q\sbra{f} - Q\sbra{f}} \leq 0.99992+\epsilon. \notag}
So, it suffices to choose $\epsilon=10^{-5}$. On each interval $\mbra{f_j-\nu, f_j+\nu}$, $f_j\in\mathcal{T}$, we also accomplish the proof by showing that
\equ{\frac{1}{2}\frac{{\text{d}^2}\twon{\overline Q\sbra{f}}^2}{\text{d}f^2}=\twon{\overline Q'\sbra{f}}^2 +\Re\lbra{\overline Q''\sbra{f}\overline Q\sbra{f}^H}<0. \notag}
According to the results in Section \ref{sec:proofcompletedata1}, it is easy to verify that
\equ{\twon{c_0^{-l}Q^{\sbra{l}}}\leq C_1, \quad l=0,1,2, \notag}
for a numerical constant $C_1$. Let $\Delta_{Q}=\overline{Q}-Q$. By Proposition \ref{prop:appendcontinuous} we have \equ{\twon{c_0^{-l}\Delta_Q^{\sbra{l}}}\leq \epsilon, \quad l=0,1,2. \notag}
Moreover, with simple derivations we have
\equ{\begin{split}
&\frac{1}{2}\frac{{\text{d}^2}\twon{\overline Q\sbra{f}}^2}{\text{d}f^2} - \frac{1}{2}\frac{{\text{d}^2} \twon{Q\sbra{f}}^2}{\text{d}f^2} \\
&= \Re\lbra{2Q'\Delta_Q'^H + Q''\Delta_Q^H+ \Delta_Q''Q^H+\Delta_Q''\Delta_Q^H} + \twon{\Delta_Q'}^2.
\end{split} \notag}
It follows that
\equ{\begin{split}
&c_0^{-2}\abs{{\frac{1}{2}\frac{{\text{d}^2}\twon{\overline Q\sbra{f}}^2}{\text{d}f^2} - \frac{1}{2}\frac{{\text{d}^2} \twon{Q\sbra{f}}^2}{\text{d}f^2} }} \\
&\leq 2\twon{c_0^{-1}Q'}\twon{c_0^{-1}\Delta_Q'} + \twon{c_0^{-2}Q''}\twon{\Delta_Q}\\
&\quad+ \twon{c_0^{-2}\Delta_Q''}\twon{Q} + \twon{c_0^{-2}\Delta_Q''}\twon{\Delta_Q} + \twon{c_0^{-1}\Delta_Q'}^2 \\
&\leq 4C_1\epsilon + 2\epsilon^2.
\end{split} \notag}
Recall that $c_0^2\leq \frac{8\pi^2}{3}n^2$ as $n\geq2$. So we have
\equ{\begin{split}
\frac{1}{2}\frac{{\text{d}^2}\twon{\overline Q\sbra{f}}^2}{\text{d}f^2}
&\leq \frac{1}{2}\frac{{\text{d}^2}\twon{Q\sbra{f}}^2}{\text{d}f^2} + \abs{{\frac{1}{2}\frac{{\text{d}^2}\twon{\overline Q\sbra{f}}^2}{\text{d}f^2} - \frac{1}{2}\frac{{\text{d}^2} \twon{Q\sbra{f}}^2}{\text{d}f^2} }} \\
&\leq -0.3910n^2 + c_0^2\sbra{4C_1\epsilon + 2\epsilon^2}\\
&\leq \mbra{-0.3910+ \frac{8}{3}\pi^2\sbra{4C_1\epsilon + 2\epsilon^2}} n^2\\
&< 0
\end{split} \notag}
as $\epsilon$ is a sufficiently small numerical value. With this choice of $\epsilon$, we obtain from (\ref{formu:Mboundcontinu}) the following (slightly stronger) condition
\equ{M\geq C\max\lbra{\log^2\frac{\sqrt{L}n}{\delta}, K\log\frac{K}{\delta}\log\frac{\sqrt{L}n}{\delta}}. \notag}
\end{proof}
\section{Numerical Simulations} \label{sec:simulation}
\subsection{Complete Data}
We consider the complete data case and test the frequency recovery performance of the proposed atomic norm method with respect to the frequency separation condition. In particular, we consider two types of frequencies, equispaced and random, and two types of source signals, uncorrelated and coherent. We fix $N=128$ and vary $\Delta_{\text{min}}$ (a lower bound of the minimum separation of frequencies) from $1.05N^{-1}$ (or $0.9N^{-1}$ for random frequencies) to $2N^{-1}$ at a step of $0.05N^{-1}$. In the case of equispaced frequencies, for each $\Delta_{\text{min}}$ we select a set of frequencies $\mathcal{T}$ of the maximal cardinality $\lfloor\Delta_{\text{min}}^{-1}\rfloor$ with the frequency separation $\Delta_{\mathcal{T}}=\frac{1}{\lfloor\Delta_{\text{min}}^{-1}\rfloor}\geq\Delta_{\text{min}}$. In the case of random frequencies, we generate the frequency set $\mathcal{T}$, $\Delta_{\mathcal{T}}\geq\Delta_{\text{min}}$, by repetitively adding new frequencies (generated uniformly at random) till no more can be added. Therefore, any two adjacent frequencies in $\mathcal{T}$ are separate by a value in the interval $[\Delta_{\text{min}},2\Delta_{\text{min}})$. It follows that $\abs{\mathcal{T}}\in\left(\frac{1}{2}\Delta_{\text{min}}^{-1},\Delta_{\text{min}}^{-1}\right]$. We empirically find that $\mathbb{E}\abs{\mathcal{T}}\approx\frac{3}{4}\Delta_{\text{min}}^{-1}$ which is the mid-point of the interval above.
\begin{figure*}
\centering
\subfigure[Equispaced frequencies, uncorrelated sources]{
\label{Fig:Complete_L_UniformFreq}
\includegraphics[width=2.8in]{Fig_Complete_L_UniformFreq.pdf}} %
\subfigure[Random frequencies, uncorrelated sources]{
\label{Fig:Complete_L_NonuniformFreq}
\includegraphics[width=2.8in]{Fig_Complete_L_NonuniformFreq.pdf}}
\subfigure[$L=5$, equispaced frequencies, $\tau$ coherent sources]{
\label{Fig:Complete_CohSource_UniformFreq}
\includegraphics[width=2.8in]{Fig_Complete_CohSource_UniformFreq.pdf}}%
\subfigure[$L=5$, random frequencies, $\tau$ coherent sources]{
\label{Fig:Complete_CohSource_NonuniformFreq}
\includegraphics[width=2.8in]{Fig_Complete_CohSource_NonuniformFreq.pdf}}
\centering
\caption{Results of probability of successful frequency recovery with respect to the minimum frequency separation $\Delta_{\text{min}}$. The vertical dotted lines in (a) indicate the necessary frequency separation condition that $\Delta_{\mathcal{T}}\geq \sbra{1+\frac{1}{2L}}N^{-1}$.} \label{Fig:Completedata}
\end{figure*}
We first consider uncorrelated sources, where the source signals $\m{S}=\mbra{s_{kt}}\in\mathbb{C}^{K\times L}$ in (\ref{formu:observmodel1}) are drawn i.i.d. from a standard complex Gaussian distribution. Moreover, we consider the number of measurement vectors $L=1,3, \text{ and }5$. For each value of $\Delta_{\text{min}}$ and each type of frequencies, we carry out 20 Monte Carlo runs and calculate the success rate of frequency recovery. In each run, we generate a set of frequencies $\mathcal{T}$ and $\m{S}\in\mathbb{C}^{K\times 5}$ and obtain the complete data $\m{Y}^o$. For each value of $L$, we attempt to recover the frequencies by the proposed atomic norm method, implemented by SDPT3 \cite{toh1999sdpt3} in Matlab, based on the first $L$ columns of $\m{Y}^o$. Given the frequency solution $\mathcal{T}^*=\lbra{f_j^*}_{j=1}^{K^*}$ and \emph{mean amplitude} $\lbra{\check c_j^*\triangleq\frac{c_j^*}{\sqrt{L}}}_{j=1}^{K^*}$ (see Section \ref{sec:freqretriev}), we denote by $\check{\m{c}}^*_K$ the vector of the largest $K$ amplitudes (the rest by $\check{\m{c}}^*_{-K}$) and by $\mathcal{T}^*_K$ the corresponding set of frequencies. We may consider the rest of the frequencies in $\mathcal{T}^*$ are spurious peaks with amplitudes $\check{\m{c}}^*_{-K}$. The recovery is considered successful if $K^*\geq K$, root mean squared error (RMSE) of frequency $\frac{\twon{\mathcal{T}^*_K-\mathcal{T}}}{\sqrt{K}}< 1\times 10^{-8}$ and maximum absolute error of amplitude $\inftyn{\begin{bmatrix}\check{\m{c}}_K^*-\check{\m{c}}\\ \check{\m{c}}_{-K}^* \end{bmatrix}}< 1\times 10^{-4}$, where $\check{\m{c}}$ denotes the true vector of amplitudes. Our simulation results are presented in Figs. \ref{Fig:Complete_L_UniformFreq} and \ref{Fig:Complete_L_NonuniformFreq}, which verify the conclusion of Theorem \ref{thm:completedata} that the frequencies can be exactly recovered by the proposed atomic norm method under a frequency separation condition. Moreover, when we take more measurement vectors, the performance of recovery improves and it seems that a weaker frequency separation condition is sufficient to guarantee exact frequency recovery in this case of uncorrelated sources. By comparing Fig. \ref{Fig:Complete_L_UniformFreq} and Fig. \ref{Fig:Complete_L_NonuniformFreq}, we also observe that a stronger frequency separation condition is required in the case of equispaced frequencies where more frequencies are present and located more closely. It is worthy noting that with the equispaced frequencies the minimum separation at which the atomic norm method starts to succeed is close to the necessary separation condition in Remark \ref{rem:necessaryseparation} that $\Delta_{\mathcal{T}}\geq \sbra{1+\frac{1}{2L}}N^{-1}$.
We next consider coherent sources. In this simulation, we fix $L=5$ and consider different percentages, denoted by $\tau$, of the $K$ source signals which are coherent (identical up to a complex scale factor). That is, $\tau=0\%$ refers exactly to the case of uncorrelated sources considered previously. $\tau=100\%$ means that all the sources signals are coherent and the MMV case is equivalent to the SMV case. For each type of frequencies, we consider five values of $\tau$ ranging from $0\%$ to $100\%$ and calculate each success rate over 20 Monte Carlo runs. Our simulation results are presented in Figs. \ref{Fig:Complete_CohSource_UniformFreq} and \ref{Fig:Complete_CohSource_NonuniformFreq}. It is shown that, as the percentage of coherent sources increases, the success rate decreases and a stronger frequency separation condition is required for exact frequency recovery. As $\tau$ equals the extreme value $100\%$, the curves of success rate approximately match those at $L=1$ in Figs. \ref{Fig:Complete_L_UniformFreq} and \ref{Fig:Complete_L_NonuniformFreq}, verifying that taking more measurement vectors does not necessarily improve the performance of frequency recovery. Finally, we report the computational speed of the proposed atomic norm method. It takes about 11s to solve one SDP on average and the CPU times differ slightly at the three values of $L$. About 22 hours are used in total to produce the data used in Fig. \ref{Fig:Completedata}.
\subsection{Incomplete Data}
For incomplete data, we study the so-called phase transition phenomenon in the $\sbra{M,K}$ plane. In particular, we fix $N=128$, $L=5$ and $\Delta_{\text{min}}=1.2N^{-1}$, and study the performance of our proposed atomic norm minimization method in signal and frequency recovery with different settings of the source signal. The frequency set $\mathcal{T}$ is randomly generated with $\Delta_{\mathcal{T}}\geq\Delta_{\text{min}}$ and $\abs{\mathcal{T}}=K$ (differently from that in the last subsection, the process of adding frequencies is terminated as $\abs{\mathcal{T}}=K$). In our simulation, we vary $M=8,12,\dots,128$ and at each $M$, $K=2,4,\dots,\min(M,84)$ since it is difficult to generate a set of frequencies with $K>84$ under the aforementioned frequency separation condition. In this simulation, we consider temporarily correlated sources. In particular, suppose that each row of source signal $\m{S}$ has a Toeplitz covariance matrix $\m{R}\sbra{r}=\begin{bmatrix}1&r&\dots&r^4\\ r&1&\dots&r^3 \\ \vdots & \vdots & \ddots & \vdots\\ r^4 & r^3 & \dots & 1\\\end{bmatrix}\in\mathbb{R}^{5\times5}$ (up to a positive scale factor). Therefore, $r=0$ means that the source signals at different snapshots are uncorrelated while $r=\pm1$ means completely correlated. In our simulation, we first generate $\m{S}_0$ from an i.i.d. standard complex Gaussian distribution and then let $\m{S}\sbra{r}=\m{S}_0\m{R}\sbra{r}^{\frac{1}{2}}$, where we consider $r=0,0.5, 0.9, 1$. For each combination $\sbra{M,K}$, we carry out 20 Monte Carlo runs and calculate the rate of successful recovery with respect to each $r$. The recovery is considered successful if $K^*\geq K$, the relative RMSE of data recovery $\frobn{\m{Y}^*-\m{Y}^o}/\frobn{\m{Y}^o}< 1\times10^{-8}$, the RMSE of frequency recovery $< 1\times 10^{-6}$ and the maximum absolute error of amplitude recovery $<1\times10^{-4}$, where $K^*$ and the last two metrics are defined as in the previous simulation, and $\m{Y}^*$ denotes the solution of $\m{Y}$.
\begin{figure*}
\centering
\subfigure[$r=0$]{
\label{Fig:pt00}
\includegraphics[width=2.5in]{Fig_pt00.pdf}} %
\subfigure[$r=0.5$]{
\label{Fig:pt50}
\includegraphics[width=2.5in]{Fig_pt50.pdf}}
\subfigure[$r=0.9$]{
\label{Fig:pt90}
\includegraphics[width=2.5in]{Fig_pt90.pdf}}%
\subfigure[$r=1$]{
\label{Fig:pt100}
\includegraphics[width=2.5in]{Fig_pt100.pdf}}
\centering
\caption{Results of phase transition with $N=128$ and $\Delta_{\text{min}}=1.2N^{-1}$.} \label{Fig:phasetransition}
\end{figure*}
Our simulation results are presented in Fig. \ref{Fig:phasetransition}, where a transition from perfect recovery to complete failure can be observed in each subfigure. More frequencies can be recovered when more samples are observed. Moreover, when the correlations between the MMVs, indicated by $r$, increase, the phase of successful recovery decreases, and at the same time the phase transition becomes less clear. In the extreme case where $r=1$, all the measurement vectors are completed correlated, which in fact is equivalent to the SMV case. Therefore, by comparing Fig. \ref{Fig:pt100} and the first three subfigures, we can conclude that the performance of frequency recovery improves in general when more measurement vectors are observed. It is also worth noting that $M=128$ corresponds to the complete data case considered in the previous simulation.
We also plot the line $K=\frac{1}{2}\sbra{M+L}$ in Figs. \ref{Fig:pt00}-\ref{Fig:pt90} and $K=\frac{1}{2}\sbra{M+1}$ in Fig. \ref{Fig:pt100} (straight gray lines) which are upper bounds of the sufficient condition in Theorem \ref{thm:AL0_guanrantee} for the atomic $\ell_0$ norm minimization. We see that successful recoveries can be obtained even above the lines, indicating good performance of the proposed atomic norm minimization method. Again, we report the computational speed. It takes about 13s on average to solve each problem and almost 200 hours in total to generate the whole data set used in Fig. \ref{Fig:phasetransition}.
\subsection{The Noisy Case}
\begin{figure*}
\centering
\subfigure[$L=1$]{
\label{Fig:noisycase1}
\includegraphics[width=2.2in]{Fig_noisycase1.pdf}} %
\subfigure[$L=5$, with uncorrelated sources]{
\label{Fig:noisycase2}
\includegraphics[width=2.2in]{Fig_noisycase2.pdf}} %
\subfigure[$L=5$, with sources 1 and 3 coherent]{
\label{Fig:noisycase3}
\includegraphics[width=2.2in]{Fig_noisycase3.pdf}}
\centering
\caption{Frequency recovery/estimation of atomic norm minimization (ANM) and MUSIC in the presence of noise, with (a) $L=1$, (b) $L=5$ and uncorrelated sources, and (c) $L=5$ and coherent sources.} \label{Fig:noisycase}
\end{figure*}
While this paper has been focused on the noiseless case, one naturally wonders the performance of the proposed method in the practical noisy case. We provide a simple simulation as follows to show this. We set $N=50$, $M=20$ with $\m{\Omega}$ randomly generated, $K=3$ sources with frequencies of $0.1$, $0.12$ and $0.3$ and powers of 2, 3 and 1 respectively, and $L=5$. The signals of each source are generated with constant amplitude and random phases. Complex white Gaussian noise is added to the observed samples with noise variance $\sigma^2=0.1$. We attempt to denoise the observed noisy signal $\m{Y}^o_{\m{\Omega}}$ and recover the frequency components by solving the following optimization problem:
\equ{\min_{\m{Y}} \norm{\m{Y}}_{\mathcal{A}}, \text{ subject to } \frobn{\m{Y}^o_{\m{\Omega}}-\m{Y}_{\m{\Omega}}}^2\leq\epsilon, }
where $\epsilon$, set to $\sbra{ML+2\sqrt{ML}}\sigma^2$ (mean + twice standard deviation), upper bounds the Frobenius norm of the noise with large probability. Our simulation results of one Monte Carlo run are presented in Fig. \ref{Fig:noisycase}. The SMV case is studied in Fig. \ref{Fig:noisycase1}, where only the first measurement vector is used for frequency recovery. It is shown that the three frequency components are correctly identified by the atomic norm minimization method while MUSIC fails. The MMV case is studied in Fig. \ref{Fig:noisycase2} with uncorrelated sources, where both atomic norm minimization and MUSIC succeed to identify the three frequency components. The coherent source case is presented in Fig. \ref{Fig:noisycase3}, where we modify source 3 in Fig. \ref{Fig:noisycase2} such that it is coherent with source 1. MUSIC fails to detect the coherent sources as expected while the proposed method still performs well. All of the three subfigures show that spurious frequency components can be present using the atomic norm minimization method, however, their powers are insignificant. To be specific, the spurious components have about $0.4\%$ of the total powers in Fig. \ref{Fig:noisycase1}, and this number is on the magnitude of $10^{-6}$ in the latter two subfigures. Since the numerical results imply that the proposed method is robust to noise, a theoretical analysis should be investigated in future studies. The proposed method takes about $1.5$s in each scenario. Finally, it is worthy noting that the proposed method requires to know the noise level while MUSIC needs the source number.
\section{Conclusion} \label{sec:conclusion}
We studied the joint sparse frequency recovery problem in this paper which arises in practical array processing applications. We presented nonconvex and convex optimization methods for its solution and analyzed their theoretical guarantees under no or very weak assumptions of the source signals. Our results extend the SMV atomic norm methods and their theoretical guarantees in \cite{candes2013towards,tang2012compressed} to the MMV case, extend the existing discrete joint sparse recovery framework to the continuous dictionary setting, and provide theoretical guidance for the array processing applications. While this paper is focused on the worst case analysis, it will be interesting to investigate in the future the average case under stronger assumptions of the source signals, in which numerical simulations of this paper suggest that a weaker frequency separation condition is sufficient for exact recovery as the number of measurement vectors increases.
|
\section{Introduction and Results}
In the study of dynamical systems, establishing the rate of mixing of a given system is of foremost importance.
It is a fundamental property describing the rate at which information about the system is lost.
More importantly the rate of mixing (or typically slightly stronger information which is obtained whilst proving the rate of mixing) can be used to prove many other statistical properties (see, for example \cite[\S9]{Keller:1989} and \cite[Chapter~7]{MR2229799}).
Furthermore, of physical relevance, these strong results associated to good rates of mixing are crucially used when studying weakly coupled systems~\cite{MR2530174,MR2842975}.
Rate of mixing results were first obtained for expanding maps and for hyperbolic maps (see \cite{Li1995} and references within),
then also for slower mixing, non-uniformly hyperbolic systems (e.g.,~\cite{Young1998, Young1999, MR1946554}).
In the case of hyperbolic flows or skew products like the one studied here one direction is completely neutral, with no expansion or contraction. These systems are not hyperbolic but merely partially hyperbolic.
In these situations there is a mechanism at work, different to hyperbolicity, but which is nonetheless sufficient for producing good statistical properties including exponential rate of mixing.
Dolgopyat~\cite{D}, extending work of Chernov~\cite{C}, succeeded in developing technology for studying this neutral mechanism and consequently proved exponential mixing for mixing Anosov flows when the stable and unstable invariant foliations are both $\cC^{1}$.
Using and developing these ideas various results followed~\cite{Li1, baladi2005edc, avila2005emt,Ts}.
However all the above systems were rather smooth or at least Markov.
Our knowledge concerning this same neutral mechanism in systems with discontinuities is less than satisfactory at present.
As far as the authors are aware, results of exponential mixing for hyperbolic flows with discontinuities are limited at present to the work of Baladi and Liverani~\cite{Baladi:2011uq} for piecewise smooth 3D hyperbolic flows which preserve a contact structure and the work of Obayashi~\cite{Obayashi:2009} in the case of suspension semiflows over expanding maps with discontinuities which admit a Young tower.
\begin{samepage}
In this article we study the 2D skew product map $F:\bT^{2} \to \bT^{2}$ defined by
\[
F: (x,u) \mapsto (f(x), u+\tau(x)).
\] \nobreak
The base map $f : \bT^{1} \to \bT^{1}$ is required to be $\cC^{2}$ except for a finite number of discontinuities and admit a $\cC^{2}$ extension to the closure of the intervals of smoothness.
\end{samepage}
Also $f$ is required to be uniformly expanding and covering.\footnote{Covering implies that the unique absolutely continuous invariant probability density is bounded away from zero~\cite{Liverani:1995aa}.}
The fibre map $\tau : \bT^{1} \to \bR$ is similarly required to be $\cC^{2}$ except for a finite number of discontinuities and admit a $\cC^{2}$ extension to the closure of the intervals of smoothness.\footnote{Here and throughout the document, if $u\in \bT^{1}$, $s\in \bR$ then we consider $u+s\in \bT^{1}$ in the natural sense that $\bT^{1} = \bR \diagup \bZ$.}
At no stage do we require the map to be Markov, nor do we work with tower constructions to reduce to the Markov case.
Since the map $f$ is piecewise $\cC^{2}$ and uniformly expanding it is known that there exists $\nu$ an $f$-invariant probability measure which is absolutely continuous with respect to Lebesgue.
Since the dynamics in the fibres is nothing more than a rigid rotation this means that $\mu := \nu \times \operatorname{Leb}$ is an $F$-invariant probability measure on $\bT^{2}$.
Given observables $g,h : \bT^{2} \to \bC$ the correlation is defined as usual
$\operatorname{Cor}_{g,h}(n)
:= \mu( g \cdot h \circ F^{n})
- \mu(g) \cdot \mu(h)$.
We say that $F:\bT^{2} \to \bT^{2}$ mixes exponentially if there exists $\zeta>0$ and for each pair of H\"older continuous observables $g,h$ there exists $C_{g,h}>0$ such that
$ \abs{ \operatorname{Cor}_{g,h}(n) }
\leq
C_{g,h}e^{-n\zeta}$, for all $n\in \bN$.
Our main result is the following.
\begin{thmm}
\label{thm:main}
Let $F: \bT^{2}\to \bT^{2}$ be a piecewise-$\cC^{2}$ skew product over an expanding map as described above.
Either $F$ mixes exponentially
or
there exists Lipschitz $\theta : \bT^{1} \to \bR$ and piecewise constant $\chi: \bT^{1} \to \bR$ such that
$\tau = \theta \circ f - \theta + \chi$. The discontinuites of $\chi$ only occur at points where either $f$ or $\tau$ are discontinuous.
\end{thmm}
\noindent
The remainder of this document is devoted to the proof of the above theorem.
The basic idea is from Dolgopyat~\cite{D}.
However we combine the best technology from the susequent articles~\cite{Li1,baladi2005edc,avila2005emt,Ts} in order to deal with the present difficulties, in particular the problems arising from the discontinuities.
We note that the issue of the discontinuities could in theory be approached by using a tower construction and so reducing to the case of a base map which is Markov. This has been done by Obayashi~\cite{Obayashi:2009} in the case of suspension semiflows over expanding maps with discontinuities. However one particular problem with such tower constructions is that the tower is very sensitive to changes in the underlying system and so important questions, for example determining the behaviour of statistical properties under perturbation of the original system, become completely unapproachable.
From a technical point of view we are forced in two opposing directions. To deal with discontinuities we are forced to consider densities of rather low regularity. However we also need to take advantage of Dolgopyat's oscillatory cancellation argument which requires some good degree of regularity for the density.
The result for skew products is closely related to the analogous result for suspension semiflows. At a techincal level this can be seen from the twisted transfer operator (introduced below~\eqref{eq:twisttrans}) which is the same object used when studied in the context of skew products or flows (see, for example, \cite{oli1203}), with exactly the same estimates being required.
In section Section~\ref{sec:trans} the key notion of transversality is discussed and a certain estimate is shown to hold in the case when $\tau$ is not cohomologous to a piecewise constant. Section~\ref{sec:preparing} concerns the estimate of the norm of twisted transfer operators reducing the problem to a single key estimate (Proposition~\ref{prop:mainlem}). This key estimate is proven in Section~\ref{sec:mainestimate}, crucially using the transversality estimate from Section~\ref{sec:trans}. Finally, in Section~\ref{sec:rateofmixing}, the estimate on the twisted transfer operators is used to produce an estimate of exponential mixing.
\section{Transversality}
\label{sec:trans}
From this point onwards we will assume that $\tilde\lambda: = \inf f' > 2$. In general it would suffice to assume that there exists $n\in \bN$ such that $\inf \, (f^{n})' >1$. In that case we would simply consider a sufficiently large iterate $m$ such that $\inf \, (f^{m})' > 2$ and proceed as now.
Let $\Lambda := \sup \abs{f'} \geq \tilde\lambda$.
We may assume that $ \sup \abs{\tau'} >0$ since if this does not hold then $\tau$ is actually equal to a piecewise constant function, and in particular cohomologous to a piecewise constant function.
The first step is to define a forward invariant unstable conefield.
Let
\[
C_{1} := \frac{ 2 \sup \abs{\tau'}}{\tilde \lambda - 1 } >0.
\]
Define the constant conefield with the cones $\cK = \{ \bigl( \begin{smallmatrix} \alpha\\ \beta \end{smallmatrix} \bigr): \abs{ \smash{\frac{\beta}{\alpha}}} \leq C_{1}\}$. This conefield is strictly invariant under
\[
DF({x}) = \bigl(\begin{smallmatrix}
f'(x) & 0\\ \tau'(x) & 1
\end{smallmatrix} \bigr).
\]
To see the invariance note that
$DF(x): \bigl( \begin{smallmatrix} \alpha\\ \beta \end{smallmatrix} \bigr) \mapsto \bigl( \begin{smallmatrix} \alpha'\\ \beta' \end{smallmatrix} \bigr) $ where ${\smash{\frac{\beta'}{\alpha'}}} = ( {\tau'}(x) + {\smash{\frac{\beta}{\alpha}}} )/f'(x)$.
Let $x_{1}, x_{2} \in \bT^{1}$ be two preimages of some $y\in \bT^{1}$, i.e., $f^{n}(x_{1}) = f^{n}(x_{2}) =y$. We write $x_{1} \pitchfork x_{2}$ (meaning \emph{transversal}) if $DF^{n}_{x_{1}} \cK \cap DF^{n}_{x_{2}} \cK = \{0\}$. Note that this transversality depends on $n$ even though the dependence is suppressed in the notation.
For future convenience let $J_{n} := {\abs{(f^{n})'}}^{-1}$.
Define the quantity
\[
\varphi(n)
:=
\sup_{y \in \bT^{1}} \ \ \sup_{x_{1}\in f^{-n}(y)} \sum_{ \substack{x_{2} \in f^{-n}(y) \\ x_{1} \not\pitchfork x_{2}} } J_{n}(x_{2}).
\]
This crucial quantity gives control on the fraction of preimages which are not transversal.
In this section we prove the following which is an extension of Tsujii~{\cite[Theorem~1.4]{Ts}} to the present situation where discontinuities are permitted.
Much of the argument follows the reasoning of the above mentioned reference with some changes due to the more general setting.
\begin{prop}
\label{prop:transversality}
Let $F:(x,u) \mapsto (f(x), u+\tau(x))$ be a piecewise-$\cC^{2}$ skew product over an expanding base map as above.
\textbf{Either:}
\begin{equation}
\label{eq:transversality}
\limsup_{n\to \infty} \varphi(n)^{\frac{1}{n}}
<1,
\end{equation}
\textbf{Or:}
There exists Lipschitz $\theta : \bT^{1} \to \bR$ and piecewise constant $\chi: \bT^{1} \to \bR$ such that
$\tau = \theta \circ f - \theta + \chi$. Moreover the discontinuities of $\chi$ only occur at points where either $f$ or $\tau$ are discontinuous.
\end{prop}
Before proving the above, let us record a consequence of the transversality.
Let $\tau_{n} := \sum_{j=0}^{n-1} \tau \circ f^{j}$.
\begin{lem}\label{lem:transverse}
If $f^{n}(x_{1}) = f^{n}(x_{2})$ and $x_{1} \pitchfork x_{2}$
then
\[
\abs{({\tau_{n}'} \cdot {J_{n}})(x_{1}) -({\tau_{n}'} \cdot {J_{n}})(x_{2})}
> {C_{1}} ( J_{n}(x_{1})+ J_{n}(x_{2}) ).
\]
\end{lem}
\begin{proof}
Assume that $\frac{\tau_{n}'}{(f^{n})'}(x_{1}) \geq \frac{\tau_{n}'}{(f^{n})'}(x_{2})$, the other case being identical.
Note that
\[
DF^{n}({x_{1}})\bigl(\begin{smallmatrix}1\\ -C_{1}\end{smallmatrix}\bigr) = \left(\begin{smallmatrix}(f^{n})'(x_{1})\\ \tau_{n}'(x_{1}) -C_{1}\end{smallmatrix}\right),
\quad
DF^{n}({x_{2}})\bigl(\begin{smallmatrix}1\\ C_{1}\end{smallmatrix}\bigr) = \left(\begin{smallmatrix}(f^{n})'(x_{2})\\ \tau_{n}'(x_{2}) + C_{1}\end{smallmatrix}\right).
\]
Transversality implies that
$({\tau_{n}'(x_{1}) -C_{1} })/{( f^{n})'(x_{1}) } > ({\tau_{n}'(x_{2}) + C_{1}}) / {( f^{n})'(x_{2}) }$.
\end{proof}
The remainder of this section is devoted to the proof of Proposition~\ref{prop:transversality}.
As mentioned in the introduction it is known that there exists an $f$-invariant probability measure $\nu$ which is absolutely continuous with respect to Lebesgue.
Let $h_{\nu}$ denote the density of $\nu$.
It is convenient to introduce the quantity
\begin{equation}
\label{eq:defphitilde}
\tilde\varphi(n,L,y)
:=
\sum_{ \substack{x \in f^{-n}(y) \\ DF^{n}(x)\cK \supset L} } J_{n}(x) \cdot \frac{h_{\nu}(x)}{h_{\nu}(y)}
\end{equation}
where $L \in \bR\bP^{1}$ (an element of real projective space, i.e., a line in $\bR^{2}$ which passes through the origin).
Let $\tilde\varphi(n) := \sup_{y} \sup_{L} \tilde\varphi(n,L,y)$.
The benefit of this definition is that $\tilde\varphi(n) $ is submultiplicative,
i.e., $\tilde\varphi(n+m) \leq \tilde\varphi(n) \tilde\varphi(m)$ for all $n,m\in \bN$;
and $\tilde\varphi(n) \leq 1$ for all $n\in \bN$.
\begin{lem}
\label{lem:allequivalent}
The following statements are equivalent.
\begin{enumerate}[label={(\roman*)}, font=\normalfont]
\item
$\displaystyle\limsup_{n\to\infty} \varphi(n)^{\frac{1}{n}} =1$
\item
$\displaystyle\lim_{n\to\infty} \tilde\varphi(n)^{\frac{1}{n}} =1$,
\item
For all $n\in \bN$, $y\in \bT^{1}$ there exists $L_{n}(y) \in \bR\bP^{1}$ such that,
for every $x \in f^{-n}(y)$,
$DF^{n}(x) \cK \supset L_{n}(y)$,
\item
There exists a measurable $F$-invariant unstable direction,
i.e., there exists $\ell: \bT^{1} \to \bR$ such that
$\tau' = f' \cdot \ell \circ f - \ell$
and so
\[
DF(x) \left(\begin{smallmatrix} 1 \\ \ell(x) \end{smallmatrix}\right)
= f'(x) \left(\begin{smallmatrix} 1 \\ \ell\circ f(x) \end{smallmatrix}\right).
\]
\item
Statement \textnormal{(iv)} holds with $\ell$ of bounded variation.
\item
There exists $ \theta : \bT^{1} \to \bT^{1}$ such that $\tau - \theta\circ f + \theta$ is piecewise constant (discontinuities only where either $f$ or $\tau$ are discontinuous). Moreover $\theta$ is differentiable with derivative of bounded variation.
\end{enumerate}
\end{lem}
Since $\limsup_{n\to\infty} \varphi(n)^{\frac{1}{n}} \leq 1$
the above lemma immediately implies Proposition~\ref{prop:transversality}.
In the remainder of this section we prove the above lemma.
First a simple fact that we will use repeatedly.
\begin{lem}\label{lem:tauprime}
$\abs{J_{n} \cdot \tau_{n}'} \leq \frac{1}{2} C_{1}$.
\end{lem}
\begin{proof}
First observe that
$\tau_{n}' = \sum_{i=0}^{n-1} \tau' \circ f^{i} \cdot (f^{i})'$.
Consequently $\abs{J_{n} \cdot \tau_{n}'} \leq \abs{\tau'} \sum_{i=0}^{n-1} \tilde\lambda^{-i}$.
For all $n\in \bN$ the sum $\sum_{i=0}^{n-1} \tilde\lambda^{-i}$ is bounded from above by $(\tilde\lambda - 1)^{-1}$.
And so, using also the definition of $C_{1}$,
we know that $\abs{J_{n} \cdot \tau_{n}'} \leq {\sup \abs{\tau'} }/ ({\tilde\lambda - 1}) = \frac{1}{2} C_{1}$.
\end{proof}
\begin{proof}[\bfseries Proof of (i) $\Longrightarrow$ (ii)]
Suppose that $m \in \bN$, $n = n(m) = \lceil 2 \frac{ \ln \Lambda }{ \ln \tilde\lambda} m \rceil$,
$y\in \bT^{1}$ and $x_{1}, x_{2} \in f^{-n}(y)$.
Note that $n > m$ since $\Lambda \geq \tilde\lambda$.
Let $p=n-m$.
Further suppose that
\[
DF^{n}(x_{1})\cK \cap DF^{n}(x_{2})\cK \neq \{0\}.
\]
The slopes of the edges of $DF^{n}(x_{1})\cK$ are
$\frac{\tau_{n}'}{(f^{n})'}(x_{1}) \pm C_{1} J_{n}(x_{1})$.
Let
\[
L(x_{1}):= DF^{n}(x_{1})( \bR \times \{0\}).
\]
The slope of $L$ is $\frac{\tau_{n}'}{(f^{n})'}(x_{1})$.
Since we assume the cones $DF^{n}(x_{1})\cK$ and $ DF^{n}(x_{2})\cK$ are not transversal this implies that
the difference in slope between one of the edges of $DF^{n}(x_{2})\cK$ and $L$ is not greater than
\begin{equation}
\label{eq:rain}
C_{1} J_{n}(x_{1}) \leq C_{1} \tilde\lambda^{-n}.
\end{equation}
Now consider the cone $DF^{n}(x_{2})\cK$ and the cone $DF^{m}(f^{p}x_{2})\cK \supset DF^{n}(x_{2}) \cK$.
The slopes of the edges of the first are
\[
\frac{\tau_{n}'}{(f^{n})'}(x_{2}) \pm C_{1} J_{n}(x_{2})
= \frac{\tau_{m}'}{(f^{m})'}\circ f^{p}(x_{2}) + \frac{\tau_{p}'}{(f^{m})'\circ f^{p} \cdot (f^{p})'}(x_{2}) \pm C_{1} J_{n}(x_{2}),
\]
whilst the slopes of the edges of the second are
\[
\frac{\tau_{m}'}{(f^{m})'}\circ f^{p}(x_{2}) \pm C_{1} J_{m}\circ f^{p}(x_{2}).
\]
Consequently the slopes of the edges of the two cones are separated by at least
\[
J_{m}\circ f^{p}(x_{2}) \left( C_{1} - \sup \abs{\tau_{p}' \cdot J_{p}} \right) - C_{1} J_{n}(x_{2}).
\]
By Lemma~\ref{lem:tauprime} we know that $\abs{\tau_{p}' \cdot J_{p}} \leq \frac{1}{2} C_{1}$.
This means that the above term is bounded from below by
\[
\tfrac{1}{2} C_{1} \Lambda^{-m} - C_{1} \tilde\lambda^{-n}
\geq \tfrac{1}{2} C_{1} \tilde\lambda^{-\frac{n}{2}} - C_{1} \tilde\lambda^{-n},
\]
where we used that the assumed relation between $n$ and $m$ implies that $m \leq \frac{n}{2} \frac{\ln \lambda}{\ln \Lambda}$ and so $\Lambda^{-m} \geq \tilde\lambda^{-\frac{n}{2}}$.
Recall now~\eqref{eq:rain}. For all $n$ sufficiently large then
$ \tfrac{1}{2} C_{1} \tilde\lambda^{-\frac{n}{2}} - C_{1} \tilde\lambda^{-n}
\geq C_{1} \tilde\lambda^{-n}$.
To conclude, we have shown that
$DF^{n}(x_{1})\cK \cap DF^{n}(x_{2})\cK \neq \{0\}$
implies that
$DF^{m}(f^{p}x_{2})\cK \supset L(x_{1})$ where $L(x_{1})$ is defined as before.
This means that
\[
\sum_{ \substack{x_{2} \in f^{-n}(y) \\ x_{1} \not\pitchfork x_{2}} } J_{n}(x_{2})
\leq
\sup_{L} \sum_{ \substack{x \in f^{-m}(y) \\ DF^{m}(x)\cK \supset L} } J_{m}(x) .
\]
Finally this implies that $\varphi(n) \leq C_{2} \tilde \varphi(m(n))$, where $C_{2}:= {\sup h_{\nu}}/{\inf h_{\nu}} >0$.
\end{proof}
\begin{proof}[\bfseries Proof of (ii) $\Longrightarrow$ (iii)]
By submultiplicativity and the fact that $\tilde \varphi(n) \leq 1$ for all $n\in \bN$ the assumption $\lim_{n\to\infty} \tilde\varphi(n)^{\frac{1}{n}} =1$
implies that $ \tilde\varphi(n) =1$ for all $n\in \bN$.
Consequently the following statement holds:
\begin{quotation}
For all $n$ there exists $y_{n}\in \bT^{1}$ and $L_{n} \subset \bR\bP^{1}$
such that, for all $x\in f^{-n}(y_{n})$,
$DF^{n}(x)\cK \supset L_{n}$.
\end{quotation}
It remains to prove that this above statement implies statement~(iii).
We will prove the contrapositive. Suppose the negation of statement~(iii). I.e.,
there exists $n_{0}\in \bN$, $y\in \bT^{1}$, $x_{1}, x_{2} \in f^{-n_{0}}(y)$ such that $DF^{n_{0}}(x_{1}) \cK \cap DF^{n_{0}}(x_{2}) \cK = \{0\} $.
Let $g_{1}, g_{2}$ denote the two inverse maps corresponding to $x_{1},x_{2}$. These inverses are defined on some interval containing $y$ and due to the openness of the transversality of cones we can assume that
$DF^{n_{0}}(g_{1}(y)) \cK \cap DF^{n_{0}}(g_{2}(y)) \cK = \{0\} $ for all $y \in \omega_{*}$ where $\omega_{*} \subset \bT^{1}$ is an open interval.
Since $f$ is covering there exists $m_{0}\in \bN$ such that
$ f^{m_{0}}(\omega_{*})
=\bT^{1}$.
Let $m= m_{0}+n_{0}$.
For all $y\in \bT^{1}$ there exists $z\in f^{-m_{0}}(y)$ and there exists $x_{1}, x_{2} \in f^{-n_{0}}(z)$ with the above transversality property.
This means that for all $y\in \bT^{1}$ there exist $x_{1}, x_{2} \in f^{-m}(y)$ such that
$DF^{m}(x_{1}) \cK \cap DF^{m}(x_{2}) \cK = \{0\} $, since
\[
DF^{m}(x_{1}) \cK \cap DF^{m}(x_{2}) \cK
= DF^{m_{0}}(y) ( DF^{n_{0}}(x_{1}) \cK \cap DF^{n_{0}}(x_{2}) \cK).
\]
This contradicts the above statement concerning the existence of some $L_{n}$ such that $DF^{n}(x)\cK \supset L_{n}$ for all $x\in f^{-n}(y)$.
\end{proof}
\begin{proof}[\bfseries Proof of (iii) $\Longrightarrow$ (iv)]
For all $x\in \bT^{1}$ let $\ell_{n}(x) $ denote the slope of $L_{n}(x)$.
I.e., $\left(\begin{smallmatrix}
1 \\ \ell_{n}(x)
\end{smallmatrix}\right)
\in L_{n}(x)$.
The uniform expansion means that the image of unstable cones contracts and consequently for each $x$ then $\ell_{n}(x) \to \ell(x)$ as $n\to \infty$.
The function $\ell(x)$
enjoys the property that $\tau_{n}'(x) + \ell(x) = (f^{n})'(x) \cdot \ell(f^{n}x)$.
\end{proof}
\begin{proof}[\bfseries Proof of (iv) $\Longleftrightarrow$ (v)]
The implication (v) $\Longrightarrow$ (iv) is immediate.
Assume that statement (iv) holds.
Since $\ell$ is invariant
we know that for any $n\in \bN$, $x \in f^{-n}(y)$ that
\[
\ell(y) = \frac{ \tau_{n}'}{(f^{n})'} (x)+ \frac{\ell}{(f^{n})'} (x).
\]
For large $n$ the second term on the right hand side becomes very small.
Note that, because we assume (iv) holds, if we want to calculate $\ell$ at $y$ it does not matter which preimage $x$ we consider.
Fix some $\omega_{0} \subset \bT^{1}$ a disjoint union of intervals and a bijection $g: \omega_{0} \to \bT^{1} $ such that $f \circ g$ is the identity. We can do this in such a way that $g$ is $\cC^{2}$ on each component of $\omega_{0}$\footnote{Note that if the map $f$ was full branch we could choose $g$ to be $\cC^2$ but this cannot be expected in general.}.
Of course $f^{n} \circ g^{n} = \mathbf{id}$ for all $n\in \bN$.
Consequently
\[
\begin{aligned}
\ell & = \frac{\tau_{n}'}{(f^{n})'}\circ g^{n} + \frac{\ell}{(f^{n})'}\circ g^{n}
\\
& =
\sum_{j=0}^{n-1}
\frac{\tau'}{(f^{n-j})'}\circ g^{n-j}
+ \frac{\ell}{(f^{n})'}\circ g^{n}.
\end{aligned}
\]
Note that
$\norm{\smash{\frac{\ell}{(f^{n})'}\circ g^{n}}}_{\L{\infty}{\bT^{1}}} \to 0$
as $n\to\infty$.
Also note that $ \sum_{j=0}^{\infty}
\frac{\tau'}{(f^{j})'}\circ g^{j}$ is of bounded variation. Indeed each term in this infinite sum is piecewise $\cC^{2}$ and has only a finite number of discontinuities.
Moreover the $\mathbf{BV}$ norm of the terms is exponentially decreasing due to the uniform expansion and so the sum converges in $\mathbf{BV}$.
Consequently $\ell$ must be of bounded variation.
\end{proof}
\begin{proof}[\bfseries Proof of (v) $\Longleftrightarrow$ (vi)]
First we prove (v) $\Longrightarrow$ (vi).
For all $y\in \bT^{1}$ let
\[
\theta(y) := \int_{0}^{y} \ell(x) \ dx.
\]
This defines a Lipschitz function on $\bT^{1}$, differentiable in the sense that the derivative is of bounded variation.
There exists a partition ${\{\omega_{m}\}}_{m}$ such that $\tau$ and $f$ are $\cC^{2}$ when restricted to each element of the partition.
Write $\omega_{m} = (a_{m}, b_{m})$.
If $y\in \omega_{m}$ then
$\tau(y) = \tau(a_{j}) + \int_{a_{j}}^{y} \tau'(x) \ dx$.
Substituting the equation $\tau' = f' \cdot \ell \circ f - \ell$ we obtain
\[
\begin{aligned}
\tau(y)
&= \tau(a_{j})
+ \int_{a_{j}}^{y} f' \cdot \ell \circ f(x) \ dx - \int_{a_{j}}^{y} \ell(x) \ dx\\
&= \tau(a_{j})
+ \int_{f(a_{j})}^{f(y)} \ell (x) \ dx - \int_{a_{j}}^{y} \ell(x) \ dx \\
&= \theta \circ f(y) - \theta(y)
+\left( \theta(a_{j}) - \theta \circ f(a_{j})
+ \tau(a_{j})\right)\\
& = \theta\circ f(y) - \theta(y) + \chi_{j}.
\end{aligned}
\]
Let $\chi$ denote the piecewise constant function equal to $\chi_{j}$ on each $\omega_{j}$.
The implication (vi) $\Longrightarrow$ (v) follows by differentiating $\tau - \theta\circ f + \theta = \chi$.
\end{proof}
\begin{proof}[\bfseries Proof of (iv) $\Longrightarrow$ (i)]
The vector $\left(\begin{smallmatrix}
1 \\ \ell(y)
\end{smallmatrix} \right)$
is contained within $DF^{n}(x)\cK$ for all $x\in f^{-n}(y)$ since
$\left(\begin{smallmatrix}
1 \\ \ell(x)
\end{smallmatrix} \right) \in \cK$
and $DF^{n}(x) \left(\begin{smallmatrix}
1 \\ \ell(x)
\end{smallmatrix} \right)
= (f^{n})'(x)
\left(\begin{smallmatrix}
1 \\ \ell(y)
\end{smallmatrix} \right)$.
Consequently
$x_{1} \not\pitchfork x_{2}$, i.e., $ DF^{n}(x_{1})\cK \cap DF^{n}(x_{2})\cK \neq \{0\}$, for every $x_{1}, x_{2} \in f^{-n}(y)$.
\end{proof}
\section{Preparation for the Main Estimate}
\label{sec:preparing}
Throughout this section and the next we assume that the first alternative of Proposition~\ref{prop:transversality} holds.
Let $\gamma_{1} := \limsup_{n\to \infty} \varphi(n)^{\frac{1}{n}}
<1$,
and fix $\gamma \in (0, \gamma_{1})$.
There exists $C_{\gamma}>0$ such that
\begin{equation}
\label{eq:transdecay}
\varphi(n) \leq C_{\gamma} e^{-n\gamma}
\quad
\text{for all $n\in \bN$}.
\end{equation}
The twisted transfer operator, for all $b\in \bR$, $n\in \bN$ is given by the formula
\begin{equation}
\label{eq:twisttrans}
\cL_{b}^{n}h(y) = \sum_{x\in f^{-n}(y)} J_{n} (x) \cdot h(x)\cdot e^{ib \tau_{n}(x)}.
\end{equation}
A simple estimate shows that $ \norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} } \leq \norm{h}_{\L{1}{\bT^{1}} }$.
We will work extensively with functions of \emph{bounded variation} due to the suitability of this function space for discontinuities. The Banach space is denoted $(\mathbf{BV} ,\norm{\cdot}_{\mathbf{BV}})$, variation is denoted by $\operatorname{Var}(\cdot)$, and $\norm{\cdot}_{\mathbf{BV}} : = \operatorname{Var}(\cdot) + \norm{\cdot}_{\L{1}{\bT^{1}} }$ as usual.
We have the following \emph{Lasota-Yorke} inequality.
\begin{lem}\label{lem:LY}
There exists $\lambda>0$, $C_{\lambda}>0$ such that, for all $n\in \bN$, $b\in \bR$, $h\in \mathbf{BV}$
\[
\norm{\cL_{b}^{n} h}_{\mathbf{BV}} \leq C_{\lambda} \lambda^{-n} \norm{h}_{\mathbf{BV}} + C_{\lambda}(1 + \abs{b}) \norm{h}_{\L{1}{\bT^{1}} }
\]
\end{lem}
\begin{proof}
The proof is essentially standard (see for example~\cite{Keller:1989}) but it is important to note the factor of $\abs{b}$ which appears in front of the $\mathbf{L^{1}}$ norm.
We already know that $ \norm{\cL_{b}h}_{\mathbf{L^{1}} } \leq \norm{h}_{\mathbf{L^{1}} }$.
Note that
\[
\operatorname{Var}(h) = \sup \left\{ \norm{h\cdot \eta' }_{\mathbf{L^{1}} } : \eta \in \cC^{1}(\bT^{1}, \bC), \abs{\eta} \leq 1 \right\}.
\]
Consequently we must estimate $\norm{\cL_{b} h\cdot \eta' }_{ \mathbf{L^{1}} } = \norm{h\cdot (e^{ib\tau} \cdot \eta'\circ f) }_{ \mathbf{L^{1}} }$. In order to do this note that (for convenience we denote $J := J_{1} = 1/\abs{f'}$)
\[
\left[ J \cdot \eta \circ f \cdot e^{ib\tau} \right]'
= J' \cdot \eta \circ f \cdot e^{ib\tau}
+ ib \tau' \cdot J \cdot \eta \circ f \cdot e^{ib\tau}
+ (e^{ib\tau} \cdot \eta'\circ f).
\]
This means that
\[
\operatorname{Var}(\cL_{b}h)
\leq
\norm{J'}_{\mathbf{L^{\infty}}}
\norm{h}_{\mathbf{L^{1}}}
+
\abs{b} \norm{\tau' \cdot J}_{\mathbf{L^{\infty}}}
\norm{h}_{\mathbf{L^{1}}}
+
\norm{h \cdot \smash{ \left[ J \cdot \eta \circ f \cdot e^{ib\tau} \right]' } }_{\mathbf{L^{1}}}
\]
The remaining problem is that $[ J \cdot \eta \circ f \cdot e^{ib\tau} ]$ could be discontinuous. Therefore we introduce the quantity $\phi : \bT^{1} \to \bR$ which is piecewise affine (discontinuous only where $[ J \cdot \eta \circ f \cdot e^{ib\tau} ]$ is discontinuous) and such that
$([ J \cdot \eta \circ f \cdot e^{ib\tau} ] - \phi)(x)$ tends to $0$ as $x$ approaches any discontinuity point.
This means that $[ J \cdot \eta \circ f \cdot e^{ib\tau} ] $ is continuous and piecewise\footnote{That $[ J \cdot \eta \circ f \cdot e^{ib\tau} ] $ is continuous and piecewise~$\cC^{1}$ means that it may be approximated by a $\cC^{1}$ function with error small in the appropriate sense that makes no difference to the final estimate.}~$\cC^{1}$.
Note that $\norm{\phi}_{\mathbf{L^{\infty}}} \leq \norm{J}_{\mathbf{L^{\infty}}}$ and so
$\norm{[ J \cdot \eta \circ f \cdot e^{ib\tau} ] - \phi}_{\mathbf{L^{\infty}}} \leq 2 \norm{J}_{\mathbf{L^{\infty}}}$.
On the other hand, taking advantage of the finite number of discontinuities in this setting, we know that $\norm{\phi'}_{\mathbf{L^{\infty}}} $ is bounded by some constant which depends on the size of the smallest image of an element of the partition of smoothness.
We have shown that
\[
\operatorname{Var}(\cL_{b}h)
\leq
2 \norm{J}_{\mathbf{L^{\infty}}} \operatorname{Var}(h)
+
( \norm{J'}_{\mathbf{L^{\infty}}} + \abs{b} \norm{\tau' \cdot J}_{\mathbf{L^{\infty}}} + \norm{\phi'}_{\mathbf{L^{\infty}}} )
\norm{h}_{\mathbf{L^{1}}}.
\]
This suffices\footnote{By considering higher iterates of the same argument, if one were interested in optimal estimates, $\lambda$ can be chosen arbitrarily close to $ \limsup_{n\to \infty} \smash{ \abs{J_{n}}^{\frac{1}{n}} }$.}
since we assumed that $\inf\abs{f'}>2$ and so $2 \norm{J}_{\mathbf{L^{\infty}}} <1$.
Consequently the above estimate may be iterated to produce an estimate for all $n\in \bN$.
\end{proof}
These estimates and the compactness of the embedding $\mathbf{BV} \hookrightarrow \L{1}{\bT^{1}}$, by the usual arguments (see, for example~\cite{keller2005sgo}), imply that the operator $\cL_{b} : \mathbf{BV} \to \mathbf{BV}$ has spectral radius not greater than $1$ and essential spectral radius not greater than $\lambda^{-1} \in (0,1)$.
The spectral radius of $\cL_{0} : \mathbf{BV} \to \mathbf{BV}$ is equal to $1$.
It is convenient to introduce the equivalent norm
\[
\norm{h}_{(b)} := (1+\abs{b})^{-1} \norm{h}_{\mathbf{BV}} + \norm{h}_{\L{1}{\bT^{1}} }.
\]
The main purpose of this section is to
prove the following result.
\begin{prop}\label{prop:mainuse}
There exists $b_0 >0$, $\rho>0$, and $\gamma_{2}>0$ such that
\[
\norm{\smash{\cL_{b}^{n(b)}} }_{(b)}
\leq
e^{-n(b) \gamma_{2} },
\quad \quad\text{for all $\abs{b} \geq b_0$, $n(b) := \lceil \rho \ln \abs{b} \rceil$}.
\]
\end{prop}
\noindent
The remainder of the section will be devoted to the proof of the above. The proof is self contained apart from using Proposition~\ref{prop:mainlem} (see below) whose proof is postponed to Section~\ref{sec:mainestimate}.
\begin{lem}\label{lem:LY2}
For all $n\in \bN$, $h\in \mathbf{BV}$, $b\in \bR$
\[
\norm{\cL_{b}^{n} h}_{(b)} \leq C_{\lambda} \lambda^{-n} \norm{h}_{(b)} + C_{\lambda} \norm{h}_{\L{1}{\bT^{1}} }.
\]
\end{lem}
\begin{proof}
This is a direct result of the definition of the norm and the Lasota-Yorke estimate (Lemma~\ref{lem:LY}).
\end{proof}
First we deal with the easy case when $ \norm{ h}_{\mathbf{BV}} $ is large in comparison to $\norm{h}_{\L{1}{\bT^{1}} } $.
Let $n_{0}:= \lceil \ln (4 C_{\lambda}) / \ln \lambda \rceil$.
\begin{lem}\label{lem:easy}
Suppose that $h\in \mathbf{BV}$, satisfying
$2 C_{\lambda} (1+\abs{b}) \norm{h}_{\L{1}{\bT^{1}} } \leq \norm{h}_{\mathbf{BV}}$.
Then $\norm{\cL_{b}^{n_{0}} h}_{(b)} \leq \frac{3}{4} \norm{ h}_{(b)} $.
\end{lem}
\begin{proof}
The definition of $n_{0} \in \bN$ is such that $C_{\lambda} \lambda^{-n_{0}} + \frac{1}{2} \leq \frac{3}{4}$.
The conclusion then follows from Lemma~\ref{lem:LY2}.
\end{proof}
This means that we only need to worry about estimating in the case where $2 C_{\lambda} (1+\abs{b}) \norm{h}_{\L{1}{\bT^{1}} } > \norm{h}_{\mathbf{BV}}$.
This is the case where the density can be considered to be ``almost constant'' as long as we look on the scale of $\abs{b}^{-1}$.
Furthermore it will suffice to estimate the $\mathbf{L^{1}}$ norm and not the $\mathbf{BV}$ norm as demonstrated by the following calculation.
Using Lemma~\ref{lem:LY2}, for any $n\in \bN$
\begin{equation}\label{eq:iterate}
\begin{aligned}
\norm{\cL_{b}^{2n} h}_{(b)}
&\leq C_{\lambda} \lambda^{-n} \norm{\cL_{b}^{n} h}_{(b)} + C_{\lambda} \norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} }\\
&\leq C_{\lambda}^{2} \lambda^{-2n} \norm{ h}_{(b)} + C_{\lambda}^{2} \lambda^{-n} \norm{h}_{\L{1}{\bT^{1}} } + C_{\lambda} \norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} }\\
&\leq 2 C_{\lambda}^{2} \lambda^{-n} \norm{ h}_{(b)} + C_{\lambda} \norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} }.
\end{aligned}
\end{equation}
It therefore remains to obtain exponential contraction of $ \norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} }$ in terms of $\norm{ h}_{(b)}$ in the case when $ 2 C_{\lambda} (1+\abs{b} )\norm{h}_{\L{1}{\bT^{1}} } > \norm{h}_{\mathbf{BV}}$.
In order to later deal with discontinuities we now introduce a ``growth lemma'' suitable for this setting.
Fix $\delta>0$ such that, for any interval $\omega \subset \bT^{1}$ of size $\abs{\omega}\leq \delta$ the image $f\omega$ consists of at most two connected components.
We will define unions of open intervals $\Omega_{n}$ for all $n\in \bN$ iteratively.
Let $\Omega_{0} \subset \bT^{1}$ be an interval, $\abs{\Omega_{0}}\leq \delta$.
Suppose that $\Omega_{n}$ is already defined.
Let $\omega$ be one of the connected components of $\Omega_{n}$.
The image $f \omega$ is the union of intervals, some could be large, some could be small. It is convenient to maintain all intervals of size less than $\delta$ and so we artificially chop long intervals so that they are always of size greater than $\delta/2 $ and less than $\delta$.
In this fashion let the set ${\{\omega_{k}\}}_{k}$ be a set of open intervals which exhaust $\omega$ except for a zero measure set and such that each $f\omega_{k}$ is a single interval of length not greater than $\delta$.
The set $\Omega_{n+1} $ is defined to be partition of $\Omega_{n}$ produced by following the same procedure for each connected component of $\Omega_{n}$.
We must control the measure of points close to the boundaries of $\Omega_{n} = {\{ \omega_{j} \}}_{j}$.
For any $n\in \bN$, $x\in \Omega_{n}$, let $r_{n}(x) := d(f^{n}(x), f^{n}( \partial \Omega_{n}) )$,
and hence let ($\mathbf{m}$ denotes Lebesgue measure)
\[
\cZ_{\epsilon}\Omega_{n} :=
\mathbf{m}({ \{ x\in \Omega_{n} : r_{n}(x) \leq \epsilon \} }).
\]
Let $C_{\beta}:= {4 \Lambda \beta} {\delta^{-1} \lambda^{-1} (\beta -{1})^{-1}}$.
\begin{lem}
\label{lem:growth}
There exists $ \beta >1$ such that, for all $n\in \bN$, $\epsilon>0$,
\[
\cZ_{\epsilon}\Omega_{n} \leq
\beta^{-n} \lambda^{n} \cZ_{\epsilon/ \lambda^{n}}\Omega_{0}
+ \epsilon C_{\beta} \abs{ \Omega_{0} }.
\]
\end{lem}
\begin{proof}
Suppose for the moment that $\Omega_{n}$ consists of just one element, i.e., $\Omega_{n} = \{\omega\}$.
We will estimate $\cZ_{\epsilon}\Omega_{n+1} $.
The image $f \omega$ consists of at most two connected components.
But some of these connected components could be large in which case they will be cut into smaller pieces of size between $\delta/2 $ and $\delta$.
The set $ \partial \Omega_{n+1}$ consists of points which come from one of three different origins:
from $ \partial \Omega_{n}$; from a cut due to the discontinuities of the map; or from the artificial cuts. The first two possibilities are bounded by $2 \cZ_{\epsilon/ \lambda}\Omega_{n} $. The total length of $f \omega$ is not greater than $\Lambda \mathbf{m}( \omega )$ and so the total number of artificial cuts is not greater than $2 \delta^{-1} \Lambda \mathbf{m}( \omega )$. Summing these terms we obtain the estimate
\[
\cZ_{\epsilon}\Omega_{n+1} \leq
2 \cZ_{\epsilon/ \lambda}\Omega_{n}
+ 4 \epsilon \frac{ \Lambda } {\delta \lambda}\mathbf{m}( \omega ).
\]
The equivalent estimate holds, even when $\Omega_{n}$ consists of more than one element.
Choose $ \beta > 1$ such that $2 = \beta^{-1} \lambda$ and so the above estimate reads as
\[
\cZ_{\epsilon}\Omega_{n+1} \leq
\beta^{-1} \lambda \cZ_{\epsilon/ \lambda}\Omega_{n}
+ 4 \epsilon \frac{ \Lambda } {\delta \lambda}\mathbf{m}( \Omega_{n} ),
\]
and iterated produces the estimate (since $\sum_{j=0}^{\infty}\beta^{-n} = \frac{\beta}{\beta-{1}}$)
\[
\cZ_{\epsilon}\Omega_{n} \leq
\beta^{-n} \lambda^{n} \cZ_{\epsilon/ \lambda^{n}}\Omega_{0}
+ \epsilon \frac{4 \Lambda \beta } {\delta \lambda (\beta -{1})}\mathbf{m}( \Omega_{0} ).\qedhere
\]
\end{proof}
The argument will depend crucially on the three quantities $\rho_{1}, \xi, \rho_{2}>0$.
Let
\begin{equation}
\label{eq:defconsts}
\rho_{1} := \frac{2}{\ln \lambda},
\quad \quad
\xi := \frac{\ln \beta}{2 \ln \lambda},
\quad \quad
\rho_{2} := \frac{\xi}{2\ln \Lambda},
\end{equation}
and hence let $n_{1}(b) := \left\lceil \rho_{1} \ln \abs{b} \right\rceil $,
$n_{2}(b) := \left\lceil \rho_{2} \ln \abs{b} \right\rceil $.
Let $n(b) := n_{1}(b) + n_{2}(b)$.
For notational simplicity we will often suppress the dependence on $b$ of $n, n_{1}, n_{2}$.
Note that $\lambda \beta^{-1} =2$ and so $\ln \beta < \ln \lambda$ and hence $\xi < \frac{1}{2}$.
We use two time scales:
The first $n_{1}$ iterates are for a small interval of length $\abs{b}^{-(1+\xi)}$ to expand to a reasonable size,
then we take $n_{2} $ iterates to see oscillatory cancelations.
The argument will also depend on the choice of $b_{0}>0$. As several points during the argument this quantity will be chosen sufficiently large.
Denote by ${\{H_{\ell}\}}_{\ell}$ the partition of $\bT^{1}$ into subintervals
of equal length
such that
\begin{equation}
\label{eq:Hpart}
\abs{b}^{-(1+\xi)} \leq \abs{ H_{\ell} } \leq 2\abs{b}^{-(1+\xi)}.
\end{equation}
We use this partition to approximate the density $h$.
Denote by $h_{b}$ the density which is constant on each $H_{\ell}$
and equal to the average value of $h$ on $H_{\ell}$.
Note that $\norm{ h }_{\L{1}{\bT^{1}}} = \norm{ h_{b} }_{\L{1}{\bT^{1}}} $.
\begin{lem}\label{lem:approx}
Let $h\in \mathbf{BV}$, $b\in \bR$, $\abs{b} \geq b_{0}$ such that $ 2 C_{\lambda} (1+\abs{b} )\norm{h}_{\L{1}{\bT^{1}} } > \norm{h}_{\mathbf{BV}}$ and let $ h_{b}$ be the piecewise constant function as defined in the above paragraph.
Then
\[
\norm{\smash{h - h_{b}}}_{\L{1}{\bT^{1}} } \leq 4 C_{\lambda} e^{-n\frac{\xi}{2(\rho_{1} + \rho_{2})}} \norm{h}_{\L{1}{\bT^{1}}}.
\]
\end{lem}
\begin{proof}
Standard approximation results for $\mathbf{BV}$ and Sobolev functions imply that
$\norm{\smash{h - h_{b}}}_{\L{1}{\bT^{1}} } \leq 2 \abs{b}^{-(1+\xi)} \norm{h}_{\mathbf{BV}}$
since $\abs{H_{\ell}} \leq 2 \abs{b}^{-(1+\xi)}$.
Substituting the control on $ \norm{h}_{\mathbf{BV}}$ which is assumed we have
\[
\norm{\smash{h - h_{b}}}_{\L{1}{\bT^{1}} } \leq 4 \abs{b}^{-(1+\xi)} C_{\lambda} (1+\abs{b} )\norm{h}_{\L{1}{\bT^{1}} }.
\]
Ensuring that $b_{0}>1$ then $(1+ \abs{b} )\leq 2 \abs{b}$.
Increasing $b_{0}$ more if required we may assume that
$n(b) \leq 2(\rho_{1}+\rho_{2}) \ln \abs{b}$.
This means that $\abs{b}^{-\xi} \leq e^{-n(b)\frac{\xi}{2(\rho_{1} + \rho_{2})}}$.
Consequently
$\norm{\smash{h - h_{b}}}_{\L{1}{\bT^{1}} } \leq 8 C_{\lambda} e^{- n(b) \frac{ \xi}{2(\rho_{1} + \rho_{2})}}$.
\end{proof}
Using Lemma~\ref{lem:approx} we know that in the case $ 2 C_{\lambda} (1+\abs{b} )\norm{h}_{\L{1}{\bT^{1}} } > \norm{h}_{\mathbf{BV}}$ then
\[
\norm{\cL_{b}^{n} h}_{\L{1}{\bT^{1}} }
\leq
\norm{\cL_{b}^{n} h_{b}}_{\L{1}{\bT^{1}} }
+
4 C_{\lambda} e^{-n \frac{\xi}{2(\rho_{1} + \rho_{2})}} \norm{h}_{\L{1}{\bT^{1}}},
\]
since $h_{b} = \sum_{\ell} h_{b} \mathbf{1}_{\ell}$ where $h_{b}$ is constant on each interval $H_{\ell}$ and that $\norm{ h }_{\L{1}{\bT^{1}}} = \norm{ h_{b} }_{\L{1}{\bT^{1}}} $.
We now take advantage of the following result. This is the main estimate which takes advantage of the oscillatory cancellation mechanism which is present in this setting.
\begin{prop}\label{prop:mainlem}
There exists $C_{3}>0$, $\gamma_{3}>0$ such that, for all $\abs{b}\geq b_{0}$ and $\ell$,
\[
\norm{\smash{ \cL_{b}^{n(b)}\mathbf{1}_{H_{\ell}} }}_{\L{1}{\bT^{1}}}
\leq
C_{3} e^{-n(b) \gamma_{3}}
\norm{\mathbf{1}_{H_{\ell}}}_{\L{1}{\bT^{1}}}.
\]
\end{prop}
\noindent
The proof of the above is postponed to Section~\ref{sec:mainestimate}.
Combining Lemma~\ref{lem:approx} and Proposition~\ref{prop:mainlem} we obtain the estimate
\[
\begin{aligned}
\norm{\smash{ \cL_{b}^{n(b)} h }}_{\L{1}{\bT^{1}} }
&\leq
\left( 4 C_{\lambda} e^{-n(b)\frac{\xi}{2(\rho_{1} + \rho_{2})}}
+
C_{3} e^{-n(b) \gamma_{3}}
\right)
\norm{h}_{\L{1}{\bT^{1}}}\\
& \leq
C_{4} e^{-n(b) \gamma_{4}}
\norm{h}_{\L{1}{\bT^{1}}}
\end{aligned}
\]
where $\gamma_{4}:= \min(\frac{\xi}{2(\rho_{1} + \rho_{2})}, \gamma_{3})$
and $C_{4}:= 4 C_{\lambda} + C_{3}$.
We now substitute these estimates into \eqref{eq:iterate}.
\[
\begin{aligned}
\norm{\smash{ \cL_{b}^{2n(b)} h }}_{(b)}
&\leq
\left(
2 C_{\lambda}^{2} \lambda^{-n(b)}
+ C_{\lambda} C_{4} e^{-n(b) \gamma_{4}}
\right)
\norm{ h}_{(b)}\\
&\leq
C_{\lambda}(2C_{\lambda} + C_{4}) e^{-n(b) \gamma_{5}}
\norm{ h}_{(b)},
\end{aligned}
\]
where $\gamma_{5} := \min( \ln \lambda, \gamma_{4})$.
To complete the proof of Proposition~\ref{prop:mainuse} we must combine the above estimate with Lemma~\ref{lem:easy}.
We choose $b_{0}>0$ sufficiently large such that
\[
\norm{\smash{ \cL_{b}^{2n(b)} h }}_{(b)}
\leq
e^{-n(b) \frac{\gamma_{5}}{2}}
\norm{ h}_{(b)}
\]
for all $\abs{b} \geq b_{0}$.
Note that the estimate of Lemma~\ref{lem:easy} cannot be simply iterated since the assumption of the estimate is not invariant. However we can argue as follows: Either the estimate can be interated or the above estimate applies. Consequently we obtain the exponential rate as required and complete the proof of Proposition~\ref{prop:mainuse}.
\section{The Main Estimate}
\label{sec:mainestimate}
This section is devoted to the proof of Proposition~\ref{prop:mainlem} which was stated in Section~\ref{sec:preparing}.
In order to prove this proposition we must estimate $ \norm{\smash{ \cL_{b}^{n(b)}\mathbf{1}_{\Omega} }}_{\L{1}{\bT^{1}}}$ where $\Omega$ is an interval such that $\abs{b}^{-(1+\xi)} \leq \abs{ \Omega} \leq 2\abs{b}^{-(1+\xi)}$.
Let $\Omega_{0} = \Omega$ and, using the notation of Lemma~\ref{lem:growth},
denote by $ {\{\omega_{j}\}}_{j}$ the connected components of $\Omega_{n}$.
Let $h_{j} := \left. f^{n(b)} \right|_{\omega_{j}}^{-1}$.
Note that $\norm{\mathbf{1}_{\Omega}}_{\L{1}{\bT^{1}}} =\abs{{\Omega} }$.
We must estimate
\begin{equation}
\label{eq:termtoestimate}
\norm{\cL_{b}^{n}\mathbf{1}_{\Omega}}_{\L{1}{\bT^{1}}}
=
\int_{\bT^{1}} \left| \vphantom{\sum } \smash{ \sum_{j} \left( { J_{n}} \cdot e^{ib\tau_{n} } \right) \circ h_{j}(z) \cdot \mathbf{1}_{f^{n}\omega_{j}}}(z) \right| \ dz.
\end{equation}
Introduce a partition of $\bT^{1}$ into equal sized subintervals ${\{I_{p}\}}_{p}$ such that
\begin{equation}
\label{eq:Ipart}
\abs{b}^{-(1-\xi)}
\leq
\abs{I_{p}}
\leq 2 \abs{b}^{-(1-\xi)}.
\end{equation}
For each $p$, fix some $y_{p} \in I_{p}$ as a reference.
To proceed we would like to ensure that the subintervals $f^{n}\omega_{j}$ make full crossings of the intervals $I_{p}$.
For each $p$ let $G_{p}$ denote the set of indexes $j$ such that $f^{n}\omega_{j} \supset I_{p}$.
Let $G_{p}^{\complement}$ denote the complement of $G_{p}$.
The integrals associated to indexes in the set $G_{p}^{\complement}$ are estimated as follows.
\[
\sum_{p}
\int_{I_{p}} \left| \vphantom{\sum } \smash{ \sum_{j \in G_{p}^{\complement}} \left( { J_{n}} \cdot e^{ib\tau_{n} } \right) \circ h_{j}(z) \cdot \mathbf{1}_{f^{n}\omega_{j}}}(z) \right| \ dz
\leq
\sum_{p} \sum_{j \in G_{p}^{\complement}}
\abs{ \omega_{j} \cap f^{-n}I_{p} }.
\]
That $j \in G_{p}^{\complement}$
implies that one of the end points of $f^{n}\omega_{j}$
is contained within $I_{p}$.
Consequently $ \omega_{j} \cap f^{-n}I_{p} $ is contained within the set
$\{ x\in \Omega_{n} : r_{n}(x) < \epsilon \}$
where $\epsilon = \abs{I_{p}} \leq 2 \abs{b}^{-(1-\xi)}$.
This means that
\[
\sum_{p} \sum_{j \in G_{p}^{\complement}}
\abs{ \omega_{j} \cap f^{-n}I_{p} }
\leq
\cZ_{\epsilon}\Omega_{n}.
\]
Applying the estimate of Lemma~\ref{lem:growth} gives a bound of
\[
\begin{aligned}
\cZ_{\epsilon}\Omega_{n}
&\leq
\beta^{-n}\lambda^{n} \frac{2\epsilon}{\lambda^{n}} +
\epsilon C_{\beta} \abs{ \Omega }\\
&\leq
8 \abs{b}^{-(1+\xi)} \left( e^{-n \ln \beta}e^{n \frac{2\xi}{\rho_{1} + \rho_{2}}} + 2 C_{\beta} e^{-n \frac{1-\xi}{\rho_{1} + \rho_{2}}} \right).
\end{aligned}
\]
Recalling the definitions of $\xi$ and $\rho_{1}$, note that $ \frac{2\xi}{\rho_{1} + \rho_{2}} < \frac{2\xi}{\rho_{1}} = \xi \ln \lambda$ and so
$ \ln \beta- \frac{2\xi}{\rho_{1} + \rho_{2}} > \ln \beta- \xi \ln \lambda > 0$.
Let $\gamma_{6} := \min( \ln \beta- \frac{2\xi}{\rho_{1} + \rho_{2}}, \frac{1-\xi}{\rho_{1} + \rho_{2}} ) >0$,
$C_{5}:= 8(1+ 2 C_{\beta} )$.
This means that
\begin{equation}
\label{eq:surplus}
\sum_{p}
\int_{I_{p}} \left| \vphantom{\sum } \smash{ \sum_{j \in G_{p}^{\complement}} \left( { J_{n}} \cdot e^{ib\tau_{n} } \right) \circ h_{j}(z) \cdot \mathbf{1}_{f^{n}\omega_{j}}}(z) \right| \ dz
\leq
\abs{\Omega} C_{5} e^{-\gamma_{6} n}.
\end{equation}
Now we may proceed to estimate \eqref{eq:termtoestimate} summing only over the indexes $j \in G_{p}$.
Since $\abs{\sum_{k} a_{k}}^{2} = \sum_{jk} a_{j}\overline{a_{k}} $, using also Jensen's inequality, we have
\begin{equation}\label{eq:jensen}
\begin{aligned}
\sum_{p} \int_{I_{p}} \left| \vphantom{\sum } \smash{ \sum_{j\in G_{p}} }
\left( { J_{n}} \cdot e^{ib\tau_{n} }\right)
\circ h_{j}(z) \right| \ dz
&=
\sum_{p} \int_{I_{p}} \left( \sum_{j,k \in G_{p}} (K_{j,k} \cdot e^{ib \theta_{j,k} })(z) \right)^{\frac{1}{2}} \ dz\\
&\leq
\left( \sum_{p} \sum_{j,k \in G_{p}} \left| \int_{I_{p}} (K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right| \right)^{\frac{1}{2}}\\
\end{aligned}
\end{equation}
where $K_{j,k}:= { J_{n}}\circ h_{j} \cdot { J_{n}}\circ h_{k}$ and we define the following crucial quantity related to the phase difference between different preimages of the same point:
\[
\theta_{j,k}(x) := \left( \tau_{n} \circ h_{j } - \tau_{n} \circ h_{k } \right)(x).
\]
\begin{lem}
\label{lem:estimateJp}
There exists $C_{6} >0$ such that
$J_{n}'(x) \leq C_{6}$ for all $x \in \bT^{1}$, $n\in \bN$.
\end{lem}
\begin{proof}
Note that
$ J_{n} = \prod_{j=0}^{n-1} \frac{1}{f'}\circ f^{j} $.
Consequently
$J_{n}' = \sum_{j=0}^{n-1} \frac{f''}{f'} \circ f^{j} \cdot J_{n-j}\circ f^{j}$.
And so $\abs{J_{n}'} \leq \sup \abs{f''} /(\lambda -1)$ for any $n\in \bN$.
\end{proof}
\begin{lem}\label{lem:taupprime}
There exists $C_{7}>0$, independent of $n \in \bN$, such that
$
\abs{\smash{\theta_{j,k}''}} \leq C_{7}$.
\end{lem}
\begin{proof}
Suppose that $g: \omega \to \bT^{1}$ such that
$g \circ f^{n} = \mathbf{id}$.
Let $g^{(j)} := f^{n-j} \circ g$.
Note that
\[
(\tau_{n}\circ g)' = \sum_{j=0}^{n-1} (\tau' \cdot J_{j})\circ g^{(j)}.
\]
Consequently
\[
(\tau_{n} \circ g)'' =
\sum_{j=0}^{n-1}
\left( J_{j}^{2}\cdot \tau'' + \tau' \cdot J_{j}' \cdot J_{j} \right)\circ g^{(j)}.
\]
By Lemma~\ref{lem:estimateJp} we know that $J_{n}' \leq C_{6}$.
Since $\tau$ is $\cC^{2}$ and $J_{n} \leq \lambda^{n}$ the above term is uniformly bounded for any $n\in \bN$.
\end{proof}
Let $g_{j} := f^{n_{1}} \circ h_{j}$.
For each $p$ let $A_{p}$ denote the set of denote the set of pairs $(j,k) \in G_{p} \times G_{p}$ such that $g_{j}(x_{p}) \pitchfork g_{k}(x_{p})$ (this is the case where we see oscillatory cancelations since the two preimages are transversal at iterate $n_{2}$).
Let $A_{p}^{\complement}$ denote the complement set, i.e., the set of pairs $(j,k)$ such that $g_{j}(x_{p}) \not\pitchfork g_{k}(x_{p})$.
\begin{equation}
\label{eq:thetwopieces}
\begin{aligned}
\sum_{j,k \in G_{p} } \left| \int_{I_{p}} (K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right|
&\leq \sum_{j} \sum_{k: (j,k) \in A_{p}} \left| \int_{I_{p}} (K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right| \\
& \ \ + \sum_{j} \sum_{ k : (j,k) \in A_{p}^{\complement} } \int_{I_{p}} K_{j,k}(z) \ dz.
\end{aligned}
\end{equation}
Before estimating the above it is convenient to give the following distortion estimates.
\begin{lem}
\label{lem:estK}
$K_{j,k}' \leq 2 C_{6} K_{j,k}$.
\end{lem}
\begin{proof}
Differentiating we obtain
$ K_{j,k}' = K_{j,k} \left( J_{m}'\circ h_{j} + J_{m}'\circ h_{k} \right) $.
By Lemma~\ref{lem:estimateJp} we know that $J_{n}' \leq C_{6}$.
\end{proof}
Recall that that $\abs{b}^{-(1+\xi)} \leq \abs{ \Omega} \leq 2\abs{b}^{-(1+\xi)}$ and
$n_{1} = \lceil \rho_{1} \ln \abs{b} \rceil$.
\begin{lem}
\label{lem:otherdistortion}
For all $y\in \bT^{1}$,
\[
\sum_{x \in f^{-n_{1}}(y) \cap \Omega} J_{n_{1}}(x) \leq 6 C_{\lambda} \abs{\Omega}.
\]
\end{lem}
\begin{proof}
First note that
\[
\sum_{x \in f^{-n_{1}}(y) \cap \Omega} J_{n_{1}}(x)
=
\sum_{x \in f^{-n_{1}}(y) } J_{n_{1}}(x) \cdot \mathbf{1}_{\Omega}
= (\cL_{0}^{n_{1}} \mathbf{1}_{\Omega} )(y),
\]
and so it suffices to estimate $\norm{ \cL_{0}^{n_{1}} \mathbf{1}_{\Omega} }_{\L{\infty}{\bT^{1}}}$.
In one dimension $\norm{ \cdot }_{\L{\infty}{\bT^{1}}} \leq 2 \norm{ \cdot }_{\mathbf{BV}}$.
Moreover $ \norm{ \mathbf{1}_{\Omega} }_{\L{1}{\bT^{1}}} = \abs{\Omega}$ and
$ \norm{ \mathbf{1}_{\Omega} }_{\mathbf{BV}} = 2$.
So, with the help of the estimate from Lemma~\ref{lem:LY},
\[
\norm{ \cL_{0}^{n_{1}} \mathbf{1}_{\Omega} }_{\L{\infty}{\bT^{1}}}
\leq 2(2 C_{\lambda} \lambda^{-n_{1}} + C_{\lambda} \abs{\Omega}).
\]
Note that $\lambda^{-n_{1}} = \abs{b}^{-2}$ since $n_{1} \geq \rho_{1} \ln \abs{b}$ and $\rho_{1} = 2 / \ln \lambda$.
Consequently ($\xi \leq \frac{1}{2}$) the above quantity is bounded by $6 C_{\lambda} \abs{b}^{-(1+\xi)}$.
\end{proof}
In a similar way to the above
$\sum_{x \in f^{-n}(y)} J_{n}(x)
= (\cL_{0}^{n} \mathbf{1} )(y)$ for all $y\in \bT^{1}$, $n\in \bN$.
We may again apply the estimate from Lemma~\ref{lem:LY}
and so
\begin{equation}
\label{eq:distortion}
\sum_{x \in f^{-n}(y)} J_{n}(x) \leq C_{\lambda}.
\end{equation}
By Lemma~\ref{lem:estK} we know that
$K_{j,k}' \leq 2 C_{6} K_{j,k}$.
Hence, by Gronwall's inequality,
\begin{equation}
\label{eq:gronwall}
K_{j,k}(z) \leq e^{2 C_{6} \abs{I_{p}}} K_{j,k}(x_{p})
\end{equation}
for all $z\in I_{p}$.
Choosing $b_{0}$ large insures that $\abs{I_{p}}$ is small and so
\begin{equation*}
\int_{I_{p}} K_{j,k}(z) \ dz
\leq
2
\abs{{I_{p}}} K_{j,k}(x_{p}).
\end{equation*}
Now we consider the sum in~\eqref{eq:thetwopieces} corresponding to the non-cancelling pairs (this is the second of the two terms on the right hand side).
Using the above estimates
\[
\sum_{p} \sum_{j} \sum_{ k : (j,k) \in A_{p}^{\complement} } \int_{I_{p}} K_{j,k}(z) \ dz
\leq
\sum_{j} \sum_{ k : (j,k) \in A_{p}^{\complement} } K_{j,k}(x_{p}) \ dz.
\]
Note that
\[
\sum_{j} J_{n} \circ h_{j}(x_{p})
\leq
\sum_{y \in f^{-n}(x_{p}) \cap \Omega} J_{n}(y)
\leq
\sum_{z \in f^{-n_{2}}(x_{p}) } J_{n_{2}}(z)
\sum_{y \in f^{-n_{1}}(z) \cap \Omega} J_{n_{1}}(y)
\]
where $n_{1} = \rho_{1} \ln \abs{b}$, $n_{2} = \rho_{2} \ln \abs{b}$.
Using also the estimate of Lemma~\ref{lem:otherdistortion}
\begin{equation}
\label{eq:toshow}
\sum_{j} J_{n} \circ h_{j}(x_{p})
\leq
6 C_{\lambda} \abs{\Omega}
\sum_{z \in f^{-n_{2}}(x_{p}) } J_{n_{2}}(z).
\end{equation}
Using the above estimates, together with \eqref{eq:distortion} and \eqref{eq:transdecay},
\begin{equation}
\label{eq:estimate2}
\begin{aligned}
\sum_{p}\sum_{j} \sum_{k: (j,k) \in A_{p}^{\complement}} \int_{I_{p}} K_{j,k}(z) \ dz
& \leq
(6 C_{\lambda} )^{2} \abs{\Omega}^{2}
C_{\lambda}
\sum_{\substack{z_{2} \in f^{-n_{2}}(x_{p})\\ z_{1} \not\pitchfork z_{2}} } J_{n_{2}}(z_{2})\\
& \leq (6 C_{\lambda} )^{2} \abs{\Omega}^{2} C_{\lambda} C_{\gamma} e^{-n_{2}\gamma}.
\end{aligned}
\end{equation}
Let us now consider the case where
$(j,k) \in A_{p}$
and so estimate the remaining term of \eqref{eq:thetwopieces}.
\begin{lem}\label{lem:yep}
Suppose that $(j,k) \in A_{p}$. Then
\[
\abs{\smash{\theta_{j,k}'(x_{p})}} > \tfrac{1}{2} C_{1} (J_{n_{2}} \circ g_{j} + J_{n_{2}} \circ g_{k})(x_{p}).
\]
\end{lem}
\begin{proof}
Differentiating, since $\tau_{n}\circ h_{j} = \tau_{n_{1}}\circ h_{j} + \tau_{n_{2}}\circ f^{n_1} \circ h_{j}$, we obtain
\begin{equation*}
\theta_{j,k}'
=
(\tau_{n_{1}}' \cdot J_{n})\circ h_{j} - ( \tau_{n_{1}}' \cdot J_{n}) \circ h_{k}
+ (\tau_{n_{2}}'\cdot J_{n_{2}}) \circ g_{j} - ( \tau_{n_{2}}' \cdot J_{n_{2}}) \circ g_{k}.
\end{equation*}
Applying the estimate of Lemma~\ref{lem:tauprime} means that the first two terms can be estimated as
\[
\abs{(\tau_{n_{1}}' \cdot J_{n})\circ h_{j} - ( \tau_{n_{1}}' \cdot J_{n}) \circ h_{k}) }
\leq
\tfrac{1}{2} C_{1} (J_{n_{2}} \circ g_{j} + J_{n_{2}} \circ g_{k}).
\]
Using the estimate of Lemma~\ref{lem:transverse} we have that
\[
\abs{(\tau_{n_{2}}'\cdot J_{n_{2}}) \circ g_{j} - ( \tau_{n_{2}}' \cdot J_{n_{2}}) \circ g_{k} }
>
C_{1} (J_{n_{2}} \circ g_{j} + J_{n_{2}} \circ g_{k}).\qedhere
\]
\end{proof}
The above lemma says that we have the required transversality at the point $x_{p}$. The following lemma says that the interval $I_{p}$ has been chosen sufficiently small such that this same transversality holds for the entire interval $I_{p}$.
\begin{lem}\label{lem:yepyep}
Suppose that $(j,k) \in A_{p}$. Then $\abs{\smash{ \theta_{j,k}'(y) }} > \frac{1}{2} C_{1} \Lambda^{-n_{2}}$ for all $y\in I_{p}$.
\end{lem}
\begin{proof}
By Lemma~\ref{lem:taupprime} and Lemma~\ref{lem:yep}
we know that
$\abs{\smash{ \theta_{j,k}'(y) }}
>
C_{1} \Lambda^{-n_{2}}
- \abs{I_{p}} C_{7}$.
To complete the proof it remains to show that $ \abs{I_{p}} \leq \frac{C_{1}}{2 C_{7}} \Lambda^{-n_{2}}$.
Recall that, by choice of the partition, $\abs{I_{p}} \leq 2 \abs{b}^{-(1-\xi)}$
and note that $\Lambda^{-n_{2}} \leq \abs{b}^{-\rho_{2} \ln \Lambda}$.
This means that
$\abs{I_{p}} \leq \Lambda^{n_{2}} \abs{b}^{-(1- \xi - \rho_{2} \ln\Lambda )}$.
Furthermore, by choice of $\xi$ and $\rho_{2}$ we have $\rho_{2} = \frac{\xi}{2\ln\Lambda}$ and $\xi \leq \frac{1}{2}$.
Consequently
$\abs{I_{p}} \leq \Lambda^{n_{2}} \abs{b}^{-\frac{1}{4}}$
and so, again increasing $b_{0}$ if required,
$ \abs{I_{p}} \leq \frac{C_{1}}{2 C_{7}} \Lambda^{-n_{2}}$
for all $\abs{b}\geq b_{0}$.
\end{proof}
The key part of the argument is the following lemma concerning oscillatory integrals.
\begin{lem}\label{lem:vandercorput}
Suppose $J$ is an interval, $\theta \in \cC^{2}(J,\bR)$, $K \in \cC^{1}(J,\bC)$, $b\in \bR \setminus \{0\}$ and there exists $\kappa>0$ such that $\inf \abs{\theta'} \geq \kappa$.
Then
\[
\abs{ \int_{J} K \cdot e^{ib \theta (x) } \ dx }
\leq
\frac{1}{\abs{b}} \left( \tfrac{1}{ \kappa} \sup \abs{K} + \tfrac{1} { \kappa^2} \sup \abs{K} \sup \abs{\theta''} \abs{J} + \tfrac{1 }{\kappa} \sup \abs{ K'} \abs{J} \right).
\]
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lem:vandercorput}]
First change variables, $y=\theta(x)$, then integrate by parts
\[
\begin{aligned}
\int_{J} K \cdot e^{ib \theta (x) } \ dx
& = \int_{\theta(J)} \frac{K}{\theta'}\circ \theta^{-1}(y) e^{ib y } \ dy\\
& =
\frac{i}{b} \left[ \frac{K}{\theta'}\circ \theta^{-1}(y) e^{ib y } \right]_{\theta(J)} \\
& \ \ \ \ + \frac{i}{b} \int_{\theta(J)} \left( \frac{K \theta''}{(\theta')^2 \cdot \theta'} + \frac{K'}{(\theta')^2} \right) \circ \theta^{-1}(y) e^{ib y } \ dy.
\end{aligned}
\]
Changing variables again, we obtain
\[
\int_{J} K \cdot e^{ib \theta (x) } \ dx
=
\frac{i}{b} \left[ \frac{K}{\theta'} e^{ib \theta } \right]_{J}
+ \frac{i}{b} \int_{J} \left( \frac{ K\theta''}{(\theta')^2}+ \frac{K'}{\theta'} \right) (x) e^{ib \theta(x) } \ dx.
\]
The required estimate follows immediately.
\end{proof}
In preparation of apply the above lemma, note that \eqref{eq:gronwall} implies
$\sup_{I_{p}} K_{j,k} \leq 2 K_{j,k}(x_{p})$
and similarly
$\sup_{I_{p}} K_{j,k}' \leq 4 C_{6} K_{j,k}(x_{p})$.
By Lemma~\ref{lem:yepyep} we know that
$\abs{\smash{ \theta_{j,k}' }} > \frac{1}{2} C_{1} \Lambda^{-n_{2}}$.
By Lemma~\ref{lem:taupprime} we know that
$\abs{\smash{\theta_{jk}''}} \leq C_{7}$.
Using these estimates with Lemma~\ref{lem:vandercorput} we obtain
\[
\left| \int_{I_{p}} ( K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right|
\leq
\frac{1}{\abs{b}} K_{j,k}(x_{p})
C_{8} \Lambda^{ n_{2}}
\]
where $C_{8} := 8C_{6} (1+C_{7} + 2 C_{6})C_{1}^{-2}$.
This means that, for the first sum in \eqref{eq:thetwopieces}, we obtain, using \eqref{eq:distortion} and Lemma~\ref{lem:otherdistortion}, the estimate
\begin{equation*}
\sum_{p} \sum_{\substack{j,k \\ (j,k) \in A_{p}}} \left| \int_{I_{p}} ( K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right|
\leq
\frac{1}{\abs{b}} \abs{b}^{1-\xi} C_{8} \Lambda^{n_2} C_{\lambda}^{2} (6 C_{\lambda} )^{2} \abs{\Omega}^{2}
\end{equation*}
(the term $\abs{b}^{1-\xi}$ comes from the sum over $p$).
Since $\abs{b} \geq e^{n_{2} \frac{2 \xi}{3 \rho_{1}}}$ (increasing $b_{0}$ again if required),
\[
\frac{1}{\abs{b}} \abs{b}^{1-\xi} \Lambda^{n_2}
=
\frac{1}{\abs{b}^{\xi}} \Lambda^{n_2}
\leq e^{-n_{2}(\frac{2\xi}{3\rho_{2}} - \ln \Lambda )}.
\]
Let $\gamma_{7} := \frac{2\xi}{3\rho_{2}} - \ln \Lambda > 0$. This means that
\begin{equation}
\label{eq:estimate1}
\sum_{p} \sum_{\substack{j,k \\ (j,k) \in A_{p}}} \left| \int_{I_{p}} ( K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right|
\leq
C_{8} C_{\lambda}^{2} (6 C_{\lambda} )^{2} \abs{\Omega}^{2} e^{-n_{2}\gamma_{7}}.
\end{equation}
In order to estimate the final term in~\eqref{eq:jensen} we use~\eqref{eq:thetwopieces}, sum the estimates \eqref{eq:estimate2} and \eqref{eq:estimate1}, to obtain
\[
\sum_{p} \sum_{j,k \in G_{p}} \left| \int_{I_{p}} (K_{j,k} \cdot e^{ib\theta_{j,k} })(z) \ dz \right|
\leq
\abs{\Omega}^{2}
e^{-n \gamma_{8}}
\left( (6 C_{\lambda} )^{2} C_{\lambda} C_{\gamma} + C_{8} C_{\lambda}^{2} (6 C_{\lambda} )^{2} \right),
\]
where $\gamma_{8} := \frac{\rho_{1}}{\rho_{1} + \rho_{2}} \min(\gamma , \gamma_{7}) >0$.
Let $C_{3}:=6 C_{\lambda} \left( C_{\lambda} C_{\gamma} + C_{8} C_{\lambda}^{2}\right)^{\frac{1}{2}}$. Taking the square root
of the above
\[
\begin{aligned}
\sum_{p} \int_{I_{p}} \left| \vphantom{\sum } \smash{ \sum_{j\in G_{p}} }
\left( { J_{n}} \cdot e^{ib\tau_{m} }\right)\circ h_{j}(z) \right| \ dz
&\leq \abs{\Omega} e^{-n \frac{\gamma_{8}}{2}} C_{3}
\end{aligned}
\]
Including also the estimate \eqref{eq:surplus} we have shown that (let $\gamma_{3}:= \min (\gamma_{6},\frac{\gamma_{8}}{2} )>0$)
\[
\int_{\bT^{1}} \left| \vphantom{\sum } \smash{ \sum_{j} \left( { J_{n}} \cdot e^{ib\tau_{n} } \right) \circ h_{j}(z) \cdot \mathbf{1}_{f^{n}\omega_{j}}}(z) \right| \ dz
\leq
(C_{3} + C_{5}) \abs{\Omega} e^{-n(b) \gamma_{3}}.
\]
This completes the proof of Proposition~\ref{prop:mainlem}.
\section{Rate of Mixing}
\label{sec:rateofmixing}
Here we use the estimates of Proposition~\ref{prop:mainuse} concerning the twisted transfer operators in order to estimate the rate of mixing.
Let $g,h : \bT^{2} \to \bC$ be two observables. We assume, without loss of generality, that $g$ is mean zero.
Denote by $\hat{g}_{b}$ and $\hat{h}_{b}$ their Fourier components (in the fibre coordinate), i.e.,
\[
g(x,u) = \sum_{b\in \bZ} \hat{g}_{b}(x) e^{-ibu},
\]
and similarily for $h(x,u)$.
Using the regularity of the observables (in particular the smoothness in the fibre direction) we have that $\norm{ \hat{g}_{b} }_{\mathbf{BV}} \leq (1+\abs{b})^{-1} \norm{g}_{\cC^{1}}$ (and for $h$ similarily).
By simple manipulations we obtain the formula
\[
\int_{\bT^{2}} \left( g \cdot h\circ F\right)(x,u) \ dx \ du
=
\sum_{b\in \bZ} \int_{\bT^{1}} \cL_{b}^{n} \hat{g}_{b}(x) \cdot \hat{h}_{b}(x) \ dx.
\]
We separate the sum into a finite number of terms where $\abs{b} \leq b_{0}$ and the infinite sum of the remaining terms. For the finite number of terms it suffices to use the quasi-compactness in a standard way using that the base map is mixing.
Recall that in Proposition~\ref{prop:mainuse} we obtained the estimate
$ \norm{\smash{\cL_{b}^{n(b)}} }_{(b)}
\leq
e^{-n(b) \gamma_{2} }$
where $ \rho \ln \abs{b} \leq n(b) \leq \rho \ln \abs{b} +2$.
We may assume that $\gamma_{2} >0$ is sufficiently small that $\rho \gamma_{2} <1$.
Consequently the above estimate implies that there exists $\alpha\in (0,1)$ such that
$ \norm{\smash{\cL_{b}^{n}} }_{(b)}
\leq
\abs{b}^\alpha e^{-n \gamma_{2} }$
for all $n\in \bN$, $\abs{b} \geq b_{0}$.
Note that
\[
\abs{\int_{\bT^{1}} \cL_{b}^{n} \hat{g}_{b}(x) \cdot \hat{h}_{b}(x) \ dx}
\leq 2 \norm{\smash{\cL_{b}^{n}} }_{(b)} \norm{\smash{\hat{g}_{b} } }_{\mathbf{BV}} \norm{\smash{\hat{h}_{b} } }_{\mathbf{BV}}
\]
since the $\mathbf{BV}$ norm dominates the $\mathbf{L^{\infty}}$ norm.
It remains to observe that
\[
\sum_{\abs{b} \geq b_{0}}
\norm{\smash{\cL_{b}^{n}} }_{(b)}
\norm{\smash{ \hat{g}_{b} }}_{\mathbf{BV}}
\norm{\smash{ \hat{h}_{b}} }_{\mathbf{BV}}
\leq \sum_{\abs{b} \geq b_{0}} \abs{b}^{-(2-\alpha)} e^{-n \gamma_{2} }
\norm{g}_{\cC^{1}}\norm{h}_{\cC^{1}}.
\]
Crucially $(2-\alpha) >1$ and so this is summable.
This proves exponential mixing for $\cC^{1}$ observables which, by the usual argument~\cite[footnote~2]{oli1401}, implies exponential mixing for H\"older observables.
|
\section{Introduction}
\label{intro}
In 1811, Sir Humphrey Davy (Davy 1811) was the first to report the existence of clathrate, a variety of compounds in which water forms a continuous and known crystal structure with small cages. These cages trap guests, such as methane or ethane, needed to stabilize the water lattice. The two most common clathrate structures found in nature are known as structures I and II, which differ in the type of water cages present in the crystal lattice (Sloan \& Koh 2008). Structure I has two types of cages, a small pentagonal dodecahedral cage, denoted 5$^{12}$ (12 pentagonal faces in the cage) and a large tetrakaidecahedral cage, denoted 5$^{12}$6$^2$ (12 pentagonal faces and 2 hexagonal faces in the cage). Structure II also has two types of cages, a small 5$^{12}$ cage and a large hexakaidecahedral cage, denoted 5$^{12}$6$^4$ (12 pentagonal faces and 4 hexagonal faces in the cage). The type of structure that forms depends largely on the size of the guest molecule. For example, methane and ethane induce water to form structure I clathrate and propane structure II clathrate (Sloan \& Koh 2008).
On Titan, the temperature and atmospheric pressure conditions prevailing at the ground level permit clathrates formation when liquid hydrocarbons enter in contact with the exposed water ice (Mousis \& Schmitt 2008). Assuming an open porosity for Titan's upper crust, clathrates made from hydrocarbons are even expected to be stable down to several kilometers from the surface (Mousis \& Schmitt 2008). An interesting feature of clathrates is that their formation induces a fractionation of the trapped molecules (van der Waals \& Platteuw 1959; Lunine \& Stevenson 1985; Mousis et al. 2010). This property has been used to suggest that the noble gas depletion observed in Titan's atmosphere could result from their efficient sequestration in a global clathrate layer located in the near subsurface (Thomas et al. 2007; Mousis et al. 2011). It has also recently been used to propose that the satellite's polar radius, which is smaller by several hundred meters than the value predicted by the flattening due to its spin rate, would result from the substitution of methane by ethane percolating in liquid form in clathrate layers potentially existing in the polar regions (Choukroun \& Sotin 2012).
In this paper, we investigate the interplay that may happen between an alkanofer, namely a reservoir of liquid hydrocarbons located in Titan's subsurface, and a hypothetical clathrate reservoir that progressively forms when the liquid mixture diffuses throughout a preexisting porous icy layer. This porous layer might have been generated by cryovolcanic events resulting from the ascent of liquid from subsurface ocean (Mitri et al. 2006) or from the destabilization of clathrates in the ice shell of Titan (Tobie et al. 2006). In both cases, a highly porous icy material in contact with the atmosphere is generated, probably similar to basaltic lava flows (Mousis \& Schmitt 2008). The cooling of cryolava is expected to take less than 1 yr to decrease down to Titan's surface temperature (Lorenz 1996), implying that it should be fast enough to allow preservation of the created porosity.
Hundreds of lakes and a few seas are observed to cover the polar regions of Titan (Stofan et al. 2007). Kraken Mare and Ligeia Mare, namely the two largest seas of Titan, have surface areas estimated to be $\sim$400,000 km$^2$ (Jaumann et al. 2010) and 126,000 km$^2$ (Mastrogiuseppe et al. 2014), respectively. With an average depth of $\sim$70 m, Ligeia Mare contains $\sim$5 $\times$ 10$^{15}$ kg of hydrocarbons, about 100 times the known terrestrial oil and gas reserves, but still only $\sim$1.4\% of Titan's atmospheric methane (Mastrogiuseppe et al. 2014). While a significant number of these lakes and seas should be regularly filled by hydrocarbon rainfalls (Turtle et al. 2011), some of them could be also renewed via their interconnection with alkanofers. Using porous media properties inferred from Huygens probe observations, Hayes et al. (2008) found that the timescales for flow into and out of observed lakes via subsurface transport are order of tens of years. Because the porosity is not expected to evolve significantly over $\sim$20 Myr within the subsurface of Titan (Kossacki \& Lorenz 1996), clathrates might form and equilibrate with liquid hydrocarbons well prior that the porosity reaches its close-off value. A fraction of these lakes may then result from the interaction between alkanofers and clathrate reservoirs through the ice porosity and possess a composition differing from that of lakes and rivers sourced by precipitation. As a liquid reservoir occupies a finite volume, the progressive transfer and fractionation of the molecules in the forming clathrate reservoir could alter the lakes' chemical composition.
In order to explore this possibility, we use a statistical--thermodynamic model derived from the approach of van der Waals \& Platteuw (1959) to compute the composition of the clathrate reservoir that forms as a result of the progressive entrapping of the species present in the liquid mixture. The major ingredient of our model is the description of the guest--clathrate interaction by a spherically averaged Kihara potential with a set of potential parameters based on the literature. This allows us to track the evolution of the mole fractions of species present in the liquid reservoir as a function of their progressive entrapment in the clathrate layer. Section 2 is devoted to the description of our computational approach and the physical ingredients of our model. We also discuss the underlying assumptions of our approach in this Section. The results concerning the composition of lakes interacting with clathrate reservoirs at polar or equatorial zones are presented in Section 3. Section 4 is devoted to discussion and conclusions.
\section{Modeling the clathrate--liquid equilibrium}
\label{mode}
\subsection{Computational approach}
We assume that the liquid reservoir is in contact with porous ice and that clathrates form at the liquid/ice interface. We consider an isolated system composed of a clathrate reservoir that progressively forms and replaces the H$_2$O crustal material with time and a liquid reservoir that correspondingly empties due to the net transfer of molecules to the clathrate reservoir (see concept pictured in Fig. \ref{Draw}). Based on this approach, we have elaborated a computational procedure with the intent to determine the mole fractions of each species present in the liquid reservoir and trapped in the forming clathrate reservoir, as a function of the fractions of the initial liquid volume (before volatile migration) remaining in lake and present in clathrates, respectively. At the beginning of our computations, the liquid reservoir's composition is derived from those computed by Cordier et al. (2009, 2013) for lakes at polar and equator temperatures, which result from models assuming thermodynamic equilibrium between the atmosphere and the lakes (see Table \ref{lake0}).
Because the clathration kinetics of hydrocarbons is poorly constrained at the Titan's surface temperatures considered, the present calculations use an iterative process for which the number of molecules in the liquid phase being trapped in clathrates between each iteration is equal to 10$^{-4}$ the total number of moles. Initially, all molecules are in the liquid phase. The mole fraction $x^{\rm lake}_K (0)$ of species $K$ in lake, is given in Table \ref{lake0}. The corresponding number of moles of species $K$ is defined by $n^{\rm lake}_K (0)$~=~$x^{\rm lake}_K (0)$~$\times$~$n_{\rm lake} (0)$, with $n_{\rm lake} (0)$ the number of moles of liquid available in the lake at this time. The mole fraction $x_{K}^{\rm{clat}}(0)$ of the enclathrated species $K$ and the number of moles $n_{\rm clat} (0)$ of liquid trapped in clathrate are set to zero.
At iteration $i$, the mole fraction $x_{K}^{\rm{clat}}(i)$ of each enclathrated guest $K$ is calculated by using the statistical-thermodynamic model described in section \ref{stat} and the relative abundances in the liquid phase of the previous iteration. The new numbers of moles in the lake $n^{\rm lake}_K (i)$ and in clathrate $n^{\rm clat}_K (i)$ are calculated for each species $K$, with $n^{\rm lake}_K (i)$~=~$n^{\rm lake}_K (i~-~1)$~-~$x^{\rm clat}_K (i)$~$\times$~10$^{-4}$ and $n^{\rm clat}_K (i)$~=~$n^{\rm clat}_K (i~-~1)$~+~$x^{\rm clat}_K (i)$~$\times$~10$^{-4}$. The mole fraction of each species $K$ present at iteration $i$ in the lake and clathrate are defined by $x_{K}^{\rm{lake}}(i)~=~\frac{n_{K}^{\rm{lake}}(i)}{n_{\rm lake}(i)}$ and $f_{K}^{\rm{clat}}(i)~=~\frac{n_{K}^{\rm{clat}}(i)}{n_{\rm clat}(i)}$, respectively, with $n_{\rm lake}$($i$)~=~$\displaystyle\sum_K$ $n^{\rm lake}_K$($i$) and $n_{\rm clat}$($i$)~=~$\displaystyle\sum_K$~ $n^{\rm clat}_K$($i$). At any iteration $i$, $n_{\rm tot}$ = $n_{\rm lake}$($i$) + $n_{\rm clat}$($i$). The new values of $n^{\rm lake}_K (t)$ and $n^{\rm clat}_K (t)$ are introduced in the next loop and the process is run until $n_{\rm lake}$ eventually gets to zero.
\subsection{The statistical-thermodynamic model}
\label{stat}
To calculate the relative abundances of guest species incorporated in the clathrate phase at given temperature and pressure, we use a model applying classical statistical mechanics that relates the macroscopic thermodynamic properties of clathrates to the molecular structure and interaction energies (van der Waals \& Platteuw 1959; Lunine \& Stevenson 1985; Mousis et al. 2010). It is based on the original ideas of van der Waals and Platteeuw for clathrate formation, which assume that trapping of guest molecules into cages corresponds to the three-dimensional generalization of ideal localized adsorption (see Sloan \& Koh (2008) for an exhaustive description of the statistical thermodynamics of clathrate equilibra).
In this formalism, the fractional occupancy of a guest molecule $K$ for a given type $q$ ($q$~=~small or large) of cage can be written as
\begin{equation}
y_{K,q}=\frac{C_{K,q}f_K}{1+\sum_{J}C_{J,q}f_J} ,
\label{eq1}
\end{equation}
\noindent where the sum in the denominator includes all the species which are present in the liquid phase. $C_{K,q}$ is the Langmuir constant of species $K$ in the cage of type $q$, and $f_K$ the fugacity of species $K$ in the mixture. Using the Redlich-Kwong equation of state (Redlich and Kwong 1949) in the case of a mixture dominated by C$_2$H$_6$, we find that the coefficient of fugacity $\phi$ of the mixture (defined as the ratio of the mixture's fugacity to its vapor pressure) converges towards 1 at Titan's surface temperatures and corresponding C$_2$H$_6$ vapor pressures. In our approach, the value $f_K$ of each species $K$ is calculated via the Raoult's law, which states
\begin{equation}
f_K \simeq P_K = x^{lake}_K \times P^*_K,
\label{Raoult}
\end{equation}
\noindent with $P_K$ the vapor pressure of species $K$ in the mixture and $P^*_K$ the vapor pressure of pure component $K$. $P^*_K$ is defined via the Antoine equation
\begin{equation}
log_{\rm 10}(P^*_K)~=~{\rm A}~-~\frac{{\rm B}}{T~+~ {\rm C}},
\label{Ant}
\end{equation}
\noindent with the parameters A, B and C listed in Table \ref{ABC} ($P^*_K$ is expressed in bar and $T$ in K).
The Langmuir constant depends on the strength of the interaction between each guest species and each type of cage, and can be determined by integrating the molecular potential within the cavity as
\begin{equation}
C_{K,q}=\frac{4\pi}{k_B T}\int_{0}^{R_c}\exp\Big(-\frac{w_{K,q}(r)}{k_B T}\Big)r^2dr ,
\label{eq2}
\end{equation}
\noindent where $R_c$ represents the radius of the cavity assumed to be spherical, $k_B$ the Boltzmann constant, and $w_{K,q}(r)$ is the spherically averaged Kihara potential representing the interactions between the guest molecules $K$ and the H$_2$O molecules forming the surrounding cage $q$. This potential $w(r)$ can be written for a spherical guest molecule, as (McKoy \& Sinano\u{g}lu 1963)
\begin{eqnarray}
w(r) = 2z\epsilon\Big[\frac{\sigma^{12}}{R_c^{11}r}\Big(\delta^{10}(r)+\frac{a}{R_c}\delta^{11}(r)\Big) \\ \nonumber
- \frac{\sigma^6}{R_c^5r}\Big(\delta^4(r)+\frac{a}{R_c}\delta^5(r)\Big)\Big],
\label{eq3}
\end{eqnarray}
\noindent with
\begin{equation}
\delta^N(r)=\frac{1}{N}\Big[\Big(1-\frac{r}{R_c}-\frac{a}{R_c}\Big)^{-N}-\Big(1+\frac{r}{R_c}-\frac{a}{R_c}\Big)^{-N}\Big].
\label{eq4}
\end{equation}
\noindent In Eq. 5, $z$ is the coordination number of the cell. This parameter depends on the structure of the clathrate (I or II; see Sloan \& Koh 2008) and on the type of the cage (small or large). The Kihara parameters $a$, $\sigma$ and $\epsilon$ for the molecule-water interactions, given in Table \ref{Kihara}, have been taken from the recent compilation of Sloan \& Koh (2008) when available and from Parrish \& Prausnitz (1972) for the remaining species.
Finally, the mole fraction $x^{clat}_K$ of a guest molecule $K$ in a clathrate can be calculated with respect to the whole set of species considered in the system as
\begin{equation}
x^{clat}_K=\frac{b_s y_{K,s}+b_\ell y_{K,\ell}}{b_s \sum_J{y_{J,s}}+b_\ell \sum_J{y_{J,\ell}}},
\label{eq5}
\end{equation}
\noindent where $b_s$ and $b_l$ are the number of small and large cages per unit cell respectively, for the clathrate structure under consideration, and with ${\sum_{K}} f_{K}~=~1$. Values of $R_c$, $z$, $b_s$ and $b_l$ are taken from Parrish \& Prausnitz (1972).
Among the different species considered in the present study, C$_3$H$_8$ is the only molecule whose size is too large to be trapped either in small or large cages of structure I clathrate. Because C$_3$H$_8$ can only be trapped in the large cages of structure II clathrate (Sloan \& Koh 2008), we assume that this molecule remains in the liquid phase in the case of structure I clathrate formation.
$\sigma_{K-W}$ is the Lennard-Jones diameter, $\epsilon_{K-W}$ is the depth of the potential well, and $a_{K-W}$ is the radius of the impenetrable core, for the guest-water pairs.
\subsection{Model uncertainties}
\label{uncert}
The scenario we propose is based on some underlying assumptions, indicated below:
\begin{itemize}
\item {\it Statistical thermodynamic model}. The predictive capabilities of our model, which is derived from the approach of van der Waals \& Platteeuw, rely on four key assumptions: (i) the host molecules contribution to the free energy is independent of the clathrate occupancy (the guest species do not distort the cages); (ii) the cages are singly occupied; (iii) guest molecules rotate freely within the cage and they do not interact with each other; (iv) classical statistics is valid, i.e., quantum effects are negligible. However, these assumptions are subject to caution since encaged molecules with larger dimensions may distort the cages. Also, for certain small-sized molecules (like H$_2$) multiple gas occupancy can occur, and non spherical molecules may not be free to rotate in the entire cavity. Molecular dynamics simulations (Erfan-Niya et al. 2011; Fleischer \& Genda 2013) are typically used to investigate these effects but, due to the amount of time they require, these computations do not easily provide quantitative estimates on the fractionation of the different species encaged in clathrates at the macroscopic level, in particular in systems considering a large number of species. For these reasons, and because it is often based on interaction parameters fitted on laboratory measurements of phase equilibria, allowing accurate prediction when compared to experiments, the approach of van der Waals \& Platteeuw remains the main tool employed in industry and research to determine clathrate composition (Sloan \& Koh 2008).
\item {\it Kinetics of clathrate formation.} Kinetics data concerning clathrate formation are scarce and mostly concern gas/ice interaction (see Sloan \& Koh for a review of measurements). In the present case, clathrates form from the interaction between liquid hydrocarbons and ice. To the best of our knowledge, the kinetics measurements of the closest system reported so far are those concerning clathrate formation from a mixture of liquid methane and ethane at temperatures ranging from 260 to 280 K (Murshed et al. 2010). Because of the large temperature difference between Titan's surface and these experiments, the uncertainty is too large to make use of these kinetics data. Kinetics measurements remain to be done at Titan's conditions.
\item {\it Assumption of equilibrium.} Our model is restricted to equilibrium calculations between subsurface alkanofers and coexisting clathrate layers. Hence it should be applied with caution to the case where lakes and rivers located on Titan's surface directly equilibrate with a clathrate layer located beneath. To do so, we would need to compute the simultaneous equilibrium between the clathrate reservoir, lake and atmosphere. However, our computations are a good approximation if the reequilibration timescale between the lake and the atmosphere is short compared to the timescale of clathrate formation.
\end{itemize}
\section{Results}
Figure \ref{lake} represents the evolution of the mole fractions of species present in subsurface alkanofers of Titan, starting with the one given in Table \ref{lake0}, as a function of their progressive entrapping in structures I and II clathrate reservoirs located at the poles ($T$ $\sim$90 K) and at the equator ($T$ $\sim$93.6 K). The evolution of the liquid reservoir's composition varies significantly if one assumes the formation of a structure I or a structure II clathrate reservoir. In particular, the change of clathrate structure in our model alters the number of dominant species present in the liquid phase at high mole fractions of entrapped liquid. It also drastically affects the evolution of the abundances of secondary species during the progressive liquid entrapping.
When considering the formation of a structure I clathrate, and irrespective of the liquid reservoir's temperature, the dominating species is C$_2$H$_6$ until that a liquid mole fraction of $\sim$0.85 has been trapped into clathrate. Above this value, and because of the entrapping of the other molecules in structure I clathrate, C$_3$H$_8$ becomes the only remaining species in the liquid reservoir. At both temperatures considered, the initial abundance of CH$_4$ in the liquid phase is close to that of C$_3$H$_8$ (see Table \ref{lake0}). However, as the liquid progressively forms clathrate with ice, the mole fraction of CH$_4$ rapidly decreases and finally converges towards zero after a mole fraction of $\sim$0.3--0.5 of the initial liquid reservoir has been entrapped. Meanwhile, the mole fractions of N$_2$, Ar and CO form plateaus and finally drop towards zero at the same mole fraction of entrapped liquid that corresponds to the disappearance of CH$_4$.
When considering the formation of a structure II clathrate, the dominant species remains C$_2$H$_6$, irrespective of the mole fraction of liquid entrapped in clathrate and the temperature of the reservoirs. Instead of increasing with the progressive liquid entrapment in clathrate as in the previous case, the abundance of C$_3$H$_8$ decreases and suddenly drops when the mole fraction of entrapped liquid is $\sim$0.15--0.23. The abundances of CH$_4$, N$_2$, Ar and CO also decrease with the progressive formation of structure II clathrate and suddenly drop at mole fractions of entrapped liquid in the $\sim$0.11--0.18 range. The temperature of the liquid and clathrate reservoirs also plays a role in the determination of their composition, but in a less important manner than the modification of the clathrate structure. Figure \ref{lake} shows that the rise of temperature decreases by several per cents the mole fraction of entrapped liquid at which the abundances of minor species drop in the solution. Compared to the change of clathrate structure, a temperature variation also affects (to a lower extent) the mole fractions of secondary species in the liquid reservoir but does not influence those of major compounds.
Figure \ref{clat} represents the evolution of the composition of structures I and II clathrate reservoirs on Titan as a function of the progressive entrapping of subsurface alkanofers located at the poles and at the equator. As noted for the composition of the liquid reservoirs, the structure change affects the number of dominant species in clathrate. It also significantly affects the evolution of the mole fractions of secondary species during their progressive trapping. The mole fractions of the different entrapped species strongly differ from those in solution at the beginning of their entrapping, as a result of the fractionation occurring during the clathration. However, irrespective of the structure considered and because of the conservation of the total number of moles in our system, the mole fraction of each species trapped in clathrate converges towards its initial abundance in solution when the fraction of entrapped liquid approaches 1.
In the case of structure I clathrate formation, C$_2$H$_6$ remains the dominant volatile. On the other hand, with a mole fraction ranging between 0.11 and 0.25, CH$_4$ is the second most abundant volatile present in clathrate, except at fractions of entrapped liquid close to 1 and at equator temperatures. With a mole fraction in the 10$^{-3}$--10$^{-2}$ range, N$_2$ is the third most abundant volatile trapped in clathrate at fractions of entrapped liquid lower than $\sim$0.9. The mole fractions of Ar and CO are in the 10$^{-6}$--10$^{-5}$ and 10$^{-9}$--10$^{-6}$ ranges, respectively, making them the less abundant species present in the forming structure I clathrate. As mentioned in Sec. \ref{mode}, C$_3$H$_8$ is not trapped in structure I clathrate, due to its large size compared to those of small and large cages.
When considering the formation of structure II clathrate, the two most abundant species present in clathrate are CH$_4$ and C$_3$H$_8$ for mole fractions of entrapped liquid lower than $\sim$0.2--0.3. Above this range of values, C$_2$H$_6$ becomes again the most abundant volatile present in clathrate. On the other hand, the mole fractions of N$_2$, Ar and CO are in ranges close to those computed in the case of structure I clathrate formation. In both clathrate structures, the temperature plays the same role as the one noted for liquid reservoirs. At a similar species abundance, an increase of temperature decreases by several percent the corresponding mole fraction of entrapped liquid.
\section{Discussion and conclusion}
Because of the large uncertainties on the kinetics of clathrate formation (see Sec. 2.3), the reliability of our conceptual model requires future laboratory experiments at conditions close to those encountered on Titan in order to be assessed. If alkanofers equilibrated with clathrate layers, then our computations should allow disentangling in situ measurements of lakes and rivers flowing from alkanofers from those of liquid areas directly sourced by precipitation. In the case of structure I clathrate formation, and irrespective of the temperature considered, the solution is dominated by ethane at mole fractions of the initial liquid reservoir trapped in clathrate lower than $\sim$0.9. At higher mole fractions, propane becomes the only species remaining in the liquid phase. In the case of structure II clathrate formation, ethane is the only dominant species in solution, irrespective of the temperature considered and mole fraction of entrapped liquid. These trends can be explained by the fact that ethane naturally enters the large cages of a structure I clathrate, while the large size of propane only allows this molecule to enter the large cages of a structure II clathrate. As both guests are very strong clathrate formers (e.g., Sloan \& Koh 2008), they compete for the same site. Therefore, in the case of formation of a structure I clathrate, propane remains in the liquid phase. Conversely, in the case of formation of a structure II clathrate, propane dominates in the large cages, forcing ethane to remain in the liquid phase. Our model then suggests that ethane and propane should be the discriminating markers of the clathrate structure forming from the solution when a significant fraction ($>$ 0.9) of the initial liquid reservoir has been entrapped.
In our model, any river or lake emanating from alkanofers possessing these characteristics should present a similar composition. On the other hand, lakes and rivers sourced by precipitation should contain substantial fractions of CH$_4$ and N$_2$, as well as minor traces of Ar and CO (see Table \ref{lake0}). Here, the lake compositions have been computed in the cases of liquid trapping in structures I and II clathrate reservoirs. However, it has been shown that a mixture dominated by ethane in presence of methane leads to the formation of a structure I clathrate (Takeya et al. 2003). We then believe that clathrate reservoirs formed from the lakes on Titan should be essentially of structure I.
Interestingly, if one postulates that the pores are fully filled by liquid hydrocarbons, it is possible to estimate the maximum porosity of the alkanofer that is consistent with a full enclathration of the solution. Assuming that the composition of the liquid reservoir is dominated by C$_2$H$_6$ in the case of structure I clathrate, the number of molecules present in the liquid is $N_L = \frac{\rho_{C_2H_6}~x V}{M_{C_2H_6}}$, with $x$ the porosity, $\rho_{C_2H_6}$ the density of the liquid, $M_{C_2H_6}$ the molecular mass of C$_2$H$_6$ and $V$ the volume of the alkanofer. On the other hand, the maximum number of available clathrate cages in the porous matrix is $N_c = \frac{\rho_{H2O}~(1-x) V}{7.66 M_{H_2O}}$, with $\rho_{H_2O}$ the density of solid water ice, $M_{H_2O}$ the molecular mass of H$_2$O, and 7.66 the number of H$_2$O molecules per enclathrated C$_2$H$_6$ molecule (Sloan \& Koh 2008). The value of $x$ that satisfies the conditions $N_L$/$N_c$ = 1 is the maximum porosity for which the number of trapped molecules is lower than or equal to the number of available clathrate cages. A larger value implies that all the liquid cannot be trapped as clathrates and a fraction of the liquid remains in the alkanofer. Assuming that the volume of ice remains constant during clathrate formation (Erfan-Niya et al. 2011), we find a maximum porosity value of $\sim$0.23. This value is larger than some estimates of 10\% to 15\% for the porosity of Titan's subsurface (Kossacki \& Lorenz 1996). The reservoir would be filled up until the ice is fully transformed into clathrates. Liquids filling the reservoir would then react with the clathrate matrix with exchange mechanisms such as those described in Choukroun \& Sotin (2012).
As mentioned in Sec. 2.3, our model can apply to the case where lakes and rivers located on Titan's surface directly equilibrate with a clathrate layer located beneath if the reequilibration timescale between the liquid and the atmosphere is short compared to the clathration timescale. Indeed, the massive atmosphere would serve only to buffer methane and N$_2$ (and the minor species CO and Ar), but not ethane and propane. Because the vapor pressures of ethane and propane are so small, the atmosphere is not a reservoir of those species. Hence methane and other minor gaseous compounds would draw into the atmosphere when the ethane/propane go into the clathrate, and would be again introduced into the lake/sea when ethane/propane are added. So, the methane abundance in the seas would adjust so as to be in a thermodynamically correct proportion to the ethane and propane in the lakes/seas for the given methane atmospheric humidity. In these conditions, the abundances of volatiles trapped in clathrate would correspond to the values computed when the mole fraction of entrapped liquid is very low and the compositions of coexisting lakes/seas would be very close to those given in Table \ref{lake0} when the clathrate reservoir is absent.
In any case, it must be borne in mind that the possible range of initial compositions of the liquid in equilibrium with the atmosphere (the one used in the starting composition of our liquid reservoirs) remains poorly constrained at present. Mole fractions predictions of current lake models vary by tens of percents (e.g., Cordier et al. 2012; Tan et al. 2013; Glein \& Shock 2013). Our predictions of clathrate/liquid equilibrium compositions are valid in any case where the mole fraction of C$_2$H$_6$ is prominent in the initial liquid reservoir. For example, similar conclusions are found when using the lake composition determined by Tan et al. (2013) at 93.7 K as the starting one of the alkanofer ($\sim$54\% of C$_2$H$_6$, 32\% of CH$_4$, 7\% of C$_3$H$_8$, and 7\% of N$_2$). On the other hand, our results strongly vary when using the liquid composition calculated by Tan et al. (2013) at 90 K ($\sim$8\% of C$_2$H$_6$, 69\% of CH$_4$, 1\% of C$_3$H$_8$, and 22\% of N$_2$), which is very different from the one they obtained at higher temperature. In this case, irrespective of the clathrate structure, the solution is dominated by N$_2$ and CH$_4$ at high mole fractions of entrapped liquid. The results of our model are wrong since the predicted mole fraction of remaining N$_2$, which can be as high as 90\%, exceeds its solubility limit in hydrocarbons ($\sim$10\%; Hibbard \& Evans 1968). To derive the correct composition of the solution, we would need to compute the simultaneous equilibrium between the clathrate reservoir, liquid and the generated N$_2$--rich atmosphere. New experimental data obtained on the atmosphere--liquid equilibrium composition at Titan's conditions are needed to refine the expected composition of the lakes. Another limitation of the model, and of all models of the composition and stability of mixed clathrates on Titan's surface or subsurface, is related to the paucity of experimental data available to constrain both the Kihara potential parameters for many clathrate formers (e.g. Choukroun et al. 2013) and the fugacity of these gases at conditions relevant to Titan's surface.
\section*{Acknowledgements}
O. M. acknowledges support from CNES. M.C. and C.S. acknowledge support from the NASA Outer Planets Research Program. Part of this work has been conducted at the Jet Propulsion Laboratory, California Institute of Technology, under contract to NASA. Government sponsorship acknowledged.
|
\section{Introduction}
A general intention of subspace signal processing is to partition the vector space of the observed data and isolate the subspace of the signal component(s) of interest from the disturbance (noise) subspace.
Subspace signal processing theory and practice rely, conventionally, on the familiar $L_2$-norm based singular-value decomposition (SVD) of the data matrix.
The SVD solution traces its origin to the fundamental problem of $L_2$-norm low-rank matrix approximation \cite{Eckart1936}, which is equivalent to the problem of maximum $L_2$-norm data projection with as many projection (``principal'') components as the desired low-rank value \cite{Golub1996}.
Among the many strengths of $L_2$-norm principal component analysis (PCA), one may point out the simplicity of the solution, scalability (new principal directions add on to the previous ones), and correspondence to maximum-likelihood estimation (MLE) under the assumption of additively Gaussian-noise corrupted data.
Practitioners have long observed, however, that $L_2$-norm PCA is sensitive to the presence of outlier values in the data matrix, that is, erroneous values that are away from the nominal data, appear only few times in the data matrix, and are not to appear again under normal system operation upon design.
Recently, there has been an --arguably small but growing-- interest in pursuing $L_1$-norm based approaches to deal with the problem of outliers in principal-components design [\citen{Ke2003}]-[\citen{Li2010}].%
\footnote{Absolute-value errors put significantly less emphasis on extreme errors than squared-error expressions.}
The growth in interest can also be credited incidentally to the popularity of compressed sensing methods [\citen{Donoho2006}]-[\citen{Gao}] that rely on $L_1$-based calculations in signal reconstruction.
This paper makes a case for $L_1$-subspace signal processing.
Interestingly, in contrast to $L_2$, subspace decomposition under the $L_1$ error minimization criterion and the $L_1$ projection maximization criterion are not the same.
A line of recent research pursues calculation of $L_1$ principal components under error minimization
[\citen{Ke2003}]-[\citen{Brooks2013b}].
The error surface is non-smooth and the problem non-convex resisting attempts to guaranteed optimization even with exponential computational cost.
Suboptimal algorithms may be developed by viewing the minimization function as a convex nondifferentiable function with a bounded Lipschitz constant \cite{Combettes2002}, \cite{Nesterov2013}.
A different approach is to calculate subspace components by $L_1$ projection maximization [\citen{Galpin1987}]-[\citen{Gu2012}].%
\footnote{A combined $L_1$/$L_2$-norm approach has been followed in [\citen{Ding2006}], [\citen{Li2010}].}
No algorithm has appeared so far with guaranteed convergence to the criterion-optimal subspace and no upper bounds are known on the expended computational effort.
In this present work, given any data matrix ${\mathbf X}\in \mathbb R^{D \times N}$ of $N$ signal samples of dimension $D$, we show that the general problem of finding the maximum $L_1$-projection principal component of $\mathbf X$ is formally NP-hard for asymptotically large $N$, $D$.
We prove, however, that the case of engineering interest of fixed given dimension $D$ is not NP-hard.
In particular, for the case where $N<D$, we present in explicit form an algorithm to find the optimal component with computational cost $2^N$.
For the case where the sample size exceeds the data dimension ($N \geq D$) --which is arguably of higher interest in signal processing applications-- we present an algorithm that computes the $L_1$-optimal principal component with complexity $\mathcal O\big(N^{\text{rank}(\mathbf X)}\big)$, $\text{rank}(\mathbf X) \leq D$.
We generalize the effort to the problem of calculating $K$, $1<K \leq \text{rank}( \mathbf X)$, $L_1$ components (necessarily a joint computational problem) and present an explicit optimal algorithm for multi-component subspace design of complexity ${\mathcal O}(N^{\text{rank}(\mathbf X)K-K+1})$.
We conclude with illustrations of the developed $L_1$ subspaces in problems from the fields of dimensionality reduction, direction-of-arrival estimation, and image reconstruction that demonstrate the inherent outlier resistance of $L_1$ subspace signal processing.
The rest of the paper is organized as follows.
Section II presents the problem statement and establishes notation.
Section III is devoted to the optimal computation of the $L_1$ principal component.
Section IV generalizes to optimal $L_1$-subspace calculation (joint multiple $L_1$ components).
Experimental illustrations are given in Section V and a few concluding remarks are drawn in Section VI.
\section{Problem Statement}
Consider $N$ real-valued measurements $\mathbf x_1, \mathbf x_2, \ldots, \mathbf x_N$ of dimension $D$ that form the $D\times N$ data matrix
\begin{equation}
{\bf X}=[{\bf x}_1\;\;{\bf x}_2\;\ldots\;{\bf x}_N].
\end{equation}
In the common version of the low-rank approximation problem, one seeks to describe (approximate) data matrix ${\bf X}$ by a rank-$K$ product ${\bf R}{\bf S}^T$ where ${\bf R} \in \mathbb R^{D \times K}$, ${\bf S} \in \mathbb R^{N \times K}$, $K \leq\min(D,N)$.
Given the observation data matrix ${\bf X}$, $L_2$-norm matrix approximation minimizes the sum of the element-wise squared error between the original matrix and its rank-$K$ surrogate in the form of Problem $\mathcal P_1^{L_2}$ defined below,
\begin{equation}
\begin{split}
{\mathcal P}^{L_2}_1:\;\;\;\;&\left({\bf R}_{L_2},{\bf S}_{L_2}\right)=\argmin_{{\bf R}\in{\mathbb R}^{D\times K},\;{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_2
\end{split}
\label{eq:RS}
\end{equation}
where $\|{\bf A}\|_2=\sqrt{\sum_{i,j}|A_{i,j}|^2}$ is the $L_2$ matrix norm (that is, Frobenius norm) of a matrix $\mathbf A$ with elements $A_{i,j}$.
Problem ${\mathcal P}^{L_2}_1$ is our most familiar $K$-singular-value-decomposition ($K$-SVD) problem solved with computational complexity $\mathcal O \big((D+N)\min^2(D,N)\big)$ \cite{Golub1996}.
$\mathcal P_1^{L_2}$ corresponds also to the statistical problem of maximum-likelihood estimation (MLE) of an unknown rank-$K$ matrix corrupted by additive element-wise independent Gaussian noise~\cite{Vantrees}.
We may expand~(\ref{eq:RS}) to $\displaystyle\min_{{\bf R}\in{\mathbb R}^{D\times K}}\min_{{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_2$ and inner minimization results to ${\bf S}={\bf X}^T{\bf R}$ for any fixed ${\bf R}$, $\mathbf R^T \mathbf R = \mathbf I_K$, by the Projection Theorem \cite{Golub1996}.
Hence, we obtain the equivalent problem
\begin{equation}
\begin{split}
{\mathcal P}^{L_2}_2:\;\;\;\;&{\bf R}_{L_2}=\argmin_{
{\bf R}\in{\mathbb R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}-{\bf R}{\bf R}^T{\bf X}\right\|_2
\end{split}
\label{eq:RR}
\end{equation}
frequently referred to as left-side $K$-SVD.
Since $\left\|{\bf A}\right\|_2^2=\text{tr}\left({\bf A}^T{\bf A}\right)$ where $\text{tr}(\cdot)$ denotes the trace of a matrix, the $L_2$ error minimization problem ${\mathcal P}^{L_2}_2$ is also equivalent to the $L_2$ projection (energy) maximization problem
\begin{equation}
\begin{split}
{\mathcal P}^{L_2}_3:\;\;\;\;&{\bf R}_{L_2}=\argmax_{{\bf R}\in{\mathbb R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_2.
\end{split}
\label{eq:R}
\end{equation}
The optimal ${\bf R}_{L_2}$ (in ${\mathcal P}^{L_2}_1$, ${\mathcal P}^{L_2}_2$, and ${\mathcal P}^{L_2}_3$) is known simply as the $K$ dominant-singular-value left singular vectors of the original data matrix or $K$ dominant-eigenvalue eigenvectors of $\mathbf X \mathbf X^T$~\cite{Eckart1936},~\cite{Golub1996}.
Note that, if $K<D$ and we possess the solution ${\bf R}_{L_2}^{(K)}$ for $K$ singular/eigen vectors in (\ref{eq:RS}), (\ref{eq:RR}), (\ref{eq:R}), then the solution for rank $K+1$ is derived readily by ${\bf R}_{L_2}^{(K+1)}=\left[{\bf R}_{L_2}^{(K)}\;\;{\bf r}_{L_2}^{(K+1)}\right]$ with
\begin{equation}
{\bf r}_{L_2}^{(K+1)}=\argmax_{{\bf r}\in{\mathbb R}^D,\;\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T \left({\bf I}_D-{\bf R}_{L_2}^{(K)}{{\bf R}_{L_2}^{(K)}}^T\right){\bf r} \right\|_2.
\end{equation}
This is known as the PCA scalability property.
$L_2$ PCA, as reviewed above in ${\mathcal P}^{L_2}_1$, ${\mathcal P}^{L_2}_2$, and ${\mathcal P}^{L_2}_3$, has a simple solution, is scalable (new principal directions add on to the previous ones), and corresponds to MLE under the assumption of Gaussian additively corrupted data.
Practitioners, however, have long noticed a drawback.
By minimizing the sum of squared errors, $L_2$ principal component calculation becomes sensitive to extreme error value occurrences caused by the presence of outlier measurements in the data matrix (measurements that are numerically distant from the nominal data, appear only few times in the data matrix, and are not to appear under normal system operation upon design).
Motivated by this observed drawback of $L_2$ subspace signal processing, in this work we study and pursue subspace-decomposition approaches that are based on the $L_1$ norm,
\begin{equation}
\left\|{\bf A}\right\|_1=\sum_{i,j}\left|A_{i,j}\right|.
\end{equation}
We may ``translate" the three equivalent $L_2$ optimization problems (\ref{eq:RS}), (\ref{eq:RR}), (\ref{eq:R}) to new problems that utilize the $L_1$ norm as follows,
\begin{align}
\begin{split}
&{\mathcal P}^{L_1}_1:\;\;\;\;\left({\bf R}_{L_1},{\bf S}_{L_1}\right)=\argmin_{{\bf R}\in{\mathbb R}^{D\times K},\;{\bf R}^T{\bf R}={\bf I}_K,\;{\bf S}\in{\mathbb R}^{N\times K}}\left\|{\bf X}-{\bf R}{\bf S}^T\right\|_1,
\end{split} \\
\label{eq:RSL1}
\begin{split}
&{\mathcal P}^{L_1}_2:\;\;\;\;{\bf R}_{L_1}=\argmin_{{\bf R}\in{\mathbb R}^{D\times K},\; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}-{\bf R}{\bf R}^T{\bf X}\right\|_1,
\end{split} \\
\begin{split}
&{\mathcal P}^{L_1}_3:\;\;\;\;{\bf R}_{L_1}=\argmax_{{\bf R}\in{\mathbb R}^{D\times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_1.
\end{split}
\label{eq:RL1}
\end{align}
A few comments appear useful at this point:
(i) ${\mathcal P}^{L_1}_1$ corresponds to MLE when the additive noise disturbance follows a Laplacian distribution \cite{Vantrees}.
(ii) The optimal metric value in ${\mathcal P}^{L_1}_3$ with a single dimension ($K=1$) is the complexity parameter for saddle-point methods when used to provide an approximate solution to the $\ell_1$/nuclear-norm Dantzig selector problem \cite{Nesterov2013}.
(iii) Under the $L_1$ norm, the three optimization problems ${\mathcal P}^{L_1}_1$, ${\mathcal P}^{L_1}_2$, and ${\mathcal P}^{L_1}_3$ are no longer equivalent.
(iv) Under $L_1$, the PCA scalability property does not hold (due to loss of the Projection Theorem).
(v) Even for reduction to a single dimension (rank $K=1$ approximation), the three problems are difficult to solve.
(vi) As of today, it is unknown which of the subspaces defined in $\mathcal P_{1}^{L_1}$, $\mathcal P_{2}^{L_1}$, and $\mathcal P_{3}^{L_1}$ exhibits stronger resistance against faulty measurements; indeed, none of these problems had been solved optimally so far for general $D,K$.
In this present work, we focus exclusively on ${\mathcal P}^{L_1}_3$.
In Section III, we seek to find efficiently the principal maximum $L_1$ projection component of ${\bf X}$.
In Section IV, we investigate the problem of calculating (jointly necessarily) multiple ($K>1$) $L_1$ projection components that maximize the $L_1$ ``energy" of the data on the projection subspace.
\section{The $L_1$-norm Principal Component}
\label{sec:OneComponent}
In this section, we concentrate on the calculation of the $L_1$-maximum-projection component of a data matrix $\mathbf X \in \mathbb R^{D \times N}$ (Problem $\mathcal P_3^{L_1}$ in (\ref{eq:RL1}), $K=1$).
First, we show that the problem is in general NP-hard and review briefly suboptimal techniques from the literature.
Then, we prove that, if the data dimension $D$ is fixed, the principal $L_1$-norm component is in fact computable in polynomial time and present an algorithm that calculates the $L_1$ principal component of $\mathbf X$ with complexity ${\mathcal O}\left(N^{\text{rank}({\bf X})}\right)$, $\text{rank}({\bf X})\leq D$.
\subsection{Hardness of Problem and an Exhaustive-search Algorithm Over the Binary Field}
We present a fundamental property of Problem ${\mathcal P}^{L_1}_3$, $K=1$, that will lead us to an efficient solution.
The property is presented in the form of Proposition \ref{prop:quad} below and interprets ${\mathcal P}^{L_1}_3$ as an equivalent quadratic-form maximization problem over the binary field.
\begin{myprop}
For any data matrix $\mathbf X \in \mathbb R^{D \times N}$, the solution to
${\mathcal P}^{L_1}_3:{\bf r}_{L_1}=\argmax_{{\bf r}\in{\mathbb R}^D,\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1$
is given by
\begin{equation}
{\bf r}_{L_1}=\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}
\label{eq:rL1}
\end{equation}
where\begin{equation}
{\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2=\argmax_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf X}{\bf b}.
\label{eq:bopt}
\end{equation}
In addition,
$\left\|{\bf X}^T {\bf r}_{L_1}\right\|_1=\left\|{\bf X}{\bf b}_\text{{opt}}\right\|_2$.
\label{prop:quad}
\end{myprop}
\noindent \emph{Proof:}
For any ${\bf z}\in{\mathbbm R}^N$, $\displaystyle\left\|{\bf z}\right\|_1=\text{sgn}\left({\bf z}\right)^T{\bf z}=\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf z}$.
Therefore, we can rewrite the optimization problem as
\begin{equation}
\max_{\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1=\max_{\left\|{\bf r}\right\|_2=1}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf r}=\max_{{\bf b}\in\{\pm1\}^N}\max_{\left\|{\bf r}\right\|_2=1}{\bf r}^T{\bf X}{\bf b}.
\label{eq:rXb}
\end{equation}
For any fixed vector ${\bf b}$, inner maximization in \eqref{eq:rXb} is solved by ${\bf r}=\frac{{\bf X}{\bf b}}{\left\|{\bf X}{\bf b}\right\|_2}$ and
\begin{equation}
\max_{\left\|{\bf r}\right\|_2=1}{\bf r}^T{\bf X}{\bf b}=\left\|{\bf X}{\bf b}\right\|_2.
\label{eq:Xb2}
\end{equation}
Combining~(\ref{eq:rXb}) and~(\ref{eq:Xb2}), we obtain
\begin{equation}
\max_{\left\|{\bf r}\right\|_2=1}\left\|{\bf X}^T{\bf r}\right\|_1=\max_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2.
\end{equation}
That is, $
\left\|{\bf X}^T{\bf r}_{L_1}\right\|_1=\left\|{\bf X}{\bf b}_\text{opt}\right\|_2$
where ${\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}\left\|{\bf X}{\bf b}\right\|_2$ and
${\bf r}_{L_1}=\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}$.
\hfill
\IEEEQEDclosed
By Proposition \ref{prop:quad}, to find the principal $L_1$-norm component ${\bf r}_{L_1}$ we solve~(\ref{eq:bopt}) to obtain ${\bf b}_\text{opt}$ and then calculate $\frac{{\bf X}{\bf b}_\text{opt}}{\left\|{\bf X}{\bf b}_\text{opt}\right\|_2}$.
The straightforward approach to solve (\ref{eq:bopt}) is an exhaustive search among all $2^N$ binary vectors of length $N$.
Therefore, with computational cost $2^N$, Proposition \ref{prop:quad} identifies the $L_1$-optimal principal component of ${\bf X}$.
As the data record size $N$ grows, calculation of the $L_1$ principal component by exhaustive search in (\ref{eq:bopt}) becomes quickly infeasible.
Proposition \ref{prop:NPhard} below declares that, indeed, in its general form $\mathcal P_3^{L_1}$, $K=1$, is NP-hard for jointly asymptotically large $N, D$. McCoy and Tropp provide an alternative proof in \cite{McCoy2011}, that is the earliest known to the authors.
\begin{myprop}
The computation of the $L_1$ principal component of ${\bf X}\in{\mathbb R}^{D\times N}$ by maximum $L_1$-norm projection (Problem $\mathcal P_{3}^{L_1}$, $K=1$) is NP-hard in jointly asymptotic $N,D$.
\label{prop:NPhard}
\end{myprop}
\noindent \emph{Proof:}
In~(\ref{eq:rXb}), for any fixed ${\bf r}\in{\mathbb R}^D$, ${\bf b}=\text{sgn}\left({\bf X}^T{\bf r}\right)$.
Hence,
\begin{equation}
{\bf b}_\text{opt}=\text{sgn}\left({\bf X}^T{\bf r}_{L_1}\right).
\label{eq:brL1}
\end{equation}
By~(\ref{eq:rL1}) and~(\ref{eq:brL1}), computation of the $L_1$ principal component of ${\bf X}$ is equivalent to computation of ${\bf b}_\text{opt}$ in~(\ref{eq:bopt}).
Consider the special case of~(\ref{eq:bopt}) where ${\bf X}^T{\bf X}={\bf I}_N-{\bf a}{\bf a}^T$, $\mathbf a \in \mathbb R^{N}$,
$\left\|{\bf a}\right\|_2=1$ (hence, $D=N-1$).
Then,
\begin{equation}
\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf X}^T{\bf X}{\bf b}=\max_{{\bf b}\in\{\pm1\}^N}\left\{\left\|{\bf b}\right\|_2^2-\left({\bf b}^T{\bf a}\right)^2\right\}=N-\min_{{\bf b}\in\{\pm1\}^N}\left({\bf b}^T{\bf a}\right)^2.
\end{equation}
But $\min_{\mathbf b \in \{\pm 1 \}^{N}}(\mathbf b^T \mathbf a)^2$ is the NP-complete equal-partition problem~\cite{Garey1979}.
We conclude that computation of the $L_1$ principal component of ${\bf X}$ is NP-hard in jointly asymptotic $N,D$.
\hfill
\IEEEQEDclosed
\subsection{Existing Approaches in Literature}
Recently there has been a growing documented effort to calculate subspace components by $L_1$ projection maximization~[\citen{Galpin1987}]-[\citen{Gu2012}].
The work in~\cite{Kwak2008} presented a suboptimal iterative algorithm for the computation of ${\bf r}_{L_1}$, which, following the formulation and notation of this present paper, initializes the solution to some arbitrary component $\mathbf r_{L_1}^{(0)}$ and executes
\begin{align}
{\bf b}^{(i+1)}&=\text{sgn}\left({\bf X}^T{\bf r}_{L_1}^{(i)}\right),
\label{eq:Kwak1}\\
{\bf r}_{L_1}^{(i+1)}&=\frac{{\bf X}{\bf b}^{(i+1)}}{\left\|{\bf X}{\bf b}^{(i+1)}\right\|_2},
\label{eq:Kwak2}
\end{align}
$i=0,1,2,\ldots$, until convergence.
The work in~\cite{Nie2011} presented an iterative algorithm for the joint computation of $K\geq1$ principal $L_1$-norm components.
For the case where $K=1$, the iteration in~\cite{Nie2011} simplifies to the iteration in~\cite{Kwak2008} (that is, (\ref{eq:Kwak1}),~(\ref{eq:Kwak2}) above).
Therefore, for $K=1$, the algorithms in~\cite{Kwak2008}, \cite{Nie2011} are identical and can, in fact, be described by the simple single iteration
\begin{equation}
{\bf b}^{(i+1)}=\text{sgn}\left({\bf X}^T{\bf X}{\bf b}^{(i)}\right),\;\;\;i=1,2,\ldots,
\label{eq:Kwak3}
\end{equation}
for the computation of ${\bf b}_\text{opt}$ in~(\ref{eq:bopt}).
Equation \eqref{eq:Kwak3}, however, does not guarantee convergence to the $L_1$-optimal component solution (convergence to one of the many local maxima may be observed).
In the following section, we present for the first time in the literature an optimal algorithm to calculate the $ L_1$ principal component of a data matrix with complexity polynomial in the sample size $N$ when the data dimension $D$ is fixed.
\subsection{Exact Computation of the $L_1$ Principal Component in Polynomial Time}
Proposition \ref{prop:NPhard} proves NP-hardness of the computation of the $L_1$ principal component ${\bf r}_{L_1}$ in $N,D$ (that is, when $N,D$ are jointly arbitrarily large).
However, of engineering interest is the case of fixed data dimension $D$.
In the following, we show for the first time in the literature that, if $D$ is fixed, then computation of ${\bf r}_{L_1}$ is no longer NP-hard (in $N$).
We state our result in the form of Proposition \ref{prop:polynomial} below.
\begin{myprop}
For any fixed data dimension $D$, computation of the $L_1$ principal component of ${\bf X}\in{\mathbb R}^{D\times N}$ has complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})}\right)$, $\text{{rank}}({\bf X})\leq D$.
\hfill
\IEEEQEDclosed
\label{prop:polynomial}
\end{myprop}
By Proposition \ref{prop:NPhard}, computation of the $ L_1$ principal component of $\mathbf X$ is equivalent to computation of $\mathbf b_{\text{opt}}$ in (\ref{eq:bopt}).
To prove Proposition \ref{prop:polynomial}, we will prove that ${\bf b}_\text{opt}$ can be computed with complexity ${\mathcal O}\left(N^{\text{rank}({\bf X})}\right)$.
We begin our developments by defining
\begin{equation}
d\stackrel{\triangle}{=}\text{rank}({\bf X})\leq D.
\label{eq:d}
\end{equation}
Then, ${\bf X}^T{\bf X}$ also has rank $d$ and can be decomposed by
\begin{equation}
{\bf X}^T{\bf X}={\bf Q}{\bf Q}^T,\;\;\;{\bf Q}_{N\times d}=\left[{\bf q}_1\;{\bf q}_2\;\ldots\;{\bf q}_d\right],\;\;\;{\bf q}_i^T{\bf q}_j=0,\;i\neq j,
\label{eq:QQ}
\end{equation}
where ${\bf q}_1$, ${\bf q}_2$, $\ldots$ , ${\bf q}_d$ are the $d$ eigenvalue-weighted eigenvectors of ${\bf X}^T{\bf X}$ with nonzero eigenvalue.
By~(\ref{eq:bopt}),
\begin{equation}
{\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf Q}{\bf Q}^T{\bf b}=\argmax_{{\bf b}\in\{\pm1\}^N}\left\|{\bf Q}^T{\bf b}\right\|_2.
\label{eq:Qb}
\end{equation}
For the case $N<D$, the optimal binary vector ${\bf b}_\text{opt}$ can be obtained directly from~(\ref{eq:bopt}) by an exhaustive search among all $2^N$ binary vectors ${\bf b}\in\{\pm1\}^N$.
Therefore, we can design the $L_1$-optimal principal component ${\bf r}_{L_1}$ with computational cost $2^N<2^D={\mathcal O}(1)$.
For the case where the sample size exceeds the data dimension ($N\geq D$),
we find it useful in terms of both theory and practice to present our developments separately for data rank $d=1$, $d=2$, and $2 < d \leq D$.\\
\emph{1) Case $d=1$:}
If the data matrix has rank $d=1$, then ${\bf Q}={\bf q}_1$ and~(\ref{eq:Qb}) becomes
\begin{equation}
{\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}\left|{\bf q}_1^T{\bf b}\right|=\text{sgn}\left({\bf q}_1\right).
\end{equation}
By~(\ref{eq:rL1}), the $L_1$-optimal principal component is
\begin{equation}
{\bf r}_{L_1}=\frac{{\bf X}\,\text{sgn}\left({\bf q}_1\right)}{\left\|{\bf X}\,\text{sgn}\left({\bf q}_1\right)\right\|_2}
\label{dequals1}
\end{equation}
designed with complexity ${\mathcal O}\left(N\right)$.
It is of notable practical importance to observe at this point that even when $\mathbf X$ is not of true rank one, \eqref{dequals1} presents us with a quality, trivially calculated approximation of the $L_1$ principal component of $\mathbf X$:
Calculate the $L_2$ principal component $\mathbf q_1$ of the $N \times N$ matrix $\mathbf X^T \mathbf X$, quantize to $\text{sgn}(\mathbf q_1)$, and project and normalize to obtain $\mathbf r_{L_1} \simeq \mathbf X\,\text{sgn}(\mathbf q_1) / \| \mathbf X\,\text{sgn}(\mathbf q_1) \|_2$.\\
\emph{2) Case $d=2$:}
If $d=2$, then ${\bf Q}=\left[{\bf q}_1\;\;{\bf q}_2\right]$ and~(\ref{eq:Qb}) becomes
\begin{equation}
{\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}\left\{\left({\bf q}_1^T{\bf b}\right)^2+\left({\bf q}_2^T{\bf b}\right)^2\right\}.
\label{eq:rank2}
\end{equation}
The binary optimization problem \eqref{eq:rank2} was seen and solved in \cite{KP} by the auxiliary-angle method \cite{mackenthun}, which was also used earlier in [\citen{sweldens}],[\citen{achilles2}].
Here, we define the $N\times1$ complex vector
\begin{equation}
{\bf z}\stackrel{\triangle}{=}{\bf q}_1+j{\bf q}_2
\end{equation}
and rewrite~(\ref{eq:rank2}) as
\begin{equation}
{\bf b}_\text{opt}=\argmax_{{\bf b}\in\{\pm1\}^N}\left|{\bf b}^T{\bf z}\right|.
\label{eq:bz}
\end{equation}
We introduce the auxiliary angle $\phi\in\left[-\pi,\pi\right)$ and note that, for any complex scalar $w$,
\begin{equation}
\text{Re}\left(we^{-j\phi}\right)\leq|w|
\end{equation}
with equality if and only if $\phi=\text{angle}\left(w\right)$.
That is,
\begin{equation}
|w|=\max_{\phi\in\left[-\pi,\pi\right)}\text{Re}\left(we^{-j\phi}\right).
\end{equation}
Therefore, the maximization in~(\ref{eq:bz}) can be rewritten as
\begin{align}
\max_{{\bf b}\in\{\pm1\}^N}\left|{\bf b}^T{\bf z}\right|&=\max_{{\bf b}\in\{\pm1\}^N}\max_{\phi\in\left[-\pi,\pi\right)}\text{Re}\left({\bf b}^T{\bf z}e^{-j\phi}\right)=\max_{\phi\in\left[-\pi,\pi\right)}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T\text{Re}\left({\bf z}e^{-j\phi}\right)
\label{eq:phib}
\end{align}
where, for any given angle $\phi\in\left[-\pi,\pi\right)$, inner maximization is achieved by
\begin{equation}
{\bf b}\left(\phi\right)=\text{sgn}\left(\text{Re}\left({\bf z}e^{-j\phi}\right)\right).
\label{eq:bphi}
\end{equation}
Then, the optimal vector ${\bf b}_\text{opt}$ in~(\ref{eq:bz}), i.e., the solution to~(\ref{eq:bopt}), is met if we scan the entire interval $\left[-\pi,\pi\right)$ and collect the locally optimal vector ${\bf b}\left(\phi\right)$ for any point $\phi\in\left[-\pi,\pi\right)$.
Interestingly, as we scan the interval $\left[-\pi,\pi\right)$, the locally optimal vector ${\bf b}\left(\phi\right)$ does not change unless the sign of $\text{Re}\left(z_ne^{-j\phi}\right)$ changes for some $n=1,2,\ldots,N$.
Since the latter happens only at $\text{angle}\left(z_n\right)$ and $\text{angle}\left(z_n\right)+\pi$, we obtain $2N$ points in total at which ${\bf b}\left(\phi\right)$ changes.
Next, we order the $2N$ points with complexity ${\mathcal O}\left(2N\log_22N\right)$ and create successively $2N$ binary vectors by changing each time the sign of $b_n$ if the $n$th element of ${\bf z}$ is the one that determines a sign change.
It is observed that the $2N$ binary vectors that we obtain this way are pair-wise opposite (the vectors that are collected when $\phi\in\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$ are opposite to the ones that are collected when $\phi\in\left[-\pi,-\frac{\pi}{2}\right)\cup\left[\frac{\pi}{2},\pi\right)$).
Since opposite vectors result in the same metric value in~(\ref{eq:bopt}), we can restrict our search to $\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$ and maintain optimality.
Therefore, with overall complexity ${\mathcal O}\left(N\log_2N\right)$, we obtain a set of $N$ binary vectors that contains ${\bf b}_\text{opt}$.
Then, we only have to evaluate the $N$ vectors against the metric of interest in~(\ref{eq:bopt}) to obtain ${\bf b}_\text{opt}$.
We conclude that the $L_1$-optimal principal component of a rank-$2$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ is designed with complexity ${\mathcal O}\left(N\log_2N\right)$.
\emph{3) Case $d>2$:}
If $d>2$, we design the $L_1$-optimal principal component of ${\bf X}$ with complexity ${\mathcal O}\left(N^d\right)$ by considering the multiple-auxiliary-angle approach that was presented in~\cite{KL} as a generalization of the work in~\cite{KP}.
Consider a unit vector ${\bf c}\in{\mathbbm R}^d$.
By the Cauchy-Schwartz inequality, for any ${\bf a}\in{\mathbbm R}^d$,
\begin{equation}
{\bf a}^T{\bf c}\leq\left\|{\bf a}\right\|_2\left\|{\bf c}\right\|_2=\left\|{\bf a}\right\|_2
\end{equation}
with equality if and only if ${\bf c}$ is codirectional with ${\bf a}$.
Then,
\begin{equation}
\max_{{\bf c}\in{\mathbbm R}^d,\;\left\|{\bf c}\right\|_2=1}{\bf a}^T{\bf c}=\left\|{\bf a}\right\|_2.
\label{eq:ca}
\end{equation}
By~(\ref{eq:ca}), the optimization problem in~(\ref{eq:Qb}) becomes
\begin{equation}
\max_{{\bf b}\in\{\pm1\}^N}\left\|{\bf Q}^T{\bf b}\right\|_2=\max_{{\bf b}\in\{\pm1\}^N}\max_{{\bf c}\in{\mathbbm R}^d,\; \left\|{\bf c}\right\|_2=1}{\bf b}^T{\bf Q}{\bf c}=\max_{{\bf c}\in{\mathbbm R}^d,\; \left\|{\bf c}\right\|_2=1}\max_{{\bf b}\in\{\pm1\}^N}{\bf b}^T{\bf Q}{\bf c}.
\label{eq:maxmax}
\end{equation}
For every ${\bf c}\in{\mathbbm R}^d$, inner maximization in~(\ref{eq:maxmax}) is solved by the binary vector
\begin{equation}
{\bf b}({\bf c}) = \text{sgn}({\bf Q}{\bf c}),
\label{eq:bc}
\end{equation}
which is obtained with complexity ${\mathcal O}(N)$.
Then, by~(\ref{eq:maxmax}), the solution to the original problem in~(\ref{eq:Qb}) is met if we collect all binary vectors ${\bf b}({\bf c})$ returned as ${\bf c}$ scans the unit-radius $d$-dimensional hypersphere.
That is, ${\bf b}_\text{opt}$ in~(\ref{eq:Qb}) is in%
\footnote{The $d$th element of vector $\mathbf c$, $c_d$, can be set nonnegative without loss of optimality, because, for any given $\mathbf c$, $\| \mathbf c\|_2=1$, the binary vectors $\mathbf b(\mathbf c)$ and $\mathbf b(\text{sgn}(c_d)\mathbf c)$ result to the same metric value in \eqref{eq:Qb}.}
\begin{equation}
{\mathcal S}_1\stackrel{\triangle}{=}\hspace{-.5cm}\bigcup_{{\bf c}\in{\mathbbm R}^d,\,\left\|{\bf c}\right\|_2=1, \,c_d\geq0}\hspace{-.5cm}{\bf b}({\bf c}).
\label{eq:S1}
\end{equation}
Two fundamental questions for the computational problem under consideration are what the size (cardinality) of set ${\mathcal S}_1$ is and how much computational effort is expended to form ${\mathcal S}_1$.
We address first the first question.
We introduce the auxiliary-angle vector ${\boldsymbol\phi} =[\phi_1,\, \phi_2,\, \ldots, \,\phi_{d-1}]^T \in\Phi^{d-1}$, $\Phi\stackrel{\triangle}{=}\left[-\frac{\pi}{2},\frac{\pi}{2}\right)$, and parametrize ${\bf c}$ as follows,
\begin{equation}
{\bf c}(\boldsymbol{\phi})\stackrel{\triangle}{=}\begin{bmatrix}
\sin\phi_1\\
\cos\phi_1\sin\phi_2\\
\cos\phi_1\cos\phi_2\sin\phi_3\\
\vdots\\
\cos\phi_1\ldots\cos\phi_{d-2}\sin\phi_{d-1}\\
\cos\phi_1\ldots\cos\phi_{d-2}\cos\phi_{d-1}
\end{bmatrix}.
\end{equation}
Then, we re-express the candidate set in~(\ref{eq:S1}) in the form
\begin{equation}
{\mathcal S}_1=\bigcup_{{\boldsymbol\phi}\in\Phi^{d-1}}{\bf b}({\boldsymbol\phi})
\label{eq:S1phi}
\end{equation}
where, according to~(\ref{eq:bc}),
\begin{equation}
{\bf b}({\boldsymbol\phi}) =
\begin{bmatrix}
b_1({\boldsymbol\phi}), & b_2({\boldsymbol\phi}), & \ldots, & b_{N}({\boldsymbol\phi})
\end{bmatrix}^T
=\text{sgn}({\bf Q}{\bf c}({\boldsymbol\phi})).
\end{equation}
We note that, for any point ${\boldsymbol\phi}$, each element $b_n({\boldsymbol\phi})$, $n=1,2, \ldots, N$, depends only on the corresponding row of ${\bf Q}$ and is determined by $b_n({\boldsymbol{\phi}})=\text{sgn}({\bf Q}_{n,:}\,{\bf c}({\boldsymbol{\phi}}))$.
Hence, the value of the binary element $b_n({\boldsymbol{\phi}})$ changes only when
\begin{equation}
{\bf Q}_{n,:}\,{\bf c}({\boldsymbol{\phi}})=0.
\label{eq:Vnc}
\end{equation}
To gain some insight into the process of introducing the auxiliary-angle vector $\boldsymbol{\phi}$, we notice that the points $\boldsymbol{\phi}$ that satisfy (\ref{eq:Vnc}) determine a hypersurface (or $(d-2)$-manifold) in the $(d-1)$-dimensional space that partitions $\Phi^{d-1}$ into two regions.
One region corresponds to $b_n=-1$ and the other corresponds to $b_n=+1$.
A key observation in the algorithm is that, as ${\boldsymbol{\phi}}$ scans any of the two regions, the decision on $b_n$ does not change.
Therefore, the $N$ rows of ${\bf Q}$ are associated with $N$ corresponding hypersurfaces that partition $\Phi^{d-1}$ into $P_1$ cells $C_1,C_2,\dots,C_{P_1}$ such that $\bigcup_{p=1}^{P_1}C_p=\Phi^{d-1}$, $C_p\cap C_q=\emptyset$ $\forall$ $p\ne q$, and each cell $C_p$ corresponds to a distinct vector ${\bf b}_p\in\{\pm1\}^N$.
As a result, the candidate vector set is $S_1=\bigcup_{p=1}^P\{{\bf b}_p\}$.
In~\cite{KL}, it was shown that $P_1=\sum_{g=0}^{d-1}\binom{N-1}{g}$ if pairs of cells that correspond to opposite binary vectors (hence, equivalent vectors with respect to the metric of interest in~(\ref{eq:Qb})) are considered as one.
Therefore, the candidate vector set ${\mathcal S}_1$ has cardinality $|{\mathcal S}_1|=\sum_{g=0}^{d-1}\binom{N-1}{g}={\mathcal O}\left(N^{d-1}\right)$.
Fig.~\ref{fig:surfaces} presents a visualization of the algorithm/partition for the case of a data matrix $\mathbf X_{D \times N}$ of $N=8$ samples with rank $d=3\leq D\leq N$.
Since $d=3$, the hypersurfaces (or $(d-2)$-manifolds) are, in fact, curves in the $2$-dimensional space that partition $\Phi^2$ into cells.
The $P=\binom{7}{0} + \binom{7}{1} + \binom{7}{2} = 29 $ cells and associated binary candidate vectors are formed by the eight-row three-column eigenvector matrix ${\bf Q}$ of $\mathbf X^T \mathbf X$ and the scanning angle vector $\mathbf c ({\boldsymbol\phi})=[\sin \phi_1, \; \cos \phi_1 \sin \phi_2, \; \cos \phi_1 \cos \phi_2]^T$.
Regarding the cost of calculating $\mathcal S_1$, since each cell ${ C}$ contains at least one vertex (that is, intersection of $d-1$ hypersurfaces), see for example Fig.~\ref{fig:surfaces}, it suffices to find all vertices in the partition and determine ${\bf b}$ for all neighboring cells.
Consider $d-1$ arbitrary hypersurfaces; say, for example, ${\bf Q}_{1,:}{\bf c}({\boldsymbol\phi})=0$, ${\bf Q}_{2,:}{\bf c}({\boldsymbol\phi})=0$, $\ldots$, ${\bf Q}_{d-1,:}{\bf c}({\boldsymbol\phi})=0$.
Their intersection satisfies ${\bf Q}_{1,:}{\bf c}({\boldsymbol\phi})={\bf Q}_{2,:}{\bf c}({\boldsymbol\phi})=\ldots={\bf Q}_{d-1,:}{\bf c}({\boldsymbol\phi})=0$ and is computed by solving the equation
\begin{equation}
{\bf Q}_{1:d-1,:}{\bf c}({\boldsymbol\phi})={\bf 0}.
\label{eq:Qc}
\end{equation}
The solution to~(\ref{eq:Qc}) consists of the spherical coordinates of the unit vector in the null space of the $(d-1)\times d$ matrix ${\bf Q}_{1:d-1,:}$.%
\footnote{If ${\bf Q}_{1:d-1,:}$ is full-rank, then its null space has rank $1$ and ${\bf c}({\boldsymbol\phi})$ is uniquely determined (within a sign ambiguity which is resolved by $c_d\geq0$).
If, instead, ${\bf Q}_{1:d-1,:}$ is rank-deficient, then the intersection of the $d-1$ hypersurfaces (i.e., the solution of~(\ref{eq:Qc})) is a $p$-manifold (with $p\geq1$) in the $(d-1)$-dimensional space and does not generate a new cell.
Hence, linearly dependent combinations of $d-1$ rows of ${\bf Q}$ are ignored.}
Then, the binary vector ${\bf b}$ that corresponds to a neighboring cell is computed by
\begin{equation}
\text{sgn}({\bf Q}\,{\bf c}({\boldsymbol\phi}))
\label{eq:sgnQc}
\end{equation}
with complexity ${\mathcal O}(N)$.
Note that~(\ref{eq:sgnQc}) presents ambiguity regarding the sign of the intersecting $d-1$ hypersurfaces.
A straightforward way to resolve the ambiguity%
\footnote{An alternative way of resolving the sign ambiguities at the intersections of hypersurfaces was developed in~\cite{KL} and led to the direct construction of a set ${\mathcal S}_1$ of size $\sum_{g=0}^{d-1}\binom{N-1}{g}={\mathcal O}(N^{d-1})$ with complexity ${\mathcal O}(N^d)$.}
is to consider all $2^{d-1}$ sign combinations for the corresponding elements $b_1,b_2,\ldots,b_{d-1}$ and obtain the binary vectors of all $2^{d-1}$ neighboring cells.
Finally, we repeat the above procedure for any combination of $d-1$ intersecting hypersurfaces among the $N$ ones.
Therefore, the total number of binary candidates that we obtain (i.e., the cardinality of ${\mathcal S}_1$) is upper bounded by $2^{d-1}\binom{N}{d-1}={\mathcal O}(N^{d-1})$.
Since complexity ${\mathcal O}(N)$ is required for each combination of $d-1$ rows of ${\bf Q}$ to solve~(\ref{eq:sgnQc}), the overall complexity of the construction of ${\mathcal S}_1$ is ${\mathcal O}(N^d)$ for any given matrix ${\bf Q}_{N\times d}$.
Our complete, new algorithm for the computation of the $L_1$-optimal principal component of a rank-$d$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ that has complexity ${\mathcal O}\left(N^d\right)$ is presented in detail in Fig.~\ref{fig:algo}.
Computation of each element of $S_1$ (i.e., column of ${\bf B}$ in the algorithm) is performed independently of each other.
Therefore, the proposed algorithm is fully parallelizable.
The space complexity of the algorithm is $O(N)$, since after every computation of a new binary candidate the best binary candidate needs to be stored.
We note that the required optimal binary vector in (\ref{eq:Qb}) can, alternatively, be computed through the algorithm in \cite{Allemand2001}, \cite{Ferrez2005} with time complexity ${\mathcal O}\left(N^{d+1}\right)$ and space complexity at least ${\mathcal O}(N)$ based on the reverse search for cell enumeration in arrangements~\cite{Avis1996} or with time complexity ${\mathcal O}(N^{d-1})$ but space complexity proportional to ${\mathcal O}\left(N^{d-1}\right)$ based on the incremental algorithm for cell enumeration in arrangements \cite{Edelsbrunner1986}, \cite{Edelsbrunner1987}.
Another algorithm that can solve (\ref{eq:Qb}) with polynomial complexity is in \cite{BenAmeur2010}.
Its time complexity is ${\mathcal O}(N^{d-1}\log N)$, while its space complexity is polynomially bounded by the output size (i.e., ${\mathcal O}(N^{d-1})$).
In comparison to the above approaches, the algorithm in Fig. 2 is the fastest known
with smallest (linear) space complexity.
We conclude that the $L_1$-optimal principal component of a rank-$d$ data matrix ${\bf X}\in{\mathbb R}^{D\times N}$, $d \leq D \leq N$, is obtained with time complexity ${\mathcal O}\left(N^d\right)$ and space complexity ${\mathcal O}(N)$.
That is, the time complexity is polynomial in the sample size with exponent equal to the rank of the data matrix, which is at most equal to the data dimension $D$.
The space complexity is linear in the sample size.
\section{Multiple $L_1$-norm Principal Components}
In this section, we switch our interest to the joint design of $K>1$ principal $L_1$ components of a $D\times N$ data matrix ${\bf X}$.
After we review suboptimal approaches from the recent literature, we generalize the result of the previous section and prove that, if the data dimension $D$ is fixed, then the $K$ principal $L_1$ components of ${\bf X}$ are computable in polynomial time ${\mathcal O}\left(N^{K \text{rank}(\mathbf X)-K+1}\right)$.
\subsection{Exact Exhaustive-search Computation of Multiple $L_1$ Principal Components}
For any $D\times K$ matrix ${\bf A}$,
\begin{equation}
\max_{{\bf R}\in{\mathbbm R}^{D\times K},\,{\bf R}^T{\bf R}={\bf I}_K}\text{tr}\left({\bf R}^T{\bf A}\right)=\left\|{\bf A}\right\|_*
\label{eq:RAA}
\end{equation}
where $\left\|{\bf A}\right\|_*$ denotes the nuclear norm (i.e., the sum of the singular values) of ${\bf A}$.
Maximization in~(\ref{eq:RAA}) is achieved by ${\bf R}={\bf U}{\bf V}^T$ where ${\bf U}{\bf\Sigma}{\bf V}^T$ is the ``compact'' SVD of ${\bf A}$, ${\bf U}$ and ${\bf V}$ are $D\times d$ and $K\times d$, respectively, matrices with ${\bf U}^T{\bf U}={\bf V}^T{\bf V}={\bf I}_d$, ${\bf\Sigma}$ is a nonsingular diagonal $d\times d$ matrix, and $d$ is the rank of ${\bf A}$.
This is due to the trace version of the Cauchy-Schwarz inequality~\cite{Cauchy} according to which
\begin{align}
\text{tr}\left({\bf R}^T{\bf A}\right)&=\text{tr}\left({\bf R}^T{\bf U}{\bf\Sigma}{\bf V}^T\right)=\text{tr}\left({\bf U}{\bf\Sigma}^\frac{1}{2}\cdot{\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T\right) \nonumber \\
&\leq\left\|{\bf U}{\bf\Sigma}^\frac{1}{2}\right\|_2\left\|{\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T\right\|_2=\left\|{\bf\Sigma}^\frac{1}{2}\right\|_2^2=\text{tr}\left({\bf\Sigma}\right)=\left\|{\bf A}\right\|_*
\label{eq:SA}
\end{align}
with equality if $\left({\bf U}{\bf\Sigma}^\frac{1}{2}\right)^T={\bf\Sigma}^\frac{1}{2}{\bf V}^T{\bf R}^T$ which is satisfied by ${\bf R}={\bf U}{\bf V}^T$.
To identify the optimal $L_1$ subspace for any number of components $K$, we begin by presenting a property of ${\mathcal P}^{L_1}_3$ in the form of Proposition~\ref{prop:nuclear} below. Proposition \ref{prop:nuclear} is a generalization of Proposition~\ref{prop:quad} and interprets ${\mathcal P}^{L_1}_3$ as an equivalent nuclear-norm maximization problem over the binary field.
\begin{myprop}
For any data matrix $\mathbf X \in \mathbb R^{D \times N}$, the solution to
${\mathcal P}^{L_1}_3:\;\;\;\;{\bf R}_{L_1}=\argmax_{{\bf R}\in{\mathbb R}^{D\times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf R}^T{\bf X}\right\|_1$
is given by
\begin{equation}
{\bf R}_{L_1}={\bf U}{\bf V}^T
\end{equation}
where ${\bf U}$ and ${\bf V}$ are the $D\times K$ and $N\times K$ matrices that consist of the $K$ dominant-singular-value left and right, respectively, singular vectors of ${\bf X}{\bf B}_\text{{opt}}$ with
\begin{equation}
{\bf B}_\text{{opt}}=\argmax_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*.
\label{eq:Bopt}
\end{equation}
In addition,
$\left\|{\bf R}_{L_1}^T{\bf X}\right\|_1=\left\|{\bf X}{\bf B}_\text{{opt}}\right\|_*$.
\label{prop:nuclear}
\end{myprop}
\noindent \emph{Proof:} We rewrite the optimization problem in~(\ref{eq:RL1}) as
\begin{align}
\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\left\|{\bf X}^T{\bf R}\right\|_1 &=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\sum_{k=1}^K\left\|{\bf X}^T{\bf r}_k\right\|_1=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\sum_{k=1}^K\max_{{\bf b}_k\in\{\pm1\}^N}{\bf b}_k^T{\bf X}^T{\bf r}_k \nonumber \\
&=\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\max_{{\bf B}\in\{\pm1\}^{N\times K}}\text{tr}\left({\bf B}^T{\bf X}^T{\bf R}\right) \nonumber \\
&=\max_{{\bf B}\in\{\pm1\}^{N\times K}}\max_{\mathbf R \in \mathbb R^{D \times K}, \; {\bf R}^T{\bf R}={\bf I}_K}\text{tr}\left({\bf R}^T{\bf X}{\bf B}\right) = \max_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*.
\label{eq:RXB}
\end{align}
That is, $\left\|{\bf R}_{L_1}^T{\bf X}\right\|_1=\left\|{\bf X}{\bf B}_\text{{opt}}\right\|_*$ where ${\bf B}_\text{opt}=\argmax_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf X}{\bf B}\right\|_*$ and, by~(\ref{eq:RAA}) and~(\ref{eq:SA}), ${\bf R}_{L_1}={\bf U}{\bf V}^T$ where
${\bf U}{\bf\Sigma}{\bf V}^T$ is the ``compact'' SVD of ${\bf X}{\bf B}_\text{opt}$.
\hfill
\IEEEQEDclosed
By Proposition~\ref{prop:nuclear}, to find exactly the optimal $L_1$-norm projection operator ${\bf R}_{L_1}$ we can perform the following steps:
\begin{enumerate}
\item
Solve~(\ref{eq:Bopt}) to obtain ${\bf B}_\text{opt}$.
\item
Perform SVD on ${\bf X}{\bf B}_\text{opt}={\bf U}{\bf\Sigma}{\bf V}^T$.
\item
Return ${\bf R}_{L_1}={\bf U}_{:,1:K}{\bf V}^T$.
\end{enumerate}
Steps $1$ - $3$ offer for the first time a direct approach for the computation of the $K$ jointly-optimal $L_1$ principal components of ${\bf X}$.
Step $1$ can be executed by an exhaustive search among all $2^{NK}$ binary matrices of size $N\times K$ followed by evaluation in the metric of interest in~(\ref{eq:Bopt}).
That is, with computational cost $\mathcal O(2^{NK})$ we identify the $L_1$-optimal $K$ principal components of ${\bf X}$.
\subsection{Existing Approaches in Literature}
For the case $K>1$,~\cite{Kwak2008} proposed to design the first $L_1$ principal component ${\bf r}_{L_1}$ by the coupled iteration~(\ref{eq:Kwak1})-(\ref{eq:Kwak2}) (which does not guarantee optimality) and then project the data onto the subspace that is orthogonal to ${\bf r}_{L_1}$; design the $L_1$ principal component of the projected data by the same coupled iteration; and continue similarly.
To avoid the above suboptimal projection-greedy approach,~\cite{Nie2011} presented an iterative algorithm for the computation of ${\bf R}_{L_1}$ altogether (that is the joint computation of the $K$ principal $L_1$ components). In the language of Proposition \ref{prop:nuclear},
the algorithm can be described as arbitrary
initialization at some ${\bf R}_{L_1}^{(0)}$ followed by updates
\begin{align}
{\bf B}^{(i+1)}&=\text{sgn}\left({\bf X}^T{\bf R}_{L_1}^{(i)}\right),\label{eq:UTA1}\\
\left({\bf U}^{(i+1)},{\bf\Sigma}^{(i+1)},{\bf V}^{(i+1)}\right)&=\text{SVD}\left({\bf X}{\bf B}^{(i+1)}\right),\\
{\bf R}_{L_1}^{(i+1)}&={\bf U}_{:,1:K}^{(i+1)}{{\bf V}^{(i+1)}}^T\label{eq:UTA3},
\end{align}
for $i=0, 1,2,\ldots$, until convergence.
Similar to the work in~\cite{Kwak2008}, the above iteration does not guarantee convergence to the $L_1$-optimal subspace.
\subsection{Exact Computation of Multiple $L_1$ Principal Components in Polynomial Time}
By the proof of Proposition~\ref{prop:nuclear}, for any given ${\bf R}\in{\mathbb R}^{D\times K}$ the corresponding metric-maximizing binary matrix is ${\bf B}=\text{sgn}\left({\bf X}^T{\bf R}\right)$. Hence,
\begin{equation}
{\bf B}_\text{opt}=\text{sgn}\left({\bf X}^T{\bf R}_{L_1}\right).
\label{eq:BRL1}
\end{equation}
By Proposition~\ref{prop:nuclear} and~(\ref{eq:BRL1}), computation of the $K$ principal $L_1$ components of ${\bf X}_{D \times N}$ is equivalent to computation of ${\bf B}_\text{opt}$ in~(\ref{eq:Bopt}), which indicates NP-hardness in $N,D$ (that is, when $N,D$ are arbitrarily large). As before,
in this section we consider the case of engineering interest of fixed data dimension $D$.
As in Section~\ref{sec:OneComponent}, we show that, if $D$ is fixed, then computation of the $K$ principal $L_1$ components of ${\bf X}$ is no longer NP-hard (in $N$).
We state our result in the form of the following proposition.
\begin{myprop}
For any fixed data dimension $D$, optimal computation of the $K$ principal $L_1$ components of ${\bf X}\in{\mathbb R}^{D\times N}$ can be carried out with complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})K-K+1}\right)$, $\text{{rank}}({\bf X})\leq D$.
\hfill
\IEEEQEDclosed
\label{prop:polynomial_multiple}
\end{myprop}
To prove Proposition~\ref{prop:polynomial_multiple}, it suffices to prove that ${\bf B}_\text{opt}$ can be computed with complexity ${\mathcal O}\left(N^{\text{{rank}}({\bf X})K-K+1}\right)$.
As in~(\ref{eq:d}), (\ref{eq:QQ}), let $d$ denote the rank of ${\bf X}$ and ${\bf Q}{\bf Q}^T$ where ${\bf Q}\in{\mathbbm R}^{N\times d}$ is the eigen-decomposition matrix of ${\bf X}^T{\bf X}$.
By~(\ref{eq:Bopt}),
\begin{equation}
{\bf B}_\text{{opt}}=\argmax_{{\bf B}\in\{\pm1\}^{N\times K}}\sum_{k=1}^K\sqrt{\lambda_k\left[{\bf B}^T{\bf X}^T{\bf X}{\bf B}\right]}=\argmax_{{\bf B}\in\{\pm1\}^{N\times K}}\sum_{k=1}^K\sqrt{\lambda_k\left[{\bf B}^T{\bf Q}{\bf Q}^T{\bf B}\right]}=\argmax_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf Q}^T{\bf B}\right\|_*
\label{eq:QB}
\end{equation}
where $\lambda_k [{\bf A}]$ denotes the $k$th eigenvalue of matrix ${\bf A}$, $k=1, \ldots, K$.
For the case $N<D$, the optimal binary matirx ${\bf B}_\text{opt}$ can be obtained directly from~(\ref{eq:Bopt}) by an exhaustive search among all $2^{NK}$ binary matrices ${\bf B}\in\{\pm1\}^{N\times K}$.
Therefore, we can design the $L_1$-optimal $K$ principal components with computational cost $2^{NK}<2^{DK}={\mathcal O}(1)$.
For the (certainly more interesting) case where the sample size exceeds the data dimension, $N\geq D$, we present for the first time a generalized version of the approach in~\cite{KP},~\cite{KL} that introduces an orthonormal scanning matrix to maximize a rank-deficient nuclear norm.
In particular, we observe by~(\ref{eq:QB}) that we need ${\bf B}_\text{{opt}}$ that solves
\begin{equation}
\max_{{\bf B}\in\{\pm1\}^{N\times K}}\left\|{\bf Q}^T{\bf B}\right\|_*\overset{(\ref{eq:RAA})}{=} \max_{{\bf B}\in\{\pm1\}^{N\times K}}\max_{{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K}\text{tr}\left({\bf C}^T{\bf Q}^T{\bf B}\right)=\max_{{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K}\max_{{\bf B}\in\{\pm1\}^{N\times K}}\text{tr}\left({\bf B}^T{\bf Q}{\bf C}\right).
\label{eq:BQC}
\end{equation}
By interchanging the maximizations in~(\ref{eq:BQC}), for any fixed $d\times K$ matrix ${\bf C}$ the inner maximization with respect to $\mathbf B \in \{ \pm 1 \}^{N \times K}$ is solved by
\begin{equation}
{\bf B}({\bf C}) = \left[\text{sgn}({\bf Q}{\bf C}_{:,1}), \;\text{sgn}({\bf Q}{\bf C}_{:,2}), \;\ldots, \;\text{sgn}({\bf Q}{\bf C}_{:,K})\right],
\label{eq:BC}
\end{equation}
which is obtained with complexity linear in $N$.
Then, by~(\ref{eq:BQC}), the solution to our original problem in~(\ref{eq:QB}) is met if we collect all possible binary matrices ${\bf B}({\bf C})$ returned as the columns of ${\bf C}$ scan the unit-radius $d$-dimensional hypersphere while maintaining orthogonality among them.
That is, ${\bf B}_\text{opt}$ in~(\ref{eq:QB}) is in%
\footnote{Without loss of optimality, we set $C_{d,k} \geq 0$, $k =1, 2, \ldots, K$, since, for any given $\mathbf C$, $\mathbf C^T \mathbf C = \mathbf I_K$, the binary matrices $\mathbf B(\mathbf C)$ and $\mathbf B\big( \mathbf C \, \text{diag} ( \text{sgn}(\mathbf C_{d, :}) ) \big) $ result to the same metric value in \eqref{eq:Bopt}.}
\begin{equation}
{\mathcal S}_K\stackrel{\triangle}{=}\hspace{-1cm}\bigcup_{\begin{smallmatrix}{\bf C}\in{\mathbbm R}^{d\times K},\,{\bf C}^T{\bf C}={\bf I}_K,\\C_{d,k}\geq0,\,k=1,2,\ldots,K\end{smallmatrix}}\hspace{-1cm}{\bf B}({\bf C}).
\end{equation}
Then, by relaxing orthogonality among the columns of ${\bf C}$,
\begin{equation}
{\mathcal S}_K\subset\hspace{-1cm}\bigcup_{\begin{smallmatrix}{\bf C}\in{\mathbbm R}^{d\times K},\,\left[{\bf C}^T{\bf C}\right]_{k,k}=1,\\C_{d,k}\geq0,\,k=1,2,\ldots,K\end{smallmatrix}}\hspace{-1cm}{\bf B}({\bf C}){=}\Big(\bigcup_{\begin{smallmatrix}{\bf c}\in{\mathbbm R}^d,\,\left\|{\bf c}\right\|_2=1,\\c_d\geq0\end{smallmatrix}}\hspace{-.5cm}{\bf b}({\bf c})\Big)^K{=}{\mathcal S}_1\times{\mathcal S}_1\times\ldots\times{\mathcal S}_1={\mathcal S}_1^K,
\label{eq:S1K}
\end{equation}
which implies that
\begin{equation}
\left|{\mathcal S}_K\right|\leq\left|{\mathcal S}_1\right|^K=\left({\mathcal O}\left(N^{d-1}\right)\right)^K={\mathcal O}\left(N^{dK-K}\right).
\label{eq:SKO}
\end{equation}
From~(\ref{eq:SKO}), we observe that the number of binary matrices that we collect as the columns of ${\bf C}$ scan the unit-radius $d$-dimensional hypersphere --with or without maintaining orthogonality-- is polynomial in $N$.
After ${\bf C}$ has finished scanning the hypersphere, all collected binary matrices in ${\mathcal S}_K$ are compared to each other against the metric of interest in~(\ref{eq:QB}) with complexity ${\mathcal O}(N)$ per matrix.
Therefore, the complexity to solve~(\ref{eq:Bopt}) is determined by the complexity to build ${\mathcal S}_K$ or at most $\mathcal S_1^K$ since $\mathcal S_K \subset \mathcal S_1^K$ by \eqref{eq:S1K}.
Since ${\bf B}_\text{opt}\in{\mathcal S}_K{\subset}{\mathcal S}_1^K$, we already have a direct way to solve~(\ref{eq:QB}).
First, we construct ${\mathcal S}_1$ with complexity ${\mathcal O}\left(N^d\right)$ as described in Section~\ref{sec:OneComponent}.
We note that ${\mathcal S}_1$ contains ${\mathcal O}\left(N^{d-1}\right)$ binary vectors.
Then, we construct ${\mathcal S}_1^K$ which consists of all selections of $K$ elements of ${\mathcal S}_1$ allowing repeated elements.
The order of the elements in each selection can be disregarded, since the order of the columns of ${\bf B}$ does not affect the metric in~(\ref{eq:QB}).
Hence, the total number of selections that we need to consider is the number of possible ways one can choose $K$ elements from a set of $\left|{\mathcal S}_1\right|$ elements disregarding order and allowing repetitions (i.e., the number of size-$K$ multisets of all ${\mathcal S}_1$), which equals~\cite{stanley}
\begin{equation}
P_K=\binom{\left|{\mathcal S}_1\right|+K-1}{K}={\mathcal O}\left(N^{dK-K}\right)
\end{equation}
since $|{\mathcal S}_1|={\mathcal O}\left(N^{d-1}\right)$.
For each one of the $P_K$ binary matrices, we evaluate the corresponding metric $\left\|{\bf Q}^T{\bf B}\right\|_*$ in~(\ref{eq:QB}) with complexity ${\mathcal O}(N)$.
Then, we identify the optimal matrix ${\bf B}_\text{opt}$ by comparing the calculated metric values.
Therefore, the overall complexity to solve~(\ref{eq:Bopt}) is ${\mathcal O}\left(N^{Kd-K}\right)\cdot{\mathcal O}\left(N\right)={\mathcal O}\left(N^{dK-K+1}\right)$.
The complete algorithm for the computation of the optimal $K$-dimensional ($K>1$) $L_1$-principal subspace of a rank-$d$ matrix ${\bf X}\in{\mathbb R}^{D\times N}$ with complexity ${\mathcal O}\left(N^{dK - K +1 }\right)$ is given in Fig.~\ref{fig:algo_multi}.
As a simple illustration of the practical computational cost of the presented algorithm, in Table I we show the average CPU time expended by an Intel$^{{\tiny \textregistered}}$ Core$^{{\tiny \texttrademark}}$ i5 Processor at 3.40 GHz running the algorithm of Fig. \ref{fig:algo_multi} in Matlab$^{{\tiny \textregistered}}$ R2012a to calculate the $K=2$ principal components of a $d \times N$ rank-$d$ data matrix for $d=3,4, 5, 6$ and $N = 4,6, \ldots, 14$ (we consider only the cases $N>d$).
The presented CPU time for each $(d, N)$ case is the average over $100$ data matrix realizations
created with independent zero-mean unit-variance Gaussian drawn entries.
Importantly, per Figs. 2 and 3, both visiting the ${N \choose d-1}$ manifold-intersection points for constructing $\mathcal S_1$ (lines 2-8 of function \emph{compute\_candidates} in Fig. \ref{fig:algo}) and constructing $\mathcal S_K$ given $\mathcal S_1$ (line 4 of the $L_1$-principal subspace algorithm in Fig. \ref{fig:algo_multi}) are fully parallelizable actions that can be distributed over multiple processing units.
Thus, the entire subspace calculation is fully parallelizable and the expended calculation time can be divided down by the number of available processors (plus necessary inter-processor communication overhead).
\section{Experimental Studies}
In this section, we carry out a few experimental studies on $L_1$-subspace signal processing to motivate and illustrate the theoretical developments in the previous sections.
Examples are drawn from the research fields of dimensionality reduction, data restoration, direction-of-arrival estimation, and image conditioning/reconstruction.
\subsection*{Experiment 1 - Data Dimensionality Reduction}
We generate a nominal data set $\mathbf X_{D \times N}$ of $N=50$ two-dimensional ($D=2$) observation points drawn from the Gaussian distribution $\mathcal N \left( \mathbf 0_2,
\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}
\right)$ as seen in Fig. \ref{fig:dr1}.
We calculate and plot in Fig. \ref{fig:dr1} the $L_2$ (by standard SVD) and $L_1$ (by Section III.C, Case $d=2$, complexity about $50 \log_2 50$) principal component of the data matrix $\mathbf X$.%
\footnote{We note that without the presented algorithm, computation of the $L_1$ principal component of $\mathbf X_{2 \times 50}$ would have required complexity proportional to $2^{50}$ (by \eqref{eq:rank2}), which is of course infeasible.}
For reference purposes, we also plot the true nominal data maximum-variance direction, i.e., the
dominant eigenvector of the autocorrelation matrix
$\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}$.
Then, we assume that our data matrix is corrupted by four outlier measurements, $\mathbf o_1, \mathbf o_2, \mathbf o_3, \mathbf o_{4}$, shown in the bottom right corner of Fig. \ref{fig:dr2}.
We recalculate the $L_2$ and $L_1$ principal component of the corrupted data matrix $\mathbf X^{\text{CRPT}} = [\mathbf X, \mathbf o_1, \mathbf o_2 , \mathbf o_3, \mathbf o_{4}]$ and notice (Fig. \ref{fig:dr1} versus Fig. \ref{fig:dr2}) how strongly the $L_2$ component responds to the outliers compared to $L_1$.
To quantify the impact of the outliers, in Fig. \ref{fig:dr3} we generate $1000$ new independent evaluation data points from $\mathcal N \left( \mathbf 0_2,
\begin{bmatrix}
15 & 13 \\ 13 & 26
\end{bmatrix}
\right)$
and estimate the mean square-fit-error $\text{E} \left\{ \|\mathbf x - \mathbf r\mathbf r^T \mathbf x \|_2^2 \right\}$ when $\mathbf r=\mathbf r_{L_2}(\mathbf X^{\text{CRPT}})$ or $\mathbf r_{L_1}(\mathbf X^{\text{CRPT}})$.
We find $\frac{1}{1000} \sum_{i=1}^{1000} \| {\mathbf x}_i - \mathbf r_{L_2}(\mathbf X^{\text{CRPT}})\mathbf r_{L_2}(\mathbf X^{\text{CRPT}})^T {\mathbf x}_i \|_2^2 = 34.417$ versus $\frac{1}{1000} \sum_{i=1}^{1000} \| {\mathbf x}_i - \mathbf r_{L_1}(\mathbf X^{\text{CRPT}})\mathbf r_{L_1}(\mathbf X^{\text{CRPT}})^T {\mathbf x}_i \|_2^2 = 11.555$.
In contrast, when the principal component is calculated from the clean training set, $\mathbf r= \mathbf r_{L_2}(\mathbf X)$ or $ \mathbf r_{L_1}(\mathbf X)$, we find estimated mean square-fit-error $6.077$ and $ 6.080$, correspondingly.
We conclude that dimensionality reduction by $L_1$ principal components may loose only minimally in mean-square fit compared to $L_2$ when the designs are from clean training sets, but can protect significantly when training is carried out in the presence of erroneous data.
Next, we will compare the dimensionality-reduction performance of the proposed $L_1$-principal subspace with that of other subspaces in the literature obtained by means of $L_1$-norm based methods.
Specifically, alongside the $L_2$ (SVD) and $L_1$-principal component (proposed),
we calculate the $R_1$-principal component \cite{Ding2006} as well as the direction obtained by means of $L_1$-factorization through alternating weighted median calculation \cite{Ke2003}, \cite{Ke2005}.\footnote{Notice that for $R_1$-PCA \cite{Ding2006} and $L_1$-factorization \cite{Ke2003}, \cite{Ke2005}, no optimal solution exists in the literature so far.}
All directions are calculated from an $(N=20)$-point corrupted data set $\mathbf X^{\text{CRPT}} \in \mathbb R^{2 \times 20}$ with $N_{\mathrm{out}}$ outliers drawn from $\mathcal N \left( \begin{bmatrix} 20 \\ -20 \end{bmatrix}, \begin{bmatrix} 5.73 & -4.494 \\ -4.494 & 5.27 \end{bmatrix} \right)$ and $N - N_{\mathrm{out}}$ nominal points drawn from
$\mathcal N \left( \mathbf 0_{3}, \begin{bmatrix} 15 & 13 \\ 13 & 26 \end{bmatrix} \right)$.
In Fig. \ref{fig:linefitting_mse}, we plot the mean-squared-fit-error
averaged over $10 000$ independent corrupted training data-set experiments
as a function of the number of outlying points in the data set $N_{\mathrm{out}}$. We notice that, when designed on nominal data, all examined subspaces differ little, if any, from the $L_2$-principal subspace in mean-square fit error.
However, when designed on outlier-corrupted data sets, the $L_1$-principal subspace exhibits
notable robustness outperforming uniformly and significantly all other subspaces, especially
in the $15\%$ - $40\%$ mid-range of corruption. Given that $L_1$ and $L_2$ start very near
each other
in mean-square-fit-error at $0\%$ corruption and meet again only at $100\%$ corruption,
one is tempted
to say that the $L_1$ subspaces are to be uniformly preferred over $L_2$ if the associated
computational cost can be afforded.
\renewcommand{\arraystretch}{0.6}
\subsection*{Experiment 2 - Data Restoration}
As a toy numerical example, consider a hypothetical case where we collect from a sensor system
eight samples of five-dimensional data. Due to the nature of the sensed source, the data are to
lie in a lower-than-five dimensional space, say a plane. Say, then, the true data are given by the
$\text{rank}$-$2$ data matrix below
{\footnotesize
\begin{align}
\mathbf X_{5 \times 8} =
\begin{bmatrix}
2.0724 & -1.2024 & 1.2956 & 2.8719 & 1.5637 & -2.9323 & -3.1792 & -1.4152 \\
-0.5233 & 0.2595 & -0.3298 & -0.7562 & -0.4087 & 0.7973 & 0.8235 & 0.4155 \\
0.0185 & -0.8158 & -0.0367 & -0.5406 & -0.2380 & 1.0108 & 0.3502 & 1.0487 \\
-0.6424 & 0.1476 & -0.4151 & -1.0486 & -0.5552 & 1.1989 & 1.0913 & 0.7355 \\
-2.1289 & 2.2734 & -1.2687 & -2.2200 & -1.2814 & 1.6751 & 2.7777 & 0.0851 \\
\end{bmatrix}. \nonumber
\end{align}
}
Assume that due to sensor malfunction or data transfer error or data storage failure,
we are presented instead with
{\footnotesize
\begin{align}
\mathbf X_{5 \times 8}^{\text{CRPT}} =
\begin{bmatrix}
2.0724 & {\color{red}\it{ 8.9538}} & 1.2956 & 2.8719 & {\color{red}\it{ 10.6817}} & -2.9323 & -3.1792 & -1.4152 \\
-0.5233 & {\color{red}\it{ 10.6187}} & -0.3298 & -0.7562 & {\color{red}\it{ 11.0235}} & 0.7973 & 0.8235 & 0.4155 \\
0.0185 & {\color{red}\it{ 11.3050}} & -0.0367 & -0.5406 & -0.2380 & 1.0108 & 0.3502 & 1.0487 \\
-0.6424 & 0.1476 & -0.4151 & -1.0486 & {\color{red}\it{ 7.8846}} & 1.1989 & 1.0913 & 0.7355 \\
-2.1289 & 2.2734 & -1.2687 & -2.2200 & -1.2814 & 1.6751 & 2.7777 & 0.0851 \\
\end{bmatrix}\nonumber
\end{align}
}\\
where six of the original entries in two of the data points have been
altered/overwritten and $\mathbf X^{\text{CRPT}} $ spans now a four-dimensional subspace of $\mathbb R^{5}$.
Our objective is to ``restore'' $\mathbf X^{\text{CRPT}} $ to $\mathbf X$ taking advantage of
our knowledge (or assumption) of the rank of the original data. Along these lines,
we project $\mathbf X^{\text{CRPT}}$ onto the span of its $K=2$ $L_2$- or $L_1$-principal components,
\begin{align}
\hat{\mathbf X}=\mathbf R \mathbf R^T \mathbf X^{\text {CRPT}}
\end{align}
where $\mathbf R_{5 \times 2}= [\mathbf r_{L_2}^{(1)}, \mathbf r_{L_2}^{(2)}]$ or $ [\mathbf r_{L_1}^{(1)}, \mathbf r_{L_1}^{(2)}]$.
The resulting $L_2$- and $L_1$-derived representations of $\mathbf X$ are
{\footnotesize
\begin{align}
\hat{\mathbf X}_{L_2}=
\begin{bmatrix}
0.8029 & 8.2311 & 0.4919 & 0.9945 & 11.8445 & -0.9197 & -1.1528 & -0.3268\\
0.4839 & 11.0891 & 0.2888 & 0.5096 & 10.2500 & -0.3897 & -0.6347 & -0.0285\\
-0.5922 & 11.1679 & -0.3843 & -0.9862 & 0.0165 & 1.1412 & 1.0192 & 0.7148\\
0.6521 & 0.8969 & 0.4067 & 0.8926 & 6.6810 & -0.9024 & -0.9930 & -0.4245\\
-0.3868 & 2.8347 & -0.2455 & -0.5789 & -2.2540 & 0.6257 & 0.6220 & 0.3444
\end{bmatrix} \nonumber
\end{align}
}
and
{\footnotesize
\begin{align}
\hat{\mathbf X}_{L_1}=
\begin{bmatrix}
2.0724 & -0.0303 & 1.2956 & 2.8719 & 2.9321 & -2.9323 & -3.1792 & -1.4152\\
-0.5233 & 0.1880 & -0.3298 & -0.7562 & -0.7283 & 0.7973 & 0.8235 & 0.4155\\
0.0185 & 3.2915 & -0.0367 & -0.5406 & 0.2476 & 1.0108 & 0.3502 & 1.0487\\
-0.6424 & 0.9300 & -0.4151 & -1.0486 & -0.8469 & 1.1989 & 1.0913 & 0.7355\\
-2.1289 & -4.2139 & -1.2687 & -2.2200 & -3.2976 & 1.6751 & 2.7777 & 0.0851
\end{bmatrix}, \nonumber
\end{align}
}
\noindent{respectively}.
In Fig.~\ref{fig:error}, we plot the element-by-element and per-measurement square-restoration error for the two projections.
The relative superiority of $L_1$-subspace data representation is clearly captured and documented.
\subsection*{Experiment 3 - Direction-of-Arrival Estimation}
We consider a uniform linear antenna array of $D=5$ elements that takes $N=10$ snapshots of two incoming signals with angles of arrival $\theta_1 = -30^\circ$ and $\theta_2 = 50^\circ$,
\begin{align}
\mathbf x_n= A_1 \mathbf s_{\theta_1} + A_2 \mathbf s_{\theta_2} + \mathbf n_n, ~n=1, \ldots, 10, \label{cleanmes}
\end{align}
where $A_1, A_2$ are the received-signal amplitudes with array response vectors $\mathbf s_{\theta_1}$ and $\mathbf s_{\theta_2}$, correspondingly, and
$\mathbf n \sim \mathcal {CN}\left(\mathbf 0_{5}, \sigma^2 \mathbf I_{5} \right)$ is additive white complex Gaussian noise.
We assume that the signal-to-noise ratio (SNR) of the two signals is $\text{SNR}_1=10 \log_{10}\frac{A_1^2}{\sigma^2}\text{dB}=2\text{dB}$ and $\text{SNR}_2=10 \log_{10}\frac{A_2^2}{\sigma^2}\text{dB}=3\text{dB}$.
Next, we assume that one arbitrarily selected measurement out of the ten observations $\mathbf X_{5 \times 10}=[\mathbf x_1, \ldots, \mathbf x_{10}] \in \mathbb C^{5 \times 10}$ is corrupted by a jammer operating at angle $\theta_J= 20^\circ$ with amplitude $A_J = A_2$.
We call the resulting corrupted observation set $\mathbf X^{\text{CRPT}} \in \mathbb C^{5 \times 10}$ and create the real-valued version $\tilde{\mathbf X}^{\text{CRPT}} = [ \text{Re} \{\mathbf X^{\text{CRPT}}\}^T, ~ \text{Im} \{\mathbf X^{\text{CRPT}}\}^T ]^T \in \mathbb R^{10 \times 10}$ by $\text{Re}\{ \cdot\}, \text{Im}\{ \cdot\} $ part concatenation.
We calculate the $K=2$ $L_2$-principal components of $\tilde{\mathbf X}^{\text{CRPT}}$, $\mathbf R_{L_2} = [\mathbf r_{L_2}^{(1)}, \mathbf r_{L_2}^{(2)}] \in \mathbb R^{10 \times 2}$, and the $K=2$ $L_1$-principal components of $\tilde{\mathbf X}^{\text{CRPT}}$, $\mathbf R_{L_1} = [\mathbf r_{L_1}^{(1)}, \mathbf r_{L_1}^{(2)}] \in \mathbb R^{10 \times 2}$.
In Fig. \ref{fig:DOA}, we plot the standard $L_2$ MUSIC spectrum \cite{schmidt}
\begin{align}
P(\theta) \overset{\triangle}{=} \frac{1}{ \tilde{\mathbf s}_{\theta}^T (\mathbf I_{2D}- \mathbf R_{L_2} \mathbf R_{L_2}^T) \tilde{\mathbf s}_{\theta}}, ~\theta \in \left( -\frac{\pi}{2}, \frac{\pi}{2}\right),
\end{align}
where $ \tilde{\mathbf s}_{\theta} = [ \text{Re}\{{\mathbf s}_{\theta}\}^T, ~ \text{Im}\{{\mathbf s}_{\theta}\}^T ]^T$, as well as what we may call ``$L_1$ MUSIC spectrum'' with $\mathbf R_{L_1}$ in place of $\mathbf R_{L_2}$.
It is interesting to observe how $L_1$ MUSIC (in contrast to $L_2$ MUSIC) does not respond to the one-out-of-ten outlying jammer value in the data set and shows only the directions of the two actual nominal signals.
\subsection*{Experiment 4 - Image Reconstruction}
Consider the ``clean'' $ 100 \times 64 $ gray-scale image $\mathbf A \in \{0, 1, \ldots, 255\}^{100 \times 64}$ of Fig.~\ref{fig:alexis1}.
We assume that $\mathbf A$ is not available and instead we have a data set of $N=10$ corrupted/occluded versions of $\mathbf A$, say $\mathbf A_{1}, \mathbf A_{2}, \ldots, \mathbf A_{10}$.
Each corrupted instance $\mathbf A_i$, $i=1, \ldots, 10$, is created by partitioning the original
image $\mathbf A$ into sixteen tiles of size $25 \times 16$ and replacing three arbitrarily selected tiles by $25 \times 16$ grayscale-noise patches as seen, for example, in Fig.~\ref{fig:alexis2}.
The $10$ corrupted instances are vectorized to form the data matrix
\begin{align}
\mathbf M= [ \text{vec}(\mathbf A_1), \ldots, \text{vec}(\mathbf A_{10}) ] \in \{0, \ldots, 255\}^{6400 \times 10}.
\end{align}
Next, we ``condense'' $\mathbf M$ to a rank-$2$ representation by both $L_2$- and $L_1$-subspace projection,
\begin{align}
\hat{\mathbf M}_{L_{2/1}}= \mathbf R_{L_{2/1}} \mathbf R_{L_{2/1}}^T \mathbf M,
\end{align}
where $\mathbf R_{L_{2/1}} \in \mathbb R^{6400 \times 2}$ consists of the $K=2$ $L_2$ or $L_1$, accordingly, principal components of $\mathbf M$.
In Fig.~\ref{fig:alexis3} we show the projection of the corrupted image of Fig.~\ref{fig:alexis2} onto the $L_2$-derived rank-$2$ subspace (maximum-$L_2$-projection reconstruction).
In Fig.~\ref{fig:alexis4}, we show the projection of the same image onto the $L_1$-derived rank-$2$ subspace (maximum-$L_1$-projection reconstruction).
Figs. \ref{fig:alexis3} and (d) offer a perceptual (visual) interpretation of the difference between $L_2$ and $L_1$-subspace rank reduction. It is apparent that maximum-$L_1$-projection reconstruction offers a much clearer image representation of $\mathbf A$ than maximum-$L_2$-projection reconstruction. This is another result that highlights the resistance of $L_1$-principal subspaces against outlying data corruption.
\section{Conclusions}
We presented for the first time in the literature optimal (exact) algorithms for the calculation of maximum-$L_1$-projection subspaces of data sets with complexity polynomial in the sample size (and exponent equal to the data dimension).
It may be possible in the future to develop an $L_1$ principal-component-analysis (PCA) line of research that parallels the enormously rewarding $L_2$ PCA/feature-extraction developments.
When $L_1$ subspaces are calculated on nominal ``clean" training data, they differ little --arguably-- from their $L_2$-subspace counterparts in least-squares fit.
When, however, subspaces are calculated from data sets with possible erroneous, out-of-line, ``outlier" entries, then $L_1$ subspace calculation offers significant robustness/resistance to the presence of inappropriate data values.\\
\mbox{}\\
\begin{center}
\bf{ACKNOWLEDGEMENT}\\
\end{center}
The authors would like to thank the Associate Editor and the four %
anonymous reviewers for their comments and suggestions that helped %
improve this manuscript significantly, both in presentation and content.
|
\subsubsection{Abstract}
One reason why classical phylogenetic reconstruction methods fail to correctly infer the underlying topology is because they assume oversimplified models. In this paper we propose a topology reconstruction method consistent with the most general Markov model of nucleotide substitution, which can also deal with data coming from mixtures on the same topology. It is based on an idea of Eriksson on using phylogenetic invariants and provides a system of weights that can be used as input of quartet-based methods.
We study its performance on real data and on a wide range of simulated 4-taxon data (both time-homogeneous and nonhomogeneous, with or without among-site rate heterogeneity, and with different branch length settings). We compare it to the classical methods of neighbor-joining (with paralinear distance), maximum likelihood (with different underlying models), and maximum parsimony. Our results show that this method is accurate and robust, has a similar performance to ML when data satisfies the assumptions of both methods, and outperforms all methods when these are based on inappropriate substitution models or when both long and short branches are present. If alignments are long enough, then it also outperforms other methods when some of its assumptions are violated.\\
\noindent[\textbf{Keywords}: phylogenetic invariants, topology reconstruction, general Markov model, heterogeneity across lineages, heterogeneity across sites, yeast]\\
\section{Introduction}
Classical methods of phylogenetic tree topology reconstruction are known to have limitations. For example, maximum likelihood (\texttt{ML}\xspace) is known to fail when data violates some of the underlying model assumptions \citep{Swofford2001, Kuck2012, Ho2004}; maximum parsimony (\mp) is statistically inconsistent in the Felsenstein zone \citep{felsenstein1978}; and neighbor-joining (\texttt{NJ}\xspace) is subject to the choice of an unbiased distance and it is not as accurate as \texttt{ML}\xspace when both methods can be applied \citep{Tateno1994}. When trying to estimate distant phylogenies, neglecting
heterogeneity in the substitution process across lineages (HAL from now on, as denoted in \citet{jayaswal2014})
or heterogeneity across sites (HAS) may result in inaccurate phylogenetic estimates
(see \citet{Yang1995,Ho2004,foster2004, galtier1998, felsenstein1978, yang1994, fitch1986, Stefankovic2007, Kolaczkowski2004} among others).
\textit{Phylogenetic invariants} were first introduced by \citet{cavender1987} and \citet{lake1987} as a non-parametric method of phylogenetic reconstruction: they are equations satisfied by any possible joint distribution of character patterns at the leaves of a tree evolving under an evolutionary Markov model. The potential of phylogenetic invariants was the ability of dealing with more general models and of detecting the topology without estimating branch lengths or substitution parameters \citep[see][chapter 22]{Felsenstein2004}. In particular, they can handle HAL
better than other methods \citep{casanellas2007,holland2013} and (some) could deal with
HAS, as Lake's invariants did \citep{lake1987}. Nevertheless, only a few phylogenetic invariants
were known by that time, it was not clear how to use them \citep{Felsenstein2004}, they seemed useless for large trees, and the approach was laid aside by the bad results obtained in simulations \citep{huelsenbeck1995}.
\citet{eriksson2005} proposed a new topology reconstruction method, \texttt{ErikSVD}\xspace, based on the work on invariants of
\citet{AllmanRhodeschapter4}. The underlying idea is that organizing the joint distribution of character patterns
according to a bipartition of the set of taxa gives a \textit{bipartition matrix} (\ref{flattening}) of rank $\leq 4$ if the bipartition is induced by an edge of the tree (and otherwise, the rank is higher). This result holds for any set of DNA
sequences evolving under the most general Markov model (GMM), also known as Barry Hartigan's model \citep{barryhartigan87trans, allman2008, jayaswal2005}. This is the most general HAL model as
it allows different instantaneous
rate matrices and heterogeneous composition at different parts of the tree, even locally along each branch \citep{Jayaswal2011}.
\texttt{ErikSVD}\xspace does not use phylogenetic invariants directly but computes the Frobenius distance of the bipartition matrices to the set of matrices of rank $\leq 4$.
Although it is nowadays clear that phylogenetic invariants derived from rank conditions on matrices
are the only relevant invariants for reconstructing the topology \citep{casfer2010},
the original method \texttt{ErikSVD}\xspace turned out not to be
accurate enough to compete against standard methods \citep{eriksson2005}, especially in the presence of long branches and short alignments.
Here we revisit \texttt{ErikSVD}\xspace by correcting the target matrix: in the method \texttt{Erik+2}\xspace proposed here we consider the two possible transition matrices from the
states of one side of the bipartition to the other (that is, we normalize by column and row sum the
bipartition matrix of \texttt{ErikSVD}\xspace). This correction is made to take into account that the rank of the bipartition matrix obtained from empirical distributions could be affected by the presence of long-branch attraction situations
(see Appendix 1). The original
\texttt{ErikSVD}\xspace was already \textit{statistically consistent} (that is, as the empirical distribution approaches the theoretical distribution, the probability of correctly reconstructing the tree goes to one) and so is the new method (see Materials and Methods).
\texttt{Erik+2}\xspace is model-based as it assumes a general Markov model of evolution (and it could also be redesigned to incorporate more restrictive Markov models or even aminoacid substitution models),
but is non-parametric in the sense that it does not attempt to recover the parameters of the model. Moreover, the theoretical background of \texttt{Erik+2}\xspace permits to apply it on HAS data evolved on the same tree topology under GMM \citep{jayaswal2014}: that is, a parameter $m$ can be introduced so that \texttt{Erik+2}\xspace considers the sites of the alignment to be divided into $m$ categories, each evolving on the same topology but with (possibly) different Markov substitution matrices --this is called an \textit{m-mixture} \citep{Stefankovic2007}. For example, discrete-gamma rates or the heterogeneous tree in \citep{Kolaczkowski2004} are instances of mixtures, and \texttt{ML}\xspace is known to fail under these conditions even when consistent underlying homogeneous models are considered \citep{Kuck2012,Kolaczkowski2004}. For $m$-mixtures, the rank of the bipartition matrix induced by an edge is not larger than $4m$ \citep[e.g.][]{rhodessullivant} so that in this case we use the distance to matrices of rank $\leq 4m$.
We develop \texttt{Erik+2}\xspace on 4-taxon trees and study its performance on simulated and real data. Using computer simulations we compare it to the classical methods \texttt{ML}\xspace, \texttt{NJ}\xspace, \mp and to the original \texttt{ErikSVD}\xspace in many different scenarios. We chose quartets because
they are the smallest building blocks of phylogenetic reconstruction \citep{Ranwez2001} and they
are widely used as a hint of efficiency and robustness of the method under study \citep{huelsenbeck1995}. \texttt{Erik+2}\xspace evaluates the three possible quartet topologies and returns a system of weights that can be used as input for quartet-based methods (see the Methods section).
Some of our computer simulations are generated under the general Markov process that underlies \texttt{Erik+2}\xspace
and some are based on themost general time-reversible (GTR) and homogeneous across lineages model (homGTR from now on). We also simulate HAS data by generating either 2-mixtures on the same topology evolving under GMM or Gamma continuously-distributed rates across sites under the homGTR model.
Throughout the paper \texttt{NJ}\xspace has been considered with the paralinear distance,
and \texttt{ML}\xspace computations have been based on continuous-time models (with parameters to be estimated by the method)
considering homogeneity or heterogeneity across lineages and sites depending on the situation (we detail it explicitly in figure captions).
The performance of \texttt{Erik+2}\xspace on real data is analyzed on the eight species of yeast studied in \citet{Rokas2003} with the concatenated alignment provided by \citet{jayaswal2014}. We investigate whether the quartets output by \texttt{Erik+2}\xspace and \texttt{ErikSVD}\xspace support the tree $T$ of \citet{Rokas2003} or the alternative tree $T'$ of \citet{Phillips2004}, and the mixture model proposed by \citet{jayaswal2014}.
\FloatBarrier
\section{Results}
We present the performance of the new method \texttt{Erik+2}\xspace, the original method of Eriksson (\texttt{ErikSVD}\xspace), and the classical
methods \texttt{ML}\xspace, \texttt{NJ}\xspace and \mp, on quartet reconstruction on different simulated data.
\texttt{Erik+2}\xspace is publicly available at the webpage \newline
http://www.pagines.ma1.upc.edu/$\sim$casanellas/Erik+2.html.
\subsection{Homogeneity across sites}
First of all we consider a tree subject to long branch attraction. On the tree of Figure \ref{tree}.a
we fix $a= 0.05$, $b=0.75,$
and let the internal branch length $c$ vary in the range $[0.01,0.4]$ so that the tree lies in the Felsenstein zone.
Alignments of lengths 1~000, 10~000 and 100~000 base pairs (\emph{bp}.) were generated under GMM according to this tree.
The results obtained for \texttt{Erik+2}\xspace, \texttt{ML}\xspace , \mp and \texttt{ErikSVD}\xspace on these data are shown in Figure \ref{single}. In this figure
two models underlying \texttt{ML}\xspace computations have been considered:
the most general homogeneous continuous-time model, \texttt{ML}(hom) from now on, and a HAL GTR model, \texttt{ML}(HALGTR) henceforth. \texttt{ML}\xspace has a similar performance with both models.
We observe that, \texttt{Erik+2}\xspace is more accurate than \texttt{ErikSVD}\xspace in general and
especially when the
interior branch length is short
(only for length 1~000 and $c\in(0.13,0.25)$ \texttt{ErikSVD}\xspace outperforms slightly \texttt{Erik+2}\xspace). For 1~000 sites, both versions of \texttt{ML}\xspace perform better than \texttt{Erik+2}\xspace, but when more data is available
\texttt{Erik+2}\xspace outperforms \texttt{ML}\xspace (in its both versions). Notice incidentally that the accuracy of \texttt{ML}\xspace or \mp does not seem to increase as the length of the alignment grows, while it certainly does for \texttt{ErikSVD}\xspace and \texttt{Erik+2}\xspace.
In a more complete study, we adopted a similar approach to \citet{huelsenbeck1995} to test
different methods. More precisely, we evaluate the methods on a \textit{tree space} (see Figure \ref{tree}.b) where the quartets are as in Figure \ref{tree}.a with $c=a$, and the branch lengths $a$ and $b$ vary between 0 and 1.5 in steps of 0.02.
For each pair $a,b$ we generated 100 alignments of a fixed length and represented in a gray scale the success of different methods in recovering the right topology (black means 100 \% of success, and white 0 \%).
The methods \texttt{Erik+2}\xspace, \texttt{ML}\xspace, and \texttt{NJ}\xspace with paralinear distance have been tested according to this approach.
The results for 1~000 and 10~000 bp. are shown in Figure \ref{treespaceGMM} for data generated under GMM and \texttt{ML}\xspace estimating the most general homogeneous continuous-time model \texttt{ML}(hom), and in Figure \ref{treespaceGTR} for data generated under homogeneous GTR model and \texttt{ML}\xspace estimating exactly the same model, \texttt{ML}(homGTR).
In Figure \ref{treespaceGMM}, we see that both \texttt{ML}(hom) and
\texttt{NJ}\xspace have lower accuracy than \texttt{Erik+2}\xspace, as it was expected under data that violates the assumptions of \texttt{ML}\xspace and \texttt{NJ}\xspace.
In both figures \ref{treespaceGMM} and \ref{treespaceGTR} we observe that, while \texttt{Erik+2}\xspace and \texttt{NJ}\xspace drastically increase their accuracy when alignment length is multiplied by
10, \texttt{ML}\xspace does not significantly improve with alignment length (especially in Figure \ref{treespaceGMM} when the substitution model assumed by \texttt{ML}\xspace is incorrect).
In Figure S1 of the Appendix 2 the reader can find the performance of \texttt{ErikSVD}\xspace on the tree space of Figure \ref{tree}.b, confirming the improvement of \texttt{Erik+2}\xspace over the original method. The average success and standard deviation achieved by these methods on this tree space are shown in Table \ref{tab:mean_sd} (where alignment length 500 bp. is also included).
It is worth pointing out that, for alignments of 1~000 bp. evolving under homogeneous GTR model, \texttt{ML}(homGTR) seems to perform better than \texttt{Erik+2}\xspace in the Felsenstein zone. However, for length 10~000, \texttt{Erik+2}\xspace already outperforms \texttt{ML}(homGTR) (Fig. \ref{treespaceGTR}). Moreover, the global accuracy of \texttt{ML}(hom) drastically drops when applied to data obtained under GMM (see Fig. \ref{treespaceGMM}).
Notice also that whereas the accuracy of \texttt{NJ}\xspace and \texttt{ML}\xspace
drops when all branches are long (top right corner), the performance of \texttt{Erik+2}\xspace seems less sensitive to long branches.
We have also evaluated the version of \texttt{Erik+2}\xspace with 2-mixtures ($m=2$) on the same data (see Fig. S3 in Appendix 2). The accuracy obtained for alignments of 1~000 bp. is similar to that of \texttt{Erik+2}\xspace with $m=1$ (the means are 0.790 and 0.803, respectively), and hence the choice $m=2$ appears as a good option when alignments are long enough and we ignore whether the data comes from mixtures or not (see also Fig.
S4 in Appendix~2).
\subsection{Heterogeneity across sites (HAS)}
On the same tree of Figure \ref{tree}.a,
we
generated data under homogeneous GTR model with sites varying according to a Gamma distribution with parameter $\alpha=\beta$ in the range $(0,2]$ varying in steps of $0.1$. Small values of this parameter indicate a lot of variation
across sites \citep{yang1993}. While this setting violates the hypotheses of the model underlying \texttt{Erik+2}\xspace and \texttt{ErikSVD}\xspace, in this case maximum likelihood is estimating a homogeneous GTR model with rates varying according to the auto-discrete Gamma model \texttt{ML}(homGTR+$\Gamma$) \citep{yang1994}.
The results appear in Figure \ref{2partitions}.a, where we observe that \texttt{Erik+2}\xspace manages to overcome the violation of its hypotheses giving 100\% success
already for 10~000 bp., while \texttt{ErikSVD}\xspace gives notably worse results. \texttt{ML}\xspace is more successful than \texttt{Erik+2}\xspace for 1~000 bp., but both methods have a similar performance on longer alignments.
On the same data we also tested \mp, obtaining in all cases the incorrect tree $13|24$ (and therefore we do not represent the corresponding 0\% line in the figure).
As mentioned above, one of the main features of \texttt{Erik+2}\xspace is that it can deal with different categories of
evolutionary rates. In order to test its accuracy on such setting, we used the approach of \citet{Kolaczkowski2004}. We considered two categories of the
same size both evolving under GMM on the tree of Figure \ref{tree}.a: the first category corresponds to branch lengths $a=0.05$, $b=0.75$, while the second corresponds to $a=0.75$ and $b=0.05$. The internal branch length was set to the
same value in both categories and varied from 0.01 to 0.4.
In Figure \ref{2partitions}.b) we present the performance of \texttt{Erik+2}\xspace (with $m=1$ and $m=2$), \mp, \texttt{ML}(hom), and \texttt{ML}\xspace estimating a HAL GTR model with discrete Gamma rates with 2 categories, \texttt{ML}(HALGTR+2$\Gamma$) henceforth.
We included \mp in this study because, as stated in \cite{Kolaczkowski2004}, it performs better than \texttt{ML}\xspace estimating a single category model. This claim is confirmed by the results in our simulations with both versions of \texttt{ML}\xspace.
It is worth pointing out that even \texttt{Erik+2}\xspace with $m=1$ performs better than \texttt{ML}(hom) for internal branch length $\leq 0.25$, and than \texttt{ML}(HALGTR+2$\Gamma$) for internal branch length $\leq 0.15$. Also, notice
that for length 10~000 and larger, the accuracy of \texttt{Erik+2}\xspace with $m=2$ is always greater than 33\%, even if the internal branch length is close to zero. This does not happen for \texttt{ML}\xspace, \mp, which are clearly inconsistent in this setting
\subsection{Performance on real data}
We considered the data provided by \citet{jayaswal2014} with 42~337 second codon positions of 106 orthologous genes of \emph{Saccharomyces
cerevisiae}, \emph{S. paradoxus}, \emph{S. mikatae}, \emph{S. kudriavzevii}, \emph{S. castellii}, \emph{S. kluyveri}, \emph{S. bayanus}, and \emph{Candida albicans}. The phylogenetic tree of these species was originally studied in \citet{Rokas2003}, where a tree topology $T$ was identified with 100\% bootstrap support for the concatenated alignment of these genes. This tree is widely accepted by the community but its correct inference is known to depend on the consideration of HAL \citep{Rokas2003, Phillips2004, jayaswal2014}. For example, \citet{Phillips2004} obtain an alternative tree $T'$ with 100\% bootstrap using the method of minimum evolution, but identified the incorrect handling of compositional bias as responsible for this inconsistency. Moreover, according to \citet{jayaswal2014} these data is best modeled by taking into account HAL plus two different rate categories and invariable sites. In our setting, this would involve three mixtures. We apply \texttt{ErikSVD}\xspace and \texttt{Erik+2}\xspace with $m=1,2,3$ to 4-taxon subalignments and investigate the proportion of output quartets that are compatible with $T$ or $T'$. The results displayed in Table \ref{tab:realdata} show that \texttt{Erik+2}\xspace supports the tree $T$ and the model suggested by \citet{jayaswal2014} ($m=3$), whereas \texttt{ErikSVD}\xspace gives more support to the alternative tree $T'$.
\subsection{Time of execution}
We have compared the time of execution of the different reconstruction methods used in our simulations with 100 alignments of length 1~000 bp on a 3.2GHz processor.
The results obtained show that \texttt{NJ}\xspace is the fastest method, 1.324s. \texttt{ErikSVD}\xspace and \texttt{Erik+2}\xspace take 1.928s and 2.148s respectively, and \mp takes 3.984s. Finally, \texttt{ML}\xspace is the slowest method by far because it has to infer the model parameters: using PAML software, \texttt{ML}(hom) and \texttt{ML}(homGTR) need about 10 seconds, and using \texttt{bppml}\xspace of Bio++ package, \texttt{ML}(HALGTR) and \texttt{ML}(HALGTR+2$\Gamma$) need about 200 minutes.
\section{Discussion}
The simulation studies show that \texttt{Erik+2}\xspace is an accurate and robust topology reconstruction method on quartets, especially in situations where other methods systematically fail (compositional heterogeneity and/or rate heterogeneity across lineages, or long branch attraction). In such scenarios, \texttt{Erik+2}\xspace outperforms the method of Eriksson, \texttt{ErikSVD}\xspace, and classical methods like \mp, \texttt{NJ}\xspace and \texttt{ML}\xspace based on models that cannot accommodate these assumptions. \texttt{Erik+2}\xspace is based on the most general Markov model and hence accounts for HAL data, even locally at each edge. When its assumptions are violated, for example in the presence of continuous Gamma-distributed rates among sites, we have shown that it is highly accurate if there is enough data. As observed, \texttt{Erik+2}\xspace can also deal with $m$-mixtures on the same tree topology (although for quartets the limit is $m=3$). Even more, using \texttt{Erik+2}\xspace with $m=2$ is probably the best option for large alignments
when mixed/unmixed nature of data is unknown.
On the experiments we presented, the overall performance of \texttt{ML}\xspace is quite accurate if model assumptions are not violated, confirming the conclusions of \citet{Kolaczkowski2004, Kuck2012}. Also in line with these papers, we corroborate that long sequences do not improve \texttt{ML}\xspace performance on data that do not satisfy the hypothesis of the underlying model. Moreover, \texttt{ML}\xspace is by far the slowest among the methods tested here,
while \texttt{Erik+2}\xspace is slightly twice slower than \texttt{NJ}\xspace. Another drawback of \texttt{ML}\xspace is that, quite often, it does not converge when it is computed on the incorrect topology, which makes the comparison of likelihoods impossible. Whereas the goal of \texttt{Erik+2}\xspace is to reconstruct the topology, \texttt{ML}\xspace is designed to estimate the parameters of the substitution matrices and, it would probably be a good choice to use first \texttt{Erik+2}\xspace and then \texttt{ML}\xspace to estimate the parameters.
In our simulation study, \texttt{NJ}\xspace (with paralinear distance) and \mp have been the methods with least success, which is not so surprising if one takes into account that they are also the less adaptable to general data.
We have only developed \texttt{Erik+2}\xspace for quartets with the aim of validating it as a successful method, and it is still a work in progress to further develop it for larger number of taxa. Using \texttt{Erik+2}\xspace to evaluate the confidence of particular bipartitions of large sets of taxa is already a viable option, and in this case one can deal with a larger number $m$ of categories (the maximum $m$ allowed depends on the size of the subsets $A$, $B$ of taxa involved in the bipartition: $4^{min\{|A|,|B|\}-1}-1$). We have also started testing its weights as input of weighted quartet-based methods with high success (unpublished). In particular, it outperforms global \texttt{NJ}\xspace, which makes \texttt{Erik+2}\xspace a potential input method for quartet-based methods \citep{stjohn2003}.
\newpage
\FloatBarrier
\section{Materials and Methods}
\subsection{\texttt{ErikSVD}\xspace and \texttt{Erik+2}\xspace methods}
\texttt{Erik+2}\xspace arises as a variation of the method described by \citet{eriksson2005} by normalizing certain bipartition matrices obtained from an alignment of nucleotide sequences. As in the original method, the information contained in the alignment is recorded as a vector $\tilde{p}$ whose coordinates are the observed relative frequencies of possible patterns at the leaves. In the case of an alignment of four taxa 1,2,3,4, each possible (trivalent) topology is determined by a bipartition of thetaxa: $12|34$, $13|24$ or $14|23$. For each bipartition $A|B$, a matrix $M_{A\mid B}(\tilde{p})$ is considered by rearranging the coordinates of $\tilde{p}$ according to it, so that the rows of the matrix $M_{12|34}(\tilde{p})$ are indexed by all possible observations $(x_1,x_2)$ at the leaves $1,2$, and similarly for columns and observations $(x_3,x_4)$ at the leaves $3,4$.
For example, the $(AG,CT)$-entry of $M_{12|34}(\tilde{p})$ is the relative frequency $\tilde{p}_{AGCT}$ of the pattern $AGCT$ in the alignment. The same entry in $M_{13|24}(\tilde{p})$ corresponds to the relative frequency $\tilde{p}_{ACGT}$ of $ACGT$.
\begin{eqnarray}\label{flattening}
M_{12\mid 34}(\tilde{p})=\left (
\begin{array}{ccccc}
\tilde{p}_{AAAA} & \tilde{p}_{AAAC} & \tilde{p}_{AAAG} & \ldots & \tilde{p}_{AATT} \\
\tilde{p}_{ACAA} & \tilde{p}_{ACAC} & \tilde{p}_{ACAG} & \ldots & \tilde{p}_{ACTT}\\
\tilde{p}_{AGAA} & \tilde{p}_{AGAC} & \tilde{p}_{AGAG} & \ldots & \tilde{p}_{AGTT}\\
\ldots & \ldots & \ldots & & \ldots \\
\end{array}
\right )
\end{eqnarray}
Assume that the coordinates of $\tilde{p}$ are the empirical estimates of the theoretical joint distribution $p$ at the leaves of a tree $T$ evolving under GMM, say $T=12|34$. Then
the key point is Theorem 19.5 of \citet{eriksson2005} (see \citet{casfer2010} for a complete proof) that claims that
the rank of $M_{A|B}(p)$ is $4$ if $A|B=12|34$, and $4^2$ otherwise (if the substitution matrices that generated $p$ were general enough). Eriksson's idea is to
compute the \emph{Frobenius distance} \citep[that is, the euclidean distance if we view the matrices as elements in $\mathbb{R}^{4^2\times 4^2}$,][]{Demmel} $\mathrm{d_4}$ of
the three matrices $M_{12|34}(\tilde{p})$, $M_{13|24}(\tilde{p})$ and $M_{14|23}(\tilde{p})$ to the space of matrices of rank $\leq 4$. In this manner, one derives which of the three matrices is closer to having rank $\leq 4$. The Frobenius distance of a matrix $A$ to $k$-rank matrices, $\mathrm{d_k}(A)$, is easily computed in terms of the singular values of $A$ \citep{Eckart1936}.
The main motivation for the variation introduced in \texttt{Erik+2}\xspace arises from the observation that the presence of short branches may seriously affect this distance when its computed from short alignments.
For example, taking the tree of Figure 1.a with small $a$ and large $b$ (a tree corresponding to the so-called Felsenstein's zone), the distance of $M_{13|24}(\tilde{p})$ to 4-rank matrices is smaller than that of $M_{12|34}(\tilde{p})$ (see Appendix 1 for an example). The reason is that a small $a$ implies that the probability of observing the same nucleotide at leaves 2 and 4 is high,
so columns in $M_{13|24}(\tilde{p})$ indexed by $AA$, $CC$, $GG$, or $TT$ capture most of the non-zero entries in the matrix, while other columns may only have few nonzero entries. This makes the matrix to be close to a rank $\leq 4$ matrix, even if $13|24$ is not the correct topology.
By dividing any non-zero column by the sum of its entries, we make all the non-zero columns to have the same weight. As the same situation may occur with rows, we also need to correct the matrix by row sums.
In this way, each matrix $M_{A|B}(\tilde{p})$ gives rise to a pair of \emph{transition} matrices $M_{A\rightarrow B}(\tilde{p})$ and $M_{A\leftarrow B}(\tilde{p})$, obtained by column and row sum correction, respectively:
\begin{footnotesize}
\begin{eqnarray*}\label{transmat}
M_{12\rightarrow 34}(\tilde{p})=\left (
\begin{array}{cccc}
\frac{\tilde{p}_{AAAA}}{\tilde{p}_{AA++}} & \frac{\tilde{p}_{AAAC}}{\tilde{p}_{AA++}} & \ldots & \frac{\tilde{p}_{AATT}}{\tilde{p}_{AA++}} \\
\frac{\tilde{p}_{ACAA}}{\tilde{p}_{AC++}} & \frac{\tilde{p}_{ACAC}}{\tilde{p}_{AC++}} & \ldots & \frac{\tilde{p}_{ACTT}}{\tilde{p}_{AC++}}\\
\frac{\tilde{p}_{AGAA}}{\tilde{p}_{AG++}} & \frac{\tilde{p}_{AGAC}}{\tilde{p}_{AG++}} & \ldots & \frac{\tilde{p}_{AGTT}}{\tilde{p}_{AG++}}\\
\ldots & \ldots & & \ldots \\
\end{array}
\right ) \quad
M_{12\leftarrow 34}(\tilde{p})=\left (
\begin{array}{cccc}
\frac{\tilde{p}_{AAAA}}{\tilde{p}_{++AA}} & \frac{\tilde{p}_{AAAC}}{\tilde{p}_{++AC}} & \ldots & \frac{\tilde{p}_{AATT}}{\tilde{p}_{++TT}} \\
\frac{\tilde{p}_{ACAA}}{\tilde{p}_{++AA}} & \frac{\tilde{p}_{ACAC}}{\tilde{p}_{++AC}} & \ldots & \frac{\tilde{p}_{ACTT}}{\tilde{p}_{++TT}}\\
\frac{\tilde{p}_{AGAA}}{\tilde{p}_{++AA}} & \frac{\tilde{p}_{AGAC}}{\tilde{p}_{++AC}} & \ldots & \frac{\tilde{p}_{AGTT}}{\tilde{p}_{++TT}}\\
\ldots & \ldots & & \ldots \\
\end{array}
\right ).
\end{eqnarray*}
\end{footnotesize}
We give a score to any tree $T_{A|B}$ as
\begin{eqnarray*}\label{formula_cft}
\mathrm{sc}(T_{A\mid B}):=\frac{\mathrm{d_4}\left(M_{A\rightarrow B}(\hat{p}) \right) + \mathrm{d_4}\left(M_{B\rightarrow A}(\hat{p}) \right)}{2}.
\end{eqnarray*}
Notice that the smaller the score is, the more reliable the topology $T_{A|B}$ is and \texttt{Erik+2}\xspace outputs as correct tree the topology with smallest score.
As the empirical distribution $\hat{p}$ approaches the theoretical distribution $p$, the transition matrices $M_{A\rightarrow B}(\hat{p})$ and $M_{B\rightarrow A}(\hat{p})$ approach the theoretical transition matrices. These have rank 4 for the correct topology because they have the same rank as the theoretical bipartition matrices (as they are obtained from them by dividing rows/columns by scalars). Therefore $\mathrm{d_4}\left(M_{A\rightarrow B}(\hat{p}) \right)$ and $\mathrm{d_4}\left(M_{B\rightarrow A}(\hat{p})\right)$ tend to 0 when $\hat{p}$ approaches the theoretical distribution (as the Frobenius distance is a continuous function) and thus \texttt{Erik+2}\xspace is statistically consistent.
\texttt{Erik+2}\xspace provides also normalized weights that can be used into weighted quartet-based methods.
Indeed, the score above is turned into a confidence weight by inverting it and normalizing so that the overall sum of weights is 1:
\[w(T_{A|B}):=\frac{\mathrm{sc}(T_{A|B})^{-1}}{\sum_{T\in \mathcal{T}_4}\mathrm{sc}(T)^{-1}}.\]
The basic model underlying \texttt{Erik+2}\xspace and \texttt{ErikSVD}\xspace assumes that all sites in the alignment evolve independently and identically distributed according to a general Markov model. There is no extra assumption about the shape of substitution matrices (nor stationarity, nor time-reversibility, nor global or local homogeneity). But in \texttt{Erik+2}\xspace we relax the i.i.d hypotheses and allow HAS by considering mixtures in the sense of \citep{Kolaczkowski2004} and \citep{Stefankovic2007}. That is, a single tree topology $T$ is considered but we allow $m$ categories of Markov processes on $T$ defined by $m$ sets $(\sigma_1,\dots,\sigma_m)$ of substitution parameters.
The proportion of sites contributed by the
$i$-th tree $(T,\sigma_i)$ is denoted by $p_i$ and the joint distribution at the leaves of $T$ follows an $m$-\textit{mixture distribution}: $\sum_i p_i P(T,\sigma_i)$.
A parameter $m \in \{1,2,3\}$ can be passed to \texttt{Erik+2}\xspace to adapt the method to consider $m$ categories (in this case, we compute the distance $d_{4m}$ to matrices of rank $\leq 4m$). The restriction to 3 categories at most is only due to theoretical results about non-identifiability for quartet trees with four or more partitions (there would be 255 parameters in a 4-mixture, which already fills the whole space of pattern distributions, see \citet{casferked}).
We had also developed different modifications of the original method of Eriksson, all of them
showing lower success than the version considered here. Therefore in this paper we only present the
results corresponding to \texttt{Erik+2}\xspace.
\subsection{\texttt{ML}\xspace and \texttt{NJ}\xspace}
Software PAML \citep{yang1997} was used to estimate the likelihood under time-homogeneous models. Depending on the simulations we used either the most general continuous-time homogeneous model, denoted as \texttt{ML}(hom) throughout the paper (model UNREST in PAML documentation), or the homogeneous time-reversible model denoted as \texttt{ML}(homGTR). Rate matrix entries and root distribution had to be estimated by the software. We waited up to 60 seconds for convergence on each tree topology and if it did not converge, we treated it as failed (because we cannot compare likelihoods in this case). It is worth pointing out that, usually, \texttt{ML}\xspace was not convergent only for the incorrect topologies.
In order to estimate HAL time-reversible model we used the software \texttt{bppml}\xspace of the Bio++ package \citep{Dutheil2008} for the inference of HAL models with homogeneity across sites, \texttt{ML}(HALGTR), and with discrete Gamma rates with two categories, \texttt{ML}(HALGTR+2$\Gamma$).
As far as Neighbor-joining is concerned, the paralinear distance \citep{lake1994} was always used to estimate pairwise divergence.
\subsection{Simulations}
To generate data under the general Markov model, we have used
\texttt{GenNon-h} \citep{GenNonh}. Given a set of branch lengths and a tree topology, this software generates random root distribution and substitution matrices with the expected substitutions per site, and lets nucleotides evolve according to this Markov process on the tree.
In order to generate data evolving under homogeneous GTR model (with or without continuous Gamma-rates) we have used \texttt{Seq-gen} \citep{Rambaut1997}. We used uniform root distribution, and the rate matrix underlying \texttt{Seq-gen} alignments on the tree space (Figure \ref{treespaceGTR} and Appendix 2.S2) had rates 2 (A$\leftrightarrow$C), 7 (A$\leftrightarrow$G), 4 (A$\leftrightarrow$T), 3 (C$\leftrightarrow$G), 1 (C$\leftrightarrow$T), 5 (G$\leftrightarrow$T), while the rate matrix underlying GTR+Gamma-rates had rates 2 (A$\leftrightarrow$C), 5 (A$\leftrightarrow$G), 3 (A$\leftrightarrow$T), 4 (C$\leftrightarrow$G), 1 (C$\leftrightarrow$T), 2 (G$\leftrightarrow$T).
\section{Acknowledgments}
We are indebted to B. Misof and C. Mayer for encouraging us to pursue this project and for their very useful comments. We wish to thank them and their group for their warm hospitality during our stay at the Alexander Koenig Zoological Museum.
Both authors are partially supported by
Spanish government MTM2012-38122-C03-01/FEDER and Generalitat de Catalunya 2009SGR1284.
\section{Author's contributions}
Both authors contributed equally to the development of this work.
|
\section{Introduction}\label{intro}
The study of young stellar objects is important for understanding the phenomena
that occur in star and planet formation, including our solar system. T Tauri
stars are young ($\sim 10^{6}$ years), low-mass stars (M $\leq$ 2 M\sun) in the
pre-main sequence (PMS) phase, which are of great interest as prototypes of
young solar-type stars. They emit X-rays and have strong magnetic fields
\citep[$\sim$ 2 kG,][]{johnskrull01art}. Based on their H$\alpha$ emission
strength, they are classified as classical (CTTSs) or weak-line T Tauri stars
(WTTSs). CTTSs have a flux excess with respect to the stellar photosphere at
infrared, optical, and ultraviolet wavelengths, which is not observed in WTTSs.
The CTTSs also exhibit irregular photometric and spectroscopic
variability, with broad emission lines, strong H$\alpha$ emission, and forbidden
emission lines. Most of these general features are reproduced by magnetospheric
accretion models \citep{shu94,hartmann94,muzerolle01,kurosawa06,lima10}, in
which a young magnetized star accretes material from a circumstellar disk. The
stellar magnetic field is strong enough to disrupt the disk at a distance of few
stellar radii from the star. Ionized material from the inner disk is then
channeled onto the stellar surface along field lines, creating accretion
funnels. Hot spots are produced at the stellar surface by the strong shock of
material at free-fall velocity. Material is also ejected from the system in
the form of a stellar wind and as a disk wind along open magnetic field lines.
From this model, we can explain some characteristics observed in CTTSs. Hot
spots are responsible for optical and ultraviolet excess emission. The broad
emission lines, which usually present redshifted absorptions, are predominantly
produced by accelerated material in the accretion funnels. Blueshifted
absorptions and forbidden emission lines are formed in the low-density wind.
Infrared excess emission comes from the reprocessing by the disk of radiation
generated in the system.
Although the magnetospheric accretion scenario describes the general CTTSs
properties, some observational results indicate that this axisymmetric, stable
model is not completely correct \citep{bouvier07b}. Some studies demonstrated
that the outflow and inflow processes are intimately connected
\citep{cabrit90,hartigan95}. The interaction between the stellar magnetosphere
and the disk is expected to be very dynamic, as shown by magnetohydrodynamical
(MHD) simulations \citep{goodson99,romanova02}. As a result of differential
rotation between the star and the inner disk region where the accretion flux
originates, the field lines may be distorted after a few rotational periods,
eventually reconnecting and restoring the initial field configuration. This
process repeats as the star rotates. Furthermore, a misalignment between the
rotation and magnetic axes creates a deformation in the inner disk, leading to
the formation of non-axisymmetric regions where accretion is favored
\citep{terquem00,romanova03}.
One of the best-studied CTTSs is AA Tau, observed for a month during three
different campaigns \citep{bouvier99,bouvier03,bouvier07}. This star exhibits a
peculiar photometric behavior, with almost constant brightness interrupted by
quasi-cyclical and irregular episodes of attenuation. The brightness decrease
occurs in absence of significant color variation and with an increase of
polarization \citep{menard03}. This was interpreted as occultation of the
stellar photosphere by circumstellar dusty material present in an inner disk
warp. The warp is produced by the interaction between the disk and the stellar
magnetic field, misaligned with respect to the rotation axis, as suggested by
MHD simulations. Simultaneous high-resolution spectroscopy has shown signs of
correlation between mass accretion and ejection processes. During the second
observing campaign, the photometric and spectroscopic variability was
drastically reduced for a few days, revealing an episode of disruption of
magnetic configuration at the inner edge of the disk, which suppressed accretion
temporarily. The observed variations in H$\alpha$ absorption components
showed the cyclical process of inflation and reconnection of field lines
caused by the differential rotation between the star and the inner disk region,
once again in agreement with MHD simulations. The same evidence of this
dynamical interaction was observed again in the third campaign, five years
after the second one.
Even though the photometric behavior of AA Tau was atypical at that time, the
characteristics of the structure of its accretion zone might be similar in
other CTTSs. To confirm this hypothesis, it would be necessary to obtain
high-precision photometry for many CTTSs during many rotational periods.
This corresponds to a continuous observation of at least a month, based on
typical rotational periods of eight days for CTTSs in Taurus. Such a monitoring
from the ground is complicated because of telescope-time allocation and weather
conditions. The CoRoT satellite additional program to observe the star formation
region NGC 2264 during 23 days uninterruptedly in 2008 allowed this analysis to
be performed.
NGC 2264 is a well-studied young stellar cluster, located in the Mon OB1
association at $\sim$ 760 pc \citep{sung97,gillen14}, with evidence of
active star formation \citep{dahm08}. The pioneering work of \citet{herbig54},
\citet{walker56}, and others established this region as an important laboratory
for studies of star formation and evolution of young stars. From the light
curves of 83 CTTSs identified among the 301 cluster members observed with CoRoT
in 2008, 23 have been classified as exhibiting the same type of variability as
AA Tau \citep{alencar10}. This revealed that the photometric behavior of AA Tau
is common in young stellar objects; it was found in 28\% $\pm$ 6\% of the CTTSs
in NGC 2264 observed with CoRoT. These systems are essential to test predictions
of MHD simulations.
The main goal of the work presented in this paper is to characterize the mass
accretion and ejection properties of the CTTS V354 Mon, a member of NGC 2264,
from simultaneous high-resolution photometric and spectroscopic observations.
V354 Mon is one of the 23 cluster members whose light curves have been
classified as AA Tau type. From the analysis of these data, we identify
correlations between emission line variability and light-curve modulation. It is
possible to investigate the dynamical processes of mass accretion and outflow
that occur in this system, as well as the interaction between the stellar
magnetic field and the circumstellar disk. We also test predictions of
magnetospheric accretion models and MHD simulations, constructing plausible
scenarios for the phenomena observed in this young object.
The paper is organized as follows: Sect. 2 describes the observations and
data reduction. In Sect. 3 we present the data analysis and results. In
Sect. 4 we discuss the possible causes of the observed photometric and
spectroscopic variations. The conclusions are depicted in Sect. 5.
\section{Observations}\label{obs}
CoRoT observed NGC 2264 from March 7 to 30, 2008. For V354 Mon, a cluster member
in the CoRoT observational program, high-resolution spectroscopy and \BVRI
photometry were obtained simultaneously with the SOPHIE \'echelle spectrograph
at the Observatoire de Haute-Provence (OHP - CNRS, France) and the 60 cm
telescope equipped with a CCD camera and Johnson/Cousins filters at the
Observat\'orio Pico dos Dias (OPD - LNA, Brazil), respectively. The journal of
observations is given in Table \ref{t:regobs}. V354 Mon is a CTTS with a
well-determined photometric period, which increases the chances that this system
is being seen edge-on, and exhibits a large \ha equivalent width, which
facilitates the study of its spectroscopic variability with good signal-to-noise
ratio, at least in this line. Information from the literature about V354 Mon is
gathered in Table \ref{t:infov354}.
\begin{table}[ht]
\caption{Journal of observations.}
\label{t:regobs}
\centering
\begin{tabular}{lccc}
\hline\hline
Dates (March 2008) & Instrument & Exp. time (s) & N$_{obs}$ \\
\hline
{\bf Spectroscopy} & & & \\
12, 14-1, 14-2, & SOPHIE & 3600 & 10\\
18-1, 18-2, 20, & (OHP) & & \\
23, 25, 27, and 28 & & & \\
\hline
{\bf Photometry} & & & \\
7 to 30 & CoRoT & 512 & cont.\\
19, 20, 25, 26, and 27 & 60 cm & B 400 & 10\\
& (OPD) & V 300 & 8\\
& & R 200 & 8\\
& & I 200 & 10\\
\hline
\end{tabular}
\tablefoot{On March 14 and 18, we obtained two spectra of V354 Mon, referred to
as 14-1, 14-2 and 18-1, 18-2.}
\end{table}
Many results presented in this work are dependent on the effective temperature
of the star. \citet{lamm05} assigned a K4 spectral type to V354 Mon, which
agrees with the $T_{\rm eff}$ value of 4590 K obtained by \citet{flaccomio06},
using the \citet{kenyon95} spectral type/intrinsic color scale that is, however,
more appropriate for main-sequence dwarfs than young stars. V354 Mon was
observed in 2011 with the FLAMES spectrograph (VLT/ESO) during a multiwavelength
campaign of the NGC 2264 star-forming region \citep{cody14} when 22
medium-resolution spectra (R = 17\,000) were acquired. To check the V354 Mon
spectral type, we estimated the effective temperature of V354 Mon with the
FLAMES spectra, using line ratios. A series of synthetic spectra were calculated
with the code Spectroscopy Made Easy (SME) \citep{valenti96}, using the same
resolution as the FLAMES observations, within the spectral domain of the FLAMES
data (6440 \ang $< \lambda <$ 6820 \ang) and the 3500 K to 6000 K temperature
range, with $\log{g}=4.0$ and 4.5 and solar metallicity. In this range of
$T_{\rm eff}$ and $\log{g}$, FeI lines are very $T_{\rm eff}$ dependent, while CaI
lines are not. The CaI/FeI ratio is therefore a good $T_{\rm eff}$ indicator that
is also independent of veiling. We analyzed the CaI 6717.7 \ang and FeI 6546.2
\ang lines that were always present and easily identified in the spectra.
Comparing the line ratios measured in the observed and theoretical spectra, we
obtained for V354 Mon $T_{\rm eff}=4647 \pm 161$ K with $\log{g}=4.5$ (dwarfs)
and $T_{\rm eff}=4434 \pm 133$ K with $\log{g}=4.0$ (young stars). The errors come
from the standard deviation of the values obtained with all the FLAMES spectra.
Recently, \citet{pecaut13} compiled tables of effective temperature vs. spectral
type for both dwarfs and young (5-30 Myr) stars. We compared the $T_{\rm eff}$
values we obtained for V354 Mon with their respective tables, and they are both
consistent with a K4 spectral type. Recently, \citet{marinas13} assigned a K7
spectral type to V354 Mon (their target 26) based on an [OH]/[MgI] line ratio
calibration obtained from {\it H} band low-resolution spectra. They acknowledged
an error of two subclasses to spectral types obtained with their calibration.
However, in the K7-K3 spectral range, the OH (1.69 $\mu$m) line is about 5 to 10
times shallower than the MgI (1.50 $\mu$m) line, which can complicate its
evaluation in low-resolution spectra. This, together with the scatter present in
their calibration, possibly adds some more uncertainty to their spectral type
determination in the K7-K3 region. Our results based on the FLAMES spectra are
not compatible with a K7 spectral type, which would require much lower
$T_{\rm eff}$ values \cite[4050 K for dwarfs and 3970 K for young stars, according
to the tables of][]{pecaut13} than we obtained. We decided therefore to adopt a
K4 spectral type for V354 Mon, in agreement with the results of \citet{lamm05}
and \citet{flaccomio06}.
\begin{table}[h]
\caption{Characteristics of V354 Mon from the literature.}
\label{t:infov354}
\centering
\begin{tabular}{lc}
\hline\hline
Data & Ref.\\
\hline
V = 14.45 mag & 1\\
spectral type: K4V & 2\\
photometric period = $5.22 \pm 0.87$ days & 2\\
\ha equivalent width = $16.60$ \ang & 1\\
\ha width at $10\% > 270$ \kms & 3\\
log(T$_{\textrm{eff}}$/K) = $3.66$ & 4\\
log(L$_{\textrm{bol}}$/\ls) = $0.34$ & 4\\
mass = $1.50$ \ms & 4\\
log(age/years) = $6.42$ & 4\\
heliocentric radial velocity (Dec 2004): $19.42$ \kms & 3\\
\hline
\end{tabular}
\tablebib{(1) \citet{dahm05};
(2) \citet{lamm05}; (3) \citet{furesz06}; (4) \citet{flaccomio06}.
}
\end{table}
\subsection{Photometry}
The CoRoT data were reduced using a standard procedure
\citep{samadi07,auvergne09} and delivered in the form of a light curve. The
light curve was processed using a sigma-clipping filter that removed outliers
(mostly related to South Atlantic Anomaly crossings) and hot pixels, and
corrected for the effects associated with entrance into and exit from Earth
eclipses. The light curve was also rebinned to 512 s and corresponds to the
integrated flux in the CoRoT mask.
The OPD photometric observations were carried out on five nights between March
19 and 27, 2008. The 60 cm telescope was equipped with the CCD camera \#106
(SITe SI003AB) of $1024\times1024$ pixels and Johnson/Cousins \BVRI filters.
Integration times ranged from 200 s to 400 s, depending on the filter. We
processed raw images with the usual techniques within the IRAF environment,
which included bias subtraction and flat-field calibration. Differential
photometry between V354 Mon and a reference star was also performed using IRAF.
The reference star is TYC 750-1637-1 (spectral type F7V, V = 11.609;
\citealt{sung97}), located less than 2.2\arcmin\, away from V354 Mon and
recorded on the same images.
\subsection{Spectroscopy}\label{obs_spec}
The spectroscopic observations were collected on eight nights between March 12
and 28, 2008. We obtained ten high-resolution spectra at the 1.93 m OHP
telescope with the SOPHIE dual fiber échelle spectrograph \citep{perruchot08},
which yields 41 orders covering the 3872 \ang to 6943 \ang domain at a mean
spectral resolution of $\lambda / \Delta \lambda \simeq 75\,000$
(high-resolution mode), and simultaneously records the object and the
neighboring sky. All spectra have the same exposure time of 3600 s. The data
were automatically reduced via a standard procedure, adapted from the software
used with the HARPS spectrograph designed at the Geneva Observatory. The
reduction procedure includes bias subtraction, optimum extraction of the orders,
removal of cosmic rays, flat-fielding achieved through a tungsten lamp exposure,
wavelength calibration with a thorium lamp exposure, cross-correlation with an
appropriate numerical mask, and reconnection of spectral orders with barycentric
correction.
Unfortunately, the nebular emission present in the stellar spectrum and in the
sky spectrum are not equivalent, because the fibers used in the observations do
not respond equally to the same light source\footnote{This problem was first
investigated by selecting only nebular emission lines and calculating the ratio
between the highest intensity of these lines in the sky and stellar spectra. It
was later confirmed by OHP technicians through private communication.}. In
addition, nebular emission is highly variable on scales of arcseconds or less,
according to studies of HII regions \citep{henney99,mccollum04}. For these
reasons, contamination by nebular emission was hard to exclude properly and some
emission lines could not be analyzed. For each spectrum, we subtracted the sky
continuum level, normalized the stellar continuum to unity, and corrected from
the star radial velocity. Then we subtracted a template spectrum (HD190007) of
same spectral type as V354 Mon, resulting in circumstellar spectral profiles. We
also calculated the average profile of the spectra obtained on the same night,
since there is no significant variation between them, resulting in one spectrum
for each night of observation.
\section{Results}
\subsection{Photometry}\label{res_phot}
\begin{figure}
\centering
\includegraphics[width=9cm]{corotcl1.eps}
\includegraphics[width=9cm]{corotcl2.eps}
\caption{CoRoT light curve, continuous (top) and folded in phase (bottom)
with a period of $5.26 \pm 0.50$ days. Different colors represent different
rotational cycles in the order indicated in the lower right corner of the bottom
panel. Magnitudes are given on an arbitrary scale.}
\label{f:corotlc}
\end{figure}
The broadband, white-light curve of V354 Mon obtained with CoRoT (Fig.
\ref{f:corotlc}, top panel) displays a typical CTTS behavior, noticeably
periodic and with well-defined minima and maxima that vary in depth and width
from one rotational cycle to the other. A periodogram analysis \citep{scargle82}
of the light curve reveals a period of $5.26 \pm 0.50$ days, which is consistent
with the value obtained by \citet{lamm05}, $5.22 \pm 0.87$ days. This indicates
that the dominant source of photometric variability did not significantly change
on a timescale of a few years. The varying depth and width at minimum is more
evident when the light curve is folded in phase with the calculated period
(Fig. \ref{f:corotlc}, bottom panel).
We measured the percentage variability amplitude of the light curve as
[(Flux$_{max}$ - Flux$_{min}$)/Flux$_{median}$] $\times$ 100, obtaining a value
of 115\% for V354 Mon. This is one of the largest photometric variations in the
CTTSs observed with CoRoT, which range between 3\% and 137\% \citep{alencar10}.
As observed in Fig. \ref{f:corotlna}, the OPD data are consistent
with the CoRoT photometry. From the calculation of magnitude differences
between the observations of March 25 and 27, we note that the photometric
amplitude decreases toward longer wavelengths: $1.5 \pm 0.3$ in B,
$1.15 \pm 0.05$ in V, $0.6 \pm 0.5$ in R, and $0.4 \pm 0.1$ in I. Analyzing the
color variation (Fig. \ref{f:corotlna}), we see that the system becomes slightly
bluer as the flux increases.
\begin{figure}
\centering
\includegraphics[width=9cm]{corotlna3.eps}
\caption{CoRoT light curve, marked with OPD photometric observation dates,
compared with the V-band photometry and color variations of V354 Mon. Vertical
bars indicate the errors of each measurement. The magnitude scale on the
vertical axis is the same for each color plot.}
\label{f:corotlna}
\end{figure}
\subsection{Spectroscopy}
For the reasons stated in Sect. \ref{obs_spec}, contamination by nebular
emission present in the stellar spectrum was hard to exclude properly,
which restricted the analysis of emission lines. In this work we focus on the
\ha region. To exclude the nebular emission from calculations or identify
its contribution in the results, we defined the range of the contaminated
spectral region. In each sky spectrum we determined the wavelength values
that constrain the region around 6562.85 \ang where the flux is higher than the
normalized background continuum. The mean values obtained are 6561.4 and 6563.4
\ang, corresponding to velocity values of $-64.6$ and $27.1$ \kms, respectively,
relative to the spectral line center at the stellar rest frame.
The circumstellar profiles exhibit a remarkable variability (Fig.
\ref{f:haspec}), with a well-distinguishable nebular emission in the central
part of the plots. We show in Fig. \ref{f:medvarha} the \ha line average profile
with its normalized variance \citep{johns95}, which measures how much each
region of the observed profiles varies with respect to the average profile. We
note that the blueshifted region is more variable than the redshifted one. The
blueshifted absorption seen in the average profile is generally associated with
winds in the system.
\begin{figure}
\centering
\includegraphics[angle=+90,width=8cm]{haspecnew.eps}
\caption{\ha circumstellar spectra for each observed night, indicated at the
right of each profile. The continuum of each spectrum has been normalized to
unity and the profiles have been shifted for clarity.}
\label{f:haspec}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=9cm]{medvarha2.eps}
\caption{Average profile (solid line) and normalized variance (gray shaded area)
of \ha line. Dashed red lines delimit the region dominated by nebular emission.}
\label{f:medvarha}
\end{figure}
Using the same method as was applied to calculate the light-curve period
discussed in Sect. \ref{res_phot}, we investigated the \ha normalized flux
periodicity through a periodogram analysis of the observed time series, which we
did independently in each velocity bin of 0.5 \kms along the profile. The
results were disposed side by side in space velocity to form an image. The
two-dimensional periodogram is shown in Fig. \ref{f:hapercorr}, where the
normalized power scales from zero (white) to the highest value (black). The \ha
redshifted side displays periodicity in a broad region centered on 5.3 days,
which is close to the photometric period, while the blueshifted side is
variable at around 3.1 days.
\begin{figure}
\centering
\includegraphics[width=9cm]{perHa_bin0_5_2_08.eps}
\caption{Two-dimensional periodogram of \ha line flux. The power scale varies
from zero (white) to the highest value (black). Dashed red lines delimit the
region dominated by nebular emission. The solid green line marks the period of
5.26 days.}
\label{f:hapercorr}
\end{figure}
The period obtained for the variation of the \ha redshifted side is close to
the photometric period, which is interesting from the point of view of
magnetospheric accretion with a misalignment between the rotation and magnetic
axes. In this scenario, material from the inner part of the disk is more easily
channeled along the field lines that are deflected from the disk mid-plane
towards the stellar surface because of the tilt of the stellar magnetosphere.
Thus there is a preferential region of accretion, causing the \ha redshifted
absorption to vary with the stellar rotational period. This absorption will be
more pronounced and at higher velocities when we see the accretion region
projected onto the stellar photosphere along our line of sight.
The modulation of the \ha blueshifted side presents a period shorter than the
photometric period. In this case, the existence of only one preferential region
of ejection of material from the disk is not physically acceptable, as the wind
generation is located immediately beyond the accretion region, at a slightly
larger distance from the star, which corresponds to a Keplerian period also
slightly longer. The period of $\sim$ 3 days may indicate that in fact we
observe two major contributions of the wind, at opposite sides. Consequently,
the formation region of these components is related to a Keplerian period of
$\sim$ 6 days, slightly longer than the period of the variation of \ha
redshifted side, which is connected with the accretion process.
From analyzing the \ha profiles, we observe that the spectrum obtained on March
25 shows a strong emission in the blueshifted side and no emission in the
redshifted side. Because no other profile resembles it, this blueshifted
emission might be a sign of a larger, occasional ejection of material, maybe
related to the disruption of the magnetic field configuration. Since this
isolated event is not linked to the rotational modulation of the system, we made
a new analysis of the \ha flux periodicity excluding the spectrum of March 25.
We recovered the results obtained considering all spectra, but the blueshifted
side also presented a broad region around the photometric period, while a
periodicity of about 3 days in the redshifted side becomes evident, making the
new 2D periodogram symmetric. Then all \ha components shows periodic variation
according to the photometric modulation. However, the period of 3 days has to be
confirmed with a richer dataset because we have only a few points to investigate
additional periods.
The \ha profiles are shown as a function of rotational phase with photometric
period (Fig. \ref{f:hanadfase}) to identify similarities and differences between
the spectra that are close in phase position. The top plot helps to locate the
spectroscopic observations in the light curve and to identify a correlation
between them. The profile of March 28 resembles the profile of March 18, with a
slight difference in the redshifted absorption component. They are close to each
other in phase, 0.11 and 0.20, and at similar points of the light-curve profile,
in which there is a flux decrease before a small local peak. Similarly, the
spectra of days 12, 28, 23, and 18 are similar, with an asymmetric profile and a
more intense redshifted side. All of them are located in the first half of the
photometric cycle, in which the stellar brightness is reduced. On the other
hand, the spectra of days 25, 20, and 27 are also similar, with an asymmetric
profile as well, but are more intense in the blueshifted side. These
observations are located in the flux increase of the light curve. The spectrum
of March 14 seems to be a transition between these two situations, because it is
situated in a photometric minimum and presents a distinct profile. We conclude
that there is a correlation between the light-curve modulation and the spectral
line variability. The phenomenon that produces these variations appears to be
asymmetric, since the brightness increase in the light curve seems to be slower
than the decrease and the emission profiles observed in these two phase ranges
exhibit different characteristics.
\begin{figure}
\centering
\includegraphics[width=8.5cm]{clfase.eps}
\includegraphics[angle=+90,width=7.5cm]{hafase.eps}
\caption{\ha profiles ordered according to rotational phase (middle number in
each panel). The observation date and CoRoT JD are also displayed (top and
bottom numbers). The color code is the same as in Fig. \ref{f:corotlc}, bottom
panel. The dashed vertical line marks the central position of the line,
corresponding to 6562.85 \ang. The top panel helps to locate the spectroscopic
observations in the light curve and to identify the correlation between them.}
\label{f:hanadfase}
\end{figure}
\section{Discussion}
The most often discussed causes of photometric variations at optical
wavelengths in young stars are cold spots produced by magnetic activity, hot
spots created by the shock of the accretion flow at the stellar photosphere, and
partial occultation of the star by inhomogeneous circumstellar material
\citep{herbst94}. Of these, only cold spots are not associated with an accretion
disk. We analyze below the photometric and spectroscopic variability expected in
each one of these scenarios and compare this with the observed characteristics
of V354 Mon in an attempt to identify the physical phenomenon that dominates the
observed optical variability in this system.
\subsection{Cold and hot spots}
Cold spots are one of the most common sources of photometric variability in
low-mass stars at all ages. They are good indicators of magnetic activity, since
they are associated with the eruption of magnetic flux from the stellar interior
out into the atmosphere. Hot spots in young stars are caused by infalling gas at
the stellar photosphere, which is a direct consequence of accretion.
Spots rotate with the star, as they are located in the photosphere, and generate
periodic photometric variability on timescales of the stellar rotational period.
If a cold spot is the main cause of light-curve modulation, it is fully visible
to the observer at light-curve minimum, while a hot spot is fully visible at
maximum. Both spots produce a modulation in the photometric amplitude that
increases toward shorter wavelenghts because of the temperature difference
between the spot and the photosphere, which makes the star bluer as it becomes
brighter. But this effect is more pronounced for hot than for cold spots. In
this manner, we can estimate the spot temperature and size by comparing the
total photometric amplitude of variation at different wavelengths.
We applied the model developed by \citet{bouvier93}, which derives the
temperature and the smallest size of a spot that is responsible for the
modulation of the stellar brightness through the fit of amplitude variability as
a function of wavelength. This model searches for the best spot configuration
that reproduces the observed amplitudes using a $\chi^2$ method, taking into
account limb-darkening effects and simultaneously fitting the amplitudes in all
bands. No assumption is made about the spot number and shape, but it is assumed
that all have the same temperature and the temperature distribution in each
individual spot is uniform. Therefore this model does not determine the spot
location over the star and provides only a lower limit for the fractional area
of the visible stellar hemisphere covered by spots.
We considered that the star has an effective temperature of 4500 K
\citep{flaccomio06}, with limb-darkening coefficients consistent with this
temperature given by \citet{claret00} for $\log g \sim 4.0$ and $\log$[M/H]
$\sim$ 0. Evaluating spots with temperatures of
$3.0 \leq \log(T_s/\textrm{K}) \leq 4.0$ and fractional area $f$ of the visible
stellar hemisphere between 0.1\% and 90\%, the model converged to a spot of
$T_s=10\,000$ K and $f=5\%$ as the best configuration that reproduces the
observed amplitudes of V354 Mon. Even though this solution is at the boundary of
model parameters because it presents the highest possible spot temperature, this
result is plausible because it agrees with hot spot parameters derived for
other T Tauri stars \citep{bouvier95}.
Although the spot model indicates that hot spots can reproduce the amplitudes of
variation better than cold spots, we restricted the fit to spots with
$T_s < 4500$ K to confirm that cold spots cannot be considered as the main cause
of photometric modulation. In this case, we obtained $T_s=3981$ K and $f=89\%$
as the best configuration. But this result is implausible because cold spots in
TTSs typically cover less than $50\%$ of the projected stellar disk
\citep{bouvier95}. We conclude that cold spots probably exist on the stellar
surface of V354 Mon, but are not the main cause of the light-curve variability.
The effective temperature of V354 Mon adopted from \citet{flaccomio06} was
derived based on the \citet{kenyon95} spectral type/intrinsic color scale, which
is more adequate to main-sequence dwarfs than young stars. As discussed in
Sect. \ref{obs}, \citet{pecaut13} have recently compiled a table of effective
temperature vs. spectral type for 5-30 Myr old PMS stars. According to this
scale, a K4 star presents an effective temperature of 4330 K. Considering this
value, the spot model yields results that are very similar to those obtained for
4500 K: $T_s=10\,000$ K and $f=4\%$ as the best configuration that reproduces
the observed amplitudes of V354 Mon, and $T_s=3852$ K and $f=89\%$ if we
restrict the model to cold spots only.
A hot spot influences not only the light curve of the star, but also its
spectrum because the accretion shock produces a continuum flux excess with
respect to the stellar photosphere that veils the observed photospheric lines,
reducing their depth. During the process of data reduction, we found no
evidence of veiling, as shown in Fig. \ref{f:compad}. We calculated the flux
generated by a hot spot of $10\,000$ K and $f=5\%$, the best spot
configuration that reproduces the observed amplitudes of V354 Mon in the \BVRI
bands, through a blackbody curve integration for $10\,000$ K convolved with
passband response functions \citep{bessell83}. The ratio between the values
obtained and the stellar blackbody flux without a spot at 4500 K was 2.84 in
B, 1.34 in V, 0.86 in R, and 0.55 in I. This means that this hot spot should
produce a remarkable veiling that would be measurable when comparing absorption
lines observed at photometric maximum and minimum. We note in the top panel of
Fig. \ref{f:liveil} that the LiI line, one of the best-defined absorptions in
the spectrum, displays no difference in depth between these two occasions. To
exemplify that the flux of a spot of $10\,000$ K and $f=5\%$ would cause a
visible effect, we added the calculated veiling to the LiI absorption observed
at photometric maximum, when the hot spot would be fully visible. We applied the
values obtained for R and I bands, as they have characteristic wavelengths of
5925 \ang and 7900 \ang, and the LiI line is located at 6707.8 \ang. The result
is shown in the bottom panel of Fig. \ref{f:liveil}, in which we observe that
there is a noticeable decrease in the line depth due to the added veiling. The
fact that we do not observe this in the spectra of V354 Mon refutes a hot spot
as the main cause of photometric variability.
\begin{figure}
\centering
\includegraphics[width=9cm]{compad.eps}
\includegraphics[width=9cm]{compad2.eps}
\caption
{Two regions of the V354 Mon spectrum (black line) obtained at a photometric
maximum on March 12, superimposed on the standard star spectrum (red line),
corrected for its radial velocity and rotationally broadened to the V354 Mon
value. The photospheric absorptions in the spectrum of both stars agree well,
indicating that there is no measurable veiling in the spectrum of V354 Mon. In
the top plot, which represents the \ha region, we observe the nebular
contribution to this line in the V354 Mon spectrum and the photospheric
contribution to \ha in the standard spectrum.}
\label{f:compad}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=9cm]{li12_14.eps}
\includegraphics[width=9cm]{liveil2.eps}
\caption
{Comparison between LiI absorption lines observed on March 12 and 14,
corresponding to photometric maximum and minimum. In the top plot there is no
difference in line depth, showing that no veiling is detectable with brightness
increase. In the bottom plot we added veiling to the spectrum at photometric
maximum to demonstrate how a spot with temperature of $10\,000$ K and $f=5\%$
would affect the line depth if it were the main cause of the observed
photometric variations.}
\label{f:liveil}
\end{figure}
Furthermore, if a hot spot were responsible for the observed photometric
variations, the light-curve maximum would correspond to the moment when the hot
spot is fully visible to the observer and the minimum would correspond to the
opposite situation, when it is completely hidden. The material in free fall in
the accretion funnel absorbs photons emitted by the hot spot. Thus, the \ha
redshifted absorption should occur at light-curve maximum, since the spot would
be in our line of sight at this moment. The opposite is observed in V354 Mon
spectra, however. The \ha redshifted absorption is visible in the March 14
spectrum, located at a light-curve minimum, but is not present in the March 12
spectrum, corresponding to a light-curve maximum, as seen in Fig.
\ref{f:spec12_14}. According to this result, the hot spot should be facing the
observer at light-curve minima, which is not consistent with a hot spot as the
main cause of the photometric variation of V354 Mon. Nevertheless, the existence
of a hot spot is confirmed by the \ha redshifted absorption, since it indicates
that there is material at high velocity that falls onto the star and hits the
photosphere, which certainly produces a hot spot at the stellar surface. But
this hot spot does not generate a measurable veiling, at least within the error
of our measurements, or it may be hidden by the accretion column.
\begin{figure}
\centering
\includegraphics[width=9cm]{ha12_14.eps}
\caption
{Comparison between \ha circumstellar spectrum observed on March 12 and 14,
corresponding to photometric maximum and minimum. The dashed vertical line
indicates the central position of the \ha line. The appearance of a redshifted
absorption during the brightness minimum of the star is evident.}
\label{f:spec12_14}
\end{figure}
\subsection{Occultation by circumstellar material}
For young stars with an accretion disk seen at high inclination with respect to
the line of sight, disk material may absorb part of the stellar flux. If the
material is nonuniformly distributed, the light curve is modulated according to
the disk structure. Inhomogeneities can be caused by azimuthal asymmetries, a
disk with the outer part partially expanded (flared), or the inner part
distorted (warped), or even by dust in an inhomogeneous disk wind. In these
cases, the timescale of photometric variation depends on the disk angular
velocity at the radius where the inhomogeneity is located. Similarly to spots,
occultation by circumstellar disk makes the star bluer as it becomes brighter.
But there is no color variation with the brightness modulation if the disk
material is completely opaque, as observed in AA Tau
\citep{bouvier99,bouvier03,bouvier07}. The 3D MHD numerical simulations of
\citet{romanova04} naturally produced deformations in the inner part of the disk
when the magnetic axis did not coincide with the rotation axis, creating regions
where accretion is favored. This prediction has been confirmed by the study of
AA Tau, in which a dynamic interaction between the stellar magnetic field and
the inner disk was observed.
The presence of dust around V354 Mon is indicated by observations with the
InfraRed Array Camera (IRAC) of Spitzer \citep{teixeira08}, used to identify
infrared excess emission from hot circumstellar dust. The
$\alpha_{\textrm{\footnotesize IRAC}}$ index represents the slope of spectral energy
distribution between 3.6 $\mu$m and 8 $\mu$m, which is used to classify the
inner disk structure, following the criteria proposed by \citet{lada06}. V354
Mon presents an $\alpha_{\textrm{\footnotesize IRAC}} = -1.72$, indicating that this
star has an optically thick inner disk. The asymmetric shape of minima and
maxima observed in the light curve of V354 Mon obtained with CoRoT favors the
idea that an irregular structure obscures the light emitted by the star,
probably a circumstellar disk with nonuniformly distributed material, where a
denser region periodically intercepts the star in our line of sight. This
phenomenon can only occur if the disk is seen at high inclination. We calculated
the star inclination $i$ through the relation
\begin{equation}
\sin i = \frac {P(v \sin i)}{2 \pi R_*},
\end{equation}
\noindent
where $P$ is the rotation period, $v \sin i$ is the projected rotational
velocity, and $R_*$ is the stellar radius. Using the effective temperature
and luminosity of V354 Mon given by \citet{flaccomio06} and the PMS theoretical
evolutionary tracks for solar metallicity computed by \citet{siess00}, we
inferred a radius of 2.39 R\sun, a mass of 1.49 M\sun, and an age of
2.4$\times$10${^6}$ years. From a cross-correlation analysis of each spectrum
with an appropriate numerical mask, we derived a mean rotational velocity of
$v \sin i = 22.4 \pm 1.2$ \kms. Considering the CoRoT photometric period,
$P = 5.26 \pm 0.50$ days, we obtained 77\degr\,for the inclination $i$ of the
stellar rotation axis relative to the line of sight. This means that the system
presents a high inclination, which supports the possibility of star occultation
by circumstellar material.
The radius, mass, and age of V354 Mon were also estimated from \citet{landin06},
using a PMS model with non-gray atmosphere, solar metallicity, initially without
rotation and applying mixing length theory with $\alpha = 2$ to treat the
convective energy transport. The result (Table \ref{t:paramV354}) is slightly
different from that obtained with the \citet{siess00} model, mainly in mass and
age, but the values are on the same order of magnitude.
To investigate the possibility of occultation by circumstellar material as the
main cause of photometric variability, we applied the model originally developed
for AA Tau \citep{bouvier99} to V354 Mon. The obscuring region is identified as
a warp, a vertical deformation in the inner part of the circumstellar disk
produced by the interaction with the dipole magnetic field inclined with respect
to the rotation axis. The occultation model generates a synthetic light curve,
assuming that the height of the inner disk varies with the azimuthal position
according to
\begin{equation}
h(\phi)=h_{max}\left |\cos \frac{\pi (\phi - \phi_0)}{2 \phi_c} \right|,
\end{equation}
\noindent
where $\phi_0$ is the azimuth of maximum disk height, corresponding to the
middle of the eclipse in photometric phase, and $\phi_c$ is the warp azimuthal
semi-extension. Therefore, the height in the inner part of the disk decreases
gently from its highest value $h_{max}$ at $\phi_0$ to zero at
$\phi_0 \pm \phi_c$.
Assuming that the photometric period corresponds to the stellar rotation period
and the inner disk warp co-rotates with the star, like in AA Tau, the observed
photometric period of 5.26 $\pm$ 0.50 days would locate the warp at a distance
\begin{equation}
r_c=\left (\frac{P}{2\pi}\right)^{2/3}(GM_*)^{1/3}=6.1\,R_*
\end{equation}
\noindent
from the star, with $M_*$ and $R_*$ obtained from the model of \citet{siess00}.
At this distance, the magnetic field affects the accretion dynamics and the disk
is truncated. Internally to this radius, the flux of material is directed to the
star following the field lines. The inclination of the magnetic axis with
respect to line of sight can be quantified from the redshifted absorptions in
the emission line profiles because they are produced by the accretion funnel
material that falls at free-fall velocity in the stellar photosphere. In the
spectrum of V354 Mon obtained on March 14, located at the photometric minimum,
we observe an \ha redshifted absorption with a maximum velocity, projected on
the line of sight, of $125 \pm 25$ \kms. The free-fall velocity of the accretion
flow material from a truncation radius of 6.1 $R_*$ is
\begin{equation}
\textrm{v}_{ff}=\left [\frac{2GM_*}{R_*}\left (1-\frac{R_*}{r_c}\right)\right]^{1/2}\simeq 450 \textrm{ \kms},
\end{equation}
\noindent
with $M_*$ and $R_*$ obtained from the model of \citet{siess00}. Assuming that
the accreting material follows the dipole field lines\footnote{Given the
inferred mass and age of V354 Mon, it has already developed a small radiative
core. According to \citet{gregory12}, stars with this configuration present
large-scale magnetic fields that are axisymmetric and have high-order
components, typically the octupole, that dominate over the dipole at the stellar
surface. However, the dipolar component is generally stronger than the octupolar
at large distances from the star and then matter flows from the disk towards the
star along dipolar lines. The octupole is more important near the stellar
surface, redirecting the accretion streams towards the octupolar pole and
shaping the hot spots
\cite[see the case of V2129 Oph:][]{donati11,romanova11,alencar12}.}, the
inclination $i_m$ of the magnetic axis is calculated through the relation
$\cos i_m = $ v$_{max}/$v$_{ff}$, which gives a value of $\sim 0.28$,
i.e., $i_m \simeq $74\degr. As the inclination of the rotation axis is $\sim$
77\degr, the angle between this and the magnetic axis is only a few degrees at
most. The \citet{romanova03} MHD simulations have shown that for very small
angles of misalignment between the rotation and the magnetic field axis the
accretion flows become asymmetric.
We recalculated the inclinations of the rotation and the magnetic field axis
using the mass and radius values obtained from \citet{landin06} to show the
dependence of the results with the adopted theoretical evolutionary model. Even
though the radius values given by the two models are very close, the system
inclination is affected by the small difference between them. We obtained
$i= 82^\circ$ for $R_*=2.35$ \rs. The inclination of the magnetic field axis with
respect to the line of sight is $i_m= 71$\degr\,and the misalignment with the
rotation axis is 11\degr. We note that the differences between the system
parameters obtained from the two theoretical models (Table \ref{t:paramV354})
are negligible, given the expected uncertainties. The results are on the same
order of magnitude, confirming that the system is seen at high inclination and
the magnetic axis is only slightly tilted with respect to the rotation axis. The
misalignment between the magnetic and rotation axis is also observed in other
CTTSs \citep{donati10,donati11}.
\begin{table}[htp]
\caption{V354 Mon parameters obtained from two PMS evolutionary models,
using the effective temperature and luminosity given by \citet{flaccomio06}.}
\label{t:paramV354}
\centering
\begin{tabular}{l c c}
\hline
\hline
Model & \citet{siess00} & \citet{landin06} \\\hline
$R_*$ (\rs) & 2.39 & 2.35 \\
$M_*$ (\ms) & 1.49 & 1.12 \\
Age (years) & $2.4 \times 10^6 $ & $1.2 \times 10^6 $ \\
$i$ & $77^{\circ \: +13^\circ}_{\;\;\, -22^\circ}$ & $82^{\circ \: +8^\circ}_{\: -25^\circ}$ \\
$r_c$ ($R_*$) & $6.1 \pm 1.0 $ & $5.6 \pm 1.1$ \\
v$_{ff}$ (\kms) & $450 \pm 70$ & $390 \pm 80$ \\
$i_m$ & $74^\circ \pm 7^\circ$ & $71^\circ \pm 10^\circ$ \\\hline
\end{tabular}
\tablefoot{We consider the difference between both models as the error in the
radius and mass determination. Therefore, we calculated the errors of $i$,
$r_c$, v$_{ff}$, and $i_m$ taking into account the errors in the period,
rotational velocity, mass, radius, and \ha redshifted absorption v$_{max}$.}
\end{table}
To reproduce the V354 Mon photometric variability observed with CoRoT
by employing the occultation model, we considered the mass and radius values
derived from \citet{siess00}. The model free parameters are the warp maximum
height $h_{max}$ and azimuthal semi-extension $\phi_c$, which influence the
amplitude of variability and the duration of the eclipse, respectively. With
these parameters fixed, the best fit was obtained with $h_{max} =$ 0.3 $r_c$ and
a total azimuthal extension of 360\degr, represented in Fig. \ref{f:occmodel}
(top). The synthetic light curve closely follows the large-scale photometric
variation and reproduces the largest amplitudes observed, but does not fit the
change in the maxima and minima phase to phase. This is to be expected with
fixed warp parameters. However, it is interesting to note that the warp
parameters are similar to those obtained in the model of the variability of
AA Tau, with a maximum height larger than the value traditionally used in disk
models, $\sim$ 0.05 - 0.1 $r_c$ \citep{bertout88,duchene10}.
\begin{figure}
\centering
\includegraphics[width=11cm,angle=90]{grafocc2.eps}
\caption{Best fit of the occultation model with fixed parameters (top),
corresponding to a warp with maximum scale height of 0.3 and azimuthal extension
of 360\degr. Individual fit of the model to light-curve minima (bottom). The
corresponding properties of the warp are indicated in Table
\ref{t:paramoccm_ind}.}
\label{f:occmodel}
\end{figure}
The analysis of AA Tau photometric and spectroscopic variability indicated that
the large-scale stellar magnetosphere configuration changed over a month because
of the differential rotation between the star and the disk, showing that the
circumstellar disk dynamically interacts with the misaligned magnetic field.
Therefore, the deformation in the inner disk, which results from this
interaction, is also expected to change its characteristics on a short
timescale. Based on this and the fact that the depth and width of V354 Mon
photometric minima vary considerably, we modeled each light-curve minimum
individually. The warp properties are presented in Table \ref{t:paramoccm_ind}
for each cycle and the corresponding synthetic light curve is shown in
Fig. \ref{f:occmodel} (bottom). Despite the irregularity of the photometric
modulation, the parameters of the inner disk warp for each cycle are quite
similar, indicating that small changes in its height and azimuthal extension can
reproduce the large amplitude of variability in the observed light curve.
\begin{table}[htp]
\caption{Occultation model parameters from individual fit of light-curve
minima.}
\label{t:paramoccm_ind}
\centering
\begin{tabular}{c c c}
\hline
\hline
Minimum & $h_{max}$ ($r_c$) & $2 \phi_c$ (\degr) \\\hline
$1^{st}$ & 0.31 & 320 \\
$2^{nd}$ & 0.23 & 320 \\
$3^{rd}$ & 0.25 & 240 \\
$4^{th}$ & 0.28 & 320 \\
$5^{th}$ & 0.28 & 280 \\\hline
\end{tabular}
\end{table}
We derived new values of effective temperature and luminosity using the recently
published spectral type vs. effective temperature/intrinsic color scale of
\citet{pecaut13}, which is more appropriate for young stars than the
dwarf-temperature scales used by \citet{flaccomio06}. As the {\it J}-band
photometry is less affected by the disk and accretion column emissions than
other bands, we used the available 2MASS {\it J} magnitude \citep{skrutskie06}
and bolometric correction in this band to compute a luminosity of
log(L$_{\textrm{bol}}$/\ls) = 0.11. Considering this value and an effective
temperature of 4330 K, we recalculated the stellar parameters from the PMS
models of \citet{siess00} and \citet{landin06} (Table \ref{t:paramV354_2}).
Although the mass and radius have decreased compared with the values derived
using the effective temperature and luminosity given by \citet{flaccomio06}
(Table \ref{t:paramV354}), the overall characteristics of the system are on the
same order of magnitude. This corroborates the scenario of a young star
surrounded by a disk seen at high inclination, with a small misalignement
between the magnetic and rotation axis.
\begin{table}[htp]
\caption{V354 Mon parameters obtained from two PMS evolutionary models,
using the effective temperature and luminosity derived from \citet{pecaut13}.}
\label{t:paramV354_2}
\centering
\begin{tabular}{l c c}
\hline
\hline
Model & \citet{siess00} & \citet{landin06} \\\hline
$R_*$ (\rs) & 1.98 & 2.03 \\
$M_*$ (\ms) & 1.09 & 0.88 \\
Age (years) & $2.5 \times 10^6$ & $1.4 \times 10^6$ \\
$i$ & $85^\circ \pm 5^\circ$ & $82^\circ \pm 8^\circ$ \\
$r_c$ ($R_*$) & $6.6 \pm 1.0$ & $6.0 \pm 1.0$ \\
v$_{ff}$ (\kms) & $420 \pm 50$ & $370 \pm 60$ \\
$i_m$ & $73^\circ \pm 7^\circ$ & $70^\circ \pm 9^\circ$ \\\hline
\end{tabular}
\tablefoot{The process to calculate the errors was the same as described in
Table \ref{t:paramV354}.}
\end{table}
The change in the stellar parameters affected the result of the occultation
model as the inclination of the system and the corotation radius increased.
Considering the mass and radius obtained from \citet{siess00}, we reproduced
exactly the same fit as shown in Fig. \ref{f:occmodel} (top) with fixed
parameters with a warp of maximum scale height of 0.15 $r_c$ and total azimuthal
extension of 280\degr. The individual fit of the model to light-curve minima
(Fig. \ref{f:occmodel}, bottom) is also recovered with a warp that changes its
characteristics from 0.10 to 0.16 $r_c$ in maximum height and from 180\degr\,to
260\degr\,in azimuthal extension. As the inclination of the system is slightly
steeper in this case, the warp does not need to be too high and too extended to
generate the observed amplitude of variations. While the stellar parameters
derived using the scale of \citet{pecaut13} are expected to be more suitable for
this system, the results in the literature are largely based on the relations
compiled by \citet{kenyon95}. Therefore we kept the results obtained based on
the latter because they are easier to compare with similar published works.
According to the results obtained, occultation by circumstellar material can be
the main cause of the photometric variability observed in V354 Mon.
Spectroscopic evidence also favors this interpretation. Photons emitted by the
accretion shock or lower accretion column are absorbed by the funnel material in
free fall, producing a redshifted absorption in H$\alpha$. As seen from Fig.
\ref{f:spec12_14}, this absorption is more pronounced in the spectrum that
corresponds to a photometric minimum, but it does not appear in the spectrum
obtained in the maximum. This is evidence for a correlation between the position
of the accretion flow onto the star and the decreasing of stellar brightness.
The MHD simulations of \citet{romanova03} indicated that for a misalignment
between the magnetic field and the rotation axis smaller than 30\degr, the
densest regions of accreting material to the star are located in two main
funnels, following the field lines to the closest magnetic pole. Each region is
located in one hemisphere, one above the disk and the other below, and rotates
with the star. The region located above the disk periodically occults the star
for an observer that views the system at medium to high inclination. Therefore,
the spectroscopic observations and the CoRoT photometry, along with predictions
of numerical simulations, corroborate the idea of stellar occultation by
circumstellar material in this system.
A possible explanation for the color variation observed in V354 Mon \BVRI
photometry is extinction by circumstellar dust. The star becomes redder when
fainter because the warp may not be totally opaque, different from what was
observed in AA Tau. An additional analysis is necessary to confirm this
hypothesis as the real cause of V354 Mon color variation, which cannot be done
based on the data presented in this study.
\section{Conclusions}
From simultaneous high-resolution spectroscopic and photometric observations,
we analyzed the CTTS V354 Mon. This star exhibits a large, periodic brightness
variation with minima that change in shape from one rotational cycle to the
next. A periodogram analysis of the light curve obtained with CoRoT provided a
photometric period of 5.26 $\pm$ 0.50 days, close to the value derived by
\citet{lamm05}, 5.22 $\pm$ 0.87 days, which indicates that the main structure
that produces the photometric modulation did not significantly change over a few
years. Observations at \BVRI bands showed that there is also a small color
variation and the system becomes slightly bluer as the flux increases.
The spectrum of V354 Mon is variable on a timescale of a few days. The
periodicity of \ha redshifted and blueshifted sides supports the magnetospheric
accretion scenario, in which the star accretes material from the circumstellar
disk while it ejects mass through a disk wind that originates close to the
accretion region.
We investigated the possibility that spots at the stellar surface are the main
source of photometric variations. A spot with a temperature of $10\,000$ K and
occupying 5\% of stellar hemisphere is the spot configuration that best
reproduces the variability amplitudes observed in \BVRI bands, which means that
the possibility of cold spots is discarded. Such a hot spot would produce a
significant veiling in photospheric lines, which is not observed. In addition,
the occurrence of pronounced \ha redshifted absorptions seen only in light-curve
minima indicates that the accretion funnel and, consequently, the hot spot are
visible at this phase, which invalidates this phenomenon as the main cause of
photometric modulation.
We found evidence that the emission lines vary in a cyclic manner according to
the photometric modulation. The asymmetric shape of light-curve minima and the
difference in the spectral line profile at brightness increase and decrease show
that an irregular structure produces these variations, probably a circumstellar
disk with nonuniformly distributed material. This is supported by the fact that
the system is seen at high inclination. According to MHD simulations
\citep{romanova04}, the small misalignment between magnetic and rotational axes
observed in this star may create a distortion in the inner disk, producing a
warp that periodically occults part of the stellar photosphere. We used an
occultation model to determine the general parameters of this structure, which
exhibits a maximum scale height of 0.3 located near the disk corotation point,
with an azimuthal extension of 360\degr. These features are similar to those
obtained in the fit of this model to the variability of AA Tau
\citep{bouvier99}. The warp in the disk of V354 Mon seems to modify its shape at
each cycle, revealing a dynamical interaction between the stellar magnetosphere
and the disk inner part, as predicted by MHD simulations
\citep{goodson99,romanova02}. Nevertheless, the parameters obtained from the
individual fitting of the model to the light-curve minima are not very
different from each other. This result shows that, despite the irregularity of
the brightness modulation, small variations in the warp characteristics are
capable of reproducing the large amplitude of the observed photometric
variability. The presence of an \ha redshifted absorption component more
pronounced only at photometric minimum points to a connection between the
accretion funnel and the flux variations of the system, which corroborates that
occultation by circumstellar material is the main cause of photometric
modulation in V354 Mon.
\begin{acknowledgements}
We thank Suzanne Aigrain for pre-processing the CoRoT light curve and the
anonymous referee for many useful comments and suggestions. NNJF acknowledges
support from CAPES (fellowship process n$_{o}$.~18697-12-7) and CNPq. SHPA
acknowledges support from CAPES, CNPq, and Fapemig. JB acknowledges funding from
CNES and from ANR Toupie 2011 Blanc SIMI5-6 020 01. This research was developed
within the scope of CAPES-Cofecub project and is based on data collected in the
CoRoT satellite.
\end{acknowledgements}
\bibliographystyle{aa}
\input{v354mon_fin.bbl}
\end{document}
|
\section{Introduction}
\label{sect:intro}
Small-scale turbulent magnetic field is ubiquitous on the solar surface and provides the dominant contribution to the
magnetic energy in the quiet Sun photosphere (see, e.g., \citet{Lites:1996:small_scale_field,
Khomenko:2003:quiet_sun_infrared,Dominguez:2003:quiet_sun,Almeida:2003:internetwork_sol_min,Trujillo:etal:2004,
Dominguez:2006:quiet_sun,Orozco-Suarez:etal:2007,Lites:etal:2008,Bellot:2012:pervasive_lin_pol} and recent reviews by
\citet{deWijn:2009:SSRv,Mpillet:2013:SSRv}). Several investigations
found that inter-network magnetic field shows only little dependence on the solar cycle \citep{Trujillo:etal:2004,Buehler:2013:cyc_dep} and also little correlation
with the strength of the surrounding network field \citep{Lites:2011:hinode_ssd,Ishikawa:2009:cmp_QS_plage}. This points toward
an origin of the quiet Sun magnetic field largely independent from the global solar dynamo responsible for the solar
cycle.
It was suggested by \citet{Petrovay:1993:SSD}, based on a simplified transport model for signed and unsigned flux in the convection zone,
that a small-scale dynamo is the key process maintaining turbulent magnetic field in the quiet Sun. Small-scale dynamos were first
studied through MHD simulations in incompressible setups by \citet{Cattaneo:1999} and later with stratification (anelastic approximation)
by \citet{Bercik:2005:dynamo}. \citet{Voegler:Schuessler:2007} used the most "solar-like" setup by including realistic physics in terms
of equation of state and 3-dimensional radiative transfer. They were able to demonstrate that despite the lack of significant recirculation
within the computational domain (use of open bottom boundary conditions that mimic the deep convection zone), a considerable amount
of magnetic field can be maintained in the photosphere. It was found later that the photospheric field strength falls still short by a
factor of $2-3$ compared to observations based on Zeeman diagnostics \citep{Danilovic:2010:zeeman_dynamo}; an even more dramatic
shortfall by about one order of magnitude was found by \citet{Shchukina:2011:hanle_dynamo} based on Hanle-effect diagnostics. To
which degree the discrepancy between the magnetic field strength found in solar photospheric dynamo simulations and observations is
due to boundary conditions and resolution (which relates to the magnetic Reynolds numbers reached) is still an open issue, which we aim
to adress primarily in this paper.
On a more fundamental level the operation as well as non-linear saturation of small-scale dynamos at very small magnetic Prandtl
numbers ($P_M=\nu/\eta \ll 1$, with viscosity $\nu$ and magnetic diffusivity $\eta$) as encountered on the Sun ($P_M\sim 10^{-5}$)
remains an open question. While the regime $P_M > 1$
has been studied in great detail including the non-linear saturation phase \citep{Schekochihin:2004:SSD_high_Pm}, the regime $P_M\ll 1$
is less accessible by direct numerical simulations. Small-scale dynamos at low $P_M$ were studied in the kinematic regime by
\citet{Iskakov:2007:SSD_low_Pm,Schekochihin:2007:SSD_low_PM} and in the non-linear regime by \citet{Brandenburg:2011:SSD_low_Pm}.
These investigations indicate that the threshold for dynamo action increases moderately when approaching the $P_M\ll 1$ regime, but
small-scale dynamo action remains possible for large values of the magnetic Reynolds number typically found in astrophysical
systems \citep{Tobias:Cattaneao:Boldyrev:SSD_review,Brandenburg:2011:SSD_low_Pm}.
We present here an investigation which follows along the lines of comprehensive photospheric MHD simulations similar to the work presented by \citet{Voegler:Schuessler:2007,Pietarila-Graham:etal:2010:SSD}. In particular we address the problem that these simulations fall short by
a factor of about $2-3$ in field strength when compared to observations as pointed out by \citet{Danilovic:2010:zeeman_dynamo}. We explore
here the potential role of two factors: 1. Numerical resolution and magnetic diffusivities, 2. Influence from the bottom boundary condition.
While numerical resolution and treatment of diffusivities determine the dynamo growth rate in the kinematic phase, $\gamma_K$, the bottom boundary
condition determines the amount of magnetic energy that recirculates within the computational domain. It was first pointed out by
\citet{Stein:2003:recirculation} that the rather small recirculation of plasma found in the top layers of the convection zone could be a major
hurdle for a small-scale dynamo to exist locally in the photosphere. While the recirculation is small it is not zero, since
some level of turbulent mixing between up- and downflow regions is unavoidable. Magnetic energy loss due to overturning convection happens on
a rather slow time-scale $\sim H_{\varrho}/v_{z {\rm RMS}}$ (with density scale height $H_{\varrho}$ and vertical RMS velocity $v_{z {\rm RMS}}$),
which can be compensated by a sufficiently efficient small-scale dynamo. \citet{Voegler:Schuessler:2007} showed that a local
dynamo in the photosphere can operate despite small recirculation if the value of the magnetic diffusivity $\eta$ is smaller than
$2.5\cdot 10^{10}\mbox{cm}^2\mbox{s}^{-1}$ (for a magnetic Prandtl number close to unity).
\citet{Pietarila-Graham:etal:2010:SSD} studied photospheric dynamos in the kinematic growth phase using values of $\eta$ as low as
$4\cdot 10^{9}\mbox{cm}^2\mbox{s}^{-1}$, but did not study the non-linear saturation. The setup of
\citet{Voegler:Schuessler:2007,Pietarila-Graham:etal:2010:SSD} is conservative in terms of the bottom boundary condition. There is no vertical
Poynting flux in upflow regions and in addition an enhanced magnetic diffusivity near the bottom boundary. Since it is likely that the bulk of the
convection zone has strong magnetic field, it is reasonable to assume that upflow regions are magnetized and transport magnetic energy into the
photosphere.
In order to adress these questions we consider models that differ from previous studies in the following aspects:\vspace{0.1cm}\\1. We use only numerical diffusivities in an
attempt to minimize the influence from dissipation for a given numerical resolution. In comparison to direct numerical simulations (DNS) that use
explicit diffusivities and fully resolve the dissipation range, our setup is more along the lines of large eddy simulations (LES) in which a high Reynolds number
regime is realized on large scales, while the dissipation range is truncated through the use of a subgrid-scale model. In our case the latter is entirely based
on monotonicity constraints through the use of a slope-limited diffusion scheme (see Section \ref{sect:setup} for detail).
This is an attempt to make the small-scale dynamo maximally efficient, i.e. to maximize the kinematic growth rate $\gamma_K$ for a given numerical resolution.
This setup allows us to study a regime with strong non-linear feedback (saturated phase of the dynamo), but it does not allow us to address questions related
to the magnetic Prandtl number (implicitly set by the numerical dissipation terms, in general close to unity). \vspace{0.1cm}\\2. We use generalized open boundary conditions
which allow also for the presence of (mixed polarity) horizontal field in upflow regions. These are setups that explore a stronger coupling between the top
layers (including photosphere) and the bulk of the convection zone. Formally these boundary conditions lead to different magnetic energy loss rates at the bottom boundary. While they do not strongly affect the kinematic growth phase of an efficient dynamo with $\gamma_K \gg v_{z {\rm RMS}}/H_{\varrho}$, they do become relevant in the non-linear
saturation regime, which we primarily focus on here. \vspace{0.1cm}\\
In addition we compare models with different resolutions as well as domain sizes to
evaluate the robustness of results.
Our aim is to investigate with this setup a small-scale dynamo operating in a regime that is consistent with observational constraints
on the quiet Sun magnetic field strength, like those inferred by \citet{Danilovic:2010:zeeman_dynamo}. This regime is currently not
accessible with comprehensive solar MHD simulations that use only physical diffusion terms (DNS) or even operate in a low $P_M$ regime.
The remainder of the paper is organized as follows: In Section \ref{sect:setup} we describe in detail the numerical setup in terms of the equations
solved, the formulation of numerical diffusivites, boundary conditions and domain sizes used. In Section \ref{sect:results} we present
the results, subsections describe in detail the resolution dependence of results, the dependence on domain size and boundary conditions, and a
detailed analysis of the dynamo process based on transfer functions in spectral space. In Section \ref{sec:discussion} we discuss our main findings
in relation to observational constraints on quiet Sun magnetism, a detailed comparison with observations through forward modeling of spectral lines
is deferred to future publications. Concluding remarks are presented in Section \ref{sect:conclusion}.
\section{Numerical setup}
\label{sect:setup}
\subsection{Numerical scheme}
\label{sect:numdiff}
We use for our simulations the {\em MURaM} radiative MHD code
\citep{Voegler:etal:2005,Rempel:etal:2009}. This code uses a $4^{th}$
order accurate (in space and time) conservative, centered finite
difference scheme for discretization of the MHD equations, combined
with a short characteristics approach for radiative transfer. The code uses a tabulated
{\em OPAL} equation of state \citep{Rogers:opal:1996}. We solve the MHD equations in the following form:
\begin{eqnarray}
\frac{\partial \varrho}{\partial t}&=&-\nabla\cdot\left(\varrho \vec{v}\right)\label{eq:cont}\\
\frac{\partial \varrho\vec{v}}{\partial t}&=&-\nabla\cdot(\varrho\vec{v}\vec{v}) +\frac{f_{v_A}}{4\pi}\nabla\cdot\left(\vec{B}\vec{B}-\frac{1}{2}\vec{I}\,B^2\right)\nonumber\\
&&-\nabla P +\varrho\vec{g}\label{eq:mom}\\
\frac{\partial E_{\rm HD}}{\partial t}&=&-\nabla\cdot\left[\vec{v}\,(E_{\rm HD} + P)\right]+\varrho\vec{v}\cdot\vec{g}+\frac{\eta}{4\pi}(\nabla\times\vec{B})^2 \nonumber\\
&&+\vec{v}\cdot \frac{f_{v_A}}{4\pi}\nabla\cdot\left(\vec{B}\vec{B}-\frac{1}{2}\vec{I}\,B^2\right)+Q_{\rm rad}\label{eq:ener}\\
\frac{\partial \vec{B}}{\partial t}&=&\nabla\times\left(\vec{v}\times\vec{B}-\eta\nabla\times\vec{B}\right)\label{eq:ind}
\end{eqnarray}
Here $\varrho$, $P$, $\vec{v}$, and $\vec{B}$ denote mass density, pressure, velocity and magnetic field. For the
gravitational acceleration $\vec{g}$ we use a constant value of $-2.74\cdot 10^4 \mbox{cm}^2\mbox{s}^{-1}$ in the vertical direction (small local domains),
$\eta$ is an optional magnetic diffusivity. Furthermore $Q_{\rm rad}$ denotes the radiative heating term. We use
in the energy equation Eq. \ref{eq:ener} a treatment that is conservative for the quantity $E_{\rm HD}=E_{\rm int}+0.5\varrho v^2$
($E_{\rm int}$ denotes the internal energy). We separated out magnetic energy to avoid numerical problems in regions with
small values of the plasma beta $8\pi P/B^2$, which can be encountered above the photosphere in strong field regions.
In addition the Lorentz force pre-factor $f_{v_A}$ can be used to artificially limit the Alfv\'en velocity in those regions in order to
prevent severe time step constraints. We use here the same functional form as \citet{Rempel:etal:2009}
\begin{equation}
f_{v_A}=\frac{1}{\sqrt{1+\left(\frac{v_A}{v_{\rm max}}\right)^4}}\;,
\end{equation}
where $v_A=B/\sqrt{4\pi\varrho}$ and $v_{\rm max}$ denotes the maximum permissible Alfv\'en velocity.
While the latter two features were mostly implemented for sunspot simulations,
we limit also here in all simulations the maximum Alfv\'en velocity to $31.6\,\mbox{km s}^{-1}$ in order to prevent severe numerical time step constraints
that can arise from strong magnetic field near the top boundary (in particular in the two simulations that extend $1.5$~Mm above the photosphere).
This does not impact any of the results presented here for which mostly the sub-photospheric dynamics matter ($f_{v_A}=1$ until about a few $100$ km above the
photosphere), but it dramatically reduces the required
computing time by more than a factor of $10$ in some cases. This allows us to focus our study on higher resolution setups.
We do not consider explicit viscosity and also set $\eta$ to zero except for one control experiment. As a
consequence we require additional artificial diffusion terms in order to maintain numerical stability. These terms are
computed as described below.
We use here a modified version of the scheme first introduced
by \citet{Rempel:etal:2009}, which we explain here in detail. Our
approach is based on a slope-limited diffusion scheme that uses a
piecewise linear reconstruction of the discrete solution $u_i$ to compute
extrapolated values at cell interfaces:
\begin{eqnarray}
u_l&=&u_i+0.5\,\Delta u_i\\
u_r&=&u_{i+1}-0.5\,\Delta u_{i+1}\;.
\end{eqnarray}
Here $\Delta u_i$ denotes the reconstruction slope for the $i^{th}$ cell,
$u_l$ ($u_r$) are the interface values extrapolated from the cells on the
left (right). The reconstruction slopes $\Delta u_i$ are computed using the
monotonized central difference limiter, given by
\begin{eqnarray}
\Delta u_i&=&\mbox{minmod}\left[(u_{i+1}-u_{i-1})/2,\right.\\ \nonumber
&&\left. 2\,(u_{i+1}-u_i),2\,(u_i-u_{i-1})\right]\;.
\end{eqnarray}
Numerical diffusive fluxes at cell interfaces are computed from the
extrapolated values through the expression
\begin{equation}
f_{i+\frac{1}{2}}=-\frac{1}{2}\,c_{i+\frac{1}{2}}\,
\Phi_h(u_r-u_l,u_{i+1}-u_i)\cdot(u_r-u_l)\;.
\end{equation}
Here $c_{i+\frac{1}{2}}$ is a characteristic velocity at the cell interface,
the function $\Phi_h$ is given by
\begin{equation}
\Phi_h=\mbox{max}\left[0,1+h\left(\frac{u_r-u_l}{u_{i+1}-u_i}-1\right)\right]
\label{phi_lim}
\end{equation}
in regions with $(u_r-u_l)\cdot(u_{i+1}-u_{i})>0$, while $\Phi_h=0$ if
$(u_r-u_l)\cdot(u_{i+1}-u_{i})\leq 0$ (no anti-diffusion). Here $h$ is
a parameter that allows to control the (hyper-) diffusive character of the
scheme. A choice of $h=0$ reduces the diffusive flux to that of a standard
second order Lax-Friedrichs scheme. For $h>0$ the diffusivity is reduced
for smooth regions in which $\vert(u_r-u_l)/(u_{i+1}-u_{i})\vert<1$, while
the maximum diffusivity of $0.5\,c_{i+\frac{1}{2}}\,\Delta x$ is always kept
in regions with $\vert(u_r-u_l)/(u_{i+1}-u_{i})\vert=1$. For values of
$h>1$ the diffusive fluxes are switched off in regions with
$\vert(u_r-u_l)/(u_{i+1}-u_{i})\vert<1-1/h$, leading to a diffusivity that
is concentrated to monotonicity changes or features resolved by only a few
grid points. For the work presented here we use a choice of $h=2$.
\citet{Rempel:etal:2009} used a different functional form of
$\Phi_h=[(u_r-u_l)/(u_{i+1}-u_{i})]^2$ in regions with
$(u_r-u_l)\cdot(u_{i+1}-u_{i})>0$ that suppresses, but does not completely
disable diffusion for well-resolved features.
The above describe diffusion scheme is applied to the variables
$\{\log(\varrho), v_x, v_y, v_z, \varepsilon, B_x, B_y, B_z\}$, where
$\varepsilon=E_{\rm int}/\varrho$. In addition we make the assumption that
the diffusive mass flux also transports momentum and internal energy, i.e.,
we add to the momentum flux a term $\vec{f_{\varrho}}\vec{v}$ and to the
energy flux a term $\vec{f_{\varrho}}\varepsilon$, where $\vec{f_{\varrho}}$
denotes
the diffusive mass flux. This correction is identical with the assumption
that momentum and energy are transported by the total mass flux
$\varrho\vec{v}+\vec{f_{\varrho}}$. Since at the same time the induction
equation uses only the velocity $\vec{v}$ without a contribution from the
diffusive mass flux, the presence of mass diffusion mimics to some degree
ambipolar diffusion.
For enhanced stability we also implemented a switch, which limits the maximum
density contrast between neighboring grid cells to $10$. If the density
contrast exceeds that threshold we disable the piecewise linear reconstruction
and set the diffusivity to the maximum value allowed for by the CFL condition
to prevent a further increase.
We also added an additional optional $4^{th}$ hyper-diffusion term that scales
with the advection velocity and acts only in the vertical direction on the
quantities $\log(\varrho)$, $v_z$, and $\varepsilon$. This term allows to damp
some low level spurious oscillations on the grid scale that are too small to
cause monotonicity changes in the presence of a background gradient
(stratification) and go mostly undetected by the slope-limited diffusion scheme.
The numerical diffusion scheme is implemented in a dimensional split way to
ensure maximum stability and is applied to the solution in a separate filtering
step after a full time-step update of our $4^{th}$-order time integration
scheme. In the energy equation we account for artificial viscous and ohmic
heating.
Errors caused in ${\rm div}\vec{B}$ are controlled with the help of an iterative
hyperbolic divergence cleaning approach \citep{Dedner:etal:2002:divB}.
Estimating the effective diffusivity of our numerical scheme is not a trivial task. The numerical diffusivity
is in general highly intermittent and inhomogeneous as well as scale-dependent (see Section \ref{sec:transfer} for
further detail). Comparing results obtained at $4$~km grid spacing with simulations that use only a physical
magnetic diffusivity of $\eta=5\cdot 10^{9}\mbox{cm}^2\mbox{s}^{-1}$ (which is the minimum value
required for numerical stability in that case) we find an about $6$ times larger kinematic growth rate
with numerical diffusivity, indicating a significantly lower effective diffusivity.
\subsection{Domain size, boundary conditions, simulation setup}
We present numerical simulations in two domains:
$6.144\times 6.144\times 3.072\,\mbox{Mm}^3$ and $24.576\times
24.576\times 7.680\,\mbox{Mm}^3$. In the smaller domain the top boundary condition
is located about $700$~km above the average $\tau=1$ level, in the large domain about
$1.5$~Mm. This leads to depths of the convective part of about $2.3$~Mm and $6.2$~Mm,
respectively. In addition we performed also a series of simulations in a
$98.304\times 98.304\times 18.432\,\mbox{Mm}^3$ sized domain, but we will not
discuss them in great detail in this publication.
All simulations presented here use a setup with no vertical netflux, i.e.
$\langle B_z\rangle=0$. Since we use for most setups open boundary conditions
and allow for the transport of horizontal flux across the bottom boundary, the domain
averaged horizontal flux
can fluctuate, but stays on average close to zero. Our primary aim is to study the
contributions from a small-scale dynamo to quiet Sun magnetism separate from
potential contributions of a large-scale dynamo. We will discuss how both
dynamos could be coupled in Section \ref{sect:discuss_LSD-SSD}.
In the horizontal direction the domains are periodic, the top boundary is
semi-transparent (open for outflows, closed for inflows). For the magnetic
field we use two top boundary conditions: vertical magnetic field and a potential
field extrapolation.
Since the details of the formulation of the bottom boundary condition
have significant influence on the solutions in terms of the saturation field
strength reached, we explore here a total of $5$ different boundary
conditions. These boundary conditions are a balance between a self-contained
dynamo problem (best achieved with closed boundaries) and an attempt to
capture the deep convection zone (open boundaries).
In our numerical formulation
we have 2 ghost cells and the position of the domain boundary is between the first domain
and first ghost cell. For many variables we use boundary conditions which prescribe a
symmetric or anti-symmetric behavior across the boundary. If $v_1$ and $v_2$ are the
values in the first and second domain cell and $v_1^*$ and $v_2^*$ are the corresponding
quantities in the first and second ghost cell ($v_1^*$ is the ghost cell closest to the boundary),
a symmetric boundary implies $v_1^*=v_1$ and $v_2^*=v_2$, an anti-symmetric boundary implies
$v_1^*=-v_1$ and $v_2^*=-v_2$.
Most of our simulations use open hydrodynamic boundary conditions, which aim
to mimic the presence of a deep convection zone beneath the domain boundary. We use here
two different formulations for open and one formulation for a closed boundary condition,
which we describe first before we detail the magnetic boundary conditions:
\begin{enumerate}
\item[\it HD1:]
All three mass flux components are symmetric with respect to the boundary. The pressure
$P_{\rm BND}=P_{\rm gas}+B_z^2/(8\pi)$ is uniform and fixed at the boundary.
If $P_1$ and $P_2$ are the values of the gas pressure in the first and second domain cell, we assign
the ghost cell values as follows (linear extrapolation into ghost cells):
\begin{eqnarray}
P_1^*&=&1.5\,P_{\rm gas}-0.5\,\sqrt{P_1\,P_2}\\
P_2^*&=&2.5\,P_{\rm gas}-1.5\,\sqrt{P_1\,P_2}
\end{eqnarray}
The entropy is symmetric in downflow regions and is specified in upflow regions such that the
resulting radiative losses in the photosphere lead to a solar-like energy flux (within a few $\%$).
The corresponding values for density and internal energy follow from the equation of state.
In addition upflow velocities are capped at $1.5$ times the vertical RMS velocity at the boundary to
prevent extreme events.
\item[\it HD2:]
All three mass flux components are symmetric with respect to the boundary. We decompose the gas pressure
into mean pressure and fluctuation, $P=\bar{P}+P^{\prime}$. The mean pressure is extrapolated into the ghost cells such that its
value at the boundary is fixed, while the pressure fluctuations are damped in the ghost cells. This is achieved the following way:
\begin{eqnarray}
\bar{P}_1^*&=&\bar{P}_1 \cdot\frac{P_{\rm BND}}{\sqrt{\bar{P}_1 \bar{P}_2}}\\
\bar{P}_2^*&=&\bar{P}_1 \cdot\frac{P_{\rm BND}^2}{\bar{P}_1 \bar{P}_2}\\
P_1^{\prime *}&=&P_1^{\prime}\cdot C_{\rm dmp}\label{p1_dmp}\\
P_2^{\prime *}&=&P_1^{\prime}\cdot C_{\rm dmp}^2\label{p2_dmp}\
\end{eqnarray}
We use a value of $C_{\rm dmp}=0.95$. We used first a symmetric boundary condition for $P^{\prime}$, but found
problems with over-excited standing pressure waves in deeper domains.
The entropy is symmetric in downflow regions and is specified in upflow regions such that the
resulting radiative losses in the photosphere lead to a solar-like energy flux (within a few $\%$).
The corresponding values for density and internal energy follow from the equation of state.
\item[\it HD3:]
This is a closed boundary condition. The vertical mass flux is antisymmetric, the horizontal velocity components are
symmetric (closed for vertical mass flux and stress free for horizontal motions).
The gas pressure is extrapolated into the ghost cells as follows:
\begin{eqnarray}
P_1^*&=&P_1^2/P_2\\
P_2^*&=&P_1^3/P_2^2\,.
\end{eqnarray}
The entropy is symmetric across the boundary. We added a heating term in the lower
$10\%$ of the domain to replenish the energy radiated away in the photosphere.
\end{enumerate}
We used in our investigation initially the boundary {\it HD1}. Since the pressure at the boundary is fixed, this
boundary condition does not allow for pressure differences between up- and downflow regions, which are expected
for dynamical reasons. As a consequence this boundary condition underestimates the value of horizontal flow divergence
in upflow regions when compared to a deeper reference run. The boundary condition {\it HD2} puts less constraints on the
pressure at the boundary and does allow for systematic pressure differences between up- and downflow regions and improves
the properties of the flow at the boundary. While {\it HD1} accounts only for magnetic pressure from vertical field, {\it HD2}
incorporates the total magnetic pressure to the degree it is reflected in the gas pressure perturbation $P^{\prime}$ (we exclude
magnetic pressure contributions from the damping in Eqs. \ref{p1_dmp} and \ref{p2_dmp}).
The boundary {\it HD3} is used for control experiments using a closed domain.
In addition to the above described hydrodynamic boundary conditions we implement the following magnetic boundary conditions
in our experiments:
\begin{enumerate}
\item[\it OV:]
(Open boundary/vertical field)
We use {\it HD1}, the magnetic field is vertical at the boundary ($B_z$ symmetric, $B_x$ and $B_y$ antisymmetric).
\item[\it OSa:]
(Open boundary/symmetric field)
We use {\it HD1}, all three magnetic field components are symmetric. We impose an upper limit of
$600$~G for the horizontal RMS field strength in inflow regions and limit the maximum horizontal magnetic field
strength to $3$ times the RMS value. We set net horizontal magnetic flux in inflow regions to zero and
rescale the vertical magnetic field such that the horizontal and vertical RMS field strength are identical
in inflow regions (since we consider here only situations with $\langle B_z\rangle=0$ the rescaling
of $B_z$ does not affect the vertical net flux).
\item[\it OSb:]
(Open boundary/symmetric field)
We use {\it HD2}, all three magnetic field components are symmetric.
\item[\it OZ:]
(Open boundary/zero field)
We use {\it HD2}, similar to {\it OSb}, but we set $\vec{B}=0$ in inflow regions, i.e. $\vec{B}$ is antisymmetric
in inflow and symmetric in outflow regions.
\item[\it CH:]
(Closed boundary/horizontal field)
We use {\it HD3}, the magnetic field is horizontal at the bottom boundary ($B_z$ antisymmetric,
$B_x$ and $B_y$ symmetric).
\end{enumerate}
The boundary condition {\it OV} is similar to that used by \citet{Voegler:Schuessler:2007} and
we included one simulation with this boundary condition to better connect our results to
previous work. We started our investigation with {\it OSa}, but found that we had to implement
several corrections to the magnetic field to prevent runaway solutions when we also allow for
a horizontal magnetic to be present in inflow regions in combination with the hydrodynamic boundary
condition {\it HD1}. Most importantly, we limit the horizontal
RMS field strength to $600$~G (for the $6.144\times 6.144\times 3.072\,\mbox{Mm}^3$ domain),
which corresponds to a solution in which $B_{\rm RMS}$ increases with depth approximately at the
same rate as the equipartition field strength $B_{\rm eq}=\sqrt{4\pi \varrho}\,v_{\rm RMS}$
(see Section \ref{sec:subsurf} for further detail). Using the hydrodynamic boundary condition {\it HD2}
resolves most of these issues and a much simpler magnetic boundary conditions is sufficient ({\it OSb}).
The differences between boundary conditions {\it OSa,b} affect mostly the first
pressure scale height above the bottom boundary, boundary {\it OSb} performs overall
better when comparing simulations with different domain depths (see Section \ref{sec:subsurf}).
The boundary {\it OZ} is a control experiment making the very conservative (and likely
unrealistic) assumption that the deep convection zone is unmagnetized. We use boundary {\it CH}
as an additional control experiment to study a setup in which we have a complete recirculation
of mass and all magnetic induction effects are confined to the simulation domain.
As a general note we want to point out that none of the above boundary conditions is "perfect''.
Closed boundary are not a representation for the deep solar convection zone and open boundaries suffer
all from the same problem that the properties of quantities leaving the domain are well determined by
the solution, while the properties of quantities entering the domain have to be assumed, i.e. these
boundary conditions cannot be free from implicit or explicit assumptions. It is therefore crucial
to compare simulations with different boundary conditions as well as domain depths in order to quantify their
potential influence on solution properties.
\begin{table}
\begin{center}
\begin{tabular}{ c | c | c | c | c }
\hline
ID & Size [Mm$^3$] & Res [km] & Bot & Top \\ \hline\hline
V16 & $6.144\times 6.144\times 3.072$ & $16$ & {\it OV} & V\\ \hline\hline
O32a & $6.144\times 6.144\times 3.072$ & $32$ & {\it OSa} & V\\ \hline
O16a & $6.144\times 6.144\times 3.072$ & $16$ & {\it OSa} & V\\ \hline
O8a & $6.144\times 6.144\times 3.072$ & $8$ & {\it OSa} & V\\ \hline
O4a & $6.144\times 6.144\times 3.072$ & $4$ & {\it OSa} & V\\ \hline
O2a & $6.144\times 6.144\times 3.072$ & $2$ & {\it OSa} & V\\ \hline\hline
O32b & $6.144\times 6.144\times 3.072$ & $32$ & {\it OSb} & P\\ \hline
O16b & $6.144\times 6.144\times 3.072$ & $16$ & {\it OSb} & P\\ \hline
O8b & $6.144\times 6.144\times 3.072$ & $8$ & {\it OSb} & P\\ \hline\hline
Z32 & $6.144\times 6.144\times 3.072$ & $32$ & {\it OZ} & P\\ \hline
Z16 & $6.144\times 6.144\times 3.072$ & $16$ & {\it OZ} & P\\ \hline
Z8 & $6.144\times 6.144\times 3.072$ & $8$ & {\it OZ} & P\\ \hline\hline
C32 & $6.144\times 6.144\times 3.072$ & $32$ & {\it CH} & V\\ \hline
C16 & $6.144\times 6.144\times 3.072$ & $16$ & {\it CH} & V\\ \hline
C8 & $6.144\times 6.144\times 3.072$ & $8$ & {\it CH} & V\\ \hline
C8$\eta$ & $6.144\times 6.144\times 3.072$ & $8$ & {\it CH} & V\\ \hline \hline
O16bM & $24.576\times 24.576\times 7.68$ & $16$ & {\it OSb} & P\\ \hline
Z16M & $24.576\times 24.576\times 7.68$ & $16$ & {\it OZ} & P\\ \hline \hline
O32bSG & $98.304\times 98.304\times 18.432$ & $32$ & {\it OSb} & P\\ \hline \hline
\end{tabular}
\end{center}
\caption{Overview of numerical simulations discussed in this publication. See text
for further explanation.}
\label{tab:t1}
\end{table}
In Table \ref{tab:t1} we present all the simulations we discuss in this publication.
With the exception of O2a, C8, and C8$\eta$ all simulations were started from a thermally relaxed
non-magnetic convection simulation after addition of a $\sim 10^{-3}$~G random
field (pointing in the z-direction, random in the horizontal plane and uniform in the vertical direction).
The run O2a was restarted from a saturated snapshot of O4a and evolved for
an additional 5 minutes to further explore the resolution dependence. The simulation
C8 and C8$\eta$ were restarted from C16. C8$\eta$ uses a Laplacian diffusivity of
$\eta=10^{10}\,\mbox{cm}^2\,\mbox{s}^{-1}$ for the magnetic field instead of numerical
diffusivity (we kept numerical diffusivity for all other variables).
The simulation O32bSG was restarted from a sequence of lower resolution runs we do not list in Table \ref{tab:t1}.
As a consequence the spectral energy distribution is in this run likely biased toward larger scales. We use
this simulation here mostly to explore the connection toward deeper layers of the convection zone through
comparison of horizontally averaged mean quantities.
The column "Bot" refers to the boundary condition used at the bottom boundary,
the column "Top" to the magnetic field boundary condition used at the top. Here
"V" and "P" refer to vertical magnetic field and potential field extrapolation.
The hydrodynamical boundary condition at the top boundary is in all cases open for upflows (i.e. upward
directed shocks can leave the domain) and closed for downflows. In the simulations
O16bM and Z16M the top boundary is about $1.5$~Mm above the photosphere, while it is about
$700$~km in all other simulations.
\begin{figure*}
\centering
\resizebox{0.85\hsize}{!}{\includegraphics{fig01}}
\caption{Comparison of kinematic growth phase (panels a-c) and saturated
phase (panels d-f) of the simulation run O4a. Shown are the bolometric
intensity (panels a, d), the vertical velocity at $\tau=1$ (panels b, e),
and the magnetogram at $\tau=1$ (panels c, f). Two animations of this figure
for the kinematic and saturated phase are provided with the online
material. During the kinematic growth phase we clipped the displayed values
of $B_z$ in the movie at $\pm 5 \langle \vert B_z\vert\rangle$ to follow the
fast growth of the field; in the saturated phase we display values in the range
$\pm 400$~G similar to panel f.
}
\label{fig:IVzBz_kin_sat}
\end{figure*}
\subsection{Scope of the simulations presented here}
Are the simulations we present here small-scale dynamos? This question arises because
of two aspects of our setup: open boundary conditions and the use of (unphysical)
numerical diffusivities. The open boundaries we use allow for a magnetic energy flux across domain
boundaries,
which implies that the maintenance of the magnetic field is not restricted to processes
within the simulation domain. Although, as we show later, the Poynting flux transports
significantly more energy out of the domain than is returning back in inflow regions.
We have conducted experiments that use closed boundary conditions and only a physical
Laplacian diffusivity for the magnetic field (run C8$\eta$) and we confirmed that we have a small
scale dynamo operating under these conditions. In addition a comparison of the spectral
energy transfers presented in Section \ref{sec:transfer} does not reveal any significant differences
(apart from the saturation field strength reached) between this reference simulation and a simulation
solely based on numerical diffusivities.
While our numerical experiments should be more carefully labeled as large eddy simulations of
photospheric magneto-convection with zero imposed magnetic flux, we did not find any indication
that they are not small-scale dynamos.
Since we apply the same numerical dissipation scheme to all MHD variables, the resulting
"numerical magnetic Prandtl number'' is close to 1 in all our experiments. We do not address here
the role of the magnetic Prandtl number for the small-scale dynamo process.
\section{Results}
\label{sect:results}
In the following subsections we analyze our simulations by presenting quantities in the photosphere
on constant $\tau$ levels. Since we use here only simulations computed with gray radiative transfer
these layers refer to a $\tau$-scale computed with mean opacities. Further $\tau$ levels always refer
to warped $\tau=\mbox{const.}$ surfaces and not the constant geometric height surface with the corresponding
average $\tau$ value.
We further discuss in detail (mostly) photospheric power spectra and probability distribution functions
of magnetic field. On the one hand we use these quantities to simply compare different simulations, on the
other hand they have a strong connection to results from observational studies of quiet Sun magnetism.
We refer the reader to Sections \ref{sec:discuss_PDF} and \ref{sec:discuss_power} for a summary and discussion
of their importance.
\begin{figure*}[ht!]
\centering
\resizebox{0.8\hsize}{!}{\includegraphics{fig02}}
\caption{a) Photospheric ($\tau=1$) power spectra for magnetic energy (red)
and kinetic energy (blue) in simulation O4a. The dashed lines correspond to the kinematic
growth phase during times when Lorentz-force feedback was negligible. We
scaled up the magnetic energy spectrum in order to show clearly in the same figure.
The solid red lines correspond bottom to top
to snapshots with $20, 40, 60$, and $80$~G vertical mean field strength
at $\tau=1$. For increasing field strength the peak of magnetic power moves
toward larger scales. The solid blue line shows the kinetic energy spectrum
for the $80$~G case. The dotted line in panel a) indicates a Kolmogorov slope
of $-5/3$. b) Probability distribution functions for
$\vert B\vert$ at $\tau=1$ for vertical mean field strength from $20$ to $80$~G.
}
\label{fig:pwr_pdf_sat}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig03}}
\caption{a) Resolution dependence of magnetic (red) and kinetic (blue) energy
power spectra comparing the simulations O32a-O2a. All simulations reached about
$80$~G vertical mean field strength at $\tau=1$, the grid spacing was varied from $32$ to $2$~km.
The dotted line in panel a) indicates a Kolmogorov slope of $-5/3$.
b) Ratio of magnetic to kinetic energy as function
of resolution. For grid spacings smaller than $16$~km the magnetic energy is
in super-equipartition on scales smaller than about $100$~km. For the highest
resolution cases ($4$ and $2$~km) we see some indication that the ratio
might reach asymptotically a value around $2$. c) Resolution dependence
of the probability distribution function for $\vert B_x\vert$ and $\vert B_z\vert$.
}
\label{fig:pwr_pdf_res}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.75\hsize}{!}{\includegraphics{fig04}}
\caption{a) Normalized integrated magnetic energy spectra from Figure \ref{fig:pwr_pdf_res}.
For the highest resolution case (grid spacing of $2$~km) $50\%$ of the magnetic
energy found at the $\tau=1$ surface is present on scales smaller than $100$~km.
b) Normalized integrated magnetic energy distribution functions. $50\%$ of the
magnetic energy on the $\tau=1$ surface is found in field with less than $500$~G
strength. Solid lines correspond to simulations O32a-O2a with
$\langle\vert B_z\vert\rangle=80$~G, dashed (dotted) lines correspond to simulations
O16b and O16bM with $\langle\vert B_z\vert\rangle=60$~G.
}
\label{fig:energy_dist}
\end{figure*}
\subsection{Kinematic to saturated phase}
\label{sec:kin_sat}
We start our discussion of results with numerical simulations using the
$6.144\times 6.144\times 3.072\,\mbox{Mm}^3$ domain and the boundary condition {\it OSa}.
We limit the horizontal RMS field strength in inflow regions to $600$~G, which corresponds
approximately to a solution in which the RMS field strength increases with depth as the
same rate as the equipartition field strength. As we will discuss in Section \ref{sec:subsurf}, these
solutions are close to an upper limit for the quiet Sun field strength.
We use the small domain to explore the resolution dependence of the results and repeated the
same experiment with grid spacings of $32$, $16$, $8$, and $4$~km. All simulations were started
from a thermally relaxed $B=0$ G convection simulation to which we added a $10^{-3}$~G seed field
(pointing in the z-direction, random in the horizontal plane and uniform in the vertical direction). In
addition we present a simulation with $2$~km grid spacing, which was restarted from the $4$~km case.
Figure \ref{fig:IVzBz_kin_sat} presents for the simulation with $4$~km grid spacing (O4a) two snapshots, one
during the early growth phase (panels a-c) at a time when
$\langle\vert B_z\vert\rangle(\tau=1) = 8.4$~G and one during during a later phase (panels d-f) when
$\langle\vert B_z\vert\rangle(\tau=1) = 86$~G. The panels a) and d) show the intensity
for a vertical ray, panels b) and e) the vertical velocity at $\tau=1$, and the panels
c) and f) the vertical magnetic field at $\tau=1$. While the snapshot with
$\langle\vert B_z\vert\rangle(\tau=1) = 8.4$~G shows magnetic field organized on scales close
to the grid spacing of the simulation, the snapshot with
$\langle\vert B_z\vert\rangle(\tau=1) = 86$~G shows magnetic field organized more on the
scale of granular downflows with a mostly sheet-like appearance. Several downflow lanes show
sheets with opposite polarity nearby. Panel d) shows also several brightness enhancements
associated with strong field on scales of $100$~km and less. We provide also 2 animations of
Figure \ref{fig:IVzBz_kin_sat} in the online material (one for the kinematic and one for the saturated
phase). These animation show the same quantities as presented in Figure \ref{fig:IVzBz_kin_sat}.
Figure \ref{fig:pwr_pdf_sat}a) shows kinetic and magnetic energy spectra (at $\tau=1$), which were computed for
the $4$~km grid spacing case. As the solution is evolving from the kinematic growth phase to
the saturated regime, the peak of the magnetic energy spectrum is moving toward larger
scales. At the same time kinetic energy becomes suppressed by about a factor of $2$ on
scales smaller than $100$~km as a consequence of Lorentz-force feedback. We will discuss the saturation
process further in Section \ref{sec:transfer}. For the solution
reaching a vertical mean field strength of $80$~G in the photosphere, the magnetic energy is in
super-equipartition by about a factor of $2$ on scales smaller than $100$~km.
For the case with $4$~km grid spacing presented here the e-folding time scale for magnetic energy in the photosphere
is about $50$ sec during the kinematic growth phase. The growth rate is strongly resolution dependent,
we find time scales of $120$, $350$, and $850$ sec for grid spacings of $8$, $16$, and $32$~km. This
leads on average to a resolution dependence of the kinematic growth rate $\gamma_K\sim \Delta x^{-1.36}$.
This resolution dependence is significantly steeper compared to simple estimates that yield for
$P_M \ll 1$ $\gamma_K\sim Re_M^{1/2}$, where $Re_M=v L/\eta$, $v$ and $L$ are typical velocity and length scales of
the problem (for $P_M \gg 1$ $Re_M$ has to be replaced by $Re=v L/\nu$) .
Since we do not have explicit viscosity and magnetic diffusivity we further assume that $Re_M$ is linked to the scale separation allowed for
by the numerical simulation. Assuming a 5/3 Kolmogorov spectrum, we would expect $Re_M\sim \Delta x^{-4/3}$,
leading to $\gamma_K \sim \Delta x^{-2/3}$. The growth rate is more consistent with a $\gamma_K\sim Re_M$
dependence, which was also found by \citet{Pietarila-Graham:etal:2010:SSD}. Since we do not use here
any explicit numerical diffusivity a detailed interpretation is difficult.
Figure \ref{fig:pwr_pdf_sat}b) shows the corresponding probability distribution functions (PDFs) for
$\vert B\vert$ at $\tau=1$. The PDF has a peak at around $30$~G and a nearly exponential
drop for stronger field. For the snapshots with $\langle\vert B_z\vert\rangle>60$~G strong kG
field concentrations cause a bulge for $\vert B\vert>1000$~G. In snapshots with
$\langle\vert B_z\vert\rangle=80$~G we find at $\tau=1$ field concentrations with more than
$2$~kG.
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig05}}
\caption{Comparison of a) power spectra and b), c) probability
distribution functions for $\vert B_x\vert$ and $\vert B_z\vert$ for
3 different height levels in simulation O4a. Solid lines show quantities in $1$~Mm
depth, dotted lines on the $\tau=1$ and dashed lines on the $\tau=0.01$
levels. Panel c) compares normalized PDFs, which allow for a direct
comparison of the shape regardless of field strength. In panel a) the
black dotted, dashed and dashed-dotted lines indicate $-1.4$, $-2.2$ and $-2.7$
power law slopes, respectively.
}
\label{fig:pwr_pdf_3lev}
\end{figure*}
\subsection{Resolution dependence}
Figure \ref{fig:pwr_pdf_res} compares simulations with 5 different grid spacings ranging from $32$ to $2$~km.
The simulation with $2$~km grid spacing was restarted from the saturated $4$~km simulation
and evolved for an additional $5$ minutes. Power spectra and PDFs were averaged over snapshots with
values of $\langle\vert B_z\vert\rangle(\tau=1)$ between $75$ and $85$~G. Panel a) shows kinetic
(blue) and magnetic (red) energy spectra for the simulations O32a-O2a. Increasing the resolution
leads to a convergence of the power spectra on the large scales while smaller scales are added.
The simulations with $32$ and $16$~km grid spacing show excess power on large scales, since the same amount of magnetic energy
is distributed over less wave numbers. The simulations with $8$ to $2$~km grid spacing do not show a significant
difference indicating that a grid spacing of $8$~km or smaller is required to properly represent the energy
distribution on larger scales in the photosphere. The dotted line in panel a) indicates a Kolmogorov slope
of $-5/3$ as a rough reference. Over the scale-range explored we don't see a clear
indication of a power law for the magnetic energy, there is some indication of power laws for the kinetic energy
(see also Figure \ref{fig:pwr_pdf_3lev}). Panel b) shows the ratio of magnetic to kinetic energy
as function of scale. For grid spacings smaller than $16$~km we find a super-equipartition regime on scales
smaller than $100$~km and see some indication that the ratio of magnetic to kinetic energy asymptotically
reaches a factor of about $2-2.5$. Panel c) shows the PDF for $\vert B_x\vert$ and $\vert B_z\vert$ at $\tau=1$.
We do not see a systematic dependence on resolution, differences for stronger field are mostly realization noise.
For field with less than $500$ G
strength the PDFs for $\vert B_x \vert$ and $\vert B_z \vert$ are essentially identical. Note that we show
here the PDFs for the absolute values of the field components since the simulations do not have any net magnetic
flux, leading to symmetric PDFs with respect to $B=0$.
Figure \ref{fig:energy_dist} shows normalized integrated magnetic energy spectra and distribution functions for
the simulations O32a-O2a (solid lines) . The quantities shown are defined as
\begin{eqnarray}
S_k&=&\frac{\int_0^k E_M(k)dk}{\int_0^{k_{max}} E_M(k)dk}\\
S_B&=&\frac{\int_0^B PDF(B) B^2dB}{\int_0^{B_{max}} PDF(B) B^2dB}
\end{eqnarray}
The quantity $S_k$ shows resolution dependence as expected from Figure \ref{fig:pwr_pdf_res}a). In the highest
resolution case about $50\%$ of the magnetic energy in the photosphere on the $\tau=1$ level is found
on scales smaller than about $100$~km. Properly resolving the spectral magnetic energy distribution in the
photosphere requires grid spacings of $8$~km or smaller. In contrast to this the quantity $S_B$ shows
only little resolution dependence. In all cases $50\%$ of the magnetic energy is found in regions with
$\vert B \vert$ of less than $500$~G. Kilo-Gauss field contributes about $10\%$ to the total magnetic energy.
For comparison we also show these quantities for the simulation
O16b (dotted) and O16bM (dashed). Both simulations have $25\%$ less unsigned flux in the photosphere.
The differences in $S_k$ are very minor. $S_B$ is shifted for the simulation O16b to the left toward weaker field.
In contrast to that the simulation O16bM (larger domain) is very similar to the $80$~G cases and has an even larger
contribution from kG field. We will discuss kG field concentrations further in Section \ref{sec:kG}.
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig06}}
\caption{Results from run O2a with a grid spacing of $2$~km. a) Intensity, b) $B_z(\tau=1)$, c) Intensity
patterns magnified for lower left corner of domain. Contour lines indicate regions with $\vert B\vert>1$~kG,
d) $B_z(\tau=1)$ magnified for lower left corner.
}
\label{fig:IBz_2km}
\end{figure*}
\subsection{Height dependence}
\label{sec:height}
Figure \ref{fig:pwr_pdf_3lev} presents how power spectra and probability distribution functions for the
magnetic field strength depend on the vertical position in the simulation domain (based on run O4a). Here we focus on
three levels that are indicated by different line styles (solid: $-1$~Mm depth, dotted: $\tau=1$,
dashed: $\tau=0.01$). Panel a) presents the kinetic and magnetic energy spectra for the three levels. They
show the same overall behavior with a super-equipartition regime toward small scales. While the super-equipartition
regime is reached at $\tau=1$ for scales smaller than $100$~km, it extends to $500$~km at $\tau=0.01$ since the
kinetic energy drops more rapidly than magnetic energy above the photosphere (short density scale height). The
super-equipartition regime extends also to moderately larger scales beneath the photosphere, since the overall
scale of convective motions increases with the increasing scale height, although the difference is small
between $\tau=1$ and $1$~Mm deeper. The black dotted, dashed and dashed-dotted lines indicate power law slopes
of $-1.4$, $-2.2$ and $-2.7$, respectively. While we do not find a clear power law for $E_M$ at any height level there is some
indication of a power law for $E_k$ on scales smaller than downflow lanes (few $100$~km). At $1$~Mm depth
and $\tau=0.01$ we find slopes of about $-1.4$, while the $\tau=1$ level is with $-2.2$ substantially steeper.
Extrapolating the approximate slopes to smaller scales implies that the spectra of $E_k$ on the $\tau=1$ and $\tau=0.01$ levels
will cross unless there is a change of slope toward smaller scales in either layer, which is more
likely. For all three layers we also find steeper slopes on scales larger than a few $100$~km. At $\tau=0.01$ we find
with $-2.7$ the steepest slope.
Comparing the PDFs for $\vert B_x\vert$ and $\vert B_z\vert$ (panel b) shows
systematic differences in the overall shape between the distribution for vertical field at $\tau=1$ and
$\tau=0.01$ and the rest. This difference is most obvious if we consider PDFs for the normalized magnetic field
components (panel c). Here the PDFs for $\vert B_x\vert$ at all three height levels and the PDF for $\vert B_z\vert$
in $1$~Mm depth are essentially identical, while the PDFs for $\vert B_z\vert$ at $\tau=1$ and $\tau=0.01$ show a
much more extended tail toward stronger field. This is a strong indication for the presence of a distinct
amplification process operating only on vertical field in the photosphere, while the distribution of $B_x$ in all
three levels and $B_z$ beneath the photosphere is of mostly turbulent origin. While it is non-trivial to separate
out the additional amplification process in the photosphere, we conjecture that it is related to a process along
the lines of "convective intensification" \citep{Schuessler:1990:IAUS}, which is a combination of flux-expulsion,
back-reaction of magnetic field leading to partial evacuation, enhanced radiative cooling and related downflows.
These processes go beyond the idealized picture of "convective collapse" \citep{Spruit:1979:convcollapse}.
\begin{figure}
\centering
\resizebox{0.85\hsize}{!}{\includegraphics{fig07}}
\caption{Filling factor of kG field at $\tau=1$. Solid lines are computed from simulations
O32a-O4a, the dotted (dashed) lines corresponds to O16b (O16bM). Field concentrations with more than
$1$~kG strength appear independent of resolution once $\langle\vert B_z\vert\rangle$ exceeds
about $30-40$~G. For $80$~G the filling factor reaches about $0.45\%$. This fraction is not
systematically dependent on resolution, but does increase with domain size (dashed line).
}
\label{fig:ff_kG}
\end{figure}
\begin{figure*}
\centering
\resizebox{\hsize}{!}{\includegraphics{fig08}}
\caption{a) Intensity and b) $B_z$ at $\tau=1$ taken from the simulation O16bM. The lower left corner
($x,y<6.144$~Mm) shows $I$ and $B_z$ from the simulation O16b for comparison. The larger
domain shows indications of a network structure significantly larger than
granulation that cannot form in the small domain. An animation of this figure displaying only O16bM is
provided with the online material.
}
\label{fig:IBz_gran_meso}
\end{figure*}
\subsection{kG flux concentrations}
\label{sec:kG}
Figure \ref{fig:IVzBz_kin_sat} shows the presence of several kG field concentrations in the photosphere
that lead to brightness enhancements in the downflow lanes. Here we analyze in more detail how
these flux concentrations depend on the overall field strength as well as domain size.
Figure \ref{fig:IBz_2km} shows examples of kG flux concentrations in the highest resolution simulation O2a. The panels a)
and b) show $I$ and $B_z(\tau=1)$ for the full horizontal domain extent, while panels c) and d) show a magnification
of the lower left corner of the domain. In panel c) contour lines highlight regions with $\vert B\vert>1$~kG. Many kG
field concentrations exist on scales smaller than $100$~km down to scales comparable to the grid resolution. Strong magnetic field
is typically organized in sheets, often with alternating polarities. kG flux concentrations are small knots along these sheets in which
the field strength is increased temporarily due to dynamical effects. Some longer lived flux concentrations may be found in
granular downflow vertices. We do not present here a detailed analysis of the temporal evolution of kG field concentrations, but refer
the interested reader to the animations of Figure \ref{fig:IVzBz_kin_sat} provided with the online material.
Figure \ref{fig:ff_kG} presents how the filling factor of kG field at $\tau=1$ depends on the
vertical mean field strength, resolution as well as domain depth. To this end we computed for the simulations
O32a-O4a the filling factor of regions with $\vert B\vert > 1$~kG at $\tau=1$, while these
simulations were evolving from the kinematic phase into the saturated phase. The data points presented in Figure
\ref{fig:ff_kG} result from binning snapshots in $\pm 5$~G intervals. For all 4 simulations
we find regardless of the resolution that kG flux concentrations appear when the vertical mean field strength
exceeds about $30-40$~G at $\tau=1$. For $\langle\vert B_z\vert\rangle=80$~G around $0.45\%$ of the area
is occupied by kG flux concentrations. While the results show some scatter due to realization noise in the simulation
domain with small horizontal extent, there is no indication of a systematic resolution dependence of this result.
For comparison we also show the simulations O16b (dotted green) and O16bM (dashed green). While O16b
is comparable to O16a, for the same field strength O16bM shows about twice the filling factor. This difference
is related to the formation of a larger scale magnetic network structure we discuss further in Section \ref{sec:meso}.
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig09}}
\caption{a) Photospheric ($\tau=1$) power spectra for magnetic energy (red)
and kinetic energy (blue) in simulations O16b and O16bM (same resolution, but
different domain size). Solid (dashed) lines refer to OS16b (OS16bM).
b) Probability distribution functions for $\vert B_x\vert$ and
$\vert B_z\vert$at $\tau=1$. Increasing the domain size increases the
magnetic power at larger scales and leads to a significantly higher fraction
of field with more than $1$~kG field strength.
c) Distribution of magnetic energy. In the larger domain the quantity
$PDF(\vert B_z\vert)\, B_z^2$ shows a plateau toward $1700$~G.
}
\label{fig:pwr_pdf_gran_meso}
\end{figure*}
\subsection{From granular to meso-granular scales}
\label{sec:meso}
Figure \ref{fig:IBz_gran_meso} presents a comparison of snapshots from simulations O16b and O16bM.
Both simulations have a grid spacing of $16$~km and differ only in domain size. Presented are intensity and the $\tau=1$ magnetograms.
In the larger domain magnetic field becomes organized on a scale larger then granulation. We find more pronounced kG flux
concentrations that show up mostly as bright features in the intensity image. We do not find the spontaneous formation of larger
pore-like field concentrations in O16bM. We provide also an animation of Figure \ref{fig:IBz_gran_meso} in the online material.
The animation shows only the simulation O16bM, but otherwise the same quantities as presented in Figure \ref{fig:IBz_gran_meso}.
Figure \ref{fig:pwr_pdf_gran_meso} compares the magnetic and kinetic energy spectra as well as probability
distribution function for the simulations O16b and O16bM. We compare here time averages of snapshots with
values of $\langle\vert B_z\vert\rangle$ from $55$ to $65$~G. In the larger domain the magnetic power spectrum extends
toward larger scales, while the kinetic energy spectrum continues to fall off. We see an increase of magnetic power
on scales larger than about $300$~km, while there is no significant change on smaller scales.
The PDF for $\vert B_x\vert$ remains mostly unchanged, while the PDF for $\vert B_z\vert$ shows a significant increase toward
kG fields in O16bM. We find that the filling factor of kG field is in O16bM with $0.4\%$ more than
twice as large as in O16b ($0.16\%$). Computing the the distribution of energy from vertical magnetic field, $PDF(\vert B_z\vert)\, B_z^2$,
leads to a plateau toward $1700$~G in O16bM that is not present in the smaller domain. We find the plateau only in the contribution from $B_z$.
In terms of the fraction of the total magnetic energy that is present in kG field at $\tau=1$ we find the values $9\%$ (O16b) and $23\%$ (O16bM).
In addition we studied also similar setups in larger domain ($98.304\times 98.304\times 18.432\,\mbox{Mm}^3$) at lower resolution and found
that the trends indicated here (mostly flat magnetic energy spectrum on scales larger than granulation, increasing fraction of kG field) continue.
In the simulation O32bSG we find a filling factor of $1.1\%$ for kG field, which contribute around $50\%$ to the magnetic energy at $\tau=1$.
The differences we see between the small and large domain arise from the presence of longer-lived,
larger-scale convection flows present in the larger domain, which lead to the formation longer-lived flux
concentrations. The trend of an increasing fraction of kG field with domain size indicates that perhaps even
a super-granular network structure could be maintained by a small-scale dynamo, provided the domain
is large enough. While the fraction of kG field increases with domain size, we did not find any indication
for a secondary peak in the probability distribution functions of the magnetic field (including O32bSG).
\begin{figure*}[ht!]
\centering
\resizebox{0.85\hsize}{!}{\includegraphics{fig10}}
\caption{Comparison of simulations in domains of different depth and
horizontal extent. Both simulations use the boundary {\it OSb} and
have a grid spacing of $16$~km (simulations O16b and O16bM).
Panels a) and b) show $B_{\rm eq}$ and $B_{\rm RMS}$ as
function of depth. Solid lines correspond to horizontal averages,
while dotted (dashed) lines are averages over
up- (down-) flow regions. Panel a) shows results for the simulation O16bM
($\sim 6.2$~Mm deep domain), panel b) the same quantities for simulation O16b
($\sim 2.3$~Mm deep domain). Both simulations are consistent with each other in terms
of the RMS field strength. Black (dotted) dashed lines
indicate in panel b) the profiles of $B_{\rm eq}$ and $B_{\rm RMS}$ from
panel a) for comparison.
Panels c) and d) show the magnetic field structure at and above the
photosphere. Red (green) lines indicate the RMS (mean) field strengths, while blue
lines show the equipartition field strength. The meaning of line styles is different from
panels a) and b): dashed (dotted) lines refer to the corresponding averages of vertical
(horizontal) field components.
}
\label{fig:BeqBrms}
\end{figure*}
\subsection{Subsurface field structure, role of boundary conditions}
\label{sec:subsurf}
Figure \ref{fig:BeqBrms} presents for the simulations O16b and O16bM the vertical profiles of the equipartition field strength $B_{\rm eq}$ (blue) and $B_{\rm RMS}$ (red). The equipartition field strength $B_{\rm eq}=\sqrt{4\pi \varrho}\,v_{\rm RMS}$ is a measure for energy available in convective motions.
Different line styles correspond to upflow regions (dotted),
downflow regions (dashed) and the averages over the whole domain (solid). In Panel b) black lines indicate the
profiles from Panel a) for better comparison ($B_{\rm RMS}$ is dashed, $B_{\rm eq}$ dotted).
The simulations O16b and O16bM show a lot of similarity, in terms of the total $B_{\rm RMS}$ both simulations match
each other in the part of the domain where they overlap. Differences are present when we compare $B_{\rm RMS}$ in
up and downflow regions in separation. Also $B_{\rm eq}$ is lower throughout most of the shallow domain, except for the
near photospheric layers. The fact that the average magnetic properties in the shallow domain stay very
close to those in the deep domain is an indication that the bottom boundary condition {\it OSb} does perform fairly well
in "mimicking" a deep convection zone and leads to consistent results independent from the location of the
bottom boundary. The panels c) and d) give a more
detailed view of the magnetic field structure in and above the photosphere. For both simulations we find a secondary peak of the horizontal
mean field strength about $450$~km above $\tau=1$ (green dotted line). In the larger domain, panel c),
this translates into a secondary peak of the total field
strength, while the RMS field strength continues to drop monotonically above the photosphere. We discuss the inclination
of magnetic field above the photosphere in more detail in Figure \ref{fig:Inclination}.
Figure \ref{fig:BeqBrms_BND} presents a comparison of simulations with different bottom boundary conditions. Panel a) compares
the boundaries {\it OSb} and {\it OZ} in the deep (runs O16bM and Z16M) and shallow domain (runs O16b and Z16). Changing from
{\it OSb} to {\it OZ} drops the field strength in the bulk of the convection zone by about a factor of $2$. The difference
between both boundary conditions does not depend on the domain depth within the range explored here. The boundary condition
{\it OZ} is very conservative in the sense that it assumes that the deeper convection zone is field free, which is unlikely
to be the case. But even with this assumption a still considerable amount of magnetic field is maintained within the
computational domain, although the mean vertical magnetic field at $\tau=1$ levels out at about $30$~G. This value is not
strongly dependent on resolution as long as a critical value is passed (i.e. the kinematic dynamo growth rate has to be
sufficiently large compared to the flux loss rate $\sim v_{z {\rm RMS}}/H_{\varrho}$).
We repeated this experiment with the resolutions from $32$ to $8$km. While we find for $32$~km resolution only a vertical magnetic field
strength of $20$~G at $\tau=1$, the simulations with $16$ and $8$~km grid spacing reach both values around $30$~G.
The simulations with the boundary condition {\it OSb} reach $>0.5 B_{\rm eq}$ in the deeper parts of the domain. These solutions are not
far from an upper bound for the field strength in which $B_{\rm RMS}$ and $B_{\rm eq}$ increase with depth at the same rate. To better illustrate
this asymptotic limit we show also the results from O32bSG, which uses a $18$~Mm deep domain. The dotted lines indicate
$B_{\rm RMS}$ profiles for O16bM and O32bSG that are rescaled by a factor of $1.5$ to illustrate this asymptotic limit. Substantially stronger field
would require a $B_{\rm RMS}$ increasing with depth faster than $B_{\rm eq}$ and exceeding equipartition in only a few Mm of depth. This
asymptotic limit corresponds to a solution with $\langle\vert B_z\vert\rangle=85$~G, $\langle B\rangle=160$~G, and $B_{\rm RMS}=275$~G at $\tau=1$.
\begin{figure*}[ht!]
\centering
\resizebox{0.85\hsize}{!}{\includegraphics{fig11}}
\caption{Comparison of different bottom boundary conditions for a fixed grid spacing of $16$~km.
Solid (dashed) lines correspond to equipartition (RMS) field strength, the color indicates
simulations with different bottom boundary conditions. In panel a) we compare
boundary {\it OSb}: O16bM (blue) and O16b (green) with boundary {\it OZ}: Z16M (red) and Z16 (orange).
The black line show the results from O32bSG, for which the bottom boundary is in about $18$~Mm depth.
In addition the dotted lines show scaled subsurface $B_{\rm RMS}$ profiles for O16bM and O32bSG to indicate a
solution we consider the upper limit ($B_{\rm RMS}$ increases with depth at the same rate as $B_{\rm eq}$). This
solution corresponds to about $\langle\vert B_z\vert\rangle=85$~G at $\tau=1$ (based on O16bM).
Panel b) presents additional experiments in the shallow domain.
Here we present simulations with the boundaries {\it OSa} (O16a, black), {\it OSb} (O16b, green),
{\it OZ} (Z16, orange), {\it OV} (V16, red) and {\it CH} (C16, blue).
}
\label{fig:BeqBrms_BND}
\end{figure*}
\begin{figure*}[ht!]
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig12}}
\caption{Resolution dependence of $B_{\rm RMS}$ (dashed) and $B_{\rm eq}$ (solid) for simulations with the boundary
a): {\it OSb}, runs O32b-O8b; b) {\it CH}, runs C32-C8; and c) {\it OZ}, runs Z32-Z8.
}
\label{fig:BeqBrms_Res}
\end{figure*}
Figure \ref{fig:BeqBrms_BND}b) compares results of all 5 boundary conditions considered here for the shallow domain. We do not
find a significant difference between zero field in inflows {\it OZ} (orange) and vertical field everywhere at the bottom boundary {\it OV} (red). Due to the strong horizontal divergence in upflows, vertical magnetic field present at the
bottom boundary condition becomes quickly expelled from upflows. Solutions with stronger magnetic field require the
presence of horizontal field in upflow regions, which is less affected by horizontally divergent flows. The solutions with
the boundary conditions {\it OSa,b} (black, green) are very similar to a solution computed with a closed bottom boundary condition {\it CH} (blue). The saturation field strength for the latter is fully determined by processes within the computational domain, while the former exchange magnetic field through the bottom boundary.
Figure \ref{fig:BeqBrms_Res} analyses further the resolution dependence of the saturation field strength for the simulations using the boundaries
{\it OSb}, {\it CH}, and {\it OZ}. In all three cases we find a similar trend of increasing $B_{\rm RMS}$ with resolution. While the saturation field
strength is not yet fully converged, it cannot grow much further in the simulations with the boundary conditions {\it OSb} and {\it CH}
without creating a super-equipartition regime near the bottom of the domain.
\subsection{Subsurface Poynting flux and energy conversion rates}
\label{sec:energetics}
Figure \ref{fig:Poynting}a) shows the Poynting flux for the
simulations O16bM and Z16bM. The flux is normalized by the solar photospheric energy flux of
$F_{\odot}=6.3\cdot 10^{10}\,\mbox{erg}\,\mbox{cm}^{-2}\,\mbox{s}^{-1}$. Although we have magnetic energy entering the
domain in upflow regions in simulation O16bM (solid lines) using the boundary condition {\it OSb}, the magnetic energy leaving
the domain in downflow regions over-compensates this contribution by more than a factor of $6$. In the simulation Z16M (dashed lines)
the Poynting flux is zero in upflow regions at the bottom boundary by construction (boundary condition {\it OZ}). However,
about $1-2$~Mm above the bottom boundary mixing between up and downflows provided enough field in upflow regions to have also
here an upward directed Poynting flux. The relative contributions from up and downflows in case Z16M are almost
identical with case O16bM if we stay $1-2$~Mm away from the bottom boundary condition. Panel b) shows the resulting magnetic energy loss rates for both simulations, which are defined as (with the Poynting flux $P(z)$)
\begin{figure*}[ht!]
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig13}}
\caption{Comparison of Poynting flux and associated time scales for the simulations O16bM (solid) and
Z16M (dashed). a) Black lines show the horizontally averaged Poynting flux, blue
and red lines present the contributions from up and downflows. b) Energy loss
rate due to Poynting flux ($\zeta(z)$, Eq. \ref{tau_poyn1}). The red dotted line indicates a convective overturning
time scale $v_{z {\rm RMS}}/H_{\varrho}$. c) Fraction of energy transported by Poynting
flux relative to energy converted by Lorentz force ($\varepsilon(z) $, Eq. \ref{tau_poyn2}).
}
\label{fig:Poynting}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.85\hsize}{!}{\includegraphics{fig14}}
\caption{a) Ratio of horizontal to vertical field strength as
function of height. Different colors refer to simulations with the
average vertical field strength at $\tau=1$ as indicated. The ratio
of horizontal to vertical field has a maximum about $450$~km above
$\tau=1$ and is strongly dependent on the overall field strength of
the simulation and decreases with increasing field strength. b)
Probability distribution functions for the field inclination with respect
to the vertical. Solid lines refer to the deep photosphere around $\tau=1$,
dashed lines to about $450$~km height as indicated in panel a). The black
solid line indicates an isotropic distribution of field inclinations.
}
\label{fig:Inclination}
\end{figure*}
\begin{equation}
\zeta(z)=\frac{-P(z)}{\int_z^{z_{top}}E_M\,{\rm d}z}\;,\label{tau_poyn1}
\end{equation}
i.e. we compare the Poynting flux at a height $z$ to the total magnetic energy of the domain above $z$. For this analysis we can ignore
the Poynting flux at the top boundary, which is around $10^{-5} F_{\odot}$. In simulation Z16M with zero magnetic field in inflow regions
we find an about $1.8$ times larger loss rate. The vertical profile of $\zeta$ agrees very well with a convective time scale
$v_{z {\rm RMS}}/H_{\varrho}$, indicated by a red dotted line. The kinematic growth phase of the dynamo is not affected by details of the
bottom boundary condition as long as the growth rate fullfils $\gamma_K\gg v_{z {\rm RMS}}/H_{\varrho}$ (this condition is fulfilled well by
the higher resolution cases with a grid spacing of $8$~km or smaller, for the cases with $16$ and $32$~km grid spacing this condition is
fulfilled in the lower parts of the domain, but not in the photosphere). The bottom boundary matters when non-linear saturation
effects cause $\gamma(B) \rightarrow v_{z {\rm RMS}}/H_{\varrho}$. The simulation Z16M presents a setup with the maximum possible energy loss
at the bottom boundary, since the time-scale of magnetic energy loss is the same as the time scale for mass exchange. In that sense
this setup presents a lower limit for an efficient dynamo with $\gamma_K\gg v_{z {\rm RMS}}/H_{\varrho}$ (a less efficient dynamo could have
of course an even lower saturation field strength). Stronger saturation field strengths require less leaky bottom boundary conditions, which require
the presence of a Poynting flux in upflow regions, like in O16bM.
Panel c) compares the energy lost by the Poynting flux to the energy converted via the Lorentz force in the domain above a height $z$:
\begin{equation}
\varepsilon(z)=\frac{P(z)}{\int_z^{z_{top}}\vec{v}\cdot(\vec{j}\times\vec{B})\,{\rm d}z}\;.\label{tau_poyn2}
\end{equation}
\begin{figure*}
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig15}}
\caption{Energy transfer functions for the simulation O16a with $4$~km grid spacing.
Panel a) shows the kinematic growth phase, panel b) the saturated phase.
The line color corresponds to contributions from pressure/buoyancy (black),
advection of momentum (yellow), Lorentz force (green), stretching of magnetic field
(red), and advection of magnetic field (blue). Contributions from the
magnetic compression term are split 50/50 among stretching and
advection. Solid lines correspond to positive contributions, dashed lines
to negative contributions. Contributions from numerical diffusivity in the
induction equation are indicated by the purple lines. Here the dotted line
shows a Laplacian diffusivity with $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$
for comparison.
}
\label{fig:Transfer_kin_sat}
\end{figure*}
This fraction is lower in Z16M because of non-linear saturation effects in O16bM, which affect $\vec{v}\cdot(\vec{j}\times\vec{B})$
more strongly than $(\vec{v}\times\vec{B})\times\vec{B}$ (see also Figure \ref{fig:Saturation}).
While the simulation domain of O16bM contains almost $4$ times the magnetic energy
of Z16M, the average amount of energy converted from kinetic to magnetic energy is in both simulations comparable within a few
$\%$, i.e. $85\,\mbox{erg}\,\mbox{cm}^{-3}\,\mbox{s}^{-1}$ (O16bM) and $80\,\mbox{erg}\,\mbox{cm}^{-3}\,\mbox{s}^{-1}$ (Z16M),
which is about $50\%$ of the energy conversion by pressure/buoyancy forces in the domain, see also Section \ref{sec:transfer} for further detail.
Integrated over the depth of the domain this energy conversion rate equals to about $80\%$ of the energy flux through the domain.
Note that we discuss here conversion rates between energy reservoirs and not true sinks of energy, since the energy is returned to
internal energy through dissipation processes. The conversion rates can be comparable or even exceed the energy flux through the system. Most of the
energy converted from kinetic to magnetic is preferentially dissipated in downflow regions, while work against the Lorentz force reduces the kinetic
energy there. This changes the overall balance of convective energy transport by reducing the contribution from the kinetic energy flux. We find
in a non-magnetic convection simulation in $6$~Mm depth a downward directed kinetic energy flux of about $-0.3\,F_{\odot}$, this value is
reduced to $-0.2\,F_{\odot}$ in simulation O16bM.
Recently \citet{Hotta:2014:Global_norot} presented small-scale dynamo simulations in a global setup covering the convection zone up to $7$~Mm
beneath the photosphere. Using a similar numerical approach, but a substantially lower grid spacing of $1100$~km horizontally and $375$~km vertically,
they were able to maintain a field with $0.15-0.25\,B_{\rm eq}$ throughout the convection zone. The maintenance of the field requires in their setup around
$5-10\,\mbox{erg}\,\mbox{cm}^{-3}\,\mbox{s}^{-1}$, which is consistent with our results considering the differences in the overall field strength reached
(their field near the top boundary falls short of our values by a factor of $4$, which is reflected in a more than a factor of $10$ lower energy conversion rate).
Integrated over the entire convection zone the energy conversion rate by a small-scale dynamo could account to as much as few
$L_{\odot}$ ($L_{\odot}=3.84\cdot 10^{33} \mbox{erg}\,\mbox{s}^{-1}$). Compared to that the energy extracted from large-scale mean flows in mean field dynamo
models \citep{Rempel:2006:dynamo} as well as 3D global dynamo simulations \citep{Nelson:2013:Wreaths&Cycles} is about 2 orders of magnitude smaller.
\subsection{Horizontal magnetic field above $\tau=1$}
In Figure \ref{fig:Inclination} we further analyze how the ratio of horizontal and vertical field
as well as distribution of inclination angles varies as function of height. In Panel a)
we present the quantity $\langle(B_x^2+B_y^2)^{0.5}\rangle/\langle\vert B_z\vert\rangle$ as
function of height for the simulation O16bM. Since this simulation was started from a weak
seed field we selected during the growth phase 3 snapshots with the field strength of
$\langle\vert B_z\vert\rangle=20$, $40$, and $60$~G at $\tau=1$. We find that independent
from the field strength
$\langle(B_x^2+B_y^2)^{0.5}\rangle/\langle\vert B_z\vert\rangle$ peaks about $450$~km above
$\tau=1$. The peak value reached drops monotonically with increasing field strength from about
$5$ at $20$~G to $2.75$ at $60$~G. The dashed vertical lines indicate regions for which we computed the
PDFs for the inclination angle with respect to the vertical direction (panel b). Solid lines
refer to the PDFs around $\tau=1$, while dashed lines correspond to about $450$~km height.
For reference the black line indicates the distribution for an isotropic field. For all three
field strengths shown the PDFs are close to isotropic in the deep photosphere, but strongly
skewed toward horizontal field in $450$~km height. The contribution from horizontal field
is strong enough to create a distinct peak in the field strength about $450$~km above $\tau=1$
as presented in Figure \ref{fig:BeqBrms}, panel c).
\begin{figure*}[ht!]
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig16}}
\caption{Saturation of the dynamo comparing the kinematic growth phase with the
saturated phase of O4a. Panel a) shows the vorticity spectrum $\sqrt{k\,P(\omega)}$
(blue) and $T_{MS}/E_M$ (red). Solid (dashed) lines indicate the kinematic
(saturated) phase. Panel b) shows the corresponding kinetic energy and magnetic energy
power spectra.}
\label{fig:Saturation}
\end{figure*}
\subsection{Transfer functions and saturation process}
\label{sec:transfer}
Figure \ref{fig:Transfer_kin_sat} presents energy transfer functions computed for the simulation O4a (4 km grid spacing).
We compare here the kinematic growth phase (panel a) with the saturated phase (panel b).
The transfer functions are computed at a depth of about $800$~km beneath the photosphere.
They are averaged over a depth range of $160$~km. In addition we conducted a time average and applied
smoothing to the transfer functions in order
to suppress realization noise. Since the magnetic energy is varying rapidly during the
kinematic growth phase we normalized the transfer functions that depend on the magnetic field
by the total magnetic energy in each time step and averaged the normalized transfer functions
(the kinematic growth is self-similar, i.e. only the amplitude and not the shape of the
transfer functions is changing). In Figure \ref{fig:Transfer_kin_sat}a) we scaled the corresponding
transfer functions arbitrarily to show them on the same scale as the non-magnetic ones.
Colors
refer to the transfer function defined in the Appendix. The energy transfers to the kinetic energy reservoir
are $T_{KP}$ (black, energy transfer by pressure and buoyancy), $T_{KA}$ (orange, kinetic energy transfer by advection),
and $T_{KL}$ (green, kinetic energy transfer by the Lorentz force). The energy transfers to the magnetic energy
reservoir are $T_{MS}+0.5\,T_{MC}$ (red, energy transfer by stretching and compression), $T_{MA}+0.5\,T_{MC}$
(blue, energy transfer by advection and compression), and $T_{MD}$ (purple, energy transfer due to magnetic numerical
diffusivity). Solid (dashed) lines indicate positive (negative) contributions, the purple dotted lines indicate
the transfer of a Laplacian magnetic diffusivity with $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$ for
comparison (the simulations were only run with numerical diffusivity). We do not show the terms $T_{KD}$ for
better readability of the figures. We also split the term $T_{MC}$ $50/50$ among $T_{MA}$ and $T_{MS}$. The
reason for this (apart from reducing the number of quantities shown in Figure \ref{fig:Transfer_kin_sat}) is that we can
expand the underlying terms as:
\begin{eqnarray}
-\vec{B}&\cdot&\left((\vec{v}\cdot\nabla)\vec{B}+\frac{\vec{B}}{2}\nabla\cdot\vec{v}\right)=
-\nabla\cdot\left(\vec{v}\frac{B^2}{2}\right)\label{eq:tma}\\
\vec{B}&\cdot&\left((\vec{B}\cdot\nabla)\vec{v}-\frac{\vec{B}}{2}\nabla\cdot\vec{v}\right)=
\nabla\cdot\left(\vec{B}(\vec{v}\cdot\vec{B})-\vec{v}\frac{B^2}{2}\right)\nonumber\\
&&-\vec{v}\cdot\nabla\cdot\left(\vec{B}\vec{B}-\frac{1}{2}\vec{I}\,B^2
\right)\label{eq:tms}
\end{eqnarray}
i.e. the terms underlying $T_{MA}+0.5 T_{MC}$ (Eq. \ref{eq:tma}) can be identified with an advective energy
transport within the magnetic energy reservoir analogous to the terms underlying
$T_{KA}$ that refer to the turbulent momentum cascade. The terms underlying $T_{MS}+0.5 T_{MC}$ (Eq. \ref{eq:tms})
describe in part a transport within the magnetic energy reservoir (remaining non-advective
terms of Poynting flux) and in part the energy transfer with the kinetic energy reservoir (via Lorentz force).
During the kinematic growth phase (panel a) $T_{MS}+0.5 T_{MC}$ peaks on a scale of about $25-30$~km, which
is about about $6-8\,\Delta x$, i.e. close to the smallest features that can be resolved with the given
grid spacing. With increasing resolution this scale is decreasing as it stays near $6-8\,\Delta x$.
The corresponding Lorentz force related energy transfer $T_{KL}$ shows two peaks, one at around
$50$~km and one around $2$~Mm. The peak at around $50$~km is related to the magnetic tension force, while
the peak at $2$~Mm is caused by magnetic pressure. The dominant contribution to the energy exchange comes
from the peak at small scales. On scales larger than $20$~km $T_{MS}+0.5 T_{MC}$ is partially opposed
by the transport term $T_{MA}+0.5 T_{MC}$ and a numerical diffusion term of similar amplitude. The remainder
leads to an exponential growth of magnetic energy with a e-folding time scale of about $50$~sec. On
scales smaller than $20$~km positive contributions from $T_{MS}+0.5 T_{MC}$ and $T_{MA}+0.5 T_{MC}$ are opposed
by numerical diffusivity. The contribution from numerical magnetic diffusivity (dashed purple line) is on scales
larger than $100$~km very similar to a Laplacian diffusivity with $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$
(dotted purple line), moderate differences exist on smaller scales. However, replacing our numerical diffusivity with
Laplacian diffusivity of $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$ leads to an about $6$ times smaller kinematic
growth rate of the dynamo, which implies that it is non-trivial to estimate the effective numerical diffusivity by looking
at transfer functions or energy dissipation rates.
In the saturated phase (panel b) $T_{MS}+0.5 T_{MC}$ peaks on a scale of about $250-300$~km, about a factor of $10$
larger than during the kinematic growth phase. Similarly $T_{KL}$ peaks now at a scale of $500$~km, i.e. most of the
energy transfers from kinetic to magnetic energy happen on a scale comparable to downflow lanes. Unlike the kinematic
growth phase these scales are independent of resolution and realized in all simulations presented here regardless of their
resolution. On scales larger than $100$~km $T_{MS}+0.5 T_{MC}$ is in balance with $T_{MA}+0.5 T_{MC}$, contributions from
numerical diffusivity, $T_{MD}$, are about 2 orders of magnitude smaller. While $T_{MD}$ was close to Laplacian during the
kinematic growth phase, it differs substantially during the saturated phase. Contributions on scales larger than about
$70$~km are in amplitude about a factor of $10$ smaller than a Laplacian with $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$
(dotted purple line) and have the opposite sign. The latter is related to the cutoff we introduced in Eq. (\ref{phi_lim}) for
a setting of $h=2$. Integrated over all scales the positive contribution accounts to about $1.7\%$ of the total unsigned dissipation.
The sign change is not present for grid spacings of $8$~km and larger and can be avoided by using a setting
of $h=1$ in higher resolution cases (with no significant difference to the obtained results). On the smallest scales the
contribution from numerical diffusivity remains similar to a Laplacian diffusivity with $\eta=5\cdot 10^{9}\,\mbox{cm}^2\mbox{s}^{-1}$.
\begin{figure*}[ht!]
\centering
\resizebox{0.95\hsize}{!}{\includegraphics{fig17}}
\caption{Energy transfer functions for simulations C8 and C8$\eta$ with $8$~km grid spacing
using the boundary {\it CH}. The meaning of the line styles is the same as
in Figure \ref{fig:Transfer_kin_sat}.
Panel a) shows a simulation computed with artificial magnetic
diffusivity, panel b) a simulation computed with a Laplacian
diffusivity using a values of $\eta=10^{10}\,\mbox{cm}^2\mbox{s}^{-1}$. The
dotted purple line in panel a) indicates a Laplacian diffusivity with
$\eta=10^{10}\,\mbox{cm}^2\mbox{s}^{-1}$ for comparison.
}
\label{fig:Transfer_eta_hyp}
\end{figure*}
In the saturated phase the Lorentz force is a dominant contributor in the momentum equation and dominates energy transfers
on scales smaller than $500-1000$~km. While we find in the kinematic growth phase a balance between pressure/buoyancy driving,
$T_{KP}$, and the kinetic energy cascade, $T_{KA}$, down to scales of about $150$~km, this balance is only realized on scales
larger than $1000$~km in the saturated phase. From scales of $500$~km down to scales of $50$~km the term $T_{KA}$ balances
mostly $T_{KL}$ and as a consequence the amount of kinetic energy that is transported to small scales by $T_{KA}$ is significantly
reduced. Not all of the energy that is extracted by the Lorentz force is transferred into magnetic energy. On scales smaller
than $40$~km $T_{KL}$ becomes the dominant source of kinetic energy. Overall $30\%$ of the energy extracted from kinetic energy on
large scales is returned to kinetic energy on small scales. Comparing transfer functions computed for simulations with different
resolution, we find that this fraction increases with resolution.
The total amount of energy that is dissipated numerically is within $10\%$ the same between kinematic and saturated phase.
In the saturated phase about $55\%$ of the energy dissipation happens through the magnetic channel. That we find about the same
energy dissipation rates from magnetic and viscous dissipation is possibly related to an intrinsic numerical $P_M$ close to $1$.
\citet{Brandenburg:2011:SSD_low_Pm} found that this ratio depends on $P_M$ and that more energy is dissipated through the
magnetic channel for low $P_M$.
Figure \ref{fig:Saturation}a) analyzes the saturation mechanism of the dynamo. To this end we compare the vorticity spectrum
and the normalized transfer function $T_{MS}/E_M$ between kinematic growth phase and
saturated phase. We present these quantities averaged over the same depth range as the transfer functions in Figure \ref{fig:Transfer_kin_sat}.
We see only a moderate reduction of the overall shear by about $40\%$, mostly on smaller scales.
$T_{MS}/E_M$ drops by more than a factor of $10$, which indicates that most of the saturation happens though a
misalignment of magnetic field and velocity shear. Panel b) shows the corresponding kinetic and magnetic energy spectra,
which are not very different from the photospheric spectra shown in Figure \ref{fig:pwr_pdf_sat}. In the saturated phase kinetic
energy is suppressed on scales smaller than $100$~km and magnetic energy is in super equipartition by about a factor of $1.7$.
The sum of kinetic and magnetic energy in the saturated phase is similar to the kinetic energy during the kinematic phase.
Figure \ref{fig:Transfer_eta_hyp} compares a solution computed with numerical diffusivity (panel a) to a solution computed
with a Laplacian magnetic diffusivity using $\eta=10^{10}\,\mbox{cm}^2\mbox{s}^{-1}$. These simulations are both
computed with a closed bottom boundary, i.e. the saturation level reached is determined by processes
within the computational domain and not sensitive to the magnetic bottom boundary condition. We present the
same quantities as in Figure \ref{fig:Transfer_kin_sat}. In panel a) the dotted purple line shows the transfer functions
of a Laplacian diffusivity with the same values as in panel b) for comparison. The solution computed with
a Laplacian diffusivity saturates at about half the field strength compared to the case with only numerical diffusivity.
The differences seen in the transfer functions for both cases are mostly a reflection of the differences in
overall field strength, i.e. we see the peak of $T_{MS}+0.5 T_{MC}$ shifted toward smaller scales as expected.
The contribution from numerical diffusivity on large scales in panel a) is by about one order of magnitude smaller than the
contribution from the Laplacian diffusivity in panel b) (relative to the terms $T_{MS}+0.5 T_{MC}$. The value of
$\eta=10^{10}\,\mbox{cm}^2\mbox{s}^{-1}$ is very close to the smallest value we can use for a grid spacing of $8$~km without
excessive ringing and numerical instability on small scales. Achieving a regime similar to that shown in panel a) on large scales
using a Laplacian magnetic diffusivity would likely require about 10 times higher resolution, i.e. $10^4$ times more computing time.
\section{Discussion}
\label{sec:discussion}
Since a significant fraction of the magnetic energy in the presented simulations is found on scales smaller than the resolution
of current instrumentation, a detailed comparison with observations is only possible through spectro-polarimetric
forward modeling. We defer such in-depth comparison to future publications, and limit the discussion here to a more
qualitative level.
\subsection{Field strength of quiet Sun}
We presented a series of simulations that lead to saturation field strength about a factor of $2-3$ higher than
previously found by \citet{Voegler:Schuessler:2007}. We achieved that by a combination of a sufficiently high resolution
in combination with an LES approach
(i.e. sufficient low numerical diffusivities) and different bottom boundary conditions. Higher resolution alone is
not sufficient, we found that the bottom boundary condition plays a crucial role. Using a conservative boundary assuming
zero magnetic field in inflow regions or only vertical magnetic field (similar to \citet{Voegler:Schuessler:2007}), we obtain
a lower limit of about $\langle\vert B_z \vert\rangle=30$~G at $\tau=1$. The term "lower limit'' refers here to efficient
dynamos, which have a growth rate $\gamma_K\gg v_{z {\rm RMS}}/H_{\varrho}$ during their kinematic phase. Less efficient dynamos
can have of course a lower saturation field strength. We derive an upper asymptotic limit of about $85$~G based on
the assumption that $B_{\rm RMS}$ cannot increase faster with depth than $B_{\rm eq}$. While our lower limit is possibly affected by
the overall dynamo efficiency including potential magnetic Prandtl-number effects we do not account for, the upper limit
is only set by the available kinetic energy of the near surface convection zone. Similar values are also found in a setup with
a closed bottom boundary, which, while less solar-like, provides a better posed dynamo problem. If we take our simulation
O16bM as basis for the extrapolation, our "upper limit" of $\langle\vert B_z \vert\rangle=85$~G implies a value of
$B_{\rm RMS}=275$~G, $\langle B \rangle=160$~G, and $\langle B_h \rangle=120$~G (all at $\tau=1$). These values are
similar to those found by \citet{Danilovic:2010:zeeman_dynamo} through spectropolarimetric forward modeling
of rescaled dynamo models and comparison with observations ($\langle\vert B_z \vert\rangle=84$~G, $\langle B \rangle=170$).
We have repeated their analysis with non-grey versions of some of the simulations presented here (Danilovic $\&$ Rempel in prep.) and
found that a model with $\langle\vert B_z \vert\rangle=60$~G agrees best with the data used in \citet{Danilovic:2010:zeeman_dynamo}.
The difference comes from the fact that due to non-linear feedback a rescaled weak-field solution is not identical to the fully non-linear
strong field solutions we consider here.
Inversion results by \citet{Orozco:Rubio:2012:QS_LP} ($\langle\vert B_z \vert\rangle=64$~G, $\langle B \rangle=220$) lead to
similar values for $\langle\vert B_z \vert\rangle$, but significantly stronger field strength, which is due to an about a factor of $2$
stronger horizontal field ($198$~G) in their case..
\citet{Trujillo:etal:2004} and \citet{Shchukina:2011:hanle_dynamo} inferred from Hanle depolarization measurements values of
$\langle B \rangle$ around $130$~G a few $100$~km above $\tau=1$. If we use our upper limit as reference, we find a field strength
in that height range around $80-90$~G, about a factor of $1.5$ less. If we use a model with $\langle\vert B_z \vert\rangle=60$~G
(best fit to {\it Hinode} Zeeman data) the disagreement is almost a factor of $2$.
Overall these results indicate that the quiet Sun is magnetized near the upper limit we find, i.e. the observed field strength implies
that the subphotospheric layers have to be magnetized close to equipartition. The experiments with $B=0$ at the bottom boundary
indicate that a small-scale dynamo restricted to the top $1-2$~Mm of the convection zone could only explain about $50\%$ of the
observationally inferred field strength, i.e. the origin and strength of the quiet Sun magnetic field cannot be understood in separation
from the deeper layers of the convection zone.
\subsection{Scales beyond granulation}
\label{sect:disscuss_large_scale}
We presented one simulation on meso-granular scales using a domain $25$~Mm wide and $6$~Mm deep.
Compared to the smaller domains in which no scales larger than granulation are allowed for, this simulation shows an organization
of magnetic field on scales larger than granulation and also a stronger contribution from kG field. On a fundamental level this shows
that the organization of magnetic field on a wide range of scales is not inconsistent with a small-scale dynamo, provided that the
small-scale dynamo operates itself on a wide range of scales. The latter is a natural consequence of stratified convection in larger
domains and does not require the contribution from a global dynamo, although that contribution becomes unavoidable once the
domain size approaches scales on which rotation and shear become important. In that case a separation of contributions from a
small-scale and large-scale dynamo is not trivial and not necessarily meaningful.
\subsection{Coupling between the large- and small-scale dynamos}
\label{sect:discuss_LSD-SSD}
In our numerical experiments we focused entirely on setups with no imposed netflux, which leads to the question of how
the presence of a (cyclic) mean field produced by a large-scale dynamo might influence these results. We conducted an additional
experiment (similar to O16bM) in which we imposed initially a vertical mean field of $30$~G. We found roughly a doubling of the
photospheric magnetic energy, which is mostly due to the formation of a strong meso-granular network with $\vert B\vert >1$~kG, while
the core of the probability distribution functions remains unchanged (the magnetic energy in regions with $\vert B\vert <500$~G
is unchanged, in regions with $\vert B\vert<1$~kG we find only a $10\%$ increase). This result is expected since there is little
recirculation of mass in the top layers of the convection zone. The imposed netflux is expelled quickly into longer lived downflow
regions forming a magnetic network structure. The resulting inter-network regions are mostly void of netflux and have properties
similar to a simulation without any netflux. A similar weak dependence of the strength of inter-network field and the strength of the
surrounding network field was also found by \citet{Lites:2011:hinode_ssd} in {\it Hinode} data. Overall this indicates that the strength
of mixed polarity field in the photosphere is only weakly influenceable by a vertical mean field. A much stronger modulation is possible
through the properties of horizontal field in upflows regions, which is reflected in the strong dependence of our results on the details
of the bottom boundary condition. How much the strength of the quiet Sun is modulated by such a coupling depends ultimately
on how strongly the large-scale mean and small-scale mixed polarity field in the bulk of the convection zone vary throughout the
cycle. Current global dynamo models do not have sufficient resolution to properly capture small-scale field and likely underestimate
its contribution. As discussed in Section \ref{sec:energetics} our estimates indicate that the energy converted by small-scale induction
effects exceeds the induction by large-scale mean flows (mostly differential rotation) by about two orders of magnitude.
There are also possible feedbacks from the small-scale on the large-scale dynamo as well as convective dynamics. The presence of
small-scale field suppresses turbulent motions and reduces the kinetic energy flux. The amount of energy taken out of convective
motions through the Lorentz-force is substantial, if we equate this energy transfer with a viscous energy transfer we would require
an effective viscosity of a few times $10^{11}\mbox{cm}^2\mbox{s}^{-1}$ to mimic this effect. Recently \citet{Fan:Fang:2014:dynamo}
showed that the presence of magnetic field in the convection zone can be crucial for maintaining a solar-like differential rotation and
that the contribution of the magnetic field can be approximated to some degree by an enhanced effective viscosity of the flow. To which
degree small- and large-scale field components contribute to this effects requires further investigation.
\subsection{Distribution functions, kG field concentrations}
\label{sec:discuss_PDF}
We find that probability distribution functions are very robust, i.e. they barely depend on numerical resolution (we varied
the grid spacing by a factor of $16$!). The shape of PDFs is mostly determined by the average field strength and domain size.
For a given average field strength we find more strong field in a larger domain. Comparing photospheres with
$\langle\vert B_z \vert\rangle=80$~G we find that $50\%$ of the energy comes from field with less than $500$~G, kG field
concentrations contribute about $16\%$ to the total energy. The latter drops to $9\%$ for a $60$~G case and increases to
$23\%$ for a $60$~G case in a larger domain. In our simulations the filling factor of kG field concentrations is strongly
field strength dependent. More than $\langle\vert B_z \vert\rangle\sim 30-40$~G is required at $\tau=1$ before they form and
the filling factor increases steeply as the field strength increases beyond that threshold.
Comparing the shape of normalized PDFs for vertical and
horizontal field components we find that the PDFs for vertical field in the photosphere deviate substantially from those of
horizontal field as well as vertical field beneath the photosphere (see Figure \ref{fig:pwr_pdf_3lev}). This is a strong hint for
the presence of a convective intensification mechanism \citep{Schuessler:1990:IAUS} that is restricted to the photosphere and
mostly affects vertical field. While we find strong field at $\tau=1$ reaching up $2.5$~kG, this field component is not organized
in form of flux tubes, nor does it have a preferred scale around $100$~km. Strong magnetic field is typically organized in sheets,
often with alternating polarities. kG flux concentrations are small knots along these sheets in which the field strength is increased
temporarily due to dynamical effects. We find kG flux concentrations down to the smallest scales we can resolve. The kG field present
in our simulations does not produce a distinct feature in the PDFs like a secondary peak around kG field strength found in many
observations \citep{Ishikawa:2009:cmp_QS_plage,Lites:2011:hinode_ssd,Orozco:Rubio:2012:QS_LP}. These observation also indicate that
the second peak is possibly caused by contributions from network field and may not be present for inter-network field alone.
\subsection{Power spectra}
\label{sec:discuss_power}
In our highest resolution simulation (grid spacing of $2$~km) we find that about $50\%$ of the magnetic energy in the deep
photosphere is found on scales smaller than $100$~km. Therefore properly reproducing the spectral energy distribution
requires the highest possible resolution. Performing simulations with lower resolution will artificially move magnetic energy
toward larger scales in spectral space, for example we find $50\%$ on scales smaller than $300$~km for a grid spacing
of $32$~km. On scales smaller than $100$~km magnetic energy is in super-equipartition buy about a factor of $2$. A similar
feature has been found in several small-scale dynamo simulations and LES models of MHD turbulence
\citep[see, e.g., review by][]{Brandenburg:2012SSRv}. We further find that the sum of kinetic and magnetic energy power spectrum
in the saturated state is similar to the kinetic energy power spectrum of a pure HD simulation, i.e. kinetic energy is suppressed
on scales smaller than $100$~km and that gap is filled with magnetic energy. Even in the highest resolution case it is difficult
to identify power laws, in particular for magnetic energy. We see some indication for power laws in the kinetic energy. Steeper
slopes (as steep as $-2.7$) are typically found for scales larger than a few $100$~km (width of downflow lanes). On smaller scales
the slope is height dependent, we find $-1.4$ at $\tau=0.01$ and $1$~Mm beneath $\tau=1$, while we see a steeper
$-2.2$ slope at $\tau=1$ in between these layers. This likely indicates that the slope at $\tau=1$ will change when
approaching smaller scale, otherwise the kinetic energy at $\tau=1$ would drop below that at $\tau=0.01$. Recently
\citet{Katsukawa:2012:power} derived power spectra for kinetic and magnetic energy from {\it Hinode} data. They found
in the frequency range from $1.5$ to $3.5\,\mbox{Mm}^{-1}$ kinetic energy spectra with slopes around $-3.3$ to $-3.6$,
while the slope of the magnetic energy spectrum is less steep with a slope around $-1.4$. While a detailed comparison
of these slopes is likely difficult without properly accounting for resolution and noise effects, we see in our simulations at
least some indication for the substantially different slopes for kinetic and magnetic energy. In the frequency range
$1.5$ to $3.5\,\mbox{Mm}^{-1}$ we find at $\tau=1$ and $0.01$ steep kinetic energy spectra with slopes as steep as
$-2.7$, while magnetic energy spectra are flat with slopes of less than $-1$. Based on these simulations we caution not to
extrapolate these slopes as they still change significantly toward smaller scales.
We see no indication that kG magnetic field present in our simulations would create a secondary peak in the magnetic power spectrum
around $100$~km as suggested by \citet{Stenflo:2012:scaling}.
\subsection{Field inclination}
We find in the deep photosphere a close to isotropic magnetic field distribution, while higher layers are dominated by horizontal field.
The ratio of horizontal to vertical field peaks about $450$~km above $\tau=1$. The exact value of this ratio is strongly field strength
dependent. While we can find ratios as high as $5$ for a solution with $\langle\vert B_z \vert\rangle=20$~G,
the ratio drops to less than $3$ for field strength values that are most compatible with observations. Over the height range where for
example {\it Hinode} observations are taken the ratio is more close to $2$. These values are consistent with those reported by
\citet{Schuessler:Voegler:2008:bhorz} when we take into account that they considered a dynamo model reaching only
$\langle\vert B_z \vert\rangle=25$~G, and that they used in their estimates the horizontal RMS field strength which is up to a
factor of $2$ stronger than the mean horizontal field we use. On the observational side \cite{Lites:etal:2008} found a ratio of $5$, while
recently \citet{Orozco:Rubio:2012:QS_LP} deduced a lower value of $3.1$. In contrast to that other investigations such as \citet{MartinezGonzales:2008:iso,AsensioRamos:2009:iso} find a mostly isotropic field. Recently \citet{Stenflo:2013:inc} showed that
the angular distribution of magnetic field in the quiet Sun varies with height. While the deep photosphere is more vertical, the upper photosphere
tends to be more horizontal. At least on a qualitative level we find a similar result, the deep photosphere is close to isotropic and higher
layers are more horizontally inclined, with the most horizontal distribution found $450$~km above $\tau=1$.
\subsection{Spectral energy transfers}
We computed for our models spectral energy transfer functions in order to analyze in detail the operation of the dynamo during kinematic
and saturated phase.
During the kinematic growth phase most energy transfers happen on scales $\sim 6-8\,\Delta x$, i.e. this scale is resolution
dependent. Overall the transfer functions for the kinematic growth phase are similar to those presented by
\citet{Pietarila-Graham:etal:2010:SSD}. In the saturated phase energy transfers happen on a scale of $250-500$~km, i.e. the
scale of downflow lanes. This scale is found independent of the adopted numerical resolution. Contributions from numerical diffusivity
on scales larger than $10\,\Delta x$~km are negligible and a balance between stretching and non-linear transport terms is achieved, i.e.
the regime expected for $R_M\gg 1$ (which is the main reason for using an LES approach in the first place). In the saturated phase the
Lorentz force becomes the major player in the momentum equation on scales smaller than $1$~Mm.
Comparing models using physical and numerical magnetic diffusivity we do not find a significant difference in the transfer functions, but
differences exist in the kinematic growth rate (about $6$ times larger with numerical diffusivity) and saturation field strength
(about $2$ times stronger with numerical diffusivity).
Ultimately the validity of a LES approach for a small-scale dynamo problem has to be tested
against high resolution DNS models. While the latter is becoming feasible for simplified setups, it will remain a substantial challenge for realistic
solar-like MHD simulations as discussed here.
\section{Concluding remarks}
\label{sect:conclusion}
We presented a series of MHD simulations of the solar photosphere with the aim of understanding better the origin of
small-scale mixed polarity magnetic field in the solar photosphere. Such field is ubiquitous on the solar surface and provides the
dominant contribution to the magnetic energy in the quiet Sun photosphere. Previous simulations of small-scale dynamos fell short by a factor of
$2-3$ in terms of the field strength required to explain the observed level of Zeeman polarisation\citep{Danilovic:2010:zeeman_dynamo}.
In our study we showed that increasing resolution and reducing numerical diffusivities alone is not sufficient to reach the observationally
inferred field strength. The saturated solutions remain strongly dependent on the (open) bottom boundary condition typically used in photospheric
MHD simulations (note: the kinematic growth phase is not very sensitive to the choice of the bottom boundary condition). Physically this boundary
condition dependence implies a strong coupling between the photosphere and the deeper
convection zone, i.e. the magnetism of the photosphere cannot be understood in separation from the rest of the convection zone.
Solutions that agree with observational constraints on the field strength in the photosphere imply a subsurface magnetic field with an
energy density comparable to the kinetic energy density. By that measure the small-scale magnetic field is far from being a weak field.
The energy conversion through the associated Maxwell-stresses accounts to about a solar luminosity when integrated over the top $10$~Mm
of the convection zone, which is about $50\%$ of the energy conversion by pressure/buoyancy forces. The Lorentz-force feedback on the flow
leads to a significant reduction of kinetic energy flux as well as kinetic energy on small scales. Such feedback has potentially dynamical consequences
for the convective dynamics of the convection zone including the maintenance of mean flows and operation of a large-scale dynamo
\citep{Fan:Fang:2014:dynamo}.
We further studied how the strength and distribution of magnetic field in the photosphere is influence by numerical resolution. Since about
$50\%$ of the magnetic energy resides in the deep photosphere on scale smaller than $100$~km, a sufficiently high numerical resolution
(ideally a grid spacing smaller than $8$~km) is required for properly capturing the spectral energy distribution. Probability distribution
functions of the magnetic field strength are on the other hand insensitive to numerical resolution in the explored range (grid spacings from
$32$ down to $2$~km) and can be considered converged. Our models are all based on the strong assumption that an LES approach
(using only numerical diffusivities) with an intrinsic $P_M$ close to unity is a proper way to deal with the regime found on the Sun. Within
this approach results are very consistent and robust, which implies that they are either correct or systematically wrong. A detailed comparison
with observations through forward modeling, which is work in progress, should tell which is the case.
\acknowledgements
The National Center for Atmospheric Research (NCAR) is sponsored by the National Science Foundation. The author thanks M. Sch{\"u}ssler,
A. DeWijn and Kyle Augustson for helpful discussions and comments on the manuscript. We would like to acknowledge high-performance computing support from Yellowstone
( http://n2t.net/ark:/85065/d7wd3xhc) provided by NCAR's Computational and Information Systems Laboratory, sponsored by the National
Science Foundation, under project NHAO0002; from the NASA High-End Computing (HEC) Program through the NASA Advanced
Supercomputing (NAS) Division at Ames Research Center under project s9025; and by the National
Science Foundation and the University of Tennessee through the use of the Kraken computing resource at the National Institute
for Computational Sciences (http://www.nics.tennessee.edu) under grant AST100005.
This research has been partially supported through NASA contracts NNH09AK02I, NNH12CF68C and NASA grant NNX12AB35G. M. Rempel is grateful
to NAOJ for support of a Visiting Professorship during November 2011. Many fruitful discussions with the Japanese solar physics community during
that time sparked my interest in quiet Sun magnetism.
\bibliographystyle{natbib/apj}
|
\section{INTRODUCTION}
\label{sec:intro}
Nulling interferometry\cite{bra78} has been for long identified as a unique and competitive technique to perform mid-infrared spatially resolved spectroscopy of Earths and Super-Earths orbiting nearby main-sequence stars\cite{lab14}. The original Bracewell setup consisted in phase-shifting by 180 degrees one of the beam of a 2-telescope Michelson interferometer in order to cancel out the light of the on-axis bright star and reveal the dim glow of the off-axis companion. By combining in nulling mode more than two telescopes, a broader central dark fringe can generally be obtained\cite{men97}, which allows us to further reduce the stellar leaks resulting from the star being resolved by the interferometer, and furthermore to modulate the exo-zodiacal signal. In addition, a broader null allows higher instrumental tolerances with respect to the residual piston phase and photometric unbalance. Infrared space-based interferometry has been historically the preferential way to implement the nulling technique since optimal sensitivity and spectral coverage in the mid-infrared can be guaranteed from space to detect the photons from the faint planet. The technique has benefited from numerous technical and feasibility studies to investigate the requirements for space operation\cite{coc09}. The inherent complexity of the beam combination stage in a multi-aperture interferometer has always been a pitfall for nulling, and for space interferometry in general. It has been often suggested that the integrated optics (IO) approach would be able to highly simplify the optical design of such an instrument\cite{lab08}, but was experimentally tested only at the level of a simple two-aperture Bracewell setup\cite{web04}, which nevertheless mitigates the advantage of the multi-aperture interferometry approach. While four-beam interferometric nulling concepts have been successfully studied at JPL with the Planet Detection Testbed\cite{mar12}, the potential of the IO for implementing a four-aperture - or plus - design has hardly been investigated both theoretically and experimentally.
Here we concentrate on exploring a four-beam combination nulling scheme based on cascaded 2x2 directional couplers, which have inherent $\pi$-phase shifted outputs. Combining together the nulled outputs is the condition for producing the necessary broad dark fringe. In this paper, the accent is set on showing the proof-of-concept of an integrated optics four-telescope nuller rather than on achieving and optimizing deep level of nulling. We therefore report, for the first time to our knowledge, the characterization of an integrated optical component designed to combine in nulling mode 4 telescopes. Our investigation has been conducted at visible wavelength (632\,nm), but the emergence of new photonics solutions operational in the 10-microns spectral range\cite{lab09} justifies the interest of our approach in preparation of future mid-infrared instrumentation.
\section{EXPERIMENTAL SETUP}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=5cm]{Angel-Wolf.eps}
&
\includegraphics[height=5cm]{integrated_beam_combiners.eps}
\end{tabular}
\end{center}
\caption[Telescope setup]
{ \label{fig:teles}
Left: Telescope arrangement, for which the nulling interferometry was simulated. The mirror diameter of the telescopes at the edge (telescopes 1 and 2) are half the diameter of the inner ones (telescopes 3 and 4). Right: Setup of the photonic chip, using 3 beam combiners ($\alpha$, $\beta$, $\gamma$) with 50/50 coupling ratio each. The numbers of the input channels correspond to the number of the telescopes.}
\end{figure}
Our goal is to investigate the potential of the integrated optics approach for multi telescope nulling interferometry. To this end, we manufactured and tested an optical chip which reproduces the functionalities of the beam combiner of the classical Angel \& Woolf\cite{ang97} configuration for multi-telescope nulling. In its basic form, this configuration (illustrated in Fig.~\ref{fig:teles}.a) consists in a linear array of 4 telescopes located at coordinate 0,$d$, $3d$ and $4d$, respectively. The two inner telescopes (3 and 4) collect a flux which is four times the one collected by the two external telescopes (1 and 2). The beam combiner of the interferometer consist of three 50/50 beam splitters arranged on two levels. At the first level, two beam splitters are used to combine pairwise the internal and the external telescopes in nulling mode, \textit{i.e.} the phase delay between the telescope pairs is arranged such that at each beam splitter one output exhibits a destructive interference (nulled output).
At the second level, a third beam splitter is used to combine the nulled outputs of the previous level of beam combination.
The implementation in terms of integrated optics of the 4-telescopes nulling beam combiner is straightforward. The scheme of the integrated optical circuit is shown in Fig.~\ref{fig:teles}.b. Light propagates from left to right. The beam splitters are replaced by 50/50 waveguide couplers and are arranged in a two-level cascade, the third coupler connecting two of the outputs of the first two. The four input waveguides are connected to the respective telescopes according to the numeration of Fig.~\ref{fig:teles}.a.
The IO component was fabricated by femtosecond laser inscription in silica substrate\cite{per04} and has the physical dimensions of $360\,\mu$m$\times 50$\,mm, as indicated in the Figure. Waveguides are strictly single mode for the wavelength we have chosen for the experiment ($\lambda=633$\,nm, corresponding to the emission line of a He-Ne laser).
\begin{figure}[t]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{setup_telescopes.eps}
\end{tabular}
\end{center}
\caption[Setup]
{ \label{fig:setup_tel}
Scheme of the setup used to simulate the four telescopes including the phase modulation and flux control. The flux of the beams 1 and 2 are adjusted to 25\% of the beams of 3 and 4. The integrated optics device for the beam combination is shown also in Fig.~\ref{fig:teles}.b.}
\end{figure}
The simulation of the telescopic array and the motion of the star in the sky required the implementation of a rather complex setup which is schematically outlined in Fig.~\ref{fig:setup_tel}. It consists basically in a modified Mach-Zehnder interferometer which is used to prepare 4 beams out of a collimated and astigmatically shaped laser beam (as mentioned, we used a He-Ne laser as a source). The four beams are recombined with a small tilt to each other in the horizontal direction so that they can be focused at the input of the corresponding waveguides of the photonic chip by means of a NA=0.25 microscope objective (the numbering of the beams in Fig.~\ref{fig:setup_tel} corresponds to the telescopes/waveguides illustrated in Fig.~\ref{fig:teles}).
Astigmatic beam shaping of the laser beam was necessary to improve the coupling efficiency into the laser-written waveguides of the photonic chip, which support elliptical propagating modes\cite{min12}.
The brightnesses of the different beams is controlled by a combination of rotatable $\lambda$/2 plates and a polarizer. As expected for the Angel\&Wolf configuration, the input beams 1 and 2 are set to a quarter of the flux of the beams 3 and 4. In order to include the effect of the coupling efficiency in the waveguides (which we did not measure), this calibration was performed by measuring the total flux at the output of waveguides ABCD for single beam injection. The measurement of the output fluxes (described in the next section) was carried out on images of the output facet of the photonic chip recorded with a high-resolution, 10-bit CCD camera attached to a microscope.
To simulate the motion of a target star in our model interferometer, we control the relative optical path difference (OPD) between the 4 beams with three movable mirrors, which we fabricated by attaching the mirrors to the membrane of loudspeakers.
If the star transits near the zenith, the $\mathrm{OPD}$ measured at each telescope respect to the first one is given by $\mathrm{OPD_n}=b_\mathrm{n}\cdot\theta$, where $b_\mathrm{n}$ is the baseline between the first and the $\mathrm{n^{th}}$ telescope (n=2,3,4) and $\theta$ is the zenital angle (see Fig.~\ref{fig:teles}.a).
In our setup, we simulate the linear dependence of the OPD on the angle $\theta$ by feeding the loudspeakers with a sawtooth driving voltage whose amplitude is proportional to the baseline.
Therefore, we kept fixed the path of telescope 1 (our reference) and modulated the delay for the beams 2, 3 and 4 with a signal of OPD amplitude $2\,\lambda$, $0.5\,\lambda$, and $1.5\,\lambda$ (corresponding to baseline $4d$, $1d$, and $3d$), respectively. We chose to map the baseline $d$ on an amplitude of $0.5\,\lambda$ as we are limited by the linear response of the loudspeakers in respect to the input voltage.
The phase delay control of the setup as well as the recording of the output of the IO are automatized with a \texttt{LabVIEW} program.
\section{CHARACTERISATION OF THE PHOTONIC CHIP}
\begin{figure}[b]
\begin{center}
\begin{tabular}{c c}
\includegraphics[width=0.45\textwidth]{image_output_field-inp1.eps} & \includegraphics[width=0.45\textwidth]{image_output_field-inp4.eps}
\end{tabular}
\end{center}
\caption[Output field]
{ \label{fig:output_field}
Logarithmic output field for the channels A, B, C, D. Light was coupled into the input channels 1 (left image) and 4 (right image). The aperture (10\,px radius, corresponding to $10.2\,\mu$m), which includes 90\% of the light, as well es ring shaped area for the background measurements (25-35\,px) are shown in the images. Grayscale in decadic logarithm of the counts.}
\end{figure}
The characterization of the laser-written photonic chip consisted in measuring the splitting ratio of the couplers.
As mentioned in Section 2, the 4-telescope nuller requires beam splitting ratios of exactly 50\%.
In order to cope with the fabrication uncertainties associated with laser writing, several samples of the beam combiner were written in a glass substrate and tested. Samples differ from each other by the separation of the waveguides at the center of the couplers, which controls the coupling ratio between the waveguides. Samples with waveguide separations between 17 and $19\,\mu$m in steps of $0.5\,\mu$m were written and characterized individually.
The splitting ratios were inferred from the photometry of the output waveguides ABCD for single waveguide excitation of the photonic chip.
We employed the aperture photometry method, that is measuring on CCD images the flux inside a circular aperture around the center of an output waveguide and correct it with the background flux. The latter one is calculated in a ring around the aperture with a gap between ring and aperture. We used a radius of 10\,pixel (px) for the aperture, which corresponds to $10.2\,\mu$m at the photonic device, and a ring starting at a distance 25\,px from the center and a width of 10\,px.
Fig.~\ref{fig:output_field} shows two typical images from the output facet of the photonic chip and overlaid the photometric region of interest corresponding of output waveguide B.
With this method, we can estimate the light power carried by each waveguide, regardless of distortions of the mode shape associated to
interfering stray light (estimated to be about 1\% of the peak intensity of the mode).
The properties of the best integrated beam combiner are given in Table~\ref{tab:beam_comb}.
In the left part of the table, we give the values of the normalized photometric outputs for waveguides A-D (columns) for individual input waveguide excitation (1-4, lines). Subsequently, the splitting ratios of the couplers were calculated (right part of the table) under the assumption that losses from all connecting waveguides are negligible. We reach nearly 50\% beam splitting ratios for all 3 beam splitters, the largest difference to that is only 2.5\%. Interestingly, the values in each of columns B and C differ slightly, depending on the injected input channel. The largest difference occurs at output B when switching between inputs 1 and 2. As the light passes the same paths after the beam splitter $\alpha$, one would expect the same flux ratios between output B and C. The difference is most probably due to an interference between coherent stray light, which differs between different excitation of the waveguides, and the guided signal in the waveguides.
\begin{table}[h]
\caption{Normalized output flux for single input injection and the calculated beam splitter efficiencies. See Fig.~\ref{fig:teles} for the definition of the variables, the flux ratio is given for the upper waveguide. The ratio and uncertainty of the splitter efficiency are the average and standard deviation of the output channels, respectively.}
\label{tab:beam_comb}
\begin{center}
\begin{tabular}{|r|cccc|l| @{\hspace{10mm}} rcl |}
\hline
Input & \multicolumn{4}{c|}{Output channels} & & \multicolumn{3}{c|}{beam splitter ratios} \\
channel & A & B & C & D & sum & beam splitter & ratio & uncertainty \\
\hline
1 & 0.495 & 0.258 & 0.247 & 0.001 & 1.001 & $\alpha$ & 0.503 & 0.011 \\
2 & 0.511 & 0.224 & 0.264 & 0.001 & 1.000 & $\beta$ & 0.476 & 0.004 \\
3 & 0.002 & 0.274 & 0.250 & 0.474 & 1.000 & $\gamma$ & 0.502 & 0.030 \\
4 & 0.000 & 0.270 & 0.251 & 0.479 & 1.000 & & & \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{INTERFEROMETRIC RESULTS}
The interferometric trace of the deep nulling outputs B and C is illustrated in Fig.~\ref{fig:outBC}.a.
As the angular coordinate advances (in our case, the time), we have an alternating sequence of narrow and broad minima for each of the outputs.
The logarithmic plot shows that the depth of the nulling signal ranges form 1:30 to 1:100.
This signal sequence is expected from the theory of the 4-telescope nulling. For applications to astronomy, the broad minima are the ones of interest, since they can mitigate the stellar leakage and the phase delay errors which would occur in a real world interferometer.
In fact, Angel \& Wolf predict that for small angular detunings $\theta$ from the center of the broad null the 4-wave interference signal should scale as $\theta^6$, thus suppressing efficiently the light over a large range of angles.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.48\textwidth]{null-d-sh8-ap10-sm3-lin.eps}
&
\includegraphics[width=0.48\textwidth]{null-d-sh8-ap10-sm3-log.eps}
\end{tabular}
\end{center}
\caption[Setup]
{ \label{fig:outBC}
Output B and C as a function of time, in linear (left) and logarithmic (right) scale. A smoothing over 3 sampling points (0.1 s) is applied to the data to reduce noise.}
\end{figure}
To test the power law describing the angular dependence of the interferometric signal in proximity of the center of the broad minimum, we plot the output signal of channel B as a function of the absolute value of the phase detuning (defined as $\phi=k\,\theta\,d$, $k$ being the wavenumber of light) in a bi-logarithmic plot (see Fig.~\ref{fig:loglogplot}). We found that for delays larger than 0.5 rad the interferometric signal follows a $\theta^4$ power law. For smaller delays, the interferometric signal is relaxing to a constant value of about 1:100. Simulations show that for a perfect combiner the $\theta^6$ behavior should become apparent for phase detunings smaller than 0.2-0.3 rad. At that point, however, the nulling depth would already reach levels below $10^{-4}$, requiring a detection system with a much higher dynamic range than the employed 10-bit CCD.
\begin{figure}[b]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{loglogplot.eps}
\end{tabular}
\end{center}
\caption
{ \label{fig:loglogplot} Bi-logarithmic plot showing the power dependence of the interferometric signal as a function of the detuning from the center of a broad minimum. The high background level prevents to observe the $\theta^6$ dependence of the interferometric signal. Overlaid in full line, the simulated fringe from an imperfect nulling beam combiner featuring the splitting ratios we measured in the chosen combiner.}
\end{figure}
\section{Discussion}
Even considering the reduced dynamics of our measurement system, the background level observed in the experiments is significantly higher than the noise floor, estimated in about 0.1-0.2\%. To account for this observation, there are a few possible causes. From one side, we have seen (see Tab.1) that the splitting ratio of the couplers is not the same and coupler $\beta$ deviates by nearly 3\% from the ideal value of 50\% splitting ratio.
Simulations show that fabrication errors that lead to a variation of the splitting ratio of the couplers can result indeed in a constant noise floor that can reduce the contrast of the fringes. One of such simulations featuring the measured coupling ratios is illustrated in Fig.~\ref{fig:loglogplot}. We see the appearance of a constant floor for $\phi<0.1$ rad, albeit with a much deeper null than observed in the experiment.
A second source of noise comes from the stray light which is generated by 1) the non-perfect coupling of the light into the waveguides due to mode mismatch and 2) radiation losses from the bended waveguides.
As seen in the logarithmic intensity maps displayed in Fig.~\ref{fig:output_field}, these two sources of stray light have an intensity of about 0.1\% the peak value of the guided mode intensity distribution. While the incoherent part of this background is taken into account by the employed photometric method (see Section 3), it is not possible to remove its coherent contribution to the signal.
Better control of this stray light may be accomplished by bending the input/output waveguides so that the injection light beam direction is not collinear to the observation direction, or by implanting suitable slits in the chip. Both methods were successfully used in the past to achieve deep nulling levels with integrated optics\cite{web04}.
Finally, we mention that other sources of noise may come from small photomertric and phase variations which prevent the achievement of a deep nulling. Indeed, small displacements of the beam induced by the phase modulators could alter the coupling ratio in the waveguide by a fraction of percent, thus preventing exact nulling at the waveguide couplers.
Also the fact that we do not control with a feedback-loop the optical path difference between the beams
could play a role. Previous characterization of a similar Mach-Zehnder interferometer installed in our lab, showed that the r.m.s. phase noise is about 0.08 rad\cite{min12}, or 8 nm in terms of OPD.
Better phase control could be achieved by using light sources of longer wavelength and/or a feedback-loop control of the OPD based on laser metrology\cite{mar12}.
\section{Conclusions}
We have characterized for the first time a 4-telescope integrated optics beam combiner for nulling interferometry.
In our setup we simulated the operation of a linear array of telescopes, as proposed by Angel\&Wolf \cite{ang97} in the context of exoplanet detection.
We showed that broad (in terms of the zenithal angle $\theta$) nulls can be generated, which could in principle mitigate the stellar leakage of partially resolved stars and/or exo-zodiacal light. While our experiments were primarily focused on the proof-of-concept of the IO component, we achieved a nulling depth from 1:30 to 1:100. Deeper nulling was prevented by a number of technical issues (some of them related to the test bench stability) which we will address in future experiments.
The importance of our work resides in demonstrating the potential of integrated optics for space-based, multi-telescope nulling interferometry. In fact, the use of integrated optical technology removes several degrees of freedom (alignment and positioning) which are present in bulk-optics setups\cite{mar12}, sensibly reducing the complexity and maintenance procedures of the beam combiner. This is of particular importance for space missions where automated alignment procedures are costly and subject to failure.
We highlight that the chosen manufacturing platform (direct laser writing) is suitable also for the production of integrated components in mid-infrared \cite{rod12}, where exo-planet detection is most favored.
\acknowledgments
RE would like to thank DFG for support in the Priority Programme SPP 1385 in project NE 515 / 34-1 and the Abbe School of photonics for the Ph.D. grant.
|
\section{Conclusion}
In this article, a map of heart-induced tissue displacements in the thorax is presented. Due to limited resolution, movements of scale less than 1 mm were not detected.
The largest heart movements were found in the anterior and left regions in the heart; the scale of these in-plane displacements were on the order of 1 cm and caused artery-lung boundary displacements on the order of 2-3mm. These induced displacements were far stronger in the left than right lungs due to the proximity to the heart left ventricular.
Mechanical coupling between the heart and aorta caused aorta centroid displacements with axial variations. This coupling could be seen to affect the aorta wall in a complex manner. However, due to lacking correlation with displacement, similarity with typical blood pressure curves and expected range of diameter variations, aorta radius variations was concluded to be dominated by internal pressure variations.
\section{Results}
\label{section:results}
The variation in aortic radius was virtually equal for all slices with very similar profiles (to within the noise of the measurements): this profile was saw-tooth shaped (\Figref{Aorta_size_var}). Neither the dicrotic notch nor shoulder was clearly apparent in these measurements - this may have been due to lack of segmentation accuracy.
Through the heart cycle, the aorta moved along a grossly diagonal path in the sagital planes as demonstrated by the aorta centroids and particularly so in the upper torso (correlation between $x$ and $y$ in \figref{Aorta_centroid_var}). This was due to the aorta being pushed against the spine which constrained the movements. The amplitude of the movement in $y$ is roughly 1.5 mm. Lower down the torso, a more complex pattern emerges.
The difference in mechanical coupling with the heart as a function of z-level does not appear as a difference in profile of the aorta size.
\Fig[-0.5 cm]{Overview}{1.1\ImageW}{Vc_encapsulated}{A sagittal section through the subject and which is centered on the aorta ($x = -1.0cm$) visible as a long, tube beneath the spine. The heart can be seen as an oval, oblique structure in the center with several visible heart chambers and just below the aorta. The positive $x$ axis points out of the image.}
\Fig[-1 cm]{Aorta_centroid_var}{1.2\ImageW}{aorta_centroid_variations_encapsulated}{Aorta centroid (top: $x$, bottom: $y$) displacements as a function of time through a heart cycle and subtracted mean values ($\mu_x$, $\mu_y$) to ease comparison. Upper z levels red; lower levels blue: $z \in \intervalCC{-6.9,2.3}$ cm.}
\Fig[-1 cm]{Aorta_size_var}{1.2\ImageW}{aorta_size_variations_encapsulated}{Aorta radius dilatations as a function of time through a heart cycle and subtracted mean values to ease comparison. Upper z levels red; lower levels blue: $z \in \intervalCC{-6.9,2.3}$ cm. The radius profiles indicate temporal shifts while descending the aorta: red traces appear to precede blue traces.}
\Fig{aortic_blood_pressure}{\ImageW}{aortic_waveforms_Human}{The series of blood pressure cycles is reproduced from Murakami \cite{Murakami2005} and was measured in a human while gradually removing the catheter from close to the aortic valve. Typical features of the aortic pressure wave is shown: the dicrotic notch at the closing of the aortic valve, the ``shoulder'' associated with the reflected wave from further down the aorta. Finally, the form of the pressure pulse changes as the reflection arrives at different time within the cycle. The actual blood pulse's shape depends on the individual.}
The accumulated edges $M_z$ shows large movements of the heart with induced movements in neighboring tissues, particularly large blood vessels in the lungs. These movements were strongest to the left, probably due to left ventricular activity. Edges appeared also to be detected between the compartments of the heart, and several of these internal interfaces show large displacements. The right side of the heart showed comparatively smaller displacements, therefore with smaller induced movements in neighboring tissues.
Very little traces of movements were detected between the lungs and the thoracic cage.
\Fig{Other-Movements}{\ImageW}{Edges_Ic}{To left, accumulated edges $M_z$; to right average intensities ${\expected[t]{J_{z,t}}}$. All images span the same $xy$ region; in images 1 through 5 $z$ was in the range $\intervalCC{-9.6,1.0}$ cm. In the edge images, the red lines represent boundaries that essentially did not move (more than 24 of 30 images in the same point). Sets of neighboring, similar curves are likely due to tissue boundaries that were moving. The blue arrows are qualitative measures of movements based on edge ``movies'' ($E_{z,t}$ as time sequences). In image 1, blood flow in the aorta created artifacts.}
\section{Acknowledgments}
This work is part of the MELODY project, which is funded by the Research Council of Norway under the contract number 187857/S10.
The Department of Radiology and Nuclear Medicine, Oslo University Hospital in Oslo, Norway provided the MR images.
\section{Introduction}
This article addresses the movements of tissues in a human thorax as detected by magnetic resonance images (MRI) of a single, healthy individual. Motion in the thorax is primarily due to respiration of the lungs and to the beating heart.
The motion induced by respiration is extensively studied in the literature. Several studies have quantified the precision of motion compensation in imaging processes \cite{Siebenthal2007,Hinkle2009,McLeish2008,Paganelli2013}, although a quantized map of movements is generally not presented. F. Odille and colleagues have, however, presented a map in a sagital plane for one study \cite{Odille2008}. Larger movements due to respiration is found along the superior-inferior direction with anterior-posterior being the second most significant axis \cite{Brandner2006} \cite{Weber2009b}, \cite{Wang1995}, \cite{Langen2001}.
Dilatations of the aorta are found to be due to the blood flow through the aorta, and the aortic displacements are both a consequence of respiration and the heart beats \cite{Biesdorf2011}, \cite{Rengier2012}, \cite{Weber2009}. Motion induced specifically by the heart activity, e.g. while suspending respiration, is reviewed in \cite{Scott2009} with an emphasis on improving cardiac imaging, however, no map is provided.
While investigating radar-based techniques for estimating aortic blood pressure based on characteristics of the aorta geometry a map in the axial plane of heart-induced motions was needed, as was a more detailed description of the aortic behavior. We have not been successful in finding an adequate map in the literature. This article hence provides a simple map of organ boundary movements in general and aortic dilatations and displacements in particular. During all acquisitions, respiration was suspended.
\section{Discussion}
The largest movements were observed in the heart and dominated by the left ventricular. Outside the heart, the tissue boundaries displaced by the heart were generally those between lungs and arteries. This coupling has two possible causes: the blood flow through the arterial tree and the pressure propagation through intermediate tissues. Given that the blood flow through the arterial tree should act similarly in the left and right sides while apparent movements were strongest to the left, the dominant effect was interpreted to be due to the propagation of the left ventricular movements.
Regarding the scale of in-plane displacements of the heart walls, large movements were generally on the order of 0.5 cm to 1 cm with the largest exceeding 1 cm, perhaps as much as 2 cm (point (6) in image 3, \figref{Other-Movements}). The displacements were typically divided into 4 phases: long even compression, long retraction, still, short retraction. Outside the heart, the displacement scales were significantly less and the stronger, observable displacements appeared to be on the order of 2-3 mm.
Although the scale of the lung arteries' movements are greater than that of the aorta, the former constitutes a chaotic tree structure with relatively thin branches as opposed to the aorta which is larger and runs almost straight down the back.
Close to the aorta, heart wall displacements appeared to have lesser amplitude. This movement was coupled with the aorta dilatations and in some instances (e.g. images 3, 4 in \figref{Other-Movements}) resulted in complex deformations and displacements. This was probably due to an interaction between internal blood pressure pulses and the mechanical coupling.
The segmentation of the aorta resulted in measures of both radius and centroid. The segmentation followed a semi-automatic method and although the aorta was a fairly easy organ to extract in MRIs, artifacts are known to affect the images in particular due to strong blood flow or any large movement in general. This was also the reason why the lower-most z-levels were excluded from \figsref{Aorta_centroid_var} and \ref{figure:Aorta_size_var}. For those slices included in these figures, the continuity of the traces across slices and relatively low noise along a trace both indicated an acceptable precision of the segmentation.
As mentioned in the previous section, the aorta centroid followed a grossly diagonal path, especially at upper z levels. In \figref{Other-Movements}, the aorta is located between the heart wall and the spine - at least in the upper z levels. Heart pulsations exhorted pressure on the aorta which in turn was constrained by the spine: this may explain the diagonal movement. The tendency to adjust to the heart pressure is seen in \figref{Other-Movements}: the accumulated edges indicates stronger movements of the aorta wall along this SW-NE diagonal. These movements were partially correlated with the aorta size: compare the relative $x$ movements with aorta radius. The reason for this correlation was likely the small lag between the blood pressure pulse wave through the aorta and the mechanical coupling with the nearby heart walls.
In \cite{Stefanidis1995} the radius dilatations of the aorta were measured using a precise and invasive method based on pressure and diameter sensors introduced through catheters. It concluded that typical radius peak-to-peak amplitudes for a normal population is 1.09 $\pm$ 0.22 mm. In \cite{Rengier2012}, similar results were obtained at several locations along the aorta. The scale of dilatations in \figref{Aorta_size_var} are of the same size.
The linear relationship between relative changes of blood pressure and cross sectional area was studied in \cite{Sugawara2000}: it provided a link between the radius changes and blood pressure variations. A similar study \cite{Olsen1972} also shows the similarity of pressure and diameter using trans-oesophagal ultrasound. The form of the aorta radius dilatations was similar to common central arterial blood pressure curves (\figref{aortic_blood_pressure}).
Furthermore, the lack of correlation between aorta size and $y$ position as functions of the $z$ level pointed out in the results section proved at least that size and position are partially decorrelated. Given the lack of visible influence, if not independent, the influence of $y$ on size was small. Finally, we saw no reason to assume that the influence of $x$ on aorta size should be stronger then that of $y$.
Therefore, we concluded that the relationship between size and internal pressure was likely to be valid in spite of the pressure exerted on the aorta from the heart.
The analysis in this article has been based on sequences of MRI images. Given that the in-plane resolution of the images were approximately 1 mm, the scale of movements detectable in the accumulated edges images was on the order of this resolution. Given that the aorta was segmented in each MRI acquisition and that the measures of size and centroid were based on the entire segmented regions, their accuracy was significantly higher than the 1 mm bound.
The observations of this article relied only on a single individual in prone position. The posture of a given individual is likely to affect certain aspects - in particular how the aorta is squeezed between the heart and the spine. However, the general map of displacements is likely to remain valid.
\section{Methods}
\label{section:methods}
Two sets of magnetic resonance images (MRI) were recorded: one set (I) for constructing a 3D volume representation; one set of axial planes (II) for analyzing tissue movements through the heart cycle. A three Tesla MR system (Achieva, Phillips Healthcare, Best, The Netherlands) was used. In this presentation, a right-handed coordinate system was used with 'x' pointing from left to right (lateral), 'y' pointing from posterior towards anterior (depth) and 'z' pointing from lower-back towards the neck (longitudinal) (\figref{Overview}).
In set I, a voxel element spanned a volume of $0.89 \times 0.89 \times 1 \unit{\mbox{mm}^3}$ using a transverse, Turbo Field Echo procedure. For image acquisition the respiration was not suspended, but synchronization with respiration was implemented at close to minimal inhalation.
For set II, a balanced, Turbo Field Echo procedure was used where the heart cycle was divided into 30 phases, axially oriented, for each of the longitudinal positions. Each slice had a 2D slice pixel size of $0.85 \times 0.85 \unit{\mbox{mm}^2}$, at 6.6 mm slice spacing from just below the heart to slightly below the neck. For each acquisition, respiration was suspended; however, only a partial inhalation was done in order to allow for good ECG quality.
In both sets, MRI images were recorded in synchronization with the ECG signal and the sets were both T1-weighted.
The images of set II were further segmented with relation to the aorta in order to have precise ``position'' and ``size'' data. The segmentation was based on a manually initialized, elliptic seed with automatic growth in a transformed neighborhood while attempting to maintain region homogeneity. The ``position'' of the aorta was given through the centroid of the segmented pixels in each image. The ``size'' was defined as the radius of the circle with equal area as the segmented ellipsis: $\hat{r} = \sqrt{ab}$, where $a,b$ are the minor and major axes. This procedure was not considered successful in the lower most axial slices: in these slices, strong artifacts in the MR images due to blood flow precluded reliable segmentation.
For other structures, segmentation was considered too difficult and a different approach was adopted in order to illustrate the tissue displacements: this was based on the extraction of intensity gradients (edges) and which were assumed to be tissue boundaries. First the edges ($E$) of the compressed MRI images ($J$) in set II ($S$) were calculated; these are binary images. These were then accumulated ($M$) over the set of time instants.
\eqn{\forall I_{z,t}\in S, \s J_{z,t} &=& \sqrt{I_{z,t}} \in \R^{N_x \times N_y} \\
E_{z,t} &=& edge(J_{z,t};\theta,\sigma) \in \set{0,1}^{N_x \times N_y} \\
M_{z} &=& \Sigma_{t=1}^{N_t} E_{z,t} \in \set{0,\cdots,N_t}^{N_x \times N_y}}
where $t$ is here an index into the heart cycle (with $N_t$ instants) and the variables are of size $N_x \times N_y$. The edge detection was the Canny edge detector, with thresholds set by $\theta$ and size set by $\sigma$.
Ideally, movements created sets of similar edges in a neighborhood with extent depending on the scale of movement, and static edges would all be identified at the same point $(x,y)$ and $M_{z}(x,y)=N_t$. In practice, the accumulated edges are not unambiguous indications of movements (see results in \figref{Other-Movements}):
\enmbe[i]{\label{es:spurr}``thin'', gray lines represented spurious detections (e.g. point(3)),
\item\label{es:ill_def} a dark gray, narrow ($\approx$2 pixels) line may represent ill-defined boundaries (examples along the lung-thoracic cage boundary),
\item \label{es:plane} intensity changes and boundaries which passed in/out of the plane may have created numerous, ``parallel'' lines which nonetheless did not represent the movement of a single boundary (e.g. point(6)).
}
Regarding \ref{es:spurr} and \ref{es:ill_def}, edges with poor quality were not always detected, or not always at the same position.
The accumulated edges image was chosen to represent movements because it allowed the sequence of edges to be displayed in a single two-dimensional image. A movie representation avoids some of the aforementioned problems and was the basis of adding qualitative indications of displacements in the form of blue arrows to these images.
Remark, displacements assessed in a two dimensional plane can only provide reliable estimations of the component in the plane, not orthogonal to the plane (z): actual boundary movements were necessarily less.
|
\section{Introduction}
The tachyon field can play an important role in inflationary models \cite{Pad}-\cite{Jain} as well as in the present accelerated expansion, simulating the effect of the dark energy \cite{Pad}-\cite{Abramo}, \cite{Aguirre}-\cite{LuisTaq2} depending upon the form of the tachyon potential \cite{Pad}-\cite{Abramo}, \cite{Bagla}-\cite{Gilberto}. The tachyon is an unstable field which has becomes important in string theory through its role in the Dirac-Born-Infeld Lagrangian, because it is used to describe the D-brane action \cite{TaS}. It was shown that the tachyon field could play a useful role in cosmology independent of the fact that it can be an unstable field \cite{TaC}. Besides, it was pointed out in \cite {Pad} that the tachyon Lagrangian can be accommodated into a quintessence form when the derivates of the fields are small.
Several years ago, it was proposed that the tachyon Lagrangian could be extended in such a way to allow the barotropic index takes any value \cite{LuisTE} generating new species of tachyon called phantom and complementary tachyons in addition to the ordinary one \cite{phank}, \cite{LuisTaq2}, \cite{LuisTE}. The standard tachyon field can also describe a transition from an accelerated to a decelerated regime, behaving as an inflaton field at early times and as a matter field at late times. The complementary tachyon field always behaves as a matter field. The phantom tachyon field is characterized by a rapid expansion where its energy density increases with time \cite{LuisTaq2}, \cite{Shi}-\cite{Novos}.
On the other hand, form invariance transformations involve internal or external variables in such a way that the transformations preserve the form of the dynamical equations, i.e., they have a form invariance symmetry (FIS) \cite{symm}. Particularly useful are the T-duality \cite{Polchinsky} or ``scale-factor duality'' \cite{VET}.
A new kind of internal symmetry that preserves the form of the spatially flat Friedmann cosmology was found by one of the authors \cite{Luis}-\cite{upc}. There it was shown that the equations governing the evolution of FRW cosmologies have a FIS group. The FIT which preserves the form of those equations relates quantities of the fluid, energy density and pressure, with geometrical quantities such as the scale factor and Hubble expansion rate. The FIS introduce an alternative concept of equivalence between different physical problems meaning that essentially a set of cosmological models are equivalent when their dynamical equations are form invariant under the action of some internal symmetry group \cite{FIS}.
The FIS makes possible to find exact solutions in several contexts and generate new cosmologies from a known ``seed'' one \cite{upc}-\cite{CM1}.
In this paper we will show that the FIT applied to the standard tachyon field, used as a seed, can generate the complementary tachyon field and the phantom tachyon field. Our main goal is to show that the extended tachyon field, is a consequence of the internal symmetry that preserves the form of the Einstein equations, in a FRW space-time. Additionally we will shown that the FIT allow us to pass from a non-stable cosmology to a stable one and vice-versa \cite{Gibb}, \cite{Barrow}. In particular we will analyze the tachyon field, driven by a potential depending inversely on the square of the scalar field. We will start with some seed cosmology and use FIT to obtain a different new one, for example passing form a accelerated to a super-accelerated scenario.
\section{FIS in flat FRW cosmology and linear FIT}
We will investigate an internal symmetry contained in the Einstein equations for a spatially flat FRW space-time
\begin{equation}
\label{E1a}
3H^{2}=\rho\,
\end{equation}
\begin{equation}
\label{E1b}
\dot{\rho}+ 3H(\rho+p)=0\,
\end{equation}
where $H=\dot a/a$ is the Hubble expansion rate and $a(t)$ is the scale factor. We assume that the universe is filled with a perfect fluid having energy density $\rho$ and pressure $p$. The two independent Einstein equations have three unknown quantities $(H, p, \rho)$, hence the system of equations (\ref{E1a})-(\ref{E1b}) has one degree of freedom. This allows to introduce FIT which involves those quantities,
\begin{equation}
\label{tr}
\bar\rho=\bar\rho(\rho),
\end{equation}
\begin{equation}
\label{th}
\bar H=\left(\frac{\bar\rho}{\rho}\right)^{1/2}H,
\end{equation}
\begin{equation}
\label{tp+r}
\bar p+\bar\rho=\left(\frac{\rho}{\bar\rho}\right)^{1/2}\frac{d\bar\rho}{d\rho}\,\,(\rho+p).
\end{equation}
Hence, the FIT (\ref{tr})-(\ref{tp+r}), generated by the invertible function $\bar\rho(\rho)$, make the job of preserving the form of the system of equations (\ref{E1a})-(\ref{E1b}) and the FRW cosmology has a FIS. The FIT (\ref{tr})-(\ref{tp+r}) map solutions of a define cosmology, through the variables $(H,p,\rho)$, into solutions of other system of equations, defining a different cosmology identified with the barred variables $(\bar{H}, \bar{p},\bar{\rho})$, forming a Lie group structure as is demonstrated in \cite{FIS}.
We present FIT induced by the linear generating function $\bar\rho=n^2\rho$ being $n$ a constant. After this choice Eqs. (\ref{tr})-(\ref{tp+r}) become
\begin{equation}
\label{trn}
\bar\rho=n^2\rho,
\end{equation}
\begin{equation}
\label{thn}
\bar H=nH,\quad \Rightarrow \quad \bar a=a^n,
\end{equation}
\begin{equation}
\label{tpn}
(\bar\rho+\bar p)=n(\rho+p).
\end{equation}
Hence, the linear transformation (\ref{trn}) leads to a linear combinations of the variables $\rho, H, p$ and a power transformation of the scale factor, obtained after having integrated $\bar H=nH$. Finally, the Eq. (\ref{tpn}) gives the transformation rule for the pressure of the fluid
\begin{equation}
\label{tp}
\bar p=-n^2\rho+n(\rho+p).
\end{equation}
In the case of considering two universes, each one of them filled with a perfect fluid for which we assume equations of state $\bar p=(\bar\gamma-1)\bar\rho$ and $p=(\gamma-1)\rho$ respectively, the barotropic index $\gamma$ transforms as
\begin{equation}
\label{tg}
\bar\gamma=\frac{(\bar\rho+\bar p)}{\bar\rho}=\frac{\rho+p}{n\,\rho}=\frac{\gamma}{n},
\end{equation}
after using Eq. (\ref{trn}) along with Eq. (\ref{tpn}).
The existence of a Lie group structure opens the possibility of connecting the scale factor $a$ of a seed cosmology with the scale factor $\bar a =a^n$ of a different cosmology.
\section{The extended tachyon cosmology}
We turn our attention to the tachyon field and will show how it transforms under the FIT (\ref{trn})-(\ref{tpn}). We consider a scalar field $\phi$ of the tachyon-type with the self-interaction potential $V (\phi)$. The background energy density and pressure of the tachyon condensate, for a flat FRW cosmology, are
\begin{equation}
\rho_\phi=\frac{V}{\sqrt{1-\dot{\phi}^2}}, \qquad \quad p_\phi=-V\sqrt{1-\dot{\phi}^2}, \label{roE}
\end{equation}
respectively. The corresponding Einstein-Klein-Gordon (EKG) equations are
\begin{equation}
\label{A}
3H^2=\frac{V}{\sqrt{1-\dot{\phi}^2}},
\end{equation}
\begin{equation}
\label{kg}
\ddot\phi+3H\dot\phi(1-\dot{\phi}^2)+\frac{1-\dot{\phi}^2}{V}\frac{dV}{d\phi}=0.
\end{equation}
The equation of state for the tachyon is $p=(\gamma -1))\rho$, so the barotropic index is
\begin{equation}
\label{gs}
\gamma=\dot\phi^2.
\end{equation}
with $0<\gamma<1$ for Eqs. (\ref{roE}). The sound speed is $c^{2}_{s}=1-\gamma>0$, and using (\ref{gs}), we can write
\begin{equation}
\label{sos}
c^{2}_{s}=1-\dot\phi^2.
\end{equation}
From equations (\ref{trn}) and (\ref{tp}), the transformed energy density and pressure of the tachyon field are given by
\begin{equation}
\label{trt}
\bar{\rho}=\frac{\bar{V}}{\sqrt{1-\dot{\bar{\phi}}^2}}=\frac{n^{2}V}{\sqrt{1-\dot{\phi}^2}},
\end{equation}
\begin{equation}
\label{tpt}
\bar{p}=-\bar{V}\sqrt{1-\dot{\bar{\phi}}^2}=-\left(1-\frac{\dot{\phi}^2}{n}\right)\frac{n^{2}V}{\sqrt{1-\dot{\phi}^2}}.
\end{equation}
So, we find that the tachyon field, the potential, the barotropic index and the sound speed transform linearly under the FIT (\ref{trn})-(\ref{tpn}),
\begin{equation}
\label{tfv}
\dot{\bar{\phi}}^{2} =\frac{\dot{\phi}^{2}}{n}, \qquad \bar{V}=n^2 V \sqrt{\frac{1-\dot{\phi}^{2}/n}{1-\dot{\phi}^{2}}},
\end{equation}
\begin{equation}
\label{tgs}
\bar\gamma=\frac{\gamma}{n}, \qquad \bar{c}^{2}_{s}=\frac{n-\dot\phi^2}{n}.
\end{equation}
and the scalar field transforms as $\bar\phi=\phi/\sqrt{n}$.
We consider the Eqs. (\ref{roE}) with a barotropic index $0<\gamma<1$ as a seed tachyon, called the ordinary tachyon. Using the first equation of (\ref{tgs}) we can get a barotropic index $\bar{\gamma}=\gamma/n<0$. Then, the energy density and the pressure of the barred cosmology are given by
\begin{equation}
\label{trtp}
\bar{\rho}=\frac{\bar{V}}{\sqrt{1+\dot{\bar{\phi}}^2}}, \qquad \bar{p}=-\bar{V}\sqrt{1+\dot{\bar{\phi}}^2},
\end{equation}
These fluids represented by the Eq. (\ref{trtp}) with negative pressure and negative barotropic index describe phantom cosmologies. Moreover, we can get $1<\bar{\gamma}=\gamma/n$ under de condition $n<\gamma$ applying the transformed rule (\ref{tgs}) to the barotropic index of the seed tachyon $0<\gamma<1$. So, the energy density and the pressure of the barred fluid are
\begin{equation}
\label{trtc}
\bar{\rho}=\frac{i\mid\bar{V}\mid}{i\sqrt{\dot{\bar{\phi}}^{2}-1}}=\frac{\mid\bar{V}\mid}{\sqrt{\dot{\bar{\phi}}^{2}-1}},
\end{equation}
\begin{equation}
\label{tptc}
\bar{p}=-i\mid\bar{V}\mid i\sqrt{\dot{\bar{\phi}}^2-1}=\mid\bar{V}\mid \sqrt{\dot{\bar{\phi}}^2-1},
\end{equation}
while, these fluids described by the Eqs. (\ref{trtc}) and (\ref{tptc}) give rise to nonaccelerated expanding evolutions.
We used the ordinary tachyon field, Eqs. (\ref{roE}) with $0<\gamma<1$, as a seed. With the application of the FIS Eqs. (\ref{trt})-(\ref{tgs}) we found the two species of tachyon fields, as in \cite{LuisTE}, the phantom tachyon Eqs. (\ref{trtp}) with a $\gamma<0$ and the complementary tachyon Eqs. (\ref{trtc}) and (\ref{tptc}) with $1<\gamma$. Therefore, the form invariance transformations allow us to extend the family of tachyon field.
Following Gibbons \cite{Gibb} and Barrow \textit{et. al.} \cite{Barrow}, the Einstein static universe containing a perfect fluid is always neutrally stable for the condition $c^{2}_{s}>1/5$. Therefore, the FIT Eq. (\ref{tgs}) allow us to pass from a non-stable cosmology to a stable one and vice-versa. For example, if we use a barotropic index $\gamma_{0}=6/7$ as a seed solution with $c^{2}_{s}=1/7$, using the transformation rule Eq. (\ref{tgs}), we can get a stable cosmology with $c^{2}_{s}=5/7>1/5$ if $n=3$.
\subsection{Power-law expansion for the tachyon field}
Let us assume that the potential is an inverse square in terms of the tachyon field,
\begin{equation}
\label{pot}
V(\phi)=\frac{V_0}{\phi^2},
\end{equation}
with $V_0$ a constant. This potential, that diverges at $\phi=0$, fairly mimics the behavior of a typical potential in the condensate of bosonic string theory. The Eq. (\ref{pot}), leads to the power law expansion $a(t)=kt^{\delta}$, with $k$ a constant, if $\phi$ is the only source \cite{Pad} \cite{FE}. The tachyon field and the barotropic index are
\begin{equation}
\label{CTqPL}
\phi=\left(\frac{2}{3\delta}\right)^{1/2}t, \qquad 0<\gamma_{0}<1,
\end{equation}
with
\begin{equation}
\label{DTqPL}
\delta=\frac{1}{3}\left[1+\sqrt{1+4\beta}\right], \qquad \beta=\left(\frac{3V_{0}}{4}\right)^{2}.
\end{equation}
For this reason the power law expansion appears to be a good example to illustrate how from a seed solution, characterized by particular values of the parameters $V_0$, $\gamma_0$ and $k$, the FIS helps us to find the scalar field and the scale factor driven by inverse square potential (\ref{pot}) for any other value of those parameters. Applying the FIT (\ref{trn})-(\ref{tpn}) to the seed solution (\ref{CTqPL}), (\ref{DTqPL}) and using Eqs. (\ref{gs}), (\ref{tfv}) and (\ref{tgs}) we obtain the transformation rules for $V_0$ and $\gamma_0$
\begin{equation}
\label{ga0}
\bar{\gamma_0}=\frac{\gamma_0}{n},
\end{equation}
\begin{equation}
\label{V0}
\bar{V}_{0}= nV_{0}\sqrt{\frac{1-\frac{\gamma_0}{n}}{1-\gamma_0}}.
\end{equation}
Therefore, the transformed tachyon field, for a barotropic index $\bar{\gamma}<0$, is given by
\begin{equation}
\label{plp}
\bar{\phi}=\left(\frac{2}{-3\mid\bar{\delta}\mid}\right)^{1/2}t, \qquad \bar{\delta}=\frac{1}{3}\left[1-\sqrt{1+4\bar{\beta}}\right].
\end{equation}
These tachyon field solutions, Eq. (\ref{plp}), describe phantom cosmologies. Note that if $n=-1$ in Eqs. (\ref{ga0}) and (\ref{V0}) we can get the results of \cite{phank} for the phantom tachyon.
On the other hand, if the transformed barotropic index is $1<\bar{\gamma}$, we get
\begin{equation}
\label{plc}
\bar{\phi}=\left(\frac{2}{3\bar{\delta}}\right)^{1/2}t, \qquad \bar{\delta}=\frac{1}{3}\left[1\pm\sqrt{1-4\mid\bar{\beta}\mid}\right].
\end{equation}
This type of tachyon field solution with $1<\gamma_{0}$ is called the complementary tachyon solution, which represents stiff matter with a deceleration cosmology.
The scale factor $a(t)=kt^{\delta}$ transforms as $\bar{a}=a^{n}$, so the transformed scalar field is $\bar{a}=\bar{k}t^{\bar\delta}$ with $\bar{k}=k^{n}$ and $\bar{\delta}=n\delta$. The condition to has an inflation solution is that $1<\delta$ and it is represented by the solutions (\ref{CTqPL}). Notice that the exponent of the power law solution can takes positive or negative values provided that $n \in \Re$. We can see that this exponent is directly related with the barotropic index of the tachyon fluid, $\delta=2/3\gamma_0$, therefore changing $n$ it is equivalent to allow that the $\bar{\gamma}_0$ varies over $\Re$. This simple fact leads us to a remarkable conclusion, there are new species of tachyons and the FIS has revealed their existence to us.
\section{conclusion}
As part of a long-term investigation \cite{Luis}-\cite{FIS} we have shown here that form invariance transformations can be used as tools for generating new solutions to the Einstein field equations, in this case the existence of two new kinds of extended tachyon fields were derived from the standard tachyon field ($0<\gamma<1$): the complementary ($1<\gamma$) and the phantom tachyon ($\gamma<0$) fields, confirming the work made by one of the authors \cite{LuisTE}. In addition we see that the form invariance transformations allow us to pass from a neutrally unstable universe to a stable one \cite{Gibb}, \cite{Barrow}.
In particular, we have applied the method to obtaining phantom and complementary versions of FRW tachyon cosmologies, with an accent on power-law space-times generated by an inverse-square potential. We have found that the FIT transform the seed scale factor $a=kt^{\delta}$ into the power law solution $a=k^{n}t^{n\delta}$. For illustration purposes, if we start from a decelerated model with $2/3<\delta<1$, we can get a power-law inflation model with $\delta>1$ or a super-accelerated model (phantom model) with $\delta<0$. So, we have shown how FIT generate new cosmologies from a seed one.
\acknowledgments
This work was supported by Doctoral program of the Consejo Nacional de Investigaciones Cient\'{\i}ficas y T\' ecnicas (CONICET). I.S.G thanks IMAS, Math. Department, FyCEN-University of Buenos Aires. Also acknowledges Dr. Guillem Per\' ez-Nadal for useful discussions.
\vskip 1cm
|
\section{Introduction}
In this paper, we study a class of models that combine several mechanisms of dissipation:
plasticity, visco-plasticity, visco-elasticity and fracture.
Our goal is to investigate whether models from solid mechanics could be pertinent to describe geophysical materials
(and particularly the lithosphere on continental scales), as advocated by Peltzer and Tapponnier \cite{PeltzerTapponnier}, while others (see for example~\cite{England1}, \cite{England2})
prefer descriptions based on fluid mechanics.
The solid mechanics approach would have advantage to account for cracks in the formation of geological faults. They illustrate their claim with analogue experiments, where a rigid indenter deforms a layer of plasticine, to model the action of the Indian sub-continent on
the Tibetan plateau.
The plasticine experiments seems to reproduce the geophysical scenario of creation of the Asian faults and the extrusion of the South-Asian block (including Vietnam).
No general consensus prevails on the modeling of crack initiation and propagation, even in homogeneous materials.
The popular Griffith model, much in use in the engineering community, suffers from various shortcomings.
For instance it does not account for crack nucleation, and assumes pre-determined crack paths.
In the last decade, a series of investigation initiated by Francfort and Marigo \cite{FrancfortMarigo}
has addressed the mathematical foundation of fracture mechanics, using new concepts that have emerged
from the mathematical modeling of composite materials and from the calculus of variation.
This approach postulates that crack evolution is governed by the minimization of a total energy, among all
possible crack states.
Our models of fracture are inspired by this work, though we only consider fracture via a phase-field approximation.
In other words, the geometry of possible cracks is captured by a function $v$ with values between $0$ and $1$,
$v=1$ in the healthy parts that do not contain cracks. The length of the cracks, a quantity that contributes to the
total energy, is approximated via a functionnal introduced by Ambrosio and Tortorelli~\cite{Ambrosio} and Bourdin~\cite{Bourdin}.
The numerical simulations of fracture in a purely elastic medium, using such phase-field approximation,
was carried out by Bourdin, Francfort and Marigo~\cite{Bourdin2}, \cite{BourdinFrancfortMarigo1}, \cite{BourdinFrancfortMarigo}.
A model combining elasticity, visco-elasticity and fracture regularized via phase-field, is analyzed in~\cite{Larsen},
where the main point is how to define a consistent evolution as the limit of semi-discrete approximations
in time.
Our class of models extends this work to the case when plastic behavior and viscoplastic behavior can occur.
From a thermodynamical point of view, we interpret the phase field function $v$, that tracks the location and
propagation of cracks, not only as a variable for numerical approximation,
but as a global thermodynamical internal variable.
We show that our models are consistent with thermodynamics, in the
sense that they satisfy a Clausius-Duhem type inequality.
We propose a numerical scheme for a space-time discretization of the evolution,
and analyze its advantages and shortcomings on 1D et 2D traction experiments and on the experiment by Peltzer and Tapponnier~\cite{PeltzerTapponnier}.
The paper is organized as follows. In Section 2, we describe the proposed models
with regularized fracture and define their evolution in time.
Section 3 is dedicated to showing that they satisfy a Clausius-Duhem type inequality.
In Section 4, we introduce a semi-discrete time evolution, which is the base, in the final section, for numerical experiments
in the case of 1D, 2D traction and 2D plasticine experiment.
In particular we show that several dissipation mechanisms can be expressed according to the choice of parameters.
\section{Description of models with several dissipation mechanisms.}
\subsection{Notations.}
Throughout the paper, $\Omega$ denotes a bounded connected open set in $\mathbb{R}^{2}$ with Lipschitz boundary
$\partial\Omega=\partial\Omega_D\cup\partial\Omega_N$, where $\partial\Omega_D,\, \partial\Omega_N$ are
disjoint measurable sets. We denote time derivatives with a dot and $\operatornamewithlimits{argmin}_{v\in V}\mathcal{F}(v)$ denotes a function $u$ that minimizes
$\mathcal{F}$ over $V$.
Given $T_f>0$, we denote by $L^p((0,T_f),X)$, $W^{k,p}((0,T_f),X)$, the Lebesgue and Sobolev spaces involving time [see \cite{Evans} p. 285],
where X is a Banach space. The set of symmetric $2 \times 2$ matrices is denoted by $\mathbb{M}_{\text{sym}}^{2\times2}$ .
For $\xi,\zeta \in \mathbb{M}_{\text{sym}}^{2\times2}$ we define the scalar product between matrices
$\zeta:\xi:=\sum_{ij}\zeta_{ij}\xi_{ij}$, and the associated matrix norm by $|\xi|:=\sqrt{\xi:\xi}$.
Let A be the fourth order tensor of Lam\'e coefficients and B a suitable symmetric-fourth order tensor.
We assume that for some constants $0<\alpha_1 \leq\ \alpha_2<\infty$, they satisfy the ellipticity conditions
\begin{eqnarray*}
\forall\; e \in \mathbb{M}_{\text{sym}}^{2\times2},\quad \alpha_1|e|^{2}\leq Ae:e\leq \alpha_2|e|^{2}\quad\text{and}\quad\alpha_1|e|^{2}\leq Be:e\leq \alpha_2|e|^{2}
\end{eqnarray*}
The mechanical unknowns of our model are the displacement field $ u: \Omega \times[0,T_f]\rightarrow \mathbb{R}^2$,
the elastic strain $e: \Omega \times[0,T_f]\rightarrow \mathbb{M}_{\text{sym}}^{2\times2}$, the plastic strain $p: \Omega \times[0,T_f]\rightarrow \mathbb{M}_{\text{sym}}^{2\times2}$.
We assume $u$ and $\nabla u$ remain small. So that the relation between the deformation tensor $E$ and the displacement field is given by
\begin{eqnarray*}
Eu := \frac{1}{2}(\nabla u + \nabla u^T).
\end{eqnarray*}
We also assume that $Eu$ decomposes as an elastic part and
a plastic part
\begin{eqnarray*}
Eu &=& e + p.
\end{eqnarray*}
For $w \in H^1(0,T_f, H^1(\Omega, \mathbb{R}^2))$, which represents an applied boundary displacement, we define for $t\in[0,T_f]$ the set of kinematically admissible fields by
\begin{eqnarray*}
A_{adm}(w(t)) &:=& \{(u,e,p) \in H^{1}(\Omega,\mathbb{R}^{2}) \times L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\times L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})~: \\
&& Eu= e+p\, \quad a.e.\,\, \text{in} \,\, \Omega,\;u=w(t) \quad a.e.\;\; \text{on} \; \; \partial \Omega_D\}.
\end{eqnarray*}
For $f \in C^1([0,T_f], L^2(\Omega)^2)$, and $g \in C^1([0,T_f],H^{-1/2}(\Omega)^2)$,
the external forces at time $t \in [0,T_f]$ are collected into
$$
\langle l(t), u \rangle := \int_\Omega f(t).u\,dx+\int_{\partial\Omega_N} g(t).u\,ds.
$$
For a fixed constant $\tau>0$, we define $\mathbb{K}:=\{q\in\mathbb{M}_{\text{sym}}^{2\times2};\, |q|\leq \tau\quad a.e.\, \text{in} \, \Omega\}$.
We define $H:\mathbb{M}_{\text{sym}}^{2\times2}\rightarrow[0,\infty]$ the support function of $\mathbb{K}$ by
$$
H(p):=\sup_{\theta\in \mathbb{K}}\,\, \theta:p=\tau|p|,
$$
and a perturbed dissipation potential $H_{\beta}$ by
$$
H_{\beta}(p):=H(p)+\dfrac{\beta}{2}|p|^2,
$$
where $\beta>0$ plays the role of a regularization parameter.
The variational approach to fracture \cite{FrancfortMarigo}, \cite{BourdinFrancfortMarigo}
is based on Griffith's idea that the crack growth and crack path are determined by the competition between
the elastic energy release, when the crack increases, and the energy
dissipated to create a new crack.
We approximate the fracture (see Figure~\ref{rupture}) by a phase field function
$v: \Omega\times[0,T_f]\rightarrow [0,1]$ that depends on two parameters:
\begin{itemize}
\item $\epsilon>0$, the parameter of space regularization, relates to the width of the generalized fracture,
\item $\eta>0$ is a parameter, that preserves the ellipticity of the elastic energy. In \cite{Ambrosio}, $\eta$ scales as $o(\varepsilon)$ as $\varepsilon\rightarrow0$
in the approximation of a true crack by a phase-field function.
\end{itemize}
The nucleation and propagation of cracks, and the material deformation
result from minimizing at each time a global energy, that contains several terms:
\begin{eqnarray*}
\mathcal{E}_{total}:= \mathcal{E}_{el}+ \mathcal{E}_{p}+ \mathcal{E}_{h}+ \mathcal{E}_{ve}+ \mathcal{E}_{vp}+ \mathcal{E}_{S}.
\end{eqnarray*}
The elastic energy is defined as
\begin{align*}
&\mathcal{E}_{el}\,:\, L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2}) \times H^1(\Omega,\mathbb{R})\rightarrow \mathbb{R}\\
&(e,v)\longmapsto \mathcal{E}_{el}(e,v) = \dfrac{1}{2} \int_{\Omega}\left( v^{2}+\eta\right)Ae:e \,dx.
\end{align*}
The plastic dissipated energy is defined, by
\begin{align*}
&\mathcal{E}_{p}\,:\, L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\times L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\rightarrow \mathbb{R}\\
&(p,p_0)\longmapsto \mathcal{E}_{p}(p,p_0) = \int_{\Omega}H(p-p_0)\,dx,
\end{align*}
and the hardening energy by
\begin{align*}
&\mathcal{E}_{h}\,:\, L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\rightarrow \mathbb{R}\\
&p\longmapsto \mathcal{E}_{h}(p) = \dfrac{1}{2}\int_{\Omega}Bp:p\,dx.
\end{align*}
Given $\beta_1>0$, $\beta_2>0$ and $h>0$, the visco-elastic energy is
\begin{align*}
&\mathcal{E}_{ve}\,:\, H^{1}(\Omega,\mathbb{R}^2)\times H^{1}(\Omega,\mathbb{R}^2)\rightarrow \mathbb{R}\\
&(u,u_0)\longmapsto \mathcal{E}_{ve}(u,u_0) = \dfrac{\beta_1}{2h}\int_{\Omega}(E(u)-E(u_0)):(E(u)-E(u_0))\,dx.
\end{align*}
and the viscoplastic energy is defined by
\begin{align*}
&\mathcal{E}_{vp}\,:\, L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\times L^{2}(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})\rightarrow \mathbb{R}\\
&(p,p_0)\longmapsto \mathcal{E}_{vp}(p,p_0) = \dfrac{\beta_2}{2h}\int_{\Omega}(p-p_0):(p-p_0)\,dx.
\end{align*}
The Griffith surface energy is approximated by the phase-field surface energy
\begin{align*}
&\mathcal{E}_{S} \,:\,H^{1}(\Omega,\mathbb{R})\rightarrow \mathbb{R}\\
&v\quad\longmapsto \,\mathcal{E}_{S}(v)= \int_{\Omega} \varepsilon \vert\nabla v\vert^{2} dx+ \int_{\Omega}\frac{\left(1-v\right)^{2}}{4\varepsilon} \,dx.
\end{align*}
It is shown in \cite{Bourdin} that in the elastic anti-plane case, where the displacement reduces to a scalar and $Eu$ reduces to $\nabla u$,
the Ambrosio-Tortorelli functional
$$
\mathcal{E}_{\varepsilon}(\nabla u,v) =\mathcal{E}_{el}(\nabla u,v)+\mathcal{E}_{S}(v),
$$
$\Gamma$-converges, as $0<\eta\ll\epsilon\rightarrow0$, to the Griffith energy $\mathcal{G}$, where
\begin{eqnarray*}
\mathcal{G}(u):=\dfrac{1}{2}\int_{\Omega}A |\nabla u|^2\,dx+\mathcal{H}^{N-1}(S(u)).
\end{eqnarray*}
Here, $S(u)$ denotes the discontinuity set of u, and $\mathcal{H}^{N-1}$ is the $(N-1)$- dimensional Hausdorff measure.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}[scale=1]
\def\VDepth{2};
\def\VDefaultLevel{2.5};
\def\VEpsilon{0.5};
\def\VPower{0.35};
\def\VBorder{2};
\draw[line width=1pt] ({-\VEpsilon - \VBorder}, \VDefaultLevel) node[left] {$1+\eta$}
-- ++(\VBorder, 0)
.. controls +(0:\VPower) and +(180:\VPower) .. ++(\VEpsilon, -\VDepth)
.. controls +(0:\VPower) and +(180:\VPower) .. ++(\VEpsilon, \VDepth)
-- ++(\VBorder, 0);
\draw [dashed] (0, {\VDefaultLevel - \VDepth}) -- ++({-\VEpsilon - \VBorder}, 0) node[left] {$\eta$};
\draw[dashed] (-\VEpsilon, 0) -- (-\VEpsilon, {\VDefaultLevel + 1});
\draw[dashed] (\VEpsilon, 0) -- (\VEpsilon, {\VDefaultLevel + 1});
\draw[<->] (-\VEpsilon, {\VDefaultLevel + 0.75}) -- ++({2*\VEpsilon}, 0) node[pos=0.5, above] {$\approx\circ(\varepsilon)$};
\draw[->, line width=1pt] ({-\VEpsilon - \VBorder}, -0.5) -- ++(0, {\VDefaultLevel + 2}) node[right] {$v(x) + \eta$};
\draw[->, line width=1pt] ({-\VEpsilon - \VBorder}, 0) node[left] {$0$} -- ++({2*\VBorder + 2*\VEpsilon + 0.5}, 0) node[above] {$x$};
\end{tikzpicture}
\end{center}
\caption{In the generalized crack model, a crack is replaced by a thin region of very compliant material.}
\label{rupture}
\end{figure}
Note that in the region around an approximate crack, where $v$ is close to 0,
the effective Lam\'e tensor is $(v^2+\eta)A$ : the elastic material is replaced there by a
very compliant medium. We define $a:=v^2+\eta$.
\medskip
\subsection{Formulation of the models.}
We now propose 3 models that combine the various ingredients that we are interested in.
\begin{itemize}
\item Model 1 contains: elasticity, plasticity, visco-elasticity and fracture,
\item Model 2: elasticity, plasticity, visco-plasticity, fracture,
\item Model 3: elasticity, plasticity, kinematic hardening, fracture.
\end{itemize}
We define a time evolution for our models to be a quadruplet of functions
$(u,e,p,v): \Omega\times[0,T_f]\rightarrow \mathbb{R}^{2} \times\mathbb{M}_{\text{sym}}^{2\times2}\times\mathbb{M}_{\text{sym}}^{2\times2} \times \mathbb{R}$
satisfying the following conditions:
\begin{itemize}
\item[(E1)] Initial condition: $(u(0),v(0),e(0),p(0))=(u_0,v_0,e_0,p_0)$ with \\
$(u_0,e_0,p_0) \in A_{adm}(w(0))$. We also suppose that $(v_0^2+\eta)|Ae_0|\leq\tau$ and
$v_0 =1$ in $\Omega$ (the medium at $t=0$ does not contain any crack).
\item[(E2)] Kinematic compatibility: for $t\in[0,T_f]$,
\begin{eqnarray*}
(u(t),e(t),p(t)) \in A_{adm}(w(t)) \label{B}
\end{eqnarray*}
\item [(E3)] Equilibrium condition: for $t\in[0,T_f]$,
\begin{equation}\label{bebe}
\left\{
\begin{array}{ll}
-\text{div}(\sigma(t))=f(t),& a.e.\,\,\text{in}\quad \Omega,\\
\sigma(t).\vec{n}=g(t), & \text{on}\quad\partial\Omega_N,\notag\\
(u(t),v(t))=(w(t),1), \quad &\text{on} \quad \partial\Omega_D. \notag
\end{array}
\right.
\end{equation}
\item [(E4)] Constitutive relations: for $t\in[0,T_f]$,
\begin{itemize}
\item Model 1: $\sigma(t)= (v(t)^2 + \eta)Ae(t) + \beta_1 E\dot{u}(t)$.
The first term represents the stress due to elastic deformation, while the second
represents viscous dissipation.
\item Model 2 and Model 3: $\sigma(t)= (v(t)^2 + \eta)Ae(t).$
There the stress is only related to elastic deformation. We recall the notation $a(t):=v(t)^2 + \eta$.
\end{itemize}
\item[(E5)] Plastic flow rule: for a.e. $t\in[0,T_f]$,
\begin{itemize}
\item Model 1:
\begin{eqnarray}
a(t)Ae(t) \in \partial H(\dot{p}(t)) \quad \text{for}\,\,a.e.\quad x\in \Omega. \label{model1}
\end{eqnarray}
\item Model 2:
\begin{eqnarray}
a(t)Ae(t)\in \partial H_{\beta_2}(\dot{p}(t)) \quad \text{for}\,\,a.e.\quad x\in \Omega. \label{model2}
\end{eqnarray}
\item Model 3:
\begin{eqnarray}
a(t)Ae(t)-Bp(t) \in \partial H(\dot{p}(t)) \quad \text{for}\,\,a.e.\quad x\in \Omega. \label{model3}
\end{eqnarray}
\end{itemize}
\item[(E6)] Crack stability condition: for $t\in[0,T_f]$,
\begin{eqnarray*}
\mathcal{E}_{el}(e(t),v(t))+\mathcal{E}_{S}(v(t))= \inf_{v=1\,\text{sur} \,\partial\Omega_D,v\leq v(t)} \mathcal{E}_{el}(e(t),v)+\mathcal{E}_{S}(v).\label{ruptura}
\end{eqnarray*}
The crack stability condition implies that a fracture can only grow, and cannot disappear.
\item[(E7)] Energy balance formula: for every $T\in[0,T_f]$, the energy dissipates in the medium so as to satisfy
\begin{itemize}
\item Model 1:
\begin{eqnarray}
\lefteqn{\mathcal{E}_{el}(e(T),v(T))+\mathcal{E}_{S}(v(T))-\langle l(T), u(T)\rangle}\notag\\
&=&
\mathcal{E}_{el}(e(0),v(0))+\mathcal{E}_{S}(v(0))-\langle l(0), u(0)\rangle\notag \\
&-&
\beta_1\int_0^T \parallel E\dot{u}(t)\parallel_2^{2} dt - \tau \int_0^T \int_\Omega |\dot{p}| dx\,dt\notag\\
&-&
\int_0^T\langle\dot{l},u\rangle\,dt+\int_0^T\int_{\partial\Omega_D}\sigma(t)\vec{n}.\dot{w}(t) \,ds\,dt.\label{energy1}
\end{eqnarray}
\item Model 2:
\begin{eqnarray}
\lefteqn{\mathcal{E}_{el}(e(T),v(T))+\mathcal{E}_{S}(v(T))-\langle l(T), u(T)\rangle}\notag\\
&=&
\mathcal{E}_{el}(e(0),v(0))+\mathcal{E}_{S}(v(0))-\langle l(0), u(0)\rangle\notag \\
&-&
\beta_2\int_0^T \parallel\dot{p}(t)\parallel_2^{2} dt - \tau \int_0^T \int_\Omega |\dot{p}| dx\,dt\notag\\
&-&
\int_0^T\langle\dot{l},u\rangle\,dt+\int_0^T\int_{\partial\Omega_D}\sigma(t)\vec{n}.\dot{w}(t) \,ds\,dt.\label{energy2}
\end{eqnarray}
\item Model 3:
\begin{eqnarray}
\lefteqn{\mathcal{E}_{el}(e(T),v(T))+\mathcal{E}_{S}(v(T))+\mathcal{E}_{h}(p(t))-\langle l(T), u(T)\rangle}\notag\\
&=&
\mathcal{E}_{el}(e(0),v(0))+\mathcal{E}_{S}(v(0))+\mathcal{E}_{h}(p(0))-\langle l(0), u(0)\rangle\notag\\
&-&
\tau \int_0^T \int_\Omega |\dot{p}| dx\,dt\notag-\int_0^T\langle\dot{l},u\rangle\,dt+\int_0^T\int_{\partial\Omega_D}\sigma(t)\vec{n}.\dot{w}(t) \,ds\,dt.\notag\\
\label{energy3}
\end{eqnarray}
\end{itemize}
\end{itemize}
In the rest of paper, we suppose $l\equiv0$.
\section{Consistency of models with thermodynamics.}
In this section, we show that our models can be set in a thermodynamical framework
which resembles that of the Generalized Standard Materials of Halphen and Nguyen~\cite{HalphenSon},
see also Le Tallec~\cite{LeTallec}.
To this end, we introduce for $t\in [0,T]$, a free energy density $w(E(t),v(t),p(t))$
which depends on $E(t):=Eu(t)$, and on $v(t), p(t)$, the latter considered as internal variables (see~~\cite{HalphenSon}).
We also introduce a free energy functional $\mathcal{W}(t)$
\begin{eqnarray*}
\mathcal{W}(E,p,v)(t)&=&\int_{\Omega} w(E,p,v)(t)\,dx,
\end{eqnarray*}
and thermodynamic forces as operators associated with the internal variables
\begin{eqnarray*}
T_p(t)\tilde{p}:=-\frac{\partial w}{\partial p}(E,p,v)(t)\tilde{p}\quad\text{and}\quad T_v(t)\tilde{v}:=-\frac{\partial \mathcal{W}}{\partial v}(E,p,v)(t)\tilde{v}.\label{thermo3}
\end{eqnarray*}
Note that in our case, the thermodynamic force associated with the phase field
is defined as a global operator $H^1(\Omega) \longrightarrow \mathbb{R}$.
We also postulate the existence of a dissipation potential $\phi(t)=\phi(\dot{E}(t),\dot{p}(t),\dot{v}(t))$
which is a convex function of its arguments and minimal at
$(\dot{E}(t),\dot{p}(t),\dot{v}(t))=(0,0,0)$.
According to the principle of conservation of linear momentum,
we recall that the Cauchy theorem implies the existence of symmetric stress tensor $\sigma(t)$ which satisfies under the hypothesis
of small deformations the equilibrium condition (E3).
The stress tensor is split into an irreversible and a reversible part by setting
\begin{eqnarray}
\sigma^{rev}(t):=\frac{\partial w}{\partial E}(E(t),p(t),v(t))\quad\text{and}\quad \sigma^{irrev}(t):=\sigma(t)-\sigma^{rev}(t).\label{baba}
\end{eqnarray}
Following the work of Halphen and Nguyen \cite{HalphenSon} and LeTallec \cite{LeTallec}, we make the constitutive hypothesis that the thermodynamic forces are
related to the dissipation potential by
\begin{eqnarray}
(\sigma^{irrev}(t),T_p(t), T_v(t)) \in \partial \phi (\dot{E}(t), \dot{p}(t), \dot{v}(t)). \label{SE8}
\end{eqnarray}
Our goal is to show that our models are consistent with this thermodynamic
framework, in the sense that if one assumes (\ref{SE8}) and if the equilibrium condition (E3) is verified, then one recovers the relations (E4)-(E7),
and further, a form of the Clausius-Duhem inequality holds.
We state this result for Model 1 only. The same analysis carries on
for Models 2 and 3 (see the remark below).
We also define the fracture dissipation potential by
$$
D_{\text{S}} (\xi)= \left\{
\begin{array}{ll}
0 & \mbox{if} \quad\xi\leq 0, \,\, a.e.\quad\text{in} \,\Omega, \,\xi\in H^1_D(\Omega),\, \\
\infty & \mbox{elsewhere},
\end{array}
\right.
$$
with $H^1_D(\Omega):=\{z\in H^1(\Omega); z=0\,\, \text{on}\,\, \partial\Omega_D\}$.
\begin{theorem}\label{Theorem1}
Suppose that $(u,v,e,p)$ in $C^1(0,T_f, H^1(\Omega))\times C^1(0,T_f, H^1(\Omega))\times C^1(0,T_f, L^2(\Omega))\times C^1(0,T_f, L^2(\Omega))$
satisfy for all $t\in [0,T]$, $(\dot{u}(t),\dot{e}(t),\dot{p}(t))\in A_{adm}(\dot{w}(t))$, $\dot{v}(t)\leq 0 \, a.e.\,\text{in} \,\,\Omega, \,\dot{v}(t)\in H^1_D(\Omega)$, $v(t)=1$ on $\partial\Omega_D$,
(E1), (E2), (E3), and (\ref{SE8}). Let
\begin{eqnarray*}
\mathcal{W}_1(t,Eu(t),p(t),v(t))&:=& \dfrac{1}{2} \int_{\Omega}\left( v(t)^{2}+\eta\right)A(Eu(t)-p(t)):(Eu(t)-p(t)) \,dx\\
&+& \int_{\Omega} \varepsilon \vert\nabla v(t)\vert^{2} dx+ \int_{\Omega}\frac{\left(1-v(t)\right)^{2}}{4\varepsilon} \,dx,
\end{eqnarray*}
and the potential of dissipation
\begin{eqnarray*}
\phi_1(t,E\dot{u}(t),\dot{p}(t),\dot{v}(t))= \dfrac{1}{2} \beta_1E\dot{u}(t):E\dot{u}(t)+ \tau |\dot{p}(t)|+D_{\text{S}} (\dot{v}(t)).
\end{eqnarray*}
Then $(u,v,e,p)$ satisfies (E4), (E5), (E6), (E7). Furthermore, for all $t\in [0,T_f]$,
\begin{eqnarray}
\mathcal{D}(t):=\int_{\Omega}\sigma(t):E\dot{u}(t)\,dx-\dot{\mathcal{W}}_1(t)\geqslant 0.
\end{eqnarray}
\end{theorem}
\begin{proof}:
Let $t\in[0,T_f]$. The relations (\ref{baba}) and (\ref{SE8}) lead to (E4)
\begin{eqnarray}
\sigma(t)=a(t)Ae(t)+\beta_1E\dot{u}(t). \label{SE10}
\end{eqnarray}
We also deduce from (\ref{SE8}) that
\begin{eqnarray}
(v(t)^2+\eta)Ae(t) \in \partial H(\dot{p}(t)). \label{SE11}
\end{eqnarray}
which prove (E5).
From (\ref{SE8}) we see that for every $\xi\leq0$ and $\xi=0$ on $\partial\Omega_D$
\begin{eqnarray}
\lefteqn{-\frac{\partial \mathcal{W}_1}{\partial v}(E,p,v)(t)(\xi-\dot{v}(t))}\notag\\
&=&
-\int_{\Omega}v(t)A(Eu(t)-p(t)):(Eu(t)-p(t))(\xi-\dot{v}(t))\notag\\
&+&
(2\varepsilon)^{-1}(v(t)-1)(\xi-\dot{v}(t))+2\varepsilon\nabla v(t)\nabla (\xi-\dot{v}(t))\,dx\leq0,\label{SE12}
\end{eqnarray}
so that we obtain
\begin{eqnarray}
\lefteqn{\int_{\Omega}v(t)A(Eu(t)-p(t)):(Eu(t)-p(t))(\dot{v}(t)-\xi)}\notag\\
&+&
(2\varepsilon)^{-1}(v(t)-1)(\dot{v}(t)-\xi)\,dx+\int_{\Omega}2\varepsilon\nabla v(t)\nabla(\dot{v}(t)-\xi)\,dx\leq0\label{thermottt}.
\end{eqnarray}
Testing (\ref{thermottt}) with $\xi=\dot{v}(t)+\varphi-v(t)$ where $\varphi \in H^1(\Omega)$, $\varphi \leq v(t)$, and $\varphi=1$ on $\partial\Omega_D$, implies that
\begin{eqnarray}
&&2\varepsilon\int_{\Omega} \nabla v(t)\nabla(v(t)-\varphi) \,dx+\int_{\Omega}v(t)A e(t):e(t)(v(t)-\varphi)\,dx\notag\\
&+&(2\varepsilon)^{-1}\int_{\Omega}(v(t)-1)(v(t)-\varphi)\,dx \leq 0, \label{thermot4}
\end{eqnarray}
for every $\varphi \in H^1(\Omega)$, $\varphi \leq v(t)$, and $\varphi =1$ on $\partial\Omega_D$.
We rewrite (\ref{thermot4}) as follows
\begin{eqnarray}
\lefteqn{2\varepsilon\int_{\Omega} \nabla v(t)\nabla v(t) \,dx+\int_{\Omega}v(t)A e(t):e(t)v(t)\,dx+(2\varepsilon)^{-1}\int_{\Omega}(v(t)-1)v(t)\,dx } \notag\\
&\leq&
2\varepsilon\int_{\Omega} \nabla v(t)\nabla \varphi \,dx+\int_{\Omega}v(t)A e(t):e(t)\varphi\,dx+(2\varepsilon)^{-1}\int_{\Omega}(v(t)-1)\varphi\,dx.\notag\\\label{thermot5}
\end{eqnarray}
Using the Cauchy inequality yields
\begin{eqnarray}
2\varepsilon\int_{\Omega} \nabla v(t)\nabla \varphi \,dx\leq\varepsilon\int_{\Omega} \arrowvert\nabla v(t)\arrowvert^2\,dx+\varepsilon\int_{\Omega} \arrowvert\nabla \varphi\arrowvert^2\,dx,\notag
\end{eqnarray}
\begin{eqnarray}
\int_{\Omega}v(t)A e(t):e(t)\varphi\,dx\leq\dfrac{1}{2}\int_{\Omega}v^2(t)A e(t):e(t)\,dx+\dfrac{1}{2}\int_{\Omega}\varphi^2 A e(t):e(t)\,dx.\notag
\end{eqnarray}
We rewrite
\begin{eqnarray*}
&&(v(t)-1)\varphi=(v(t)-1)(\varphi-1)+(v(t)-1),\\
&&(v(t)-1)v(t)-(v(t)-1)=(v(t)-1)^2,
\end{eqnarray*}
and it follows that
\begin{eqnarray}
\mathcal{E}_{el}(e(t),v(t))+ \mathcal{E}_{S}(v(t))\leq \mathcal{E}_{el}(e(t),\varphi)+\mathcal{E}_{S}(\varphi)
\end{eqnarray}
for all $\varphi \in H^1(\Omega)$, $\varphi \leq v(t)$, and $\varphi =1$ on $\partial\Omega_D$, which proves (E6).
We now prove the energy balance formula. First we differentiate $\mathcal{W}_1(t,E(u(t)),p(t),v(t))$ in time:
\begin{eqnarray}
\lefteqn{\dfrac{d}{dt} \mathcal{W}_1(t,E(u(t)),p(t),v(t))=\int_{\Omega}a(t)A(Eu(t)-p(t)):(E\dot{u}(t)-\dot{p}(t))\,dx}\notag\\
&+&\int_{\Omega}v(t)A(Eu(t)-p(t)):(Eu(t)-p(t))\dot{v}(t)+(2\varepsilon)^{-1}(v(t)-1)\dot{v}(t)\,dx\notag\\
&+&2\int_{\Omega}\varepsilon\nabla v(t)\nabla \dot{v}(t)\,dx\label{I1}.
\end{eqnarray}
Testing inequality (\ref{SE12}) with $\xi=0$ and $\xi=2\dot{v}(t)$ leads to
\begin{eqnarray}
-\frac{\partial \mathcal{W}_1}{\partial v}(E,p,v)(t)\dot{v}(t)=0.\label{I2}
\end{eqnarray}
From (\ref{I1}) and (\ref{I2}) we deduce that
\begin{eqnarray}
\lefteqn{\dfrac{d}{dt} \mathcal{W}_1(t,Eu(t),p(t),v(t))}\notag\\
&=&
\int_{\Omega}a(t)A(Eu(t)-p(t)):(E\dot{u}(t)-\dot{p}(t))\,dx \notag\\
&=&\int_{\Omega}\sigma(t):E\dot{u}(t)\,dx-\int_{\Omega}\beta_1 E\dot{u}(t):E\dot{u}(t)\,dx\notag\\
&-&\int_{\Omega} a(t)A(Eu(t)-p(t)):\dot{p}(t)\,dx.\label{I3}
\end{eqnarray}
The equilibrium equation (E3) gives
\begin{eqnarray}
\int_{\Omega}\sigma(t):E\dot{u}(t)\,dx=\int_{\partial\Omega_D}\sigma(t)\vec{n}.\dot{w}(t) \,ds.\label{I4}
\end{eqnarray}
By definition of the subgradient, (\ref{SE11}) leads to a variational inequality: for all admissible $q \in L^2(\Omega,\mathbb{M}_{\text{sym}}^{2\times2})$ we have
\begin{eqnarray}
\tau\int_{\Omega}|q|\,dx \ge \tau\int_{\Omega}|\dot{p}(t)|\,dx+\int_{\Omega} a(t)A(Eu(t)-p(t)) :(q-\dot{p}(t))\,dx \label{model4}
\end{eqnarray}
Testing (\ref{model4}) with $q=0$ et $q=2\dot{p}(t)$ implies that
\begin{eqnarray}
\int_{\Omega} a(t)A(Eu(t)-p(t)) :\dot{p}(t)\,dx=\tau\int_{\Omega}|\dot{p}(t)|dx.\label{mod2}
\end{eqnarray}
So that we deduce from (\ref{I3}), (\ref{I4}), (\ref{mod2}) that
\begin{eqnarray}
\dfrac{d}{dt} \mathcal{W}_1(t,Eu(t),p(t),v(t))&=& -\int_{\Omega}\beta_1 E\dot{u}(t):E\dot{u}(t)\,dx-\tau\int_{\Omega}|\dot{p}(t)|\,dx\notag\\
&+&\int_{\partial\Omega_D}\sigma(t)\vec{n}.\dot{w}(t) \,ds. \label{mod3}
\end{eqnarray}
Integrating (\ref{mod3}) over $[0, T]$, for every $0\leq T\leq T_f$ shows that the balance formula (E7) holds.
Finally, from (\ref{I4}) and (\ref{mod3}) we deduce that
\begin{eqnarray*}
\lefteqn{\mathcal{D}(t):=\int_{\Omega}\sigma(t):E\dot{u}(t)\,dx-\dot{\mathcal{W}_1}(t)}\\
&=&
\int_{\Omega}\beta_1 E\dot{u}(t):E\dot{u}(t)\,dx+\tau\int_{\Omega}|\dot{p}(t)|\,dx\geqslant 0.
\end{eqnarray*}
\hfill$\square$
\end{proof}
\begin{remark}
\begin{enumerate}
\item The assumption (\ref{SE8}) is stronger than (E6).
\item Theorem~\ref{Theorem1} also holds for Models 2 and 3 with the following choices of free energies and dissipation potentials:
\begin{itemize}
\item for Model 2:
\begin{eqnarray*}
\lefteqn{\mathcal{W}_2(t,Eu(t),p(t),v(t))}\\
&:=&
\dfrac{1}{2} \int_{\Omega}\left( v(t)^{2}+\eta\right)A(Eu(t)-p(t)):(Eu(t)-p(t)) \,dx\\
\quad\quad&+&
\int_{\Omega} \varepsilon \vert\nabla v(t)\vert^{2} dx+ \int_{\Omega}\frac{\left(1-v(t)\right)^{2}}{4\varepsilon} \,dx.
\end{eqnarray*}
and
\begin{eqnarray*}
\phi_2(t,E\dot{u}(t),\dot{p}(t),\dot{v}(t))= \dfrac{1}{2} \beta_2\dot{p}(t):\dot{p}(t)+ \tau |\dot{p}(t)|+D_{S}(\dot{v}(t)).
\end{eqnarray*}
\item for Model 3:
\begin{eqnarray*}
\lefteqn{\mathcal{W}_3(t,Eu(t),p(t),v(t))}\\
&:=& \dfrac{1}{2} \int_{\Omega}\left( v(t)^{2}+\eta\right)A(Eu(t)-p(t)):(Eu(t)-p(t)) \,dx\notag\\
&+&
\dfrac{1}{2} \int_{\Omega} Bp(t):p(t)\,dx+\int_{\Omega} \varepsilon \vert\nabla v(t)\vert^{2} dx+ \int_{\Omega}\frac{\left(1-v(t)\right)^{2}}{4\varepsilon} \,dx,
\end{eqnarray*}
and
\begin{eqnarray*}
\phi_3(t,E\dot{u}(t),\dot{p}(t),\dot{v}(t))= \tau |\dot{p}(t)|+D_{S} (\dot{v}(t)).
\end{eqnarray*}
\end{itemize}
\end{enumerate}
\end{remark}
\section{Discrete-time evolutions for Models 1-3.}
We now approximate the continuous-time evolutions of the constructed models via discrete time evolutions obtained by solving incremental variational problems.
We describe the discrete-time evolution of the medium as follows:
we consider a partition of the time interval $[0,T_f]$ into $N_f$ sub-intervals of equal length $h$:
\begin{eqnarray*}
0=t_h^0<t_h^1<...<t_h^n<...<t_h^{N_f}=T_f, \quad \text{with}\quad h=\dfrac{T_f}{N_f}= t_h^n-t_h^{n-1}\rightarrow0.
\end{eqnarray*}
We define
\begin{eqnarray*}
\mathcal{B}(t_h^n)&:=&\left\{(z,q,\varphi)\in H^1(\Omega)\times L^2(\Omega)\times H^1(\Omega);
\begin{array}{ll}
z=w_h^n ,&\text{on}\quad \partial \Omega_D,\\
\varphi=1, & \text{on}\quad\partial \Omega_D,\notag\\
\varphi\leq v_h^{n-1},&\text{in}\quad\Omega.\notag
\end{array}
\right\}.
\end{eqnarray*}
Let us assume that for $n\geqslant1$, the approximate evolution $(u_h^{n-1}, v_h^{n-1}, p_h^{n-1})\in\mathcal{B}(t_h^{n-1})$ is known at $t_h^{n-1}$.
We seek $(u_h^{n}, v_h^{n}, p_h^{n})$ at time $t_h^n$ as the solution to the following variational problem:
\begin{eqnarray}
\min_{(z,q,\varphi)\in \mathcal{B}(t_h^n)}\mathcal{E}_{total}(z,q,\varphi,u_h^{n-1}, v_h^{n-1}, p_h^{n-1}),\label{Problem1}
\end{eqnarray}
where, for each model, $\mathcal{E}_{total}$ is defined as follows:
\begin{enumerate}
\item Model 1: Elasto-plasticity, viscoelasticity and fracture:
\begin{eqnarray*}
\mathcal{E}^1_{total}(z,q,\varphi,u_h^{n-1}, v_h^{n-1}, p_h^{n-1})&=&\mathcal{E}_{el}(\varphi,E(z)-q)+\mathcal{E}_{p}(q,p_h^{n-1})\\
&+&\mathcal{E}_{ve}(z,u_h^{n-1})+\mathcal{E}_{S}(\varphi).
\end{eqnarray*}
\item Model 2: Elasto-plasticity, viscoplasticity and fracture:
\begin{eqnarray*}
\mathcal{E}^2_{total}(z,q,\varphi,u_h^{n-1}, v_h^{n-1}, p_h^{n-1})&=&\mathcal{E}_{el}(\varphi,E(z)-q)+\mathcal{E}_{p}(q,p_h^{n-1})\\
&+&\mathcal{E}_{vp}(q,p_h^{n-1})+\mathcal{E}_{S}(\varphi).
\end{eqnarray*}
\item Model 3: Elasto-plasticity, linear kinematic hardening and fracture:
\begin{eqnarray*}
\mathcal{E}^3_{total}(z,q,\varphi,u_h^{n-1}, v_h^{n-1}, p_h^{n-1})&=&\mathcal{E}_{el}(\varphi,E(z)-q)+\mathcal{E}_{p}(q,p_h^{n-1})\\
&+&\mathcal{E}_{h}(q)+\mathcal{E}_{S}(\varphi).
\end{eqnarray*}
One can easily prove that for $i=1,2,3$ the variational problem
\begin{eqnarray}
\min_{(z,q,\varphi)\in \mathcal{B}(t_h^n)}\mathcal{E}^i_{total}(z,q,\varphi,u_h^{n-1}, v_h^{n-1}, p_h^{n-1}).\label{Problem2}
\end{eqnarray}
has at least one solution. If $(z_n,q_n,\varphi_n)_n$ is a minimizing subsequence, that one easily checks that $\parallel z_n \parallel_{H^1}$, $\parallel p_n \parallel_{L^2}$,
$\parallel \varphi_n \parallel_{H^1}$ are uniformly bounded, so that a subsequence converges weakly to some $(z,q,\varphi)$.
The only difficulty in passing to the limit in $\mathcal{E}_{total}$ comes from the term
\begin{eqnarray*}
\int_{\Omega} \varphi_n^2 A(Ez_n-q_n):(Ez_n-q_n)\,dx,
\end{eqnarray*}
which can be rewritten as
\begin{eqnarray*}
\int_{\Omega} A\varphi_n(Ez_n-q_n):\varphi_n(Ez_n-q_n)\,dx,
\end{eqnarray*}
and one can use the fact that since
\begin{eqnarray*}
& \displaystyle Ez_n \rightharpoonup Ez& \quad \text{weakly in} \quad L^{2}, \\
& \displaystyle \varphi_n \rightharpoonup \varphi& \quad \text{weakly in} \quad H^{1}, \\
& \displaystyle q_n \rightharpoonup q& \quad \text{weakly in} \quad L^{2},
\end{eqnarray*}
one has
\begin{eqnarray*}
\displaystyle \varphi_n(Ez_n-q_n) \rightharpoonup \varphi(Ez-q)& \quad \text{weakly in} \quad L^{2}.
\end{eqnarray*}
Note however that $\mathcal{E}_{total}$ is not convex because of the no quadratic term $\varphi^2 (E(z)-q):(E(z)-q)$ in the elastic energy.
So that there might be several solutions to (\ref{Problem2}).
\end{enumerate}
\subsection{An alternate minimization algorithm and backtracking for materials with memory}
A solution $(u_h^n, v_h^n, p_h^n)$ of problem (\ref{Problem2}) is characterized by a system of one equality and two variational inequalities. Such a system is not easy to solve numerically.
For this reason we propose to solve (\ref{Problem2}) at each time step $t_h^n$ using an alternate minimization algorithm. The advantage of this approach
is that the problem (\ref{Problem2}) is separately strictly convex in each variable, so that each alternating step has a unique solution.
\subsubsection{An alternate minimization algorithm}
\begin{algorithm}
\caption{Alternate minimization algorithm}
Let $\delta_1>0$ and $\delta_2>0$ be fixed tolerance parameters.
\begin{enumerate}
\item Let $m=0$, $v^n_{(m=0)}=v_h^{n-1}$
\item \textbf{iterate}
\item Find $(u^n_{(m)}, p^n_{(m)}):={\operatornamewithlimits{argmin}}_{(u,p)} \mathcal{E}_{total}(u,p,v^n_{(m-1)})$
\item \hspace*{2cm} Let $l=0$, $p^n_{(l=0)}=p_h^{n-1}$
\item \hspace*{2cm} \textbf{iterate}
\item \hspace*{2cm} $u^n_{(l)}:={\operatornamewithlimits{argmin}}_{u} \mathcal{E}_{total}(u,p^n_{(l-1)},v^n_{(m-1)} )$
\item \hspace*{2cm} $p^n_{(l)}:={\operatornamewithlimits{argmin}}_{p} \mathcal{E}_{total}(u^n_{(l)},p,v^n_{(m-1)} )$
\item \hspace*{2cm} \textbf{until} $\parallel u^n_{(l)}-u^n_{(l-1)}\parallel_{H^1}\leq\delta_1$
\item \hspace*{2cm} We define $u^n_{(m)}:=u^n_{(l)}$ and $p^n_{(m)}:=p^n_{(l)}$ at convergence
\item Find $v^n_{(m)}:={\operatornamewithlimits{argmin}}_{v} E_{total}(u^n_{(m)},p^n_{(m)},v)$
\item \textbf{until} $\parallel v^n_{(m)}-v^n_{(m-1)}\parallel_{H^1}\leq\delta_2$
\item We define $u_h^n:=u^n_{(m)}$, $p_h^n:=p^n_{(m)}$ and $v_h^n:=v^n_{(m)}$ at convergence
\end{enumerate}
\end{algorithm}
In practice, it is not exactly the variational problems described
in the alternating procedure above, that one solves, but the associated
first-order optimality conditions of problems that have been discretized
in space too. This may introduce local minima, as the following example of a traction
experiment of a 1D bar illustrates.
Assume that $u, v \in W^{1,\infty}(0,T_f,\Omega)$ represent the displacement
and phase-field marker of a 1D bar that lies in $\Omega = (0,L)$. We consider
a simple model of evolution with only elasticity and fracture where the total energy writes
\begin{eqnarray*}
\mathcal{E}_{total}(u,v)&=&\mathcal{E}_{el}(v,u^{'}) + \mathcal{E}_S(v)\\
&=&\dfrac{1}{2} \int_{\Omega} (v^2 + \eta) K (u^{'})^2 \,dx+\int_{\Omega }\varepsilon(v^\prime)^2 +\frac{(1-v)^2}{4 \varepsilon}\,dx,
\end{eqnarray*}
where $K > 0$ is a fixed Young modulus and where the primes denote derivatives
with respect to $x$. The bar is crack-free at $t=0$ and thus $v(x,0) = 1$.
It is fixed at $x=0$, while a uniform traction $u(L,t) = tL$ is applied at the other extremity.
If $u^\prime$ is close to a constant at time $t$, say $u^\prime(x,t) \sim t$, then the Euler-Lagrange optimality
condition for minimization of the total energy with respect to $v$ amounts to solving
\begin{eqnarray*}
v^{\prime\prime} -\left(\frac{1}{4\varepsilon^2}+\dfrac{Kt^2}{2\varepsilon}\right) v+\frac{1}{4\varepsilon^2}=0,
\end{eqnarray*}
the solution of which is
\begin{eqnarray*}
v(x,t)=C_1e^{-\frac{x\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}+C_2e^{\frac{x\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}},
\end{eqnarray*}
with
\begin{eqnarray*}
C_1:=\frac{e^{\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}-1}{\left(2kt^2\varepsilon+1\right) \left( e^{-\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}-e^{\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}\right)},
\end{eqnarray*}
and
\begin{eqnarray*}
C_2:=\frac{e^{-\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}-1}{\left(2kt^2\varepsilon+1\right) \left( e^{-\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}-e^{\frac{L\sqrt{2}\sqrt{4Kt^2\varepsilon+2}}{4\varepsilon}}\right)}.
\end{eqnarray*}
These profiles are indeed what one obtains in the course of the numerical
computations according to the algorithm described above,
until $t$ reaches a sufficiently large value so that the
term $\int_\Omega (v^2 + \eta) K (u^{'})^2\,dx$ dominates $\int_\Omega \frac{(1-v)^2}{4\varepsilon}\,dx$
in the energy, see Figures~\ref{PV}, ~\ref{DERIVEEU} and \ref{ener} (we use the same parameters as in \cite{BourdinFrancfortMarigo} by Bourdin).
Note that, due to the presence of the exponentials in the expression of $v$,
these profiles vary significantly near the extremities $x=0$ and $x=L$ of the beam,
but are quite flat otherwise, and do not correspond to the picture of a generalized
crack as that depicted in Figure~\ref{rupture}.
Further, the corresponding states $u,v$ are only local minima,
as one can build states with lower total energy, as the examples below show.
We note that choosing $\varepsilon$ smaller does not improve the situation for that matter.
This hurdle had been noticed earlier by Bourdin~\cite{Bourdin2}, \cite{Bourdin}, who suggested to complement
the numerical algorithm with a supplementary step called backtracking,
where after each iteration in time, one imposes a necessary condition derived
from the definition~(\ref{Problem1}) of the discrete-time evolution. We extend this idea in the context
of our models, where plasticity and viscous dissipation may occur as well.
\begin{figure}[htbp!]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-8mm}
\includegraphics[width=8cm]{profilv.pdf}
\caption{Profile of $v(t,.)$ during an elastic evolution with fracture at time $t=4$.}
\label{PV}
\end{minipage}
\hspace*{10mm}
\begin{minipage}[c]{.46\linewidth}
\includegraphics[width=8cm]{profilderivee.pdf}
\caption{Profile of $u{'}(t,.)$ during an elastic evolution with fracture at time $t=4$.}
\label{DERIVEEU}
\end{minipage}
\end{figure}
\begin{figure}[htbp!]
\centering
\hspace*{-8mm}
\includegraphics[width=8cm]{energy1.pdf}
\caption{Evolution of the total, elastic, and surface energies for the 1D traction experiment without backtracking.}
\label{ener}
\end{figure}
\subsubsection{Backtracking}
Because the loading is monotonous, if $(u_h^n,p_h^n,v_h^n)$ is a solution of (\ref{Problem1}) at time $t_h^n$,
then $(\frac{t_h^j}{t_h^n} u^n,\frac{t_h^j}{t_h^n} p^n,v_h^n)$ is admissible at time $t_h^j$. Thus, we must have
\begin{eqnarray}
\mathcal{E}_{total}(u_h^j,p_h^j,v_h^j)\leq \mathcal{E}_{total}((\frac{t_h^j}{t_h^n} u^n,\frac{t_h^j}{t_h^n} p^n,v_h^n)).
\end{eqnarray}
Numerically we check this condition for all $t_h^j\leq t_h^n$. If there exists some j such that this condition is not verified
$(u_h^j,p_h^j,v_h^j)$ cannot be a global minimizer for time $t_h^j$; and we backtrack to time $t_h^j$ for the alternate minimization algorithm with initialization $v^j_{(m=0)}=v_h^n$ and $p^j_{(l=0)}=p_h^{j-1}$.
\section{Dissipation phenomena appearing during deformation - 1D and 2D numerical experiments}
In this section, we study some evolution problems for Models 1-3 in terms of their mechanical parameters, to check if
during evolution several dissipation phenomena can be observed.
\subsection{1D-traction numerical experiments with fracture}
We consider a beam $\Omega=(0,L)$ of length L, the Young modulus $K>0$. It is clamped at $x=0$.
A Dirichlet boundary condition $u(L,t)$=tL is imposed at its right extremity $x=L$.
At each time step $t_h^n$ we use P1-elements to approximate $u$ and $v$, and P0-elements for $p$.
\subsubsection{Elasto-perfectly plastic case with fracture}\label{Model0}
If we take $\beta_1=0$ in Model 1 or $\beta_2=0$ in Model 2, these models reduce to perfect plasticity with numerical fracture.
\begin{eqnarray}
\min_{(z,q,\varphi)\in \mathcal{B}(t_h^n)} \mathcal{E}_{el}(\varphi,E(z)-q)+\mathcal{E}_{p}(q,p_h^{n-1})+\mathcal{E}_{S}(\varphi).\label{backtrack}
\end{eqnarray}
In this example, we illustrate the importance of the backtracking step.
We apply the alternate minimization algorithm without backtracking with the following parameters: $L=10$, $K=4$, $\tau=1.5$, the space discretization mesh size $\bigtriangleup x=0.015$, the time step $h=0.025$, $\eta=10^{-6}$, $\varepsilon=0.094$.
With this choice of parameters, we observe in Figure~\ref{Figure1} that if the beam is elastic ($v=1$, $p=0$) at time $t=0$, it remains elastic until the time $t\simeq0.5$ when the beam becomes plastic, then a crack appears at $t\simeq3$.
Because the loading is monotonous, if the system is crack free at $t=0$ and if $p=0$,
it should remain in the elastic regime until the stress reaches the yield surface or until a crack appears. It is easy to check that if there is no crack,
the yield stress should be reached at time $t_p=\tau/K$, while if there is no plastic deformation, a crack should appear at time $t_c=\sqrt{2/KL}$.
With the given choice of parameters, we obtain $t_p=0.375$ and $t_c=0.224$.
When we compare with Figure~\ref{Figure1}, we see that the beam deforms elastically until plastic deformation takes place at time $t\simeq0.5$,
far from the predicted value.
Figure~\ref{Figure2} shows the same traction experiment computed with the backtracking step. We see that the elastic medium cracks at the computed time $t\simeq0.25$
which is close to the expected theoretical
crack time $t_c=0.224$. Before plastic deformation takes place.
\begin{figure}[htbp!]
\begin{center}
\centering
\includegraphics[width=8cm]{Image1.pdf}
\end{center}
\caption{Evolution of the total, elastic, plastic and surface energies for the 1D traction experiment without backtracking, $\tau=1.5$.}
\label{Figure1}
\end{figure}
We now change the plastic parameter to $\tau=0.8$ so that the expected plastic time should be $t_p=0.2$.
Figure~\ref{Figure3} shows that, as expected since now $t_p<t_c$, plastic deformation occurs first. As this model does not allow the elastic energy
to grow once plastic deformation has taken place, no crack appears after $t_p$.
In all the following experiments, the backtracking strategy is used.
\begin{figure}[htbp!]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image2.pdf}
\caption{Evolution of the total, elastic, plastic and surface energies for the 1D traction experiment with backtracking, $\tau=1.5$.}
\label{Figure2}
\end{minipage} \hfill
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image3.pdf}
\caption{Evolution of the total, elastic, plastic and surface energies for the 1D traction experiment with backtracking, $\tau=0.8$.}
\label{Figure3}
\end{minipage}
\end{figure}
In the sequel, the same discretization parameters are used, and the Young Modulus is chosen as in \ref{Model0}.
\subsubsection{Model 1 - Elasto-plastic model with visco-elasticity and fracture.}
As can be seen, Model 1 can express the various dissipation mechanisms: elasto-plasticity only (Figure.~\ref{first}), elasticity with fracture (Figure.~\ref{second}),
viscoelasticity with fracture (Figure.~\ref{third}), and elasto-visco-plasticity with fracture (Figure.~\ref{four}).
\begin{figure}[htbp!]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image7.pdf}
\caption{Evolution of the total, elastic, plastic, viscoelastic and surface energies for the 1D traction experiment with backtracking, $\tau=1$, $\beta_1=0.01$.}
\label{first}
\end{minipage}
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image5.pdf}
\caption{Evolution of the total, elastic, plastic, viscoelastic and surface energies for the 1D traction experiment with backtracking, $\tau=1.5$, $\beta_1=0.0001$.}
\label{second}
\end{minipage}
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image6.pdf}
\caption{Evolution of the total, elastic, plastic, viscoelastic and surface energies for the 1D traction experiment with backtracking, $\tau=5$, $\beta_1=0.01$.}
\label{third}
\end{minipage} \hfill
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image4.pdf}
\caption{Evolution of the total, elastic, plastic, viscoelastic and surface energies for the 1D traction experiment with backtracking, $\tau=1.5$, $\beta_1=0.01$.}
\label{four}
\end{minipage} \hfill
\end{figure}
\subsubsection{Model 2 - The elasto-viscoplastic model with fracture}
We cannot exclude complex regimes, however, in all our numerical experiments, we only
observed that after the initial elastic regime, either plastic deformation takes place (Figure~\ref{Figure8}), or a crack may appear (Figure~\ref{Figure11}).
\begin{figure}[htbp!]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image8.pdf}
\caption{Evolution of the total, elastic, plastic, viscoplastic and surface energies for the 1D traction experiment with backtracking, $\tau=1$, $\beta_2=0.1$.}
\label{Figure8}
\end{minipage} \hfill
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image9.pdf}
\caption{Evolution of the total, elastic, plastic, viscoplastic and surface energies for the 1D traction experiment with backtracking, $\tau=1$, $\beta_2=1$.}
\label{Figure11}
\end{minipage}
\end{figure}
\subsubsection{Model 3- Elasto-plastic model with linear kinematic hardening and fracture.}\label{Hardening3}
We choose the hardening parameter $k=0.5$.
For $\tau=1$, the Figure~\ref{Figure4} shows the elastic behavior with fracture.
For $\tau=0.7$, the medium firstly plastifies and then cracks as depicted in
Figure~\ref{Figure12}.
Indeed, kinematic hardening allows the translation of the yield surface and thus the elastic energy can increase after the plastification, so that cracks can appear.
\begin{figure}[htbp!]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image10.pdf}
\caption{Evolution of the total, elastic, plastic, hardening and surface energies for the 1D traction experiment with backtracking, $\tau=1$.}
\label{Figure4}
\end{minipage} \hfill
\begin{minipage}[c]{.46\linewidth}
\hspace*{-10mm}
\includegraphics[width=8cm]{Image11.pdf}
\caption{Evolution of the total, elastic, plastic, hardening and surface energies for the 1D traction experiment with backtracking, $\tau=0.7$.}
\label{Figure12}
\end{minipage}
\end{figure}
\subsection{2D-traction numerical experiments-Model 3}
From the numerical 1D-traction experiments of Models 1-3 we conclude that the Models 1 and 3 are those allowing the more complex
evolutions, as all their dissipative mechanisms can be expressed. Those two models have the capacity to plastify the body and then crack during evolution.
This behavior strongly depends on the choice of the mechanical parameters.
Here, we illustrate the behaviour of Model 3 in 2D traction numerical experiments.
We remark that contrarily to the 1D case, the minimization with respect to p is not explicit.
We compute $p$ with a standard gradient descent method. We consider a beam of length L, and cross section $S=1$ (so that $\Omega=(0,L)\times(0,1))$, which is clamped at $(x,y)=(0,y)$ for $y\in (0,S)$.
The elastic parameters are the Young modulus $K$ and $\nu$ the Poisson coefficient. The elastic matrix A is defined via Lam\'e's coefficients associated with $K$ and $\mu$.
For $y\in (0,S)$, we impose at time $t_h^n$ a constant displacement $t_h^nW_0=(t_h^nU_0,0)$ with $U_0>0$ at the right extremity $(x,y)=(L,y)$ of the beam.
We consider the hardening tensor $B$ as a diagonal matrix $B=k\mathbb{I}_2$ with $k>0$ a hardening parameter.
We report numerical experiments with the following parameters: $h=0.1$, $\bigtriangleup x=0.05$, $\varepsilon=0.25$, $K=10$, $k=100$, $\tau=1$, $\nu=0.252$, $U_0=1$.
\begin{figure}[htbp]
\begin{minipage}[c]{.46\linewidth}
\hspace*{-8mm}
\includegraphics[width=6cm]{plot4.pdf}
\caption{Evolution of the elastic, plastic, hardening and surface energies for the 2D traction experiment with backtracking.}
\label{2D1}
\end{minipage}
\hspace*{10mm}
\begin{minipage}[c]{.46\linewidth}
\includegraphics[width=5cm]{rupture_test4.pdf}
\caption{Profile of the v(t,.)-fracture approximation at time t=5.}
\label{2D2}
\vspace*{1cm}
\includegraphics[width=5cm]{plasticite_test4.pdf}
\caption{Profile of the matrix norm of plastic strain $|p(t,.)|$ at time t=5, ($|p(t,.)|=0.008$ (yellow), $|p(t,.)|=0.009$ (green)).}
\label{2D3}
\end{minipage}
\end{figure}
The evolution of Model 3 shows that with this choice of parameters the material is deformed plastically and then cracks.
We reproduce qualitatively the same behavior as that the of 1D traction experiment of \ref{Hardening3} (see Figure~\ref{2D1}).
In Figure~\ref{2D2}, the yellow zone represents the cracked zone.
The magenta zone represents the crack-free zone where $v\sim1$ (see also Figure~\ref{2D3} for plastic strain p).
\subsection{Numerical simulation of the Peltzer and Tapponnier plasticine experiment}
Using Model 3, we reproduce numerically the first stages of the plasticine experiment, see Figure~\ref{rupture1}.
This experiment is meant to model the action of India (as an indenter) on the Tibetan Plateau.
We consider a square domain $\Omega=(0,1)\times (0,1)$, that represents the layer of plasticine, see Figure~\ref{figomega}.
At time $t_h^n$, the indenter is modeled by a Dirichlet boundary condition $u=t_h ^nU_0=(0,t_h^nU_0)$ with $U_0>0$ on $\partial\Omega_3$.
We set $u=0$ on $\partial\Omega_6$ and $u.\vec{n}=0$ on $\partial\Omega_1$.
We use following parameters:
$h=0.05$, $\bigtriangleup x=0.017$, $\varepsilon=0.15$, $K=100$, $k=100$, $\tau=1$, $\nu=0.252$, $U_0=1$.
In Figure~\ref{2D4} we show the fracture profile at $t=2$, which is in good agreement with the plasticine experiment. In Figure~\ref{2D5},
we observe that Model 3 deforms plastically the layer of plasticine and then cracks. Figure~\ref{2D7} indicates the regions of plastic deformation at time $t=2$.
\begin{figure}[htbp!]
\begin{minipage}[c]{.44\linewidth}
\includegraphics[width=5cm]{etape4.pdf}
\caption{Tapponnier and Peltzer's indentation experiment.}
\label{rupture1}
\vspace{0.5cm}
\includegraphics[width=5cm]{omega.pdf}
\caption{Domain $\Omega$ with boundary partition.}
\label{figomega}
\end{minipage} \hfill
\begin{minipage}[c]{.44\linewidth}
\includegraphics[width=5cm]{test4_plasticine.pdf}
\caption{Profile of the v(t,.)-fracture approximation at time t=2.}
\label{2D4}
\end{minipage}
\end{figure}
\begin{figure}[htbp!]
\begin{minipage}[c]{.44\linewidth}
\hspace*{-10mm}
\includegraphics[width=7cm]{plot4_plasticine.pdf}
\caption{Evolution of the elastic, plastic, hardening and surface energies for the 2D plasticine experiment with backtracking.}
\label{2D5}
\end{minipage} \hfill
\begin{minipage}[c]{.46\linewidth}
\includegraphics[width=5cm]{plasticine4_test.pdf}
\caption{Profile of the matrix norm of plastic strain $|p(t,.)|$ at time t=2. ($|p(t,.)|=0.01$ (blue), $|p(t,.)|=0.003$ (green)).}
\label{2D7}
\end{minipage}
\end{figure}
\section{Conclusion}
In this work, we study 3 models of evolution for materials that can exhibit several dissipation mechanisms: fracture, plasticity, viscous dissipation.
The evolution is defined via a time discretization: at each time step, we seek to minimize a global energy with respect to the variables $(u,p,v)$.
We have reported numerical experiments that show that Models 1 and 3 are most versatile: in particular we can observe evolutions where such materials become
plastic and then crack
In a forthcoming work, we show that we can pass to the limit in Model 1 as time step tends to 0 and give an existence result for a continuous evolution (E1)-(E6).
\section*{Acknowledgements}
The authors wish to express their gratitude to G. Francfort for the fruitful and enlightening discussions.
This work has been supported by a grant from Labex OSUG@2020 (Investissements d'\,avenir – ANR10 LABX56).
|
\section{Introduction}
The \emph{Eulerian polynomial}, $A_n(x)$, $n \in \mathbb{N}$, may be defined as the generating polynomial for the \emph{descent statistic} over the symmetric group $\mathfrak{S}_n$;
$$
A_n(x) = \sum_{\sigma \in \mathfrak{S}_n} x^{\des(\sigma)+1},
$$
where $\des(\sigma)= |\{ i : \sigma(i)>\sigma(i+1)\}|$. Another important statistic which has the same distribution as descents is the number of \emph{excedances},
$\exc(\sigma)= |\{ i : \sigma(i) >i\}|$.
Eulerian polynomials are among the most studied families of polynomials in combinatorics. There are several multivariate extensions of the Eulerian polynomials. We are interested in the one that refines
the excedance statistic over permutations by the position of excedances. This multivariate refinement was used in
conjunction with stable polynomials in recent papers to solve the monotone column permanent conjecture \cite{Mon}. Similar methods were employed
to generalize properties---such as recurrences, and zero location---for several variants of Eulerian polynomials
for Stirling permutations \cite{HV2}, barred multiset permutations \cite{Chen13}, signed and colored permutation \cite{VW13}.
In statistical mechanics the same multivariate Eulerian polynomials appear in connection with an important and much studied Markov process called the
\emph{asymmetric exclusion process} (ASEP), which models particles hopping left and right on a one-dimensional lattice, see \cite{CW1,CW2,ansatz,Lig} and the references therein.
The ASEP has a unique stationary distribution, and Corteel and Williams \cite{CW1} showed that its partition function is a $q$--analog of the above mentioned multivariate
Eulerian polynomial (whenever $\alpha=\beta=1$ and $\gamma=\delta=0$, where $\alpha, \beta, \gamma$ and $\delta$ are as described in Section \ref{secEP}). \\[1ex]
In this paper we show that the multivariate partition function of the stationary distribution of the ASEP with parameters $\alpha = r/(r-1)$ and $\beta = r$
is a $q$-analog of the multivariate Eulerian polynomials for $r$-colored permutations recently introduced in \cite{VW13}. This contains the signed permutations as special case for $r=2$. We also give a new combinatorial interpretation of the stationary distribution in terms of colored permutations for all $q, \alpha, \beta \geq 0$ and $\gamma=\delta=0$. Previous combinatorial interpretations of the stationary distributions have been proved using either permutation tableaux \cite{CW1}, or staircase tableaux \cite{CW2}, while our proof uses directly the (matrix) ansatz for the ASEP of Liggett \cite{Lig}, which was later rediscovered by Derrida, Evans, Hakim and Pasquier \cite{ansatz}.
Our methods extend certain properties of the excedance set statistic studied for the case of permutations by Ehrenborg and Steingr\'imsson \cite{ES00}, as well as the
plane alternative trees of Nadeau \cite{Nad11}, to the case of $r$-colored permutations.
We also point to negative dependence properties and zero restrictions of the multivariate partition function satisfied by stationary distribution for $q=1$ which follow from general theorems obtained by Borcea, Liggett and the first author in \cite{BBL}, and Wagner \cite{Wa3}. These negative dependence properties and zero restrictions generalize recent results of Hitczenko and Janson \cite{HiJa}. We speculate in what negative dependence properties may hold for $q \neq 0$.
\section{Exclusion processes}
\label{secEP}
We focus on a class of exclusion processes that model particles jumping on a finite set of sites, labeled by $[n] := \{1,2,\ldots, n\}$. Given a matrix $Q= (q_{ij})_{i,j=1}^n$ of nonnegative numbers and vectors $b=(b_i)_{i=1}^n$ and $d=(d_i)_{i=1}^n$ of nonnegative numbers, define a continuous time Markov chain on $\{0,1\}^n$ as follows. Let $\eta \in \{0,1\}^n$ represent the configuration of the particles, with $\eta(i)=1$ meaning that site $i$ is occupied, and $\eta(i)=0$ that site $i$ is vacant. Particles at occupied sites jump to vacant sites at specified rates. More precisely, these are the transitions in the Markov chain:
\begin{itemize}
\item[(J)] A particle jumps from site $i$ to site $j$ at rate $q_{ij}$: The configuration $\eta$ is unchanged unless $\eta(i)=1$ and $\eta(j)=0$, and then only $\eta(i)$ and $\eta(j)$ are exchanged.
\item[(B)] A particle at site $i$ is created (is born) at rate $b_i$: The configuration $\eta$ is unchanged unless $\eta(i)=0$, and then only $\eta(i)$ is changed from a zero to a one.
\item[(D)] A particle at site $i$ is annihilated (dies) at rate $d_i$: The configuration $\eta$ is unchanged unless $\eta(i)=1$, and then only $\eta(i)$ is changed from a one to a zero.
\end{itemize}
The (multivariate) \emph{partition function} of a discrete probability measure $\mu$ on $\{0,1\}^n$ is the polynomial in $\mathbb{R}[x_1,\ldots, x_n]$ defined by
\begin{equation}\label{partitioneq}
Z_\mu(\mathbf{x})= \sum_{\eta \in \{0,1\}^n} \mu(\eta) \mathbf{x}^\eta:=\sum_{\eta \in \{0,1\}^n} \mu(\eta) x_1^{\eta(1)}\cdots x_n^{\eta(n)}.
\end{equation}
Hence a discrete probability measure can be recovered from its partition function.
A special case of the Markov chain described by (J), (B) and (D) above which has been much studied by combinatorialists is the ASEP (on a line). Here
$$
q_{ij} =
\begin{cases}
1 &\mbox{ if } j=i+1, \\
q &\mbox{ if } j=i-1, \mbox{ and }\\
0 &\mbox{ if } |j-i|>1,
\end{cases}
$$
where $q\geq 0$ is a parameter.
Moreover, particles are only allowed to leave and enter at the ends $i=1, n$ ($b_1= \alpha, d_1=\gamma, b_n=\delta, d_n=\beta$). The stationary distributions of the ASEP have been explicitly solved by \cite[Theorem~3.2]{Lig} and later by the Matrix Ansatz of \cite{ansatz}, and also by beautiful combinatorial models such as permutation tableaux, staircase tableaux, and alternative tableaux \cite{CW1,CW2,Vi}.
The following theorem is essentially the Matrix Ansatz of \cite{ansatz} for $\gamma =\delta =0$.
\begin{theorem}\label{az}
Let $\alpha, \beta, q \geq 0$ and $\xi > 0$, and let $\{0,1\}^*$ be the set of words of finite length of zeros and ones.
Define a function $\langle \cdot \rangle : \{0,1\}^* \mapsto \mathbb{R}$ recursively by $\langle \varepsilon\rangle =1$ if $\varepsilon$ is the empty word, and
\[
\langle u10v\rangle =q\langle u01v\rangle + \alpha\beta \xi \langle u1v\rangle + \alpha\beta \xi \langle u0v\rangle , \quad \langle 0v\rangle=\beta\xi\langle v\rangle, \quad \langle u1\rangle =\alpha\xi\langle u\rangle ,
\]
for any $u,v \in \{0,1\}^*$.
Then the partition function of the ASEP on $n$ sites with parameters $\alpha, \beta, q \geq 0$ and $\gamma=\delta = 0$ is equal to a constant multiple of
$$
\sum_{\eta \in \{0,1\}^n} \langle \eta \rangle \mathbf{x}^\eta,
$$
where we identify $\eta$ with the corresponding word $\eta(1)\cdots \eta(n) \in \{0,1\}^*$.
\end{theorem}
In the next few sections we use Theorem \ref{az} to give a purely combinatorial interpretation of the partition function of the
stationary distribution of the ASEP in terms of permutation statistics.
\section{Alternative trees}
Alternative trees were introduced by Nadeau \cite{Nad11} in connection with alternative tableaux, which were used by Viennot \cite{Vi} as combinatorial model for the ASEP. Throughout this section and the next, $S$ is a finite, nonempty and totally ordered set. An unordered rooted tree with vertex set $S$ is called an \emph{alternative tree} if the following three conditions are satisfied
\begin{itemize}
\item[(i)] The root is either $\min S$ or $\max S$.
\item[(ii)] If a vertex is larger than its parent, then it is larger than all of its descendants.
\item[(iii)] If a vertex is smaller than its parent, then it is smaller than all of its descendants.
\end{itemize}
See Fig.~\ref{AltTree} for an example.
\begin{figure}[htp]
$$
\begin{tikzpicture}[scale=0.75, inner sep=1pt]
\node (0) at (2,3) {$0$};
\node (1) at (2,1) {$1$};
\node (2) at (3,0) {$2$};
\node (3) at (0,1) {$3$};
\node (4) at (3,1) {$4$};
\node (5) at (4,1) {$5$};
\node (6) at (1,0) {$6$};
\node (7) at (3,2) {$7$};
\node (8) at (1,2) {$8$};
\draw (3) -- (8) -- (0) -- (7) -- (1) -- (6);
\draw (1) -- (2);
\draw (7) -- (4);
\draw (7) -- (5);
\end{tikzpicture}
$$
\caption{ An alternative tree on the set $[0,8]$ corresponding to the marked cycle $((386214570),0)$, or the permutation $\sigma=
(38)(621457)$. }
\label{AltTree}
\end{figure}
If we write a cycle as a word $a_1 \cdots a_k$, we mean $a_1 \to a_2 \to \cdots \to a_k \to a_1$.
A \emph{marked cycle} on $S$ is a pair $(\sigma, s)$ where $\sigma$ is a cycle on $S$ and
$s \in S$ is either the maximum or minimum of $S$. We write a marked cycle $(\sigma,s)$ as a word
\begin{equation}\label{ws}
\sigma = Ws=W_kx_k\cdots W_2x_2W_1x_1s,
\end{equation}
where the $x_i$'s are defined as follows.
\begin{itemize}
\item
If $s=\min{S}$, then $x_k>\cdots >x_1$ are the \emph{right-to-left maxima} of the word $W$, that is, $x_i$ is larger than all letters in $W_i W_{i-1} x_{i-1}\cdots W_1x_1$ for all $1\leq i \leq k$.
\item If $s=\max{S}$, then $x_k<\cdots <x_1$ are the \emph{right-to-left minima} of $W$, that is, $x_i$ is smaller than all letters in $W_i W_{i-1} x_{i-1}\cdots W_1x_1$ for all $1\leq i \leq k$.
\end{itemize}
Note that the words $W_i$ are implicitly defined.
We will recursively describe a map $T$ from marked cycles on $S$ to alternative trees on $S$, and then extend the construction to a bijection between permutations and alternative trees.
If $\sigma=s$ is a marked cycle on one letter, then $T(\sigma, s)$ is a vertex labeled with $s$. Otherwise write $\sigma =Ws$ as in \eqref{ws}. The root of $T(\sigma,s)$ is $s$, the children of $s$ are $x_1,\ldots, x_k$, and the subtree with $x_i$ as a root is $T(W_i x_i, x_i)$, where we consider $(W_i x_i, x_i)$ a marked cycle on the set of letters of the word $W_ix_i$ for all $1\leq i \leq k$, see Fig.~\ref{AltTree}.
By a straightforward induction argument, it follows that the map $T$ described above is a bijection between alternative trees on $S$ and marked cycles on $S$.
\begin{proposition}\label{cycleprop}
Let $(\sigma , s)$ be a marked cycle of length greater than $1$, and let $p(i)$ denote the parent of $i \in S \setminus \{s\}$ in $T=T(\sigma, s)$. If $i<\sigma(i)$, then
$$
p(i)= \max\{\sigma(i), \sigma^2(i), \ldots, \sigma^{k}(i)\}, \mbox{ where } k+1=\min\{j>0 : i\geq \sigma^j(i)\}.
$$
If $i > \sigma(i)$, then
$$
p(i)= \min\{\sigma(i), \sigma^2(i), \ldots, \sigma^{k}(i)\}, \mbox{ where } k+1=\min\{j>0 : \sigma^j(i)\geq i\}.
$$
\end{proposition}
\begin{proof}
Let $p(i)$ denote the parent of $i \in S \setminus \{s\}$ in $T=T(\sigma, s)$. Note that either
\begin{itemize}
\item[(a)]
$i$ is one of the $x_j$'s in \eqref{ws}, or;
\item[(b)]
the parent of $i$ in $T(\sigma, s)$ is also the parent of $i$ in $T(W_jx_j, x_j)$ for the word $W_j$ which $i$ is a letter of.
\end{itemize}
In case (a) the description of $p(i)$ is obviously correct. But then it is also true in general (by induction on $|S|$) in light of (b).
\end{proof}
It will be convenient to depict (the tree associated to) a marked cycle as a \emph{diagram} of arcs: Order $S$ on a line. If $i<j$ and $i$ is a child of $j$, we draw an arc between $i$ and $j$ \emph{above} the line. If $i<j$ and $j$ is a child of $i$, we draw an arc between $i$ and $j$ \emph{below} the line. Hence to each marked cycle we associate a diagram, see Fig.~\ref{Figdi}.
\begin{figure}[htp]
\begin{center}
\includegraphics[width=7cm]{diagram}
\end{center}
\caption{ The diagram of the marked cycle $((386214570), 0)$, or the permutation $\sigma=
(38)(621457)$. }
\label{Figdi}
\end{figure}
Let $\mathcal{O}_n$ be the set of alternative trees on $[0,n+1]:=\{0,1,\ldots, n+1\}$ with $0$ as a root. The set $\mathcal{O}_n$ is in bijection with $\mathfrak{S}_{n+1}$ by the map $T' : \mathfrak{S}_{n+1} \rightarrow \mathcal{O}_n$ defined as follows. Let $\sigma_1, \ldots, \sigma_k$ be the cycles of $\pi \in \mathfrak{S}_{n+1}$, and let $s_i$ be the maximal element of $\sigma_i$ for each $1 \leq i \leq k$. Then the root of $T'(\pi)$ is $0$, and the children of the root are $s_1, \ldots, s_k$. The maximal subtree with root $s_i$ is defined to be
$T(\sigma_i,s_i)$, where $T$ is defined above, see Fig.~\ref{AltTree}.
Hence
$
p(i) = 0$ if $i$ is the maximal element in its cycle, and if $i<\sigma(i)$, then by Proposition~\ref{cycleprop}
$$
p(i)= \max\{\sigma(i), \sigma^2(i), \ldots, \sigma^{k}(i)\}, \mbox{ where } k+1=\min\{j>0 : i\geq\sigma^j(i)\},
$$
and if $i>\sigma(i)$, then
$$
p(i)= \min\{\sigma(i), \sigma^2(i), \ldots, \sigma^{k}(i)\}, \mbox{ where } k+1=\min\{j>0 : \sigma^j(i)\geq i\}.
$$
Recall that $i \in [n]$ is an \emph{excedance} of $\sigma \in \mathfrak{S}_{n+1}$ if $\sigma(i)>i$. Let $\mathcal{X}(\sigma)$ denote the set of excedances of $\sigma$.
A non-root vertex $i$ of a alternative tree $\mathcal T$ is an \emph{excedance} if
$i$ is smaller than its parent $p(i)$. Define three statistics on alternating trees. A pair $1\leq i<j \leq n$ is \emph{yin-yang} in $\mathcal T$ if $p(j)<i<j<p(i)$, see Fig.~\ref{picyy}. Let $\yy(\mathcal T )$ denote the number of yin-yang pairs in $\mathcal T$. Let $c_0(\mathcal T)$ be the number of children of $0$, and $c_1(\mathcal T)$ the number of children of $n+1$.
\begin{figure}[htp]
\setlength{\unitlength}{1.0cm}
\begin{picture}(4,2)(0,-.5)
\put(0,0){\circle*{0.2}}
\put(1,0){\circle*{0.2}}
\put(2,0){\circle*{0.2}}
\put(3,0){\circle*{0.2}}
\put(0.2,0.5){\makebox(0,0){$p(j)$}}
\put(1.3,0){\makebox(0,0){$i$}}
\put(2.3,0){\makebox(0,0){$j$}}
\put(3.2,0.5){\makebox(0,0){$p(i)$}}
\qbezier(0,0)(1,-1)(2,0)
\qbezier(1,0)(2,1)(3,0)
\end{picture}
\caption{ A yin-yang pair $(i,j)$ in the diagram of a permutation/tree. }
\label{picyy}
\end{figure}
Define $\mathcal{X}(\mathcal T) \in \{0,1\}^n$ by $\mathcal{X}(\mathcal T)(i)=1$ if and only if $i$ is an excedance in $\mathcal T$, and define a map $[\cdot]: \{0,1\}^* \rightarrow \mathbb{R}[a,b,q]$ by
\begin{equation}\label{brack1}
[\eta] = \sum a^{c_0(\mathcal T)}b^{c_1(\mathcal T)}q^{\yy(\mathcal T)},
\end{equation}
where the sum is over all $\mathcal T \in \cup_{n \geq 0} \mathcal{O}_n$ with $\mathcal{X}(\mathcal T) = \eta$.
Let $\sigma = Ws$, $s=n+1$, be the (marked) cycle of $\pi \in \mathfrak{S}_{n+1}$ containing $n+1$. Then
\begin{align*}
c(\pi) &:= \# \mbox{ cycles of } \pi, \\
c'(\pi) &:= \# \mbox{ right-to-left minima of } W \\
\yy(\pi) &:= \yy(T'(\pi)).
\end{align*}
Note that $\yy(\pi)$ may be intrinsically defined by the description of $p\circ T'$ above.
The following theorem is now an immediate consequence of the above bijection.
\begin{theorem}
Let $\mathcal T=T'(\pi)$, where $\pi \in \mathfrak{S}_{n+1}$. Then
$$(\mathcal{X}(\mathcal T), c_0(\mathcal T), c_1(\mathcal T), \yy(\mathcal T)) = (\mathcal{X}(\pi), c(\pi), c'(\pi), \yy(\pi)),$$
where we consider $\mathcal{X}(\pi)$ as an element of $\{0,1\}^n$.
\end{theorem}
\begin{theorem}\label{solsn}
Let $u,v \in \{0,1\}^*$ and let $[\cdot]$ be defined as in \eqref{brack1}. Then
$$
[u10v]=q[u01v] +[u1v]+[u0v] \quad \mbox{ and } \quad [0v]=a[v], [u1] = b[u].
$$
\end{theorem}
\begin{proof}
Two vertices in a tree are \emph{comparable} if one of them is a descendant of the other. Let $i=|u|+1$ and
suppose $\eta(i) \neq \eta(i+1)$.
We claim that if $i$ and $i+1$ are comparable, then
\begin{itemize}
\item[(a)] $i$ is a leaf, $p(i)=i+1$ and $p(i+1)<i$, or;
\item[(b)] $i+1$ is a leaf, $p(i+1)=i$ and $p(i)>i+1$.
\end{itemize}
To prove the claim first note that $i$ and $i+1$ cannot be comparable if $\eta(i)=0$ and $\eta(i+1)=1$, since then $i$ is larger than it's descendants while $i+1$ is smaller than it's descendants. Consider the case when $\eta(i)=1$ and $\eta(i+1)=0$. Then $p(i)>i$ and $p(i+1)<i+1$. If $i$ is a descendent of $i+1$, then $i$ has to be a leaf since $i+1$ is larger then it's descendants while $i$ is smaller than all it's descendants. But then $p(i)=i+1$, since otherwise $p(i)>i+1$ which means that $i+1$ has a descendant which is larger than $i+1$. By a similar argument if $i+1$ is a descendant of $i$, then $i+1$ is a leaf and $p(i+1)=i$.
For $\eta \in \{0,1\}^*$, let $\mathcal{O}(\eta)=\{\mathcal T \in \cup_{n \geq 0} \mathcal{O}_n : \mathcal{X}(\mathcal T)=\eta\}$. Hence in both (a) and (b) above we have $\mathcal T \in \mathcal{O}(u10v)$.
Define a map
$\phi : \mathcal{O}(u10v) \rightarrow \mathcal{O}(u01v) \cup \mathcal{O}(u1v) \cup \mathcal{O}(u0v)$ as follows.
If $i$ and $i+1$ are non-comparable, then $\phi(\mathcal T)$ is obtained by switching the labels $i$ and $i+1$ in the tree. By the above claim this is a bijection between set of trees in $\mathcal{O}(u10v)$ for which $i$ and $i+1$ are non-comparable and $\mathcal{O}(u01v)$. Moreover, the yin-yang
$p(i+1)<i<i+1<p(i)$ is destroyed by $\phi$ and no other yin-yangs are destroyed or created.
If $i$ and $i+1$ are comparable, then $\phi(\mathcal T)$ is obtained by contracting the edge between $i$ and $i+1$ while keeping the label $i$ and then relabel the vertices with $[0,n]$ so that the relative order is preserved. Then $\phi$ is a bijection between the set of trees satisfying
(a) and $\mathcal{O}(u0v)$, and $\phi$ is a bijection between the set of trees satisfying
(b) and $\mathcal{O}(u1v)$. No yin-yangs are created or destroyed. This establishes the first equation.
If $1$ is not an excedance, then $1$ is a leaf and a child of the root. Hence we may contract the edge between $0$ and $1$ and relabel the vertices, which shows $[0v]=a[v]$.
Similarly, if $n$ is an excedance, then $n$ is a leaf and a child of $n+1$. Hence we may contract this edge and relabel the vertices, which proves $[u1]=b[u]$.
\end{proof}
\begin{theorem}\label{asep}
The multivariate partition function of the stationary distribution of the ASEP on $n$ sites with parameters $\alpha, \beta >0$, $q\geq 0$ and $\gamma=\delta=0$ is a constant multiple of
$$
\sum_{\pi \in \mathfrak{S}_{n+1}} \alpha^{-c(\pi)} \beta^{-c'(\pi)} q^{\yy(\pi)} \prod_{i \in \mathcal{X}(\pi)} x_i.
$$
\end{theorem}
\begin{proof}
The theorem follows immediately by comparing Theorem~\ref{az} and Theorem~\ref{solsn}.
\end{proof}
Note that a different combinatorial interpretation in terms of permutation statistics of the univariate partition function (equivalent to setting all $x_i$'s equal in Theorem~\ref{asep}) of the ASEP on $n$ sites with parameters $\alpha, \beta >0$, $q\geq 0$ and $\gamma=\delta=0$ was recently obtained by Josuat-Verg\`es \cite{J-V}.
\section{Decorated alternative trees and colored permutations}
Let $r$ be a positive integer. Consider
$$
\mathbb{Z}_r \wr \mathfrak{S}_n=\{ (\kappa, \sigma) \mid \kappa : [n] \rightarrow \mathbb{Z}_r \mbox{ and } \sigma \in \mathfrak{S}_n\},
$$ the \emph{wreath product} of the symmetric group of order $n$ with a cyclic group of order $r$. The elements
of the $\mathbb{Z}_r\wr \mathfrak{S}_n$ are often referred to as $r$-\emph{colored permutations}.
There are several different ways of defining excedances for wreath products. For our purposes, the definition of Steingr\'imsson \cite{Stein94} is the most suitable choice.
Let $\pi =(\kappa, \sigma) \in \mathbb{Z}_r \wr \mathfrak{S}_n$. Define the \emph{excedance set}, $\mathcal{X}(\pi)$, and the \emph{anti-excedance set}, $\mathcal{Y}(\pi)$, by
$$i \in \mathcal{X}(\pi) \ \ \ \ \mbox{ if and only if } \ \ \ \ \begin{cases}
\sigma(i) > i, \mbox{ or}; \\
\sigma(i)=i \mbox{ and } \kappa_i \neq 0.
\end{cases}
$$
and
$$\sigma(i) \in \mathcal{Y}(\pi) \ \ \ \ \mbox{ if and only if } \ \ \ \ \begin{cases}
\sigma(i) < i, \mbox{ or}; \\
\sigma(i)=i \mbox{ and } \kappa_i=0.
\end{cases}
$$
Let $\pi = (\kappa , \sigma) \in \mathbb{Z}_r \wr \mathfrak{S}_n$ and consider the cycle decomposition of $\sigma$. A cycle $c$ of $\sigma$ is called a \emph{zero cycle} if $\kappa_i=0$ for the maximal element $i$ of $c$, otherwise $c$ is called a \emph{non-zero cycle}.
A \emph{decorated alternative tree} on a finite non-empty set of integers $S$ is an alternative tree on $S$ where the vertices are also colored with $0, \ldots, r-1$, where $r>1$. Hence each vertex in the tree is labeled with an element from $S\times \{0,\ldots, r-1\}$. The coloring should obey the following restrictions. Let $s$ and $t$ be the smallest and largest vertex (with respect to the total order on $S$), respectively.
\begin{itemize}
\item[(a)] The children of the root (which is $s$) all have color zero.
\item[(b)] The largest vertex of a maximal subtree whose root is a child of $t$ has non-zero color.
\item[(c)] The root $s$ has color $1$, and $t$ has color $0$.
\end{itemize}
Let $\mathcal{O}_n^r$ be the set of decorated trees on $S=[0,n+1]$ with permitted colors $\{0, \ldots, r-1\}$. Then
$\mathcal{O}_n^r$ is in bijection with $\mathbb{Z}_r \wr \mathfrak{S}_n$ by the mapping $T''$ described below. Start with the tree consisting of a single vertex $0$ (as a root) and attach $n+1$ to it. Give them colors $1$ and $0$, respectively.
Let $\sigma_1, \ldots, \sigma_k$ be the zero cycles, and let $s_i$ be the maximal element of $\sigma_i$. The children of $0$ are $n+1$ and $s_1, \ldots, s_k$. The maximal subtree with root $s_i$ is defined to be $T(\sigma_i,s_i)$, where $T$ is defined above. Now assign colors according to $\kappa$.
Let $\tau_1, \ldots, \tau_\ell$ be the non-zero cycles, and let $t_i$ be the minimal element of
$\tau_i$. The children of $n+1$ are $t_1, \ldots, t_\ell$.
The maximal subtree with root $t_i$ is defined to be $T(\tau_i,t_i)$, where $T$ is defined above.
Assign colors according to $\kappa$. See Fig.~\ref{DecTree} for an example.
\begin{figure}[htp]
$$
\begin{tikzpicture}[scale=1, inner sep=2pt]
\node (0) at (2,3) {$0^1$};
\node (1) at (4.5,1) {$7^4$};
\node (2) at (2.5,1) {$2^2$};
\node (3) at (3.5,2) {$1^8$};
\node (4) at (4,0) {$4^2$};
\node (5) at (5,0) {$5^0$};
\node (6) at (3.5,1) {$6^1$};
\node (7) at (1,2) {$3^0$};
\node (8) at (3.5,2.8) {$8^0$};
\draw (0)--(7);
\draw(0) -- (8) --(3) -- (2);
\draw (3) -- (6);
\draw (3) -- (1);
\draw (1) -- (5);
\draw (1) -- (4);
\end{tikzpicture}
$$
\caption{ An alternative tree on the set $[0,8],$ with root $0,$ corresponding to the colored permutation $\pi=(3^0)(4^2 5^0 7^4 6^1 2^2 1^8)$. }
\label{DecTree}
\end{figure}
The definitions of $\mathcal{X}$ and $\yy$ are the same as for the non-decorated trees. For a decorated tree $\mathcal{T} \in \mathcal{O}_n^r$, $c_0(\mathcal{T})$ is defined as the number of children of 0 minus one, while $c_1(\mathcal{T})$ is the number of children of $n+1$. Define $[\cdot]_r: \{0,1\}^* \rightarrow \mathbb{R}[a,b,q]$ by
\begin{equation}\label{brackr}
[\eta]_r = \sum a^{c_0(\mathcal T)}b^{c_1(\mathcal T)}q^{\yy(\mathcal T)},
\end{equation}
where the sum is over all $\mathcal T \in \cup_{n \geq 0} \mathcal{O}_n^r$ with $\mathcal{X}(\mathcal T) = \eta$.
\begin{theorem}\label{dequi}
Let $\mathcal T=T''(\pi)$, where $\pi \in \mathbb{Z}_r \wr \mathfrak{S}_n$. Then
$$(\mathcal{X}(\mathcal T), c_0(\mathcal T), c_1(\mathcal T), \yy(\mathcal T)) = (\mathcal{X}(\pi), \nc(\pi), \zc(\pi), \yy(\pi)),$$
where we consider $\mathcal{X}(\pi)$ as an element of $\{0,1\}^n$.
\end{theorem}
\begin{theorem}\label{slow} Let $u,v \in \{0,1\}^*$, and let $[\cdot]_r$ be defined as in \eqref{brackr}. Then
$$
[u10v]_r=q[u01v]_r+ r[u1v]_r+r[u0v]_r, \quad [0v]_r=a[v]_r, \quad [u1]_r=b(r-1)[u]_r.
$$
\end{theorem}
\begin{proof}
The definition of $\phi : \mathcal{O}(u10v) \rightarrow \mathcal{O}(u01v) \cup \mathcal{O}(u1v) \cup \mathcal{O}(u0v)$ is almost the same as for the non-decorated case, we just have to specify how the colors are effected. For the non-comparable case, the color of $i$ and $i+1$ are swapped, so that the color stay at the same place in the tree. When we contract an edge in the comparable case we keep the color of $i+1$. Hence the map in the comparable case is $r$ to $1$, which explains the factor $r$ in the equation.
If $1$ is not an excedance, then $1$ is a leaf, has color zero, and $1$ is a child of the root. Hence we may contract the edge between $0$ and $1$ and relabel the vertices, which shows $[0v]_r=a[v]_r$.
Similarly, if $n$ is an excedance, then $n$ is a leaf of non-zero color, and $n$ is a child of $n+1$. Hence we may contract this edge and relabel the vertices, which proves $[u1]_r=b(r-1)[u]_r$.
\end{proof}
\begin{theorem}\label{maincolor}
The multivariate partition function of the stationary distribution of the ASEP on $n$ sites with parameters $\alpha, \beta >0$, $q\geq 0$ and $\gamma=\delta=0$ is a constant multiple of
$$
\sum_{\sigma \in \mathbb{Z}_r \wr \mathfrak{S}_{n}}
\left(\frac{r}{\alpha}\right)^{\nc(\sigma)}
\left(\frac{r}{(r-1)\beta}\right)^{\zc(\sigma)} q^{\yy(\sigma)} \prod_{i \in \mathcal{X}(\sigma)} x_i,
$$
where $r\geq 2$ is an integer.
In particular, when $\alpha = r$ and $\beta = r/(r-1)$, then the partition function is a constant multiple of
$$
\sum_{\sigma \in \mathbb{Z}_r \wr \mathfrak{S}_{n}}
q^{\yy(\sigma)} \prod_{i \in \mathcal{X}(\sigma)} x_i,
$$
\end{theorem}
\begin{proof}
The theorem follows immediately by using Theorem \ref{dequi}, and comparing Theorem~\ref{az} with Theorem~\ref{slow}.
\end{proof}
\section{Multivariate Eulerian polynomials and stability}
The Eulerian polynomials are important in enumerative and algebraic combinatorics and their generalizations to finite
Coxeter groups, wreath products and partially ordered sets have been studied frequently. An important property, first noted by Frobenius \cite{Fro}, is that all zeros of $A_n(x)$ are real. This result has subsequently been lifted to different generalizations of Eulerian polynomials \cite{Brenti, SV13}.
Recently efforts have been made to generalize Frobenius' result in yet another direction, namely to multivariate polynomials \cite{VW13}. A notion of ``real-rootedness'' that has been fruitful in several settings is the following. A polynomial $P(x_1, \ldots, x_n)$ is \emph{stable} if $P(x_1, \ldots, x_n) \neq 0$ whenever $\Im(x_j)>0$ for all $1 \leq j \leq n$. For applications to the ASEP we find it convenient to define the multivariate generalization in terms of excedances.
There is a strong relationship between symmetric exclusion processes and stability which was first proved in \cite{BBL} (without (B) and (D)) and in \cite{Wa3} (with (B) and (D)).
\begin{theorem}\label{prest}
Consider the Markov chain described by (J), (B) and (D) in Section~\ref{secEP}, with $Q$ symmetric. If the partition function of the initial distribution is stable, then the distribution is stable for all $t \geq 0$.
\end{theorem}
\begin{corollary}\label{stabcor}
Consider the Markov chain described by (J), (B) and (D) in Section~\ref{secEP}, with $Q$ symmetric. If the Markov chain is irreducible and positive recurrent, then the partition function of the (unique) stationary distribution is stable.
\end{corollary}
\begin{proof}
Choose an initial distribution with stable partition function, for example a product measure. Then the partition function, $Z_t(\mathbf{x})$, of the distribution at time $t$ is stable for all $t>0$ by Theorem \ref{prest}. The partition function of the stationary distribution is given by $\lim_{t \to \infty}Z_t(\mathbf{x})$, and hence the corollary follows from Hurwitz' theorem on the continuity of zeros (see \cite[Footnote 3, p.~96]{COSW} for a multivariate version).
\end{proof}
As an immediate corollary of Theorem \ref{maincolor} and Corolllary \ref{stabcor} we have.
\begin{corollary}\label{consta}
Let $n$ and $r$ be positive integers and $a$ and $b$ nonnegative real numbers. Then the polynomial
\begin{equation}\label{brackxi}
\sum_{ \pi \in \mathbb{Z}_r \wr \mathfrak{S}_n}a^{\nc(\pi)}b^{\zc(\pi)}\prod_{i \in \mathcal{X}(\pi)}x_{i},
\end{equation}
where $\zc(\pi)$ and $\nc(\pi)$ denotes the number of zero cycles and non-zero cycles of $\pi$, respectively, is stable.
\end{corollary}
This corollary generalizes a recent theorem of Hitczenko and Janson \cite[Theorem 4.5]{HiJa} who proved that the univariate polynomials obtained by setting all the $x_i$'s in \eqref{brackxi} equal are real--rooted.
We shall now see how Corollary \ref{consta} can be generalized further by introducing a new set of variables.
Let
$$
F_n = F_{n,r}(\mathbf{x},\mathbf{y}, a,b)= \sum_{ \pi \in \mathbb{Z}_r \wr \mathfrak{S}_n}a^{\nc(\pi)}b^{\zc(\pi)}\prod_{i \in \mathcal{X}(\pi)}x_{i} \prod_{j \in \mathcal{Y}(\pi) } y_j.
$$
\begin{theorem}\label{difff}For positive integers $n$ and $r$,
\begin{equation}\label{nrxyab}
F_n = (a(r-1)x_1 + by_1)F^*_{n-1}+ rx_1y_1\sum_{j=2}^{n} \left(\frac{ \partial }{\partial x_j} + \frac{ \partial }{\partial y_j}\right) F^*_{n-1},
\end{equation}
where $F^*_{n-1}$ is obtained from $F_{n-1}$ by the changes of variables $x_i \to x_{i+1}$ and $y_i \to y_{i+1}$ for all $1\leq i \leq n-1$, and $F_0 =1$.
\end{theorem}
\begin{proof} Consider a colored permutation $\pi^* = (\kappa^*,\sigma^*)$, where $\sigma^*$ is a permutation of $\{2,\ldots, n\}$ and $\kappa^* : \{2, \ldots, n\} \to \mathbb{Z}_r$. When $\sigma^*$ is written as a product of cycles and $i \mapsto j$ (i.e., $\sigma^*(i) = j$), then we record $x_i$ if $i<j$ and $y_j$ if $j<i$. If $i \mapsto i$ is a fixed point, then we record $x_i$ if $\kappa^*_i \neq 0$, and $y_i$ otherwise.
To create an element $\pi=(\kappa, \sigma) \in \mathbb{Z}_r \wr \mathfrak{S}_n$ from $\pi^*$ we insert $1$ into $\sigma^*$ and choose its color $\kappa_1$.
Hence inserting $1$ between $i \mapsto j$ in an existing cycle will have the effect of taking the derivative with respect to $x_i$ or $y_j$ depending on
whether $i \in \mathcal{X}(\pi^*)$ or $ j \in \mathcal{Y}(\pi^*)$, respectively,
and multiplying by $x_1y_1$ since $1 \in \mathcal{X}(\pi) \cap \mathcal{Y}(\pi)$.
Note also that the assignment of variables to the arrows $i \mapsto j$ is injective. This explains the second term on the right hand side of \eqref{nrxyab}. If we make $1$ a fixed point we either create a new non-zero cycle and an excedance (if $\kappa_1 \neq 0$), or a zero cycle and an anti-excedance (if $\kappa_1 =0$). This explains the first term on the right hand side of \eqref{nrxyab}.
\end{proof}
\begin{remark} Theorem \ref{difff} can be seen as a relation satisfied by the stationary distributions of the ASEP with $q-1=\gamma=\delta =0$. Is there a similar relation for general $\alpha, \beta, \gamma, \delta$?
\end{remark}
\begin{remark}
In \cite{VW13} a multivariate extension of the Eulerian polynomials for wreath products was defined in terms of descent-- and ascent bottoms.
By the recursion given in the proof of Theorem 3.15 in \cite{VW13}, we see that for $a=b=1$, their polynomials are the same as ours up to a reindexing of the variables.
\end{remark}
\begin{theorem}\label{stabF}
Let $r>1$ and $n$ be positive integers and $a,b \geq 0$. Then $F_{n,r}(\mathbf{x},\mathbf{y}, a,b)$ is stable.
\end{theorem}
\begin{proof}
By \eqref{nrxyab},
$
F_n = T(F_{n-1}^*)
$,
where $T$ is the linear operator
$$
T= a(r-1)x_1 + by_1+ rx_1y_1\sum_{j=2}^{n} \left(\frac{ \partial }{\partial x_j} + \frac{ \partial }{\partial y_j}\right).
$$
Since $F_1= a(r-1)x_1+by_1$ is obviously stable it remains to prove that $T$ preserves stability. By the characterization of stability preservers, \cite[Theorem~2.2]{LYPSI}, this is the case if the polynomial
$$
G_T = T\left( (x_1+z_1)\cdots (x_n+z_n)(y_1+w_1)\cdots (y_n+w_n)\right),
$$
is stable (in $4n$ variables). Here $T$ acts on the $x$- and $y$-variables
and treats the $z$- and $w$-variables as constants. Now
$$
\frac { G_T } { x_1y_1 \prod_{j=1}^n(x_j+z_j)(y_j+w_j)}= \frac {a(r-1)}{y_1}+ \frac b {x_1}+ \sum_{j=2}^n \left( \frac 1 {x_j + z_j} + \frac 1 {y_j + w_j} \right).
$$
Each term on the right hand side of the above equation has negative imaginary part whenever all variables have positive imaginary parts. Hence $G_T$ is stable and the theorem follows.
\end{proof}
\section{Negative dependence}
Negative dependence in probability theory models repelling particles. There are many correlation inequalities of varying strength that model negative dependence, see \cite{BBL,Pem}. For example, a discrete probability measure $\mu$ on $\{0,1\}^n$ is \emph{negatively associated} if
\begin{equation}\label{na}
\int f g d\mu \leq \int f d\mu \int g d\mu,
\end{equation}
whenever $f,g : \{0,1\}^n \rightarrow \mathbb{R}$ depend on disjoint sets of variables (i.e., $f$ depends only on $\{\eta_i : i \in A\}$ and $g$ depends only on $\{\eta_j : j \in B\}$, where $A\cap B = \emptyset$). In particular if $\mu$ is negatively associated, then it is \emph{pairwise negatively correlated}, i.e., for distinct $i,j \in [n]$:
$$
\mu(\eta(i)=\eta(j)=1) \leq \mu(\eta(i)=1) \mu(\eta(j)=1),
$$
which is obtained from \eqref{na} by setting
$$
f(\eta) = \begin{cases} 1 &\mbox{ if } \eta(i)=1 \\
0 &\mbox{ otherwise}
\end{cases} \ \ \ \mbox{ and } \ \ \
g(\eta) = \begin{cases} 1 &\mbox{ if } \eta(j)=1 \\
0 &\mbox{ otherwise}
\end{cases}.
$$
It was proved in \cite{BBL} that if the multivariate partition function of a discrete probability measure $\mu$ is stable (such measures are called \emph{strong Rayleigh}), then it satisfies several of the strongest correlation inequalities known to model negative dependence. In particular $\mu$ is negatively associated.
\begin{corollary}
The stationary distribution of the ASEP with $q=1$ and $\alpha,\beta, \gamma, \delta \geq 0$ is negatively associated.
\end{corollary}
Hitczenko and Janson \cite{HiJa} used the real--rootedness of the univariate partition function (obtained by setting $x_1=\cdots=x_n=x$ in \ref{partitioneq}) of the stationary distribution of the ASEP with $q=1$, $\alpha,\beta \geq 0$ and $\gamma=\delta= 0$, to prove concentration inequalities for the corresponding measures. Since we now know that the multivariate partition functions are stable whenever $q=1$ and $\alpha,\beta, \gamma, \delta \geq 0$ there are more refined concentration inequalities available due to Pemantle and Peres \cite{PePe}. A function
$f : \{0,1\}^n \rightarrow \mathbb{R}$ is \emph{Lipschitz-1} if
$$
|f(\eta)-f(\xi)| \leq d(\eta,\xi), \ \ \mbox{ for all } \eta, \xi \in \{0,1\}^n,
$$
where $d$ is the \emph{Hamming distance}.
\begin{theorem}[Pemantle and Peres, \cite{PePe}]
Suppose $\mu$ is a probability measure on $\{0,1\}^n$ whose partition function is stable and has mean $m = \mathbb{E}(\sum_{i=1}^n\eta_i)$. If $f$ is any Lipschitz-1 function on $\{0,1\}^n$, then
$$
\mu( \eta : |f(\eta)-\mathbb{E}f|>a) \leq 5 \exp\left( - \frac {a^2}{16(a+2m)}\right).
$$
\end{theorem}
The case when $f(\eta)=\sum_{i=1}^n \eta_i$ (the number of particles) corresponds to the univariate partition function.
We pose as an open problem to investigate negative dependence properties when $q \neq 1$. In particular:
\begin{question}\label{stbq}
Consider the ASEP on $n$ sites with $\gamma=\delta=0$, $\alpha=\beta=1$ and $q \neq 1$. Is the multivariate partition function of the stationary distribution stable? Is the stationary distribution negatively associated?
\end{question}
Question \ref{stbq} is open even for the case $q = 0$. It is also open whether the univariate partition function is real--rooted, for $\gamma=\delta=0$, $\alpha=\beta=1$ and $q \neq 1$.
However, for $q=0$, we get the $(n+1)$st \emph{Narayana polynomial}, which is known to be real--rooted. This can be seen as supporting evidence for a affirmative answer to Question \ref{stbq} when $q=0$.
It would be interesting if one could find explicit combinatorial models for the exclusion process for classes of labeled graphs which are not necessarily lines. In the symmetric case the partition functions of the stationary distributions (if unique) will be stable by Corollary \ref{stabcor}, and thus the stationary distributions will be negatively associated.
|
\section{Introduction}
\label{sec:intro}
Automatic species classification of birds from their sounds has many potential applications in conservation, ecology and archival \citep{Laiolo:2010,Digby:2013,Ranft:2004}.
However, to be useful it must work with high accuracy across large numbers of possible species, on noisy outdoor recordings and at big data scales.
The ability to scale to big data is crucial:
remote monitoring stations can generate huge volumes of audio recordings \citep{Aide:2013},
and audio archives contain large volumes of audio, much of it without detailed labelling.
For example the British Library Sound Archive holds over 100,000 recordings of bird sound in digital format, from various sources \citep{Ranft:2004}.
Big data scales also imply that methods must work without manual intervention,
in particular without manual segmentation of recordings into song syllables, or into vocal/silent sections.
The lack of segmentation is a pertinent issue for both remote monitoring and archive collections,
since many species of bird may be audible for only a minority of the recorded time,
and therefore much of the audio will contain irrelevant information.
The task of classifying bird sounds by species has been studied by various authors, at least as far back as
\cite{McIlraith:1997}.
(See \cite{Stowell:2010e} for a survey.)
Many of the early studies used small datasets, often noise-free and/or manually-segmented
and with a small number of species, so their practical applicability for ecological applications is unclear.
More recent studies have fewer such limitations, and introduce useful methods customised to the task
\citep{Lakshminarayanan:2009
\citep{Damoulas:2010}%
\citep{Briggs:2012
. However, there remain questions of scalability,
due to the computational intensity of algorithms or to procedures such as all-pairs comparisons which cannot be applied to arbitrarily large datasets without modification \citep{Damoulas:2010}.
In addition to noise-robustness and scalability issues, one further issue
is the number of species considered by a classifier:
certain classification systems may be developed to distinguish among ten or twenty species,
but in many parts of the world there are hundreds of species that might be heard \citep{BTO:2013}.
Further, typical recordings in the wild contain sounds from more than one bird, and so it is advantageous to consider the task as a \textit{multi-label} task,
in which the classifier must return not one label but a set of labels representing all species that are present \citep{Briggs:2012}.
One recent research project (named ``SABIOD'') has provided a valuable stimulus to the research community
by conducting classification challenges evaluated on large datasets of bird sounds collected in the wild,
and with large numbers of species to recognise
\citep{Glotin:2013
\citep{Fodor:2013
\citep{birdclef2014}%
.
The research reported in this paper benefits from the datasets made available through that project,
as well as other datasets, to evaluate bird sound classification suitable for large-scale practical deployments.
Some previous work has compared the performance of different classification algorithms for the task \citep{Acevedo:2009,Briggs:2009}.
In the present work, we instead use a standard but powerful classification algorithm,
and focus on the choice of audio features used as input data.
We introduce the concept of \textit{feature learning} which has been applied in other machine learning domains,
and show that in most cases it can lead the classifier to strongly outperform those using common MFCC and Mel spectrum features.
We also evaluate the role of other aspects such as noise reduction in the feature preprocessing;
however, the strongest effect of the parameters we study comes from replacing MFCCs with learned features.
In the following, we use four large and diverse birdsong datasets with varying characteristics to evaluate classifier performance.
Overall, feature learning enables a classifier to perform very strongly on large datasets with large numbers of species,
and achieves this boost with very little computational cost after the training step.
Three of the four datasets demonstrate clearly the boost attained through feature learning,
attaining very strong performance in both single-label and multi-label classification tasks.
One dataset, consisting of long dawn-chorus recordings with a substantial amount of audio but few annotations,
does not yield a significant benefit from the improved feature representation.
We explore the reasons for this in follow-up experiments in which the training data is augmented or substituted with other data.
Before describing our experiment, however, we discuss the use of spectral features and feature learning for audio classification.
\subsection{Spectral features and feature learning}
\label{sec:kfl}
Raw audio data is not generally suitable input to a classification algorithm:
even if the audio inputs were constrained to a fixed duration (so that the data dimensionality was constant),
the dimensionality of an audio signal (considered as a vector) would be extremely large,
and would not represent sound in such a way that acoustically/perceputally similar sounds would generally be near neighbours in the vector space.
Hence audio data is usually converted to a spectrogram-like representation before processing,
i.e.\ the magnitudes of short-time Fourier transformed (STFT) frames of audio, often around 10 ms duration per frame.
(Alternatives to STFT which have been considered for bird sound classification include linear prediction \citep{Fox:2008}, wavelets \citep{Selin:2007} and chirplets \citep{Stowell:2014}).
An STFT spectrum indicates the energy present across a linear range of frequencies.
This linear range might not reflect the perceptual range of a listener, and/or the range of frequencies at which the signal carries information content,
so it is common to transform the frequency axis to a more perceptual scale, such as the Mel scale originally intended to represent the approximately logarithmic sensitivity of human hearing.
This also reduces the dimensionality of the spectrum, but even the Mel spectrum has traditionally been considered rather high-dimensional for automatic analysis.
A convention, originating from speech processing, is to transform the Mel spectrum using a cepstral analysis and then to keep the lower coefficients (e.g.\ the first 13) which typically contain most of the energy %
\citep{Davis:1980
. These coefficients, the Mel frequency cepstral coefficients (MFCCs), became widespread in applications of machine learning to audio, including bird vocalisations \citep{Stowell:2010e}.
MFCCs have some advantages, including that the feature values are approximately decorrelated from each other,
and they give a substantially dimension-reduced summary of spectral data.
Dimension reduction is advantageous for manual inspection of data,
and also for use in systems that cannot cope with high-dimensional data.
However, as we will see, modern classification algorithms can cope very well with high-dimensional data,
and dimension reduction always reduces the amount of information that can be made available to later processing,
risking discarding information that a classifier could have used.
Further, there is little reason to suspect that MFCCs should capture information optimal for bird species identification:
they were designed to represent human speech, yet humans and birds differ in their use of the spectrum both perceptually and for production.
MFCCs aside, one could use raw (Mel-)spectra as input to a classifier,
or one could design a new transformation of the spectral data that would tailor the representation to the subject matter.
Rather than designing a new representation manually, we consider automatic feature learning.
The topic of feature learning (or representation learning, dictionary learning)
has been considered from many perspectives within the realm of statistical signal processing
\citep{Bengio:2013
\citep{Jafari:2011
\citep{Coates:2012}
\citep{Dieleman:2013
. The general aim is for an algorithm to learn some transformation that, when applied to data, improves performance on tasks such as sparse coding, signal compression or classification.
This procedure may be performed in a ``supervised'' manner, meaning it is supplied with data as well as some side information about the downstream task (e.g.\ class labels), or ``unsupervised'', operating on a dataset but with no information about the downstream task.
A simple example that can be considered to be unsupervised feature learning is principal components analysis (PCA):
applied to a dataset, PCA chooses a linear projection which ensures that the dimensions of the transformed data are decorrelated \citep{Bengio:2013}.
It therefore creates a new feature set, without reference to any particular downstream use of the features,
simply operating on the basis of qualities inherent in the data.
Recent work in machine learning has shown that unsupervised feature learning
can lead to representations that perform very strongly in classification tasks,
despite their ignorance of training data labels that may be available \citep{Coates:2012,Bengio:2013}.
This rather surprising outcome suggests that feature learning methods emphasise patterns in the data that turn out to have semantic relevance,
patterns that are not already made explicit in the basic feature processing such as STFT.
A second surprising aspect is that such representations often perform the opposite of feature reduction,
increasing the dimensionality of the problem without adding any new information:
a deterministic transformation from one feature space to a higher-dimensional feature space
cannot, in an information-theoretic sense, add any information that is not present in the original space.
However, such a transformation can help to reveal the manifold structure that may be present in the data
\citep{Olshausen:2004
. Neural networks, both in machine implemetations and in animals,
perform such a dimension expansion in cases where one layer of neurons is connected as input to a larger layer of neurons
\citep{Olshausen:2004}.
In our study, however, we will not use a feature learning procedure intended to parallel a biological process.
Instead, we use \textit{spherical k-means}, a simple and highly scalable modification of the classic k-means algorithm
\citep{Coates:2012,Dieleman:2013}.
We perform a further adaptation of the algorithm to ensure that it can
run in streaming fashion across large audio datasets, to be described in Section \ref{sec:methods}.
Birdsong often contains rapid temporal modulations, and this information should be useful for identifying species-specific characteristics \citep{Stowell:2014}.
From this perspective, a useful aspect of feature learning is that it can be applied not only to single spectral frames,
but to short sequences (or ``patches'') of a few frames.
The representation can then reflect not only characteristics of instantaneous frequency patterns in the input data, but characteristics of frequencies and their short-term modulations,
such as chirps sweeping upwards or downwards.
This bears some analogy with the ``delta-MFCC'' features sometimes used by taking the first difference in the time series of MFCCs,
but is more flexible since it can represent amplitude modulations, frequency modulations, and correlated modulations of both sorts
(cf.\ \cite{Stowell:2014}).
In our study we tested variants of feature learning with different temporal structures:
either considering one frame at a time (which does not capture modulation),
multiple frames at a time,
or a variant with two layers of feature learning,
which captures modulation across two timescales.
\section{Materials and Methods}
\label{sec:methods}
Our primary experiment evaluated automatic species classification separately across four different datasets of bird sound.
For each dataset we trained and tested a random forest classifier \citep{Breiman:2001},
while systematically varying the following configuration parameters to determine their effect on performance:
\begin{itemize}
\item Choice of features (MFCCs, Mel spectra, or learned features) and their summarisation over time (mean and standard deviation, maximum, or modulation coefficients);
\item Whether or not to apply noise reduction to audio spectra as a pre-processing step;
\item Decision windowing: whether to treat the full-length audio as a single unit for training/testing purposes, or whether to divide it into shorter-duration windows (1, 5 or 60 seconds);
\item How to produce an overall decision when using decision windowing (via the mean or the maximum of the probabilities);
\item Classifier configuration: the same random forest classifier tested in single-label, multilabel or binary-relevance setting.
\end{itemize}
We will say more about the configuration parameters below.
Each of the above choices was tested in all combinations (a ``grid search'' over possible configurations) for each of our datasets separately,
thus providing a rigorous search over a vast number of classifier settings,
in hundreds of individual crossvalidated classification tests.
In follow-up experiments we explored some further issues and their effect on species recognition:
\begin{itemize}
\item We separated out two aspects of our different feature sets---their dimensionality and their intrinsic character---by projecting the feature data to the same fixed dimensionality, and then re-testing with these;
\item We tested the effect of data expansion, by training on the union of two datasets;
\item We tested the effect of cross-condition training, by training on one dataset and testing with a different dataset.
\end{itemize}
\subsection{Datasets}
\begin{table}[t]
\resizebox{\textwidth}{!}{
\begin{tabular}{ l | l r r r r l }
Dataset & Location & Items & Total duration & Mean duration & Classes & Labelling \\
\hline
nips4b & France & 687 & 0.8 hrs (125k frames) & 4 secs & 87 & multilabel \\
xccoverbl & UK/Europe & 264 & 4.9 hrs (763k frames) & 67 secs & 88 & single-label \\
bldawn & UK & 60 & 7.8 hrs (1.2M frames) & 468 secs & 77 & multilabel \\
lifeclef2014 & Brazil & 9688 & 77.8 hrs (12M frames) & 29 secs & 501 & single-label
\end{tabular}
}
\caption{Summary of bird sound datasets used.}
\label{tbl:datasets}
\end{table}
We gathered four datasets, each representing a large amount of audio data and a large number of species to classify (Table \ref{tbl:datasets}).
Two of the datasets (\textit{nips4b} and \textit{lifeclef2014}) consist of the publicly-released training data
from bird classification challenges organised by the SABIOD project \citep{Glotin:2013,birdclef2014}.
The \textit{nips4b} dataset is multilabel (median 1 species per recording, range 0--6);
the \textit{lifeclef2014} dataset is single-label but much larger.
Note that we only use the publicly-released training data from those challenges,
and not any private test data,
and so our evaluation will be similar in nature to their final results but not precisely comparable.
For evaluation we partitioned each of these datasets into two, so that we could run two-fold crossvalidation:
training on one half of the dataset and testing on the other half, and vice versa.
In addition, the British Library Sound Archive has a large collection of environmental sound recordings,
and they made available to us a subset of 60 ``dawn chorus'' recordings.
This consisted of 20 recordings each from three UK-based recordists,
ranging in duration from 2 minutes to 20 minutes,
and annotated by each recordist with a list of species heard (median 6 species per recording, range 3--12).
We refer to this dataset as \textit{bldawn}, and perform three-fold stratified crossvalidation:
for each recordist, we train the system using the data from the other two recordists, and then test on the audio from the held-out recordist.
This stratified approach is useful because it tests whether the system can generalise to recordings from unknown recordists,
rather than adapting to any specifics of the known recordists.
We also gathered a single-label dataset as a subset of the recordings available from the Xeno Canto website,%
\footnote{\url{http://www.xeno-canto.org/}}
covering many of the common UK bird species, and covering at least all the species present in the \textit{bldawn} dataset.
For each species included, we queried Xeno Canto to retrieve three different recordings,
preferring to retrieve recordings from the UK, but allowing the system to return recordings from further afield if too few UK recordings were available.
Our search query also requested high-quality recordings (quality label `A'),
and song rather than calls, where possible.
Since we retrieved three examples for each species, this enabled us to partition the dataset for three-fold crossvalidation:
not stratified into individual recordists (as was \textit{bldawn}), but sampled from a wide range of recordists.
These datasets have widely varying characteristics, for example in the typical duration of the sound files, the recording location,
and the number of classes to distinguish (Table \ref{tbl:datasets}).
Note that most of the datasets have different and irreconcilable lists of class labels:
in particular, for \textit{bldawn} and \textit{xccoverbl} the class label is the species,
whereas \textit{nips4b} and \textit{lifeclef2014} use separate labels for song and calls.
Of our datasets only \textit{bldawn} and \textit{xccoverbl} have strong overlap in their species lists.
Therefore only these datasets could be combined to create larger pools of training data.
In this work we performed automatic classification for each audio file, without any segmentation procedure
to select region(s) of bird vocalisation in the file.
The only segmentation that is done is implicit in the collection processes for the dataset:
for the two datasets originating from Xeno Canto, each audio clip might or might not contain a large amount of silence or other noise, depending on the contributor;
for \textit{nips4b} the audio is collected from remote monitoring stations with no manual selection;
for \textit{bldawn} the audio is selected by the contributor, but not trimmed to a specific vocalisation, instead selected to present a long dawn chorus audio recording.
\subsection{Feature learning method}
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.5\textwidth,clip, trim=40mm 10mm 40mm 10mm]{plots/oskmeansexample-datacent}
\end{center}
\caption{Example of spherical k-means applied to a simple two-dimensional dataset. We generated synthetic 2D data points by sampling from three clusters which were each Gaussian-distributed in terms of their angle and log-magnitude (coloured dots), and then applied our online spherical k-means algorithm to find 10 unit vectors (crosses). These unit vectors form an overcomplete basis with which one could represent this toy data, projecting two-dimensional space to ten-dimensional space.}
\label{fig:oskmeansexample}
\end{figure}
As discussed in Section \ref{sec:kfl}, the aim of unsupervised feature learning is to find some transformation of a dataset,
driven only by the characteristics inherent in that dataset.
For this we use a method that has has shown promise in previous studies, and can be run effectively at big data scales:
\textit{spherical k-means}, described by \cite{Coates:2012} and first applied to audio by \cite{Dieleman:2013}.
Spherical k-means is related to the well-known k-means clustering algorithm \citep{Lloyd:1982},
except that instead of searching for cluster centroids which minimise the Euclidean distance to the data points,
we search for unit vectors (directions) to minimise their angular distance from the data points.
This is achieved by modifying the iterative update procedure for the k-means algorithm:
for an input data point, rather than finding the nearest centroid by Euclidean distance and then moving the centroid towards that data point,
the nearest centroid is found by cosine distance,
\begin{equation}
\text{cosine distance} = 1-\cos(\theta) = 1-{A \cdot B \over \|A\| \|B\|} \, ,
\end{equation}
where $A$ and $B$ are vectors to be compared, $\theta$ is the angle between them, and $\|\cdot\|$ is the Euclidean vector norm.
The centroid is renormalised after update so that it is always a unit vector.
Fig. \ref{fig:oskmeansexample} shows an example of spherical k-means applied to synthetic data.
Spherical k-means thus finds a set of unit vectors which represent the distribution of directions found in the data:
it finds a basis (here an overcomplete basis) so that data points can in general be well approximated as a scalar multiple of one of the basis vectors.
This basis can then be used to represent input data in a new feature space which reflects the discovered regularities,
in the simplest case by representing every input datum by its dot product with each of the basis vectors \citep{Coates:2012,Dieleman:2013}:
\begin{equation}
x'(n,j) = \sum_{i=1}^M{b_j(i) x(n,i)} \, ,
\label{eqn:dotprod1frame}
\end{equation}
where $x$ represents the input data indexed by time frame $n$ and feature index $i$ (with $M$ the number of input features, e.g.\ the number of spectral bins),
$b_j$ is one of the learnt basis vectors (indexed by $j \in [1,k]$),
and $x'$ is the new feature representation.
In our case, the data on which we applied the spherical k-means procedure consisted of Mel spectral frames ($M=40$ dimensions),
which we first normalised and PCA-whitened as in \cite{Dieleman:2013}.
We also tested configurations in which the input data was not one spectral frame but a sequence of them---e.g.\ a sequence of four spectral frames at a time---allowing the clustering to respond to short-term temporal patterns as well as spectral patterns.
We can write this as
\begin{equation}
x'(n,j) = \sum_{\delta=0}^\Delta{\sum_{i=1}^M{b_j(\delta,i) x(n+\delta,i)}} \, ,
\label{eqn:dotprodmultiframe}
\end{equation}
where $\Delta$ is the number of frames considered at a time,
and the $b$ are now indexed by a frame-offset as well as the feature index.
(See Fig. \ref{fig:basesexample} to preview examples of such bases.)
Alternatively, this can be thought of as ``stacking'' frames, e.g.\
stacking each sequence of four 40-dimensional spectral frames to give a 160-dimensional vector,
before applying \eqref{eqn:dotprod1frame} as before.
In all our experiments we used a fixed $k=500$, a value which has been found useful in previous studies \citep{Dieleman:2013}.
The standard implementation of k-means clustering requires an iterative batch process which considers all data points in every step.
This is not feasible for high data volumes.
Some authors use ``minibatch'' updates, i.e.\ subsamples of the dataset.
For scalability as well as for the potential to handle real-time streaming data,
we instead adapted an online streaming k-means algorithm, ``online Hartigan k-means'' \cite[Appendix B]{McFee:2012}.
This method takes one data point at a time, and applies a weighted update to a selected centroid dependent on the amount of updates that the centroid has received so far.
We adapted the method of \cite[Appendix B]{McFee:2012} for the case of spherical k-means.
K-means is a local optimisation algorithm rather than global, and may be sensitive to the order of presentation of data.
Therefore in order to minimise the effect of order of presentation for the experiments conducted here,
we did not perform the learning in true single-pass streaming mode.
Instead, we performed learning in two passes:
a first streamed pass in which data points were randomly subsampled (using reservoir sampling) and then shuffled before applying PCA whitening and starting the k-means procedure,
and then a second streamed pass in which k-means was further trained by exposing it to all data points.
Our Python code implementation of online streaming spherical k-means is available as supplementary information.
As a further extension to the method, we also tested a \textit{two-layer} version of our feature-learning method, intended to reflect detail across multiple temporal scales.
In this variant, we applied spherical k-means feature learning to a dataset,
and then projected the dataset into that learnt space.
We then downsampled this projected data by a factor of 8 on the temporal scale (by max-pooling, i.e.\ taking the max across each series of 8 frames),
and applied spherical k-means a second time.
The downsampling operation means that the second layer has the potential to learn regularities that emerge across a slightly longer temporal scale.
The two-layer process overall has analogies to deep learning techniques,
most often considered in the context of artificial neural networks \citep{Erhan:2010,Bengio:2013},
and to the progressive abstraction believed to occur towards the higher stages of auditory neural pathways.
\subsection{Classification and evaluation}
Our full classification workflow started by converting each audio file to a standard sample-rate of 44.1 kHz.
We then calculated Mel spectrograms for each file, using a frame size of 1024 frames with Hamming windowing and no overlap.
We filtered out spectral energy below 500 Hz, a common choice to reduce the amount of environmental noise present,
and then normalised the root-mean-square (RMS) energy in each spectrogram.
For each spectrogram we then optionally applied the noise-reduction procedure that we had found to be useful in our NIPS4B contest submission \citep{Stowell:2013n},
a simple and common median-based thresholding.
This consists of finding the median value for each spectral band in a spectrogram,
then subtracting this median spectrum from every frame,
and setting any resulting negative values to zero.
This therefore preserves only the spectral energy that rises above the median bandwise energy.
In principle it is a good way to reduce the stationary noise background,
but is not designed to cope well with fluctuating noise.
However its simplicity makes it easy to apply across large datasets efficiently.
The Mel spectrograms, either noise-reduced or otherwise, could be used directly as features.
We also tested their reduction to MFCCs (including delta features, making 26-dimensional data),
and their projection onto learned features,
using the spherical k-means method described above.
For the latter option, we tested projections based on single frame as well as on sequences of
2, 3, 4 and 8 frames,
to explore the benefit of modelling short-term temporal variation.
We also tested the two-layer version based on the repeated application to 4-frame sequences across two timescales.
The feature representations thus derived were all time series.
In order to reduce them to summary features for use in the classifier,
we tested two common and simple techniques:
summarising each feature dimension independently by its mean and standard deviation,
or alternatively by its maximum.
These are widespread but are not designed to reflect any temporal structure in the features
(beyond the fine-scale temporal information that is captured by some of our features).
Therefore, for the Mel and MFCC features we also tested summarising by modulation coefficients:
we took the short-time Fourier transform (STFT) along the time axis of our features,
and then downsampled the spectrum to a size of 10 to give a compact representation of the temporal evolution of the features (cf.\ \cite{Lee:2008}).
The multi-frame feature representations already intrinstically included short-term summarisation of temporal variation,
so to limit the overall size of the experiment, for the learned feature representations we only applied the mean-and-standard-deviation summarisation.
Overall we tested six types of non-learned representation against six types of learned representation (Table \ref{tbl:featurecombis}).
To perform classification on our temporally-pooled feature data,
then, we used a random forest classifier
\citep{Breiman:2001
. Random forests and other tree-ensemble classifiers perform very strongly in a wide range of empirical evaluations
\citep{Caruana:2006
, and were used by many of the strongest-performing entries to the SABIOD evaluation contests
\citep{Glotin:2013,Fodor:2013,Potamitis:2014}.
For this experiment we used the
implementation from the Python \texttt{scikit-learn} project \citep{Pedregosa:2011}.
Note that \texttt{scikit-learn} v0.14 was found to have a specific issue preventing training on large data,
so we used a pre-release v0.15 after verifying that it led to the same results with our smaller datasets.
We did not manually tune any parameters of the classifier.
However, since our experiment covered both single-label and multilabel classification,
we did test three different ways of using the classifier to make decisions:
\begin{enumerate}
\item Single-label classification: this assumes that there is only one species present in a recording. It therefore cannot be applied to multilabel datasets, but for single-label datasets it may benefit from being well-matched to the task.
\item Binary relevance: this divides the multilabel classification task into many single-label tasks, training one separate classifier for each of the potential output labels \citep{Tsoumakas:2010}. This strategy ignores potential correlations between label occurrence, but potentially allows a difficult task to be approximated as the combination of more manageable tasks. Binary relevance is used e.g.\ by \cite{Fodor:2013}.
\item Full multilabel classification: in this approach, a single classifier (here, a single random forest) is trained to make predictions for the full multi-label situation. Predicting presence/absence of every label simultaneously can be computationally difficult compared against a single-label task, and may require larger training data volumes, but represents the full situation in one model \citep{Tsoumakas:2010}.
\end{enumerate}
For each of these methods the outputs from the classifier are per-species probabilities.
We tested all of our datasets using the full multilabel classifier,
then for comparison we tested the single-label datasets using the single-label classifier,
and the multi-label datasets using the binary-relevance classifier.
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.48\textwidth,clip, trim=0mm 0mm 0mm 0mm]{figs/kflworkflowb}
\end{center}
\caption{Summary of the classification workflow, here showing the case where single-layer feature learning is used.}
\label{fig:kflworkflow}
\end{figure}
Some of our datasets contain long audio recordings, yet none of the annotations indicate which point(s) in time each species is heard.
This is a common format for annotations: for example, the \textit{bldawn} annotations are derived directly from the archival metadata,
which was not designed specifically for automatic classification.
Long audio files present an opportunity to make decisions either for the entire file as one scene,
or in smaller ``decision windows'', for which the decisions are then pooled to yield overall decisions.
We tested this empirically, using decision windows of length 1, 5 or 60 seconds or the whole audio.
Each decision window was treated as a separate datum for the purposes of training and testing the classifier,
and then the decisions were aggregated per audio file using either mean or maximum.
The mean probability of a species across all the decision windows is a reasonable default combination;
we compared this against the maximum with the motivation that if a bird is heard only at one point in the audio,
and this leads to a strong detection in one particular decision window, then such a strong detection should be the overriding factor in the overall decision.
For some datasets (\textit{nips4b}) we did not test long windows since all audio files were short;
while for \textit{lifeclef2014} we used only whole-audio classification because of the runtime costs of evaluating these combinations over this largest dataset.
We performed feature learning, training and testing separately for each of our four datasets,
using the appropriate two- or threefold crossvalidation described above,
and across all combinations of the feature settings we have just described.
Fig. \ref{fig:kflworkflow} summarises the main stages of the workflow described.
We evaluated the performance in each experimental run using two measures, the Area Under the ROC Curve (AUC) and the Mean Average Precision (MAP).
The AUC statistic is an evaluation measure for classification/detection systems which has many desirable properties \citep{Fawcett:2006}:
unlike raw accuracy, it is not affected by ``unbalanced'' datasets having an uneven mixture of true-positive and true-negative examples;
and it has a standard probabilistic interpretation,
in that the AUC statistic tells us the probability that the algorithm will rank a random positive instance
higher than a random negative instance.
Chance performance is always 50\% for the AUC statistic.
The AUC is a good statistic to use to evaluate the output from a probabilistic classifier in general,
but for user-facing applications, which may for example show a ranked list of possible hits,
the Mean Average Precision (MAP) statistic leads to an evaluation which relates more closely to user satisfaction with the ranked list.
The key difference in effect is that the AUC statistic applies an equal penalty for a mis-ordering at any position on the ranked list,
whereas the MAP statistic assigns greater penalties higher up the ranking \citep{Yue:2007}.
We therefore calculate both evaluation statistics.
\begin{table}[t]
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{ l | l l r }
Label & Features & Summarisation & Dimension \\
\hline
\texttt{mfcc-ms} & MFCCs (+deltas) & Mean \& stdev & 52 \\
\texttt{mfcc-maxp} & MFCCs (+deltas) & Max & 26 \\
\texttt{mfcc-modul} & MFCCs (+deltas) & Modulation coeffs & 260 \\
\texttt{melspec-ms} & Mel spectra & Mean \& stdev & 80 \\
\texttt{melspec-maxp} & Mel spectra & Max & 40 \\
\texttt{melspec-modul} & Mel spectra & Modulation coeffs & 400 \\
\texttt{melspec-kfl1-ms} & Learned features, 1 frame & Mean \& stdev & 1000 \\
\texttt{melspec-kfl2-ms} & Learned features, 2 frames & Mean \& stdev & 1000 \\
\texttt{melspec-kfl3-ms} & Learned features, 3 frames & Mean \& stdev & 1000 \\
\texttt{melspec-kfl4-ms} & Learned features, 4 frames & Mean \& stdev & 1000 \\
\texttt{melspec-kfl8-ms} & Learned features, 8 frames & Mean \& stdev & 1000 \\
\texttt{melspec-kfl4pl8kfl4-ms} & Learned features, 4 frames, two-layer & Mean \& stdev & 1000 \\
\end{tabular}
}
\end{center}
\caption{The twelve combinations of feature-type and feature-summarisation tested. The feature-type and feature-summarisation method jointly determine the dimensionality of the data input to the classifier.}
\label{tbl:featurecombis}
\end{table}
To test for significant differences in the performance statistics, we applied a general linear model (GLM)
using the \texttt{lme4} package \citep{Bates:2014} for R 2.15.2 \citep{R}.
We focused primarily on the AUC for significance testing,
since the AUC and MAP statistics are related analyses of the same data.
Since AUC is bounded in the range [0,1], we applied the GLM in the logistic domain:
note that given the probabilistic interpretation of AUC,
the logistic model is equivalent to applying a linear GLM to odds ratios,
a fact which facilitates interpretation.
Every experimental run for a given dataset used the same set of folds, so we used a repeated-measures version of the GLM with the ``fold index'' as the grouping variable.
We tested for individual and pairwise interactions of our five independent categorical variables, which were as follows:
\begin{itemize}
\item choice of feature set and temporal summarisation method, testing the 12 configurations listed in Table \ref{tbl:featurecombis}:
\item noise reduction on vs.\ off;
\item classifier mode (multilabel vs.\ either single-label or binary-relevance);
\item decision pooling window duration (1, 5, or 60 seconds or whole audio);
\item decision pooling max vs.\ mean.
\end{itemize}
Combinatorial testing of all these configurations resulted in $12 \times 2 \times 2 \times 4 \times 2 = 384$ crossvalidated classification experiments for the \texttt{bldawn} and \texttt{xccoverbl} datasets.
For the other datasets we did not test all four pooling durations, for reasons given above:
the number of crossvalidated experiments was thus $192$ for \texttt{nips4b} and $96$ for \texttt{lifeclef2014}.
Since the tests of \texttt{lifeclef2014} did not vary decision pooling, decision pooling factors were not included in that GLM.
We considered effects to be significant when the 95\% confidence interval calculated from the GLM excluded zero,
in which cases we report the estimated effects as differences in odds-ratios.
\subsection{Additional tests}
The \textit{bldawn} dataset has relatively few annotations, since it only consists of 60 items.
We therefore wanted to explore the use of auxiliary information from other sources to help improve recognition quality,
in particular using the \textit{xccoverbl} dataset, which has strong overlap in the list of species considered.
In further tests we tested three ways of using this additional data:
\begin{enumerate}
\item \textbf{Cross-condition training}, meaning training on one dataset and testing on the other.
The two datasets have systematic differences---for example, \textit{xccoverbl} items are annotated with only one species each, and are generally shorter---and so we did not expect this to yield very strong results.
\item \textbf{Data augmentation} for the feature learning step, meaning that feature learning is conducted using the training data for the \textit{bldawn} as well as all of the \textit{xccoverbl} data. This gives a larger and more varied pool of data for the feature learning step, which we expected to give a slight improvement to the results of feature learning.
\item \textbf{Data augmentation} for feature learning and also for training. Although the systematic differences mean the \textit{xccoverbl} training data might not guide the classifier in the correct way,
it holds many more species annotations for the species of interest, in a wider set of conditions.
We expected the combined training would provide stronger generalisation performance.
\end{enumerate}
We evaluated these train/test conditions as separate evaluation runs.
We also wanted to distinguish between two possible explanations for any difference between the performance of the different feature sets:
was it due to intrinsic characteristics of the features, or more simply due to the dramatic differences in feature dimensionality (which ranged 26 to 1000; see Table \ref{tbl:featurecombis})?
Differences in dimensionality might potentially give more degrees of freedom to the classifier without necessarily capturing information in a useful way.
In order to test this, we ran a version of our test in which for each experimental run we created a random projection matrix which projected the feature set to a fixed dimensionality of 200.
For MFCC/Mel features this was a simple form of data expansion,
while for learned features it was a form of data reduction.
By standardising the feature dimensionality,
this procedure decoupled the nature of the feature set from the degrees of freedom available to the classifier.
We ran this test using the \textit{nips4b} dataset.
\section{Results}
\label{sec:results}
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_xc3fold_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_xc3fold_rfall_marg}
\\
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_lifeclef2014xvsy_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_lifeclef2014xvsy_rfall_marg}
\end{center}
\caption{AUC and MAP statistics, summarised for each feature-type tested---here for the two single-label datasets, %
using the full multilabel classifier. %
Each column in the boxplot summarises the crossvalidated scores attained over many combinations of the other configuration settings tested %
(for the full multi-class classifier only). %
The ranges indicated by the boxes therefore do not represent random variation due to training data subset, but systematic variation due to classifier configuration. %
Figure \ref{fig:resultsfeatm} plots the same for the multilabel datasets. %
}
\label{fig:resultsfeats}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_nips4bXvsY_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_nips4bXvsY_rfall_marg}
\\
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_bl3fold_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_bl3fold_rfall_marg}
\end{center}
\caption{As Fig. \ref{fig:resultsfeats}, but for the two multilabel datasets, with the multilabel classifier.}
\label{fig:resultsfeatm}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_nips4bXvsY_rfeach_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_nips4bXvsY_rfeach_marg}
\\
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_bl3fold_rfeach_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_bl3fold_rfeach_marg}
\end{center}
\caption{As Fig. \ref{fig:resultsfeats}, but for the two multilabel datasets, with the binary relevance classifier.}
\label{fig:resultsfeatrfeach}
\end{figure}
\begin{table}[t]
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{ l l | r r r r }
Factor & Factor value & \texttt{nips4b} & \texttt{xccoverbl} & \texttt{bldawn} & \texttt{lifeclef2014} \\
\hline
\hline
featureset & {mfcc-maxp} & \textbf{* -0.59} & \textbf{* -0.29} & -0.03 & \textbf{* -0.30} \\
(vs.\ {mfcc-ms}) & {mfcc-modul} & \textbf{* -0.55} & \textbf{* -0.45} & \textbf{* -0.12} & \textbf{* -0.27} \\
& {melspec-ms} & \textbf{* 0.10} & \textbf{* 1.01} & \textbf{* -0.04} & \textbf{* 0.73} \\
& {melspec-maxp} & -0.01 & \textbf{* 0.82} & -0.03 & \textbf{* 0.52} \\
& {melspec-modul} & 0.02 & \textbf{* 0.82} & -0.03 & \textbf{* 0.67} \\
& {melspec-kfl1-ms} & \textbf{* 0.26} & \textbf{* 1.43} & 0.01 & \textbf{* 0.86} \\
& {melspec-kfl2-ms} & \textbf{* 0.25} & \textbf{* 1.36} & -0.02 & \textbf{* 0.90} \\
& {melspec-kfl3-ms} & \textbf{* 0.23} & \textbf{* 1.44} & -0.00 & \textbf{* 0.92} \\
& {melspec-kfl4-ms} & \textbf{* 0.20} & \textbf{* 1.40} & -0.00 & \textbf{* 0.91} \\
& {melspec-kfl8-ms} & \textbf{* 0.21} & \textbf{* 1.39} & -0.01 & \textbf{* 0.91} \\
& {melspec-kfl4pl8kfl4-ms} & \textbf{* -0.36} & \textbf{* 1.40} & -0.00 & \textbf{* 0.95} \\
\hline
noisered. & {on} & \textbf{* -0.64} & \textbf{* -0.20} & -0.01 & \textbf{* 0.06} \\
\hline
pooldur & {1} & \textbf{* -0.23} & \textbf{* -0.15} & -0.02 & \\
(vs.\ none) & {5} & & \textbf{* -0.15} & -0.04 & \\
& {60} & & 0.04 & -0.00 & \\
\hline
dpoolmode & {mean} & 0.03 & \textbf{* 0.07} & 0.01 & \\
\hline
classif & {binary relevance} & \textbf{* -0.28} & & \textbf{* -0.05} & \\
(vs.\ multi) & {single-label} & & 0.05 & & 0.01 \\
\end{tabular}
}
\end{center}
\caption{First-order effect sizes, estimated by our GLM as linear changes to the AUC odds ratio. %
Significant results are marked in bold with an asterisk, %
judged relative to a baseline category indicated in the first column. %
Positive values represent an improvement over the baseline. %
Empty cells indicate combinations that were not tested, as described in the text. %
}
\label{tbl:effectsizes}
\end{table}
Recognition performance was generally strong (Fig. \ref{fig:resultsfeats}, Fig. \ref{fig:resultsfeatm}, Table \ref{tbl:effectsizes}), given the very large number of classes to distinguish (at least 77).
The AUC and MAP performance measures both led to very similar rankings in our experiments.
The strongest effect found in our tests was the effect of feature type,
with a broad tendency for MFCCs to be outperformed by Mel spectra, and both of these outperformed by learned features.
For the largest dataset, \textit{lifeclef2014}, feature learning led to classification performance up to 85.4\% AUC,
whereas without feature learning the performance peaked at 82.2\% for raw Mel spectra or 69.3\% for MFCCs.
This pattern was clear for all datasets except \textit{bldawn}.
Compared against the baseline standard configuration \texttt{mfcc-ms},
switching to learned features provided all the strongest observed boosts in recognition performance
(Table \ref{tbl:effectsizes}).
The effect was particularly strong for the two single-label datasets, \textit{xccoverbl} and \textit{lifeclef2014} (effect size estimates $\geq 0.86$ for all feature-learning variants).
For \textit{nips4b} there was a milder effect ($\approx 0.25$),
except for the two-layer version which had a significant negative effect (-0.36).
Conversely, the two-layer version achieved strongest performance on the largest dataset (\textit{lifeclef2014}).
These facts together suggest that the performance impairment for \textit{nips4b} was due to the relatively small size of the dataset,
since deeper models typically require more data \citep{Coates:2012}.
That aside, the performance differences between variants of our feature-learning method were small.
For all datasets except \textit{bldawn},
the switch from MFCCs to raw Mel spectral features also provided a strong boost in performance,
though not to the same extent as did the learned features.
Across those three datasets, mean-and-standard-deviation summarisation consistently gave the strongest performance
over our two alternatives (i.e.\ maximum or modulation coefficients).
None of the above tendencies are discernible in the results for \textit{bldawn},
for which all methods attain the same performance.
The classifier can reach over 80\% AUC (50\% MAP), which is far above chance performance,
but not as strong as for the other datasets,
nor showing the pattern of differentiation between the configurations.
The relatively low scores reflect a relatively low ability to generalise,
demonstrated by the observation that the trained systems attained very strong scores when tested on their training data ($>99.95$\% in all cases, and 100\% in most).
This outcome is typical of systems trained with an amount of ground-truth annotations which is insufficient to represent the full range of variation within the classes.
The choice of classifier mode showed moderate but consistent effects across all the combinations tested.
For multilabel datasets, a decrease in AUC was observed
by switching to the ``binary relevance'' approach to classification.
Note however that this difference is more pronounced for AUC than for MAP (Fig. \ref{fig:resultsfeatrfeach}).
For single-label datasets, no significant effect was observed,
with a very small boost in AUC
by switching from the multilabel classifier to the single-label classifier.
Splitting the audio into decision windows and then combining the outcomes generally had a negative or negligible effect on outcomes;
however, using mean (rather than maximum) to aggregate such decisions had a mild positive effect (significant only for \textit{xccoverbl}).
Looking at the second-order interaction did not find any synergistic positive effects of using mean-pooling and a particular window length.
Activating noise reduction showed an inconsistent effect, significantly impairing performance on \textit{nips4b} (and to a lesser extent \textit{xccoverbl})
while slightly improving performance on \textit{lifeclef2014}.
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 67.7mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_bl3foldxconly_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 67.7mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_bl3foldxconly_rfall_marg}
\\
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 67.7mm 2.5mm 7.5mm]{plots/auceval_xval_featureset_bl3foldxcfl_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 67.7mm 2.5mm 7.5mm]{plots/mapateval_xval_featureset_bl3foldxcfl_rfall_marg}
\\
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 7.5mm]{plots/auceval_xval_featureset_bl3foldxc_rfall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 7.5mm]{plots/mapateval_xval_featureset_bl3foldxc_rfall_marg}
\end{center}
\caption{As Fig. \ref{fig:resultsfeats}, for the \textit{bldawn} dataset, except that the input data for training is either: %
(a) a different dataset, \textit{xccoverbl} (cross-condition scenario; upper); %
(b) the union of the two datasets for feature learning, but \textit{bldawn} alone for training (middle); %
(c) the union of the two datasets for feature learning and for training (lower); %
}
\label{fig:resultsfeatblxc}
\end{figure}
Our follow-up data expansion tests and cross-condition test failed to improve performance for \textit{bldawn} (Fig. \ref{fig:resultsfeatblxc}).
Adding the \textit{xccoverbl} data to the feature learning step made little difference,
giving a slight but insignificant boost to the two-layer model.
This tells us firstly that the dataset already contained enough audio for feature-learning to operate satisfactorily,
and secondly that the audio from \textit{xccoverbl} is similar enough in kind that its use is no detriment.
However, the story is quite different for the cases in which we then included \textit{xccoverbl} in the training step.
With or without feature learning, using \textit{xccoverbl} to provide additional training data for \textit{bldawn} acted as a distractor in this particular classification setup, and performed uniformly poorly.
Note that the data expansion procedure augmented the audio data size by around 60\%,
but augmented the number of annotations even more substantially (since \textit{xccoverbl} contains more individual items than \textit{bldawn}),
and so the classifier may have been led to accommodate the single-label data better than the dawn chorus annotations.
We diagnosed the problem by looking at the classification quality that the classifiers attained on their \textit{bldawn} training examples:
in this case the quality was poor (AUC $<60\%$),
confirming that the single-label \textit{xccoverbl} data had acted as distractors rather than additional educational examples.
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/auceval_xval_featureset_nips4bXvsY_rfrpall_marg}
\includegraphics [width=0.49\textwidth,clip, trim=2.5mm 2.5mm 2.5mm 2.5mm]{plots/mapateval_xval_featureset_nips4bXvsY_rfrpall_marg}
\end{center}
\caption{As Fig. \ref{fig:resultsfeats}, for the \textit{nips4b} dataset, except that the feature dimensionality is standardised before train/test by the application of a random projection.}
\label{fig:resultsfeatrfrpall}
\end{figure}
Turning to the \textit{nips4b} dataset to explore the effect of feature dimensionality,
the relative performance of the different feature types was broadly preserved even after dimensionality was standardised by random projection
(Fig. \ref{fig:resultsfeatrfrpall}).
The general effect of the random projection was a modest improvement for low-dimensional features,
and a modest impairment for high-dimensional (learned) features,
but not to the extent of changing the ordering of performance.
This suggests that high dimensionality is a small,
though non-zero, part of what lends the learned features their power.
The overall effect estimated by the GLM for the random-projection modification was a small but significant impairment ($-0.07$).
\begin{table}[t]
\resizebox{\textwidth}{!}{
\begin{tabular}{ l | r r }
System variant submitted & Cross-validated MAP (\%) & Final official MAP (\%) \\
\hline
\texttt{melspec-kfl3-ms}, noise red., binary relevance & 30.56 & 36.9 \\
Average from 12 single-layer models & 32.73 & 38.9 \\
\texttt{melspec-kfl4pl8kfl4-ms}, noise red., binary relevance & \textbf{35.31} & \textbf{42.9} \\
Average from 16 single- and double-layer models & 35.07 & 41.4 \\
\end{tabular}
}
\caption{Summary of MAP scores attained by our system in the public LifeCLEF 2014 Bird Identification Task \citep{birdclef2014}. %
The first column lists scores attained locally in our two-fold \textit{lifeclef2014} split. %
The second column lists scores evaluated officially, using a classifier(s) trained across the entire training set. %
}
\label{tbl:lifeclefofficial}
\end{table}
We can compare our results against those obtained recently by others.
A formal comparison was conducted in Spring 2014 when we submitted decisions from our system to the LifeCLEF 2014 bird identification challenge
\citep{birdclef2014}.
In that evaluation, our system attained by far the strongest audio-only classification results,
with a MAP peaking at 42.9\% (Table \ref{tbl:lifeclefofficial}).
(Only one system outperformed ours, peaking at 51.1\% in a variant of the challenge which provided additional metadata as well as audio.)
We submitted the outputs from individual models, as well as model-averaging runs using the simple mean of outputs from multiple models.
Notably, the strongest classification both in our own tests and the official evaluation was attained not by model averaging,
but by a single model based on two-layer feature learning.
Also notable is that our official scores, which were trained and tested on larger data subsets,
were substantially higher than our crossvalidated scores,
corroborating our observation that the method works particularly well at high data volumes.
Considering the \textit{nips4b} dataset,
the peak result from our main tests reached a crossvalidated AUC of 89.8\%.
In the actual NIPS4B contest (conducted before our current approach was developed), the winning result attained 91.8\%;
\cite{Potamitis:2014}, developing further a model submitted to the contest, reports a peak AUC of 91.7\%.
Our results are thus slightly behind these, although note that these other reported results use the full public-and-private datasets, without crossvalidation,
whereas we restricted ourselves only to the data that were fully public
and divided this public data into two crossvalidation folds, so the comparison is not strict.
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.65\textwidth,clip, trim=0mm 0mm 0mm 0mm]{plots/plottimestaken}
\end{center}
\caption{Times taken for each step in the process, for the \textit{lifeclef2014} dataset. %
Note that these are heuristic ``wallclock'' times measured on processes across two compute servers, %
and disk read/write processes (to store state) took non-trivial time in each step. %
Each measurement is averaged across the two folds and across two settings (noise reduction on/off) across the runs using the multilabel classifier and no decision-pooling. %
}
\label{fig:timestaken}
\end{figure}
We measured the total time taken for each step in our workflow,
to determine the approximate computational load for the steps (Fig. \ref{fig:timestaken}).
The timings are approximate---in particular because our code was modularised to save/load state on disk between each process,
which impacted particularly on the ``classify'' step which loaded large random forest settings from disk before processing.
Single-layer feature learning was efficient, taking a similar amount of time as did the initial feature extraction.
Double-layer feature learning took more than double this, because of the two layers as well as performing max-pooling downsampling.
Training the random forest classifier took longer on the learned features due to the higher dimensionality.
However, once the system was trained, the time taken to classify new data was the same across all configurations.
\section{Discussion}
\label{sec:disc}
In order to be of use for applications in ecology and archival,
automatic bird species recognition from sound must work across large data volumes,
across large numbers of potential species,
and on data with a realistic level of noise and variation.
Our experiments have demonstrated that very strong results can be achieved in exactly these cases
by supplementing a classification workflow with unsupervised feature learning.
We have here used a random forest classifier, but unsupervised feature learning operates without any knowledge
of the classifier or even the training labels, so we can expect this finding to apply in other classification systems (cf.\ \cite{Erhan:2010}).
The procedure requires large data volumes in order for benefits to be apparent,
as indicated by the lack of improvement on our \textit{bldawn} dataset,
and by the failure of two-layer feature learning on the \textit{nips4b} dataset.
However, the use of single-layer feature learning creates a classifier that is equivalent to or better than manually-designed features in all our tests.
There were very few differences in performance between our different versions of feature learning.
One difference is that two-layer feature learning, while unsuccessful for \textit{nips4b},
led to the strongest performance for \textit{lifeclef2014} which is the largest dataset considered%
---largest by an order of magnitude in data volume, and by almost an order of magnitude in the number of possible species labels.
This confirms the recommendations of \cite{Coates:2012} about the synergy between feature learning and big data scales,
here for the case of ecological audio data.
However, note that a lesser but still substantial improvement over the baseline MFCC system can usually be attained simply by using the raw Mel spectral data as input rather than MFCCs.
One of the long-standing motivations for the MFCC transformation has been to reduce spectral data down to a lower dimensionality while hoping to preserve most of the implicit semantic information;
but as we have seen, the random forest classifier performs well with high-dimensional input,
and such data reduction is not necessary and often holds back classification performance.
Future investigators should consider using Mel spectra as a baseline, rather than MFCCs as is common at present.
The lack of improvement on the \textit{bldawn} dataset is notable,
along with the low-quality results obtained by simply using a different dataset for training, or augmenting the data with an additional dataset.
The availability of audio data is not the issue here, as the dataset is second-largest by audio volume,
although the availability of training annotations may be crucial, since the dataset is the smallest by an order of magnitude
in terms of individual labelled items.
Augmenting the data with \textit{xccoverbl} increased the amount of training annotations,
but its failure to boost performance may have been due to differences in kind,
since it consisted of single-label recordings of individual birds rather than multilabel recordings of dawn chorus soundscapes.
The issue is not due to differences in recording conditions (such as microphone type or recording level):
the audio went through various levels of standardisation in our workflow,
and caused no problems when added to the feature-learning step only.
We emphasise that the audio and metadata of the \textit{bldawn} dataset comes directly from a sound archive collection,
and the long-form recordings with sparse annotation are exactly the format of a large portion of the holdings in sound archives.
Our results (up to around 80\% AUC) correspond to a classifier that could provide useful semi-automation of archive labelling
(e.g.\ suggested labels to a human annotator) but not yet a fully automatic process in that case.
This outcome thus reinforces the importance of collecting and publishing annotated audio datasets which fit the intended application.
Public collections such as Xeno Canto are highly valuable, but their data do not provide the basis to solve all tasks in species recognition,
let alone the other automated tasks (censusing, individual recogntion) that we may wish to perform automatically from audio data.
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.95\textwidth,clip, trim=15mm 20mm 40mm 0mm]{plots/kfldata_melspec-kfl4-lifeclef2014y_nr0_pk0_heq0_page3annot}
\end{center}
\caption{A random subset of the spectrotemporal bases estimated during one run of feature learning, in this case using 4 frames per base and the \textit{lifeclef2014} dataset. %
Each base is visualised as a brief spectrogram excerpt, with dark indicating high values. %
The frequency axis is nonlinearly (Mel) scaled. %
}
\label{fig:basesexample}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics [width=0.65\textwidth,clip, trim=35mm 15mm 60mm 10mm]{figs/hausberger00fig4annot}
\end{center}
\caption{Spectro-temporal receptive fields (STRFs) measured from individual neurons in auditory field L of starling. %
Adapted from \cite{Hausberger:2000} for comparison with Fig. \ref{fig:basesexample}. %
Each image shows the spectro-temporal patterns that correlate with excitation (left) and inhibition (right) of a single neuron. %
The frequency axis is linearly scaled. %
}
\label{fig:hausbergerfig4}
\end{figure}
Our feature-based approach to classification is not the only approach.
Template-based methods have some history in the literature, with the main issue of concern being how to match a limited set of templates against the unbounded natural variation in bird sound realisations, in particular the dramatic temporal variability.
One technique to compensate for temporal variability is dynamic time warping (DTW)
\citep{Anderson:1996
\citep{Ito:1999
. Recent methods which performed very strongly in the SABIOD-organised contests used templates without any time-warping considerations,
making use of a large number of statistics derived from the match between a template and an example (not using just the closeness-of-match)
\citep{Fodor:2013
. Other recent methods use templates and DTW but deployed within a kernel-based distance measure,
again going beyond a simple one-to-one match \citep{Damoulas:2010}.
In light of these other perspectives, we note an analogy with our learned features.
In the workflow considered here, the new representation is calculated by taking the dot-product between the input data and each base
(such as those in Fig. \ref{fig:basesexample}), as given in \eqref{eqn:dotprodmultiframe}.
The form of \eqref{eqn:dotprodmultiframe} is the same mathematically as spectro-temporal cross-correlation,
where the $b_j$ would be thought of more traditionally as ``templates''.
Our features are thus equivalent to the output of an unusual kind of template matching by cross-correlation,
where the ``templates'' are not indivudal audio excerpts but generalisations of features found broadly across audio excerpts,
and are also of a fixed short duration (shorter than many song syllables, though long enough to encompass many calls).
A question that arises from this perspective is whether our approach should use longer series of frames, long enough to encompass many types of song syllable entirely.
In our tests we found no notable tendency for improved recognition as we increased the number of frames from one to eight,
and we also saw many temporally-compact bases learnt (Fig. \ref{fig:basesexample}),
so we do not believe lengthening the bases is the route to best performance.
Further, the advantage of using relatively short durations is that the feature learning method learns \textit{components} of bird vocalisations rather than over-specific whole units.
These components may co-occur in a wide variety of bird sounds, in temporally-flexible orders,
conferring a combinatorial benefit of broad expressivity.
Our two-layer feature learning provides a further level of abstraction over temporal combinations of energy patterns,
which is perhaps part of its advantage when applied to our largest dataset.
We have not explicitly tested our method in comparison to template-based approaches;
the relative merits of such approaches will become clear in further research.
The bases shown in Fig. \ref{fig:basesexample} also bear some likeness with the spectro-temporal receptive fields (STRFs) measured from neurons found in the early auditory system of songbirds
(e.g.\ Fig. \ref{fig:hausbergerfig4}, adapted from \cite{Hausberger:2000}
. Broadly similar generalisations seem to emerge, including sensitivity to spectrally-compact stationary tones as well as up-chirps and down-chirps, and combinations of harmonic sounds.
We do not make strong claims from this likeness:
firstly because our method (spherical k-means) is a simple sparse feature learning method with no designed resemblance to neural processes involved in natural learning,
and secondly because STRFs show only a partial summary of the nonlinear response characteristics of neurons.
However, sparse feature learning in general is motivated by considerations of the informational and energetic constraints that may have influenced the evolution of neural mechanisms \citep{Olshausen:2004}.
\cite{Theunissen:2006}
note that the sensitivities measured from neurons in the avian primary auditory forebrain generally relate not to entire song syllables, but to smaller units which may serve as building blocks for later processing.
Biological analogies are not a necessary factor in the power of machine learning methods,
but such hints from neurology suggest that the method we have used in this study fits within a paradigm that may be worth further exploration.
\section{Conclusions}
\label{sec:conc}
Current interest in automatic classification of bird sounds is motivated by the practical scientific need to label
large volumes of data coming from sources such as remote monitoring stations and sound archives.
Unsupervised feature learning is a simple and effective method to boost classification performance
by learning spectro-temporal regularities in the data.
It does not require training labels or any other side-information,
it can be used within any classification workflow,
and once trained it imposes negligible extra computational effort on the classifier.
In experiments it learns regularities in bird vocalisation data with similar qualities to the sensitivities of bird audition reported by others.
The principal practical issue with unsupervised feature learning is that it requires large data volumes to be effective,
as confirmed in our tests.
However, this exhibits a synergy with the large data volumes that are increasingly becoming standard.
For our largest dataset, feature learning led to classification performance up to 85.4\% AUC,
whereas without feature learning the performance peaked at 82.2\% for raw Mel spectra or 69.3\% for MFCCs.
In our tests, the choice of feature set made a much larger difference to classification performance than any of our other configuration choices
(such as the use of noise reduction, decision pooling, or binary relevance).
MFCCs cannot be recommended for bird species recognition since even the simple approach of using undigested Mel spectra dramatically outperforms them in most of our tests.
Although MFCCs have been widespread as baseline features, the Mel spectra themselves may be more appropriate for benchmarking.
Across our various tests in single-label and multilabel settings,
unsupervised feature learning together with a multilabel classifier achieved peak or near-peak classification quality.
This study, thanks to the large-scale data made available by others,
has demonstrated strong performance on bird sound classification is possible at very large scale, when the synergy between big data volumes and feature learning is exploited.
However, automatic classification is not yet trivial across all domains, as demonstrated by the lack of improvement on our \textit{bldawn} dataset of dawn chorus recordings.
The research community will benefit most from the creation/publication of large bird audio collections,
labelled or at least part-labelled, and published under open data licences.
\section*{Acknowledgments}
We would like to thank the people and projects which made available the data used for this research:
the Xeno Canto website and its many volunteer contributors;
the SABIOD research project and the Biotope society;
the British Library Sound Archive and its contributors, and curator Cheryl Tipp.
\section*{Funding sources}
This work was supported by EPSRC Leadership Fellowship EP/G007144/1 and EPSRC Early Career Fellowship EP/L020505/1.
\section*{Data availability}
\begin{itemize}
\item The \textit{xccoverbl} data is archived at \url{https://archive.org/details/xccoverbl_2014} --- it is composed of sound files sourced from \url{http://www.xeno-canto.org/}. The sound file authors and other metadata are listed in a CSV file included in the dataset.
\item The \textit{bldawn} dataset is available on request from the British Library Sound Archive (BLSA).
Our machine-readable version of the species metadata is downloadable from Figshare,
and lists the file identifiers corresponding to the BLSA records:
\url{http://dx.doi.org/10.6084/m9.figshare.1024549}
\item The \textit{nips4b} dataset is downloadable from the SABIOD/nips4b website:
\url{http://sabiod.univ-tln.fr/nips4b/challenge1.html}
\item The \textit{lifeclef2014} dataset is available from the Lifeclef website:
\url{http://www.imageclef.org/2014/lifeclef/bird}
\end{itemize}
\bibliographystyle{plainnat}
|
Subsets and Splits